Merge tag 'sched_ext-for-7.3' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext

Pull sched_ext updates from Tejun Heo:
 "Most of this cycle completes the enqueue-path support for hierarchical
  sub-scheduling, which makes sub-scheduler support feature complete: a
  root BPF scheduler can now hand a cgroup subtree over to a nested
  sub-scheduler together with revocable CPU grants, and the
  sub-scheduler owns all scheduling decisions for its tasks on those
  CPUs.

  Development volume was high and a number of changes plugging holes in
  the new support landed late in the cycle. Also included are core
  scheduling fixes that were completed too late for the v7.2 release and
  are routed through this pull request.

  Sub-scheduler CPU delegation:

   - Parent schedulers now grant and revoke per-CPU capabilities
     (enqueueing, preemption, CPU frequency control) on their children,
     enforced on every path a scheduler can reach a CPU through.
     Previously only dispatching could be delegated; this lets
     sub-schedulers fully schedule their CPUs.

   - Rescue execution: a task whose scheduler doesn't have access to the
     CPUs the task needs to run on starved until the watchdog ejected
     the whole scheduler. The kernel now runs such tasks directly on a
     small bandwidth budget, turning a scheduler-killing failure into
     bounded degradation.

   - Cgroup integration: tasks migrating across a sub-scheduler boundary
     weren't re-homed to the new owner, causing wrong-scheduler
     scheduling and a use-after-free. Sub-schedulers now take over their
     cgroup subtree and receive its cgroup callbacks.

   - Arena objects now cross the kernel/BPF boundary as typed pointer
     arguments, translated transparently by the BPF tree's new arena
     argument support, replacing untyped arguments with manual
     translation.

   - scx_qmap now demonstrates full hierarchical sub-scheduling.

  Other fixes and updates:

   - Robustness improvements: the abort path is now NMI-safe, fixing
     deadlocks when errors are raised from NMI context and making
     hardlockup recovery direct. Reenqueue loops that could monopolize a
     CPU ahead of the watchdog now eject the offending scheduler, and
     stalls are blamed on the scheduler actually responsible.

   - Hardening: BPF-writable arena memory is validated before kernel
     use, and task slice and vtime writes got explicit synchronization
     rules, closing corruption vectors open to buggy or malicious
     schedulers.

   - Core scheduling: sched_ext dispatching can drop the rq lock inside
     the core-wide pick, which let interleaving selections corrupt each
     other's state and hard-hang the machine. The selection now restarts
     when the lock was released. The task ordering callback was also
     invoked with its arguments swapped, and the default ordering is
     updated to work across sub-scheduler boundaries. The fixes are
     marked for stable.

   - Other fixes headed for stable: a task init leak on fork failure
     during enable, tooling compat macros that silently failed to detect
     newer kernels, and a crash on reenqueueing against a destroyed
     dispatch queue.

   - Tooling: scx_pair moves off deprecated callbacks, and the
     deprecated scx_bpf_cpu_rq() kfunc is removed"

* tag 'sched_ext-for-7.3' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext: (144 commits)
  sched_ext: Drop the dead SCX_DEQ_CORE_SCHED_EXEC test in dequeue_task_scx()
  sched_ext: Make core-sched task ordering hierarchy-aware
  sched_ext: Use runnable_at for the default core-sched task ordering
  sched_ext: Fix inverted ops.core_sched_before() invocation
  sched_ext: Move the config-off sub-cap kfunc stubs into sub.c
  sched_ext: Rename balance-era identifiers to dispatch terms
  sched_ext: Drop the stale keep_prev fixup in dispatch_pick()
  sched_ext: Keep kick_sync waiting on the rq's own CPU
  sched_ext: Make SCHED_CLASS_EXT select GENERIC_ALLOCATOR
  sched_ext/scx_flatcg: Fix cvtime true-up on slice expiry
  sched_ext: Don't BUG_ON a destroyed DSQ in process_deferred_reenq_users
  sched_ext: Fix scx_bpf_dsq_move_to_local___v2 compat detection
  sched_ext: Make scx_bpf_events() read the calling scheduler's counters
  sched_ext: Drop unlocked scx_rq_clock_invalidate() from scx_root_disable()
  selftests/sched_ext: Fix flaky ddsp failure tests on busy systems
  selftests/sched_ext: Make numa idle validation race-free
  sched_ext: Fix scx_bpf_dsq_reenq___compat kfunc extern prototype
  sched_ext/scx_flatcg: expire cached hweights on weight changes
  sched_ext: Fix exit_task leak on fork failure during enable
  sched_ext: fix stale references in doc comments
  ...

19 files changed, 6886 insertions, 2228 deletions

diff --git a/kernel/Kconfig.preempt b/kernel/Kconfig.preempt
index 35f546a042b1..f294dad43bd7 100644
--- a/kernel/Kconfig.preempt
+++ b/kernel/Kconfig.preempt
@@ -173,6 +173,7 @@ config SCHED_CORE
 config SCHED_CLASS_EXT
 	bool "Extensible Scheduling Class"
 	depends on BPF_SYSCALL && BPF_JIT && DEBUG_INFO_BTF
+	select GENERIC_ALLOCATOR
 	select STACKTRACE if STACKTRACE_SUPPORT
 	help
 	  This option enables a new scheduler class sched_ext (SCX), which
diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 98c536f8b666..c3a12fee7528 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -88,6 +88,8 @@ EXPORT_SYMBOL_GPL(css_set_lock);
 
 struct blocking_notifier_head cgroup_lifetime_notifier =
 	BLOCKING_NOTIFIER_INIT(cgroup_lifetime_notifier);
+struct blocking_notifier_head cgroup_task_notifier =
+	BLOCKING_NOTIFIER_INIT(cgroup_task_notifier);
 
 DEFINE_SPINLOCK(trace_cgroup_path_lock);
 char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
@@ -2676,14 +2678,27 @@ struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
 	return NULL;
 }
 
+static void cgroup_migrate_notify_canceled(struct css_set *src_cset,
+					   struct task_struct *task)
+{
+	struct cgroup_task_migrate_ctx ctx = {
+		.task = task,
+		.src_dcgrp = src_cset->dfl_cgrp,
+		.dst_dcgrp = src_cset->mg_dst_cset->dfl_cgrp,
+	};
+
+	blocking_notifier_call_chain(&cgroup_task_notifier,
+				     CGROUP_TASK_MIGRATE_CANCELED, &ctx);
+}
+
 /**
  * cgroup_migrate_execute - migrate a taskset
  * @mgctx: migration context
  *
- * Migrate tasks in @mgctx as setup by migration preparation functions.
- * This function fails iff one of the ->can_attach callbacks fails and
- * guarantees that either all or none of the tasks in @mgctx are migrated.
- * @mgctx is consumed regardless of success.
+ * Migrate tasks in @mgctx as setup by migration preparation functions. This
+ * function fails iff one of the ->can_attach callbacks or CGROUP_TASK_MIGRATING
+ * notifications fails and guarantees that either all or none of the tasks in
+ * @mgctx are migrated. @mgctx is consumed regardless of success.
  */
 static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
 {
@@ -2691,6 +2706,7 @@ static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
 	struct cgroup_subsys *ss;
 	struct task_struct *task, *tmp_task;
 	struct css_set *cset, *tmp_cset;
+	bool dfl_migration = false;
 	int ssid, failed_ssid, ret;
 
 	/* check that we can legitimately attach to the cgroup */
@@ -2708,6 +2724,33 @@ static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
 	}
 
 	/*
+	 * Notify each task about the impending migration. An error return fails
+	 * the migration. Only migrations on the default hierarchy are reported:
+	 * a migration modifies either every moved task's dfl cgroup or, on
+	 * cgroup1 or for subtree_control writes, none.
+	 */
+	list_for_each_entry(cset, &tset->src_csets, mg_node) {
+		if (cset->dfl_cgrp == cset->mg_dst_cset->dfl_cgrp)
+			continue;
+		dfl_migration = true;
+		list_for_each_entry(task, &cset->mg_tasks, cg_list) {
+			struct cgroup_task_migrate_ctx ctx = {
+				.task = task,
+				.src_dcgrp = cset->dfl_cgrp,
+				.dst_dcgrp = cset->mg_dst_cset->dfl_cgrp,
+			};
+
+			ret = blocking_notifier_call_chain_robust(&cgroup_task_notifier,
+								  CGROUP_TASK_MIGRATING,
+								  CGROUP_TASK_MIGRATE_CANCELED,
+								  &ctx);
+			ret = notifier_to_errno(ret);
+			if (ret)
+				goto out_cancel_migrating;
+		}
+	}
+
+	/*
 	 * Now that we're guaranteed success, proceed to move all tasks to
 	 * the new cgroup.  There are no failure cases after here, so this
 	 * is the commit point.
@@ -2750,9 +2793,41 @@ static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
 		} while_each_subsys_mask();
 	}
 
+	/*
+	 * Notify each task after successful migration. The operation can no
+	 * longer fail and the return value is ignored. The MIGRATING loop
+	 * above explains why only dfl migrations are reported. Per-task
+	 * sources are not tracked past the commit point, so src_dcgrp is
+	 * NULL.
+	 */
+	if (dfl_migration) {
+		list_for_each_entry(cset, &tset->dst_csets, mg_node) {
+			list_for_each_entry(task, &cset->mg_tasks, cg_list) {
+				struct cgroup_task_migrate_ctx ctx = {
+					.task = task,
+					.dst_dcgrp = cset->dfl_cgrp,
+				};
+
+				blocking_notifier_call_chain(
+					&cgroup_task_notifier,
+					CGROUP_TASK_MIGRATED, &ctx);
+			}
+		}
+	}
+
 	ret = 0;
 	goto out_release_tset;
 
+out_cancel_migrating:
+	list_for_each_entry_continue_reverse(task, &cset->mg_tasks, cg_list)
+		cgroup_migrate_notify_canceled(cset, task);
+	list_for_each_entry_continue_reverse(cset, &tset->src_csets, mg_node) {
+		if (cset->dfl_cgrp == cset->mg_dst_cset->dfl_cgrp)
+			continue;
+		list_for_each_entry_reverse(task, &cset->mg_tasks, cg_list)
+			cgroup_migrate_notify_canceled(cset, task);
+	}
+	failed_ssid = CGROUP_SUBSYS_COUNT;
 out_cancel_attach:
 	if (tset->nr_tasks) {
 		do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
@@ -2976,11 +3051,11 @@ int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
  * cgroup_migrate_prepare_dst() on the targets before invoking this
  * function and following up with cgroup_migrate_finish().
  *
- * As long as a controller's ->can_attach() doesn't fail, this function is
- * guaranteed to succeed.  This means that, excluding ->can_attach()
- * failure, when migrating multiple targets, the success or failure can be
- * decided for all targets by invoking group_migrate_prepare_dst() before
- * actually starting migrating.
+ * As long as a controller's ->can_attach() or a CGROUP_TASK_MIGRATING
+ * notification doesn't fail, this function is guaranteed to succeed.  This
+ * means that, excluding those failures, when migrating multiple targets,
+ * the success or failure can be decided for all targets by invoking
+ * group_migrate_prepare_dst() before actually starting migrating.
  */
 int cgroup_migrate(struct task_struct *leader, bool threadgroup,
 		   struct cgroup_mgctx *mgctx)
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 03a43d3d2616..415583c35f8c 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -4900,6 +4900,9 @@ static void __init rcu_dump_rcu_node_tree(void)
 
 struct workqueue_struct *rcu_gp_wq;
 
+static struct cpumask rcu_stall_cpumask;
+static struct cpumask rcu_exp_stall_cpumask;
+
 void __init rcu_init(void)
 {
 	int cpu = smp_processor_id();
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index 82cada459e5d..46b6907f1b09 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -578,6 +578,7 @@ static void synchronize_rcu_expedited_stall(unsigned long jiffies_start, unsigne
 			if (!(READ_ONCE(rnp->expmask) & mask))
 				continue;
 			ndetected++;
+			cpumask_set_cpu(cpu, &rcu_exp_stall_cpumask);
 			rdp = per_cpu_ptr(&rcu_data, cpu);
 			pr_cont(" %d-%c%c%c%c", cpu,
 				"O."[!!cpu_online(cpu)],
@@ -665,6 +666,8 @@ static void synchronize_rcu_expedited_wait(void)
 		if (rcu_stall_is_suppressed())
 			continue;
 
+		cpumask_clear(&rcu_exp_stall_cpumask);
+
 		nbcon_cpu_emergency_enter();
 
 		j = jiffies;
@@ -675,7 +678,7 @@ static void synchronize_rcu_expedited_wait(void)
 
 		nbcon_cpu_emergency_exit();
 
-		panic_on_rcu_stall();
+		panic_on_rcu_stall(&rcu_exp_stall_cpumask);
 	}
 }
 
diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index cf7ae51cba40..ebf381936eb1 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -159,7 +159,7 @@ static int __init check_cpu_stall_init(void)
 early_initcall(check_cpu_stall_init);
 
 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
-static void panic_on_rcu_stall(void)
+static void panic_on_rcu_stall(const struct cpumask *stalled_mask)
 {
 	static int cpu_stall;
 
@@ -167,7 +167,7 @@ static void panic_on_rcu_stall(void)
 	 * Attempt to kick out the BPF scheduler if it's installed and defer
 	 * the panic to give the system a chance to recover.
 	 */
-	if (scx_rcu_cpu_stall())
+	if (scx_rcu_cpu_stall(stalled_mask))
 		return;
 
 	if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
@@ -644,6 +644,8 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 	if (rcu_stall_is_suppressed())
 		return;
 
+	cpumask_clear(&rcu_stall_cpumask);
+
 	nbcon_cpu_emergency_enter();
 
 	/*
@@ -659,6 +661,7 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 			for_each_leaf_node_possible_cpu(rnp, cpu)
 				if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 					print_cpu_stall_info(cpu);
+					cpumask_set_cpu(cpu, &rcu_stall_cpumask);
 					ndetected++;
 				}
 		}
@@ -700,7 +703,7 @@ static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 
 	nbcon_cpu_emergency_exit();
 
-	panic_on_rcu_stall();
+	panic_on_rcu_stall(&rcu_stall_cpumask);
 
 	rcu_force_quiescent_state();  /* Kick them all. */
 }
@@ -753,7 +756,9 @@ static void print_cpu_stall(unsigned long gp_seq, unsigned long gps)
 
 	nbcon_cpu_emergency_exit();
 
-	panic_on_rcu_stall();
+	cpumask_clear(&rcu_stall_cpumask);
+	cpumask_set_cpu(smp_processor_id(), &rcu_stall_cpumask);
+	panic_on_rcu_stall(&rcu_stall_cpumask);
 
 	/*
 	 * Attempt to revive the RCU machinery by forcing a context switch.
diff --git a/kernel/sched/build_policy.c b/kernel/sched/build_policy.c
index d74b54f81992..2a828725a7f9 100644
--- a/kernel/sched/build_policy.c
+++ b/kernel/sched/build_policy.c
@@ -66,10 +66,13 @@
 # include "ext/cid.h"
 # include "ext/arena.h"
 # include "ext/idle.h"
+# include "ext/sub.h"
+# include "ext/inlines.h"
 # include "ext/ext.c"
 # include "ext/cid.c"
 # include "ext/arena.c"
 # include "ext/idle.c"
+# include "ext/sub.c"
 #endif
 
 #include "syscalls.c"
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 6544e56925ee..4520d63763d4 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -443,6 +443,17 @@ static void __sched_core_flip(bool enabled)
 
 		sched_core_lock(cpu, &flags);
 
+		/*
+		 * A core-wide selection may have the shared rq lock temporarily
+		 * released by a lock-dropping ->pick_task(). Flipping would
+		 * rebind rq_lockp() under it. Wait it out.
+		 */
+		while (cpu_rq(cpu)->core->core_pick_in_flight) {
+			sched_core_unlock(cpu, &flags);
+			cpu_relax();
+			sched_core_lock(cpu, &flags);
+		}
+
 		for_each_cpu(t, smt_mask)
 			cpu_rq(t)->core_enabled = enabled;
 
@@ -6228,7 +6239,7 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
 	unsigned long cookie;
 	int i, cpu, occ = 0;
 	struct rq *rq_i;
-	bool need_sync;
+	bool need_sync = false;
 
 	if (!sched_core_enabled(rq))
 		return __pick_next_task(rq, rf);
@@ -6247,6 +6258,8 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
 		return __pick_next_task(rq, rf);
 	}
 
+	rq->core->core_pick_in_flight++;
+
 	/*
 	 * If there were no {en,de}queues since we picked (IOW, the task
 	 * pointers are all still valid), and we haven't scheduled the last
@@ -6271,7 +6284,9 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
 	prev_balance(rq, rf);
 
 	smt_mask = cpu_smt_mask(cpu);
-	need_sync = !!rq->core->core_cookie;
+
+restart:
+	need_sync |= !!rq->core->core_cookie;
 
 	/* reset state */
 	rq->core->core_cookie = 0UL;
@@ -6306,10 +6321,15 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
 	 * and there are no cookied tasks running on siblings.
 	 */
 	if (!need_sync) {
-restart_single:
 		next = pick_task(rq, rf);
-		if (unlikely(next == RETRY_TASK))
-			goto restart_single;
+		if (unlikely(next == RETRY_TASK)) {
+			/* rq lock may have been dropped, clocks invalidated */
+			core_clock_updated = false;
+			if (!(rq->clock_update_flags & RQCF_UPDATED))
+				update_rq_clock(rq);
+			goto restart;
+		}
+
 		if (!next->core_cookie) {
 			rq->core_pick = NULL;
 			rq->core_dl_server = NULL;
@@ -6329,7 +6349,6 @@ restart_single:
 	 *
 	 * Tie-break prio towards the current CPU
 	 */
-restart_multi:
 	max = NULL;
 	for_each_cpu_wrap(i, smt_mask, cpu) {
 		rq_i = cpu_rq(i);
@@ -6343,8 +6362,13 @@ restart_multi:
 			update_rq_clock(rq_i);
 
 		p = pick_task(rq_i, rf);
-		if (unlikely(p == RETRY_TASK))
-			goto restart_multi;
+		if (unlikely(p == RETRY_TASK)) {
+			/* rq lock may have been dropped, clocks invalidated */
+			core_clock_updated = false;
+			if (!(rq->clock_update_flags & RQCF_UPDATED))
+				update_rq_clock(rq);
+			goto restart;
+		}
 
 		rq_i->core_pick = p;
 		rq_i->core_dl_server = rq_i->dl_server;
@@ -6450,6 +6474,7 @@ restart_multi:
 	}
 
 out_set_next:
+	rq->core->core_pick_in_flight--;
 	put_prev_set_next_task(rq, rq->donor, next);
 	if (rq->core->core_forceidle_count && next == rq->idle)
 		queue_core_balance(rq);
@@ -6645,6 +6670,13 @@ static void sched_core_cpu_deactivate(unsigned int cpu)
 	core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
 
 	/*
+	 * A stale leftover would bias the count forever if this CPU later
+	 * returns as its own leader. Move, don't copy.
+	 */
+	core_rq->core_pick_in_flight       = rq->core_pick_in_flight;
+	rq->core_pick_in_flight            = 0;
+
+	/*
 	 * Accounting edge for forced idle is handled in pick_next_task().
 	 * Don't need another one here, since the hotplug thread shouldn't
 	 * have a cookie.
@@ -9058,6 +9090,7 @@ void __init sched_init(void)
 		rq->core_forceidle_count = 0;
 		rq->core_forceidle_occupation = 0;
 		rq->core_forceidle_start = 0;
+		rq->core_pick_in_flight = 0;
 
 		rq->core_cookie = 0UL;
 #endif
diff --git a/kernel/sched/ext/arena.c b/kernel/sched/ext/arena.c
index 5783694ec21d..f7a9f633f435 100644
--- a/kernel/sched/ext/arena.c
+++ b/kernel/sched/ext/arena.c
@@ -6,8 +6,7 @@
  *
  * Each chunk added to @sch->arena_pool comes from one
  * bpf_arena_alloc_pages_sleepable() call and is registered at the
- * kernel-side mapping address. Callers translate to the BPF-arena form
- * themselves if needed.
+ * kernel-side mapping address.
  *
  * Allocations grow the pool on demand. Underlying arena pages are released
  * when the arena map itself is torn down.
@@ -70,8 +69,6 @@ void scx_arena_pool_destroy(struct scx_sched *sch)
  */
 static int scx_arena_grow(struct scx_sched *sch, u32 page_cnt)
 {
-	u64 kern_vm_start;
-	u32 uaddr32;
 	void *p;
 	int ret;
 
@@ -83,15 +80,8 @@ static int scx_arena_grow(struct scx_sched *sch, u32 page_cnt)
 	if (!p)
 		return -ENOMEM;
 
-	uaddr32 = (u32)(unsigned long)p;
-	/* arena.o, which defines these, is built only on MMU && 64BIT */
-#if defined(CONFIG_MMU) && defined(CONFIG_64BIT)
-	kern_vm_start = bpf_arena_map_kern_vm_start(sch->arena_map);
-#else
-	kern_vm_start = 0;
-#endif
-
-	ret = gen_pool_add(sch->arena_pool, kern_vm_start + uaddr32,
+	ret = gen_pool_add(sch->arena_pool,
+			   (unsigned long)scx_arena_to_kaddr(sch, p),
 			   page_cnt * PAGE_SIZE, NUMA_NO_NODE);
 	if (ret) {
 		bpf_arena_free_pages_non_sleepable(sch->arena_map, p, page_cnt);
diff --git a/kernel/sched/ext/cid.c b/kernel/sched/ext/cid.c
index af83084ec740..39f88deb94bc 100644
--- a/kernel/sched/ext/cid.c
+++ b/kernel/sched/ext/cid.c
@@ -11,20 +11,26 @@
 #include "cid.h"
 
 /*
- * cid tables.
- *
- * Pointers are published once on first enable and never revoked. The default
- * mapping is populated before ops.init() runs; scx_bpf_cid_override() commits
- * before it returns. As long as the BPF scheduler only uses the tables from
- * those points onward, it sees a consistent view.
+ * cid tables. The cid kfuncs are available whether the root scheduler is
+ * cid-form or cpu-form, the latter to allow gradual migration to cids, so every
+ * root builds a default mapping. Each root enable allocates a fresh set, builds
+ * it privately and publishes the __rcu globals below once the layout is final.
+ * Root disable unpublishes and RCU-frees the set. kfuncs may run before the
+ * tables are published and must check for NULL.
  */
-s16 *scx_cid_to_cpu_tbl;
-s16 *scx_cpu_to_cid_tbl;
-struct scx_cid_topo *scx_cid_topo;
+u32 scx_nr_cid_shards;
+s16 __rcu *scx_cid_to_cpu_tbl;
+s16 __rcu *scx_cpu_to_cid_tbl;
+s32 __rcu *scx_cid_to_shard;
+s32 __rcu *scx_shard_node;
+struct scx_cid_shard __rcu *scx_cid_shard_ranges;
+struct scx_cid_topo __rcu *scx_cid_topo;
+
+static struct scx_cid_tables *scx_cid_tables;	/* used only during alloc/free */
 
 #define SCX_CID_TOPO_NEG	(struct scx_cid_topo) {				\
 	.core_cid = -1, .core_idx = -1, .llc_cid = -1, .llc_idx = -1,		\
-	.node_cid = -1, .node_idx = -1,						\
+	.node_cid = -1, .node_idx = -1, .shard_cid = -1, .shard_idx = -1,	\
 }
 
 /*
@@ -43,37 +49,128 @@ static const struct cpumask *cpu_llc_mask(int cpu, struct cpumask *fallbacks)
 	return &ci->info_list[ci->num_leaves - 1].shared_cpu_map;
 }
 
-/* Allocate the cid tables once on first enable; never freed. */
-static s32 scx_cid_arrays_alloc(void)
+/*
+ * Compute per-LLC shard layout. Each shard holds at most @shard_size cids, and
+ * in any case no more than SCX_CID_SHARD_MAX_CPUS. Cores are spread as evenly
+ * as possible across shards so cpu count is balanced: the first *@nr_large_p
+ * shards get (*@cores_per_shard_p + 1) cores, the rest get *@cores_per_shard_p.
+ */
+static void calc_shard_layout(const struct cpumask *llc_cpus, u32 shard_size,
+			      u32 *cores_per_shard_p, u32 *nr_large_p)
 {
-	u32 npossible = num_possible_cpus();
-	s16 *cid_to_cpu, *cpu_to_cid;
-	struct scx_cid_topo *cid_topo;
+	u32 nr_cores = 0, nr_cpus = 0, nr_shards;
+	int cpu;
 
-	if (scx_cid_to_cpu_tbl)
-		return 0;
+	for_each_cpu(cpu, llc_cpus) {
+		nr_cpus++;
+		if (cpumask_first(topology_sibling_cpumask(cpu)) == cpu)
+			nr_cores++;
+	}
 
-	cid_to_cpu = kzalloc_objs(*scx_cid_to_cpu_tbl, npossible, GFP_KERNEL);
-	cpu_to_cid = kzalloc_objs(*scx_cpu_to_cid_tbl, nr_cpu_ids, GFP_KERNEL);
-	cid_topo = kmalloc_objs(*scx_cid_topo, npossible, GFP_KERNEL);
+	nr_shards = max_t(u32, 1, DIV_ROUND_UP(nr_cpus, shard_size));
+	nr_shards = max_t(u32, nr_shards,
+			  DIV_ROUND_UP(nr_cpus, SCX_CID_SHARD_MAX_CPUS));
 
-	if (!cid_to_cpu || !cpu_to_cid || !cid_topo) {
-		kfree(cid_to_cpu);
-		kfree(cpu_to_cid);
-		kfree(cid_topo);
-		return -ENOMEM;
+	*cores_per_shard_p = nr_cores / nr_shards;
+	*nr_large_p = nr_cores % nr_shards;
+}
+
+static void scx_cid_tables_free(struct scx_cid_tables *tbls)
+{
+	if (!tbls)
+		return;
+	kvfree(tbls->cid_to_cpu);
+	kvfree(tbls->cpu_to_cid);
+	kvfree(tbls->cid_to_shard);
+	kvfree(tbls->shard_node);
+	kvfree(tbls->shard_ranges);
+	kvfree(tbls->topo);
+	kfree(tbls);
+}
+
+static void scx_cid_tables_free_rcufn(struct rcu_head *rcu)
+{
+	scx_cid_tables_free(container_of(rcu, struct scx_cid_tables, rcu));
+}
+
+static struct scx_cid_tables *scx_cid_alloc_tables(void)
+{
+	u32 npossible = num_possible_cpus();
+	struct scx_cid_tables *tbls;
+
+	tbls = kzalloc_obj(*tbls, GFP_KERNEL);
+	if (!tbls)
+		return NULL;
+
+	tbls->cid_to_cpu = kvcalloc(npossible, sizeof(*tbls->cid_to_cpu), GFP_KERNEL);
+	tbls->cpu_to_cid = kvcalloc(nr_cpu_ids, sizeof(*tbls->cpu_to_cid), GFP_KERNEL);
+	tbls->cid_to_shard = kvcalloc(npossible, sizeof(*tbls->cid_to_shard), GFP_KERNEL);
+	tbls->shard_node = kvcalloc(npossible, sizeof(*tbls->shard_node), GFP_KERNEL);
+	tbls->shard_ranges = kvcalloc(npossible, sizeof(*tbls->shard_ranges), GFP_KERNEL);
+	tbls->topo = kvcalloc(npossible, sizeof(*tbls->topo), GFP_KERNEL);
+
+	if (!tbls->cid_to_cpu || !tbls->cpu_to_cid || !tbls->cid_to_shard ||
+	    !tbls->shard_node || !tbls->shard_ranges || !tbls->topo) {
+		scx_cid_tables_free(tbls);
+		return NULL;
 	}
 
-	WRITE_ONCE(scx_cid_to_cpu_tbl, cid_to_cpu);
-	WRITE_ONCE(scx_cpu_to_cid_tbl, cpu_to_cid);
-	WRITE_ONCE(scx_cid_topo, cid_topo);
-	return 0;
+	return tbls;
+}
+
+/**
+ * scx_cid_publish_tables - Publish the tables scx_cid_init() built
+ *
+ * Called after ops.init_cids() where the layout is final.
+ */
+void scx_cid_publish_tables(void)
+{
+	struct scx_cid_tables *tbls = scx_cid_tables;
+
+	lockdep_assert_held(&scx_enable_mutex);
+
+	scx_nr_cid_shards = tbls->nr_shards;
+	rcu_assign_pointer(scx_cid_to_cpu_tbl, tbls->cid_to_cpu);
+	rcu_assign_pointer(scx_cpu_to_cid_tbl, tbls->cpu_to_cid);
+	rcu_assign_pointer(scx_cid_to_shard, tbls->cid_to_shard);
+	rcu_assign_pointer(scx_shard_node, tbls->shard_node);
+	rcu_assign_pointer(scx_cid_shard_ranges, tbls->shard_ranges);
+	rcu_assign_pointer(scx_cid_topo, tbls->topo);
+}
+
+/**
+ * scx_cid_retire_tables - Unpublish and retire the cid tables
+ *
+ * Called by root disable after the readers which dereference without NULL
+ * checks are drained, inside cpus_read_lock() to exclude the hotplug path.
+ */
+void scx_cid_retire_tables(void)
+{
+	struct scx_cid_tables *tbls = scx_cid_tables;
+
+	lockdep_assert_held(&scx_enable_mutex);
+	lockdep_assert_cpus_held();
+
+	if (!tbls)
+		return;
+
+	scx_cid_tables = NULL;
+	RCU_INIT_POINTER(scx_cid_to_cpu_tbl, NULL);
+	RCU_INIT_POINTER(scx_cpu_to_cid_tbl, NULL);
+	RCU_INIT_POINTER(scx_cid_to_shard, NULL);
+	RCU_INIT_POINTER(scx_shard_node, NULL);
+	RCU_INIT_POINTER(scx_cid_shard_ranges, NULL);
+	RCU_INIT_POINTER(scx_cid_topo, NULL);
+	call_rcu(&tbls->rcu, scx_cid_tables_free_rcufn);
 }
 
 /**
  * scx_cid_init - build the cid mapping
  * @sch: the scx_sched being initialized; used as the scx_error() target
  *
+ * Build a fresh table set. It becomes visible through scx_cid_publish_tables()
+ * and is retired by scx_cid_retire_tables() at disable.
+ *
  * See "Topological CPU IDs" in cid.h for the model. Walk online cpus by
  * intersection at each level (parent_scratch & this_level_mask), which keeps
  * containment correct by construction and naturally splits a physical LLC
@@ -88,18 +185,32 @@ s32 scx_cid_init(struct scx_sched *sch)
 	cpumask_var_t core_scratch __free(free_cpumask_var) = CPUMASK_VAR_NULL;
 	cpumask_var_t llc_fallback __free(free_cpumask_var) = CPUMASK_VAR_NULL;
 	cpumask_var_t online_no_topo __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	struct scx_cid_tables *tbls;
 	u32 next_cid = 0;
 	s32 next_node_idx = 0, next_llc_idx = 0, next_core_idx = 0;
-	s32 cpu, ret;
+	s32 next_shard_idx = 0;
+	u32 shard_size, max_cids;
+	u32 notopo_in_shard;
+	s32 notopo_shard_cid, notopo_shard_idx;
+	s32 cpu, cid, si;
 
 	/* CMASK_MAX_WORDS in cid.bpf.h covers NR_CPUS up to 8192 */
 	BUILD_BUG_ON(NR_CPUS > 8192);
 
 	lockdep_assert_cpus_held();
+	lockdep_assert_held(&scx_enable_mutex);
 
-	ret = scx_cid_arrays_alloc();
-	if (ret)
-		return ret;
+	shard_size = sch->ops.cid_shard_size ?: SCX_CID_SHARD_SIZE_DFL;
+	max_cids = min_t(u32, shard_size, SCX_CID_SHARD_MAX_CPUS);
+
+	tbls = scx_cid_alloc_tables();
+	if (!tbls)
+		return -ENOMEM;
+
+	scx_cid_tables = tbls;
+
+	for (si = 0; si < num_possible_cpus(); si++)
+		tbls->shard_node[si] = NUMA_NO_NODE;
 
 	if (!zalloc_cpumask_var(&to_walk, GFP_KERNEL) ||
 	    !zalloc_cpumask_var(&node_scratch, GFP_KERNEL) ||
@@ -111,7 +222,7 @@ s32 scx_cid_init(struct scx_sched *sch)
 
 	/* -1 sentinels for sparse-possible cpu id holes (0 is a valid cid) */
 	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
-		scx_cpu_to_cid_tbl[cpu] = -1;
+		tbls->cpu_to_cid[cpu] = -1;
 
 	cpumask_copy(to_walk, cpu_online_mask);
 
@@ -142,36 +253,69 @@ s32 scx_cid_init(struct scx_sched *sch)
 			const struct cpumask *llc_mask = cpu_llc_mask(ncpu, llc_fallback);
 			s32 llc_cid = next_cid;
 			s32 llc_idx = next_llc_idx++;
+			u32 cores_per_shard, nr_large;
+			u32 shard_local = 0, cores_in_shard = 0, cids_in_shard = 0;
+			s32 shard_cid, shard_idx;
 
 			/* llc_scratch = node_scratch & this llc */
 			cpumask_and(llc_scratch, node_scratch, llc_mask);
 			if (WARN_ON_ONCE(!cpumask_test_cpu(ncpu, llc_scratch)))
 				return -EINVAL;
 
+			calc_shard_layout(llc_scratch, shard_size, &cores_per_shard, &nr_large);
+			shard_cid = next_cid;
+			shard_idx = next_shard_idx++;
+			tbls->shard_node[shard_idx] = nid;
+
 			while (!cpumask_empty(llc_scratch)) {
 				s32 lcpu = cpumask_first(llc_scratch);
 				const struct cpumask *sib = topology_sibling_cpumask(lcpu);
 				s32 core_cid = next_cid;
 				s32 core_idx = next_core_idx++;
 				s32 ccpu;
+				u32 max_cores, cids_in_core;
 
 				/* core_scratch = llc_scratch & this core */
 				cpumask_and(core_scratch, llc_scratch, sib);
 				if (WARN_ON_ONCE(!cpumask_test_cpu(lcpu, core_scratch)))
 					return -EINVAL;
 
+				/*
+				 * Advance to a new shard when either core or
+				 * cid count reaches max. The latter bounds
+				 * shard sizes under uneven SMT. Never start an
+				 * empty shard.
+				 */
+				cids_in_core = cpumask_weight(core_scratch);
+				max_cores = cores_per_shard + (shard_local < nr_large ? 1 : 0);
+				if (cores_in_shard &&
+				    (cores_in_shard >= max_cores ||
+				     cids_in_shard + cids_in_core > max_cids)) {
+					shard_local++;
+					cores_in_shard = 0;
+					cids_in_shard = 0;
+					shard_cid = next_cid;
+					shard_idx = next_shard_idx++;
+					tbls->shard_node[shard_idx] = nid;
+				}
+				cores_in_shard++;
+				cids_in_shard += cids_in_core;
+
 				for_each_cpu(ccpu, core_scratch) {
 					s32 cid = next_cid++;
 
-					scx_cid_to_cpu_tbl[cid] = ccpu;
-					scx_cpu_to_cid_tbl[ccpu] = cid;
-					scx_cid_topo[cid] = (struct scx_cid_topo){
+					tbls->cid_to_cpu[cid] = ccpu;
+					tbls->cpu_to_cid[ccpu] = cid;
+					tbls->cid_to_shard[cid] = shard_idx;
+					tbls->topo[cid] = (struct scx_cid_topo){
 						.core_cid = core_cid,
 						.core_idx = core_idx,
 						.llc_cid = llc_cid,
 						.llc_idx = llc_idx,
 						.node_cid = node_cid,
 						.node_idx = node_idx,
+						.shard_cid = shard_cid,
+						.shard_idx = shard_idx,
 					};
 
 					cpumask_clear_cpu(ccpu, llc_scratch);
@@ -184,21 +328,37 @@ s32 scx_cid_init(struct scx_sched *sch)
 
 	/*
 	 * No-topo section: any possible cpu without a cid - normally just the
-	 * not-online ones. Collect any currently-online cpus that land here in
-	 * @online_no_topo so we can warn about them at the end.
+	 * not-online ones. Pack into shards of up to min(@shard_size,
+	 * SCX_CID_SHARD_MAX_CPUS) cids so that every cid has a valid shard
+	 * assignment and the hard cap holds even with a large @shard_size.
+	 * Collect any currently-online cpus that land here in @online_no_topo
+	 * so we can warn about them at the end.
 	 */
-	for_each_cpu(cpu, cpu_possible_mask) {
-		s32 cid;
+	notopo_in_shard = min_t(u32, shard_size, SCX_CID_SHARD_MAX_CPUS);
+	notopo_shard_cid = -1;
+	notopo_shard_idx = -1;
 
-		if (__scx_cpu_to_cid(cpu) != -1)
+	for_each_cpu(cpu, cpu_possible_mask) {
+		if (tbls->cpu_to_cid[cpu] != -1)
 			continue;
 		if (cpu_online(cpu))
 			cpumask_set_cpu(cpu, online_no_topo);
 
 		cid = next_cid++;
-		scx_cid_to_cpu_tbl[cid] = cpu;
-		scx_cpu_to_cid_tbl[cpu] = cid;
-		scx_cid_topo[cid] = SCX_CID_TOPO_NEG;
+		tbls->cid_to_cpu[cid] = cpu;
+		tbls->cpu_to_cid[cpu] = cid;
+
+		if (notopo_in_shard >= min_t(u32, shard_size, SCX_CID_SHARD_MAX_CPUS)) {
+			notopo_shard_cid = cid;
+			notopo_shard_idx = next_shard_idx++;
+			notopo_in_shard = 0;
+		}
+		notopo_in_shard++;
+
+		tbls->cid_to_shard[cid] = notopo_shard_idx;
+		tbls->topo[cid] = SCX_CID_TOPO_NEG;
+		tbls->topo[cid].shard_cid = notopo_shard_cid;
+		tbls->topo[cid].shard_idx = notopo_shard_idx;
 	}
 
 	if (!cpumask_empty(llc_fallback))
@@ -208,6 +368,20 @@ s32 scx_cid_init(struct scx_sched *sch)
 		pr_warn("scx_cid: online cpus with no usable topology: %*pbl\n",
 			cpumask_pr_args(online_no_topo));
 
+	/*
+	 * Fill cid_shard_ranges[] from cid_to_shard[]. Shards are contiguous
+	 * cid ranges by construction: base_cid is the first cid landing in a
+	 * shard, nr_cids is the count.
+	 */
+	for (cid = 0; cid < next_cid; cid++) {
+		s32 sidx = tbls->cid_to_shard[cid];
+
+		if (tbls->shard_ranges[sidx].nr_cids == 0)
+			tbls->shard_ranges[sidx].base_cid = cid;
+		tbls->shard_ranges[sidx].nr_cids++;
+	}
+
+	tbls->nr_shards = next_shard_idx;
 	return 0;
 }
 
@@ -253,50 +427,76 @@ void scx_cmask_fill(struct scx_cmask *m)
 		m->bits[nr_words - 1] &= (1ULL << tail_bits) - 1;
 }
 
-/**
- * scx_cpumask_to_cmask - Translate a kernel cpumask into a cmask
- * @src: source cpumask
- * @dst: cmask to write
- *
- * Clear @dst's active range and set the bit for each cid whose cpu is in
- * @src and lies within that range. Out-of-range cids are silently ignored.
+/*
+ * Return the index of the largest entry in @counts, or NUMA_NO_NODE if all
+ * entries are zero. Ties resolve to the lowest index.
  */
-void scx_cpumask_to_cmask(const struct cpumask *src, struct scx_cmask *dst)
+static s32 pick_max_node(const u32 *counts, u32 n)
 {
-	s32 cpu;
-
-	scx_cmask_clear(dst);
-	for_each_cpu(cpu, src) {
-		s32 cid = __scx_cpu_to_cid(cpu);
+	s32 best = NUMA_NO_NODE;
+	u32 best_count = 0, i;
 
-		if (cid >= 0)
-			__scx_cmask_set(cid, dst);
+	for (i = 0; i < n; i++) {
+		if (counts[i] > best_count) {
+			best_count = counts[i];
+			best = i;
+		}
 	}
+	return best;
 }
 
 __bpf_kfunc_start_defs();
 
 /**
- * scx_bpf_cid_override - Install an explicit cpu->cid mapping
- * @cpu_to_cid: array of nr_cpu_ids s32 entries (cid for each cpu)
- * @cpu_to_cid__sz: must be nr_cpu_ids * sizeof(s32) bytes
+ * scx_bpf_cid_override - Install an explicit cpu->cid mapping with shard info
+ * @cpu_to_cid__arena: array of nr_cpu_ids s32 entries (cid for each cpu)
+ * @cpu_to_cid_cnt: number of entries, must be nr_cpu_ids
+ * @shard_start__arena: array of first-cid-of-each-shard, one entry per shard
+ * @shard_start_cnt: number of shards
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
- * May only be called from ops.init() of the root scheduler. Replace the
- * topology-probed cid mapping with the caller-provided one. Each possible cpu
- * must map to a unique cid in [0, num_possible_cpus()). Topo info is cleared.
- * On invalid input, trigger scx_error() to abort the scheduler.
+ * May only be called from ops.init_cids() of the root scheduler. Replace the
+ * topology-probed cid mapping and shard layout with caller-provided ones. Each
+ * possible cpu must map to a unique cid in [0, num_possible_cpus()). The shard
+ * starts must be strictly increasing with the first entry 0 and all values <
+ * num_possible_cpus(). The last shard extends to num_possible_cpus() and no
+ * shard may span more than SCX_CID_SHARD_MAX_CPUS cids. Topo info
+ * (core/LLC/node) is cleared and the shard layout is set from the input. On
+ * invalid input, abort the scheduler.
  */
-__bpf_kfunc void scx_bpf_cid_override(const s32 *cpu_to_cid, u32 cpu_to_cid__sz,
+__bpf_kfunc void scx_bpf_cid_override(const s32 *cpu_to_cid__arena, u32 cpu_to_cid_cnt,
+				      const s32 *shard_start__arena, u32 shard_start_cnt,
 				      const struct bpf_prog_aux *aux)
 {
 	cpumask_var_t seen __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	u32 *node_counts __free(kfree) = NULL;
+	s32 *cpu_to_cid __free(kfree) = NULL;
+	s32 *shard_start __free(kfree) = NULL;
+	u32 npossible = num_possible_cpus();
+	struct scx_cid_tables *tbls;
 	struct scx_sched *sch;
+	u32 nr_shards = shard_start_cnt;
 	bool alloced;
-	s32 cpu, cid;
+	s32 cpu, cid, si;
 
-	/* GFP_KERNEL alloc must happen before the rcu read section */
+	/*
+	 * GFP_KERNEL allocs must happen before the rcu read section. Snapshot
+	 * the BPF-supplied arrays so a concurrent arena write can't change
+	 * them between validation and use.
+	 *
+	 * The BPF-supplied counts size the snapshots and thus the arena reads.
+	 * Gate the copies on the count bounds, reported below once @sch is
+	 * available. The bounded reads, at most 32KB, stay within the guard
+	 * region that arena fault recovery covers.
+	 */
 	alloced = zalloc_cpumask_var(&seen, GFP_KERNEL);
+	node_counts = kcalloc(nr_node_ids, sizeof(*node_counts), GFP_KERNEL);
+	if (cpu_to_cid_cnt == nr_cpu_ids)
+		cpu_to_cid = kmemdup(cpu_to_cid__arena, cpu_to_cid_cnt * sizeof(s32),
+				     GFP_KERNEL);
+	if (nr_shards && nr_shards <= npossible)
+		shard_start = kmemdup(shard_start__arena, nr_shards * sizeof(s32),
+				      GFP_KERNEL);
 
 	guard(rcu)();
 
@@ -304,22 +504,59 @@ __bpf_kfunc void scx_bpf_cid_override(const s32 *cpu_to_cid, u32 cpu_to_cid__sz,
 	if (unlikely(!sch))
 		return;
 
-	if (!alloced) {
-		scx_error(sch, "scx_bpf_cid_override: failed to allocate cpumask");
+	/* called from ops.init_cids(), so the tables exist and are unpublished */
+	lockdep_assert_held(&scx_enable_mutex);
+	tbls = scx_cid_tables;
+
+	if (cpu_to_cid_cnt != nr_cpu_ids) {
+		scx_error(sch, "scx_bpf_cid_override: cpu_to_cid expected %u entries, got %u",
+			  nr_cpu_ids, cpu_to_cid_cnt);
+		return;
+	}
+
+	if (!nr_shards || nr_shards > npossible) {
+		scx_error(sch, "scx_bpf_cid_override: invalid shard_start count %u",
+			  nr_shards);
 		return;
 	}
 
-	if (scx_parent(sch)) {
-		scx_error(sch, "scx_bpf_cid_override() only allowed from root sched");
+	if (!alloced || !node_counts || !cpu_to_cid || !shard_start) {
+		scx_error(sch, "scx_bpf_cid_override: allocation failed");
 		return;
 	}
 
-	if (cpu_to_cid__sz != nr_cpu_ids * sizeof(s32)) {
-		scx_error(sch, "scx_bpf_cid_override: expected %zu bytes, got %u",
-			  nr_cpu_ids * sizeof(s32), cpu_to_cid__sz);
+	/* validate shard_start[]: starts at 0, strictly increasing, in range */
+	if (shard_start[0] != 0) {
+		scx_error(sch, "scx_bpf_cid_override: shard_start[0] must be 0, got %d",
+			  shard_start[0]);
+		return;
+	}
+	for (si = 1; si < nr_shards; si++) {
+		if (shard_start[si] <= shard_start[si - 1]) {
+			scx_error(sch, "scx_bpf_cid_override: shard_start not increasing at [%d]",
+				  si);
+			return;
+		}
+		if (shard_start[si] >= npossible) {
+			scx_error(sch, "scx_bpf_cid_override: shard_start[%d]=%d >= %u",
+				  si, shard_start[si], npossible);
+			return;
+		}
+		if (shard_start[si] - shard_start[si - 1] > SCX_CID_SHARD_MAX_CPUS) {
+			scx_error(sch, "scx_bpf_cid_override: shard[%d] span %d exceeds max %d",
+				  si - 1, shard_start[si] - shard_start[si - 1],
+				  SCX_CID_SHARD_MAX_CPUS);
+			return;
+		}
+	}
+	if (npossible - shard_start[nr_shards - 1] > SCX_CID_SHARD_MAX_CPUS) {
+		scx_error(sch, "scx_bpf_cid_override: shard[%d] span %d exceeds max %d",
+			  nr_shards - 1, npossible - shard_start[nr_shards - 1],
+			  SCX_CID_SHARD_MAX_CPUS);
 		return;
 	}
 
+	/* validate first so that invalid input leaves the tables untouched */
 	for_each_possible_cpu(cpu) {
 		s32 c = cpu_to_cid[cpu];
 
@@ -329,13 +566,56 @@ __bpf_kfunc void scx_bpf_cid_override(const s32 *cpu_to_cid, u32 cpu_to_cid__sz,
 			scx_error(sch, "cid %d assigned to multiple cpus", c);
 			return;
 		}
-		scx_cpu_to_cid_tbl[cpu] = c;
-		scx_cid_to_cpu_tbl[c] = cpu;
 	}
 
-	/* Invalidate stale topo info - the override carries no topology. */
-	for (cid = 0; cid < num_possible_cpus(); cid++)
-		scx_cid_topo[cid] = SCX_CID_TOPO_NEG;
+	for_each_possible_cpu(cpu) {
+		s32 c = cpu_to_cid[cpu];
+
+		tbls->cpu_to_cid[cpu] = c;
+		tbls->cid_to_cpu[c] = cpu;
+	}
+
+	/*
+	 * Derive shard_node[] by majority count: an overridden shard may
+	 * span NUMA nodes, so assign each to the node that owns the most cpus.
+	 */
+	for (si = 0; si < nr_shards; si++) {
+		u32 end = (si + 1 < nr_shards) ? shard_start[si + 1] : npossible;
+
+		memset(node_counts, 0, nr_node_ids * sizeof(*node_counts));
+		for (cid = shard_start[si]; cid < end; cid++) {
+			s32 node = cpu_to_node(tbls->cid_to_cpu[cid]);
+
+			if (numa_valid_node(node))
+				node_counts[node]++;
+		}
+		tbls->shard_node[si] = pick_max_node(node_counts, nr_node_ids);
+	}
+
+	/*
+	 * Invalidate stale topo info and install shard layout from
+	 * @shard_start. Walk shards to derive shard_cid/shard_idx for each cid.
+	 */
+	si = 0;
+	for (cid = 0; cid < npossible; cid++) {
+		if (si + 1 < nr_shards && cid >= shard_start[si + 1])
+			si++;
+		tbls->cid_to_shard[cid] = si;
+		tbls->topo[cid] = SCX_CID_TOPO_NEG;
+		tbls->topo[cid].shard_cid = shard_start[si];
+		tbls->topo[cid].shard_idx = si;
+	}
+
+	/* Rebuild shard_ranges[] for the new layout. */
+	memset(tbls->shard_ranges, 0, npossible * sizeof(*tbls->shard_ranges));
+	for (si = 0; si < nr_shards; si++) {
+		u32 end = (si + 1 < nr_shards) ? shard_start[si + 1] : npossible;
+
+		tbls->shard_ranges[si].base_cid = shard_start[si];
+		tbls->shard_ranges[si].nr_cids = end - shard_start[si];
+	}
+
+	tbls->nr_shards = nr_shards;
 }
 
 /**
@@ -395,21 +675,25 @@ __bpf_kfunc s32 scx_bpf_cpu_to_cid(s32 cpu, const struct bpf_prog_aux *aux)
  * bits outside stay untouched. In particular, scx_cmask_copy() does NOT zero
  * @dst bits that lie outside @src's range.
  *
- * The _RACY variants are otherwise identical to their non-racy counterpart but
- * read @src word-by-word via data_race(). Memory ordering with concurrent
- * writers is the caller's responsibility.
+ * Word accesses use READ_ONCE/WRITE_ONCE so a caller may read @src
+ * locklessly. Memory ordering against concurrent writers is the caller's
+ * responsibility.
  */
 enum cmask_op2 {
 	/* mutating */
 	CMASK_OP2_AND,
 	CMASK_OP2_OR,
-	CMASK_OP2_OR_RACY,
 	CMASK_OP2_COPY,
-	CMASK_OP2_COPY_RACY,
 	CMASK_OP2_ANDNOT,
 	/* predicates - short-circuit when the per-word result is true */
 	CMASK_OP2_SUBSET,
 	CMASK_OP2_INTERSECTS,
+	/*
+	 * @a is a BPF-arena cmask. Words on @a use READ_ONCE/WRITE_ONCE since
+	 * BPF may read/write concurrently. See scx_cmask_ref_or() / _copy().
+	 */
+	CMASK_OP2_REF_OR,
+	CMASK_OP2_REF_COPY,
 };
 
 static __always_inline bool cmask_op2_is_pred(const enum cmask_op2 op)
@@ -422,28 +706,28 @@ static __always_inline bool cmask_word_op2(u64 *av, const u64 *bp, u64 mask,
 {
 	switch (op) {
 	case CMASK_OP2_AND:
-		*av &= ~mask | *bp;
+		WRITE_ONCE(*av, *av & (~mask | READ_ONCE(*bp)));
 		return false;
 	case CMASK_OP2_OR:
-		*av |= *bp & mask;
-		return false;
-	case CMASK_OP2_OR_RACY:
-		*av |= data_race(*bp) & mask;
+		WRITE_ONCE(*av, *av | (READ_ONCE(*bp) & mask));
 		return false;
 	case CMASK_OP2_COPY:
-		*av = (*av & ~mask) | (*bp & mask);
-		return false;
-	case CMASK_OP2_COPY_RACY:
-		*av = (*av & ~mask) | (data_race(*bp) & mask);
+		WRITE_ONCE(*av, (*av & ~mask) | (READ_ONCE(*bp) & mask));
 		return false;
 	case CMASK_OP2_ANDNOT:
-		*av &= ~(*bp & mask);
+		WRITE_ONCE(*av, *av & ~(READ_ONCE(*bp) & mask));
 		return false;
 	case CMASK_OP2_SUBSET:
 		/* stop on the first bit in @sub not set in @super */
-		return (*bp & ~*av) & mask;
+		return (READ_ONCE(*bp) & ~READ_ONCE(*av)) & mask;
 	case CMASK_OP2_INTERSECTS:
-		return (*av & *bp) & mask;
+		return (READ_ONCE(*av) & READ_ONCE(*bp)) & mask;
+	case CMASK_OP2_REF_OR:
+		WRITE_ONCE(*av, READ_ONCE(*av) | (READ_ONCE(*bp) & mask));
+		return false;
+	case CMASK_OP2_REF_COPY:
+		WRITE_ONCE(*av, (READ_ONCE(*av) & ~mask) | (READ_ONCE(*bp) & mask));
+		return false;
 	}
 	unreachable();
 }
@@ -504,7 +788,7 @@ static __always_inline bool cmask_word_op1(const u64 *ap, u64 mask,
 {
 	switch (op) {
 	case CMASK_OP1_ANY_SET:
-		return *ap & mask;
+		return READ_ONCE(*ap) & mask;
 	}
 	unreachable();
 }
@@ -556,39 +840,12 @@ void scx_cmask_or(struct scx_cmask *dst, const struct scx_cmask *src)
 		       src->bits, src->base, src->nr_cids, CMASK_OP2_OR);
 }
 
-/**
- * scx_cmask_or_racy - OR @src into @dst, reading @src without locking
- *
- * @src is read word-by-word through data_race(). Same per-bit independence
- * rationale as scx_cmask_copy_racy(). Memory ordering with writers is the
- * caller's responsibility.
- */
-void scx_cmask_or_racy(struct scx_cmask *dst, const struct scx_cmask *src)
-{
-	cmask_walk_op2(dst->bits, dst->base, dst->nr_cids,
-		       src->bits, src->base, src->nr_cids, CMASK_OP2_OR_RACY);
-}
-
 void scx_cmask_copy(struct scx_cmask *dst, const struct scx_cmask *src)
 {
 	cmask_walk_op2(dst->bits, dst->base, dst->nr_cids,
 		       src->bits, src->base, src->nr_cids, CMASK_OP2_COPY);
 }
 
-/**
- * scx_cmask_copy_racy - Snapshot @src into @dst without locking
- *
- * @src is read word-by-word through data_race(). Head/tail masking matches
- * scx_cmask_copy(). Each bit in a cmask is independent, so partial updates
- * just leave some bits fresher than others. Memory ordering with writers is
- * the caller's responsibility.
- */
-void scx_cmask_copy_racy(struct scx_cmask *dst, const struct scx_cmask *src)
-{
-	cmask_walk_op2(dst->bits, dst->base, dst->nr_cids,
-		       src->bits, src->base, src->nr_cids, CMASK_OP2_COPY_RACY);
-}
-
 void scx_cmask_andnot(struct scx_cmask *dst, const struct scx_cmask *src)
 {
 	cmask_walk_op2(dst->bits, dst->base, dst->nr_cids,
@@ -660,33 +917,36 @@ bool scx_cmask_empty(const struct scx_cmask *m)
  *
  * Fill @out__uninit with the topology info for @cid. Trigger scx_error() if
  * @cid is out of range. If @cid is valid but in the no-topo section, all fields
- * are set to -1.
+ * are set to -1. All fields are also set to -1 when no cid tables have been
+ * published yet, which a program may observe while racing the root enable.
  */
 __bpf_kfunc void scx_bpf_cid_topo(s32 cid, struct scx_cid_topo *out__uninit,
 				  const struct bpf_prog_aux *aux)
 {
+	struct scx_cid_topo *topo;
 	struct scx_sched *sch;
 
 	guard(rcu)();
 
 	sch = scx_prog_sched(aux);
-	if (unlikely(!sch) || !cid_valid(sch, cid)) {
+	topo = rcu_dereference(scx_cid_topo);
+	if (unlikely(!sch) || !cid_valid(sch, cid) || unlikely(!topo)) {
 		*out__uninit = SCX_CID_TOPO_NEG;
 		return;
 	}
 
-	*out__uninit = READ_ONCE(scx_cid_topo)[cid];
+	*out__uninit = topo[cid];
 }
 
 __bpf_kfunc_end_defs();
 
-BTF_KFUNCS_START(scx_kfunc_ids_init)
+BTF_KFUNCS_START(scx_kfunc_ids_init_cids)
 BTF_ID_FLAGS(func, scx_bpf_cid_override, KF_IMPLICIT_ARGS | KF_SLEEPABLE)
-BTF_KFUNCS_END(scx_kfunc_ids_init)
+BTF_KFUNCS_END(scx_kfunc_ids_init_cids)
 
-static const struct btf_kfunc_id_set scx_kfunc_set_init = {
+static const struct btf_kfunc_id_set scx_kfunc_set_init_cids = {
 	.owner	= THIS_MODULE,
-	.set	= &scx_kfunc_ids_init,
+	.set	= &scx_kfunc_ids_init_cids,
 	.filter	= scx_kfunc_context_filter,
 };
 
@@ -701,9 +961,205 @@ static const struct btf_kfunc_id_set scx_kfunc_set_cid = {
 	.set	= &scx_kfunc_ids_cid,
 };
 
+/**
+ * scx_cmask_ref_init - Bind a scx_cmask_ref to a BPF-arena cmask
+ * @sch: scheduler whose arena hosts @src
+ * @src: BPF-supplied cmask, rebased to its kernel address
+ * @ref: output ref
+ *
+ * Snapshot @src's @base, @nr_cids and @alloc_words. The snapshot is necessary
+ * because BPF may mutate the live header asynchronously.
+ *
+ * Return 0 on success, -EINVAL if the range is out of bounds or @alloc_words
+ * doesn't cover it.
+ */
+int scx_cmask_ref_init(struct scx_sched *sch, const struct scx_cmask *src,
+		       struct scx_cmask_ref *ref)
+{
+	u32 base, nr_cids, alloc_words, npossible = num_possible_cpus();
+	s32 *cid_to_shard;
+
+	base = READ_ONCE(src->base);
+	nr_cids = READ_ONCE(src->nr_cids);
+	alloc_words = READ_ONCE(src->alloc_words);
+
+	if (unlikely(base >= npossible || nr_cids > npossible - base ||
+		     SCX_CMASK_NR_WORDS(nr_cids) > alloc_words))
+		return -EINVAL;
+
+	ref->sch = sch;
+	ref->src = (struct scx_cmask *)src;
+	ref->base = base;
+	ref->nr_cids = nr_cids;
+
+	cid_to_shard = rcu_dereference_all(scx_cid_to_shard);
+	ref->shard_first = cid_to_shard[base];
+	if (likely(nr_cids))
+		ref->shard_end = cid_to_shard[base + nr_cids - 1] + 1;
+	else
+		ref->shard_end = ref->shard_first;
+
+	return 0;
+}
+
+/**
+ * scx_cmask_ref_init_kern - Bind a scx_cmask_ref to a kernel-owned cmask
+ * @sch: scheduler the cmask belongs to
+ * @m: kernel address of the target cmask, storage sized for @nr_cids at @base
+ * @base: first cid of the active range
+ * @nr_cids: active range length
+ * @ref: output ref
+ *
+ * Like scx_cmask_ref_init() but the geometry is supplied by the caller, not
+ * read from @m's header, so a concurrent BPF write to the header can't steer
+ * later sizing or offsets. Rewrite the header from the trusted geometry and
+ * bind @ref to it.
+ */
+void scx_cmask_ref_init_kern(struct scx_sched *sch, struct scx_cmask *m,
+			     u32 base, u32 nr_cids, struct scx_cmask_ref *ref)
+{
+	s32 *cid_to_shard;
+
+	WRITE_ONCE(m->base, base);
+	WRITE_ONCE(m->nr_cids, nr_cids);
+	WRITE_ONCE(m->alloc_words, SCX_CMASK_NR_WORDS(nr_cids));
+
+	ref->sch = sch;
+	ref->src = m;
+	ref->base = base;
+	ref->nr_cids = nr_cids;
+
+	cid_to_shard = rcu_dereference_all(scx_cid_to_shard);
+	ref->shard_first = cid_to_shard[base];
+	if (likely(nr_cids))
+		ref->shard_end = cid_to_shard[base + nr_cids - 1] + 1;
+	else
+		ref->shard_end = ref->shard_first;
+}
+
+/**
+ * scx_cmask_ref_shard - Read one shard from @ref into @out
+ * @ref: validated ref
+ * @shard_idx: target shard, in [@ref->shard_first, @ref->shard_end)
+ * @out: output cmask whose @out->alloc_words must hold the shard
+ *
+ * Set @out to the intersection of @ref's range with @shard_idx's cid range,
+ * with bits[] read from @ref->src via READ_ONCE. Empty intersection sets
+ * @out->nr_cids to 0. scx_error()s on @ref's sched if @out can't hold the
+ * shard.
+ */
+void scx_cmask_ref_shard(const struct scx_cmask_ref *ref, s32 shard_idx,
+			 struct scx_cmask *out)
+{
+	const struct scx_cid_shard *shard =
+		&rcu_dereference_all(scx_cid_shard_ranges)[shard_idx];
+	u32 shard_base = shard->base_cid;
+	u32 shard_end = shard_base + shard->nr_cids;
+	u32 isect_base, isect_end, nr_words, src_off, wi;
+	u64 head_mask, tail_mask;
+
+	isect_base = max(ref->base, shard_base);
+	isect_end = min(ref->base + ref->nr_cids, shard_end);
+
+	if (isect_base >= isect_end) {
+		out->base = shard_base;
+		out->nr_cids = 0;
+		return;
+	}
+
+	nr_words = ((isect_end - 1) / 64) - (isect_base / 64) + 1;
+	if (nr_words > out->alloc_words) {
+		scx_error(ref->sch, "scx_cmask_ref_shard: out alloc_words=%u < %u for shard %d",
+			  out->alloc_words, nr_words, shard_idx);
+		out->base = shard_base;
+		out->nr_cids = 0;
+		return;
+	}
+
+	out->base = isect_base;
+	out->nr_cids = isect_end - isect_base;
+	src_off = (isect_base / 64) - (ref->base / 64);
+
+	for (wi = 0; wi < nr_words; wi++)
+		out->bits[wi] = READ_ONCE(ref->src->bits[src_off + wi]);
+
+	head_mask = GENMASK_U64(63, isect_base & 63);
+	out->bits[0] &= head_mask;
+	tail_mask = GENMASK_U64((isect_end - 1) & 63, 0);
+	out->bits[nr_words - 1] &= tail_mask;
+}
+
+/**
+ * scx_cmask_ref_or - OR @src into the arena cmask referenced by @ref
+ * @ref: validated ref
+ * @src: stable kernel cmask
+ *
+ * Bits inside the intersection of @ref's snapshotted range with @src's range
+ * are OR'd into @ref->src and bits outside are left unchanged. Stores on
+ * @ref->src use WRITE_ONCE since BPF may read/write concurrently.
+ */
+void scx_cmask_ref_or(const struct scx_cmask_ref *ref, const struct scx_cmask *src)
+{
+	cmask_walk_op2(ref->src->bits, ref->base, ref->nr_cids,
+		       src->bits, src->base, src->nr_cids, CMASK_OP2_REF_OR);
+}
+
+/**
+ * scx_cmask_ref_copy - Copy @src into the arena cmask referenced by @ref
+ * @ref: validated ref
+ * @src: stable kernel cmask
+ *
+ * Bits inside the intersection of @ref's snapshotted range with @src's range
+ * take @src's values and bits outside are left unchanged. Stores on @ref->src
+ * use WRITE_ONCE since BPF may read/write concurrently.
+ */
+void scx_cmask_ref_copy(const struct scx_cmask_ref *ref, const struct scx_cmask *src)
+{
+	cmask_walk_op2(ref->src->bits, ref->base, ref->nr_cids,
+		       src->bits, src->base, src->nr_cids, CMASK_OP2_REF_COPY);
+}
+
+/**
+ * scx_cmask_ref_from_cpumask - Populate @ref's arena cmask from a cpumask
+ * @ref: kern-bound ref, see scx_cmask_ref_init_kern()
+ * @cpumask: cpus to translate into cids
+ *
+ * Write @ref's active range one word at a time, setting each cid's bit when
+ * its cpu is in @cpumask. Offsets and length come from @ref's trusted geometry
+ * and stores use WRITE_ONCE since BPF may read concurrently, so the arena
+ * header is never read.
+ */
+void scx_cmask_ref_from_cpumask(const struct scx_cmask_ref *ref,
+				const struct cpumask *cpumask)
+{
+	struct scx_cmask *m = ref->src;
+	u32 base = ref->base, nr_cids = ref->nr_cids;
+	u32 wi, nr_words;
+
+	if (!nr_cids)
+		return;
+
+	nr_words = (base + nr_cids - 1) / 64 - base / 64 + 1;
+	for (wi = 0; wi < nr_words; wi++) {
+		u32 word_first_cid = (base / 64 + wi) * 64;
+		u64 word = 0;
+		u32 bit;
+
+		for (bit = 0; bit < 64; bit++) {
+			u32 cid = word_first_cid + bit;
+
+			if (cid < base || cid >= base + nr_cids)
+				continue;
+			if (cpumask_test_cpu(__scx_cid_to_cpu(cid), cpumask))
+				word |= BIT_U64(bit);
+		}
+		WRITE_ONCE(m->bits[wi], word);
+	}
+}
+
 int scx_cid_kfunc_init(void)
 {
-	return register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_init) ?:
+	return register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_init_cids) ?:
 		register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_cid) ?:
 		register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_cid) ?:
 		register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_cid);
diff --git a/kernel/sched/ext/cid.h b/kernel/sched/ext/cid.h
index 9c4f4b907f12..2fe2311a0f99 100644
--- a/kernel/sched/ext/cid.h
+++ b/kernel/sched/ext/cid.h
@@ -48,25 +48,40 @@ struct scx_sched;
  * See the comment above the table definitions in cid.c for the
  * memory-ordering and visibility contract.
  */
-extern s16 *scx_cid_to_cpu_tbl;
-extern s16 *scx_cpu_to_cid_tbl;
-extern struct scx_cid_topo *scx_cid_topo;
-extern struct btf_id_set8 scx_kfunc_ids_init;
+struct scx_cid_tables {
+	u32			nr_shards;
+	s16			*cid_to_cpu;	/* [num_possible_cpus()] */
+	s16			*cpu_to_cid;	/* [nr_cpu_ids] */
+	s32			*cid_to_shard;	/* [num_possible_cpus()] */
+	s32			*shard_node;	/* [num_possible_cpus()] */
+	struct scx_cid_shard	*shard_ranges;	/* [num_possible_cpus()] */
+	struct scx_cid_topo	*topo;		/* [num_possible_cpus()] */
+	struct rcu_head		rcu;
+};
+
+extern u32 scx_nr_cid_shards;
+extern s16 __rcu *scx_cid_to_cpu_tbl;
+extern s16 __rcu *scx_cpu_to_cid_tbl;
+extern s32 __rcu *scx_cid_to_shard;
+extern s32 __rcu *scx_shard_node;
+extern struct scx_cid_shard __rcu *scx_cid_shard_ranges;
+extern struct scx_cid_topo __rcu *scx_cid_topo;
+extern struct btf_id_set8 scx_kfunc_ids_init_cids;
+extern struct btf_id_set8 scx_kfunc_ids_cid;
 
 void scx_cmask_clear(struct scx_cmask *m);
 void scx_cmask_fill(struct scx_cmask *m);
 void scx_cmask_and(struct scx_cmask *dst, const struct scx_cmask *src);
 void scx_cmask_or(struct scx_cmask *dst, const struct scx_cmask *src);
-void scx_cmask_or_racy(struct scx_cmask *dst, const struct scx_cmask *src);
 void scx_cmask_copy(struct scx_cmask *dst, const struct scx_cmask *src);
-void scx_cmask_copy_racy(struct scx_cmask *dst, const struct scx_cmask *src);
 void scx_cmask_andnot(struct scx_cmask *dst, const struct scx_cmask *src);
 bool scx_cmask_subset(const struct scx_cmask *sub, const struct scx_cmask *super);
 bool scx_cmask_intersects(const struct scx_cmask *a, const struct scx_cmask *b);
 bool scx_cmask_empty(const struct scx_cmask *m);
 s32 scx_cid_init(struct scx_sched *sch);
+void scx_cid_publish_tables(void);
+void scx_cid_retire_tables(void);
 int scx_cid_kfunc_init(void);
-void scx_cpumask_to_cmask(const struct cpumask *src, struct scx_cmask *dst);
 
 /**
  * cid_valid - Verify a cid value, to be used on ops input args
@@ -88,14 +103,12 @@ static inline bool cid_valid(struct scx_sched *sch, s32 cid)
  * __scx_cid_to_cpu - Unchecked cid->cpu table lookup
  * @cid: cid to look up. Must be in [0, num_possible_cpus()).
  *
- * Intended for callsites that have already validated @cid and that hold a
- * non-NULL @sch from scx_prog_sched() - a live sched implies the table has
- * been allocated, so no NULL check is needed here.
+ * Intended for callsites that have already validated @cid and that run on a
+ * live scheduler, which guarantees the tables are published and stable.
  */
 static inline s32 __scx_cid_to_cpu(s32 cid)
 {
-	/* READ_ONCE pairs with WRITE_ONCE in scx_cid_arrays_alloc() */
-	return READ_ONCE(scx_cid_to_cpu_tbl)[cid];
+	return rcu_dereference_all(scx_cid_to_cpu_tbl)[cid];
 }
 
 /**
@@ -106,7 +119,7 @@ static inline s32 __scx_cid_to_cpu(s32 cid)
  */
 static inline s32 __scx_cpu_to_cid(s32 cpu)
 {
-	return READ_ONCE(scx_cpu_to_cid_tbl)[cpu];
+	return rcu_dereference_all(scx_cpu_to_cid_tbl)[cpu];
 }
 
 /**
@@ -115,15 +128,19 @@ static inline s32 __scx_cpu_to_cid(s32 cpu)
  * @cid: cid to look up
  *
  * Return the cpu for @cid or a negative errno on failure. Invalid cid triggers
- * scx_error() on @sch. The cid arrays are allocated on first scheduler enable
- * and never freed, so the returned cpu is stable for the lifetime of the loaded
- * scheduler.
+ * scx_error() on @sch. The mapping is stable while the scheduler is live.
+ *
+ * Return -EINVAL without triggering scx_error() if no tables have been
+ * published yet, which a prog-facing kfunc can observe while racing the root
+ * scheduler enable.
  */
 static inline s32 scx_cid_to_cpu(struct scx_sched *sch, s32 cid)
 {
-	if (!cid_valid(sch, cid))
+	s16 *tbl = rcu_dereference_all(scx_cid_to_cpu_tbl);
+
+	if (!cid_valid(sch, cid) || unlikely(!tbl))
 		return -EINVAL;
-	return __scx_cid_to_cpu(cid);
+	return tbl[cid];
 }
 
 /**
@@ -132,13 +149,15 @@ static inline s32 scx_cid_to_cpu(struct scx_sched *sch, s32 cid)
  * @cpu: cpu to look up
  *
  * Return the cid for @cpu or a negative errno on failure. Invalid cpu triggers
- * scx_error() on @sch. Same lifetime guarantee as scx_cid_to_cpu().
+ * scx_error() on @sch. Same usage rules as scx_cid_to_cpu().
  */
 static inline s32 scx_cpu_to_cid(struct scx_sched *sch, s32 cpu)
 {
-	if (!scx_cpu_valid(sch, cpu, NULL))
+	s16 *tbl = rcu_dereference_all(scx_cpu_to_cid_tbl);
+
+	if (!scx_cpu_valid(sch, cpu, NULL) || unlikely(!tbl))
 		return -EINVAL;
-	return __scx_cpu_to_cid(cpu);
+	return tbl[cpu];
 }
 
 /**
@@ -291,4 +310,15 @@ static inline s32 scx_cpu_ret(struct scx_sched *sch, s32 cpu_or_cid)
 	return scx_cid_to_cpu(sch, cpu_or_cid);
 }
 
+int scx_cmask_ref_init(struct scx_sched *sch, const struct scx_cmask *src,
+		       struct scx_cmask_ref *ref);
+void scx_cmask_ref_init_kern(struct scx_sched *sch, struct scx_cmask *m,
+			     u32 base, u32 nr_cids, struct scx_cmask_ref *ref);
+void scx_cmask_ref_shard(const struct scx_cmask_ref *ref, s32 shard_idx,
+			 struct scx_cmask *out);
+void scx_cmask_ref_from_cpumask(const struct scx_cmask_ref *ref,
+				const struct cpumask *cpumask);
+void scx_cmask_ref_or(const struct scx_cmask_ref *ref, const struct scx_cmask *src);
+void scx_cmask_ref_copy(const struct scx_cmask_ref *ref, const struct scx_cmask *src);
+
 #endif /* _KERNEL_SCHED_EXT_CID_H */
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index 18183062f751..10af28a9f2c0 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -19,8 +19,10 @@
 #include "cid.h"
 #include "arena.h"
 #include "idle.h"
+#include "sub.h"
+#include "inlines.h"
 
-static DEFINE_RAW_SPINLOCK(scx_sched_lock);
+DEFINE_RAW_SPINLOCK(scx_sched_lock);
 
 /*
  * NOTE: sched_ext is in the process of growing multiple scheduler support and
@@ -36,17 +38,17 @@ struct scx_sched __rcu *scx_root;
  * All scheds, writers must hold both scx_enable_mutex and scx_sched_lock.
  * Readers can hold either or rcu_read_lock().
  */
-static LIST_HEAD(scx_sched_all);
+LIST_HEAD(scx_sched_all);
 
 #ifdef CONFIG_EXT_SUB_SCHED
-static const struct rhashtable_params scx_sched_hash_params = {
+const struct rhashtable_params scx_sched_hash_params = {
 	.key_len		= sizeof_field(struct scx_sched, ops.sub_cgroup_id),
 	.key_offset		= offsetof(struct scx_sched, ops.sub_cgroup_id),
 	.head_offset		= offsetof(struct scx_sched, hash_node),
 	.insecure_elasticity	= true,	/* inserted under scx_sched_lock */
 };
 
-static struct rhashtable scx_sched_hash;
+struct rhashtable scx_sched_hash;
 #endif
 
 /* see SCX_OPS_TID_TO_TASK */
@@ -68,9 +70,9 @@ static DEFINE_RAW_SPINLOCK(scx_tasks_lock);
 static LIST_HEAD(scx_tasks);
 
 /* ops enable/disable */
-static DEFINE_MUTEX(scx_enable_mutex);
+DEFINE_MUTEX(scx_enable_mutex);
 DEFINE_STATIC_KEY_FALSE(__scx_enabled);
-DEFINE_STATIC_PERCPU_RWSEM(scx_fork_rwsem);
+DEFINE_PERCPU_RWSEM(scx_fork_rwsem);
 static atomic_t scx_enable_state_var = ATOMIC_INIT(SCX_DISABLED);
 static DEFINE_RAW_SPINLOCK(scx_bypass_lock);
 static bool scx_init_task_enabled;
@@ -79,6 +81,14 @@ DEFINE_STATIC_KEY_FALSE(__scx_switched_all);
 static DEFINE_STATIC_KEY_FALSE(__scx_tid_to_task_enabled);
 
 /*
+ * Gates cgroup ops delivery. Set at the end of the cgroup init phase of root
+ * enable and cleared before root disable starts tearing down tasks, both under
+ * scx_cgroup_lock(). Holding cgroup_lock() and seeing %true guarantees no race
+ * against root tearing down tasks.
+ */
+bool scx_cgroup_enabled;
+
+/*
  * True once SCX_OPS_TID_TO_TASK has been negotiated with the root scheduler
  * and the tid->task table is live. Wraps the static key so callers don't
  * take the address, and hints "likely enabled" for the common case where
@@ -95,13 +105,29 @@ static atomic_long_t scx_hotplug_seq = ATOMIC_LONG_INIT(0);
 /* Global cursor for the per-CPU tid allocator. Starts at 1; tid 0 is reserved. */
 static atomic64_t scx_tid_cursor = ATOMIC64_INIT(1);
 
+/* is @dsq synchronized by the containing rq lock instead of dsq->lock? */
+static bool dsq_is_rq_owned(struct scx_dispatch_q *dsq)
+{
+	switch (dsq->id) {
+	case SCX_DSQ_LOCAL:
+	case SCX_DSQ_REJECT:
+	case SCX_DSQ_RESCUE:
+		return true;
+	default:
+		return false;
+	}
+}
+
+/* Cursor for unique scx_sched instance ids. id 0 is reserved. */
+static atomic64_t scx_sched_id_cursor = ATOMIC64_INIT(0);
+
 #ifdef CONFIG_EXT_SUB_SCHED
 /*
  * The sub sched being enabled. Used by scx_disable_and_exit_task() to exit
  * tasks for the sub-sched being enabled. Use a global variable instead of a
  * per-task field as all enables are serialized.
  */
-static struct scx_sched *scx_enabling_sub_sched;
+struct scx_sched *scx_enabling_sub_sched;
 #else
 #define scx_enabling_sub_sched	(struct scx_sched *)NULL
 #endif	/* CONFIG_EXT_SUB_SCHED */
@@ -171,15 +197,6 @@ static const struct rhashtable_params dsq_hash_params = {
 
 static LLIST_HEAD(dsqs_to_free);
 
-/* string formatting from BPF */
-struct scx_bstr_buf {
-	u64			data[MAX_BPRINTF_VARARGS];
-	char			line[SCX_EXIT_MSG_LEN];
-};
-
-static DEFINE_RAW_SPINLOCK(scx_exit_bstr_buf_lock);
-static struct scx_bstr_buf scx_exit_bstr_buf;
-
 /* ops debug dump */
 static DEFINE_RAW_SPINLOCK(scx_dump_lock);
 
@@ -242,7 +259,6 @@ MODULE_PARM_DESC(bypass_lb_intv_us, "bypass load balance interval in microsecond
 
 static void run_deferred(struct rq *rq);
 static bool task_dead_and_done(struct task_struct *p);
-static void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags);
 static void scx_disable(struct scx_sched *sch, enum scx_exit_kind kind);
 
 __printf(5, 6) bool __scx_exit(struct scx_sched *sch,
@@ -272,58 +288,6 @@ static bool u32_before(u32 a, u32 b)
 	return (s32)(a - b) < 0;
 }
 
-#ifdef CONFIG_EXT_SUB_SCHED
-/**
- * scx_next_descendant_pre - find the next descendant for pre-order walk
- * @pos: the current position (%NULL to initiate traversal)
- * @root: sched whose descendants to walk
- *
- * To be used by scx_for_each_descendant_pre(). Find the next descendant to
- * visit for pre-order traversal of @root's descendants. @root is included in
- * the iteration and the first node to be visited.
- */
-static struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos,
-						 struct scx_sched *root)
-{
-	struct scx_sched *next;
-
-	lockdep_assert(lockdep_is_held(&scx_enable_mutex) ||
-		       lockdep_is_held(&scx_sched_lock));
-
-	/* if first iteration, visit @root */
-	if (!pos)
-		return root;
-
-	/* visit the first child if exists */
-	next = list_first_entry_or_null(&pos->children, struct scx_sched, sibling);
-	if (next)
-		return next;
-
-	/* no child, visit my or the closest ancestor's next sibling */
-	while (pos != root) {
-		if (!list_is_last(&pos->sibling, &scx_parent(pos)->children))
-			return list_next_entry(pos, sibling);
-		pos = scx_parent(pos);
-	}
-
-	return NULL;
-}
-
-static struct scx_sched *scx_find_sub_sched(u64 cgroup_id)
-{
-	return rhashtable_lookup(&scx_sched_hash, &cgroup_id,
-				 scx_sched_hash_params);
-}
-
-static void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch)
-{
-	rcu_assign_pointer(p->scx.sched, sch);
-}
-#else	/* CONFIG_EXT_SUB_SCHED */
-static inline struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root) { return pos ? NULL : root; }
-static inline void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch) {}
-#endif	/* CONFIG_EXT_SUB_SCHED */
-
 /**
  * scx_is_descendant - Test whether sched is a descendant
  * @sch: sched to test
@@ -331,26 +295,13 @@ static inline void scx_set_task_sched(struct task_struct *p, struct scx_sched *s
  *
  * Test whether @sch is a descendant of @ancestor.
  */
-static bool scx_is_descendant(struct scx_sched *sch, struct scx_sched *ancestor)
+bool scx_is_descendant(struct scx_sched *sch, struct scx_sched *ancestor)
 {
 	if (sch->level < ancestor->level)
 		return false;
 	return sch->ancestors[ancestor->level] == ancestor;
 }
 
-/**
- * scx_for_each_descendant_pre - pre-order walk of a sched's descendants
- * @pos: iteration cursor
- * @root: sched to walk the descendants of
- *
- * Walk @root's descendants. @root is included in the iteration and the first
- * node to be visited. Must be called with either scx_enable_mutex or
- * scx_sched_lock held.
- */
-#define scx_for_each_descendant_pre(pos, root)					\
-	for ((pos) = scx_next_descendant_pre(NULL, (root)); (pos);		\
-	     (pos) = scx_next_descendant_pre((pos), (root)))
-
 static struct scx_dispatch_q *find_global_dsq(struct scx_sched *sch, s32 cpu)
 {
 	return &sch->pnode[cpu_to_node(cpu)]->global_dsq;
@@ -369,11 +320,6 @@ static const struct sched_class *scx_setscheduler_class(struct task_struct *p)
 	return __setscheduler_class(p->policy, p->prio);
 }
 
-static struct scx_dispatch_q *bypass_dsq(struct scx_sched *sch, s32 cpu)
-{
-	return &per_cpu_ptr(sch->pcpu, cpu)->bypass_dsq;
-}
-
 static struct scx_dispatch_q *bypass_enq_target_dsq(struct scx_sched *sch, s32 cpu)
 {
 #ifdef CONFIG_EXT_SUB_SCHED
@@ -392,27 +338,7 @@ static struct scx_dispatch_q *bypass_enq_target_dsq(struct scx_sched *sch, s32 c
 		sch = scx_parent(sch);
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
-	return bypass_dsq(sch, cpu);
-}
-
-/**
- * bypass_dsp_enabled - Check if bypass dispatch path is enabled
- * @sch: scheduler to check
- *
- * When a descendant scheduler enters bypass mode, bypassed tasks are scheduled
- * by the nearest non-bypassing ancestor, or the root scheduler if all ancestors
- * are bypassing. In the former case, the ancestor is not itself bypassing but
- * its bypass DSQs will be populated with bypassed tasks from descendants. Thus,
- * the ancestor's bypass dispatch path must be active even though its own
- * bypass_depth remains zero.
- *
- * This function checks bypass_dsp_enable_depth which is managed separately from
- * bypass_depth to enable this decoupling. See enable_bypass_dsp() and
- * disable_bypass_dsp().
- */
-static bool bypass_dsp_enabled(struct scx_sched *sch)
-{
-	return unlikely(atomic_read(&sch->bypass_dsp_enable_depth));
+	return scx_bypass_dsq(sch, cpu);
 }
 
 /**
@@ -451,9 +377,9 @@ static bool rq_is_open(struct rq *rq, u64 enq_flags)
 	 * If we're in the dispatch path holding rq lock, $curr may or may not
 	 * be ready depending on whether the on-going dispatch decides to extend
 	 * $curr's slice. We say yes here and resolve it at the end of dispatch.
-	 * See balance_one().
+	 * See dispatch_one().
 	 */
-	if (rq->scx.flags & SCX_RQ_IN_BALANCE)
+	if (rq->scx.flags & SCX_RQ_IN_DISPATCH)
 		return true;
 
 	/*
@@ -461,8 +387,16 @@ static bool rq_is_open(struct rq *rq, u64 enq_flags)
 	 * so allow it to avoid spuriously triggering reenq on a combined
 	 * PREEMPT|IMMED insertion.
 	 */
-	if (enq_flags & SCX_ENQ_PREEMPT)
-		return true;
+	if (enq_flags & SCX_ENQ_PREEMPT) {
+		struct task_struct *curr = rq->curr;
+
+		/*
+		 * A protected slice refuses the preemption and the cpu stays
+		 * occupied. See rq_owned_post_enq().
+		 */
+		return curr->sched_class != &ext_sched_class ||
+			likely(!(curr->scx.flags & SCX_TASK_PROTECTED));
+	}
 
 	/*
 	 * @rq is either in transition to or running an SCX task and can't go
@@ -479,12 +413,27 @@ static bool rq_is_open(struct rq *rq, u64 enq_flags)
  */
 DEFINE_PER_CPU(struct rq *, scx_locked_rq_state);
 
+/*
+ * Under core scheduling, a pick that releases the rq lock invalidates the
+ * core-wide selection it is part of. Count the releases so that the core-sched
+ * pick can tell whether one happened across dispatch.
+ */
+static void scx_rq_lock_drop(struct rq *rq)
+{
+	lockdep_assert_rq_held(rq);
+#ifdef CONFIG_SCHED_CORE
+	if (sched_core_enabled(rq))
+		rq->scx.lock_drop_seq++;
+#endif
+}
+
 static void switch_rq_lock(struct rq *from, struct rq *to)
 {
 	bool tracked = scx_locked_rq() == from;
 
 	if (tracked)
 		update_locked_rq(NULL);
+	scx_rq_lock_drop(from);
 	raw_spin_rq_unlock(from);
 	raw_spin_rq_lock(to);
 	if (tracked)
@@ -512,27 +461,23 @@ static inline void scx_call_op_set_cpumask(struct scx_sched *sch, struct rq *rq,
 					   struct task_struct *task,
 					   const struct cpumask *cpumask)
 {
-	WARN_ON_ONCE(current->scx.kf_tasks[0]);
-	current->scx.kf_tasks[0] = task;
-	if (rq)
-		update_locked_rq(rq);
-
 	if (scx_is_cid_type()) {
 		struct scx_cmask *kern_va = *this_cpu_ptr(sch->set_cmask_scratch);
+		struct scx_cmask_ref ref;
+
 		/*
-		 * Build the per-CPU arena cmask and hand BPF its arena address.
-		 * Caller holds the rq lock with IRQs disabled, which makes us
-		 * the sole user of the scratch area.
+		 * Build the per-cpu arena cmask from kernel geometry via @ref,
+		 * never reading its BPF-writable header. set_cmask()'s __arena
+		 * argument takes the kernel address and the struct_ops
+		 * trampoline rebases it into BPF's arena pointer form. The rq
+		 * lock makes this cpu the sole kernel writer.
 		 */
-		scx_cpumask_to_cmask(cpumask, kern_va);
-		sch->ops_cid.set_cmask(task, scx_kaddr_to_arena(sch, kern_va));
+		scx_cmask_ref_init_kern(sch, kern_va, 0, num_possible_cpus(), &ref);
+		scx_cmask_ref_from_cpumask(&ref, cpumask);
+		SCX_CALL_CID_OP_TASK(sch, set_cmask, rq, task, kern_va);
 	} else {
-		sch->ops.set_cpumask(task, cpumask);
+		SCX_CALL_OP_TASK(sch, set_cpumask, rq, task, cpumask);
 	}
-
-	if (rq)
-		update_locked_rq(NULL);
-	current->scx.kf_tasks[0] = NULL;
 }
 
 enum scx_dsq_iter_flags {
@@ -688,12 +633,12 @@ struct bpf_iter_scx_dsq {
 } __attribute__((aligned(8)));
 
 
-static u32 scx_get_task_state(const struct task_struct *p)
+u32 scx_get_task_state(const struct task_struct *p)
 {
 	return p->scx.flags & SCX_TASK_STATE_MASK;
 }
 
-static void scx_set_task_state(struct task_struct *p, u32 state)
+void scx_set_task_state(struct task_struct *p, u32 state)
 {
 	u32 prev_state = scx_get_task_state(p);
 	bool warn = false;
@@ -733,23 +678,6 @@ static void scx_set_task_state(struct task_struct *p, u32 state)
 	p->scx.flags |= state;
 }
 
-/*
- * SCX task iterator.
- */
-struct scx_task_iter {
-	struct sched_ext_entity		cursor;
-	struct task_struct		*locked_task;
-	struct rq			*rq;
-	struct rq_flags			rf;
-	u32				cnt;
-	bool				list_locked;
-#ifdef CONFIG_EXT_SUB_SCHED
-	struct cgroup			*cgrp;
-	struct cgroup_subsys_state	*css_pos;
-	struct css_task_iter		css_iter;
-#endif
-};
-
 /**
  * scx_task_iter_start - Lock scx_tasks_lock and start a task iteration
  * @iter: iterator to init
@@ -778,7 +706,7 @@ struct scx_task_iter {
  * All tasks which existed when the iteration started are guaranteed to be
  * visited as long as they are not dead.
  */
-static void scx_task_iter_start(struct scx_task_iter *iter, struct cgroup *cgrp)
+void scx_task_iter_start(struct scx_task_iter *iter, struct cgroup *cgrp)
 {
 	memset(iter, 0, sizeof(*iter));
 
@@ -817,7 +745,7 @@ static void __scx_task_iter_rq_unlock(struct scx_task_iter *iter)
  * This function can be safely called anytime during an iteration. The next
  * iterator operation will automatically restore the necessary locking.
  */
-static void scx_task_iter_unlock(struct scx_task_iter *iter)
+void scx_task_iter_unlock(struct scx_task_iter *iter)
 {
 	__scx_task_iter_rq_unlock(iter);
 	if (iter->list_locked) {
@@ -860,7 +788,7 @@ static void scx_task_iter_relock(struct scx_task_iter *iter,
  * which is released on return. If the iterator holds a task's rq lock, that rq
  * lock is also released. See scx_task_iter_start() for details.
  */
-static void scx_task_iter_stop(struct scx_task_iter *iter)
+void scx_task_iter_stop(struct scx_task_iter *iter)
 {
 #ifdef CONFIG_EXT_SUB_SCHED
 	if (iter->cgrp) {
@@ -935,7 +863,7 @@ static struct task_struct *scx_task_iter_next(struct scx_task_iter *iter)
  * whether they would like to filter out dead tasks. See scx_task_iter_start()
  * for details.
  */
-static struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter)
+struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter)
 {
 	struct task_struct *p;
 
@@ -991,49 +919,13 @@ static struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter)
 }
 
 /**
- * scx_add_event - Increase an event counter for 'name' by 'cnt'
- * @sch: scx_sched to account events for
- * @name: an event name defined in struct scx_event_stats
- * @cnt: the number of the event occurred
- *
- * This can be used when preemption is not disabled.
- */
-#define scx_add_event(sch, name, cnt) do {					\
-	this_cpu_add((sch)->pcpu->event_stats.name, (cnt));			\
-	trace_sched_ext_event(#name, (cnt));					\
-} while(0)
-
-/**
- * __scx_add_event - Increase an event counter for 'name' by 'cnt'
- * @sch: scx_sched to account events for
- * @name: an event name defined in struct scx_event_stats
- * @cnt: the number of the event occurred
- *
- * This should be used only when preemption is disabled.
- */
-#define __scx_add_event(sch, name, cnt) do {					\
-	__this_cpu_add((sch)->pcpu->event_stats.name, (cnt));			\
-	trace_sched_ext_event(#name, cnt);					\
-} while(0)
-
-/**
- * scx_agg_event - Aggregate an event counter 'kind' from 'src_e' to 'dst_e'
- * @dst_e: destination event stats
- * @src_e: source event stats
- * @kind: a kind of event to be aggregated
- */
-#define scx_agg_event(dst_e, src_e, kind) do {					\
-	(dst_e)->kind += READ_ONCE((src_e)->kind);				\
-} while(0)
-
-/**
  * scx_dump_event - Dump an event 'kind' in 'events' to 's'
  * @s: output seq_buf
  * @events: event stats
  * @kind: a kind of event to dump
  */
 #define scx_dump_event(s, events, kind) do {					\
-	dump_line(&(s), "%40s: %16lld", #kind, (events)->kind);			\
+	scx_dump_line(&(s), "%40s: %16lld", #kind, (events)->kind);		\
 } while (0)
 
 
@@ -1100,28 +992,6 @@ bool scx_cpu_valid(struct scx_sched *sch, s32 cpu, const char *where)
 	}
 }
 
-/**
- * ops_sanitize_err - Sanitize a -errno value
- * @sch: scx_sched to error out on error
- * @ops_name: operation to blame on failure
- * @err: -errno value to sanitize
- *
- * Verify @err is a valid -errno. If not, trigger scx_error() and return
- * -%EPROTO. This is necessary because returning a rogue -errno up the chain can
- * cause misbehaviors. For an example, a large negative return from
- * ops.init_task() triggers an oops when passed up the call chain because the
- * value fails IS_ERR() test after being encoded with ERR_PTR() and then is
- * handled as a pointer.
- */
-static int ops_sanitize_err(struct scx_sched *sch, const char *ops_name, s32 err)
-{
-	if (err < 0 && err >= -MAX_ERRNO)
-		return err;
-
-	scx_error(sch, "ops.%s() returned an invalid errno %d", ops_name, err);
-	return -EPROTO;
-}
-
 static void deferred_bal_cb_workfn(struct rq *rq)
 {
 	run_deferred(rq);
@@ -1133,6 +1003,7 @@ static void deferred_irq_workfn(struct irq_work *irq_work)
 
 	raw_spin_rq_lock(rq);
 	run_deferred(rq);
+	scx_rq_lock_drop(rq);
 	raw_spin_rq_unlock(rq);
 }
 
@@ -1149,7 +1020,7 @@ static void schedule_deferred(struct rq *rq)
 	/*
 	 * This is the fallback when schedule_deferred_locked() can't use
 	 * the cheaper balance callback or wakeup hook paths (the target
-	 * CPU is not in balance or wakeup). Currently, this is primarily
+	 * CPU is not in dispatch or wakeup). Currently, this is primarily
 	 * hit by reenqueue operations targeting a remote CPU.
 	 *
 	 * Queue on the target CPU. The deferred work can run from any CPU
@@ -1185,31 +1056,31 @@ static void schedule_deferred_locked(struct rq *rq)
 		return;
 
 	/*
-	 * If in balance, the balance callbacks will be called before rq lock is
-	 * released. Schedule one.
+	 * If in dispatch, the balance callbacks will be called before rq lock
+	 * is released. Schedule one.
 	 *
 	 *
 	 * We can't directly insert the callback into the
 	 * rq's list: The call can drop its lock and make the pending balance
 	 * callback visible to unrelated code paths that call rq_pin_lock().
 	 *
-	 * Just let balance_one() know that it must do it itself.
+	 * Just let dispatch_one() know that it must do it itself.
 	 */
-	if (rq->scx.flags & SCX_RQ_IN_BALANCE) {
+	if (rq->scx.flags & SCX_RQ_IN_DISPATCH) {
 		rq->scx.flags |= SCX_RQ_BAL_CB_PENDING;
 		return;
 	}
 
 	/*
 	 * No scheduler hooks available. Use the generic irq_work path. The
-	 * above WAKEUP and BALANCE paths should cover most of the cases and the
-	 * time to IRQ re-enable shouldn't be long.
+	 * above WAKEUP and DISPATCH paths should cover most of the cases and
+	 * the time to IRQ re-enable shouldn't be long.
 	 */
 	schedule_deferred(rq);
 }
 
-static void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq,
-			       u64 reenq_flags, struct rq *locked_rq)
+void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq,
+			u64 reenq_flags, struct rq *locked_rq)
 {
 	struct rq *rq;
 
@@ -1223,6 +1094,18 @@ static void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 	if (dsq->id == SCX_DSQ_LOCAL) {
 		rq = container_of(dsq, struct rq, scx.local_dsq);
 
+		/*
+		 * A sub-sched lacking baseline access on the target cid has no
+		 * business triggering IPIs. The lockless test is fine: slipping
+		 * through right after a revoke is harmless and a wrong denial
+		 * can't happen - if the caller has seen its ownership, so does
+		 * this test.
+		 */
+		if (unlikely(scx_missing_caps(sch, cpu_of(rq), SCX_CAP_BASE))) {
+			__scx_add_event(sch, SCX_EV_SUB_REENQ_DENIED, 1);
+			return;
+		}
+
 		struct scx_sched_pcpu *sch_pcpu = per_cpu_ptr(sch->pcpu, cpu_of(rq));
 		struct scx_deferred_reenq_local *drl = &sch_pcpu->deferred_reenq_local;
 
@@ -1273,61 +1156,236 @@ static void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 		schedule_deferred(rq);
 }
 
-static void schedule_reenq_local(struct rq *rq, u64 reenq_flags)
-{
-	struct scx_sched *root = rcu_dereference_sched(scx_root);
+/*
+ * p->scx.slice_oob packs an out-of-band slice request into one atomic64. A zero
+ * word means no request. Otherwise the fields are:
+ *
+ *   63      SCX_SLICE_OOB_PENDING, set on every request
+ *   62-43   lower bits of issuing scheduler's id
+ *   42-0    requested slice duration in nsecs
+ *
+ * A duration of SCX_SLICE_OOB_DUR_MASK means SCX_SLICE_INF. A finite dur
+ * saturates at SCX_SLICE_OOB_DUR_MASK - 1. The id is used to detect and ignore
+ * a request that outlived a task ownership change.
+ *
+ * Only the low 20 bits of sch->id are packed, which is enough to make
+ * collisions practically impossible. A theoretical collision just lets a stale
+ * request through once.
+ */
+enum scx_slice_oob_consts {
+	SCX_SLICE_OOB_DUR_BITS	= 43,
+	SCX_SLICE_OOB_ID_BITS	= 64 - SCX_SLICE_OOB_DUR_BITS - 1,
 
-	if (WARN_ON_ONCE(!root))
-		return;
+	SCX_SLICE_OOB_DUR_MASK	= (1LLU << SCX_SLICE_OOB_DUR_BITS) - 1,
+	SCX_SLICE_OOB_ID_SHIFT	= SCX_SLICE_OOB_DUR_BITS,
+	SCX_SLICE_OOB_ID_MASK	= (1LLU << SCX_SLICE_OOB_ID_BITS) - 1,
+	SCX_SLICE_OOB_PENDING	= 1LLU << 63,
+};
 
-	schedule_dsq_reenq(root, &rq->scx.local_dsq, reenq_flags, rq);
+/*
+ * Slice and dsq_vtime write rules
+ *
+ * While @p is running, sleeping or queued on an rq-owned DSQ, both fields are
+ * protected by the rq lock. While running, the rq lock is required because
+ * update_curr_scx() RMWs the slice and the cap check for slice extension is
+ * only reliable under the rq lock.
+ *
+ * While @p is queued on a user DSQ or on the BPF side, the kernel neither
+ * consumes nor decides on the fields. Synchronizing the writers is the BPF
+ * scheduler's responsibility. An rq-locked scx_bpf_task_set_slice() write and a
+ * concurrent DSQ insertion commit can race each other and whichever lands last
+ * wins.
+ *
+ * A DSQ insert kfunc doesn't update the fields directly. The verdict carries
+ * the values and apply_slice_vtime() commits them at the insertion.
+ *
+ * scx_bpf_task_set_slice() may be called from any context and writes directly
+ * only if @p's rq lock is already held, otherwise it bounces through
+ * p->scx.slice_oob, applied under @p's rq lock at the next slice consideration.
+ *
+ * While %SCX_TASK_PROTECTED is set, every scheduler-reachable slice update is
+ * refused. See set_task_slice_keep_oob().
+ *
+ * dsq_vtime orders the next PRIQ insertion and has no running-side consumer, so
+ * scx_bpf_task_set_dsq_vtime() writes it directly. Fork-time init and direct
+ * BPF stores from non-cid-form schedulers are outside these rules.
+ */
+
+/* clear a pending slice request */
+static void clear_task_slice_oob(struct task_struct *p)
+{
+	if (unlikely(atomic64_read(&p->scx.slice_oob)))
+		atomic64_set(&p->scx.slice_oob, 0);
 }
 
 /**
- * touch_core_sched - Update timestamp used for core-sched task ordering
- * @rq: rq to read clock from, must be locked
- * @p: task to update the timestamp for
+ * dsq_insert_head - FIFO head insertion honoring %SCX_TASK_PROTECTED
+ * @dsq: DSQ to insert into
+ * @p: task being inserted
  *
- * Update @p->scx.core_sched_at timestamp. This is used by scx_prio_less() to
- * implement global or local-DSQ FIFO ordering for core-sched. Should be called
- * when a task becomes runnable and its turn on the CPU ends (e.g. slice
- * exhaustion).
+ * A HEAD insert should land behind any leading protected tasks. Return %true
+ * indicates whether @p became the first entry.
  */
-static void touch_core_sched(struct rq *rq, struct task_struct *p)
+static bool dsq_insert_head(struct scx_dispatch_q *dsq, struct task_struct *p)
 {
-	lockdep_assert_rq_held(rq);
+	struct list_head *pos = &dsq->list;
+	struct scx_dsq_list_node *node;
 
-#ifdef CONFIG_SCHED_CORE
 	/*
-	 * It's okay to update the timestamp spuriously. Use
-	 * sched_core_disabled() which is cheaper than enabled().
-	 *
-	 * As this is used to determine ordering between tasks of sibling CPUs,
-	 * it may be better to use per-core dispatch sequence instead.
+	 * Only rq-owned DSQs can hold protected tasks and the associated rq
+	 * lock keeps their flags stable.
 	 */
-	if (!sched_core_disabled())
-		p->scx.core_sched_at = sched_clock_cpu(cpu_of(rq));
-#endif
+	if (!dsq_is_rq_owned(dsq)) {
+		list_add(&p->scx.dsq_list.node, &dsq->list);
+		return true;
+	}
+
+	list_for_each_entry(node, &dsq->list, node) {
+		struct task_struct *q;
+
+		if (WARN_ON_ONCE(node->flags & SCX_DSQ_LNODE_ITER_CURSOR))
+			continue;
+
+		q = container_of(node, struct task_struct, scx.dsq_list);
+		if (!(q->scx.flags & SCX_TASK_PROTECTED))
+			break;
+
+		pos = &node->node;
+	}
+
+	list_add(&p->scx.dsq_list.node, pos);
+
+	return pos == &dsq->list;
 }
 
 /**
- * touch_core_sched_dispatch - Update core-sched timestamp on dispatch
- * @rq: rq to read clock from, must be locked
- * @p: task being dispatched
+ * set_task_slice_keep_oob - Set @p's slice, leaving any pending oob request
+ * @p: task of interest
+ * @slice: slice to set
  *
- * If the BPF scheduler implements custom core-sched ordering via
- * ops.core_sched_before(), @p->scx.core_sched_at is used to implement FIFO
- * ordering within each local DSQ. This function is called from dispatch paths
- * and updates @p->scx.core_sched_at if custom core-sched ordering is in effect.
+ * While %SCX_TASK_PROTECTED is set, BPF schedulers may not modify the slice.
+ * Refuse and return %false.
  */
-static void touch_core_sched_dispatch(struct rq *rq, struct task_struct *p)
+static bool set_task_slice_keep_oob(struct task_struct *p, u64 slice)
+{
+	lockdep_assert_rq_held(task_rq(p));
+
+	if (unlikely(p->scx.flags & SCX_TASK_PROTECTED))
+		return false;
+
+	p->scx.slice = slice;
+	return true;
+}
+
+/* set @p's slice, superseding any pending out-of-band request */
+bool scx_set_task_slice(struct task_struct *p, u64 slice)
+{
+	if (!set_task_slice_keep_oob(p, slice))
+		return false;
+	clear_task_slice_oob(p);
+	return true;
+}
+
+/**
+ * scx_task_slice_ended - @p's slice is consumed or given up
+ * @rq: rq @p is on
+ * @p: task of interest
+ *
+ * End what rides on the slice - the protection, and the rescue if @p is being
+ * rescued.
+ *
+ * A dequeue normally ends the slice too. The exception is a save/restore pair
+ * on the running task. Attribute changes like renice cycle the task through
+ * dequeue and enqueue while it keeps executing, so the slice continues. A
+ * queued task instead loses its DSQ position on any dequeue and the slice ends
+ * with it.
+ */
+void scx_task_slice_ended(struct rq *rq, struct task_struct *p)
 {
 	lockdep_assert_rq_held(rq);
 
-#ifdef CONFIG_SCHED_CORE
-	if (unlikely(SCX_HAS_OP(scx_root, core_sched_before)))
-		touch_core_sched(rq, p);
-#endif
+	p->scx.flags &= ~SCX_TASK_PROTECTED;
+	if (unlikely(p == scx_rescuee(rq)))
+		scx_rescue_end(rq);
+}
+
+/* request @p's slice to be set to @slice, see the write rules above */
+static void set_task_slice_oob(struct scx_sched *sch, struct task_struct *p, u64 slice)
+{
+	u64 dur;
+
+	if (slice == SCX_SLICE_INF) {
+		dur = SCX_SLICE_OOB_DUR_MASK;
+	} else if (unlikely(slice >= SCX_SLICE_OOB_DUR_MASK)) {
+		dur = SCX_SLICE_OOB_DUR_MASK - 1;
+		scx_add_event(sch, SCX_EV_SLICE_CLAMPED, 1);
+	} else {
+		dur = slice;
+	}
+
+	atomic64_set(&p->scx.slice_oob, SCX_SLICE_OOB_PENDING |
+		     ((sch->id & SCX_SLICE_OOB_ID_MASK) << SCX_SLICE_OOB_ID_SHIFT) | dur);
+}
+
+/*
+ * Apply a pending out-of-band slice request under @rq's lock. A request whose
+ * packed id no longer matches @p's current owner is dropped. An extension needs
+ * baseline cpu access on @p's cid, shortening is always allowed, and a
+ * protected slice refuses both. %SCX_EV_SLICE_DENIED counts the denials. See
+ * the write rules above.
+ */
+static void apply_task_slice_oob(struct rq *rq, struct task_struct *p)
+{
+	u64 oob, dur, slice;
+
+	lockdep_assert_rq_held(rq);
+
+	if (likely(!atomic64_read(&p->scx.slice_oob)))
+		return;
+
+	oob = atomic64_xchg(&p->scx.slice_oob, 0);
+	if (unlikely(!oob))
+		return;
+
+	/* the issuing scheduler no longer owns @p, drop the request */
+	if (unlikely(((oob >> SCX_SLICE_OOB_ID_SHIFT) & SCX_SLICE_OOB_ID_MASK) !=
+		     (scx_task_sched(p)->id & SCX_SLICE_OOB_ID_MASK)))
+		return;
+
+	dur = oob & SCX_SLICE_OOB_DUR_MASK;
+	slice = dur == SCX_SLICE_OOB_DUR_MASK ? SCX_SLICE_INF : dur;
+
+	if (slice > p->scx.slice &&
+	    unlikely(scx_missing_caps(scx_task_sched(p), cpu_of(rq), SCX_CAP_BASE))) {
+		__scx_add_event(scx_task_sched(p), SCX_EV_SLICE_DENIED, 1);
+		return;
+	}
+
+	if (unlikely(!set_task_slice_keep_oob(p, slice)))
+		__scx_add_event(scx_task_sched(p), SCX_EV_SLICE_DENIED, 1);
+}
+
+/*
+ * A dsq insert kfunc doesn't write slice or dsq_vtime. The verdict carries them
+ * and they are committed here, at the insertion. A zero @slice keeps the
+ * current value, floored at 1 so the task isn't treated as expired.
+ */
+static void apply_slice_vtime(struct task_struct *p, u64 slice, u64 vtime, u64 enq_flags)
+{
+	if (slice) {
+		p->scx.slice = slice;
+		/*
+		 * An explicit slice supersedes a pending oob request. A carried
+		 * default refill is not an explicit request and must keep it.
+		 */
+		if (!(enq_flags & SCX_ENQ_SLICE_DFL))
+			clear_task_slice_oob(p);
+	} else if (!p->scx.slice) {
+		p->scx.slice = 1;
+	}
+
+	if (enq_flags & SCX_ENQ_DSQ_PRIQ)
+		p->scx.dsq_vtime = vtime;
 }
 
 static void update_curr_scx(struct rq *rq)
@@ -1335,15 +1393,18 @@ static void update_curr_scx(struct rq *rq)
 	struct task_struct *curr = rq->curr;
 	s64 delta_exec;
 
+	/* apply even on 0 delta_exec, callers may still act on the slice */
+	apply_task_slice_oob(rq, curr);
+
 	delta_exec = update_curr_common(rq);
 	if (unlikely(delta_exec <= 0))
 		return;
 
-	if (curr->scx.slice != SCX_SLICE_INF) {
+	if (curr->scx.slice != SCX_SLICE_INF)
 		curr->scx.slice -= min_t(u64, curr->scx.slice, delta_exec);
-		if (!curr->scx.slice)
-			touch_core_sched(rq, curr);
-	}
+
+	if (unlikely(curr == scx_rescuee(rq)))
+		scx_rescue_charge(rq, delta_exec);
 
 	dl_server_update(&rq->ext_server, delta_exec);
 }
@@ -1369,8 +1430,8 @@ static void dsq_inc_nr(struct scx_dispatch_q *dsq, struct task_struct *p, u64 en
 	 * to the CPU or dequeued. In both cases, the only way @p can go back to
 	 * the BPF sched is through enqueueing. If being inserted into a local
 	 * DSQ with IMMED, persist the state until the next enqueueing event in
-	 * do_enqueue_task() so that we can maintain IMMED protection through
-	 * e.g. SAVE/RESTORE cycles and slice extensions.
+	 * scx_do_enqueue_task() so that we can maintain IMMED protection
+	 * through e.g. SAVE/RESTORE cycles and slice extensions.
 	 */
 	if (enq_flags & SCX_ENQ_IMMED) {
 		if (unlikely(dsq->id != SCX_DSQ_LOCAL)) {
@@ -1393,7 +1454,7 @@ static void dsq_inc_nr(struct scx_dispatch_q *dsq, struct task_struct *p, u64 en
 		 * done yet, @p can't go on the CPU immediately. Re-enqueue.
 		 */
 		if (unlikely(dsq->nr > 1 || !rq_is_open(rq, enq_flags)))
-			schedule_reenq_local(rq, 0);
+			scx_schedule_reenq_local(rq, 0);
 	}
 }
 
@@ -1415,7 +1476,11 @@ static void dsq_dec_nr(struct scx_dispatch_q *dsq, struct task_struct *p)
 
 static void refill_task_slice_dfl(struct scx_sched *sch, struct task_struct *p)
 {
-	p->scx.slice = READ_ONCE(sch->slice_dfl);
+	/*
+	 * A default refill is not an explicit request, so it must not drop a
+	 * pending out-of-band one, which is applied when @p next runs.
+	 */
+	set_task_slice_keep_oob(p, READ_ONCE(sch->slice_dfl));
 	__scx_add_event(sch, SCX_EV_REFILL_SLICE_DFL, 1);
 }
 
@@ -1450,14 +1515,23 @@ static void call_task_dequeue(struct scx_sched *sch, struct rq *rq,
 	p->scx.flags &= ~SCX_TASK_IN_CUSTODY;
 }
 
-static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq,
-			       struct task_struct *p, u64 enq_flags)
+static void rq_owned_post_enq(struct scx_sched *sch, struct rq *rq,
+			      struct scx_dispatch_q *dsq, struct task_struct *p,
+			      u64 enq_flags)
 {
-	struct rq *rq = container_of(dsq, struct rq, scx.local_dsq);
-
 	call_task_dequeue(sch, rq, p, 0);
 
 	/*
+	 * Only local inserts get the wakeup treatment below. Rejects kick the
+	 * deferred reenq and rescue parks are paced by the rescue timer.
+	 */
+	if (unlikely(dsq->id != SCX_DSQ_LOCAL)) {
+		if (dsq->id == SCX_DSQ_REJECT)
+			schedule_deferred_locked(rq);
+		return;
+	}
+
+	/*
 	 * Note that @rq's lock may be dropped between this enqueue and @p
 	 * actually getting on CPU. This gives higher-class tasks (e.g. RT)
 	 * an opportunity to wake up on @rq and prevent @p from running.
@@ -1493,34 +1567,41 @@ static void local_dsq_post_enq(struct scx_sched *sch, struct scx_dispatch_q *dsq
 		wakeup_preempt(rq, p, 0);
 
 	/*
-	 * If @rq is in balance, the CPU is already vacant and looking for the
+	 * If @rq is in dispatch, the CPU is already vacant and looking for the
 	 * next task to run. No need to preempt or trigger resched after moving
 	 * @p into its local DSQ.
 	 * Note that the wakeup_preempt() above may have already triggered
 	 * a resched if @rq->next_class was idle. It's harmless, since
 	 * need_resched is cleared immediately after task pick.
 	 */
-	if (rq->scx.flags & SCX_RQ_IN_BALANCE)
+	if (rq->scx.flags & SCX_RQ_IN_DISPATCH)
 		return;
 
 	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
 	    rq->curr->sched_class == &ext_sched_class) {
-		rq->curr->scx.slice = 0;
-		resched_curr(rq);
+		if (likely(scx_set_task_slice(rq->curr, 0)))
+			resched_curr(rq);
+		else
+			__scx_add_event(sch, SCX_EV_SLICE_DENIED, 1);
 	}
 }
 
-static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
-			     struct scx_dispatch_q *dsq, struct task_struct *p,
-			     u64 enq_flags)
+static void scx_dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
+				 struct scx_dispatch_q *dsq, struct task_struct *p,
+				 u64 slice, u64 vtime, u64 enq_flags)
 {
-	bool is_local = dsq->id == SCX_DSQ_LOCAL;
+	bool is_rq_owned = false;
+
+	if (dsq->id == SCX_DSQ_LOCAL) {
+		dsq = scx_resolve_local_dsq(sch, rq, p, &enq_flags);
+		is_rq_owned = true;
+	}
 
 	WARN_ON_ONCE(p->scx.dsq || !list_empty(&p->scx.dsq_list.node));
 	WARN_ON_ONCE((p->scx.dsq_flags & SCX_TASK_DSQ_ON_PRIQ) ||
 		     !RB_EMPTY_NODE(&p->scx.dsq_priq));
 
-	if (!is_local) {
+	if (!is_rq_owned) {
 		raw_spin_lock_nested(&dsq->lock,
 			(enq_flags & SCX_ENQ_NESTED) ? SINGLE_DEPTH_NESTING : 0);
 
@@ -1546,6 +1627,13 @@ static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
 		enq_flags &= ~SCX_ENQ_DSQ_PRIQ;
 	}
 
+	/*
+	 * @dsq is locked and @enq_flags is sanitized. Commit the carried slice
+	 * and vtime before the PRIQ insertion below reads the new dsq_vtime.
+	 */
+	if (enq_flags & SCX_ENQ_APPLY_SLICE)
+		apply_slice_vtime(p, slice, vtime, enq_flags);
+
 	if (enq_flags & SCX_ENQ_DSQ_PRIQ) {
 		struct rb_node *rbp;
 
@@ -1585,9 +1673,8 @@ static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
 				  dsq->id);
 
 		if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT)) {
-			list_add(&p->scx.dsq_list.node, &dsq->list);
 			/* new task inserted at head - use fastpath */
-			if (!(dsq->id & SCX_DSQ_FLAG_BUILTIN))
+			if (dsq_insert_head(dsq, p) && !(dsq->id & SCX_DSQ_FLAG_BUILTIN))
 				rcu_assign_pointer(dsq->first_task, p);
 		} else {
 			/*
@@ -1615,12 +1702,16 @@ static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
 	 * ops_state first, both sides would modify p->scx.flags
 	 * concurrently in a non-atomic way.
 	 */
-	if (is_local) {
-		local_dsq_post_enq(sch, dsq, p, enq_flags);
+	if (is_rq_owned) {
+		rq_owned_post_enq(sch, rq, dsq, p, enq_flags);
 	} else {
 		/*
-		 * Task on global/bypass DSQ: leave custody, task on
-		 * non-terminal DSQ: enter custody.
+		 * Global and bypass DSQs are terminal - the task leaves the
+		 * scheduler's custody, so ops.dequeue() fires here. It can run
+		 * without @p's rq lock (finish_dispatch() passes the dispatch
+		 * rq); that's safe because dequeue_task_scx() waits on
+		 * SCX_OPSS_DISPATCHING (see the ops_state note above) and so
+		 * can't race it. A non-terminal DSQ keeps the task in custody.
 		 */
 		if (dsq->id == SCX_DSQ_GLOBAL || dsq->id == SCX_DSQ_BYPASS)
 			call_task_dequeue(sch, rq, p, 0);
@@ -1638,8 +1729,7 @@ static void dispatch_enqueue(struct scx_sched *sch, struct rq *rq,
 		atomic_long_set_release(&p->scx.ops_state, SCX_OPSS_NONE);
 }
 
-static void task_unlink_from_dsq(struct task_struct *p,
-				 struct scx_dispatch_q *dsq)
+void scx_task_unlink_from_dsq(struct task_struct *p, struct scx_dispatch_q *dsq)
 {
 	WARN_ON_ONCE(list_empty(&p->scx.dsq_list.node));
 
@@ -1652,7 +1742,7 @@ static void task_unlink_from_dsq(struct task_struct *p,
 	list_del_init(&p->scx.dsq_list.node);
 	dsq_dec_nr(dsq, p);
 
-	if (!(dsq->id & SCX_DSQ_FLAG_BUILTIN) && dsq->first_task == p) {
+	if (!(dsq->id & SCX_DSQ_FLAG_BUILTIN) && rcu_access_pointer(dsq->first_task) == p) {
 		struct task_struct *first_task;
 
 		first_task = nldsq_next_task(dsq, NULL, false);
@@ -1660,10 +1750,10 @@ static void task_unlink_from_dsq(struct task_struct *p,
 	}
 }
 
-static void dispatch_dequeue(struct rq *rq, struct task_struct *p)
+void scx_dispatch_dequeue(struct rq *rq, struct task_struct *p)
 {
 	struct scx_dispatch_q *dsq = p->scx.dsq;
-	bool is_local = dsq == &rq->scx.local_dsq;
+	bool is_rq_owned = dsq && dsq_is_rq_owned(dsq);
 
 	lockdep_assert_rq_held(rq);
 
@@ -1687,7 +1777,7 @@ static void dispatch_dequeue(struct rq *rq, struct task_struct *p)
 		return;
 	}
 
-	if (!is_local)
+	if (!is_rq_owned)
 		raw_spin_lock(&dsq->lock);
 
 	/*
@@ -1696,7 +1786,7 @@ static void dispatch_dequeue(struct rq *rq, struct task_struct *p)
 	*/
 	if (p->scx.holding_cpu < 0) {
 		/* @p must still be on @dsq, dequeue */
-		task_unlink_from_dsq(p, dsq);
+		scx_task_unlink_from_dsq(p, dsq);
 	} else {
 		/*
 		 * We're racing against dispatch_to_local_dsq() which already
@@ -1709,13 +1799,13 @@ static void dispatch_dequeue(struct rq *rq, struct task_struct *p)
 	}
 	p->scx.dsq = NULL;
 
-	if (!is_local)
+	if (!is_rq_owned)
 		raw_spin_unlock(&dsq->lock);
 }
 
 /*
- * Abbreviated version of dispatch_dequeue() that can be used when both @p's rq
- * and dsq are locked.
+ * Abbreviated version of scx_dispatch_dequeue() that can be used when both
+ * @p's rq and dsq are locked.
  */
 static void dispatch_dequeue_locked(struct task_struct *p,
 				    struct scx_dispatch_q *dsq)
@@ -1723,7 +1813,7 @@ static void dispatch_dequeue_locked(struct task_struct *p,
 	lockdep_assert_rq_held(task_rq(p));
 	lockdep_assert_held(&dsq->lock);
 
-	task_unlink_from_dsq(p, dsq);
+	scx_task_unlink_from_dsq(p, dsq);
 	p->scx.dsq = NULL;
 }
 
@@ -1750,6 +1840,10 @@ static struct scx_dispatch_q *find_dsq_for_dispatch(struct scx_sched *sch,
 	else
 		dsq = find_user_dsq(sch, dsq_id);
 
+	/*
+	 * Built-in DSQs are never inserted into dsq_hash, so REJECT and RESCUE
+	 * hit the error below. They cannot be reached with an ID.
+	 */
 	if (unlikely(!dsq)) {
 		scx_error(sch, "non-existent DSQ 0x%llx", dsq_id);
 		return find_global_dsq(sch, tcpu);
@@ -1761,7 +1855,7 @@ static struct scx_dispatch_q *find_dsq_for_dispatch(struct scx_sched *sch,
 static void mark_direct_dispatch(struct scx_sched *sch,
 				 struct task_struct *ddsp_task,
 				 struct task_struct *p, u64 dsq_id,
-				 u64 enq_flags)
+				 u64 slice, u64 vtime, u64 enq_flags)
 {
 	/*
 	 * Mark that dispatch already happened from ops.select_cpu() or
@@ -1785,6 +1879,8 @@ static void mark_direct_dispatch(struct scx_sched *sch,
 	WARN_ON_ONCE(p->scx.ddsp_dsq_id != SCX_DSQ_INVALID);
 	WARN_ON_ONCE(p->scx.ddsp_enq_flags);
 
+	p->scx.ddsp_slice = slice;
+	p->scx.ddsp_vtime = vtime;
 	p->scx.ddsp_dsq_id = dsq_id;
 	p->scx.ddsp_enq_flags = enq_flags;
 }
@@ -1796,10 +1892,10 @@ static void mark_direct_dispatch(struct scx_sched *sch,
  *  - direct_dispatch(): cleared on the synchronous enqueue path, deferred
  *    dispatch keeps the state until consumed
  *  - process_ddsp_deferred_locals(): cleared after consuming deferred state,
- *  - do_enqueue_task(): cleared on enqueue fallbacks where the dispatch
+ *  - scx_do_enqueue_task(): cleared on enqueue fallbacks where the dispatch
  *    verdict is ignored (local/global/bypass)
- *  - dequeue_task_scx(): cleared after dispatch_dequeue(), covering deferred
- *    cancellation and holding_cpu races
+ *  - dequeue_task_scx(): cleared after scx_dispatch_dequeue(), covering
+ *    deferred cancellation and holding_cpu races
  *  - scx_disable_task(): cleared for queued wakeup tasks, which are excluded by
  *    the scx_bypass() loop, so that stale state is not reused by a subsequent
  *    scheduler instance
@@ -1816,9 +1912,7 @@ static void direct_dispatch(struct scx_sched *sch, struct task_struct *p,
 	struct rq *rq = task_rq(p);
 	struct scx_dispatch_q *dsq =
 		find_dsq_for_dispatch(sch, rq, p->scx.ddsp_dsq_id, task_cpu(p));
-	u64 ddsp_enq_flags;
-
-	touch_core_sched_dispatch(rq, p);
+	u64 ddsp_enq_flags, slice, vtime;
 
 	p->scx.ddsp_enq_flags |= enq_flags;
 
@@ -1858,12 +1952,15 @@ static void direct_dispatch(struct scx_sched *sch, struct task_struct *p,
 	}
 
 	ddsp_enq_flags = p->scx.ddsp_enq_flags;
+	slice = p->scx.ddsp_slice;
+	vtime = p->scx.ddsp_vtime;
 	clear_direct_dispatch(p);
 
-	dispatch_enqueue(sch, rq, dsq, p, ddsp_enq_flags | SCX_ENQ_CLEAR_OPSS);
+	scx_dispatch_enqueue(sch, rq, dsq, p, slice, vtime,
+			     ddsp_enq_flags | SCX_ENQ_APPLY_SLICE | SCX_ENQ_CLEAR_OPSS);
 }
 
-static bool scx_rq_online(struct rq *rq)
+bool scx_rq_online(struct rq *rq)
 {
 	/*
 	 * Test both cpu_active() and %SCX_RQ_ONLINE. %SCX_RQ_ONLINE indicates
@@ -1875,8 +1972,8 @@ static bool scx_rq_online(struct rq *rq)
 	return likely((rq->scx.flags & SCX_RQ_ONLINE) && cpu_active(cpu_of(rq)));
 }
 
-static void do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
-			    int sticky_cpu)
+void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
+			 int sticky_cpu)
 {
 	struct scx_sched *sch = scx_task_sched(p);
 	struct task_struct **ddsp_taskp;
@@ -1898,6 +1995,24 @@ static void do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
 	p->scx.flags &= ~SCX_TASK_IMMED;
 
 	/*
+	 * A task reenqueued too many times without running means the scheduler
+	 * keeps re-deciding a placement it can't honor, e.g. re-inserting to a
+	 * cid it lacks caps on. Eject the owning scheduler and strand the task
+	 * to be picked up during sched exit.
+	 */
+	if (enq_flags & SCX_ENQ_REENQ) {
+		if (++p->scx.reenq_cnt > 1)
+			__scx_add_event(sch, SCX_EV_REENQ_REPEAT, 1);
+
+		if (unlikely(p->scx.reenq_cnt > SCX_REENQ_MAX_REPEAT)) {
+			__scx_exit(sch, SCX_EXIT_ERROR_REENQ, 0, cpu_of(rq),
+				   "%s[%d] reenqueued %u times without running",
+				   p->comm, p->pid, p->scx.reenq_cnt);
+			return;
+		}
+	}
+
+	/*
 	 * If !scx_rq_online(), we already told the BPF scheduler that the CPU
 	 * is offline and are just running the hotplug path. Don't bother the
 	 * BPF scheduler.
@@ -1917,6 +2032,7 @@ static void do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
 	if (!(sch->ops.flags & SCX_OPS_ENQ_EXITING) &&
 	    unlikely(p->flags & PF_EXITING)) {
 		__scx_add_event(sch, SCX_EV_ENQ_SKIP_EXITING, 1);
+		enq_flags |= SCX_ENQ_RESCUE;	/* avoid looping on cap rejection */
 		goto local;
 	}
 
@@ -1963,7 +2079,7 @@ direct:
 	direct_dispatch(sch, p, enq_flags);
 	return;
 local_norefill:
-	dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, enq_flags);
+	scx_dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, 0, 0, enq_flags);
 	return;
 local:
 	dsq = &rq->scx.local_dsq;
@@ -1976,15 +2092,9 @@ bypass:
 	goto enqueue;
 
 enqueue:
-	/*
-	 * For task-ordering, slice refill must be treated as implying the end
-	 * of the current slice. Otherwise, the longer @p stays on the CPU, the
-	 * higher priority it becomes from scx_prio_less()'s POV.
-	 */
-	touch_core_sched(rq, p);
 	refill_task_slice_dfl(sch, p);
 	clear_direct_dispatch(p);
-	dispatch_enqueue(sch, rq, dsq, p, enq_flags);
+	scx_dispatch_enqueue(sch, rq, dsq, p, 0, 0, enq_flags);
 }
 
 static bool task_runnable(const struct task_struct *p)
@@ -2006,20 +2116,30 @@ static void set_task_runnable(struct rq *rq, struct task_struct *p)
 	 * appended to the runnable_list.
 	 */
 	list_add_tail(&p->scx.runnable_node, &rq->scx.runnable_list);
+
+	/*
+	 * Record the rq @p is runnable on, maintained under the rq lock so it
+	 * stays valid unlike task_cpu(), which a remote wakeup can move under
+	 * pi_lock alone.
+	 */
+	WRITE_ONCE(p->scx.runnable_cpu, cpu_of(rq));
 }
 
 static void clr_task_runnable(struct task_struct *p, bool reset_runnable_at)
 {
 	list_del_init(&p->scx.runnable_node);
-	if (reset_runnable_at)
+	WRITE_ONCE(p->scx.runnable_cpu, -1);
+	if (reset_runnable_at) {
 		p->scx.flags |= SCX_TASK_RESET_RUNNABLE_AT;
+		p->scx.reenq_cnt = 0;
+	}
 }
 
 static void enqueue_task_scx(struct rq *rq, struct task_struct *p, int core_enq_flags)
 {
 	struct scx_sched *sch = scx_task_sched(p);
 	int sticky_cpu = p->scx.sticky_cpu;
-	u64 enq_flags = core_enq_flags | rq->scx.extra_enq_flags;
+	u64 enq_flags = core_enq_flags | rq->scx.remote_activate_enq_flags;
 
 	if (enq_flags & ENQUEUE_WAKEUP)
 		rq->scx.flags |= SCX_RQ_IN_WAKEUP;
@@ -2028,10 +2148,13 @@ static void enqueue_task_scx(struct rq *rq, struct task_struct *p, int core_enq_
 	 * Restoring a running task will be immediately followed by
 	 * set_next_task_scx() which expects the task to not be on the BPF
 	 * scheduler as tasks can only start running through local DSQs. Force
-	 * direct-dispatch into the local DSQ by setting the sticky_cpu.
+	 * direct-dispatch into the local DSQ by setting the sticky_cpu. Mark
+	 * IGNORE_CAPS to force entry into the local DSQ.
 	 */
-	if (unlikely(enq_flags & ENQUEUE_RESTORE) && task_current(rq, p))
+	if (unlikely(enq_flags & ENQUEUE_RESTORE) && task_current(rq, p)) {
 		sticky_cpu = cpu_of(rq);
+		enq_flags |= SCX_ENQ_IGNORE_CAPS;
+	}
 
 	if (p->scx.flags & SCX_TASK_QUEUED) {
 		WARN_ON_ONCE(!task_runnable(p));
@@ -2046,14 +2169,11 @@ static void enqueue_task_scx(struct rq *rq, struct task_struct *p, int core_enq_
 	if (SCX_HAS_OP(sch, runnable) && !task_on_rq_migrating(p))
 		SCX_CALL_OP_TASK(sch, runnable, rq, p, enq_flags);
 
-	if (enq_flags & SCX_ENQ_WAKEUP)
-		touch_core_sched(rq, p);
-
 	/* Start dl_server if this is the first task being enqueued */
 	if (rq->scx.nr_running == 1)
 		dl_server_start(&rq->ext_server);
 
-	do_enqueue_task(rq, p, enq_flags, sticky_cpu);
+	scx_do_enqueue_task(rq, p, enq_flags, sticky_cpu);
 
 	if (sticky_cpu >= 0)
 		p->scx.sticky_cpu = -1;
@@ -2148,9 +2268,9 @@ static bool dequeue_task_scx(struct rq *rq, struct task_struct *p, int core_deq_
 
 	/*
 	 * Set %SCX_DEQ_SCHED_CHANGE when the dequeue is due to a property
-	 * change (not sleep or core-sched pick).
+	 * change (not sleep).
 	 */
-	if (!(deq_flags & (DEQUEUE_SLEEP | SCX_DEQ_CORE_SCHED_EXEC)))
+	if (!(deq_flags & DEQUEUE_SLEEP))
 		deq_flags |= SCX_DEQ_SCHED_CHANGE;
 
 	if (!(p->scx.flags & SCX_TASK_QUEUED)) {
@@ -2168,13 +2288,15 @@ static bool dequeue_task_scx(struct rq *rq, struct task_struct *p, int core_deq_
 	 *
 	 * @p may go through multiple stopping <-> running transitions between
 	 * here and put_prev_task_scx() if task attribute changes occur while
-	 * balance_one() leaves @rq unlocked. However, they don't contain any
+	 * dispatch_one() leaves @rq unlocked. However, they don't contain any
 	 * information meaningful to the BPF scheduler and can be suppressed by
 	 * skipping the callbacks if the task is !QUEUED.
 	 */
-	if (SCX_HAS_OP(sch, stopping) && task_current(rq, p)) {
+	if (task_current(rq, p) &&
+	    (SCX_HAS_OP(sch, stopping) || unlikely(p == scx_rescuee(rq)))) {
 		update_curr_scx(rq);
-		SCX_CALL_OP_TASK(sch, stopping, rq, p, false);
+		if (SCX_HAS_OP(sch, stopping))
+			SCX_CALL_OP_TASK(sch, stopping, rq, p, false);
 	}
 
 	if (SCX_HAS_OP(sch, quiescent) && !task_on_rq_migrating(p))
@@ -2189,7 +2311,12 @@ static bool dequeue_task_scx(struct rq *rq, struct task_struct *p, int core_deq_
 	rq->scx.nr_running--;
 	sub_nr_running(rq, 1);
 
-	dispatch_dequeue(rq, p);
+	scx_dispatch_dequeue(rq, p);
+
+	/* see scx_task_slice_ended() for the save/restore exception */
+	if (!((deq_flags & DEQUEUE_SAVE) && task_current(rq, p)))
+		scx_task_slice_ended(rq, p);
+
 	clear_direct_dispatch(p);
 	return true;
 }
@@ -2199,10 +2326,13 @@ static void yield_task_scx(struct rq *rq)
 	struct task_struct *p = rq->donor;
 	struct scx_sched *sch = scx_task_sched(p);
 
+	/* a yield gives the slice up */
+	scx_task_slice_ended(rq, p);
+
 	if (SCX_HAS_OP(sch, yield))
 		SCX_CALL_OP_2TASKS_RET(sch, yield, rq, p, NULL);
 	else
-		p->scx.slice = 0;
+		scx_set_task_slice(p, 0);
 }
 
 static bool yield_to_task_scx(struct rq *rq, struct task_struct *to)
@@ -2210,6 +2340,9 @@ static bool yield_to_task_scx(struct rq *rq, struct task_struct *to)
 	struct task_struct *from = rq->donor;
 	struct scx_sched *sch = scx_task_sched(from);
 
+	/* like a plain yield, giving the slice up ends the protection */
+	scx_task_slice_ended(rq, from);
+
 	if (SCX_HAS_OP(sch, yield) && sch == scx_task_sched(to))
 		return SCX_CALL_OP_2TASKS_RET(sch, yield, rq, from, to);
 	else
@@ -2237,35 +2370,36 @@ static void wakeup_preempt_scx(struct rq *rq, struct task_struct *p, int wake_fl
 	 * - A higher-priority wakes up while SCX dispatch is in progress.
 	 */
 	if (rq->scx.nr_immed)
-		schedule_reenq_local(rq, 0);
+		scx_schedule_reenq_local(rq, 0);
 }
 
-static void move_local_task_to_local_dsq(struct scx_sched *sch,
-					 struct task_struct *p, u64 enq_flags,
-					 struct scx_dispatch_q *src_dsq,
-					 struct rq *dst_rq)
+void scx_move_local_task_to_local_dsq(struct scx_sched *sch, struct task_struct *p,
+				      u64 enq_flags, struct scx_dispatch_q *src_dsq,
+				      struct rq *dst_rq)
 {
-	struct scx_dispatch_q *dst_dsq = &dst_rq->scx.local_dsq;
+	struct scx_dispatch_q *dst_dsq = scx_resolve_local_dsq(sch, dst_rq, p, &enq_flags);
 
-	/* @dsq is locked and @p is on @dst_rq */
-	lockdep_assert_held(&src_dsq->lock);
+	/* @p is on @dst_rq, an rq-owned @src_dsq is covered by the rq lock */
+	if (!dsq_is_rq_owned(src_dsq))
+		lockdep_assert_held(&src_dsq->lock);
 	lockdep_assert_rq_held(dst_rq);
 
 	WARN_ON_ONCE(p->scx.holding_cpu >= 0);
 
 	if (enq_flags & (SCX_ENQ_HEAD | SCX_ENQ_PREEMPT))
-		list_add(&p->scx.dsq_list.node, &dst_dsq->list);
+		dsq_insert_head(dst_dsq, p);
 	else
 		list_add_tail(&p->scx.dsq_list.node, &dst_dsq->list);
 
 	dsq_inc_nr(dst_dsq, p, enq_flags);
 	p->scx.dsq = dst_dsq;
 
-	local_dsq_post_enq(sch, dst_dsq, p, enq_flags);
+	rq_owned_post_enq(sch, dst_rq, dst_dsq, p, enq_flags);
 }
 
 /**
  * move_remote_task_to_local_dsq - Move a task from a foreign rq to a local DSQ
+ * @sch: scheduler placing @p
  * @p: task to move
  * @enq_flags: %SCX_ENQ_*
  * @src_rq: rq to move the task from, locked on entry, released on return
@@ -2273,7 +2407,8 @@ static void move_local_task_to_local_dsq(struct scx_sched *sch,
  *
  * Move @p which is currently on @src_rq to @dst_rq's local DSQ.
  */
-static void move_remote_task_to_local_dsq(struct task_struct *p, u64 enq_flags,
+static void move_remote_task_to_local_dsq(struct scx_sched *sch,
+					  struct task_struct *p, u64 enq_flags,
 					  struct rq *src_rq, struct rq *dst_rq)
 {
 	lockdep_assert_rq_held(src_rq);
@@ -2289,15 +2424,19 @@ static void move_remote_task_to_local_dsq(struct task_struct *p, u64 enq_flags,
 	switch_rq_lock(src_rq, dst_rq);
 
 	/*
-	 * We want to pass scx-specific enq_flags but activate_task() will
-	 * truncate the upper 32 bit. As we own @rq, we can pass them through
-	 * @rq->scx.extra_enq_flags instead.
+	 * activate_task() below truncates enq_flags to 32 bits and re-derives
+	 * @p's owner, dropping our scx flags and the placing @sch. We own @rq,
+	 * so stash both across the call. The enqueue reads them back, keeping
+	 * the scx flags and checking caps against the placer, not the owner.
 	 */
 	WARN_ON_ONCE(!cpumask_test_cpu(cpu_of(dst_rq), p->cpus_ptr));
-	WARN_ON_ONCE(dst_rq->scx.extra_enq_flags);
-	dst_rq->scx.extra_enq_flags = enq_flags;
+	WARN_ON_ONCE(dst_rq->scx.remote_activate_enq_flags ||
+		     dst_rq->scx.remote_activate_sch);
+	dst_rq->scx.remote_activate_enq_flags = enq_flags;
+	dst_rq->scx.remote_activate_sch = sch;
 	activate_task(dst_rq, p, 0);
-	dst_rq->scx.extra_enq_flags = 0;
+	dst_rq->scx.remote_activate_enq_flags = 0;
+	dst_rq->scx.remote_activate_sch = NULL;
 }
 
 /*
@@ -2378,13 +2517,14 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
 }
 
 /**
- * unlink_dsq_and_lock_src_rq() - Unlink task from its DSQ and lock its task_rq
+ * unlink_dsq_and_switch_rq_lock() - Unlink task and switch to its rq lock
  * @p: target task
  * @dsq: locked DSQ @p is currently on
+ * @locked_rq: currently locked rq
  * @src_rq: rq @p is currently on, stable with @dsq locked
  *
- * Called with @dsq locked but no rq's locked. We want to move @p to a different
- * DSQ, including any local DSQ, but are not locking @src_rq. Locking @src_rq is
+ * Called with @dsq and @locked_rq locked. We want to move @p to a different DSQ,
+ * including any local DSQ, but are not locking @src_rq. Locking @src_rq is
  * required when transferring into a local DSQ. Even when transferring into a
  * non-local DSQ, it's better to use the same mechanism to protect against
  * dequeues and maintain the invariant that @p->scx.dsq can only change while
@@ -2401,43 +2541,42 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
  * values afterwards, as this operation can't be preempted or recurse, the
  * holding_cpu can never become this CPU again before we're done. Thus, we can
  * tell whether we lost to dequeue by testing whether the holding_cpu still
- * points to this CPU. See dispatch_dequeue() for the counterpart.
+ * points to this CPU. See scx_dispatch_dequeue() for the counterpart.
  *
  * On return, @dsq is unlocked and @src_rq is locked. Returns %true if @p is
  * still valid. %false if lost to dequeue.
  */
-static bool unlink_dsq_and_lock_src_rq(struct task_struct *p,
-				       struct scx_dispatch_q *dsq,
-				       struct rq *src_rq)
+static bool unlink_dsq_and_switch_rq_lock(struct task_struct *p,
+					  struct scx_dispatch_q *dsq,
+					  struct rq *locked_rq,
+					  struct rq *src_rq)
 {
 	s32 cpu = raw_smp_processor_id();
 
 	lockdep_assert_held(&dsq->lock);
+	lockdep_assert_rq_held(locked_rq);
 
 	WARN_ON_ONCE(p->scx.holding_cpu >= 0);
-	task_unlink_from_dsq(p, dsq);
+	scx_task_unlink_from_dsq(p, dsq);
 	p->scx.holding_cpu = cpu;
 
 	raw_spin_unlock(&dsq->lock);
-	raw_spin_rq_lock(src_rq);
+	switch_rq_lock(locked_rq, src_rq);
 
 	/* task_rq couldn't have changed if we're still the holding cpu */
 	return likely(p->scx.holding_cpu == cpu) &&
 		!WARN_ON_ONCE(src_rq != task_rq(p));
 }
 
-static bool consume_remote_task(struct rq *this_rq,
+static bool consume_remote_task(struct scx_sched *sch, struct rq *this_rq,
 				struct task_struct *p, u64 enq_flags,
 				struct scx_dispatch_q *dsq, struct rq *src_rq)
 {
-	raw_spin_rq_unlock(this_rq);
-
-	if (unlink_dsq_and_lock_src_rq(p, dsq, src_rq)) {
-		move_remote_task_to_local_dsq(p, enq_flags, src_rq, this_rq);
+	if (unlink_dsq_and_switch_rq_lock(p, dsq, this_rq, src_rq)) {
+		move_remote_task_to_local_dsq(sch, p, enq_flags, src_rq, this_rq);
 		return true;
 	} else {
-		raw_spin_rq_unlock(src_rq);
-		raw_spin_rq_lock(this_rq);
+		switch_rq_lock(src_rq, this_rq);
 		return false;
 	}
 }
@@ -2489,14 +2628,12 @@ static struct rq *move_task_between_dsqs(struct scx_sched *sch,
 	if (dst_dsq->id == SCX_DSQ_LOCAL) {
 		/* @p is going from a non-local DSQ to a local DSQ */
 		if (src_rq == dst_rq) {
-			task_unlink_from_dsq(p, src_dsq);
-			move_local_task_to_local_dsq(sch, p, enq_flags,
-						     src_dsq, dst_rq);
+			scx_task_unlink_from_dsq(p, src_dsq);
+			scx_move_local_task_to_local_dsq(sch, p, enq_flags, src_dsq, dst_rq);
 			raw_spin_unlock(&src_dsq->lock);
 		} else {
 			raw_spin_unlock(&src_dsq->lock);
-			move_remote_task_to_local_dsq(p, enq_flags,
-						      src_rq, dst_rq);
+			move_remote_task_to_local_dsq(sch, p, enq_flags, src_rq, dst_rq);
 		}
 	} else {
 		/*
@@ -2506,14 +2643,14 @@ static struct rq *move_task_between_dsqs(struct scx_sched *sch,
 		dispatch_dequeue_locked(p, src_dsq);
 		raw_spin_unlock(&src_dsq->lock);
 
-		dispatch_enqueue(sch, dst_rq, dst_dsq, p, enq_flags);
+		scx_dispatch_enqueue(sch, dst_rq, dst_dsq, p, 0, 0, enq_flags);
 	}
 
 	return dst_rq;
 }
 
-static bool consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
-			       struct scx_dispatch_q *dsq, u64 enq_flags)
+bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
+			    struct scx_dispatch_q *dsq, u64 enq_flags)
 {
 	struct task_struct *p;
 retry:
@@ -2542,14 +2679,14 @@ retry:
 			break;
 
 		if (rq == task_rq) {
-			task_unlink_from_dsq(p, dsq);
-			move_local_task_to_local_dsq(sch, p, enq_flags, dsq, rq);
+			scx_task_unlink_from_dsq(p, dsq);
+			scx_move_local_task_to_local_dsq(sch, p, enq_flags, dsq, rq);
 			raw_spin_unlock(&dsq->lock);
 			return true;
 		}
 
 		if (task_can_run_on_remote_rq(sch, p, rq, false)) {
-			if (likely(consume_remote_task(rq, p, enq_flags, dsq, task_rq)))
+			if (likely(consume_remote_task(sch, rq, p, enq_flags, dsq, task_rq)))
 				return true;
 			goto retry;
 		}
@@ -2559,11 +2696,11 @@ retry:
 	return false;
 }
 
-static bool consume_global_dsq(struct scx_sched *sch, struct rq *rq)
+bool scx_consume_global_dsq(struct scx_sched *sch, struct rq *rq)
 {
 	int node = cpu_to_node(cpu_of(rq));
 
-	return consume_dispatch_q(sch, rq, &sch->pnode[node]->global_dsq, 0);
+	return scx_consume_dispatch_q(sch, rq, &sch->pnode[node]->global_dsq, 0);
 }
 
 /**
@@ -2572,6 +2709,8 @@ static bool consume_global_dsq(struct scx_sched *sch, struct rq *rq)
  * @rq: current rq which is locked
  * @dst_dsq: destination DSQ
  * @p: task to dispatch
+ * @slice: slice carried by the insert verdict, 0 keeps the current value
+ * @vtime: vtime carried by the insert verdict, committed on PRIQ inserts
  * @enq_flags: %SCX_ENQ_*
  *
  * We're holding @rq lock and want to dispatch @p to @dst_dsq which is a local
@@ -2582,8 +2721,8 @@ static bool consume_global_dsq(struct scx_sched *sch, struct rq *rq)
  * %SCX_OPSS_DISPATCHING).
  */
 static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
-				  struct scx_dispatch_q *dst_dsq,
-				  struct task_struct *p, u64 enq_flags)
+				  struct scx_dispatch_q *dst_dsq, struct task_struct *p,
+				  u64 slice, u64 vtime, u64 enq_flags)
 {
 	struct rq *src_rq = task_rq(p);
 	struct rq *dst_rq = container_of(dst_dsq, struct rq, scx.local_dsq);
@@ -2596,8 +2735,8 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 	 * If dispatching to @rq that @p is already on, no lock dancing needed.
 	 */
 	if (rq == src_rq && rq == dst_rq) {
-		dispatch_enqueue(sch, rq, dst_dsq, p,
-				 enq_flags | SCX_ENQ_CLEAR_OPSS);
+		scx_dispatch_enqueue(sch, rq, dst_dsq, p, slice, vtime,
+				     enq_flags | SCX_ENQ_APPLY_SLICE | SCX_ENQ_CLEAR_OPSS);
 		return;
 	}
 
@@ -2610,7 +2749,7 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 	 * As DISPATCHING guarantees that @p is wholly ours, we can pretend that
 	 * we're moving from a DSQ and use the same mechanism - mark the task
 	 * under transfer with holding_cpu, release DISPATCHING and then follow
-	 * the same protocol. See unlink_dsq_and_lock_src_rq().
+	 * the same protocol. See unlink_dsq_and_switch_rq_lock().
 	 */
 	p->scx.holding_cpu = raw_smp_processor_id();
 
@@ -2634,16 +2773,18 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 		 */
 		if (src_rq == dst_rq) {
 			p->scx.holding_cpu = -1;
-			dispatch_enqueue(sch, dst_rq, &dst_rq->scx.local_dsq, p,
-					 enq_flags);
+			scx_dispatch_enqueue(sch, dst_rq, &dst_rq->scx.local_dsq, p,
+					     slice, vtime, enq_flags | SCX_ENQ_APPLY_SLICE);
 		} else if (unlikely(!task_can_run_on_remote_rq(sch, p, dst_rq, true))) {
 			p->scx.holding_cpu = -1;
 			fallback = true;
-			dispatch_enqueue(sch, src_rq, find_global_dsq(sch, task_cpu(p)),
-					 p, enq_flags | SCX_ENQ_GDSQ_FALLBACK);
+			scx_dispatch_enqueue(sch, src_rq, find_global_dsq(sch, task_cpu(p)),
+					     p, slice, vtime,
+					     enq_flags | SCX_ENQ_APPLY_SLICE |
+					     SCX_ENQ_GDSQ_FALLBACK);
 		} else {
-			move_remote_task_to_local_dsq(p, enq_flags,
-						      src_rq, dst_rq);
+			apply_slice_vtime(p, slice, vtime, enq_flags);
+			move_remote_task_to_local_dsq(sch, p, enq_flags, src_rq, dst_rq);
 			/* task has been moved to dst_rq, which is now locked */
 			locked_rq = dst_rq;
 		}
@@ -2660,6 +2801,7 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 
 /**
  * finish_dispatch - Asynchronously finish dispatching a task
+ * @sch: the scheduler
  * @rq: current rq which is locked
  * @p: task to finish dispatching
  * @qseq_at_dispatch: qseq when @p started getting dispatched
@@ -2677,15 +2819,13 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
  * was valid in the first place. Make sure that the task is still owned by the
  * BPF scheduler and claim the ownership before dispatching.
  */
-static void finish_dispatch(struct scx_sched *sch, struct rq *rq,
-			    struct task_struct *p,
-			    unsigned long qseq_at_dispatch,
-			    u64 dsq_id, u64 enq_flags)
+static void finish_dispatch(struct scx_sched *sch, struct rq *rq, struct task_struct *p,
+			    unsigned long qseq_at_dispatch, u64 dsq_id,
+			    u64 slice, u64 vtime, u64 enq_flags)
 {
 	struct scx_dispatch_q *dsq;
 	unsigned long opss;
 
-	touch_core_sched_dispatch(rq, p);
 retry:
 	/*
 	 * No need for _acquire here. @p is accessed only after a successful
@@ -2726,10 +2866,10 @@ retry:
 		goto retry;
 	case SCX_OPSS_QUEUEING:
 		/*
-		 * do_enqueue_task() is in the process of transferring the task
-		 * to the BPF scheduler while holding @p's rq lock. As we aren't
-		 * holding any kernel or BPF resource that the enqueue path may
-		 * depend upon, it's safe to wait.
+		 * scx_do_enqueue_task() is in the process of transferring the
+		 * task to the BPF scheduler while holding @p's rq lock. As we
+		 * aren't holding any kernel or BPF resource that the enqueue
+		 * path may depend upon, it's safe to wait.
 		 */
 		wait_ops_state(p, opss);
 		goto retry;
@@ -2737,15 +2877,16 @@ retry:
 
 	BUG_ON(!(p->scx.flags & SCX_TASK_QUEUED));
 
-	dsq = find_dsq_for_dispatch(sch, this_rq(), dsq_id, task_cpu(p));
+	dsq = find_dsq_for_dispatch(sch, rq, dsq_id, task_cpu(p));
 
 	if (dsq->id == SCX_DSQ_LOCAL)
-		dispatch_to_local_dsq(sch, rq, dsq, p, enq_flags);
+		dispatch_to_local_dsq(sch, rq, dsq, p, slice, vtime, enq_flags);
 	else
-		dispatch_enqueue(sch, rq, dsq, p, enq_flags | SCX_ENQ_CLEAR_OPSS);
+		scx_dispatch_enqueue(sch, rq, dsq, p, slice, vtime,
+				     enq_flags | SCX_ENQ_APPLY_SLICE | SCX_ENQ_CLEAR_OPSS);
 }
 
-static void flush_dispatch_buf(struct scx_sched *sch, struct rq *rq)
+void scx_flush_dispatch_buf(struct scx_sched *sch, struct rq *rq)
 {
 	struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
 	u32 u;
@@ -2754,7 +2895,7 @@ static void flush_dispatch_buf(struct scx_sched *sch, struct rq *rq)
 		struct scx_dsp_buf_ent *ent = &dspc->buf[u];
 
 		finish_dispatch(sch, rq, ent->task, ent->qseq, ent->dsq_id,
-				ent->enq_flags);
+				ent->slice, ent->vtime, ent->enq_flags);
 	}
 
 	dspc->nr_tasks += dspc->cursor;
@@ -2774,123 +2915,16 @@ static inline void maybe_queue_balance_callback(struct rq *rq)
 	rq->scx.flags &= ~SCX_RQ_BAL_CB_PENDING;
 }
 
-/*
- * One user of this function is scx_bpf_dispatch() which can be called
- * recursively as sub-sched dispatches nest. Always inline to reduce stack usage
- * from the call frame.
- */
-static __always_inline bool
-scx_dispatch_sched(struct scx_sched *sch, struct rq *rq,
-		   struct task_struct *prev, bool nested)
+static enum scx_dsp_verdict dispatch_one(struct rq *rq, struct task_struct *prev)
 {
-	struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
-	int nr_loops = SCX_DSP_MAX_LOOPS;
-	s32 cpu = cpu_of(rq);
-	bool prev_on_sch = (prev->sched_class == &ext_sched_class) &&
-		scx_task_on_sched(sch, prev);
-
-	if (consume_global_dsq(sch, rq))
-		return true;
-
-	if (bypass_dsp_enabled(sch)) {
-		/* if @sch is bypassing, only the bypass DSQs are active */
-		if (scx_bypassing(sch, cpu))
-			return consume_dispatch_q(sch, rq, bypass_dsq(sch, cpu), 0);
-
-#ifdef CONFIG_EXT_SUB_SCHED
-		/*
-		 * If @sch isn't bypassing but its children are, @sch is
-		 * responsible for making forward progress for both its own
-		 * tasks that aren't bypassing and the bypassing descendants'
-		 * tasks. The following implements a simple built-in behavior -
-		 * let each CPU try to run the bypass DSQ every Nth time.
-		 *
-		 * Later, if necessary, we can add an ops flag to suppress the
-		 * auto-consumption and a kfunc to consume the bypass DSQ and,
-		 * so that the BPF scheduler can fully control scheduling of
-		 * bypassed tasks.
-		 */
-		struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
-
-		if (!(pcpu->bypass_host_seq++ % SCX_BYPASS_HOST_NTH) &&
-		    consume_dispatch_q(sch, rq, bypass_dsq(sch, cpu), 0)) {
-			__scx_add_event(sch, SCX_EV_SUB_BYPASS_DISPATCH, 1);
-			return true;
-		}
-#endif	/* CONFIG_EXT_SUB_SCHED */
-	}
-
-	if (unlikely(!SCX_HAS_OP(sch, dispatch)) || !scx_rq_online(rq))
-		return false;
-
-	dspc->rq = rq;
-
-	/*
-	 * The dispatch loop. Because flush_dispatch_buf() may drop the rq lock,
-	 * the local DSQ might still end up empty after a successful
-	 * ops.dispatch(). If the local DSQ is empty even after ops.dispatch()
-	 * produced some tasks, retry. The BPF scheduler may depend on this
-	 * looping behavior to simplify its implementation.
-	 */
-	do {
-		dspc->nr_tasks = 0;
-
-		if (nested) {
-			SCX_CALL_OP(sch, dispatch, rq, scx_cpu_arg(cpu),
-				    prev_on_sch ? prev : NULL);
-		} else {
-			/* stash @prev so that nested invocations can access it */
-			rq->scx.sub_dispatch_prev = prev;
-			SCX_CALL_OP(sch, dispatch, rq, scx_cpu_arg(cpu),
-				    prev_on_sch ? prev : NULL);
-			rq->scx.sub_dispatch_prev = NULL;
-		}
-
-		flush_dispatch_buf(sch, rq);
-
-		if ((prev->scx.flags & SCX_TASK_QUEUED) && prev->scx.slice) {
-			rq->scx.flags |= SCX_RQ_BAL_KEEP;
-			return true;
-		}
-		if (rq->scx.local_dsq.nr)
-			return true;
-		if (consume_global_dsq(sch, rq))
-			return true;
-
-		/*
-		 * ops.dispatch() can trap us in this loop by repeatedly
-		 * dispatching ineligible tasks. Break out once in a while to
-		 * allow the watchdog to run. As IRQ can't be enabled in
-		 * balance(), we want to complete this scheduling cycle and then
-		 * start a new one. IOW, we want to call resched_curr() on the
-		 * next, most likely idle, task, not the current one. Use
-		 * __scx_bpf_kick_cpu() for deferred kicking.
-		 */
-		if (unlikely(!--nr_loops)) {
-			scx_kick_cpu(sch, cpu, 0);
-			break;
-		}
-	} while (dspc->nr_tasks);
-
-	/*
-	 * Prevent the CPU from going idle while bypassed descendants have tasks
-	 * queued. Without this fallback, bypassed tasks could stall if the host
-	 * scheduler's ops.dispatch() doesn't yield any tasks.
-	 */
-	if (bypass_dsp_enabled(sch))
-		return consume_dispatch_q(sch, rq, bypass_dsq(sch, cpu), 0);
-
-	return false;
-}
-
-static int balance_one(struct rq *rq, struct task_struct *prev)
-{
-	struct scx_sched *sch = scx_root;
+	struct scx_sched *sch = scx_root_protected_live();
+	enum scx_dsp_verdict verdict;
 	s32 cpu = cpu_of(rq);
 
 	lockdep_assert_rq_held(rq);
-	rq->scx.flags |= SCX_RQ_IN_BALANCE;
-	rq->scx.flags &= ~SCX_RQ_BAL_KEEP;
+	rq->scx.flags |= SCX_RQ_IN_DISPATCH;
+
+	scx_process_sync_ecaps(rq, prev);
 
 	if ((sch->ops.flags & SCX_OPS_HAS_CPU_PREEMPT) &&
 	    unlikely(rq->scx.cpu_released)) {
@@ -2920,16 +2954,19 @@ static int balance_one(struct rq *rq, struct task_struct *prev)
 		 */
 		if ((prev->scx.flags & SCX_TASK_QUEUED) && prev->scx.slice &&
 		    !scx_bypassing(sch, cpu)) {
-			rq->scx.flags |= SCX_RQ_BAL_KEEP;
+			verdict = SCX_DSP_PREV;
 			goto has_tasks;
 		}
 	}
 
 	/* if there already are tasks to run, nothing to do */
-	if (rq->scx.local_dsq.nr)
+	if (rq->scx.local_dsq.nr) {
+		verdict = SCX_DSP_LOCAL;
 		goto has_tasks;
+	}
 
-	if (scx_dispatch_sched(sch, rq, prev, false))
+	verdict = scx_dispatch_sched(sch, rq, prev, false);
+	if (verdict != SCX_DSP_NONE)
 		goto has_tasks;
 
 	/*
@@ -2937,13 +2974,14 @@ static int balance_one(struct rq *rq, struct task_struct *prev)
 	 * %SCX_OPS_ENQ_LAST is in effect.
 	 */
 	if ((prev->scx.flags & SCX_TASK_QUEUED) &&
-	    (!(sch->ops.flags & SCX_OPS_ENQ_LAST) || scx_bypassing(sch, cpu))) {
-		rq->scx.flags |= SCX_RQ_BAL_KEEP;
+	    (!(sch->ops.flags & SCX_OPS_ENQ_LAST) || scx_bypassing(sch, cpu)) &&
+	    scx_task_can_stay_on_cpu(rq, prev)) {
 		__scx_add_event(sch, SCX_EV_DISPATCH_KEEP_LAST, 1);
+		verdict = SCX_DSP_PREV;
 		goto has_tasks;
 	}
-	rq->scx.flags &= ~SCX_RQ_IN_BALANCE;
-	return false;
+	rq->scx.flags &= ~SCX_RQ_IN_DISPATCH;
+	return SCX_DSP_NONE;
 
 has_tasks:
 	/*
@@ -2957,10 +2995,10 @@ has_tasks:
 	 *   between the IMMED queueing and the subsequent scheduling event.
 	 */
 	if (unlikely(rq->scx.local_dsq.nr > 1 && rq->scx.nr_immed))
-		schedule_reenq_local(rq, 0);
+		scx_schedule_reenq_local(rq, 0);
 
-	rq->scx.flags &= ~SCX_RQ_IN_BALANCE;
-	return true;
+	rq->scx.flags &= ~SCX_RQ_IN_DISPATCH;
+	return verdict;
 }
 
 static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
@@ -2973,7 +3011,7 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 		 * dispatched. Call ops_dequeue() to notify the BPF scheduler.
 		 */
 		ops_dequeue(rq, p, SCX_DEQ_CORE_SCHED_EXEC);
-		dispatch_dequeue(rq, p);
+		scx_dispatch_dequeue(rq, p);
 	}
 
 	p->se.exec_start = rq_clock_task(rq);
@@ -2984,6 +3022,9 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 
 	clr_task_runnable(p, true);
 
+	/* apply any pending out-of-band slice request before the tick decision */
+	apply_task_slice_oob(rq, p);
+
 	/*
 	 * @p is getting newly scheduled or got kicked after someone updated its
 	 * slice. Update SCX_RQ_CAN_STOP_TICK to reflect whether the tick can be
@@ -3035,7 +3076,7 @@ preempt_reason_from_class(const struct sched_class *class)
 
 static void switch_class(struct rq *rq, struct task_struct *next)
 {
-	struct scx_sched *sch = scx_root;
+	struct scx_sched *sch = scx_root_protected_live();
 	const struct sched_class *next_class = next->sched_class;
 
 	if (!(sch->ops.flags & SCX_OPS_HAS_CPU_PREEMPT))
@@ -3057,7 +3098,7 @@ static void switch_class(struct rq *rq, struct task_struct *next)
 	 * preempted, and it regaining control of the CPU.
 	 *
 	 * ->cpu_release() complements ->cpu_acquire(), which is emitted the
-	 *  next time that balance_one() is invoked.
+	 *  next time that dispatch_one() is invoked.
 	 */
 	if (!rq->scx.cpu_released) {
 		if (sch->ops.cpu_release) {
@@ -3076,12 +3117,24 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 			      struct task_struct *next)
 {
 	struct scx_sched *sch = scx_task_sched(p);
+	bool rescue_keep = false;
 
 	/* see kick_sync_wait_bal_cb() */
 	smp_store_release(&rq->scx.kick_sync, rq->scx.kick_sync + 1);
 
 	update_curr_scx(rq);
 
+	/*
+	 * If the slice is consumed, protection ends with it. A rescuee
+	 * preempted beforehand keeps going, see scx_rescue_keep().
+	 */
+	if (!p->scx.slice) {
+		if (unlikely(p == scx_rescuee(rq)))
+			rescue_keep = scx_rescue_keep(rq, p);
+		if (!rescue_keep)
+			scx_task_slice_ended(rq, p);
+	}
+
 	/* see dequeue_task_scx() on why we skip when !QUEUED */
 	if (SCX_HAS_OP(sch, stopping) && (p->scx.flags & SCX_TASK_QUEUED))
 		SCX_CALL_OP_TASK(sch, stopping, rq, p, true);
@@ -3095,14 +3148,34 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 		 * forcing a different task. Leave it at the head of the local
 		 * DSQ unless it was an IMMED task. IMMED tasks should not
 		 * linger on a busy CPU, reenqueue them to the BPF scheduler.
+		 *
+		 * An open rescue must keep @p on the local DSQ even if the
+		 * scheduler zeroed the slice in ops.stopping() above.
 		 */
-		if (p->scx.slice && !scx_bypassing(sch, cpu_of(rq))) {
+		if ((p->scx.slice || unlikely(p == scx_rescuee(rq))) &&
+		    !scx_bypassing(sch, cpu_of(rq))) {
 			if (p->scx.flags & SCX_TASK_IMMED) {
 				p->scx.flags |= SCX_TASK_REENQ_PREEMPTED;
-				do_enqueue_task(rq, p, SCX_ENQ_REENQ, -1);
+				scx_do_enqueue_task(rq, p, SCX_ENQ_REENQ, -1);
 				p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
 			} else {
-				dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, SCX_ENQ_HEAD);
+				u64 enq_flags = 0;
+
+				/*
+				 * Keep a preempted rescue going. If preempted
+				 * by another SCX task, append to the local DSQ,
+				 * see scx_rescue_keep().
+				 */
+				if (unlikely(p == scx_rescuee(rq))) {
+					enq_flags |= SCX_ENQ_IGNORE_CAPS;
+					if (!rescue_keep)
+						enq_flags |= SCX_ENQ_HEAD;
+				} else {
+					enq_flags |= SCX_ENQ_HEAD;
+				}
+
+				scx_dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, 0, 0,
+						     enq_flags);
 			}
 			goto switch_class;
 		}
@@ -3112,17 +3185,19 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 		 * sched_class, %SCX_OPS_ENQ_LAST must be set. Tell
 		 * ops.enqueue() that @p is the only one available for this cpu,
 		 * which should trigger an explicit follow-up scheduling event.
+		 * This doesn't apply if the baseline access on the CPU is lost.
 		 *
-		 * Core scheduling can force this CPU idle while @p stays
-		 * runnable. @p's cookie then won't match the core's, so skip
-		 * the warning in that case.
+		 * Under core scheduling, a pick dispatches only when nothing is
+		 * locally runnable and can legitimately go idle with @p still
+		 * runnable (see do_pick_task_scx()).
 		 */
-		if (next && sched_class_above(&ext_sched_class, next->sched_class)) {
-			WARN_ON_ONCE(sched_cpu_cookie_match(rq, p) &&
+		if (next && sched_class_above(&ext_sched_class, next->sched_class) &&
+		    scx_task_can_stay_on_cpu(rq, p)) {
+			WARN_ON_ONCE(!sched_core_enabled(rq) &&
 				     !(sch->ops.flags & SCX_OPS_ENQ_LAST));
-			do_enqueue_task(rq, p, SCX_ENQ_LAST, -1);
+			scx_do_enqueue_task(rq, p, SCX_ENQ_LAST, -1);
 		} else {
-			do_enqueue_task(rq, p, 0, -1);
+			scx_do_enqueue_task(rq, p, 0, -1);
 		}
 	}
 
@@ -3133,12 +3208,27 @@ switch_class:
 
 static void kick_sync_wait_bal_cb(struct rq *rq)
 {
-	struct scx_kick_syncs __rcu *ks = __this_cpu_read(scx_kick_syncs);
-	unsigned long *ksyncs = rcu_dereference_sched(ks)->syncs;
+	struct scx_kick_syncs __rcu *ks;
+	unsigned long *ksyncs;
 	bool waited;
 	s32 cpu;
 
 	/*
+	 * This callback is queued and normally flushed within @rq's own
+	 * scheduling pass. However, dispatch can drop the rq lock while it sits
+	 * queued, and lock takers in that window (the sched class change paths,
+	 * the scx task iterator) flush pending balance callbacks on release,
+	 * running this one on a foreign CPU whose snapshots are unrelated. The
+	 * kicked CPUs are already on their way to advance the kick_syncs being
+	 * waited on. Don't get in the way.
+	 */
+	if (unlikely(cpu_of(rq) != smp_processor_id()))
+		return;
+
+	ks = __this_cpu_read(scx_kick_syncs);
+	ksyncs = rcu_dereference_sched(ks)->syncs;
+
+	/*
 	 * Drop rq lock and enable IRQs while waiting. IRQs must be enabled
 	 * — a target CPU may be waiting for us to process an IPI (e.g. TLB
 	 * flush) while we wait for its kick_sync to advance.
@@ -3160,6 +3250,7 @@ retry:
 			continue;
 		}
 
+		scx_rq_lock_drop(rq);
 		raw_spin_rq_unlock_irq(rq);
 		while (READ_ONCE(cpu_rq(cpu)->scx.kick_sync) == ksyncs[cpu]) {
 			smp_store_release(&rq->scx.kick_sync, rq->scx.kick_sync + 1);
@@ -3179,27 +3270,23 @@ static struct task_struct *first_local_task(struct rq *rq)
 					struct task_struct, scx.dsq_list.node);
 }
 
-static struct task_struct *
-do_pick_task_scx(struct rq *rq, struct rq_flags *rf, bool force_scx)
+/*
+ * Run dispatch and queue the follow-up work for a pick.
+ */
+static enum scx_dsp_verdict dispatch_pick(struct rq *rq, struct rq_flags *rf,
+					  struct task_struct *prev)
 {
-	struct task_struct *prev = rq->curr;
-	bool keep_prev;
-	struct task_struct *p;
-
-	/* see kick_sync_wait_bal_cb() */
-	smp_store_release(&rq->scx.kick_sync, rq->scx.kick_sync + 1);
-
-	rq_modified_begin(rq, &ext_sched_class);
+	enum scx_dsp_verdict verdict;
 
 	rq_unpin_lock(rq, rf);
-	balance_one(rq, prev);
+	verdict = dispatch_one(rq, prev);
 	rq_repin_lock(rq, rf);
 	maybe_queue_balance_callback(rq);
 
 	/*
-	 * Defer to a balance callback which can drop rq lock and enable
-	 * IRQs. Waiting directly in the pick path would deadlock against
-	 * CPUs sending us IPIs (e.g. TLB flushes) while we wait for them.
+	 * Defer to a balance callback which can drop rq lock and enable IRQs.
+	 * Waiting directly in the pick path would deadlock against CPUs sending
+	 * us IPIs (e.g. TLB flushes) while we wait for them.
 	 */
 	if (unlikely(rq->scx.kick_sync_pending)) {
 		rq->scx.kick_sync_pending = false;
@@ -3207,10 +3294,90 @@ do_pick_task_scx(struct rq *rq, struct rq_flags *rf, bool force_scx)
 				       kick_sync_wait_bal_cb);
 	}
 
+	return verdict;
+}
+
+#ifdef CONFIG_SCHED_CORE
+/*
+ * Dispatch for a pick when core scheduling is enabled. The selection picks for
+ * all SMT siblings and the rq_i->core_pick state it builds must stay atomic
+ * throughout. If the dispatch released the rq lock, anything can have happened
+ * in between - return %SCX_DSP_RETRY to restart the selection against current
+ * state.
+ */
+static enum scx_dsp_verdict dispatch_core_pick(struct rq *rq, struct rq_flags *rf,
+					       struct task_struct *prev)
+{
+	enum scx_dsp_verdict verdict;
+	u32 seq = rq->scx.lock_drop_seq;
+
+	/* another dispatch is in flight on @rq, let that handle it */
+	if (rq->scx.flags & SCX_RQ_IN_DISPATCH)
+		return SCX_DSP_NONE;
+
+	rq_unpin_lock(rq, rf);
+
+	verdict = dispatch_one(rq, prev);
+
+	if (cpu_of(rq) == smp_processor_id()) {
+		maybe_queue_balance_callback(rq);
+
+		/* see dispatch_pick() */
+		if (unlikely(rq->scx.kick_sync_pending)) {
+			rq->scx.kick_sync_pending = false;
+			queue_balance_callback(rq, &rq->scx.kick_sync_bal_cb,
+					       kick_sync_wait_bal_cb);
+		}
+	} else if (unlikely(rq->scx.flags & SCX_RQ_BAL_CB_PENDING)) {
+		/*
+		 * Balance callbacks must run in the context that queued them,
+		 * so they can't be queued on another CPU's rq. Run the deferred
+		 * work directly instead.
+		 */
+		rq->scx.flags &= ~SCX_RQ_BAL_CB_PENDING;
+		run_deferred(rq);
+	}
+
+	rq_repin_lock(rq, rf);
+
+	/* if dispatch_one() released the rq lock, restart the selection */
+	if (rq->scx.lock_drop_seq != seq)
+		return SCX_DSP_RETRY;
+
+	return verdict;
+}
+#else	/* CONFIG_SCHED_CORE */
+static enum scx_dsp_verdict dispatch_core_pick(struct rq *rq, struct rq_flags *rf,
+					       struct task_struct *prev)
+{
+	return SCX_DSP_NONE;
+}
+#endif	/* CONFIG_SCHED_CORE */
+
+static struct task_struct *
+do_pick_task_scx(struct rq *rq, struct rq_flags *rf, bool force_scx)
+{
+	struct task_struct *prev = rq->curr;
+	enum scx_dsp_verdict verdict;
+	struct task_struct *p;
+
+	/* see kick_sync_wait_bal_cb() */
+	smp_store_release(&rq->scx.kick_sync, rq->scx.kick_sync + 1);
+
+	rq_modified_begin(rq, &ext_sched_class);
+
+	if (sched_core_enabled(rq))
+		verdict = dispatch_core_pick(rq, rf, prev);
+	else
+		verdict = dispatch_pick(rq, rf, prev);
+
+	if (verdict == SCX_DSP_RETRY)
+		return RETRY_TASK;
+
 	/*
-	 * If any higher-priority sched class enqueued a runnable task on
-	 * this rq during balance_one(), abort and return RETRY_TASK, so
-	 * that the scheduler loop can restart.
+	 * If any higher-priority sched class enqueued a runnable task on this
+	 * rq during dispatch_one(), abort and return RETRY_TASK, so that the
+	 * scheduler loop can restart.
 	 *
 	 * If @force_scx is true, always try to pick a SCHED_EXT task,
 	 * regardless of any higher-priority sched classes activity.
@@ -3218,28 +3385,23 @@ do_pick_task_scx(struct rq *rq, struct rq_flags *rf, bool force_scx)
 	if (!force_scx && rq_modified_above(rq, &ext_sched_class))
 		return RETRY_TASK;
 
-	keep_prev = rq->scx.flags & SCX_RQ_BAL_KEEP;
-	if (unlikely(keep_prev &&
-		     prev->sched_class != &ext_sched_class)) {
-		WARN_ON_ONCE(scx_enable_state() == SCX_ENABLED);
-		keep_prev = false;
-	}
-
 	/*
-	 * If balance_one() is telling us to keep running @prev, replenish slice
-	 * if necessary and keep running @prev. Otherwise, pop the first one
-	 * from the local DSQ.
+	 * If we're keeping @prev, replenish slice if necessary and keep running
+	 * @prev. Otherwise, pop the first one from the local DSQ.
 	 */
-	if (keep_prev) {
+	if (verdict == SCX_DSP_PREV) {
 		p = prev;
-		if (!p->scx.slice)
+		if (!p->scx.slice) {
+			/* the slice is consumed, protection ends */
+			scx_task_slice_ended(rq, p);
 			refill_task_slice_dfl(scx_task_sched(p), p);
+		}
 	} else {
 		p = first_local_task(rq);
 		if (!p)
 			return NULL;
 
-		if (unlikely(!p->scx.slice)) {
+		if (unlikely(!p->scx.slice) && scx_task_can_stay_on_cpu(rq, p)) {
 			struct scx_sched *sch = scx_task_sched(p);
 
 			if (!scx_bypassing(sch, cpu_of(rq)) &&
@@ -3298,33 +3460,70 @@ void ext_server_init(struct rq *rq)
  * usual sched_class'es and needs to find out the expected task ordering. For
  * SCX, core-sched calls this function to interrogate the task ordering.
  *
- * Unless overridden by ops.core_sched_before(), @p->scx.core_sched_at is used
- * to implement the default task ordering. The older the timestamp, the higher
- * priority the task - the global FIFO ordering matching the default scheduling
- * behavior.
+ * A pair of tasks owned by one scheduler is ordered by the owner's
+ * ops.core_sched_before(). A pair spanning two schedulers is ordered by their
+ * nearest common ancestor which implements the op - the one case where the op
+ * is called on tasks that the scheduler delegated to its sub-schedulers and may
+ * not be scheduling anymore.
+ *
+ * When neither applies, or the deciding scheduler is bypassing on either task's
+ * CPU, the default ordering runs the task which has been waiting longer first.
+ * A running task counts as the most recently serviced and orders after every
+ * waiting task. Waiting tasks are compared by @p->scx.runnable_at.
  *
- * When ops.core_sched_before() is enabled, @p->scx.core_sched_at is used to
- * implement FIFO ordering within each local DSQ. See pick_task_scx().
+ * Return: %true if @a should run after @b.
  */
 bool scx_prio_less(const struct task_struct *a, const struct task_struct *b,
 		   bool in_fi)
 {
 	struct scx_sched *sch_a = scx_task_sched(a);
 	struct scx_sched *sch_b = scx_task_sched(b);
+	struct scx_sched *sch = NULL;
+	bool a_running, b_running;
+
+	if (sch_a == sch_b) {
+		if (SCX_HAS_OP(sch_a, core_sched_before))
+			sch = sch_a;
+	} else {
+		s32 level;
+
+		for (level = min(sch_a->level, sch_b->level); level >= 0; level--) {
+			struct scx_sched *anc = sch_a->ancestors[level];
+
+			if (anc == sch_b->ancestors[level] &&
+			    SCX_HAS_OP(anc, core_sched_before)) {
+				sch = anc;
+				break;
+			}
+		}
+	}
 
 	/*
+	 * scx_prio_less() returns whether @a should run after @b while
+	 * ops.core_sched_before() returns whether its first argument should run
+	 * before the second. Swap the arguments.
+	 *
 	 * The const qualifiers are dropped from task_struct pointers when
 	 * calling ops.core_sched_before(). Accesses are controlled by the
 	 * verifier.
 	 */
-	if (sch_a == sch_b && SCX_HAS_OP(sch_a, core_sched_before) &&
-	    !scx_bypassing(sch_a, task_cpu(a)))
-		return SCX_CALL_OP_2TASKS_RET(sch_a, core_sched_before,
-					      task_rq(a),
-					      (struct task_struct *)a,
-					      (struct task_struct *)b);
-	else
-		return time_after64(a->scx.core_sched_at, b->scx.core_sched_at);
+	if (sch && !scx_bypassing(sch, task_cpu(a)) && !scx_bypassing(sch, task_cpu(b)))
+		return SCX_CALL_OP_2TASKS_RET(sch, core_sched_before, task_rq(a),
+					      (struct task_struct *)b,
+					      (struct task_struct *)a);
+
+	/*
+	 * runnable_at is refreshed only on enqueue, so a task which keeps
+	 * occupying its CPU carries a stale stamp. A running task is the most
+	 * recently serviced whatever its stamp says. Order it after every
+	 * waiting task.
+	 */
+	a_running = a->on_cpu;
+	b_running = b->on_cpu;
+	if (a_running != b_running)
+		return a_running;
+
+	return time_after(a->scx.runnable_at, b->scx.runnable_at);
 }
 #endif	/* CONFIG_SCHED_CORE */
 
@@ -3369,17 +3568,34 @@ static int select_task_rq_scx(struct task_struct *p, int prev_cpu, int wake_flag
 	} else {
 		s32 cpu;
 
+		/*
+		 * While bypassing, the enqueue path routes @p to a bypass DSQ
+		 * without consulting the direct-dispatch target, making the
+		 * default selection pointless. It doesn't work anyway when the
+		 * scheduler does its own idle tracking and the built-in idle
+		 * cpumasks are not updated. Leave @p on @prev_cpu.
+		 */
+		if (bypassing) {
+			__scx_add_event(sch, SCX_EV_BYPASS_DISPATCH, 1);
+			p->scx.selected_cpu = prev_cpu;
+			return prev_cpu;
+		}
+
 		cpu = scx_select_cpu_dfl(p, prev_cpu, wake_flags, NULL, 0);
 		if (cpu >= 0) {
-			refill_task_slice_dfl(sch, p);
+			/*
+			 * Carry the slice refill and let the insertion commit
+			 * it under rq lock. See the write rules.
+			 */
+			__scx_add_event(sch, SCX_EV_REFILL_SLICE_DFL, 1);
+			p->scx.ddsp_slice = READ_ONCE(sch->slice_dfl);
+			p->scx.ddsp_enq_flags = SCX_ENQ_SLICE_DFL;
 			p->scx.ddsp_dsq_id = SCX_DSQ_LOCAL;
 		} else {
 			cpu = prev_cpu;
 		}
 		p->scx.selected_cpu = cpu;
 
-		if (bypassing)
-			__scx_add_event(sch, SCX_EV_BYPASS_DISPATCH, 1);
 		return cpu;
 	}
 }
@@ -3413,8 +3629,9 @@ static void set_cpus_allowed_scx(struct task_struct *p,
 
 static void handle_hotplug(struct rq *rq, bool online)
 {
-	struct scx_sched *sch = scx_root;
+	struct scx_sched *sch = scx_root_protected();
 	s32 cpu = cpu_of(rq);
+	s32 cpu_or_cid = cpu;
 
 	atomic_long_inc(&scx_hotplug_seq);
 
@@ -3429,10 +3646,31 @@ static void handle_hotplug(struct rq *rq, bool online)
 	if (scx_enabled())
 		scx_idle_update_selcpu_topology(&sch->ops);
 
+	if (online)
+		scx_online_ecaps(rq);
+	else
+		scx_offline_ecaps(rq);
+
+	/*
+	 * The tables can't be retired while this function is running as the
+	 * retirement is inside cpus_read_lock. However, scx_cpu_arg() is
+	 * awkward here as the tables can be NULL after root enable failure and
+	 * lockdep would trigger without surrounding rcu_read_lock(). Open code
+	 * the translation. If the table is NULL, the ops are also cleared and
+	 * @cpu_or_cid goes unused.
+	 */
+	if (scx_is_cid_type()) {
+		s16 *tbl = rcu_dereference_check(scx_cpu_to_cid_tbl,
+						 lockdep_is_cpus_held());
+
+		if (tbl)
+			cpu_or_cid = tbl[cpu];
+	}
+
 	if (online && SCX_HAS_OP(sch, cpu_online))
-		SCX_CALL_OP(sch, cpu_online, NULL, scx_cpu_arg(cpu));
+		SCX_CALL_OP(sch, cpu_online, NULL, cpu_or_cid);
 	else if (!online && SCX_HAS_OP(sch, cpu_offline))
-		SCX_CALL_OP(sch, cpu_offline, NULL, scx_cpu_arg(cpu));
+		SCX_CALL_OP(sch, cpu_offline, NULL, cpu_or_cid);
 	else
 		scx_exit(sch, SCX_EXIT_UNREG_KERN,
 			 SCX_ECODE_ACT_RESTART | SCX_ECODE_RSN_HOTPLUG,
@@ -3458,6 +3696,7 @@ static void rq_online_scx(struct rq *rq)
 static void rq_offline_scx(struct rq *rq)
 {
 	rq->scx.flags &= ~SCX_RQ_ONLINE;
+	scx_rescue_flush(rq);
 }
 
 static bool check_rq_for_timeouts(struct rq *rq)
@@ -3478,8 +3717,19 @@ static bool check_rq_for_timeouts(struct rq *rq)
 
 		if (unlikely(time_after(jiffies,
 					last_runnable + READ_ONCE(sch->watchdog_timeout)))) {
+			struct scx_dispatch_q *dsq = READ_ONCE(p->scx.dsq);
 			u32 dur_ms = jiffies_to_msecs(jiffies - last_runnable);
 
+			/*
+			 * A task can be stuck on a DSQ that a sched other than
+			 * its owner is responsible for draining, e.g. an
+			 * ancestor's bypass DSQ while the owner is bypassing.
+			 * Blame the drainer. The local DSQ is consumed by the
+			 * cpu itself and keeps blame on the owner.
+			 */
+			if (dsq && dsq->sched && dsq->id != SCX_DSQ_LOCAL)
+				sch = dsq->sched;
+
 			__scx_exit(sch, SCX_EXIT_ERROR_STALL, 0, cpu_of(rq),
 				   "%s[%d] failed to run for %u.%03us",
 				   p->comm, p->pid, dur_ms / 1000,
@@ -3544,15 +3794,13 @@ static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
 	update_curr_scx(rq);
 
 	/*
-	 * While disabling, always resched and refresh core-sched timestamp as
-	 * we can't trust the slice management or ops.core_sched_before().
+	 * While disabling, always resched as we can't trust the slice
+	 * management.
 	 */
-	if (scx_bypassing(sch, cpu_of(rq))) {
-		curr->scx.slice = 0;
-		touch_core_sched(rq, curr);
-	} else if (SCX_HAS_OP(sch, tick)) {
+	if (scx_bypassing(sch, cpu_of(rq)))
+		scx_set_task_slice(curr, 0);
+	else if (SCX_HAS_OP(sch, tick))
 		SCX_CALL_OP_TASK(sch, tick, rq, curr);
-	}
 
 	if (!curr->scx.slice)
 		resched_curr(rq);
@@ -3572,15 +3820,28 @@ static struct cgroup *tg_cgrp(struct task_group *tg)
 		return &cgrp_dfl_root.cgrp;
 }
 
-#define SCX_INIT_TASK_ARGS_CGROUP(tg)		.cgroup = tg_cgrp(tg),
+#define SCX_INIT_TASK_ARGS_CGROUP(cgrp)		.cgroup = (cgrp),
 
 #else	/* CONFIG_EXT_GROUP_SCHED */
 
-#define SCX_INIT_TASK_ARGS_CGROUP(tg)
+#define SCX_INIT_TASK_ARGS_CGROUP(cgrp)
 
 #endif	/* CONFIG_EXT_GROUP_SCHED */
 
-static int __scx_init_task(struct scx_sched *sch, struct task_struct *p, bool fork)
+/**
+ * __scx_init_task - Initialize a task for a sched
+ * @sch: sched to initialize @p for
+ * @p: task of interest
+ * @cgrp: cgroup @p is joining, %NULL for @p's current task_group's cgroup
+ * @fork: %true if @p is being forked
+ *
+ * Pre-commit cgroup migration passes @cgrp explicitly as @p's task_group
+ * still reflects the source.
+ *
+ * Return 0 on success, -errno on failure.
+ */
+int __scx_init_task(struct scx_sched *sch, struct task_struct *p,
+		    struct cgroup *cgrp, bool fork)
 {
 	int ret;
 
@@ -3588,13 +3849,13 @@ static int __scx_init_task(struct scx_sched *sch, struct task_struct *p, bool fo
 
 	if (SCX_HAS_OP(sch, init_task)) {
 		struct scx_init_task_args args = {
-			SCX_INIT_TASK_ARGS_CGROUP(task_group(p))
+			SCX_INIT_TASK_ARGS_CGROUP(cgrp ?: tg_cgrp(task_group(p)))
 			.fork = fork,
 		};
 
 		ret = SCX_CALL_OP_RET(sch, init_task, NULL, p, &args);
 		if (unlikely(ret)) {
-			ret = ops_sanitize_err(sch, "init_task", ret);
+			ret = scx_ops_sanitize_err(sch, "init_task", ret);
 			return ret;
 		}
 	}
@@ -3666,7 +3927,7 @@ static void __scx_enable_task(struct scx_sched *sch, struct task_struct *p)
 		SCX_CALL_OP_TASK(sch, set_weight, rq, p, p->scx.weight);
 }
 
-static void scx_enable_task(struct scx_sched *sch, struct task_struct *p)
+void scx_enable_task(struct scx_sched *sch, struct task_struct *p)
 {
 	__scx_enable_task(sch, p);
 	scx_set_task_state(p, SCX_TASK_ENABLED);
@@ -3690,7 +3951,9 @@ static void scx_disable_task(struct scx_sched *sch, struct task_struct *p)
 	 * control, after ops.disable() has observed their final values.
 	 */
 	p->scx.dsq_vtime = 0;
-	p->scx.slice = 0;
+	scx_task_slice_ended(rq, p);
+	scx_set_task_slice(p, 0);
+	p->scx.reenq_cnt = 0;
 
 	/*
 	 * Verify the task is not in BPF scheduler's custody. If flag
@@ -3700,8 +3963,7 @@ static void scx_disable_task(struct scx_sched *sch, struct task_struct *p)
 	WARN_ON_ONCE(p->scx.flags & SCX_TASK_IN_CUSTODY);
 }
 
-static void __scx_disable_and_exit_task(struct scx_sched *sch,
-					struct task_struct *p)
+void __scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p)
 {
 	struct scx_exit_task_args args = {
 		.cancelled = false,
@@ -3735,19 +3997,19 @@ static void __scx_disable_and_exit_task(struct scx_sched *sch,
  * ran. The task state has not been transitioned, so this mirrors the
  * SCX_TASK_INIT branch in __scx_disable_and_exit_task().
  */
-static void scx_sub_init_cancel_task(struct scx_sched *sch, struct task_struct *p)
+void scx_sub_init_cancel_task(struct scx_sched *sch, struct task_struct *p)
 {
 	struct scx_exit_task_args args = { .cancelled = true };
 
 	lockdep_assert_held(&p->pi_lock);
 	lockdep_assert_rq_held(task_rq(p));
 
+	/* @p was never associated with @sch, dispatch on the explicit @sch */
 	if (SCX_HAS_OP(sch, exit_task))
-		SCX_CALL_OP_TASK(sch, exit_task, task_rq(p), p, &args);
+		__SCX_CALL_OP_TASK(sch, ops, exit_task, task_rq(p), p, &args);
 }
 
-static void scx_disable_and_exit_task(struct scx_sched *sch,
-				      struct task_struct *p)
+void scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p)
 {
 	__scx_disable_and_exit_task(sch, p);
 
@@ -3775,6 +4037,7 @@ void init_scx_entity(struct sched_ext_entity *scx)
 	RB_CLEAR_NODE(&scx->dsq_priq);
 	scx->sticky_cpu = -1;
 	scx->holding_cpu = -1;
+	scx->runnable_cpu = -1;
 	INIT_LIST_HEAD(&scx->runnable_node);
 	scx->runnable_at = jiffies;
 	scx->ddsp_dsq_id = SCX_DSQ_INVALID;
@@ -3829,12 +4092,12 @@ int scx_fork(struct task_struct *p, struct kernel_clone_args *kargs)
 
 	if (scx_init_task_enabled) {
 #ifdef CONFIG_EXT_SUB_SCHED
-		struct scx_sched *sch = kargs->cset->dfl_cgrp->scx_sched;
+		struct scx_sched *sch = scx_cgroup_sched(kargs->cset->dfl_cgrp);
 #else
-		struct scx_sched *sch = scx_root;
+		struct scx_sched *sch = scx_root_protected_live();
 #endif
 		scx_set_task_state(p, SCX_TASK_INIT_BEGIN);
-		ret = __scx_init_task(sch, p, true);
+		ret = __scx_init_task(sch, p, NULL, true);
 		if (unlikely(ret)) {
 			scx_set_task_state(p, SCX_TASK_NONE);
 			return ret;
@@ -3877,7 +4140,7 @@ void scx_post_fork(struct task_struct *p)
 
 void scx_cancel_fork(struct task_struct *p)
 {
-	if (scx_enabled()) {
+	if (scx_init_task_enabled) {
 		struct rq *rq;
 		struct rq_flags rf;
 
@@ -4058,13 +4321,15 @@ static void process_ddsp_deferred_locals(struct rq *rq)
 		struct scx_dispatch_q *dsq;
 		u64 dsq_id = p->scx.ddsp_dsq_id;
 		u64 enq_flags = p->scx.ddsp_enq_flags;
+		u64 slice = p->scx.ddsp_slice;
+		u64 vtime = p->scx.ddsp_vtime;
 
 		list_del_init(&p->scx.dsq_list.node);
 		clear_direct_dispatch(p);
 
 		dsq = find_dsq_for_dispatch(sch, rq, dsq_id, task_cpu(p));
 		if (!WARN_ON_ONCE(dsq->id != SCX_DSQ_LOCAL))
-			dispatch_to_local_dsq(sch, rq, dsq, p, enq_flags);
+			dispatch_to_local_dsq(sch, rq, dsq, p, slice, vtime, enq_flags);
 	}
 }
 
@@ -4088,16 +4353,20 @@ static void process_ddsp_deferred_locals(struct rq *rq)
  * Reenqueued tasks go through ops.enqueue() with %SCX_ENQ_REENQ |
  * %SCX_TASK_REENQ_IMMED. If the BPF scheduler dispatches back to the same local
  * DSQ with %SCX_ENQ_IMMED while the CPU is still unavailable, this triggers
- * another reenq cycle. Repetitions are bounded by %SCX_REENQ_LOCAL_MAX_REPEAT
- * in process_deferred_reenq_locals().
+ * another reenq cycle. Repetitions are bounded by %SCX_REENQ_MAX_REPEAT in
+ * scx_do_enqueue_task(), which ejects the task's owning scheduler.
  */
-static bool local_task_should_reenq(struct task_struct *p, u64 *reenq_flags, u32 *reason)
+static bool local_task_should_reenq(struct rq *rq, struct task_struct *p,
+				    u64 *reenq_flags, u32 *reason)
 {
 	bool first;
 
 	first = !(*reenq_flags & SCX_REENQ_TSR_NOT_FIRST);
 	*reenq_flags |= SCX_REENQ_TSR_NOT_FIRST;
 
+	if (unlikely((p->scx.flags & SCX_TASK_PROTECTED) || p == scx_rescuee(rq)))
+		return false;
+
 	*reason = SCX_TASK_REENQ_KFUNC;
 
 	if ((p->scx.flags & SCX_TASK_IMMED) &&
@@ -4107,6 +4376,12 @@ static bool local_task_should_reenq(struct task_struct *p, u64 *reenq_flags, u32
 		return true;
 	}
 
+	if ((*reenq_flags & SCX_REENQ_CAP_REVOKE) &&
+	    scx_task_reenq_on_cap_revoke(rq, p)) {
+		*reason = SCX_TASK_REENQ_CAP;
+		return true;
+	}
+
 	return *reenq_flags & SCX_REENQ_ANY;
 }
 
@@ -4150,10 +4425,10 @@ static u32 reenq_local(struct scx_sched *sch, struct rq *rq, u64 reenq_flags)
 		if (!scx_is_descendant(task_sch, sch))
 			continue;
 
-		if (!local_task_should_reenq(p, &reenq_flags, &reason))
+		if (!local_task_should_reenq(rq, p, &reenq_flags, &reason))
 			continue;
 
-		dispatch_dequeue(rq, p);
+		scx_dispatch_dequeue(rq, p);
 
 		if (WARN_ON_ONCE(p->scx.flags & SCX_TASK_REENQ_REASON_MASK))
 			p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
@@ -4165,25 +4440,41 @@ static u32 reenq_local(struct scx_sched *sch, struct rq *rq, u64 reenq_flags)
 	list_for_each_entry_safe(p, n, &tasks, scx.dsq_list.node) {
 		list_del_init(&p->scx.dsq_list.node);
 
-		do_enqueue_task(rq, p, SCX_ENQ_REENQ, -1);
+		scx_do_enqueue_task(rq, p, SCX_ENQ_REENQ, -1);
 
 		p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
 		nr_enqueued++;
 	}
 
+	/*
+	 * The revoke that scheduled this scan may have raced the pick: curr
+	 * may be a now-capless task, either one that kept running or one
+	 * promoted off the local DSQ between the ecaps sync and this scan.
+	 * Zero the slice to evict it. The enqueue gate blocks new capless
+	 * inserts, so no later pick can slip through after the scan.
+	 */
+	if ((reenq_flags & SCX_REENQ_CAP_REVOKE) &&
+	    rq->curr->sched_class == &ext_sched_class &&
+	    scx_task_reenq_on_cap_revoke(rq, rq->curr)) {
+		scx_set_task_slice(rq->curr, 0);
+		resched_curr(rq);
+	}
+
 	return nr_enqueued;
 }
 
 static void process_deferred_reenq_locals(struct rq *rq)
 {
-	u64 seq = ++rq->scx.deferred_reenq_locals_seq;
-
 	lockdep_assert_rq_held(rq);
 
+	/*
+	 * A task can be re-queued within this loop when a reenqueued task
+	 * bounces straight back to the local DSQ. That recursion is bounded by
+	 * the per-task reenqueue cap in scx_do_enqueue_task().
+	 */
 	while (true) {
 		struct scx_sched *sch;
 		u64 reenq_flags;
-		bool skip = false;
 
 		scoped_guard (raw_spinlock, &rq->scx.deferred_reenq_lock) {
 			struct scx_deferred_reenq_local *drl =
@@ -4202,27 +4493,12 @@ static void process_deferred_reenq_locals(struct rq *rq)
 			reenq_flags = drl->flags;
 			WRITE_ONCE(drl->flags, 0);
 			list_del_init(&drl->node);
-
-			if (likely(drl->seq != seq)) {
-				drl->seq = seq;
-				drl->cnt = 0;
-			} else {
-				if (unlikely(++drl->cnt > SCX_REENQ_LOCAL_MAX_REPEAT)) {
-					scx_error(sch, "SCX_ENQ_REENQ on SCX_DSQ_LOCAL repeated %u times",
-						  drl->cnt);
-					skip = true;
-				}
-
-				__scx_add_event(sch, SCX_EV_REENQ_LOCAL_REPEAT, 1);
-			}
 		}
 
-		if (!skip) {
-			/* see schedule_dsq_reenq() */
-			smp_mb();
+		/* see schedule_dsq_reenq() */
+		smp_mb();
 
-			reenq_local(sch, rq, reenq_flags);
-		}
+		reenq_local(sch, rq, reenq_flags);
 	}
 }
 
@@ -4258,8 +4534,10 @@ static void reenq_user(struct rq *rq, struct scx_dispatch_q *dsq, u64 reenq_flag
 		task_rq = task_rq(p);
 
 		if (locked_rq != task_rq) {
-			if (locked_rq)
+			if (locked_rq) {
+				scx_rq_lock_drop(locked_rq);
 				raw_spin_rq_unlock(locked_rq);
+			}
 			if (unlikely(!raw_spin_rq_trylock(task_rq))) {
 				raw_spin_unlock(&dsq->lock);
 				raw_spin_rq_lock(task_rq);
@@ -4280,11 +4558,12 @@ static void reenq_user(struct rq *rq, struct scx_dispatch_q *dsq, u64 reenq_flag
 			p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
 		p->scx.flags |= reason;
 
-		do_enqueue_task(task_rq, p, SCX_ENQ_REENQ, -1);
+		scx_do_enqueue_task(task_rq, p, SCX_ENQ_REENQ, -1);
 
 		p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
 
 		if (!(++nr_enqueued % SCX_TASK_ITER_BATCH)) {
+			scx_rq_lock_drop(locked_rq);
 			raw_spin_rq_unlock(locked_rq);
 			locked_rq = NULL;
 			cpu_relax();
@@ -4297,8 +4576,10 @@ static void reenq_user(struct rq *rq, struct scx_dispatch_q *dsq, u64 reenq_flag
 	raw_spin_unlock(&dsq->lock);
 
 	if (locked_rq != rq) {
-		if (locked_rq)
+		if (locked_rq) {
+			scx_rq_lock_drop(locked_rq);
 			raw_spin_rq_unlock(locked_rq);
+		}
 		raw_spin_rq_lock(rq);
 	}
 }
@@ -4309,7 +4590,7 @@ static void process_deferred_reenq_users(struct rq *rq)
 
 	while (true) {
 		struct scx_dispatch_q *dsq;
-		u64 reenq_flags;
+		u64 dsq_id, reenq_flags;
 
 		scoped_guard (raw_spinlock, &rq->scx.deferred_reenq_lock) {
 			struct scx_deferred_reenq_user *dru =
@@ -4332,7 +4613,12 @@ static void process_deferred_reenq_users(struct rq *rq)
 		/* see schedule_dsq_reenq() */
 		smp_mb();
 
-		BUG_ON(dsq->id & SCX_DSQ_FLAG_BUILTIN);
+		/* destroy_dsq() may have raced and invalidated @dsq, nothing to reenq */
+		dsq_id = READ_ONCE(dsq->id);
+		if (unlikely(dsq_id == SCX_DSQ_INVALID))
+			continue;
+
+		BUG_ON(dsq_id & SCX_DSQ_FLAG_BUILTIN);
 		reenq_user(rq, dsq, reenq_flags);
 	}
 }
@@ -4346,6 +4632,8 @@ static void run_deferred(struct rq *rq)
 
 	if (!list_empty(&rq->scx.deferred_reenq_users))
 		process_deferred_reenq_users(rq);
+
+	scx_reenq_reject(rq);
 }
 
 #ifdef CONFIG_NO_HZ_FULL
@@ -4370,6 +4658,13 @@ bool scx_can_stop_tick(struct rq *rq)
 		return false;
 
 	/*
+	 * A running rescuee's charging and expiry are tick-driven, see
+	 * scx_rescue_charge(). Keep the tick while rescue is in progress.
+	 */
+	if (unlikely(p == scx_rescuee(rq)))
+		return false;
+
+	/*
 	 * @rq can dispatch from different DSQs, so we can't tell whether it
 	 * needs the tick or not by looking at nr_running. Allow stopping ticks
 	 * iff the BPF scheduler indicated so. See set_next_task_scx().
@@ -4381,7 +4676,6 @@ bool scx_can_stop_tick(struct rq *rq)
 #ifdef CONFIG_EXT_GROUP_SCHED
 
 DEFINE_STATIC_PERCPU_RWSEM(scx_cgroup_ops_rwsem);
-static bool scx_cgroup_enabled;
 
 void scx_tg_init(struct task_group *tg)
 {
@@ -4391,14 +4685,80 @@ void scx_tg_init(struct task_group *tg)
 	tg->scx.idle = false;
 }
 
+/**
+ * scx_tg_sched - Resolve a task_group's sched
+ * @tg: task_group of interest
+ *
+ * Return the sched that @tg's ops.cgroup_init() succeeded on, %NULL if @tg
+ * isn't inited. An autogroup tg has no cgroup of its own and resolves to the
+ * root sched.
+ *
+ * When a child sched exits, its task_groups are moved to the parent and
+ * re-inited on it. A failed re-init fails the parent in turn and leaves the
+ * task_group without a sched it's inited on, resolving to %NULL. See
+ * scx_cgroup_return_subtree().
+ *
+ * Safe for callers read-locking the ops rwsem. tg->scx.sched rewrites
+ * write-lock it, and tg on/offline can't overlap such callers as a css's files
+ * are created after online and drained before offline.
+ */
+static struct scx_sched *scx_tg_sched(struct task_group *tg)
+{
+	lockdep_assert(lockdep_is_held(&cgroup_mutex) ||
+		       lockdep_is_held(&scx_cgroup_ops_rwsem));
+
+	if (!tg->css.cgroup)
+		tg = &root_task_group;
+	/* INITED means ops.cgroup_init() succeeded on @tg->scx.sched */
+	return (tg->scx.flags & SCX_TG_INITED) ? tg->scx.sched : NULL;
+}
+
+/**
+ * scx_tg_knob_sched - Resolve the sched receiving a task_group's knob updates
+ * @tg: task_group of interest
+ *
+ * Knobs of a cgroup belong to the parent. Deliver the set_* ops to the
+ * parent task_group's sched, which equals @tg's own sched everywhere except
+ * at a sub-scheduler attach point, where the sub's parent sched receives
+ * them.
+ *
+ * Return %NULL if the parent task_group has no sched. That can happen when the
+ * parent's ops.cgroup_init() fails while a sub-scheduler is being disabled.
+ *
+ * The callers sit in @tg's cgroup file writes holding the ops rwsem read
+ * side. That extends scx_tg_sched()'s file-write argument to the parent's
+ * sched read: a parent css outlives its children's files.
+ */
+static struct scx_sched *scx_tg_knob_sched(struct task_group *tg)
+{
+	lockdep_assert(lockdep_is_held(&cgroup_mutex) ||
+		       lockdep_is_held(&scx_cgroup_ops_rwsem));
+
+	if (!tg->css.cgroup || !tg->css.parent)
+		return scx_tg_sched(&root_task_group);
+	return scx_tg_sched(css_tg(tg->css.parent));
+}
+
 int scx_tg_online(struct task_group *tg)
 {
-	struct scx_sched *sch = scx_root;
 	int ret = 0;
 
 	WARN_ON_ONCE(tg->scx.flags & (SCX_TG_ONLINE | SCX_TG_INITED));
 
 	if (scx_cgroup_enabled) {
+		struct scx_sched *sch;
+
+		/*
+		 * The cgroup lifetime notifier populates cgrp->scx_sched before
+		 * css_online, but only on the default hierarchy. Sub-scheds are
+		 * attached to the cgroup2 hierarchy, so a cgroup1 task_group
+		 * always belongs to the root sched.
+		 */
+		if (cgroup_on_dfl(tg->css.cgroup))
+			sch = scx_cgroup_sched(tg->css.cgroup);
+		else
+			sch = scx_tg_sched(&root_task_group);
+
 		if (SCX_HAS_OP(sch, cgroup_init)) {
 			struct scx_cgroup_init_args args =
 				{ .weight = tg->scx.weight,
@@ -4409,10 +4769,12 @@ int scx_tg_online(struct task_group *tg)
 			ret = SCX_CALL_OP_RET(sch, cgroup_init,
 					      NULL, tg->css.cgroup, &args);
 			if (ret)
-				ret = ops_sanitize_err(sch, "cgroup_init", ret);
+				ret = scx_ops_sanitize_err(sch, "cgroup_init", ret);
 		}
-		if (ret == 0)
+		if (ret == 0) {
+			tg->scx.sched = sch;
 			tg->scx.flags |= SCX_TG_ONLINE | SCX_TG_INITED;
+		}
 	} else {
 		tg->scx.flags |= SCX_TG_ONLINE;
 	}
@@ -4422,19 +4784,30 @@ int scx_tg_online(struct task_group *tg)
 
 void scx_tg_offline(struct task_group *tg)
 {
-	struct scx_sched *sch = scx_root;
+	struct scx_sched *sch = tg->scx.sched;
 
 	WARN_ON_ONCE(!(tg->scx.flags & SCX_TG_ONLINE));
 
-	if (scx_cgroup_enabled && SCX_HAS_OP(sch, cgroup_exit) &&
-	    (tg->scx.flags & SCX_TG_INITED))
+	/* INITED implies non-NULL @sch, test before SCX_HAS_OP() derefs */
+	if (scx_cgroup_enabled && (tg->scx.flags & SCX_TG_INITED) &&
+	    SCX_HAS_OP(sch, cgroup_exit))
 		SCX_CALL_OP(sch, cgroup_exit, NULL, tg->css.cgroup);
+	tg->scx.sched = NULL;
 	tg->scx.flags &= ~(SCX_TG_ONLINE | SCX_TG_INITED);
 }
 
+/*
+ * @p's sched for the cgroup migration paths. Stable as re-homes happen either
+ * at CGROUP_TASK_MIGRATED of the same migration or under scx_cgroup_lock(),
+ * both while holding cgroup_mutex.
+ */
+static struct scx_sched *scx_cgroup_task_sched(struct task_struct *p)
+{
+	return rcu_dereference_protected(p->scx.sched, lockdep_is_held(&cgroup_mutex));
+}
+
 int scx_cgroup_can_attach(struct cgroup_taskset *tset)
 {
-	struct scx_sched *sch = scx_root;
 	struct cgroup_subsys_state *css;
 	struct task_struct *p;
 	int ret;
@@ -4443,6 +4816,7 @@ int scx_cgroup_can_attach(struct cgroup_taskset *tset)
 		return 0;
 
 	cgroup_taskset_for_each(p, css, tset) {
+		struct scx_sched *sch = scx_cgroup_task_sched(p);
 		struct cgroup *from = tg_cgrp(task_group(p));
 		struct cgroup *to = tg_cgrp(css_tg(css));
 
@@ -4456,11 +4830,22 @@ int scx_cgroup_can_attach(struct cgroup_taskset *tset)
 		if (from == to)
 			continue;
 
+		/*
+		 * The cgroup_move ops are delivered to @p's sched, and only for
+		 * moves that don't re-home @p. A re-homing move changes the dfl
+		 * cgroup's sched and is reported through the
+		 * exit_task/init_task pair that the re-homing generates.
+		 */
+		if (!sch || sch != scx_cgroup_sched(task_css_set(p)->mg_dst_cset->dfl_cgrp))
+			continue;
+
 		if (SCX_HAS_OP(sch, cgroup_prep_move)) {
 			ret = SCX_CALL_OP_RET(sch, cgroup_prep_move, NULL,
 					      p, from, css->cgroup);
-			if (ret)
+			if (ret) {
+				ret = scx_ops_sanitize_err(sch, "cgroup_prep_move", ret);
 				goto err;
+			}
 		}
 
 		p->scx.cgrp_moving_from = from;
@@ -4470,31 +4855,33 @@ int scx_cgroup_can_attach(struct cgroup_taskset *tset)
 
 err:
 	cgroup_taskset_for_each(p, css, tset) {
-		if (SCX_HAS_OP(sch, cgroup_cancel_move) &&
-		    p->scx.cgrp_moving_from)
+		struct scx_sched *sch = scx_cgroup_task_sched(p);
+
+		/* cgrp_moving_from implies non-NULL @sch, test it first */
+		if (p->scx.cgrp_moving_from && SCX_HAS_OP(sch, cgroup_cancel_move))
 			SCX_CALL_OP(sch, cgroup_cancel_move, NULL,
 				    p, p->scx.cgrp_moving_from, css->cgroup);
 		p->scx.cgrp_moving_from = NULL;
 	}
 
-	return ops_sanitize_err(sch, "cgroup_prep_move", ret);
+	return ret;
 }
 
 void scx_cgroup_move_task(struct task_struct *p)
 {
-	struct scx_sched *sch = scx_root;
+	struct scx_sched *sch;
 
 	if (!scx_cgroup_enabled)
 		return;
 
 	/*
-	 * scx_cgroup_can_attach() sets cgrp_moving_from only when the task's
-	 * cgroup changes. Migration keys off css rather than cgroup identity,
-	 * so it can hand an unchanged-cgroup task here with cgrp_moving_from
-	 * NULL. Nothing to report to the BPF scheduler then, so skip it and
-	 * keep prep_move and move paired.
+	 * Migration keys off css rather than cgroup identity, so it can hand an
+	 * unchanged-cgroup task here with cgrp_moving_from NULL. Nothing to
+	 * report to the BPF scheduler then, so skip it and keep prep_move and
+	 * move paired.
 	 */
-	if (SCX_HAS_OP(sch, cgroup_move) && p->scx.cgrp_moving_from)
+	sch = scx_cgroup_task_sched(p);
+	if (p->scx.cgrp_moving_from && SCX_HAS_OP(sch, cgroup_move))
 		SCX_CALL_OP_TASK(sch, cgroup_move, task_rq(p),
 				 p, p->scx.cgrp_moving_from,
 				 tg_cgrp(task_group(p)));
@@ -4503,7 +4890,6 @@ void scx_cgroup_move_task(struct task_struct *p)
 
 void scx_cgroup_cancel_attach(struct cgroup_taskset *tset)
 {
-	struct scx_sched *sch = scx_root;
 	struct cgroup_subsys_state *css;
 	struct task_struct *p;
 
@@ -4511,8 +4897,10 @@ void scx_cgroup_cancel_attach(struct cgroup_taskset *tset)
 		return;
 
 	cgroup_taskset_for_each(p, css, tset) {
-		if (SCX_HAS_OP(sch, cgroup_cancel_move) &&
-		    p->scx.cgrp_moving_from)
+		struct scx_sched *sch = scx_cgroup_task_sched(p);
+
+		/* cgrp_moving_from implies non-NULL @sch, test it first */
+		if (p->scx.cgrp_moving_from && SCX_HAS_OP(sch, cgroup_cancel_move))
 			SCX_CALL_OP(sch, cgroup_cancel_move, NULL,
 				    p, p->scx.cgrp_moving_from, css->cgroup);
 		p->scx.cgrp_moving_from = NULL;
@@ -4524,9 +4912,9 @@ void scx_group_set_weight(struct task_group *tg, unsigned long weight)
 	struct scx_sched *sch;
 
 	percpu_down_read(&scx_cgroup_ops_rwsem);
-	sch = scx_root;
+	sch = scx_tg_knob_sched(tg);
 
-	if (scx_cgroup_enabled && SCX_HAS_OP(sch, cgroup_set_weight) &&
+	if (scx_cgroup_enabled && sch && SCX_HAS_OP(sch, cgroup_set_weight) &&
 	    tg->scx.weight != weight)
 		SCX_CALL_OP(sch, cgroup_set_weight, NULL, tg_cgrp(tg), weight);
 
@@ -4540,9 +4928,9 @@ void scx_group_set_idle(struct task_group *tg, bool idle)
 	struct scx_sched *sch;
 
 	percpu_down_read(&scx_cgroup_ops_rwsem);
-	sch = scx_root;
+	sch = scx_tg_knob_sched(tg);
 
-	if (scx_cgroup_enabled && SCX_HAS_OP(sch, cgroup_set_idle))
+	if (scx_cgroup_enabled && sch && SCX_HAS_OP(sch, cgroup_set_idle))
 		SCX_CALL_OP(sch, cgroup_set_idle, NULL, tg_cgrp(tg), idle);
 
 	/* Update the task group's idle state */
@@ -4557,9 +4945,9 @@ void scx_group_set_bandwidth(struct task_group *tg,
 	struct scx_sched *sch;
 
 	percpu_down_read(&scx_cgroup_ops_rwsem);
-	sch = scx_root;
+	sch = scx_tg_knob_sched(tg);
 
-	if (scx_cgroup_enabled && SCX_HAS_OP(sch, cgroup_set_bandwidth) &&
+	if (scx_cgroup_enabled && sch && SCX_HAS_OP(sch, cgroup_set_bandwidth) &&
 	    (tg->scx.bw_period_us != period_us ||
 	     tg->scx.bw_quota_us != quota_us ||
 	     tg->scx.bw_burst_us != burst_us))
@@ -4585,7 +4973,7 @@ static struct cgroup *root_cgroup(void)
  * for cgroup_mutex deadlocks with cgroup teardown, which holds it while
  * draining a set_* file write blocked on the rwsem behind the writer.
  */
-static void scx_cgroup_lock(void)
+void scx_cgroup_lock(void)
 {
 	cgroup_lock();
 #ifdef CONFIG_EXT_GROUP_SCHED
@@ -4593,7 +4981,7 @@ static void scx_cgroup_lock(void)
 #endif
 }
 
-static void scx_cgroup_unlock(void)
+void scx_cgroup_unlock(void)
 {
 #ifdef CONFIG_EXT_GROUP_SCHED
 	percpu_up_write(&scx_cgroup_ops_rwsem);
@@ -4606,26 +4994,6 @@ static inline void scx_cgroup_lock(void) {}
 static inline void scx_cgroup_unlock(void) {}
 #endif	/* CONFIG_EXT_GROUP_SCHED || CONFIG_EXT_SUB_SCHED */
 
-#ifdef CONFIG_EXT_SUB_SCHED
-static struct cgroup *sch_cgroup(struct scx_sched *sch)
-{
-	return sch->cgrp;
-}
-
-/* for each descendant of @cgrp including self, set ->scx_sched to @sch */
-static void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch)
-{
-	struct cgroup *pos;
-	struct cgroup_subsys_state *css;
-
-	cgroup_for_each_live_descendant_pre(pos, css, cgrp)
-		rcu_assign_pointer(pos->scx_sched, sch);
-}
-#else	/* CONFIG_EXT_SUB_SCHED */
-static inline struct cgroup *sch_cgroup(struct scx_sched *sch) { return NULL; }
-static inline void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch) {}
-#endif	/* CONFIG_EXT_SUB_SCHED */
-
 /*
  * Omitted operations:
  *
@@ -4669,8 +5037,7 @@ DEFINE_SCHED_CLASS(ext) = {
 #endif
 };
 
-static s32 init_dsq(struct scx_dispatch_q *dsq, u64 dsq_id,
-		    struct scx_sched *sch)
+s32 scx_init_dsq(struct scx_dispatch_q *dsq, u64 dsq_id, struct scx_sched *sch)
 {
 	s32 cpu;
 
@@ -4760,7 +5127,7 @@ static void destroy_dsq(struct scx_sched *sch, u64 dsq_id)
 		goto out_unlock_dsq;
 
 	/*
-	 * Mark dead by invalidating ->id to prevent dispatch_enqueue() from
+	 * Mark dead by invalidating ->id to prevent scx_dispatch_enqueue() from
 	 * queueing more tasks. As this function can be called from anywhere,
 	 * freeing is bounced through an irq work to avoid nesting RCU
 	 * operations inside scheduler locks.
@@ -4780,8 +5147,6 @@ static void scx_cgroup_exit(struct scx_sched *sch)
 {
 	struct cgroup_subsys_state *css;
 
-	scx_cgroup_enabled = false;
-
 	/*
 	 * scx_tg_on/offline() are excluded through cgroup_lock(). If we walk
 	 * cgroups and exit all the inited ones, all online cgroups are exited.
@@ -4789,14 +5154,13 @@ static void scx_cgroup_exit(struct scx_sched *sch)
 	css_for_each_descendant_post(css, &root_task_group.css) {
 		struct task_group *tg = css_tg(css);
 
-		if (!(tg->scx.flags & SCX_TG_INITED))
-			continue;
-		tg->scx.flags &= ~SCX_TG_INITED;
-
-		if (!sch->ops.cgroup_exit)
-			continue;
-
-		SCX_CALL_OP(sch, cgroup_exit, NULL, css->cgroup);
+		/* also clear the sched of tgs whose ops.cgroup_init() failed */
+		tg->scx.sched = NULL;
+		if (tg->scx.flags & SCX_TG_INITED) {
+			tg->scx.flags &= ~SCX_TG_INITED;
+			if (sch->ops.cgroup_exit)
+				SCX_CALL_OP(sch, cgroup_exit, NULL, css->cgroup);
+		}
 	}
 }
 
@@ -4811,34 +5175,29 @@ static int scx_cgroup_init(struct scx_sched *sch)
 	 */
 	css_for_each_descendant_pre(css, &root_task_group.css) {
 		struct task_group *tg = css_tg(css);
-		struct scx_cgroup_init_args args = {
-			.weight = tg->scx.weight,
-			.bw_period_us = tg->scx.bw_period_us,
-			.bw_quota_us = tg->scx.bw_quota_us,
-			.bw_burst_us = tg->scx.bw_burst_us,
-		};
 
-		if ((tg->scx.flags &
-		     (SCX_TG_ONLINE | SCX_TG_INITED)) != SCX_TG_ONLINE)
+		if ((tg->scx.flags & (SCX_TG_ONLINE | SCX_TG_INITED)) != SCX_TG_ONLINE)
 			continue;
 
-		if (!sch->ops.cgroup_init) {
-			tg->scx.flags |= SCX_TG_INITED;
-			continue;
-		}
+		if (sch->ops.cgroup_init) {
+			struct scx_cgroup_init_args args = {
+				.weight = tg->scx.weight,
+				.bw_period_us = tg->scx.bw_period_us,
+				.bw_quota_us = tg->scx.bw_quota_us,
+				.bw_burst_us = tg->scx.bw_burst_us,
+			};
 
-		ret = SCX_CALL_OP_RET(sch, cgroup_init, NULL,
-				      css->cgroup, &args);
-		if (ret) {
-			scx_error(sch, "ops.cgroup_init() failed (%d)", ret);
-			return ret;
+			ret = SCX_CALL_OP_RET(sch, cgroup_init, NULL, css->cgroup, &args);
+			if (ret) {
+				scx_error(sch, "ops.cgroup_init() failed (%d)", ret);
+				return ret;
+			}
 		}
+
+		tg->scx.sched = sch;
 		tg->scx.flags |= SCX_TG_INITED;
 	}
 
-	WARN_ON_ONCE(scx_cgroup_enabled);
-	scx_cgroup_enabled = true;
-
 	return 0;
 }
 
@@ -4908,8 +5267,10 @@ static const struct attribute_group scx_global_attr_group = {
 
 static void free_pnode(struct scx_sched_pnode *pnode);
 static void free_exit_info(struct scx_exit_info *ei);
+static const char *scx_exit_reason(enum scx_exit_kind kind);
+static bool scx_claim_exit(struct scx_sched *sch, enum scx_exit_kind kind);
 
-static s32 scx_set_cmask_scratch_alloc(struct scx_sched *sch)
+s32 scx_set_cmask_scratch_alloc(struct scx_sched *sch)
 {
 	size_t size = struct_size_t(struct scx_cmask, bits,
 				    SCX_CMASK_NR_WORDS(num_possible_cpus()));
@@ -4959,11 +5320,13 @@ static void scx_sched_free_rcu_work(struct work_struct *work)
 	struct scx_dispatch_q *dsq;
 	int cpu, node;
 
+	irq_work_sync(&sch->propagate_exit_irq_work);
 	irq_work_sync(&sch->disable_irq_work);
 	kthread_destroy_worker(sch->helper);
 	timer_shutdown_sync(&sch->bypass_lb_timer);
 	free_cpumask_var(sch->bypass_lb_donee_cpumask);
 	free_cpumask_var(sch->bypass_lb_resched_cpumask);
+	free_cpumask_var(sch->stall_cpus);
 
 #ifdef CONFIG_EXT_SUB_SCHED
 	kfree(sch->cgrp_path);
@@ -4985,7 +5348,21 @@ static void scx_sched_free_rcu_work(struct work_struct *work)
 		 */
 		WARN_ON_ONCE(!list_empty(&pcpu->deferred_reenq_local.node));
 
-		exit_dsq(bypass_dsq(sch, cpu));
+		/* remove the queued ecaps sync so the pcpu can be freed */
+		scx_discard_ecaps_to_sync(cpu, pcpu);
+
+		/*
+		 * Bypass blocks new kicks. Flush the kick irq_work so this
+		 * pcpu's to_kick_node is off the list before it is freed.
+		 */
+		irq_work_sync(&cpu_rq(cpu)->scx.kick_cpus_irq_work);
+		WARN_ON_ONCE(!list_empty(&pcpu->to_kick_node));
+		free_cpumask_var(pcpu->cpus_to_kick);
+		free_cpumask_var(pcpu->cpus_to_kick_if_idle);
+		free_cpumask_var(pcpu->cpus_to_preempt);
+		free_cpumask_var(pcpu->cpus_to_wait);
+
+		exit_dsq(scx_bypass_dsq(sch, cpu));
 	}
 
 	free_percpu(sch->pcpu);
@@ -4994,6 +5371,8 @@ static void scx_sched_free_rcu_work(struct work_struct *work)
 		free_pnode(sch->pnode[node]);
 	kfree(sch->pnode);
 
+	scx_free_pshards(sch);
+
 	rhashtable_walk_enter(&sch->dsq_hash, &rht_iter);
 	do {
 		rhashtable_walk_start(&rht_iter);
@@ -5011,6 +5390,10 @@ static void scx_sched_free_rcu_work(struct work_struct *work)
 	scx_arena_pool_destroy(sch);
 	if (sch->arena_map)
 		bpf_map_put(sch->arena_map);
+
+	/* @sch is completely inactive by now */
+	scx_dec_has_subs(sch);
+
 	kfree(sch);
 }
 
@@ -5043,26 +5426,62 @@ static ssize_t scx_attr_events_show(struct kobject *kobj,
 	int at = 0;
 
 	scx_read_events(sch, &events);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_SELECT_CPU_FALLBACK);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_DISPATCH_KEEP_LAST);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_ENQ_SKIP_EXITING);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_ENQ_SKIP_MIGRATION_DISABLED);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_REENQ_IMMED);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_REENQ_LOCAL_REPEAT);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_REFILL_SLICE_DFL);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_BYPASS_DURATION);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_BYPASS_DISPATCH);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_BYPASS_ACTIVATE);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_INSERT_NOT_OWNED);
-	at += scx_attr_event_show(buf, at, &events, SCX_EV_SUB_BYPASS_DISPATCH);
+#define SCX_EVENT(name)	(at += scx_attr_event_show(buf, at, &events, name))
+	SCX_EVENTS_LIST(SCX_EVENT);
+#undef SCX_EVENT
 	return at;
 }
 SCX_ATTR(events);
 
+#ifdef CONFIG_EXT_SUB_SCHED
+static const char *scx_cap_names[__SCX_NR_CAPS] = {
+	[__SCX_CAP_ENQ_IMMED]	= "enq_immed",
+	[__SCX_CAP_ENQ]		= "enq",
+	[__SCX_CAP_PREEMPT]	= "preempt",
+	[__SCX_CAP_PERF]	= "perf",
+};
+
+static ssize_t scx_attr_caps_show(struct kobject *kobj,
+				  struct kobj_attribute *ka, char *buf)
+{
+	struct scx_sched *sch = container_of(kobj, struct scx_sched, kobj);
+	u32 npossible = num_possible_cpus();
+	struct scx_cmask *agg __free(kfree) =
+		kzalloc(struct_size(agg, bits, SCX_CMASK_NR_WORDS(npossible)), GFP_KERNEL);
+	unsigned long *agg_bm __free(bitmap) = bitmap_zalloc(npossible, GFP_KERNEL);
+	ssize_t count = 0;
+	s32 cap, si;
+
+	if (!agg || !agg_bm)
+		return -ENOMEM;
+
+	for (cap = 0; cap < __SCX_NR_CAPS; cap++) {
+		SCX_CMASK_DEFINE(snap, 0, SCX_CID_SHARD_MAX_CPUS);
+
+		scx_cmask_init(agg, 0, npossible);
+		for (si = 0; si < sch->nr_pshards; si++) {
+			struct scx_cmask *cm = &sch->pshard[si]->caps[cap].cmask;
+
+			scx_cmask_reframe(snap, cm->base, cm->nr_cids);
+			scx_cmask_copy(snap, cm);
+			scx_cmask_or(agg, snap);
+		}
+		/* %*pbl takes unsigned long bitmap layout, convert from u64 */
+		bitmap_from_arr64(agg_bm, agg->bits, npossible);
+		count += sysfs_emit_at(buf, count, "%s: %*pbl\n",
+				       scx_cap_names[cap], npossible, agg_bm);
+	}
+	return count;
+}
+SCX_ATTR(caps);
+#endif	/* CONFIG_EXT_SUB_SCHED */
+
 static struct attribute *scx_sched_attrs[] = {
 	&scx_attr_ops.attr,
 	&scx_attr_events.attr,
+#ifdef CONFIG_EXT_SUB_SCHED
+	&scx_attr_caps.attr,
+#endif
 	NULL,
 };
 ATTRIBUTE_GROUPS(scx_sched);
@@ -5149,6 +5568,7 @@ bool scx_allow_ttwu_queue(const struct task_struct *p)
 
 /**
  * handle_lockup - sched_ext common lockup handler
+ * @exit_cpu: CPU to record in exit_info. Pass the stalled/hung CPU, not current.
  * @fmt: format string
  *
  * Called on system stall or lockup condition and initiates abort of sched_ext
@@ -5158,7 +5578,7 @@ bool scx_allow_ttwu_queue(const struct task_struct *p)
  * resolve the lockup. %false if sched_ext is not enabled or abort was already
  * initiated by someone else.
  */
-static __printf(1, 2) bool handle_lockup(const char *fmt, ...)
+static __printf(2, 3) bool handle_lockup(int exit_cpu, const char *fmt, ...)
 {
 	struct scx_sched *sch;
 	va_list args;
@@ -5174,7 +5594,7 @@ static __printf(1, 2) bool handle_lockup(const char *fmt, ...)
 	case SCX_ENABLING:
 	case SCX_ENABLED:
 		va_start(args, fmt);
-		ret = scx_verror(sch, fmt, args);
+		ret = scx_vexit(sch, SCX_EXIT_ERROR, 0, exit_cpu, fmt, args);
 		va_end(args);
 		return ret;
 	default:
@@ -5184,6 +5604,7 @@ static __printf(1, 2) bool handle_lockup(const char *fmt, ...)
 
 /**
  * scx_rcu_cpu_stall - sched_ext RCU CPU stall handler
+ * @stalled_mask: bit mask of stalled CPUs
  *
  * While there are various reasons why RCU CPU stalls can occur on a system
  * that may not be caused by the current BPF scheduler, try kicking out the
@@ -5194,9 +5615,46 @@ static __printf(1, 2) bool handle_lockup(const char *fmt, ...)
  * resolve the reported RCU stall. %false if sched_ext is not enabled or someone
  * else already initiated abort.
  */
-bool scx_rcu_cpu_stall(void)
+bool scx_rcu_cpu_stall(const struct cpumask *stalled_mask)
 {
-	return handle_lockup("RCU CPU stall detected!");
+	struct scx_sched *sch;
+	struct scx_exit_info *ei;
+	int exit_cpu;
+
+	guard(rcu)();
+
+	sch = rcu_dereference(scx_root);
+	if (unlikely(!sch))
+		return false;
+
+	switch (scx_enable_state()) {
+	case SCX_ENABLING:
+	case SCX_ENABLED:
+		break;
+	default:
+		return false;
+	}
+
+	exit_cpu = cpumask_empty(stalled_mask) ? -1 : (int)cpumask_first(stalled_mask);
+	ei = sch->exit_info;
+
+	guard(preempt)();
+
+	if (!scx_claim_exit(sch, SCX_EXIT_ERROR))
+		return false;
+
+#ifdef CONFIG_STACKTRACE
+	ei->bt_len = stack_trace_save(ei->bt, SCX_EXIT_BT_LEN, 1);
+#endif
+	scnprintf(ei->msg, SCX_EXIT_MSG_LEN, "RCU CPU stall on CPUs (%*pbl)",
+		  cpumask_pr_args(stalled_mask));
+	ei->kind = SCX_EXIT_ERROR;
+	ei->reason = scx_exit_reason(SCX_EXIT_ERROR);
+	ei->exit_cpu = exit_cpu;
+	cpumask_copy(sch->stall_cpus, stalled_mask);
+
+	irq_work_queue(&sch->disable_irq_work);
+	return true;
 }
 
 /**
@@ -5211,53 +5669,39 @@ bool scx_rcu_cpu_stall(void)
  */
 void scx_softlockup(u32 dur_s)
 {
-	if (!handle_lockup("soft lockup - CPU %d stuck for %us", smp_processor_id(), dur_s))
+	int cpu = smp_processor_id();
+
+	if (!handle_lockup(cpu, "soft lockup - CPU %d stuck for %us", cpu, dur_s))
 		return;
 
 	printk_deferred(KERN_ERR "sched_ext: Soft lockup - CPU %d stuck for %us, disabling BPF scheduler\n",
-			smp_processor_id(), dur_s);
+			cpu, dur_s);
 }
 
-/*
- * scx_hardlockup() runs from NMI and eventually calls scx_claim_exit(),
- * which takes scx_sched_lock. scx_sched_lock isn't NMI-safe and grabbing
- * it from NMI context can lead to deadlocks. Defer via irq_work; the
- * disable path runs off irq_work anyway.
- */
-static atomic_t scx_hardlockup_cpu = ATOMIC_INIT(-1);
-
-static void scx_hardlockup_irq_workfn(struct irq_work *work)
-{
-	int cpu = atomic_xchg(&scx_hardlockup_cpu, -1);
-
-	if (cpu >= 0 && handle_lockup("hard lockup - CPU %d", cpu))
-		printk_deferred(KERN_ERR "sched_ext: Hard lockup - CPU %d, disabling BPF scheduler\n",
-				cpu);
-}
-
-static DEFINE_IRQ_WORK(scx_hardlockup_irq_work, scx_hardlockup_irq_workfn);
-
 /**
  * scx_hardlockup - sched_ext hardlockup handler
+ * @cpu: the target CPU
  *
  * A poorly behaving BPF scheduler can trigger hard lockup by e.g. putting
  * numerous affinitized tasks in a single queue and directing all CPUs at it.
  * Try kicking out the current scheduler in an attempt to recover the system to
  * a good state before taking more drastic actions.
  *
- * Queues an irq_work; the handle_lockup() call happens in IRQ context (see
- * scx_hardlockup_irq_workfn).
+ * Called from NMI. Aborting the scheduler sets ->aborting throughout the
+ * hierarchy before returning, which is what breaks the dispatch-path live-locks
+ * that can hard-lock CPUs.
  *
- * Returns %true if sched_ext is enabled and the work was queued, %false
- * otherwise.
+ * Returns %true if sched_ext is enabled and abort was initiated, which may
+ * resolve the lockup. %false if sched_ext is not enabled or abort was already
+ * initiated by someone else.
  */
 bool scx_hardlockup(int cpu)
 {
-	if (!rcu_access_pointer(scx_root))
+	if (!handle_lockup(cpu, "hard lockup - CPU %d", cpu))
 		return false;
 
-	atomic_cmpxchg(&scx_hardlockup_cpu, -1, cpu);
-	irq_work_queue(&scx_hardlockup_irq_work);
+	printk_deferred(KERN_ERR "sched_ext: Hard lockup - CPU %d, disabling BPF scheduler\n",
+			cpu);
 	return true;
 }
 
@@ -5266,7 +5710,7 @@ static u32 bypass_lb_cpu(struct scx_sched *sch, s32 donor,
 			 u32 nr_donor_target, u32 nr_donee_target)
 {
 	struct rq *donor_rq = cpu_rq(donor);
-	struct scx_dispatch_q *donor_dsq = bypass_dsq(sch, donor);
+	struct scx_dispatch_q *donor_dsq = scx_bypass_dsq(sch, donor);
 	struct task_struct *p, *n;
 	struct scx_dsq_list_node cursor = INIT_DSQ_LIST_CURSOR(cursor, donor_dsq, 0);
 	s32 delta = READ_ONCE(donor_dsq->nr) - nr_donor_target;
@@ -5314,7 +5758,7 @@ resume:
 		if (donee >= nr_cpu_ids)
 			continue;
 
-		donee_dsq = bypass_dsq(sch, donee);
+		donee_dsq = scx_bypass_dsq(sch, donee);
 
 		/*
 		 * $p's rq is not locked but $p's DSQ lock protects its
@@ -5335,7 +5779,7 @@ resume:
 		 * between bypass DSQs.
 		 */
 		dispatch_dequeue_locked(p, donor_dsq);
-		dispatch_enqueue(sch, cpu_rq(donee), donee_dsq, p, SCX_ENQ_NESTED);
+		scx_dispatch_enqueue(sch, cpu_rq(donee), donee_dsq, p, 0, 0, SCX_ENQ_NESTED);
 
 		/*
 		 * $donee might have been idle and need to be woken up. No need
@@ -5350,6 +5794,7 @@ resume:
 		if (!(nr_balanced % SCX_BYPASS_LB_BATCH) && n) {
 			list_move_tail(&cursor.node, &n->scx.dsq_list.node);
 			raw_spin_unlock(&donor_dsq->lock);
+			scx_rq_lock_drop(donor_rq);
 			raw_spin_rq_unlock_irq(donor_rq);
 			cpu_relax();
 			raw_spin_rq_lock_irq(donor_rq);
@@ -5360,6 +5805,7 @@ resume:
 
 	list_del_init(&cursor.node);
 	raw_spin_unlock(&donor_dsq->lock);
+	scx_rq_lock_drop(donor_rq);
 	raw_spin_rq_unlock_irq(donor_rq);
 
 	return nr_balanced;
@@ -5378,7 +5824,7 @@ static void bypass_lb_node(struct scx_sched *sch, int node)
 
 	/* count the target tasks and CPUs */
 	for_each_cpu_and(cpu, cpu_online_mask, node_mask) {
-		u32 nr = READ_ONCE(bypass_dsq(sch, cpu)->nr);
+		u32 nr = READ_ONCE(scx_bypass_dsq(sch, cpu)->nr);
 
 		nr_tasks += nr;
 		nr_cpus++;
@@ -5400,7 +5846,7 @@ static void bypass_lb_node(struct scx_sched *sch, int node)
 
 	cpumask_clear(donee_mask);
 	for_each_cpu_and(cpu, cpu_online_mask, node_mask) {
-		if (READ_ONCE(bypass_dsq(sch, cpu)->nr) < nr_target)
+		if (READ_ONCE(scx_bypass_dsq(sch, cpu)->nr) < nr_target)
 			cpumask_set_cpu(cpu, donee_mask);
 	}
 
@@ -5411,7 +5857,7 @@ static void bypass_lb_node(struct scx_sched *sch, int node)
 			break;
 		if (cpumask_test_cpu(cpu, donee_mask))
 			continue;
-		if (READ_ONCE(bypass_dsq(sch, cpu)->nr) <= nr_donor_target)
+		if (READ_ONCE(scx_bypass_dsq(sch, cpu)->nr) <= nr_donor_target)
 			continue;
 
 		nr_balanced += bypass_lb_cpu(sch, cpu, donee_mask, resched_mask,
@@ -5422,7 +5868,7 @@ static void bypass_lb_node(struct scx_sched *sch, int node)
 		resched_cpu(cpu);
 
 	for_each_cpu_and(cpu, cpu_online_mask, node_mask) {
-		u32 nr = READ_ONCE(bypass_dsq(sch, cpu)->nr);
+		u32 nr = READ_ONCE(scx_bypass_dsq(sch, cpu)->nr);
 
 		after_min = min(nr, after_min);
 		after_max = max(nr, after_max);
@@ -5448,7 +5894,7 @@ static void scx_bypass_lb_timerfn(struct timer_list *timer)
 	int node;
 	u32 intv_us;
 
-	if (!bypass_dsp_enabled(sch))
+	if (!scx_bypass_dsp_enabled(sch))
 		return;
 
 	for_each_node_with_cpus(node)
@@ -5514,9 +5960,9 @@ static void enable_bypass_dsp(struct scx_sched *sch)
 	 * dispatch enabled while a descendant is bypassing, which is all that's
 	 * required.
 	 *
-	 * bypass_dsp_enabled() test is used to determine whether to enter the
-	 * bypass dispatch handling path from both bypassing and hosting scheds.
-	 * Bump enable depth on both @sch and bypass dispatch host.
+	 * scx_bypass_dsp_enabled() test is used to determine whether to enter
+	 * the bypass dispatch handling path from both bypassing and hosting
+	 * scheds. Bump enable depth on both @sch and bypass dispatch host.
 	 */
 	ret = atomic_inc_return(&sch->bypass_dsp_enable_depth);
 	WARN_ON_ONCE(ret <= 0);
@@ -5536,7 +5982,7 @@ static void enable_bypass_dsp(struct scx_sched *sch)
 }
 
 /* may be called without holding scx_bypass_lock */
-static void disable_bypass_dsp(struct scx_sched *sch)
+void scx_disable_bypass_dsp(struct scx_sched *sch)
 {
 	s32 ret;
 
@@ -5553,6 +5999,38 @@ static void disable_bypass_dsp(struct scx_sched *sch)
 }
 
 /**
+ * unbypass_renotify_idle - Arm an idle re-notify for a sched leaving bypass
+ * @rq: rq of the cpu leaving bypass
+ * @pos: scheduler that just left bypass on @rq's cpu
+ * @pcpu: @pos's per-cpu state for @rq's cpu
+ *
+ * A sched leaving bypass is owed the ops.update_idle() calls suppressed while
+ * bypassing. A cpu that goes idle during the bypass window and stays idle won't
+ * produce a notification. Arm a re-notify that scx_bypass()'s resched flushes
+ * on the next idle pick.
+ *
+ * An acute case is ops.sub_attach(). If the parent grants the child cids while
+ * attaching, when attach is complete and bypass is lifted, the child may hold
+ * idle cids it never saw go idle.
+ *
+ * The root is no exception as bypass suppresses its notifications the same way.
+ * However, the root uses a separate per-rq flag so its re-notify keeps working
+ * even when !CONFIG_EXT_SUB_SCHED.
+ */
+static void unbypass_renotify_idle(struct rq *rq, struct scx_sched *pos,
+				   struct scx_sched_pcpu *pcpu)
+{
+	if (!pos->level) {
+		rq->scx.flags |= SCX_RQ_ROOT_IDLE_RENOTIFY;
+		return;
+	}
+#ifdef CONFIG_EXT_SUB_SCHED
+	pcpu->idle_renotify = true;
+	rq->scx.flags |= SCX_RQ_SUB_IDLE_RENOTIFY;
+#endif
+}
+
+/**
  * scx_bypass - [Un]bypass scx_ops and guarantee forward progress
  * @sch: sched to bypass
  * @bypass: true for bypass, false for unbypass
@@ -5573,18 +6051,18 @@ static void disable_bypass_dsp(struct scx_sched *sch)
  *
  * - ops.dispatch() is ignored.
  *
- * - balance_one() does not set %SCX_RQ_BAL_KEEP on non-zero slice as slice
+ * - dispatch_one() does not report %SCX_DSP_PREV on non-zero slice as slice
  *   can't be trusted. Whenever a tick triggers, the running task is rotated to
- *   the tail of the queue with core_sched_at touched.
+ *   the tail of the queue.
  *
  * - pick_next_task() suppresses zero slice warning.
  *
  * - scx_kick_cpu() is disabled to avoid irq_work malfunction during PM
  *   operations.
  *
- * - scx_prio_less() reverts to the default core_sched_at order.
+ * - scx_prio_less() reverts to the default runnable_at order.
  */
-static void scx_bypass(struct scx_sched *sch, bool bypass)
+void scx_bypass(struct scx_sched *sch, bool bypass)
 {
 	struct scx_sched *pos;
 	unsigned long flags;
@@ -5635,11 +6113,17 @@ static void scx_bypass(struct scx_sched *sch, bool bypass)
 
 		scx_for_each_descendant_pre(pos, sch) {
 			struct scx_sched_pcpu *pcpu = per_cpu_ptr(pos->pcpu, cpu);
+			bool was_bypassing = pcpu->flags & SCX_SCHED_PCPU_BYPASSING;
 
-			if (pos->bypass_depth)
+			if (pos->bypass_depth) {
 				pcpu->flags |= SCX_SCHED_PCPU_BYPASSING;
-			else
+			} else {
 				pcpu->flags &= ~SCX_SCHED_PCPU_BYPASSING;
+				if (was_bypassing) {
+					unbypass_renotify_idle(rq, pos, pcpu);
+					scx_unbypass_replay_ecaps(rq, pos);
+				}
+			}
 		}
 
 		raw_spin_unlock(&scx_sched_lock);
@@ -5650,6 +6134,7 @@ static void scx_bypass(struct scx_sched *sch, bool bypass)
 		 * sees scx_bypassing() before moving tasks to SCX.
 		 */
 		if (!scx_enabled()) {
+			scx_rq_lock_drop(rq);
 			raw_spin_rq_unlock(rq);
 			continue;
 		}
@@ -5666,6 +6151,13 @@ static void scx_bypass(struct scx_sched *sch, bool bypass)
 			if (!scx_is_descendant(scx_task_sched(p), sch))
 				continue;
 
+			/*
+			 * Bypass trumps protection. Cycling clears for queued
+			 * tasks but current task needs explicit stripping.
+			 */
+			if (bypass && task_current(rq, p))
+				scx_task_slice_ended(rq, p);
+
 			/* cycling deq/enq is enough, see the function comment */
 			scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
 				/* nothing */ ;
@@ -5676,12 +6168,13 @@ static void scx_bypass(struct scx_sched *sch, bool bypass)
 		if (cpu_online(cpu) || cpu == smp_processor_id())
 			resched_curr(rq);
 
+		scx_rq_lock_drop(rq);
 		raw_spin_rq_unlock(rq);
 	}
 
 	/* disarming must come after moving all tasks out of the bypass DSQs */
 	if (!bypass)
-		disable_bypass_dsp(sch);
+		scx_disable_bypass_dsp(sch);
 unlock:
 	raw_spin_unlock_irqrestore(&scx_bypass_lock, flags);
 }
@@ -5728,12 +6221,18 @@ static const char *scx_exit_reason(enum scx_exit_kind kind)
 		return "disabled by sysrq-S";
 	case SCX_EXIT_PARENT:
 		return "parent exiting";
+	case SCX_EXIT_PARENT_KILL:
+		return "killed by parent scheduler";
 	case SCX_EXIT_ERROR:
 		return "runtime error";
 	case SCX_EXIT_ERROR_BPF:
 		return "scx_bpf_error";
 	case SCX_EXIT_ERROR_STALL:
 		return "runnable task stall";
+	case SCX_EXIT_ERROR_REENQ:
+		return "reenqueue limit";
+	case SCX_EXIT_ERROR_RESCUE:
+		return "rescue bandwidth overload";
 	default:
 		return "<UNKNOWN>";
 	}
@@ -5747,6 +6246,8 @@ static void free_kick_syncs(void)
 		struct scx_kick_syncs __rcu **ksyncs = per_cpu_ptr(&scx_kick_syncs, cpu);
 		struct scx_kick_syncs *to_free;
 
+		/* flush the pending kick before freeing @ksyncs */
+		irq_work_sync(&cpu_rq(cpu)->scx.kick_cpus_irq_work);
 		to_free = rcu_replace_pointer(*ksyncs, NULL, true);
 		if (to_free)
 			kvfree_rcu(to_free, rcu);
@@ -5773,63 +6274,70 @@ static void refresh_watchdog(void)
 		cancel_delayed_work_sync(&scx_watchdog_work);
 }
 
-static s32 scx_link_sched(struct scx_sched *sch)
+s32 scx_link_sched(struct scx_sched *sch)
 {
-	const char *err_msg = "";
-	s32 ret = 0;
-
-	scoped_guard(raw_spinlock_irq, &scx_sched_lock) {
+	scoped_guard(raw_spinlock_irqsave, &scx_bypass_lock)	/* for the parent bypass check */
+	scoped_guard(raw_spinlock, &scx_sched_lock) {
 #ifdef CONFIG_EXT_SUB_SCHED
 		struct scx_sched *parent = scx_parent(sch);
 
 		if (parent) {
+			s32 ret;
+
 			/*
-			 * scx_claim_exit() propagates exit_kind transition to
-			 * its sub-scheds while holding scx_sched_lock - either
-			 * we can see the parent's non-NONE exit_kind or the
-			 * parent can shoot us down.
+			 * Bypass state is spread across per-cpu flags and a
+			 * depth count, so inheriting it is tricky and has no
+			 * valid use case. Refuse it.
 			 */
-			if (atomic_read(&parent->exit_kind) != SCX_EXIT_NONE) {
-				err_msg = "parent disabled";
-				ret = -ENOENT;
-				break;
+			if (READ_ONCE(parent->bypass_depth)) {
+				scx_error(sch, "parent bypassing (%d)", -EBUSY);
+				return -EBUSY;
 			}
 
 			ret = rhashtable_lookup_insert_fast(&scx_sched_hash,
 					&sch->hash_node, scx_sched_hash_params);
 			if (ret) {
-				err_msg = "failed to insert into scx_sched_hash";
-				break;
+				scx_error(sch, "failed to insert into scx_sched_hash (%d)",
+					  ret);
+				return ret;
 			}
 
-			list_add_tail(&sch->sibling, &parent->children);
+			list_add_tail_rcu(&sch->sibling, &parent->children);
+
+			/*
+			 * Pairs with the mb after the ->aborting assertion in
+			 * scx_claim_exit(). Either we see ->aborting and back
+			 * out, or the exit path sees us and exits us.
+			 */
+			smp_mb();
+			if (unlikely(READ_ONCE(parent->aborting))) {
+				rhashtable_remove_fast(&scx_sched_hash, &sch->hash_node,
+						       scx_sched_hash_params);
+				list_del_rcu(&sch->sibling);
+				scx_error(sch, "parent disabled (%d)", -ENOENT);
+				return -ENOENT;
+			}
+
+			sch->linked = true;
 		}
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
 		list_add_tail_rcu(&sch->all, &scx_sched_all);
 	}
 
-	/*
-	 * scx_error() takes scx_sched_lock via scx_claim_exit(), so it must run after
-	 * the guard above is released.
-	 */
-	if (ret) {
-		scx_error(sch, "%s (%d)", err_msg, ret);
-		return ret;
-	}
-
 	refresh_watchdog();
 	return 0;
 }
 
-static void scx_unlink_sched(struct scx_sched *sch)
+void scx_unlink_sched(struct scx_sched *sch)
 {
 	scoped_guard(raw_spinlock_irq, &scx_sched_lock) {
 #ifdef CONFIG_EXT_SUB_SCHED
-		if (scx_parent(sch)) {
+		if (sch->linked) {
 			rhashtable_remove_fast(&scx_sched_hash, &sch->hash_node,
 					       scx_sched_hash_params);
-			list_del_init(&sch->sibling);
+			list_del_rcu(&sch->sibling);
+			sch->linked = false;
 		}
 #endif	/* CONFIG_EXT_SUB_SCHED */
 		list_del_rcu(&sch->all);
@@ -5843,13 +6351,13 @@ static void scx_unlink_sched(struct scx_sched *sch)
  * @sch. Once @sch becomes empty during disable, there's no point in dumping it.
  * This prevents calling dump ops on a dead sch.
  */
-static void scx_disable_dump(struct scx_sched *sch)
+void scx_disable_dump(struct scx_sched *sch)
 {
 	guard(raw_spinlock_irqsave)(&scx_dump_lock);
 	sch->dump_disabled = true;
 }
 
-static void scx_log_sched_disable(struct scx_sched *sch)
+void scx_log_sched_disable(struct scx_sched *sch)
 {
 	struct scx_exit_info *ei = sch->exit_info;
 	const char *type = scx_parent(sch) ? "sub-scheduler" : "scheduler";
@@ -5869,219 +6377,6 @@ static void scx_log_sched_disable(struct scx_sched *sch)
 	}
 }
 
-#ifdef CONFIG_EXT_SUB_SCHED
-static DECLARE_WAIT_QUEUE_HEAD(scx_unlink_waitq);
-
-static void drain_descendants(struct scx_sched *sch)
-{
-	/*
-	 * Child scheds that finished the critical part of disabling will take
-	 * themselves off @sch->children. Wait for it to drain. As propagation
-	 * is recursive, empty @sch->children means that all proper descendant
-	 * scheds reached unlinking stage.
-	 */
-	wait_event(scx_unlink_waitq, list_empty(&sch->children));
-}
-
-static void scx_fail_parent(struct scx_sched *sch,
-			    struct task_struct *failed, s32 fail_code)
-{
-	struct scx_sched *parent = scx_parent(sch);
-	struct scx_task_iter sti;
-	struct task_struct *p;
-
-	scx_error(parent, "ops.init_task() failed (%d) for %s[%d] while disabling a sub-scheduler",
-		  fail_code, failed->comm, failed->pid);
-
-	/*
-	 * Once $parent is bypassed, it's safe to put SCX_TASK_NONE tasks into
-	 * it. This may cause downstream failures on the BPF side but $parent is
-	 * dying anyway.
-	 */
-	scx_bypass(parent, true);
-
-	scx_task_iter_start(&sti, sch->cgrp);
-	while ((p = scx_task_iter_next_locked(&sti))) {
-		if (scx_task_on_sched(parent, p))
-			continue;
-
-		scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
-			scx_disable_and_exit_task(sch, p);
-			scx_set_task_sched(p, parent);
-		}
-	}
-	scx_task_iter_stop(&sti);
-}
-
-static void scx_sub_disable(struct scx_sched *sch)
-{
-	struct scx_sched *parent = scx_parent(sch);
-	struct scx_task_iter sti;
-	struct task_struct *p;
-	int ret;
-
-	/*
-	 * Guarantee forward progress and wait for descendants to be disabled.
-	 * To limit disruptions, $parent is not bypassed. Tasks are fully
-	 * prepped and then inserted back into $parent.
-	 */
-	scx_bypass(sch, true);
-	drain_descendants(sch);
-
-	/*
-	 * Here, every runnable task is guaranteed to make forward progress and
-	 * we can safely use blocking synchronization constructs. Actually
-	 * disable ops.
-	 */
-	mutex_lock(&scx_enable_mutex);
-	percpu_down_write(&scx_fork_rwsem);
-	scx_cgroup_lock();
-
-	/*
-	 * An enable that failed before scx_link_sched() never owned a cgroup or
-	 * task and won't be waited on by an ancestor's drain_descendants().
-	 * Nothing to reparent and walking the tasks can misbehave as the task
-	 * ownership invariant (either owned by self or parent) does not hold.
-	 */
-	if (list_empty(&sch->sibling))
-		goto dump;
-
-	set_cgroup_sched(sch_cgroup(sch), parent);
-
-	scx_task_iter_start(&sti, sch->cgrp);
-	while ((p = scx_task_iter_next_locked(&sti))) {
-		struct rq *rq;
-		struct rq_flags rf;
-
-		/* filter out duplicate visits */
-		if (scx_task_on_sched(parent, p))
-			continue;
-
-		/*
-		 * By the time control reaches here, all linked descendant
-		 * schedulers should have been disabled.
-		 */
-		WARN_ON_ONCE(!scx_task_on_sched(sch, p));
-
-		/*
-		 * @p is pinned by the iter: css_task_iter_next() takes a
-		 * reference and holds it until the next iter_next() call, so
-		 * @p->usage is guaranteed > 0.
-		 */
-		get_task_struct(p);
-
-		scx_task_iter_unlock(&sti);
-
-		/*
-		 * $p is READY or ENABLED on @sch. Initialize for $parent,
-		 * disable and exit from @sch, and then switch over to $parent.
-		 *
-		 * If a task fails to initialize for $parent, the only available
-		 * action is disabling $parent too. While this allows disabling
-		 * of a child sched to cause the parent scheduler to fail, the
-		 * failure can only originate from ops.init_task() of the
-		 * parent. A child can't directly affect the parent through its
-		 * own failures.
-		 */
-		ret = __scx_init_task(parent, p, false);
-		if (ret) {
-			scx_fail_parent(sch, p, ret);
-			put_task_struct(p);
-			break;
-		}
-
-		rq = task_rq_lock(p, &rf);
-
-		if (scx_get_task_state(p) == SCX_TASK_DEAD) {
-			/*
-			 * sched_ext_dead() raced us between __scx_init_task()
-			 * and this rq lock and ran exit_task() on @sch (the
-			 * sched @p was on at that point), not on $parent.
-			 * $parent's just-completed init is owed an exit_task()
-			 * and we issue it here.
-			 */
-			scx_sub_init_cancel_task(parent, p);
-			task_rq_unlock(rq, p, &rf);
-			put_task_struct(p);
-			continue;
-		}
-
-		scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
-			/*
-			 * $p is initialized for $parent and still attached to
-			 * @sch. Disable and exit for @sch, switch over to
-			 * $parent and override the state to READY to account
-			 * for $p having already been initialized.
-			 */
-			scx_disable_and_exit_task(sch, p);
-			scx_set_task_state(p, SCX_TASK_INIT_BEGIN);
-			scx_set_task_state(p, SCX_TASK_INIT);
-			scx_set_task_sched(p, parent);
-			scx_set_task_state(p, SCX_TASK_READY);
-
-			/*
-			 * A task on a non-ext class, possible under an
-			 * %SCX_OPS_SWITCH_PARTIAL root, stays READY and is
-			 * enabled by switching_to_scx() if it switches over.
-			 */
-			if (p->sched_class == &ext_sched_class)
-				scx_enable_task(parent, p);
-		}
-
-		task_rq_unlock(rq, p, &rf);
-		put_task_struct(p);
-	}
-	scx_task_iter_stop(&sti);
-
-dump:
-	scx_disable_dump(sch);
-
-	scx_cgroup_unlock();
-	percpu_up_write(&scx_fork_rwsem);
-
-	/*
-	 * All tasks are moved off of @sch but there may still be on-going
-	 * operations (e.g. ops.select_cpu()). Drain them by flushing RCU. Use
-	 * the expedited version as ancestors may be waiting in bypass mode.
-	 * Also, tell the parent that there is no need to keep running bypass
-	 * DSQs for us.
-	 */
-	synchronize_rcu_expedited();
-	disable_bypass_dsp(sch);
-
-	scx_unlink_sched(sch);
-
-	mutex_unlock(&scx_enable_mutex);
-
-	/*
-	 * @sch is now unlinked from the parent's children list. Notify and call
-	 * ops.sub_detach/exit(). Note that ops.sub_detach/exit() must be called
-	 * after unlinking and releasing all locks. See scx_claim_exit().
-	 */
-	wake_up_all(&scx_unlink_waitq);
-
-	if (parent->ops.sub_detach && sch->sub_attached) {
-		struct scx_sub_detach_args sub_detach_args = {
-			.ops = &sch->ops,
-			.cgroup_path = sch->cgrp_path,
-		};
-		SCX_CALL_OP(parent, sub_detach, NULL,
-			    &sub_detach_args);
-	}
-
-	scx_log_sched_disable(sch);
-
-	if (sch->ops.exit)
-		SCX_CALL_OP(sch, exit, NULL, sch->exit_info);
-	if (sch->sub_kset)
-		kobject_del(&sch->sub_kset->kobj);
-	kobject_del(&sch->kobj);
-}
-#else	/* CONFIG_EXT_SUB_SCHED */
-static inline void drain_descendants(struct scx_sched *sch) { }
-static inline void scx_sub_disable(struct scx_sched *sch) { }
-#endif	/* CONFIG_EXT_SUB_SCHED */
-
 static void scx_root_disable(struct scx_sched *sch)
 {
 	struct scx_task_iter sti;
@@ -6119,10 +6414,11 @@ static void scx_root_disable(struct scx_sched *sch)
 	WRITE_ONCE(scx_switching_all, false);
 
 	/*
-	 * Shut down cgroup support before tasks so that the cgroup attach path
-	 * doesn't race against scx_disable_and_exit_task().
+	 * Shut down cgroup support before tasks so that the cgroup attach and
+	 * migration paths don't race against scx_disable_and_exit_task().
 	 */
 	scx_cgroup_lock();
+	scx_cgroup_enabled = false;
 	scx_cgroup_exit(sch);
 	scx_cgroup_unlock();
 
@@ -6162,12 +6458,9 @@ static void scx_root_disable(struct scx_sched *sch)
 	percpu_up_write(&scx_fork_rwsem);
 
 	/*
-	 * Invalidate all the rq clocks to prevent getting outdated
-	 * rq clocks from a previous scx scheduler.
-	 *
-	 * Also re-balance the dl_server bandwidth reservations: detach
-	 * ext_server (no more sched_ext tasks) and reinstate fair_server if it
-	 * was previously detached because we were running in full mode.
+	 * Re-balance the dl_server bandwidth reservations: detach ext_server
+	 * (no more sched_ext tasks) and reinstate fair_server if it was
+	 * previously detached because we were running in full mode.
 	 *
 	 * Unlike the enable path, this runs on a recovery path that cannot
 	 * fail, so we use dl_server_swap_bw() to atomically free ext_server's
@@ -6180,8 +6473,6 @@ static void scx_root_disable(struct scx_sched *sch)
 	for_each_possible_cpu(cpu) {
 		struct rq *rq = cpu_rq(cpu);
 
-		scx_rq_clock_invalidate(rq);
-
 		scoped_guard(rq_lock_irqsave, rq) {
 			update_rq_clock(rq);
 			if (was_switched_all) {
@@ -6210,14 +6501,23 @@ static void scx_root_disable(struct scx_sched *sch)
 	if (sch->ops.exit)
 		SCX_CALL_OP(sch, exit, NULL, sch->exit_info);
 
+	/*
+	 * @sch's non-ops programs such as timers and tracers can fire after
+	 * ops.exit(). Now that exit is complete, stop scx_prog_sched() from
+	 * resolving to @sch and drain in-flight resolvers.
+	 */
+	WRITE_ONCE(sch->dead, true);
+	synchronize_rcu();
+
 	scx_unlink_sched(sch);
 
 	/*
-	 * scx_root clearing must be inside cpus_read_lock(). See
-	 * handle_hotplug().
+	 * scx_root clearing and cid table retirement must be inside
+	 * cpus_read_lock(). See handle_hotplug().
 	 */
 	cpus_read_lock();
 	RCU_INIT_POINTER(scx_root, NULL);
+	scx_cid_retire_tables();
 	cpus_read_unlock();
 
 	/*
@@ -6229,7 +6529,9 @@ static void scx_root_disable(struct scx_sched *sch)
 	if (sch->sub_kset)
 		kobject_del(&sch->sub_kset->kobj);
 #endif
-	kobject_del(&sch->kobj);
+	/* not added if enable failed before scx_sched_sysfs_add() */
+	if (sch->kobj.state_in_sysfs)
+		kobject_del(&sch->kobj);
 
 	free_kick_syncs();
 
@@ -6240,12 +6542,36 @@ done:
 	scx_bypass(sch, false);
 }
 
+/**
+ * scx_propagate_exit_irq_workfn - Claim SCX_EXIT_PARENT on the exiting subtree
+ * @irq_work: &scx_sched.propagate_exit_irq_work
+ *
+ * Queued by scx_claim_exit() after a non-PARENT claim. Claims SCX_EXIT_PARENT
+ * on each descendant, giving every one its own disable work - most of disabling
+ * is serialized but ops.exit() can take arbitrarily long and running them in
+ * separate helper kthreads parallelizes it. No recursion as only non-PARENT
+ * claims propagate.
+ */
+static void scx_propagate_exit_irq_workfn(struct irq_work *irq_work)
+{
+	struct scx_sched *sch = container_of(irq_work, struct scx_sched,
+					     propagate_exit_irq_work);
+	struct scx_sched *pos;
+
+	scoped_guard (raw_spinlock_irqsave, &scx_sched_lock) {
+		scx_for_each_descendant_pre(pos, sch)
+			scx_disable(pos, SCX_EXIT_PARENT);
+	}
+}
+
 /*
  * Claim the exit on @sch. The caller must ensure that the helper kthread work
  * is kicked before the current task can be preempted. Once exit_kind is
  * claimed, scx_error() can no longer trigger, so if the current task gets
  * preempted and the BPF scheduler fails to schedule it back, the helper work
  * will never be kicked and the whole system can wedge.
+ *
+ * Lock-free and safe to call from any context including NMI.
  */
 static bool scx_claim_exit(struct scx_sched *sch, enum scx_exit_kind kind)
 {
@@ -6259,35 +6585,31 @@ static bool scx_claim_exit(struct scx_sched *sch, enum scx_exit_kind kind)
 	if (!atomic_try_cmpxchg(&sch->exit_kind, &none, kind))
 		return false;
 
-	/*
-	 * Some CPUs may be trapped in the dispatch paths. Set the aborting
-	 * flag to break potential live-lock scenarios, ensuring we can
-	 * successfully reach scx_bypass().
-	 */
-	WRITE_ONCE(sch->aborting, true);
+	if (kind == SCX_EXIT_PARENT) {
+		/* an ancestor is already sweeping the subtree */
+		WRITE_ONCE(sch->aborting, true);
+	} else {
+		struct scx_sched *pos;
 
-	/*
-	 * Propagate exits to descendants immediately. Each has a dedicated
-	 * helper kthread and can run in parallel. While most of disabling is
-	 * serialized, running them in separate threads allows parallelizing
-	 * ops.exit(), which can take arbitrarily long prolonging bypass mode.
-	 *
-	 * To guarantee forward progress, this propagation must be in-line so
-	 * that ->aborting is synchronously asserted for all sub-scheds. The
-	 * propagation is also the interlocking point against sub-sched
-	 * attachment. See scx_link_sched().
-	 *
-	 * This doesn't cause recursions as propagation only takes place for
-	 * non-propagation exits.
-	 */
-	if (kind != SCX_EXIT_PARENT) {
-		scoped_guard (raw_spinlock_irqsave, &scx_sched_lock) {
-			struct scx_sched *pos;
+		/*
+		 * CPUs may be live-locked in the dispatch paths of @sch or its
+		 * descendants, which ->aborting breaks. Sweep the subtree
+		 * locklessly so that this works from NMI. smp_store_mb() orders
+		 * each node's ->aborting store before its children are walked -
+		 * either we see a racing scx_link_sched() on ->children or it
+		 * sees ->aborting.
+		 */
+		scoped_guard (rcu) {
 			scx_for_each_descendant_pre(pos, sch)
-				scx_disable(pos, SCX_EXIT_PARENT);
+				smp_store_mb(pos->aborting, true);
 		}
+
+		irq_work_queue(&sch->propagate_exit_irq_work);
 	}
 
+	/* fired after ->aborting is set so callbacks can't delay recovery */
+	trace_sched_ext_exit(sch, kind);
+
 	return true;
 }
 
@@ -6329,7 +6651,7 @@ static void scx_disable(struct scx_sched *sch, enum scx_exit_kind kind)
  * as a noop. Syncing the irq_work first is required to guarantee the
  * kthread work has been queued before waiting for it.
  */
-static void scx_flush_disable_work(struct scx_sched *sch)
+void scx_flush_disable_work(struct scx_sched *sch)
 {
 	int kind;
 
@@ -6349,7 +6671,7 @@ static void dump_newline(struct seq_buf *s)
 		seq_buf_putc(s, '\n');
 }
 
-static __printf(2, 3) void dump_line(struct seq_buf *s, const char *fmt, ...)
+__printf(2, 3) void scx_dump_line(struct seq_buf *s, const char *fmt, ...)
 {
 	va_list args;
 
@@ -6381,7 +6703,7 @@ static void dump_stack_trace(struct seq_buf *s, const char *prefix,
 	unsigned int i;
 
 	for (i = 0; i < len; i++)
-		dump_line(s, "%s%pS", prefix, (void *)bt[i]);
+		scx_dump_line(s, "%s%pS", prefix, (void *)bt[i]);
 }
 
 static void ops_dump_init(struct seq_buf *s, const char *prefix)
@@ -6431,7 +6753,7 @@ static void ops_dump_flush(void)
 		 */
 		c = *end;
 		*end = '\0';
-		dump_line(dd->s, "%s%s", dd->prefix, line);
+		scx_dump_line(dd->s, "%s%s", dd->prefix, line);
 		if (c == '\0')
 			break;
 
@@ -6475,21 +6797,19 @@ static void scx_dump_task(struct scx_sched *sch, struct seq_buf *s, struct scx_d
 			  (unsigned long long)p->scx.dsq->id);
 
 	dump_newline(s);
-	dump_line(s, " %c%c %s[%d] %s%s %+ldms",
-		  marker, task_state_to_char(p), p->comm, p->pid,
-		  own_marker, sch_id_buf,
-		  jiffies_delta_msecs(p->scx.runnable_at, dctx->at_jiffies));
-	dump_line(s, "      scx_state/flags=%u/0x%x dsq_flags=0x%x ops_state/qseq=%lu/%lu",
-		  scx_get_task_state(p) >> SCX_TASK_STATE_SHIFT,
-		  p->scx.flags & ~SCX_TASK_STATE_MASK,
-		  p->scx.dsq_flags, ops_state & SCX_OPSS_STATE_MASK,
-		  ops_state >> SCX_OPSS_QSEQ_SHIFT);
-	dump_line(s, "      sticky/holding_cpu=%d/%d dsq_id=%s",
-		  p->scx.sticky_cpu, p->scx.holding_cpu, dsq_id_buf);
-	dump_line(s, "      dsq_vtime=%llu slice=%llu weight=%u",
-		  p->scx.dsq_vtime, p->scx.slice, p->scx.weight);
-	dump_line(s, "      cpus=%*pb no_mig=%u", cpumask_pr_args(p->cpus_ptr),
-		  p->migration_disabled);
+	scx_dump_line(s, " %c%c %s[%d] %s%s %+ldms",
+		      marker, task_state_to_char(p), p->comm, p->pid, own_marker, sch_id_buf,
+		      jiffies_delta_msecs(p->scx.runnable_at, dctx->at_jiffies));
+	scx_dump_line(s, "      scx_state/flags=%u/0x%x dsq_flags=0x%x ops_state/qseq=%lu/%lu",
+		      scx_get_task_state(p) >> SCX_TASK_STATE_SHIFT,
+		      p->scx.flags & ~SCX_TASK_STATE_MASK, p->scx.dsq_flags,
+		      ops_state & SCX_OPSS_STATE_MASK, ops_state >> SCX_OPSS_QSEQ_SHIFT);
+	scx_dump_line(s, "      sticky/holding_cpu=%d/%d dsq_id=%s",
+		      p->scx.sticky_cpu, p->scx.holding_cpu, dsq_id_buf);
+	scx_dump_line(s, "      dsq_vtime=%llu slice=%llu weight=%u",
+		      p->scx.dsq_vtime, p->scx.slice, p->scx.weight);
+	scx_dump_line(s, "      cpus=%*pb no_mig=%u", cpumask_pr_args(p->cpus_ptr),
+		      p->migration_disabled);
 
 	if (SCX_HAS_OP(sch, dump_task)) {
 		ops_dump_init(s, "    ");
@@ -6511,6 +6831,7 @@ static void scx_dump_cpu(struct scx_sched *sch, struct seq_buf *s,
 			 bool dump_all_tasks)
 {
 	struct rq *rq = cpu_rq(cpu);
+	struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
 	struct rq_flags rf;
 	struct task_struct *p;
 	struct seq_buf ns;
@@ -6536,28 +6857,27 @@ static void scx_dump_cpu(struct scx_sched *sch, struct seq_buf *s,
 	seq_buf_init(&ns, buf, avail);
 
 	dump_newline(&ns);
-	dump_line(&ns, "CPU %-4d: nr_run=%u flags=0x%x cpu_rel=%d ops_qseq=%lu ksync=%lu",
-		  cpu, rq->scx.nr_running, rq->scx.flags,
-		  rq->scx.cpu_released, rq->scx.ops_qseq,
-		  rq->scx.kick_sync);
-	dump_line(&ns, "          curr=%s[%d] class=%ps",
-		  rq->curr->comm, rq->curr->pid,
-		  rq->curr->sched_class);
-	if (!cpumask_empty(rq->scx.cpus_to_kick))
-		dump_line(&ns, "  cpus_to_kick   : %*pb",
-			  cpumask_pr_args(rq->scx.cpus_to_kick));
-	if (!cpumask_empty(rq->scx.cpus_to_kick_if_idle))
-		dump_line(&ns, "  idle_to_kick   : %*pb",
-			  cpumask_pr_args(rq->scx.cpus_to_kick_if_idle));
-	if (!cpumask_empty(rq->scx.cpus_to_preempt))
-		dump_line(&ns, "  cpus_to_preempt: %*pb",
-			  cpumask_pr_args(rq->scx.cpus_to_preempt));
-	if (!cpumask_empty(rq->scx.cpus_to_wait))
-		dump_line(&ns, "  cpus_to_wait   : %*pb",
-			  cpumask_pr_args(rq->scx.cpus_to_wait));
+	scx_dump_line(&ns, "CPU %-4d: nr_run=%u flags=0x%x cpu_rel=%d ops_qseq=%lu ksync=%lu",
+		      cpu, rq->scx.nr_running, rq->scx.flags, rq->scx.cpu_released,
+		      rq->scx.ops_qseq, rq->scx.kick_sync);
+	scx_rescue_dump(&ns, rq);
+	scx_dump_line(&ns, "          curr=%s[%d] class=%ps",
+		      rq->curr->comm, rq->curr->pid, rq->curr->sched_class);
+	if (!cpumask_empty(pcpu->cpus_to_kick))
+		scx_dump_line(&ns, "  cpus_to_kick   : %*pb",
+			      cpumask_pr_args(pcpu->cpus_to_kick));
+	if (!cpumask_empty(pcpu->cpus_to_kick_if_idle))
+		scx_dump_line(&ns, "  idle_to_kick   : %*pb",
+			      cpumask_pr_args(pcpu->cpus_to_kick_if_idle));
+	if (!cpumask_empty(pcpu->cpus_to_preempt))
+		scx_dump_line(&ns, "  cpus_to_preempt: %*pb",
+			      cpumask_pr_args(pcpu->cpus_to_preempt));
+	if (!cpumask_empty(pcpu->cpus_to_wait))
+		scx_dump_line(&ns, "  cpus_to_wait   : %*pb",
+			      cpumask_pr_args(pcpu->cpus_to_wait));
 	if (!cpumask_empty(rq->scx.cpus_to_sync))
-		dump_line(&ns, "  cpus_to_sync   : %*pb",
-			  cpumask_pr_args(rq->scx.cpus_to_sync));
+		scx_dump_line(&ns, "  cpus_to_sync   : %*pb",
+			      cpumask_pr_args(rq->scx.cpus_to_sync));
 
 	used = seq_buf_used(&ns);
 	if (SCX_HAS_OP(sch, dump_cpu)) {
@@ -6625,25 +6945,25 @@ static void scx_dump_state(struct scx_sched *sch, struct scx_exit_info *ei,
 
 #ifdef CONFIG_EXT_SUB_SCHED
 	if (sch->level == 0)
-		dump_line(&s, "%s: root", sch->ops.name);
+		scx_dump_line(&s, "%s: root", sch->ops.name);
 	else
-		dump_line(&s, "%s: sub%d-%llu %s",
-			  sch->ops.name, sch->level, sch->ops.sub_cgroup_id,
-			  sch->cgrp_path);
+		scx_dump_line(&s, "%s: sub%d-%llu %s",
+			      sch->ops.name, sch->level, sch->ops.sub_cgroup_id,
+			      sch->cgrp_path);
 #endif
 	if (ei->kind == SCX_EXIT_NONE) {
-		dump_line(&s, "Debug dump triggered by %s", ei->reason);
+		scx_dump_line(&s, "Debug dump triggered by %s", ei->reason);
 	} else {
 		if (ei->exit_cpu >= 0)
-			dump_line(&s, "%s[%d] triggered exit kind %d on CPU %d:",
-				  current->comm, current->pid, ei->kind,
-				  ei->exit_cpu);
+			scx_dump_line(&s, "%s[%d] triggered exit kind %d on CPU %d:",
+				      current->comm, current->pid, ei->kind,
+				      ei->exit_cpu);
 		else
-			dump_line(&s, "%s[%d] triggered exit kind %d:",
-				  current->comm, current->pid, ei->kind);
-		dump_line(&s, "  %s (%s)", ei->reason, ei->msg);
+			scx_dump_line(&s, "%s[%d] triggered exit kind %d:",
+				      current->comm, current->pid, ei->kind);
+		scx_dump_line(&s, "  %s (%s)", ei->reason, ei->msg);
 		dump_newline(&s);
-		dump_line(&s, "Backtrace:");
+		scx_dump_line(&s, "Backtrace:");
 		dump_stack_trace(&s, "  ", ei->bt, ei->bt_len);
 	}
 
@@ -6654,38 +6974,37 @@ static void scx_dump_state(struct scx_sched *sch, struct scx_exit_info *ei,
 	}
 
 	dump_newline(&s);
-	dump_line(&s, "CPU states");
-	dump_line(&s, "----------");
+	scx_dump_line(&s, "CPU states");
+	scx_dump_line(&s, "----------");
 
 	/*
-	 * Dump the exit CPU first so it isn't lost to dump truncation, then
-	 * walk the rest in order, skipping the one already dumped.
+	 * Dump stalled CPUs first so they aren't lost to dump truncation, then
+	 * walk the rest in order. Fall back to exit_cpu if no stall mask set.
 	 */
-	if (ei->exit_cpu >= 0)
-		scx_dump_cpu(sch, &s, &dctx, ei->exit_cpu, dump_all_tasks);
-	for_each_possible_cpu(cpu) {
-		if (cpu != ei->exit_cpu)
+	if (!cpumask_empty(sch->stall_cpus)) {
+		for_each_cpu(cpu, sch->stall_cpus)
 			scx_dump_cpu(sch, &s, &dctx, cpu, dump_all_tasks);
+		for_each_possible_cpu(cpu) {
+			if (!cpumask_test_cpu(cpu, sch->stall_cpus))
+				scx_dump_cpu(sch, &s, &dctx, cpu, dump_all_tasks);
+		}
+	} else {
+		if (ei->exit_cpu >= 0)
+			scx_dump_cpu(sch, &s, &dctx, ei->exit_cpu, dump_all_tasks);
+		for_each_possible_cpu(cpu) {
+			if (cpu != ei->exit_cpu)
+				scx_dump_cpu(sch, &s, &dctx, cpu, dump_all_tasks);
+		}
 	}
 
 	dump_newline(&s);
-	dump_line(&s, "Event counters");
-	dump_line(&s, "--------------");
+	scx_dump_line(&s, "Event counters");
+	scx_dump_line(&s, "--------------");
 
 	scx_read_events(sch, &events);
-	scx_dump_event(s, &events, SCX_EV_SELECT_CPU_FALLBACK);
-	scx_dump_event(s, &events, SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE);
-	scx_dump_event(s, &events, SCX_EV_DISPATCH_KEEP_LAST);
-	scx_dump_event(s, &events, SCX_EV_ENQ_SKIP_EXITING);
-	scx_dump_event(s, &events, SCX_EV_ENQ_SKIP_MIGRATION_DISABLED);
-	scx_dump_event(s, &events, SCX_EV_REENQ_IMMED);
-	scx_dump_event(s, &events, SCX_EV_REENQ_LOCAL_REPEAT);
-	scx_dump_event(s, &events, SCX_EV_REFILL_SLICE_DFL);
-	scx_dump_event(s, &events, SCX_EV_BYPASS_DURATION);
-	scx_dump_event(s, &events, SCX_EV_BYPASS_DISPATCH);
-	scx_dump_event(s, &events, SCX_EV_BYPASS_ACTIVATE);
-	scx_dump_event(s, &events, SCX_EV_INSERT_NOT_OWNED);
-	scx_dump_event(s, &events, SCX_EV_SUB_BYPASS_DISPATCH);
+#define SCX_EVENT(name)	scx_dump_event(s, &events, name)
+	SCX_EVENTS_LIST(SCX_EVENT);
+#undef SCX_EVENT
 
 	if (seq_buf_has_overflowed(&s) && dump_len >= sizeof(trunc_marker))
 		memcpy(ei->dump + dump_len - sizeof(trunc_marker),
@@ -6703,24 +7022,21 @@ static void scx_disable_irq_workfn(struct irq_work *irq_work)
 	kthread_queue_work(sch->helper, &sch->disable_work);
 }
 
-bool scx_vexit(struct scx_sched *sch,
-	       enum scx_exit_kind kind, s64 exit_code, s32 exit_cpu,
-	       const char *fmt, va_list args)
+/* finish exit_info and kick the disable work, ei->msg must already be set */
+static void scx_finish_exit(struct scx_sched *sch, enum scx_exit_kind kind,
+			    s64 exit_code, s32 exit_cpu)
 {
 	struct scx_exit_info *ei = sch->exit_info;
 
-	guard(preempt)();
-
-	if (!scx_claim_exit(sch, kind))
-		return false;
-
 	ei->exit_code = exit_code;
 #ifdef CONFIG_STACKTRACE
-	if (kind >= SCX_EXIT_ERROR)
+	/*
+	 * stack_trace_save()'s NMI-safety is arch-dependent and undocumented.
+	 * Skip the backtrace when exiting from NMI.
+	 */
+	if (kind >= SCX_EXIT_ERROR && !in_nmi())
 		ei->bt_len = stack_trace_save(ei->bt, SCX_EXIT_BT_LEN, 1);
 #endif
-	vscnprintf(ei->msg, SCX_EXIT_MSG_LEN, fmt, args);
-
 	/*
 	 * Set ei->kind and ->reason for scx_dump_state(). They'll be set again
 	 * in scx_disable_workfn().
@@ -6730,6 +7046,22 @@ bool scx_vexit(struct scx_sched *sch,
 	ei->exit_cpu = exit_cpu;
 
 	irq_work_queue(&sch->disable_irq_work);
+}
+
+bool scx_vexit(struct scx_sched *sch,
+	       enum scx_exit_kind kind, s64 exit_code, s32 exit_cpu,
+	       const char *fmt, va_list args)
+{
+	struct scx_exit_info *ei = sch->exit_info;
+
+	guard(preempt)();
+
+	if (!scx_claim_exit(sch, kind))
+		return false;
+
+	vscnprintf(ei->msg, SCX_EXIT_MSG_LEN, fmt, args);
+
+	scx_finish_exit(sch, kind, exit_code, exit_cpu);
 	return true;
 }
 
@@ -6776,7 +7108,7 @@ static struct scx_sched_pnode *alloc_pnode(struct scx_sched *sch, int node)
 	if (!pnode)
 		return NULL;
 
-	if (init_dsq(&pnode->global_dsq, SCX_DSQ_GLOBAL, sch)) {
+	if (scx_init_dsq(&pnode->global_dsq, SCX_DSQ_GLOBAL, sch)) {
 		kfree(pnode);
 		return NULL;
 	}
@@ -6785,30 +7117,12 @@ static struct scx_sched_pnode *alloc_pnode(struct scx_sched *sch, int node)
 }
 
 /*
- * scx_enable() is offloaded to a dedicated system-wide RT kthread to avoid
- * starvation. During the READY -> ENABLED task switching loop, the calling
- * thread's sched_class gets switched from fair to ext. As fair has higher
- * priority than ext, the calling thread can be indefinitely starved under
- * fair-class saturation, leading to a system hang.
- */
-struct scx_enable_cmd {
-	struct kthread_work	work;
-	union {
-		struct sched_ext_ops		*ops;
-		struct sched_ext_ops_cid	*ops_cid;
-	};
-	bool			is_cid_type;
-	struct bpf_map		*arena_map;	/* arena ref to transfer to sch */
-	int			ret;
-};
-
-/*
  * Allocate and initialize a new scx_sched. @cgrp's reference is always
  * consumed whether the function succeeds or fails.
  */
-static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
-						 struct cgroup *cgrp,
-						 struct scx_sched *parent)
+struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
+					  struct cgroup *cgrp,
+					  struct scx_sched *parent)
 {
 	struct sched_ext_ops *ops = cmd->ops;
 	struct scx_sched *sch;
@@ -6855,7 +7169,7 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 	}
 
 	for_each_possible_cpu(cpu) {
-		ret = init_dsq(bypass_dsq(sch, cpu), SCX_DSQ_BYPASS, sch);
+		ret = scx_init_dsq(scx_bypass_dsq(sch, cpu), SCX_DSQ_BYPASS, sch);
 		if (ret) {
 			bypass_fail_cpu = cpu;
 			goto err_free_pcpu;
@@ -6865,8 +7179,20 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 	for_each_possible_cpu(cpu) {
 		struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
 
+		node = cpu_to_node(cpu);
 		pcpu->sch = sch;
 		INIT_LIST_HEAD(&pcpu->deferred_reenq_local.node);
+#ifdef CONFIG_EXT_SUB_SCHED
+		init_llist_node(&pcpu->ecaps_to_sync_node);
+#endif
+		INIT_LIST_HEAD(&pcpu->to_kick_node);
+		if (!zalloc_cpumask_var_node(&pcpu->cpus_to_kick, GFP_KERNEL, node) ||
+		    !zalloc_cpumask_var_node(&pcpu->cpus_to_kick_if_idle, GFP_KERNEL, node) ||
+		    !zalloc_cpumask_var_node(&pcpu->cpus_to_preempt, GFP_KERNEL, node) ||
+		    !zalloc_cpumask_var_node(&pcpu->cpus_to_wait, GFP_KERNEL, node)) {
+			ret = -ENOMEM;
+			goto err_free_pcpu;
+		}
 	}
 
 	sch->helper = kthread_run_worker(0, "sched_ext_helper");
@@ -6882,6 +7208,7 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 		       level * sizeof(parent->ancestors[0]));
 	sch->ancestors[level] = sch;
 	sch->level = level;
+	sch->id = atomic64_inc_return(&scx_sched_id_cursor);
 
 	if (ops->timeout_ms)
 		sch->watchdog_timeout = msecs_to_jiffies(ops->timeout_ms);
@@ -6891,6 +7218,7 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 	sch->slice_dfl = SCX_SLICE_DFL;
 	atomic_set(&sch->exit_kind, SCX_EXIT_NONE);
 	sch->disable_irq_work = IRQ_WORK_INIT_HARD(scx_disable_irq_workfn);
+	sch->propagate_exit_irq_work = IRQ_WORK_INIT_HARD(scx_propagate_exit_irq_workfn);
 	kthread_init_work(&sch->disable_work, scx_disable_workfn);
 	timer_setup(&sch->bypass_lb_timer, scx_bypass_lb_timerfn, 0);
 
@@ -6902,6 +7230,10 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 		ret = -ENOMEM;
 		goto err_free_lb_cpumask;
 	}
+	if (!zalloc_cpumask_var(&sch->stall_cpus, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_free_lb_resched_cpumask;
+	}
 	/*
 	 * Copy ops through the right union view. For cid-form the source is
 	 * struct sched_ext_ops_cid which lacks the trailing cpu_acquire/
@@ -6951,37 +7283,16 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 		 * disable. Released in scx_sched_free_rcu_work().
 		 */
 		kobject_get(&parent->kobj);
-		ret = kobject_init_and_add(&sch->kobj, &scx_ktype,
-					   &parent->sub_kset->kobj,
-					   "sub-%llu", cgroup_id(cgrp));
-	} else {
-		ret = kobject_init_and_add(&sch->kobj, &scx_ktype, NULL, "root");
-	}
-
-	if (ret < 0) {
-		RCU_INIT_POINTER(ops->priv, NULL);
-		kobject_put(&sch->kobj);
-		return ERR_PTR(ret);
-	}
-
-	if (ops->sub_attach) {
-		sch->sub_kset = kset_create_and_add("sub", NULL, &sch->kobj);
-		if (!sch->sub_kset) {
-			RCU_INIT_POINTER(ops->priv, NULL);
-			kobject_put(&sch->kobj);
-			return ERR_PTR(-ENOMEM);
-		}
-	}
-#else	/* CONFIG_EXT_SUB_SCHED */
-	ret = kobject_init_and_add(&sch->kobj, &scx_ktype, NULL, "root");
-	if (ret < 0) {
-		RCU_INIT_POINTER(ops->priv, NULL);
-		kobject_put(&sch->kobj);
-		return ERR_PTR(ret);
 	}
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
 	/*
+	 * Init the kobj but don't add to sysfs yet. The enable path calls
+	 * scx_sched_sysfs_add() once @sch's sysfs-visible state is initialized.
+	 */
+	kobject_init(&sch->kobj, &scx_ktype);
+
+	/*
 	 * Consume the arena_map ref bpf_scx_reg_cid() took. Defer to here so
 	 * earlier failure paths leave cmd->arena_map set and bpf_scx_reg_cid
 	 * drops the ref. After this point, sch owns the ref and any cleanup
@@ -6998,17 +7309,27 @@ static struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
 
 #ifdef CONFIG_EXT_SUB_SCHED
 err_free_lb_resched:
-	free_cpumask_var(sch->bypass_lb_resched_cpumask);
+	free_cpumask_var(sch->stall_cpus);
 #endif
+err_free_lb_resched_cpumask:
+	free_cpumask_var(sch->bypass_lb_resched_cpumask);
 err_free_lb_cpumask:
 	free_cpumask_var(sch->bypass_lb_donee_cpumask);
 err_stop_helper:
 	kthread_destroy_worker(sch->helper);
 err_free_pcpu:
 	for_each_possible_cpu(cpu) {
+		struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
+
+		free_cpumask_var(pcpu->cpus_to_kick);
+		free_cpumask_var(pcpu->cpus_to_kick_if_idle);
+		free_cpumask_var(pcpu->cpus_to_preempt);
+		free_cpumask_var(pcpu->cpus_to_wait);
+	}
+	for_each_possible_cpu(cpu) {
 		if (cpu == bypass_fail_cpu)
 			break;
-		exit_dsq(bypass_dsq(sch, cpu));
+		exit_dsq(scx_bypass_dsq(sch, cpu));
 	}
 	free_percpu(sch->pcpu);
 err_free_pnode:
@@ -7028,6 +7349,36 @@ err_put_cgrp:
 	return ERR_PTR(ret);
 }
 
+/*
+ * Add @sch's kobject to sysfs, and create its sub_kset if the scheduler
+ * implements ops.sub_attach. Called by the enable workfns once @sch's
+ * sysfs-visible state is initialized.
+ */
+int scx_sched_sysfs_add(struct scx_sched *sch)
+{
+#ifdef CONFIG_EXT_SUB_SCHED
+	struct scx_sched *parent = scx_parent(sch);
+	int ret;
+
+	if (parent)
+		ret = kobject_add(&sch->kobj, &parent->sub_kset->kobj,
+				  "sub-%llu", cgroup_id(sch_cgroup(sch)));
+	else
+		ret = kobject_add(&sch->kobj, NULL, "root");
+	if (ret < 0)
+		return ret;
+
+	if (sch->ops.sub_attach) {
+		sch->sub_kset = kset_create_and_add("sub", NULL, &sch->kobj);
+		if (!sch->sub_kset)
+			return -ENOMEM;
+	}
+	return 0;
+#else
+	return kobject_add(&sch->kobj, NULL, "root");
+#endif
+}
+
 static int check_hotplug_seq(struct scx_sched *sch,
 			      const struct sched_ext_ops *ops)
 {
@@ -7052,7 +7403,7 @@ static int check_hotplug_seq(struct scx_sched *sch,
 	return 0;
 }
 
-static int validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops)
+int scx_validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops)
 {
 	/*
 	 * It doesn't make sense to specify the SCX_OPS_ENQ_LAST flag if the
@@ -7069,7 +7420,7 @@ static int validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops)
 	 * enabled it.
 	 */
 	if ((ops->flags & SCX_OPS_TID_TO_TASK) && scx_parent(sch) &&
-	    !(scx_root->ops.flags & SCX_OPS_TID_TO_TASK)) {
+	    !(sch->ancestors[0]->ops.flags & SCX_OPS_TID_TO_TASK)) {
 		scx_error(sch, "SCX_OPS_TID_TO_TASK requires root scheduler to enable it");
 		return -EINVAL;
 	}
@@ -7180,6 +7531,9 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 		rq->scx.cpuperf_target = SCX_CPUPERF_ONE;
 	}
 
+	scx_discard_stale_ecaps_syncs();
+	scx_rescue_set_knobs(sch);
+
 	/*
 	 * Keep CPUs stable during enable so that the BPF scheduler can track
 	 * online CPUs by watching ->on/offline_cpu() after ->init().
@@ -7187,10 +7541,9 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 	cpus_read_lock();
 
 	/*
-	 * Build the cid mapping before publishing scx_root. The cid kfuncs
-	 * dereference the cid arrays unconditionally once scx_prog_sched()
-	 * returns non-NULL; the rcu_assign_pointer() below pairs with their
-	 * rcu_dereference() to make the populated arrays visible.
+	 * Build the cid mapping into a private under-construction set. It
+	 * becomes visible to readers only through scx_cid_publish_tables() once
+	 * ops.init_cids() has finalized the layout.
 	 */
 	ret = scx_cid_init(sch);
 	if (ret) {
@@ -7212,17 +7565,24 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 
 	scx_idle_enable(ops);
 
-	if (sch->ops.init) {
-		ret = SCX_CALL_OP_RET(sch, init, NULL);
+	/*
+	 * A cid-form scheduler finalizes its cid layout in ops.init_cids(),
+	 * which may call scx_bpf_cid_override(). Run it before the caps and
+	 * shard state are built so the final layout is in effect.
+	 */
+	if (sch->is_cid_type && sch->ops_cid.init_cids) {
+		ret = SCX_CALL_OP_RET(sch, init_cids, NULL);
 		if (ret) {
-			ret = ops_sanitize_err(sch, "init", ret);
+			ret = scx_ops_sanitize_err(sch, "init_cids", ret);
 			cpus_read_unlock();
-			scx_error(sch, "ops.init() failed (%d)", ret);
+			scx_error(sch, "ops.init_cids() failed (%d)", ret);
 			goto err_disable;
 		}
-		sch->exit_info->flags |= SCX_EFLAG_INITIALIZED;
 	}
 
+	/* the cid layout is final, expose it to readers */
+	scx_cid_publish_tables();
+
 	ret = scx_arena_pool_init(sch);
 	if (ret) {
 		cpus_read_unlock();
@@ -7235,6 +7595,32 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 		goto err_disable;
 	}
 
+	ret = scx_alloc_pshards(sch);
+	if (ret) {
+		cpus_read_unlock();
+		goto err_disable;
+	}
+
+	scx_init_root_caps(sch);
+
+	/* the cid caps and shards are live now, so ops.init() can query them */
+	if (sch->ops.init) {
+		ret = SCX_CALL_OP_RET(sch, init, NULL);
+		if (ret) {
+			ret = scx_ops_sanitize_err(sch, "init", ret);
+			cpus_read_unlock();
+			scx_error(sch, "ops.init() failed (%d)", ret);
+			goto err_disable;
+		}
+		sch->exit_info->flags |= SCX_EFLAG_INITIALIZED;
+	}
+
+	ret = scx_sched_sysfs_add(sch);
+	if (ret) {
+		cpus_read_unlock();
+		goto err_disable;
+	}
+
 	for (i = SCX_OPI_CPU_HOTPLUG_BEGIN; i < SCX_OPI_CPU_HOTPLUG_END; i++)
 		if (((void (**)(void))ops)[i])
 			set_bit(i, sch->has_op);
@@ -7248,7 +7634,7 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 
 	cpus_read_unlock();
 
-	ret = validate_ops(sch, ops);
+	ret = scx_validate_ops(sch, ops);
 	if (ret)
 		goto err_disable;
 
@@ -7324,6 +7710,9 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 	if (ret)
 		goto err_disable_unlock_all;
 
+	WARN_ON_ONCE(scx_cgroup_enabled);
+	scx_cgroup_enabled = true;
+
 	scx_task_iter_start(&sti, NULL);
 	while ((p = scx_task_iter_next_locked(&sti))) {
 		/*
@@ -7347,7 +7736,7 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 		scx_set_task_state(p, SCX_TASK_INIT_BEGIN);
 		scx_task_iter_unlock(&sti);
 
-		ret = __scx_init_task(sch, p, false);
+		ret = __scx_init_task(sch, p, NULL, false);
 
 		scx_task_iter_relock(&sti, p);
 
@@ -7414,7 +7803,7 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 			queue_flags |= DEQUEUE_CLASS;
 
 		scoped_guard (sched_change, p, queue_flags) {
-			p->scx.slice = READ_ONCE(sch->slice_dfl);
+			scx_set_task_slice(p, READ_ONCE(sch->slice_dfl));
 			p->sched_class = new_class;
 		}
 	}
@@ -7425,6 +7814,7 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 
 	if (!scx_tryset_enable_state(SCX_ENABLED, SCX_ENABLING)) {
 		WARN_ON_ONCE(atomic_read(&sch->exit_kind) == SCX_EXIT_NONE);
+		ret = -EBUSY;
 		goto err_disable;
 	}
 
@@ -7490,357 +7880,6 @@ err_disable:
 	cmd->ret = 0;
 }
 
-#ifdef CONFIG_EXT_SUB_SCHED
-/* verify that a scheduler can be attached to @cgrp and return the parent */
-static struct scx_sched *find_parent_sched(struct cgroup *cgrp)
-{
-	struct scx_sched *parent = cgrp->scx_sched;
-	struct scx_sched *pos;
-
-	lockdep_assert_held(&scx_sched_lock);
-
-	/* can't attach twice to the same cgroup */
-	if (parent->cgrp == cgrp)
-		return ERR_PTR(-EBUSY);
-
-	/* does $parent allow sub-scheds? */
-	if (!parent->ops.sub_attach)
-		return ERR_PTR(-EOPNOTSUPP);
-
-	/* can't insert between $parent and its exiting children */
-	list_for_each_entry(pos, &parent->children, sibling)
-		if (cgroup_is_descendant(pos->cgrp, cgrp))
-			return ERR_PTR(-EBUSY);
-
-	return parent;
-}
-
-static bool assert_task_ready_or_enabled(struct task_struct *p)
-{
-	u32 state = scx_get_task_state(p);
-
-	switch (state) {
-	case SCX_TASK_READY:
-	case SCX_TASK_ENABLED:
-		return true;
-	default:
-		WARN_ONCE(true, "sched_ext: Invalid task state %d for %s[%d] during enabling sub sched",
-			  state, p->comm, p->pid);
-		return false;
-	}
-}
-
-static void scx_sub_enable_workfn(struct kthread_work *work)
-{
-	struct scx_enable_cmd *cmd = container_of(work, struct scx_enable_cmd, work);
-	struct sched_ext_ops *ops = cmd->ops;
-	struct cgroup *cgrp;
-	struct scx_sched *parent, *sch;
-	struct scx_task_iter sti;
-	struct task_struct *p;
-	s32 i, ret;
-
-	mutex_lock(&scx_enable_mutex);
-
-	if (!scx_enabled()) {
-		ret = -ENODEV;
-		goto out_unlock;
-	}
-
-	/* See scx_root_enable_workfn() for the @ops->priv check. */
-	if (rcu_access_pointer(ops->priv)) {
-		ret = -EBUSY;
-		goto out_unlock;
-	}
-
-	cgrp = cgroup_get_from_id(ops->sub_cgroup_id);
-	if (IS_ERR(cgrp)) {
-		ret = PTR_ERR(cgrp);
-		goto out_unlock;
-	}
-
-	raw_spin_lock_irq(&scx_sched_lock);
-	parent = find_parent_sched(cgrp);
-	if (IS_ERR(parent)) {
-		raw_spin_unlock_irq(&scx_sched_lock);
-		ret = PTR_ERR(parent);
-		goto out_put_cgrp;
-	}
-	kobject_get(&parent->kobj);
-	raw_spin_unlock_irq(&scx_sched_lock);
-
-	/* scx_alloc_and_add_sched() consumes @cgrp whether it succeeds or not */
-	sch = scx_alloc_and_add_sched(cmd, cgrp, parent);
-	kobject_put(&parent->kobj);
-	if (IS_ERR(sch)) {
-		ret = PTR_ERR(sch);
-		goto out_unlock;
-	}
-
-	ret = scx_link_sched(sch);
-	if (ret)
-		goto err_disable;
-
-	if (sch->level >= SCX_SUB_MAX_DEPTH) {
-		scx_error(sch, "max nesting depth %d violated",
-			  SCX_SUB_MAX_DEPTH);
-		goto err_disable;
-	}
-
-	if (sch->ops.init) {
-		ret = SCX_CALL_OP_RET(sch, init, NULL);
-		if (ret) {
-			ret = ops_sanitize_err(sch, "init", ret);
-			scx_error(sch, "ops.init() failed (%d)", ret);
-			goto err_disable;
-		}
-		sch->exit_info->flags |= SCX_EFLAG_INITIALIZED;
-	}
-
-	ret = scx_arena_pool_init(sch);
-	if (ret)
-		goto err_disable;
-
-	ret = scx_set_cmask_scratch_alloc(sch);
-	if (ret)
-		goto err_disable;
-
-	if (validate_ops(sch, ops))
-		goto err_disable;
-
-	struct scx_sub_attach_args sub_attach_args = {
-		.ops = &sch->ops,
-		.cgroup_path = sch->cgrp_path,
-	};
-
-	ret = SCX_CALL_OP_RET(parent, sub_attach, NULL,
-			      &sub_attach_args);
-	if (ret) {
-		ret = ops_sanitize_err(sch, "sub_attach", ret);
-		scx_error(sch, "parent rejected (%d)", ret);
-		goto err_disable;
-	}
-	sch->sub_attached = true;
-
-	scx_bypass(sch, true);
-
-	for (i = SCX_OPI_BEGIN; i < SCX_OPI_END; i++)
-		if (((void (**)(void))ops)[i])
-			set_bit(i, sch->has_op);
-
-	percpu_down_write(&scx_fork_rwsem);
-	scx_cgroup_lock();
-
-	/*
-	 * Set cgroup->scx_sched's and check CSS_ONLINE. Either we see
-	 * !CSS_ONLINE or scx_cgroup_lifetime_notify() sees and shoots us down.
-	 */
-	set_cgroup_sched(sch_cgroup(sch), sch);
-	if (!(cgrp->self.flags & CSS_ONLINE)) {
-		scx_error(sch, "cgroup is not online");
-		goto err_unlock_and_disable;
-	}
-
-	/*
-	 * Initialize tasks for the new child $sch without exiting them for
-	 * $parent so that the tasks can always be reverted back to $parent
-	 * sched on child init failure.
-	 */
-	WARN_ON_ONCE(scx_enabling_sub_sched);
-	scx_enabling_sub_sched = sch;
-
-	scx_task_iter_start(&sti, sch->cgrp);
-	while ((p = scx_task_iter_next_locked(&sti))) {
-		struct rq *rq;
-		struct rq_flags rf;
-
-		/*
-		 * Task iteration may visit the same task twice when racing
-		 * against exiting. Use %SCX_TASK_SUB_INIT to mark tasks which
-		 * finished __scx_init_task() and skip if set.
-		 *
-		 * A task may exit and get freed between __scx_init_task()
-		 * completion and scx_enable_task(). In such cases,
-		 * scx_disable_and_exit_task() must exit the task for both the
-		 * parent and child scheds.
-		 */
-		if (p->scx.flags & SCX_TASK_SUB_INIT)
-			continue;
-
-		/* @p is pinned by the iter; see scx_sub_disable() */
-		get_task_struct(p);
-
-		if (!assert_task_ready_or_enabled(p)) {
-			ret = -EINVAL;
-			goto abort;
-		}
-
-		scx_task_iter_unlock(&sti);
-
-		/*
-		 * As $p is still on $parent, it can't be transitioned to INIT.
-		 * Let's worry about task state later. Use __scx_init_task().
-		 */
-		ret = __scx_init_task(sch, p, false);
-		if (ret)
-			goto abort;
-
-		rq = task_rq_lock(p, &rf);
-
-		if (scx_get_task_state(p) == SCX_TASK_DEAD) {
-			/*
-			 * sched_ext_dead() raced us between __scx_init_task()
-			 * and this rq lock and ran exit_task() on $parent (the
-			 * sched @p was on at that point), not on @sch. @sch's
-			 * just-completed init is owed an exit_task() and we
-			 * issue it here.
-			 */
-			scx_sub_init_cancel_task(sch, p);
-			task_rq_unlock(rq, p, &rf);
-			put_task_struct(p);
-			continue;
-		}
-
-		p->scx.flags |= SCX_TASK_SUB_INIT;
-		task_rq_unlock(rq, p, &rf);
-
-		put_task_struct(p);
-	}
-	scx_task_iter_stop(&sti);
-
-	/*
-	 * All tasks are prepped. Disable/exit tasks for $parent and enable for
-	 * the new @sch.
-	 */
-	scx_task_iter_start(&sti, sch->cgrp);
-	while ((p = scx_task_iter_next_locked(&sti))) {
-		/*
-		 * Use clearing of %SCX_TASK_SUB_INIT to detect and skip
-		 * duplicate iterations.
-		 */
-		if (!(p->scx.flags & SCX_TASK_SUB_INIT))
-			continue;
-
-		scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
-			/*
-			 * $p must be either READY or ENABLED. If ENABLED,
-			 * __scx_disabled_and_exit_task() first disables and
-			 * makes it READY. However, after exiting $p, it will
-			 * leave $p as READY.
-			 */
-			assert_task_ready_or_enabled(p);
-			__scx_disable_and_exit_task(parent, p);
-
-			/*
-			 * $p is now only initialized for @sch and READY, which
-			 * is what we want. Assign it to @sch and, if it's on
-			 * the ext class, enable. A non-ext task, possible under
-			 * an %SCX_OPS_SWITCH_PARTIAL root, stays READY and is
-			 * enabled by switching_to_scx() if it switches over.
-			 */
-			scx_set_task_sched(p, sch);
-			if (p->sched_class == &ext_sched_class)
-				scx_enable_task(sch, p);
-
-			p->scx.flags &= ~SCX_TASK_SUB_INIT;
-		}
-	}
-	scx_task_iter_stop(&sti);
-
-	scx_enabling_sub_sched = NULL;
-
-	scx_cgroup_unlock();
-	percpu_up_write(&scx_fork_rwsem);
-
-	scx_bypass(sch, false);
-
-	pr_info("sched_ext: BPF sub-scheduler \"%s\" enabled\n", sch->ops.name);
-	kobject_uevent(&sch->kobj, KOBJ_ADD);
-	ret = 0;
-	goto out_unlock;
-
-out_put_cgrp:
-	cgroup_put(cgrp);
-out_unlock:
-	mutex_unlock(&scx_enable_mutex);
-	cmd->ret = ret;
-	return;
-
-abort:
-	put_task_struct(p);
-	scx_task_iter_stop(&sti);
-
-	/*
-	 * Undo __scx_init_task() for tasks we marked. scx_enable_task() never
-	 * ran for @sch on them, so calling scx_disable_task() here would invoke
-	 * ops.disable() without a matching ops.enable(). scx_enabling_sub_sched
-	 * must stay set until SUB_INIT is cleared from every marked task -
-	 * scx_disable_and_exit_task() reads it when a task exits concurrently.
-	 */
-	scx_task_iter_start(&sti, sch->cgrp);
-	while ((p = scx_task_iter_next_locked(&sti))) {
-		if (p->scx.flags & SCX_TASK_SUB_INIT) {
-			scx_sub_init_cancel_task(sch, p);
-			p->scx.flags &= ~SCX_TASK_SUB_INIT;
-		}
-	}
-	scx_task_iter_stop(&sti);
-	scx_enabling_sub_sched = NULL;
-err_unlock_and_disable:
-	/* we'll soon enter disable path, keep bypass on */
-	scx_cgroup_unlock();
-	percpu_up_write(&scx_fork_rwsem);
-err_disable:
-	mutex_unlock(&scx_enable_mutex);
-	/*
-	 * Some enable failures only return an errno (e.g. -ENOMEM from an
-	 * allocation) without calling scx_error(). Record it so
-	 * scx_flush_disable_work() runs the disable and ops.exit() fires.
-	 */
-	scx_error(sch, "scx_sub_enable() failed (%d)", ret);
-	scx_flush_disable_work(sch);
-	cmd->ret = 0;
-}
-
-static s32 scx_cgroup_lifetime_notify(struct notifier_block *nb,
-				      unsigned long action, void *data)
-{
-	struct cgroup *cgrp = data;
-	struct cgroup *parent = cgroup_parent(cgrp);
-
-	if (!cgroup_on_dfl(cgrp))
-		return NOTIFY_OK;
-
-	switch (action) {
-	case CGROUP_LIFETIME_ONLINE:
-		/* inherit ->scx_sched from $parent */
-		if (parent)
-			rcu_assign_pointer(cgrp->scx_sched, parent->scx_sched);
-		break;
-	case CGROUP_LIFETIME_OFFLINE:
-		/* if there is a sched attached, shoot it down */
-		if (cgrp->scx_sched && cgrp->scx_sched->cgrp == cgrp)
-			scx_exit(cgrp->scx_sched, SCX_EXIT_UNREG_KERN,
-				 SCX_ECODE_RSN_CGROUP_OFFLINE,
-				 "cgroup %llu going offline", cgroup_id(cgrp));
-		break;
-	}
-
-	return NOTIFY_OK;
-}
-
-static struct notifier_block scx_cgroup_lifetime_nb = {
-	.notifier_call = scx_cgroup_lifetime_notify,
-};
-
-static s32 __init scx_cgroup_lifetime_notifier_init(void)
-{
-	return blocking_notifier_chain_register(&cgroup_lifetime_notifier,
-						&scx_cgroup_lifetime_nb);
-}
-core_initcall(scx_cgroup_lifetime_notifier_init);
-#endif	/* CONFIG_EXT_SUB_SCHED */
-
 static s32 scx_enable(struct scx_enable_cmd *cmd, struct bpf_link *link)
 {
 	static struct kthread_worker *helper;
@@ -7903,6 +7942,21 @@ static bool bpf_scx_is_valid_access(int off, int size,
 	return btf_ctx_access(off, size, type, prog, info);
 }
 
+/* common to both forms: only scx.disallow is writable */
+static int bpf_scx_btf_struct_access_common(const struct bpf_reg_state *reg,
+					    int off, int size)
+{
+	const struct btf_type *t;
+
+	t = btf_type_by_id(reg->btf, reg->btf_id);
+	if (t == task_struct_type &&
+	    off >= offsetof(struct task_struct, scx.disallow) &&
+	    off + size <= offsetofend(struct task_struct, scx.disallow))
+		return SCALAR_VALUE;
+
+	return -EACCES;
+}
+
 static int bpf_scx_btf_struct_access(struct bpf_verifier_log *log,
 				     const struct bpf_reg_state *reg, int off,
 				     int size)
@@ -7911,23 +7965,22 @@ static int bpf_scx_btf_struct_access(struct bpf_verifier_log *log,
 
 	t = btf_type_by_id(reg->btf, reg->btf_id);
 	if (t == task_struct_type) {
-		/*
-		 * COMPAT: Will be removed in v6.23.
-		 */
 		if ((off >= offsetof(struct task_struct, scx.slice) &&
 		     off + size <= offsetofend(struct task_struct, scx.slice)) ||
 		    (off >= offsetof(struct task_struct, scx.dsq_vtime) &&
-		     off + size <= offsetofend(struct task_struct, scx.dsq_vtime))) {
-			pr_warn_ratelimited("sched_ext: Writing directly to p->scx.slice/dsq_vtime is deprecated, use scx_bpf_task_set_slice/dsq_vtime()\n");
-			return SCALAR_VALUE;
-		}
-
-		if (off >= offsetof(struct task_struct, scx.disallow) &&
-		    off + size <= offsetofend(struct task_struct, scx.disallow))
+		     off + size <= offsetofend(struct task_struct, scx.dsq_vtime)))
 			return SCALAR_VALUE;
 	}
 
-	return -EACCES;
+	return bpf_scx_btf_struct_access_common(reg, off, size);
+}
+
+/* cid-form rejects direct slice and dsq_vtime writes in favor of the kfuncs */
+static int bpf_scx_cid_btf_struct_access(struct bpf_verifier_log *log,
+					 const struct bpf_reg_state *reg, int off,
+					 int size)
+{
+	return bpf_scx_btf_struct_access_common(reg, off, size);
 }
 
 static const struct bpf_verifier_ops bpf_scx_verifier_ops = {
@@ -7936,6 +7989,12 @@ static const struct bpf_verifier_ops bpf_scx_verifier_ops = {
 	.btf_struct_access = bpf_scx_btf_struct_access,
 };
 
+static const struct bpf_verifier_ops bpf_scx_cid_verifier_ops = {
+	.get_func_proto = bpf_base_func_proto,
+	.is_valid_access = bpf_scx_is_valid_access,
+	.btf_struct_access = bpf_scx_cid_btf_struct_access,
+};
+
 static int bpf_scx_init_member(const struct btf_type *t,
 			       const struct btf_member *member,
 			       void *kdata, const void *udata)
@@ -7977,6 +8036,27 @@ static int bpf_scx_init_member(const struct btf_type *t,
 	case offsetof(struct sched_ext_ops, hotplug_seq):
 		ops->hotplug_seq = *(u64 *)(udata + moff);
 		return 1;
+	case offsetof(struct sched_ext_ops, cid_shard_size):
+		ops->cid_shard_size = *(u32 *)(udata + moff);
+		return 1;
+	case offsetof(struct sched_ext_ops, rescue_bandwidth_ppt): {
+		u32 bw_ppt = *(u32 *)(udata + moff);
+
+		if (bw_ppt > SCX_RESCUE_MAX_BW_PPT && bw_ppt != SCX_RESCUE_DISABLE)
+			return -E2BIG;
+		ops->rescue_bandwidth_ppt = bw_ppt;
+		return 1;
+	}
+	case offsetof(struct sched_ext_ops, rescue_quantum_us): {
+		u32 quantum_us = *(u32 *)(udata + moff);
+
+		if (quantum_us > SCX_RESCUE_MAX_QUANTUM_US)
+			return -E2BIG;
+		if (quantum_us && quantum_us < SCX_RESCUE_MIN_QUANTUM_US)
+			return -EINVAL;
+		ops->rescue_quantum_us = quantum_us;
+		return 1;
+	}
 #ifdef CONFIG_EXT_SUB_SCHED
 	case offsetof(struct sched_ext_ops, sub_cgroup_id):
 		ops->sub_cgroup_id = *(u64 *)(udata + moff);
@@ -7987,20 +8067,6 @@ static int bpf_scx_init_member(const struct btf_type *t,
 	return 0;
 }
 
-#ifdef CONFIG_EXT_SUB_SCHED
-static void scx_pstack_recursion_on_dispatch(struct bpf_prog *prog)
-{
-	struct scx_sched *sch;
-
-	guard(rcu)();
-	sch = scx_prog_sched(prog->aux);
-	if (unlikely(!sch))
-		return;
-
-	scx_error(sch, "dispatch recursion detected");
-}
-#endif	/* CONFIG_EXT_SUB_SCHED */
-
 static int bpf_scx_check_member(const struct btf_type *t,
 				const struct btf_member *member,
 				const struct bpf_prog *prog)
@@ -8016,6 +8082,7 @@ static int bpf_scx_check_member(const struct btf_type *t,
 #endif
 	case offsetof(struct sched_ext_ops, cpu_online):
 	case offsetof(struct sched_ext_ops, cpu_offline):
+	case offsetof(struct sched_ext_ops, init_cids):
 	case offsetof(struct sched_ext_ops, init):
 	case offsetof(struct sched_ext_ops, exit):
 	case offsetof(struct sched_ext_ops, sub_attach):
@@ -8039,6 +8106,11 @@ static int bpf_scx_check_member(const struct btf_type *t,
 	case offsetof(struct sched_ext_ops, dispatch):
 		prog->aux->priv_stack_requested = true;
 		prog->aux->recursion_detected = scx_pstack_recursion_on_dispatch;
+		break;
+	case offsetof(struct sched_ext_ops, sub_caps_updated):
+		prog->aux->priv_stack_requested = true;
+		prog->aux->recursion_detected = scx_pstack_recursion_on_caps_updated;
+		break;
 	}
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
@@ -8174,6 +8246,7 @@ static s32 sched_ext_ops__sub_attach(struct scx_sub_attach_args *args) { return
 static void sched_ext_ops__sub_detach(struct scx_sub_detach_args *args) {}
 static void sched_ext_ops__cpu_online(s32 cpu) {}
 static void sched_ext_ops__cpu_offline(s32 cpu) {}
+static s32 sched_ext_ops__init_cids(void) { return -EINVAL; }
 static s32 sched_ext_ops__init(void) { return -EINVAL; }
 static void sched_ext_ops__exit(struct scx_exit_info *info) {}
 static void sched_ext_ops__dump(struct scx_dump_ctx *ctx) {}
@@ -8215,6 +8288,7 @@ static struct sched_ext_ops __bpf_ops_sched_ext_ops = {
 	.sub_detach		= sched_ext_ops__sub_detach,
 	.cpu_online		= sched_ext_ops__cpu_online,
 	.cpu_offline		= sched_ext_ops__cpu_offline,
+	.init_cids		= sched_ext_ops__init_cids,
 	.init			= sched_ext_ops__init,
 	.exit			= sched_ext_ops__exit,
 	.dump			= sched_ext_ops__dump,
@@ -8238,11 +8312,13 @@ static struct bpf_struct_ops bpf_sched_ext_ops = {
 
 /*
  * cid-form cfi stubs. Stubs whose signatures match the cpu-form (param types
- * identical, only param names differ across structs) are reused; only
- * set_cmask needs a fresh stub since the second argument type differs.
+ * identical, only param names differ across structs) are reused. Some need
+ * fresh stubs, set_cmask due to an argument type difference and the sub-sched
+ * notifiers because no cpu-form stub exists to reuse.
  */
-static void sched_ext_ops_cid__set_cmask(struct task_struct *p,
-					 const struct scx_cmask *cmask) {}
+static void sched_ext_ops_cid__set_cmask(struct task_struct *p, const struct scx_cmask *cmask__arena) {}
+static void sched_ext_ops__sub_caps_updated(const struct scx_cmask *cmask__arena, u64 caps) {}
+static void sched_ext_ops__sub_ecaps_updated(s32 cid, u64 before, u64 after) {}
 
 static struct sched_ext_ops_cid __bpf_ops_sched_ext_ops_cid = {
 	.select_cid		= sched_ext_ops__select_cpu,
@@ -8264,19 +8340,22 @@ static struct sched_ext_ops_cid __bpf_ops_sched_ext_ops_cid = {
 	.enable			= sched_ext_ops__enable,
 	.disable		= sched_ext_ops__disable,
 #ifdef CONFIG_EXT_GROUP_SCHED
-	.cgroup_init		= sched_ext_ops__cgroup_init,
-	.cgroup_exit		= sched_ext_ops__cgroup_exit,
-	.cgroup_prep_move	= sched_ext_ops__cgroup_prep_move,
-	.cgroup_move		= sched_ext_ops__cgroup_move,
-	.cgroup_cancel_move	= sched_ext_ops__cgroup_cancel_move,
-	.cgroup_set_weight	= sched_ext_ops__cgroup_set_weight,
-	.cgroup_set_bandwidth	= sched_ext_ops__cgroup_set_bandwidth,
-	.cgroup_set_idle	= sched_ext_ops__cgroup_set_idle,
+	.cpuctl_init		= sched_ext_ops__cgroup_init,
+	.cpuctl_exit		= sched_ext_ops__cgroup_exit,
+	.cpuctl_prep_move	= sched_ext_ops__cgroup_prep_move,
+	.cpuctl_move		= sched_ext_ops__cgroup_move,
+	.cpuctl_cancel_move	= sched_ext_ops__cgroup_cancel_move,
+	.cpuctl_set_weight	= sched_ext_ops__cgroup_set_weight,
+	.cpuctl_set_bandwidth	= sched_ext_ops__cgroup_set_bandwidth,
+	.cpuctl_set_idle	= sched_ext_ops__cgroup_set_idle,
 #endif
 	.sub_attach		= sched_ext_ops__sub_attach,
 	.sub_detach		= sched_ext_ops__sub_detach,
+	.sub_caps_updated	= sched_ext_ops__sub_caps_updated,
+	.sub_ecaps_updated	= sched_ext_ops__sub_ecaps_updated,
 	.cid_online		= sched_ext_ops__cpu_online,
 	.cid_offline		= sched_ext_ops__cpu_offline,
+	.init_cids		= sched_ext_ops__init_cids,
 	.init			= sched_ext_ops__init,
 	.exit			= sched_ext_ops__exit,
 	.dump			= sched_ext_ops__dump,
@@ -8290,7 +8369,7 @@ static struct sched_ext_ops_cid __bpf_ops_sched_ext_ops_cid = {
  * verified to match by the BUILD_BUG_ON checks in scx_init().
  */
 static struct bpf_struct_ops bpf_sched_ext_ops_cid = {
-	.verifier_ops = &bpf_scx_verifier_ops,
+	.verifier_ops = &bpf_scx_cid_verifier_ops,
 	.reg = bpf_scx_reg_cid,
 	.unreg = bpf_scx_unreg,
 	.check_member = bpf_scx_check_member,
@@ -8354,21 +8433,23 @@ static bool can_skip_idle_kick(struct rq *rq)
 	 * We can skip idle kicking if @rq is going to go through at least one
 	 * full SCX scheduling cycle before going idle. Just checking whether
 	 * curr is not idle is insufficient because we could be racing
-	 * balance_one() trying to pull the next task from a remote rq, which
+	 * dispatch_one() trying to pull the next task from a remote rq, which
 	 * may fail, and @rq may become idle afterwards.
 	 *
 	 * The race window is small and we don't and can't guarantee that @rq is
 	 * only kicked while idle anyway. Skip only when sure.
 	 */
-	return !is_idle_task(rq->curr) && !(rq->scx.flags & SCX_RQ_IN_BALANCE);
+	return !is_idle_task(rq->curr) && !(rq->scx.flags & SCX_RQ_IN_DISPATCH);
 }
 
-static bool kick_one_cpu(s32 cpu, struct rq *this_rq, unsigned long *ksyncs)
+static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_rq,
+			 unsigned long *ksyncs)
 {
 	struct rq *rq = cpu_rq(cpu);
 	struct scx_rq *this_scx = &this_rq->scx;
 	const struct sched_class *cur_class;
 	bool should_wait = false;
+	bool kickable;
 	unsigned long flags;
 
 	raw_spin_rq_lock_irqsave(rq, flags);
@@ -8378,47 +8459,68 @@ static bool kick_one_cpu(s32 cpu, struct rq *this_rq, unsigned long *ksyncs)
 	 * During CPU hotplug, a CPU may depend on kicking itself to make
 	 * forward progress. Allow kicking self regardless of online state. If
 	 * @cpu is running a higher class task, we have no control over @cpu.
-	 * Skip kicking.
-	 */
-	if ((cpu_online(cpu) || cpu == cpu_of(this_rq)) &&
-	    !sched_class_above(cur_class, &ext_sched_class)) {
-		if (cpumask_test_cpu(cpu, this_scx->cpus_to_preempt)) {
-			if (cur_class == &ext_sched_class)
-				rq->curr->scx.slice = 0;
-			cpumask_clear_cpu(cpu, this_scx->cpus_to_preempt);
+	 * Skip kicking. A sub-sched lacking baseline access on @cid has no
+	 * business forcing a reschedule there - skip. This is the authoritative
+	 * cap check: ecaps is read here under @rq's lock.
+	 */
+	kickable = (cpu_online(cpu) || cpu == cpu_of(this_rq)) &&
+		   !sched_class_above(cur_class, &ext_sched_class);
+
+	if (kickable && !scx_missing_caps(pcpu->sch, cpu, SCX_CAP_BASE)) {
+		if (cpumask_test_cpu(cpu, pcpu->cpus_to_preempt)) {
+			if (cur_class == &ext_sched_class) {
+				u64 caps = scx_caps_for_preempt(pcpu->sch, rq, 0);
+
+				if (unlikely(scx_missing_caps(pcpu->sch, cpu, caps)))
+					__scx_add_event(pcpu->sch, SCX_EV_SUB_PREEMPT_DENIED, 1);
+				else if (unlikely(!scx_set_task_slice(rq->curr, 0)))
+					__scx_add_event(pcpu->sch, SCX_EV_SLICE_DENIED, 1);
+			}
+			cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt);
 		}
 
-		if (cpumask_test_cpu(cpu, this_scx->cpus_to_wait)) {
+		if (cpumask_test_cpu(cpu, pcpu->cpus_to_wait)) {
 			if (cur_class == &ext_sched_class) {
 				cpumask_set_cpu(cpu, this_scx->cpus_to_sync);
 				ksyncs[cpu] = rq->scx.kick_sync;
 				should_wait = true;
 			}
-			cpumask_clear_cpu(cpu, this_scx->cpus_to_wait);
+			cpumask_clear_cpu(cpu, pcpu->cpus_to_wait);
 		}
 
 		resched_curr(rq);
 	} else {
-		cpumask_clear_cpu(cpu, this_scx->cpus_to_preempt);
-		cpumask_clear_cpu(cpu, this_scx->cpus_to_wait);
+		/* a kickable cpu was skipped solely for the missing caps */
+		if (kickable)
+			__scx_add_event(pcpu->sch, SCX_EV_SUB_KICK_DENIED, 1);
+		cpumask_clear_cpu(cpu, pcpu->cpus_to_preempt);
+		cpumask_clear_cpu(cpu, pcpu->cpus_to_wait);
 	}
 
+	scx_rq_lock_drop(rq);
 	raw_spin_rq_unlock_irqrestore(rq, flags);
 
 	return should_wait;
 }
 
-static void kick_one_cpu_if_idle(s32 cpu, struct rq *this_rq)
+static void kick_one_cpu_if_idle(s32 cpu, struct scx_sched_pcpu *pcpu,
+				 struct rq *this_rq)
 {
 	struct rq *rq = cpu_rq(cpu);
 	unsigned long flags;
 
 	raw_spin_rq_lock_irqsave(rq, flags);
 
+	/* idle kicks need baseline access too, see kick_one_cpu() */
 	if (!can_skip_idle_kick(rq) &&
-	    (cpu_online(cpu) || cpu == cpu_of(this_rq)))
-		resched_curr(rq);
+	    (cpu_online(cpu) || cpu == cpu_of(this_rq))) {
+		if (likely(!scx_missing_caps(pcpu->sch, cpu, SCX_CAP_BASE)))
+			resched_curr(rq);
+		else
+			__scx_add_event(pcpu->sch, SCX_EV_SUB_KICK_DENIED, 1);
+	}
 
+	scx_rq_lock_drop(rq);
 	raw_spin_rq_unlock_irqrestore(rq, flags);
 }
 
@@ -8427,6 +8529,7 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
 	struct rq *this_rq = this_rq();
 	struct scx_rq *this_scx = &this_rq->scx;
 	struct scx_kick_syncs __rcu *ksyncs_pcpu = __this_cpu_read(scx_kick_syncs);
+	struct scx_sched_pcpu *pcpu, *tmp;
 	bool should_wait = false;
 	unsigned long *ksyncs;
 	s32 cpu;
@@ -8437,15 +8540,24 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
 
 	ksyncs = rcu_dereference_bh(ksyncs_pcpu)->syncs;
 
-	for_each_cpu(cpu, this_scx->cpus_to_kick) {
-		should_wait |= kick_one_cpu(cpu, this_rq, ksyncs);
-		cpumask_clear_cpu(cpu, this_scx->cpus_to_kick);
-		cpumask_clear_cpu(cpu, this_scx->cpus_to_kick_if_idle);
-	}
+	/*
+	 * Walk scheds with pending kicks on this cpu. scx_kick_cpu() adds to
+	 * the list under local_irq_save() and only this irq_work consumes it.
+	 * A plain list without locking is sufficient.
+	 */
+	list_for_each_entry_safe(pcpu, tmp, &this_scx->sched_pcpus_to_kick, to_kick_node) {
+		list_del_init(&pcpu->to_kick_node);
+
+		for_each_cpu(cpu, pcpu->cpus_to_kick) {
+			should_wait |= kick_one_cpu(cpu, pcpu, this_rq, ksyncs);
+			cpumask_clear_cpu(cpu, pcpu->cpus_to_kick);
+			cpumask_clear_cpu(cpu, pcpu->cpus_to_kick_if_idle);
+		}
 
-	for_each_cpu(cpu, this_scx->cpus_to_kick_if_idle) {
-		kick_one_cpu_if_idle(cpu, this_rq);
-		cpumask_clear_cpu(cpu, this_scx->cpus_to_kick_if_idle);
+		for_each_cpu(cpu, pcpu->cpus_to_kick_if_idle) {
+			kick_one_cpu_if_idle(cpu, pcpu, this_rq);
+			cpumask_clear_cpu(cpu, pcpu->cpus_to_kick_if_idle);
+		}
 	}
 
 	/*
@@ -8456,6 +8568,7 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
 		raw_spin_rq_lock(this_rq);
 		this_scx->kick_sync_pending = true;
 		resched_curr(this_rq);
+		scx_rq_lock_drop(this_rq);
 		raw_spin_rq_unlock(this_rq);
 	}
 }
@@ -8565,16 +8678,17 @@ void __init init_sched_ext_class(void)
 		int  n = cpu_to_node(cpu);
 
 		/* local_dsq's sch will be set during scx_root_enable() */
-		BUG_ON(init_dsq(&rq->scx.local_dsq, SCX_DSQ_LOCAL, NULL));
+		BUG_ON(scx_init_dsq(&rq->scx.local_dsq, SCX_DSQ_LOCAL, NULL));
+#ifdef CONFIG_EXT_SUB_SCHED
+		BUG_ON(scx_init_dsq(&rq->scx.reject_dsq, SCX_DSQ_REJECT, NULL));
+		scx_rescue_init(rq);
+#endif
 
 		INIT_LIST_HEAD(&rq->scx.runnable_list);
 		INIT_LIST_HEAD(&rq->scx.ddsp_deferred_locals);
 
-		BUG_ON(!zalloc_cpumask_var_node(&rq->scx.cpus_to_kick, GFP_KERNEL, n));
-		BUG_ON(!zalloc_cpumask_var_node(&rq->scx.cpus_to_kick_if_idle, GFP_KERNEL, n));
-		BUG_ON(!zalloc_cpumask_var_node(&rq->scx.cpus_to_preempt, GFP_KERNEL, n));
-		BUG_ON(!zalloc_cpumask_var_node(&rq->scx.cpus_to_wait, GFP_KERNEL, n));
 		BUG_ON(!zalloc_cpumask_var_node(&rq->scx.cpus_to_sync, GFP_KERNEL, n));
+		INIT_LIST_HEAD(&rq->scx.sched_pcpus_to_kick);
 		raw_spin_lock_init(&rq->scx.deferred_reenq_lock);
 		INIT_LIST_HEAD(&rq->scx.deferred_reenq_locals);
 		INIT_LIST_HEAD(&rq->scx.deferred_reenq_users);
@@ -8603,6 +8717,11 @@ static bool scx_vet_enq_flags(struct scx_sched *sch, u64 dsq_id, u64 *enq_flags)
 	bool is_local = dsq_id == SCX_DSQ_LOCAL ||
 		(dsq_id & SCX_DSQ_LOCAL_ON) == SCX_DSQ_LOCAL_ON;
 
+	if (unlikely(*enq_flags & __SCX_ENQ_INTERNAL_MASK)) {
+		scx_error(sch, "invalid enq_flags 0x%llx", *enq_flags);
+		return false;
+	}
+
 	if (*enq_flags & SCX_ENQ_IMMED) {
 		if (unlikely(!is_local)) {
 			scx_error(sch, "SCX_ENQ_IMMED on a non-local DSQ 0x%llx", dsq_id);
@@ -8612,6 +8731,11 @@ static bool scx_vet_enq_flags(struct scx_sched *sch, u64 dsq_id, u64 *enq_flags)
 		*enq_flags |= SCX_ENQ_IMMED;
 	}
 
+	if (unlikely((*enq_flags & SCX_ENQ_RESCUE) && !is_local)) {
+		scx_error(sch, "SCX_ENQ_RESCUE on a non-local DSQ 0x%llx", dsq_id);
+		return false;
+	}
+
 	return true;
 }
 
@@ -8625,11 +8749,6 @@ static bool scx_dsq_insert_preamble(struct scx_sched *sch, struct task_struct *p
 		return false;
 	}
 
-	if (unlikely(*enq_flags & __SCX_ENQ_INTERNAL_MASK)) {
-		scx_error(sch, "invalid enq_flags 0x%llx", *enq_flags);
-		return false;
-	}
-
 	/* see SCX_EV_INSERT_NOT_OWNED definition */
 	if (unlikely(!scx_task_on_sched(sch, p))) {
 		__scx_add_event(sch, SCX_EV_INSERT_NOT_OWNED, 1);
@@ -8643,14 +8762,14 @@ static bool scx_dsq_insert_preamble(struct scx_sched *sch, struct task_struct *p
 }
 
 static void scx_dsq_insert_commit(struct scx_sched *sch, struct task_struct *p,
-				  u64 dsq_id, u64 enq_flags)
+				  u64 dsq_id, u64 slice, u64 vtime, u64 enq_flags)
 {
 	struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
 	struct task_struct *ddsp_task;
 
 	ddsp_task = __this_cpu_read(direct_dispatch_task);
 	if (ddsp_task) {
-		mark_direct_dispatch(sch, ddsp_task, p, dsq_id, enq_flags);
+		mark_direct_dispatch(sch, ddsp_task, p, dsq_id, slice, vtime, enq_flags);
 		return;
 	}
 
@@ -8663,6 +8782,8 @@ static void scx_dsq_insert_commit(struct scx_sched *sch, struct task_struct *p,
 		.task = p,
 		.qseq = atomic_long_read(&p->scx.ops_state) & SCX_OPSS_QSEQ_MASK,
 		.dsq_id = dsq_id,
+		.slice = slice,
+		.vtime = vtime,
 		.enq_flags = enq_flags,
 	};
 }
@@ -8670,7 +8791,7 @@ static void scx_dsq_insert_commit(struct scx_sched *sch, struct task_struct *p,
 __bpf_kfunc_start_defs();
 
 /**
- * scx_bpf_dsq_insert - Insert a task into the FIFO queue of a DSQ
+ * scx_bpf_dsq_insert___v2 - Insert a task into the FIFO queue of a DSQ
  * @p: task_struct to insert
  * @dsq_id: DSQ to insert into
  * @slice: duration @p can run for in nsecs, 0 to keep the current value
@@ -8684,7 +8805,7 @@ __bpf_kfunc_start_defs();
  * When called from ops.select_cpu() or ops.enqueue(), it's for direct dispatch
  * and @p must match the task being enqueued.
  *
- * When called from ops.select_cpu(), @enq_flags and @dsp_id are stored, and @p
+ * When called from ops.select_cpu(), @enq_flags and @dsq_id are stored, and @p
  * will be directly inserted into the corresponding dispatch queue after
  * ops.select_cpu() returns. If @p is inserted into SCX_DSQ_LOCAL, it will be
  * inserted into the local DSQ of the CPU returned by ops.select_cpu().
@@ -8723,12 +8844,7 @@ __bpf_kfunc bool scx_bpf_dsq_insert___v2(struct task_struct *p, u64 dsq_id,
 	if (!scx_dsq_insert_preamble(sch, p, dsq_id, &enq_flags))
 		return false;
 
-	if (slice)
-		p->scx.slice = slice;
-	else
-		p->scx.slice = p->scx.slice ?: 1;
-
-	scx_dsq_insert_commit(sch, p, dsq_id, enq_flags);
+	scx_dsq_insert_commit(sch, p, dsq_id, slice, 0, enq_flags);
 
 	return true;
 }
@@ -8749,14 +8865,7 @@ static bool scx_dsq_insert_vtime(struct scx_sched *sch, struct task_struct *p,
 	if (!scx_dsq_insert_preamble(sch, p, dsq_id, &enq_flags))
 		return false;
 
-	if (slice)
-		p->scx.slice = slice;
-	else
-		p->scx.slice = p->scx.slice ?: 1;
-
-	p->scx.dsq_vtime = vtime;
-
-	scx_dsq_insert_commit(sch, p, dsq_id, enq_flags | SCX_ENQ_DSQ_PRIQ);
+	scx_dsq_insert_commit(sch, p, dsq_id, slice, vtime, enq_flags | SCX_ENQ_DSQ_PRIQ);
 
 	return true;
 }
@@ -8861,13 +8970,13 @@ static const struct btf_kfunc_id_set scx_kfunc_set_enqueue_dispatch = {
 };
 
 static bool scx_dsq_move(struct bpf_iter_scx_dsq_kern *kit,
-			 struct task_struct *p, u64 dsq_id, u64 enq_flags)
+			 struct task_struct *p, u64 dsq_id, u64 enq_flags,
+			 bool priq)
 {
 	struct scx_dispatch_q *src_dsq = kit->dsq, *dst_dsq;
 	struct scx_sched *sch;
-	struct rq *this_rq, *src_rq, *locked_rq;
+	struct rq *p_rq, *src_rq, *locked_rq;
 	bool dispatched = false;
-	bool in_balance;
 	unsigned long flags;
 
 	/*
@@ -8883,9 +8992,13 @@ static bool scx_dsq_move(struct bpf_iter_scx_dsq_kern *kit,
 	if (!scx_vet_enq_flags(sch, dsq_id, &enq_flags))
 		return false;
 
+	/* internal bit, can only go in after @enq_flags is vetted */
+	if (priq)
+		enq_flags |= SCX_ENQ_DSQ_PRIQ;
+
 	/*
 	 * If the BPF scheduler keeps calling this function repeatedly, it can
-	 * cause similar live-lock conditions as consume_dispatch_q().
+	 * cause similar live-lock conditions as scx_consume_dispatch_q().
 	 */
 	if (unlikely(READ_ONCE(sch->aborting)))
 		return false;
@@ -8897,24 +9010,28 @@ static bool scx_dsq_move(struct bpf_iter_scx_dsq_kern *kit,
 	}
 
 	/*
-	 * Can be called from either ops.dispatch() locking this_rq() or any
-	 * context where no rq lock is held. If latter, lock @p's task_rq which
-	 * we'll likely need anyway.
+	 * Can be called from either ops.dispatch() holding the dispatched rq's
+	 * lock or any context where no rq lock is held. If latter, lock @p's
+	 * task_rq which we'll likely need anyway.
 	 */
 	src_rq = task_rq(p);
 
 	local_irq_save(flags);
-	this_rq = this_rq();
-	in_balance = this_rq->scx.flags & SCX_RQ_IN_BALANCE;
 
-	if (in_balance) {
-		if (this_rq != src_rq)
-			switch_rq_lock(this_rq, src_rq);
+	/*
+	 * Under core scheduling, dispatch can run for a sibling rq, so the
+	 * locked rq is not necessarily this CPU's.
+	 */
+	locked_rq = scx_locked_rq();
+
+	if (locked_rq) {
+		if (locked_rq != src_rq)
+			switch_rq_lock(locked_rq, src_rq);
 	} else {
 		raw_spin_rq_lock(src_rq);
 	}
 
-	locked_rq = src_rq;
+	p_rq = src_rq;
 	raw_spin_lock(&src_dsq->lock);
 
 	/* did someone else get to it while we dropped the locks? */
@@ -8924,27 +9041,28 @@ static bool scx_dsq_move(struct bpf_iter_scx_dsq_kern *kit,
 	}
 
 	/* @p is still on $src_dsq and stable, determine the destination */
-	dst_dsq = find_dsq_for_dispatch(sch, this_rq, dsq_id, task_cpu(p));
+	dst_dsq = find_dsq_for_dispatch(sch, locked_rq ?: this_rq(), dsq_id, task_cpu(p));
 
 	/*
-	 * Apply vtime and slice updates before moving so that the new time is
-	 * visible before inserting into $dst_dsq. @p is still on $src_dsq but
-	 * this is safe as we're locking it.
+	 * Apply vtime and slice updates before moving. @p is still on $src_dsq
+	 * with both $src_dsq and its task_rq locked, satisfying the write
+	 * rules, and the PRIQ insertion into $dst_dsq reads the new vtime.
 	 */
 	if (kit->cursor.flags & __SCX_DSQ_ITER_HAS_VTIME)
 		p->scx.dsq_vtime = kit->vtime;
 	if (kit->cursor.flags & __SCX_DSQ_ITER_HAS_SLICE)
-		p->scx.slice = kit->slice;
+		scx_set_task_slice(p, kit->slice);
 
 	/* execute move */
-	locked_rq = move_task_between_dsqs(sch, p, enq_flags, src_dsq, dst_dsq);
+	p_rq = move_task_between_dsqs(sch, p, enq_flags, src_dsq, dst_dsq);
 	dispatched = true;
 out:
-	if (in_balance) {
-		if (this_rq != locked_rq)
-			switch_rq_lock(locked_rq, this_rq);
+	if (locked_rq) {
+		if (locked_rq != p_rq)
+			switch_rq_lock(p_rq, locked_rq);
 	} else {
-		raw_spin_rq_unlock_irqrestore(locked_rq, flags);
+		scx_rq_lock_drop(p_rq);
+		raw_spin_rq_unlock_irqrestore(p_rq, flags);
 	}
 
 	kit->cursor.flags &= ~(__SCX_DSQ_ITER_HAS_SLICE |
@@ -9000,7 +9118,7 @@ __bpf_kfunc void scx_bpf_dispatch_cancel(const struct bpf_prog_aux *aux)
 }
 
 /**
- * scx_bpf_dsq_move_to_local - move a task from a DSQ to the current CPU's local DSQ
+ * scx_bpf_dsq_move_to_local___v2 - move a task from a DSQ to the current CPU's local DSQ
  * @dsq_id: DSQ to move task from. Must be a user-created DSQ
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  * @enq_flags: %SCX_ENQ_*
@@ -9042,7 +9160,7 @@ __bpf_kfunc bool scx_bpf_dsq_move_to_local___v2(u64 dsq_id, u64 enq_flags,
 
 	dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
 
-	flush_dispatch_buf(sch, dspc->rq);
+	scx_flush_dispatch_buf(sch, dspc->rq);
 
 	dsq = find_user_dsq(sch, dsq_id);
 	if (unlikely(!dsq)) {
@@ -9050,11 +9168,11 @@ __bpf_kfunc bool scx_bpf_dsq_move_to_local___v2(u64 dsq_id, u64 enq_flags,
 		return false;
 	}
 
-	if (consume_dispatch_q(sch, dspc->rq, dsq, enq_flags)) {
+	if (scx_consume_dispatch_q(sch, dspc->rq, dsq, enq_flags)) {
 		/*
 		 * A successfully consumed task can be dequeued before it starts
 		 * running while the CPU is trying to migrate other dispatched
-		 * tasks. Bump nr_tasks to tell balance_one() to retry on empty
+		 * tasks. Bump nr_tasks to tell dispatch_one() to retry on empty
 		 * local DSQ.
 		 */
 		dspc->nr_tasks++;
@@ -9139,7 +9257,7 @@ __bpf_kfunc bool scx_bpf_dsq_move(struct bpf_iter_scx_dsq *it__iter,
 				  u64 enq_flags)
 {
 	return scx_dsq_move((struct bpf_iter_scx_dsq_kern *)it__iter,
-			    p, dsq_id, enq_flags);
+			    p, dsq_id, enq_flags, false);
 }
 
 /**
@@ -9164,47 +9282,8 @@ __bpf_kfunc bool scx_bpf_dsq_move_vtime(struct bpf_iter_scx_dsq *it__iter,
 					u64 enq_flags)
 {
 	return scx_dsq_move((struct bpf_iter_scx_dsq_kern *)it__iter,
-			    p, dsq_id, enq_flags | SCX_ENQ_DSQ_PRIQ);
-}
-
-#ifdef CONFIG_EXT_SUB_SCHED
-/**
- * scx_bpf_sub_dispatch - Trigger dispatching on a child scheduler
- * @cgroup_id: cgroup ID of the child scheduler to dispatch
- * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
- *
- * Allows a parent scheduler to trigger dispatching on one of its direct
- * child schedulers. The child scheduler runs its dispatch operation to
- * move tasks from dispatch queues to the local runqueue.
- *
- * Returns: true on success, false if cgroup_id is invalid, not a direct
- * child, or caller lacks dispatch permission.
- */
-__bpf_kfunc bool scx_bpf_sub_dispatch(u64 cgroup_id, const struct bpf_prog_aux *aux)
-{
-	struct rq *this_rq = this_rq();
-	struct scx_sched *parent, *child;
-
-	guard(rcu)();
-	parent = scx_prog_sched(aux);
-	if (unlikely(!parent))
-		return false;
-
-	child = scx_find_sub_sched(cgroup_id);
-
-	if (unlikely(!child))
-		return false;
-
-	if (unlikely(scx_parent(child) != parent)) {
-		scx_error(parent, "trying to dispatch a distant sub-sched on cgroup %llu",
-			  cgroup_id);
-		return false;
-	}
-
-	return scx_dispatch_sched(child, this_rq, this_rq->scx.sub_dispatch_prev,
-				  true);
+			    p, dsq_id, enq_flags, true);
 }
-#endif	/* CONFIG_EXT_SUB_SCHED */
 
 __bpf_kfunc_end_defs();
 
@@ -9296,10 +9375,10 @@ __bpf_kfunc s32 scx_bpf_create_dsq(u64 dsq_id, s32 node, const struct bpf_prog_a
 		return -ENOMEM;
 
 	/*
-	 * init_dsq() must be called in GFP_KERNEL context. Init it with NULL
-	 * @sch and update afterwards.
+	 * scx_init_dsq() must be called in GFP_KERNEL context. Init it with
+	 * NULL @sch and update afterwards.
 	 */
-	ret = init_dsq(dsq, dsq_id, NULL);
+	ret = scx_init_dsq(dsq, dsq_id, NULL);
 	if (ret) {
 		kfree(dsq);
 		return ret;
@@ -9353,20 +9432,60 @@ __bpf_kfunc_start_defs();
  * @slice: time slice to set in nsecs
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
- * Set @p's time slice to @slice. Returns %true on success, %false if the
- * calling scheduler doesn't have authority over @p.
+ * Set @p's time slice. @p must be on the calling scheduler. The value is
+ * applied whether or not the caller holds @p's rq lock - see the slice write
+ * rules above for the ownership model.
+ *
+ * Raising the slice is honored only while the scheduler holds %SCX_CAP_BASE on
+ * @p's cpu, otherwise it is counted in %SCX_EV_SLICE_DENIED. Shortening is
+ * always allowed. On the stashed path the slice is packed into an atomic64_t
+ * with the scheduler id and a flag bit, so a slice too large to fit is clamped
+ * and counted in %SCX_EV_SLICE_CLAMPED. %SCX_SLICE_INF is preserved.
+ *
+ * Return %true on success, %false if @p is not on the calling scheduler.
  */
 __bpf_kfunc bool scx_bpf_task_set_slice(struct task_struct *p, u64 slice,
 					const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
+	struct rq *locked_rq;
 
 	guard(rcu)();
 	sch = scx_prog_sched(aux);
 	if (unlikely(!sch || !scx_task_on_sched(sch, p)))
 		return false;
 
-	p->scx.slice = slice;
+	/*
+	 * Directly write only when we hold the lock of the rq @p is queued or
+	 * running on. See the write rules above.
+	 *
+	 * While @p is queued on a user DSQ or in the BPF scheduler,
+	 * synchronization is the scheduler's responsibility. This write can
+	 * race a concurrent dispatch's commit, see apply_slice_vtime().
+	 *
+	 * Making this kfunc always go through the oob stash would leave the
+	 * commit as the only direct writer and close the race, but that would
+	 * require two more oob application points - the dispatch keep-prev test
+	 * and the tick-time expiry check.
+	 */
+	locked_rq = scx_locked_rq();
+	if (!locked_rq ||
+	    (READ_ONCE(p->scx.runnable_cpu) != cpu_of(locked_rq) &&
+	     !task_current(locked_rq, p))) {
+		set_task_slice_oob(sch, p, slice);
+		return true;
+	}
+
+	/* under the rq lock: apply now, extensions gated on baseline access */
+	if (slice > p->scx.slice &&
+	    unlikely(scx_missing_caps(sch, cpu_of(locked_rq), SCX_CAP_BASE))) {
+		__scx_add_event(sch, SCX_EV_SLICE_DENIED, 1);
+		return true;
+	}
+
+	if (unlikely(!scx_set_task_slice(p, slice)))
+		__scx_add_event(sch, SCX_EV_SLICE_DENIED, 1);
+
 	return true;
 }
 
@@ -9393,14 +9512,25 @@ __bpf_kfunc bool scx_bpf_task_set_dsq_vtime(struct task_struct *p, u64 vtime,
 	return true;
 }
 
-static void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
+void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
 {
+	struct scx_sched_pcpu *pcpu;
 	struct rq *this_rq;
 	unsigned long irq_flags;
 
+	/*
+	 * The per-cpu kick list is guarded only by local_irq_save(), which does
+	 * not mask NMIs, so kicking from NMI could corrupt it and is unsupported.
+	 */
+	if (unlikely(in_nmi())) {
+		scx_error(sch, "scx_bpf_kick_cpu() called from NMI");
+		return;
+	}
+
 	local_irq_save(irq_flags);
 
 	this_rq = this_rq();
+	pcpu = this_cpu_ptr(sch->pcpu);
 
 	/*
 	 * While bypassing for PM ops, IRQ handling may not be online which can
@@ -9414,6 +9544,9 @@ static void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
 	 * Actual kicking is bounced to kick_cpus_irq_workfn() to avoid nesting
 	 * rq locks. We can probably be smarter and avoid bouncing if called
 	 * from ops which don't hold a rq lock.
+	 *
+	 * The kick masks are owned by @sch->pcpu, so that a preempt kick can be
+	 * attributed to @sch.
 	 */
 	if (flags & SCX_KICK_IDLE) {
 		struct rq *target_rq = cpu_rq(cpu);
@@ -9423,21 +9556,25 @@ static void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags)
 
 		if (raw_spin_rq_trylock(target_rq)) {
 			if (can_skip_idle_kick(target_rq)) {
+				scx_rq_lock_drop(target_rq);
 				raw_spin_rq_unlock(target_rq);
 				goto out;
 			}
+			scx_rq_lock_drop(target_rq);
 			raw_spin_rq_unlock(target_rq);
 		}
-		cpumask_set_cpu(cpu, this_rq->scx.cpus_to_kick_if_idle);
+		cpumask_set_cpu(cpu, pcpu->cpus_to_kick_if_idle);
 	} else {
-		cpumask_set_cpu(cpu, this_rq->scx.cpus_to_kick);
+		cpumask_set_cpu(cpu, pcpu->cpus_to_kick);
 
 		if (flags & SCX_KICK_PREEMPT)
-			cpumask_set_cpu(cpu, this_rq->scx.cpus_to_preempt);
+			cpumask_set_cpu(cpu, pcpu->cpus_to_preempt);
 		if (flags & SCX_KICK_WAIT)
-			cpumask_set_cpu(cpu, this_rq->scx.cpus_to_wait);
+			cpumask_set_cpu(cpu, pcpu->cpus_to_wait);
 	}
 
+	if (list_empty(&pcpu->to_kick_node))
+		list_add_tail(&pcpu->to_kick_node, &this_rq->scx.sched_pcpus_to_kick);
 	irq_work_queue(&this_rq->scx.kick_cpus_irq_work);
 out:
 	local_irq_restore(irq_flags);
@@ -9470,10 +9607,13 @@ __bpf_kfunc void scx_bpf_kick_cpu(s32 cpu, u64 flags, const struct bpf_prog_aux
  * @flags: %SCX_KICK_* flags
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
- * cid-addressed equivalent of scx_bpf_kick_cpu(). Return 0 on success,
- * -errno otherwise.
+ * cid-addressed equivalent of scx_bpf_kick_cpu(). An invalid @cid aborts the
+ * scheduler via scx_cid_to_cpu(). Caps are enforced on the delivery path: a
+ * kick is dropped if the caller lacks baseline access on @cid, and a
+ * %SCX_KICK_PREEMPT degrades to a plain reschedule if the caller lacks
+ * %SCX_CAP_PREEMPT for a task outside its subtree.
  */
-__bpf_kfunc s32 scx_bpf_kick_cid(s32 cid, u64 flags, const struct bpf_prog_aux *aux)
+__bpf_kfunc void scx_bpf_kick_cid(s32 cid, u64 flags, const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
 	s32 cpu;
@@ -9481,12 +9621,11 @@ __bpf_kfunc s32 scx_bpf_kick_cid(s32 cid, u64 flags, const struct bpf_prog_aux *
 	guard(rcu)();
 	sch = scx_prog_sched(aux);
 	if (unlikely(!sch))
-		return -ENODEV;
+		return;
 	cpu = scx_cid_to_cpu(sch, cid);
 	if (cpu < 0)
-		return cpu;
+		return;
 	scx_kick_cpu(sch, cpu, flags);
-	return 0;
 }
 
 /**
@@ -9496,6 +9635,10 @@ __bpf_kfunc s32 scx_bpf_kick_cid(s32 cid, u64 flags, const struct bpf_prog_aux *
  *
  * Return the number of tasks in the DSQ matching @dsq_id. If not found,
  * -%ENOENT is returned.
+ *
+ * %SCX_DSQ_LOCAL resolves to the local DSQ of the rq the current scheduler
+ * operation is locked to - e.g. the rq being dispatched for in ops.dispatch() -
+ * or the calling CPU's when no rq is locked.
  */
 __bpf_kfunc s32 scx_bpf_dsq_nr_queued(u64 dsq_id, const struct bpf_prog_aux *aux)
 {
@@ -9512,7 +9655,7 @@ __bpf_kfunc s32 scx_bpf_dsq_nr_queued(u64 dsq_id, const struct bpf_prog_aux *aux
 	}
 
 	if (dsq_id == SCX_DSQ_LOCAL) {
-		ret = READ_ONCE(this_rq()->scx.local_dsq.nr);
+		ret = READ_ONCE((scx_locked_rq() ?: this_rq())->scx.local_dsq.nr);
 		goto out;
 	} else if ((dsq_id & SCX_DSQ_LOCAL_ON) == SCX_DSQ_LOCAL_ON) {
 		s32 cpu = scx_cpu_ret(sch, dsq_id & SCX_DSQ_LOCAL_CPU_MASK);
@@ -9691,10 +9834,15 @@ __bpf_kfunc struct task_struct *scx_bpf_dsq_peek(u64 dsq_id,
  * - User DSQs
  *
  * Re-enqueues are performed asynchronously. Can be called from anywhere.
+ *
+ * %SCX_DSQ_LOCAL resolves to the local DSQ of the rq the current scheduler
+ * operation is locked to - e.g. the rq being dispatched for in ops.dispatch() -
+ * or the calling CPU's when no rq is locked.
  */
 __bpf_kfunc void scx_bpf_dsq_reenq(u64 dsq_id, u64 reenq_flags,
 				   const struct bpf_prog_aux *aux)
 {
+	struct rq *locked_rq = scx_locked_rq();
 	struct scx_sched *sch;
 	struct scx_dispatch_q *dsq;
 
@@ -9713,12 +9861,12 @@ __bpf_kfunc void scx_bpf_dsq_reenq(u64 dsq_id, u64 reenq_flags,
 	if (!(reenq_flags & __SCX_REENQ_FILTER_MASK))
 		reenq_flags |= SCX_REENQ_ANY;
 
-	dsq = find_dsq_for_dispatch(sch, this_rq(), dsq_id, smp_processor_id());
-	schedule_dsq_reenq(sch, dsq, reenq_flags, scx_locked_rq());
+	dsq = find_dsq_for_dispatch(sch, locked_rq ?: this_rq(), dsq_id, smp_processor_id());
+	schedule_dsq_reenq(sch, dsq, reenq_flags, locked_rq);
 }
 
 /**
- * scx_bpf_reenqueue_local - Re-enqueue tasks on a local DSQ
+ * scx_bpf_reenqueue_local___v2 - Re-enqueue tasks on a local DSQ
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
  * Iterate over all of the tasks currently enqueued on the local DSQ of the
@@ -9773,12 +9921,38 @@ static s32 __bstr_format(struct scx_sched *sch, u64 *data_buf, char *line_buf,
 	return ret;
 }
 
-__printf(3, 0)
-static s32 bstr_format(struct scx_sched *sch, struct scx_bstr_buf *buf,
-		       char *fmt, unsigned long long *data, u32 data__sz)
+/*
+ * Exit @sch with the reason formatted from a BPF-supplied bstr format. The exit
+ * is claimed first and the reason is formatted directly into the winner-owned
+ * exit_info buffer, which allows use from any context including NMI.
+ *
+ * @fmt_blame is the sched blamed for formatting failures through the
+ * scx_error() calls in __bstr_format() and differs from @sch when a parent
+ * supplies the kill reason for a child. A formatting failure doesn't revert the
+ * claim - @sch still exits with the claimed kind and a fallback message.
+ */
+__printf(5, 0)
+bool scx_exit_bstr(struct scx_sched *sch, enum scx_exit_kind kind,
+		   s64 exit_code, struct scx_sched *fmt_blame, char *fmt,
+		   unsigned long long *data, u32 data__sz)
 {
-	return __bstr_format(sch, buf->data, buf->line, sizeof(buf->line),
-			     fmt, data, data__sz);
+	struct scx_exit_info *ei = sch->exit_info;
+	u64 data_buf[MAX_BPRINTF_VARARGS];
+	s32 ret;
+
+	guard(preempt)();
+
+	if (!scx_claim_exit(sch, kind))
+		return false;
+
+	ret = __bstr_format(fmt_blame, data_buf, ei->msg, SCX_EXIT_MSG_LEN,
+			    fmt, data, data__sz);
+	if (ret < 0)
+		scnprintf(ei->msg, SCX_EXIT_MSG_LEN,
+			  "exit message formatting failed (%d)", ret);
+
+	scx_finish_exit(sch, kind, exit_code, raw_smp_processor_id());
+	return true;
 }
 
 __bpf_kfunc_start_defs();
@@ -9800,14 +9974,13 @@ __bpf_kfunc void scx_bpf_exit_bstr(s64 exit_code, char *fmt,
 				   const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
-	unsigned long flags;
 
-	raw_spin_lock_irqsave(&scx_exit_bstr_buf_lock, flags);
+	guard(rcu)();
+
 	sch = scx_prog_sched(aux);
-	if (likely(sch) &&
-	    bstr_format(sch, &scx_exit_bstr_buf, fmt, data, data__sz) >= 0)
-		scx_exit(sch, SCX_EXIT_UNREG_BPF, exit_code, "%s", scx_exit_bstr_buf.line);
-	raw_spin_unlock_irqrestore(&scx_exit_bstr_buf_lock, flags);
+	if (likely(sch))
+		scx_exit_bstr(sch, SCX_EXIT_UNREG_BPF, exit_code, sch, fmt,
+			      data, data__sz);
 }
 
 /**
@@ -9825,14 +9998,13 @@ __bpf_kfunc void scx_bpf_error_bstr(char *fmt, unsigned long long *data,
 				    u32 data__sz, const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
-	unsigned long flags;
 
-	raw_spin_lock_irqsave(&scx_exit_bstr_buf_lock, flags);
+	guard(rcu)();
+
 	sch = scx_prog_sched(aux);
-	if (likely(sch) &&
-	    bstr_format(sch, &scx_exit_bstr_buf, fmt, data, data__sz) >= 0)
-		scx_exit(sch, SCX_EXIT_ERROR_BPF, 0, "%s", scx_exit_bstr_buf.line);
-	raw_spin_unlock_irqrestore(&scx_exit_bstr_buf_lock, flags);
+	if (likely(sch))
+		scx_exit_bstr(sch, SCX_EXIT_ERROR_BPF, 0, sch, fmt, data,
+			      data__sz);
 }
 
 /**
@@ -9872,8 +10044,8 @@ __bpf_kfunc void scx_bpf_dump_bstr(char *fmt, unsigned long long *data,
 	ret = __bstr_format(sch, buf->data, buf->line + dd->cursor,
 			    sizeof(buf->line) - dd->cursor, fmt, data, data__sz);
 	if (ret < 0) {
-		dump_line(dd->s, "%s[!] (\"%s\", %p, %u) failed to format (%d)",
-			  dd->prefix, fmt, data, data__sz, ret);
+		scx_dump_line(dd->s, "%s[!] (\"%s\", %p, %u) failed to format (%d)",
+			      dd->prefix, fmt, data, data__sz, ret);
 		return;
 	}
 
@@ -9991,6 +10163,62 @@ __bpf_kfunc u32 scx_bpf_cidperf_cur(s32 cid, const struct bpf_prog_aux *aux)
 	return arch_scale_freq_capacity(cpu);
 }
 
+/* validate and apply a cpuperf target, see scx_bpf_cpuperf_set() */
+static s32 scx_cpuperf_set(struct scx_sched *sch, s32 cpu, u32 perf)
+{
+	struct rq *rq, *locked_rq;
+	struct rq_flags rf;
+	s32 ret;
+
+	if (unlikely(perf > SCX_CPUPERF_ONE)) {
+		scx_error(sch, "Invalid cpuperf target %u for CPU %d", perf, cpu);
+		return -EINVAL;
+	}
+
+	if (!scx_cpu_valid(sch, cpu, NULL))
+		return -EINVAL;
+
+	rq = cpu_rq(cpu);
+	locked_rq = scx_locked_rq();
+
+	/*
+	 * When called with an rq lock held, restrict the operation to the
+	 * corresponding CPU to prevent ABBA deadlocks.
+	 */
+	if (locked_rq && rq != locked_rq) {
+		scx_error(sch, "Invalid target CPU %d", cpu);
+		return -EINVAL;
+	}
+
+	/*
+	 * If no rq lock is held, allow to operate on any CPU by acquiring
+	 * the corresponding rq lock.
+	 */
+	if (!locked_rq) {
+		rq_lock_irqsave(rq, &rf);
+		update_rq_clock(rq);
+	}
+
+	/*
+	 * ecaps updates are folded under the rq lock, making this test
+	 * authoritative: a write can never land after a revoke has taken
+	 * effect on @cpu.
+	 */
+	if (likely(!scx_missing_caps(sch, cpu, SCX_CAP_PERF))) {
+		rq->scx.cpuperf_target = perf;
+		cpufreq_update_util(rq, 0);
+		ret = 0;
+	} else {
+		__scx_add_event(sch, SCX_EV_SUB_CIDPERF_DENIED, 1);
+		ret = -EACCES;
+	}
+
+	if (!locked_rq)
+		rq_unlock_irqrestore(rq, &rf);
+
+	return ret;
+}
+
 /**
  * scx_bpf_cpuperf_set - Set the relative performance target of a CPU
  * @cpu: CPU of interest
@@ -10016,39 +10244,7 @@ __bpf_kfunc void scx_bpf_cpuperf_set(s32 cpu, u32 perf, const struct bpf_prog_au
 	if (unlikely(!sch))
 		return;
 
-	if (unlikely(perf > SCX_CPUPERF_ONE)) {
-		scx_error(sch, "Invalid cpuperf target %u for CPU %d", perf, cpu);
-		return;
-	}
-
-	if (scx_cpu_valid(sch, cpu, NULL)) {
-		struct rq *rq = cpu_rq(cpu), *locked_rq = scx_locked_rq();
-		struct rq_flags rf;
-
-		/*
-		 * When called with an rq lock held, restrict the operation
-		 * to the corresponding CPU to prevent ABBA deadlocks.
-		 */
-		if (locked_rq && rq != locked_rq) {
-			scx_error(sch, "Invalid target CPU %d", cpu);
-			return;
-		}
-
-		/*
-		 * If no rq lock is held, allow to operate on any CPU by
-		 * acquiring the corresponding rq lock.
-		 */
-		if (!locked_rq) {
-			rq_lock_irqsave(rq, &rf);
-			update_rq_clock(rq);
-		}
-
-		rq->scx.cpuperf_target = perf;
-		cpufreq_update_util(rq, 0);
-
-		if (!locked_rq)
-			rq_unlock_irqrestore(rq, &rf);
-	}
+	scx_cpuperf_set(sch, cpu, perf);
 }
 
 /**
@@ -10057,10 +10253,13 @@ __bpf_kfunc void scx_bpf_cpuperf_set(s32 cpu, u32 perf, const struct bpf_prog_au
  * @perf: target performance level [0, %SCX_CPUPERF_ONE]
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
- * cid-addressed equivalent of scx_bpf_cpuperf_set().
+ * cid-addressed equivalent of scx_bpf_cpuperf_set(). A sub-sched needs
+ * SCX_CAP_PERF on @cid. Returns 0 if the target was applied, -%EACCES if
+ * the write was denied for missing caps, other -errnos if @cid didn't
+ * resolve.
  */
-__bpf_kfunc void scx_bpf_cidperf_set(s32 cid, u32 perf,
-				     const struct bpf_prog_aux *aux)
+__bpf_kfunc s32 scx_bpf_cidperf_set(s32 cid, u32 perf,
+				    const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
 	s32 cpu;
@@ -10069,11 +10268,12 @@ __bpf_kfunc void scx_bpf_cidperf_set(s32 cid, u32 perf,
 
 	sch = scx_prog_sched(aux);
 	if (unlikely(!sch))
-		return;
+		return -ENODEV;
 	cpu = scx_cid_to_cpu(sch, cid);
 	if (cpu < 0)
-		return;
-	scx_bpf_cpuperf_set(cpu, perf, aux);
+		return cpu;
+
+	return scx_cpuperf_set(sch, cpu, perf);
 }
 
 /**
@@ -10125,13 +10325,15 @@ __bpf_kfunc u32 scx_bpf_nr_online_cids(void)
  *
  * cid-addressed equivalent of bpf_get_smp_processor_id() for scx programs.
  * The current cpu is trivially valid, so this is just a table lookup. Return
- * -EINVAL if called from a non-SCX program before any scheduler has ever
- * been enabled (the cid table is still unallocated at that point).
+ * -EINVAL if called before any scheduler has ever published its cid tables.
  */
 __bpf_kfunc s32 scx_bpf_this_cid(void)
 {
-	s16 *tbl = READ_ONCE(scx_cpu_to_cid_tbl);
+	s16 *tbl;
 
+	guard(rcu)();
+
+	tbl = rcu_dereference(scx_cpu_to_cid_tbl);
 	if (!tbl)
 		return -EINVAL;
 	return tbl[raw_smp_processor_id()];
@@ -10190,47 +10392,23 @@ __bpf_kfunc s32 scx_bpf_task_cpu(const struct task_struct *p)
  * @p: task of interest
  *
  * cid-addressed equivalent of scx_bpf_task_cpu(). task_cpu(p) is always a
- * valid cpu, so this is just a table lookup. Return -EINVAL if called from
- * a non-SCX program before any scheduler has ever been enabled.
+ * valid cpu, so this is just a table lookup. Return -EINVAL if called before
+ * any scheduler has ever published its cid tables.
  */
 __bpf_kfunc s32 scx_bpf_task_cid(const struct task_struct *p)
 {
-	s16 *tbl = READ_ONCE(scx_cpu_to_cid_tbl);
+	s16 *tbl;
 
+	/* KF_RCU covers only @p - a sleepable program holds no RCU lock */
+	guard(rcu)();
+
+	tbl = rcu_dereference(scx_cpu_to_cid_tbl);
 	if (!tbl)
 		return -EINVAL;
 	return tbl[task_cpu(p)];
 }
 
 /**
- * scx_bpf_cpu_rq - Fetch the rq of a CPU
- * @cpu: CPU of the rq
- * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
- */
-__bpf_kfunc struct rq *scx_bpf_cpu_rq(s32 cpu, const struct bpf_prog_aux *aux)
-{
-	struct scx_sched *sch;
-
-	guard(rcu)();
-
-	sch = scx_prog_sched(aux);
-	if (unlikely(!sch))
-		return NULL;
-
-	if (!scx_cpu_valid(sch, cpu, NULL))
-		return NULL;
-
-	if (!sch->warned_deprecated_rq) {
-		printk_deferred(KERN_WARNING "sched_ext: %s() is deprecated; "
-				"use scx_bpf_locked_rq() when holding rq lock "
-				"or scx_bpf_cpu_curr() to read remote curr safely.\n", __func__);
-		sch->warned_deprecated_rq = true;
-	}
-
-	return cpu_rq(cpu);
-}
-
-/**
  * scx_bpf_locked_rq - Return the rq currently locked by SCX
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
@@ -10332,6 +10510,27 @@ __bpf_kfunc struct task_struct *scx_bpf_tid_to_task(u64 tid)
 	return container_of(scx, struct task_struct, scx);
 }
 
+u64 __scx_bpf_now(struct rq *rq)
+{
+	/* the caller must be on @rq's cpu or hold its lock */
+	lockdep_assert((rq == this_rq() && !preemptible()) ||
+		       lockdep_is_held(__rq_lockp(rq)));
+
+	if (smp_load_acquire(&rq->scx.flags) & SCX_RQ_CLK_VALID) {
+		/* if the rq clock is valid, use the cached rq clock */
+		return READ_ONCE(rq->scx.clock);
+	} else {
+		/*
+		 * Otherwise, return a fresh rq clock.
+		 *
+		 * The rq clock is updated outside of the rq lock.
+		 * In this case, keep the updated rq clock invalid so the next
+		 * read outside the rq lock gets a fresh rq clock.
+		 */
+		return sched_clock_cpu(cpu_of(rq));
+	}
+}
+
 /**
  * scx_bpf_now - Returns a high-performance monotonically non-decreasing
  * clock for the current CPU. The clock returned is in nanoseconds.
@@ -10362,76 +10561,47 @@ __bpf_kfunc struct task_struct *scx_bpf_tid_to_task(u64 tid)
  */
 __bpf_kfunc u64 scx_bpf_now(void)
 {
-	struct rq *rq;
-	u64 clock;
-
-	preempt_disable();
-
-	rq = this_rq();
-	if (smp_load_acquire(&rq->scx.flags) & SCX_RQ_CLK_VALID) {
-		/*
-		 * If the rq clock is valid, use the cached rq clock.
-		 *
-		 * Note that scx_bpf_now() is re-entrant between a process
-		 * context and an interrupt context (e.g., timer interrupt).
-		 * However, we don't need to consider the race between them
-		 * because such race is not observable from a caller.
-		 */
-		clock = READ_ONCE(rq->scx.clock);
-	} else {
-		/*
-		 * Otherwise, return a fresh rq clock.
-		 *
-		 * The rq clock is updated outside of the rq lock.
-		 * In this case, keep the updated rq clock invalid so the next
-		 * kfunc call outside the rq lock gets a fresh rq clock.
-		 */
-		clock = sched_clock_cpu(cpu_of(rq));
-	}
-
-	preempt_enable();
-
-	return clock;
+	/*
+	 * Note that scx_bpf_now() is re-entrant between a process context and
+	 * an interrupt context (e.g., timer interrupt). However, we don't need
+	 * to consider the race between them because such race is not observable
+	 * from a caller.
+	 */
+	guard(preempt)();
+	return __scx_bpf_now(this_rq());
 }
 
 static void scx_read_events(struct scx_sched *sch, struct scx_event_stats *events)
 {
-	struct scx_event_stats *e_cpu;
 	int cpu;
 
 	/* Aggregate per-CPU event counters into @events. */
 	memset(events, 0, sizeof(*events));
 	for_each_possible_cpu(cpu) {
-		e_cpu = &per_cpu_ptr(sch->pcpu, cpu)->event_stats;
-		scx_agg_event(events, e_cpu, SCX_EV_SELECT_CPU_FALLBACK);
-		scx_agg_event(events, e_cpu, SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE);
-		scx_agg_event(events, e_cpu, SCX_EV_DISPATCH_KEEP_LAST);
-		scx_agg_event(events, e_cpu, SCX_EV_ENQ_SKIP_EXITING);
-		scx_agg_event(events, e_cpu, SCX_EV_ENQ_SKIP_MIGRATION_DISABLED);
-		scx_agg_event(events, e_cpu, SCX_EV_REENQ_IMMED);
-		scx_agg_event(events, e_cpu, SCX_EV_REENQ_LOCAL_REPEAT);
-		scx_agg_event(events, e_cpu, SCX_EV_REFILL_SLICE_DFL);
-		scx_agg_event(events, e_cpu, SCX_EV_BYPASS_DURATION);
-		scx_agg_event(events, e_cpu, SCX_EV_BYPASS_DISPATCH);
-		scx_agg_event(events, e_cpu, SCX_EV_BYPASS_ACTIVATE);
-		scx_agg_event(events, e_cpu, SCX_EV_INSERT_NOT_OWNED);
-		scx_agg_event(events, e_cpu, SCX_EV_SUB_BYPASS_DISPATCH);
+		struct scx_event_stats *e_cpu = &per_cpu_ptr(sch->pcpu, cpu)->event_stats;
+#define SCX_EVENT(name)	(events->name += READ_ONCE(e_cpu->name))
+		SCX_EVENTS_LIST(SCX_EVENT);
+#undef SCX_EVENT
 	}
 }
 
-/*
- * scx_bpf_events - Get a system-wide event counter to
+/**
+ * scx_bpf_events - Read the event counters of the calling scheduler
  * @events: output buffer from a BPF program
- * @events__sz: @events len, must end in '__sz'' for the verifier
+ * @events__sz: @events len, must end in '__sz' for the verifier
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Read the event counters of the scheduler associated with the calling program.
+ * @events is zeroed when no scheduler can be resolved.
  */
-__bpf_kfunc void scx_bpf_events(struct scx_event_stats *events,
-				size_t events__sz)
+__bpf_kfunc void scx_bpf_events(struct scx_event_stats *events, size_t events__sz,
+				const struct bpf_prog_aux *aux)
 {
 	struct scx_sched *sch;
 	struct scx_event_stats e_sys;
 
 	rcu_read_lock();
-	sch = rcu_dereference(scx_root);
+	sch = scx_prog_sched(aux);
 	if (sch)
 		scx_read_events(sch, &e_sys);
 	else
@@ -10520,16 +10690,19 @@ BTF_ID_FLAGS(func, scx_bpf_put_cpumask, KF_RELEASE)
 BTF_ID_FLAGS(func, scx_bpf_task_running, KF_RCU)
 BTF_ID_FLAGS(func, scx_bpf_task_cpu, KF_RCU)
 BTF_ID_FLAGS(func, scx_bpf_task_cid, KF_RCU)
-BTF_ID_FLAGS(func, scx_bpf_cpu_rq, KF_IMPLICIT_ARGS)
 BTF_ID_FLAGS(func, scx_bpf_locked_rq, KF_IMPLICIT_ARGS | KF_RET_NULL)
 BTF_ID_FLAGS(func, scx_bpf_cpu_curr, KF_IMPLICIT_ARGS | KF_RET_NULL | KF_RCU_PROTECTED)
 BTF_ID_FLAGS(func, scx_bpf_cid_curr, KF_IMPLICIT_ARGS | KF_RET_NULL | KF_RCU_PROTECTED)
 BTF_ID_FLAGS(func, scx_bpf_tid_to_task, KF_RET_NULL | KF_RCU_PROTECTED)
 BTF_ID_FLAGS(func, scx_bpf_now)
-BTF_ID_FLAGS(func, scx_bpf_events)
+BTF_ID_FLAGS(func, scx_bpf_events, KF_IMPLICIT_ARGS)
 #ifdef CONFIG_CGROUP_SCHED
 BTF_ID_FLAGS(func, scx_bpf_task_cgroup, KF_IMPLICIT_ARGS | KF_RCU | KF_ACQUIRE)
 #endif
+BTF_ID_FLAGS(func, scx_bpf_sub_grant, KF_IMPLICIT_ARGS)
+BTF_ID_FLAGS(func, scx_bpf_sub_revoke, KF_IMPLICIT_ARGS)
+BTF_ID_FLAGS(func, scx_bpf_sub_caps, KF_IMPLICIT_ARGS)
+BTF_ID_FLAGS(func, scx_bpf_sub_kill_bstr, KF_IMPLICIT_ARGS)
 BTF_KFUNCS_END(scx_kfunc_ids_any)
 
 static const struct btf_kfunc_id_set scx_kfunc_set_any = {
@@ -10555,7 +10728,6 @@ static const struct btf_kfunc_id_set scx_kfunc_set_any = {
 BTF_KFUNCS_START(scx_kfunc_ids_cpu_only)
 BTF_ID_FLAGS(func, scx_bpf_kick_cpu, KF_IMPLICIT_ARGS)
 BTF_ID_FLAGS(func, scx_bpf_task_cpu, KF_RCU)
-BTF_ID_FLAGS(func, scx_bpf_cpu_rq, KF_IMPLICIT_ARGS)
 BTF_ID_FLAGS(func, scx_bpf_cpu_curr, KF_IMPLICIT_ARGS | KF_RET_NULL | KF_RCU_PROTECTED)
 BTF_ID_FLAGS(func, scx_bpf_cpu_node, KF_IMPLICIT_ARGS)
 BTF_ID_FLAGS(func, scx_bpf_cpuperf_cap, KF_IMPLICIT_ARGS)
@@ -10588,7 +10760,7 @@ BTF_KFUNCS_END(scx_kfunc_ids_cpu_only)
  */
 enum scx_kf_allow_flags {
 	SCX_KF_ALLOW_UNLOCKED		= 1 << 0,
-	SCX_KF_ALLOW_INIT		= 1 << 1,
+	SCX_KF_ALLOW_INIT_CIDS		= 1 << 1,
 	SCX_KF_ALLOW_CPU_RELEASE	= 1 << 2,
 	SCX_KF_ALLOW_DISPATCH		= 1 << 3,
 	SCX_KF_ALLOW_ENQUEUE		= 1 << 4,
@@ -10618,9 +10790,11 @@ static const u32 scx_kf_allow_flags[] = {
 #endif	/* CONFIG_EXT_GROUP_SCHED */
 	[SCX_OP_IDX(sub_attach)]	= SCX_KF_ALLOW_UNLOCKED,
 	[SCX_OP_IDX(sub_detach)]	= SCX_KF_ALLOW_UNLOCKED,
+	[SCX_OP_IDX(sub_ecaps_updated)]	= SCX_KF_ALLOW_ENQUEUE | SCX_KF_ALLOW_DISPATCH,
 	[SCX_OP_IDX(cpu_online)]	= SCX_KF_ALLOW_UNLOCKED,
 	[SCX_OP_IDX(cpu_offline)]	= SCX_KF_ALLOW_UNLOCKED,
-	[SCX_OP_IDX(init)]		= SCX_KF_ALLOW_UNLOCKED | SCX_KF_ALLOW_INIT,
+	[SCX_OP_IDX(init_cids)]		= SCX_KF_ALLOW_UNLOCKED | SCX_KF_ALLOW_INIT_CIDS,
+	[SCX_OP_IDX(init)]		= SCX_KF_ALLOW_UNLOCKED,
 	[SCX_OP_IDX(exit)]		= SCX_KF_ALLOW_UNLOCKED,
 };
 
@@ -10635,7 +10809,7 @@ static const u32 scx_kf_allow_flags[] = {
 int scx_kfunc_context_filter(const struct bpf_prog *prog, u32 kfunc_id)
 {
 	bool in_unlocked = btf_id_set8_contains(&scx_kfunc_ids_unlocked, kfunc_id);
-	bool in_init = btf_id_set8_contains(&scx_kfunc_ids_init, kfunc_id);
+	bool in_init_cids = btf_id_set8_contains(&scx_kfunc_ids_init_cids, kfunc_id);
 	bool in_select_cpu = btf_id_set8_contains(&scx_kfunc_ids_select_cpu, kfunc_id);
 	bool in_enqueue = btf_id_set8_contains(&scx_kfunc_ids_enqueue_dispatch, kfunc_id);
 	bool in_dispatch = btf_id_set8_contains(&scx_kfunc_ids_dispatch, kfunc_id);
@@ -10643,19 +10817,20 @@ int scx_kfunc_context_filter(const struct bpf_prog *prog, u32 kfunc_id)
 	bool in_idle = btf_id_set8_contains(&scx_kfunc_ids_idle, kfunc_id);
 	bool in_any = btf_id_set8_contains(&scx_kfunc_ids_any, kfunc_id);
 	bool in_cpu_only = btf_id_set8_contains(&scx_kfunc_ids_cpu_only, kfunc_id);
+	bool in_cid = btf_id_set8_contains(&scx_kfunc_ids_cid, kfunc_id);
 	u32 moff, flags;
 
 	/* Not an SCX kfunc - allow. */
-	if (!(in_unlocked || in_init || in_select_cpu || in_enqueue || in_dispatch ||
-	      in_cpu_release || in_idle || in_any))
+	if (!(in_unlocked || in_init_cids || in_select_cpu || in_enqueue || in_dispatch ||
+	      in_cpu_release || in_idle || in_any || in_cid))
 		return 0;
 
 	/* SYSCALL progs (e.g. BPF test_run()) may call unlocked and select_cpu kfuncs. */
 	if (prog->type == BPF_PROG_TYPE_SYSCALL)
-		return (in_unlocked || in_select_cpu || in_idle || in_any) ? 0 : -EACCES;
+		return (in_unlocked || in_select_cpu || in_idle || in_any || in_cid) ? 0 : -EACCES;
 
 	if (prog->type != BPF_PROG_TYPE_STRUCT_OPS)
-		return (in_any || in_idle) ? 0 : -EACCES;
+		return (in_any || in_idle || in_cid) ? 0 : -EACCES;
 
 	/*
 	 * add_subprog_and_kfunc() collects all kfunc calls, including dead code
@@ -10690,7 +10865,7 @@ int scx_kfunc_context_filter(const struct bpf_prog *prog, u32 kfunc_id)
 		return -EACCES;
 
 	/* SCX struct_ops: check the per-op allow list. */
-	if (in_any || in_idle)
+	if (in_any || in_idle || in_cid)
 		return 0;
 
 	moff = prog->aux->attach_st_ops_member_off;
@@ -10698,7 +10873,7 @@ int scx_kfunc_context_filter(const struct bpf_prog *prog, u32 kfunc_id)
 
 	if ((flags & SCX_KF_ALLOW_UNLOCKED) && in_unlocked)
 		return 0;
-	if ((flags & SCX_KF_ALLOW_INIT) && in_init)
+	if ((flags & SCX_KF_ALLOW_INIT_CIDS) && in_init_cids)
 		return 0;
 	if ((flags & SCX_KF_ALLOW_CPU_RELEASE) && in_cpu_release)
 		return 0;
@@ -10732,6 +10907,9 @@ static int __init scx_init(void)
 	CID_OFFSET_MATCH(timeout_ms, timeout_ms);
 	CID_OFFSET_MATCH(exit_dump_len, exit_dump_len);
 	CID_OFFSET_MATCH(hotplug_seq, hotplug_seq);
+	CID_OFFSET_MATCH(cid_shard_size, cid_shard_size);
+	CID_OFFSET_MATCH(rescue_bandwidth_ppt, rescue_bandwidth_ppt);
+	CID_OFFSET_MATCH(rescue_quantum_us, rescue_quantum_us);
 	CID_OFFSET_MATCH(sub_cgroup_id, sub_cgroup_id);
 	/* shared callbacks: the union view requires byte-for-byte offset match */
 	CID_OFFSET_MATCH(enqueue, enqueue);
@@ -10754,28 +10932,31 @@ static int __init scx_init(void)
 	CID_OFFSET_MATCH(dump_task, dump_task);
 	CID_OFFSET_MATCH(sub_attach, sub_attach);
 	CID_OFFSET_MATCH(sub_detach, sub_detach);
+	CID_OFFSET_MATCH(sub_caps_updated, sub_caps_updated);
+	CID_OFFSET_MATCH(sub_ecaps_updated, sub_ecaps_updated);
+	CID_OFFSET_MATCH(init_cids, init_cids);
 	CID_OFFSET_MATCH(init, init);
 	CID_OFFSET_MATCH(exit, exit);
-#ifdef CONFIG_EXT_GROUP_SCHED
-	CID_OFFSET_MATCH(cgroup_init, cgroup_init);
-	CID_OFFSET_MATCH(cgroup_exit, cgroup_exit);
-	CID_OFFSET_MATCH(cgroup_prep_move, cgroup_prep_move);
-	CID_OFFSET_MATCH(cgroup_move, cgroup_move);
-	CID_OFFSET_MATCH(cgroup_cancel_move, cgroup_cancel_move);
-	CID_OFFSET_MATCH(cgroup_set_weight, cgroup_set_weight);
-	CID_OFFSET_MATCH(cgroup_set_bandwidth, cgroup_set_bandwidth);
-	CID_OFFSET_MATCH(cgroup_set_idle, cgroup_set_idle);
-#endif
 	/* renamed callbacks must occupy the same slot as their cpu-form sibling */
 	CID_OFFSET_MATCH(select_cpu, select_cid);
 	CID_OFFSET_MATCH(set_cpumask, set_cmask);
 	CID_OFFSET_MATCH(cpu_online, cid_online);
 	CID_OFFSET_MATCH(cpu_offline, cid_offline);
 	CID_OFFSET_MATCH(dump_cpu, dump_cid);
+#ifdef CONFIG_EXT_GROUP_SCHED
+	CID_OFFSET_MATCH(cgroup_init, cpuctl_init);
+	CID_OFFSET_MATCH(cgroup_exit, cpuctl_exit);
+	CID_OFFSET_MATCH(cgroup_prep_move, cpuctl_prep_move);
+	CID_OFFSET_MATCH(cgroup_move, cpuctl_move);
+	CID_OFFSET_MATCH(cgroup_cancel_move, cpuctl_cancel_move);
+	CID_OFFSET_MATCH(cgroup_set_weight, cpuctl_set_weight);
+	CID_OFFSET_MATCH(cgroup_set_bandwidth, cpuctl_set_bandwidth);
+	CID_OFFSET_MATCH(cgroup_set_idle, cpuctl_set_idle);
+#endif
 	/* @priv tail must align since both share the same data block */
 	CID_OFFSET_MATCH(priv, priv);
 	/*
-	 * cid-form must end exactly at @priv - validate_ops() skips
+	 * cid-form must end exactly at @priv - scx_validate_ops() skips
 	 * cpu_acquire/cpu_release for cid-form because reading those fields
 	 * past the BPF allocation would be UB.
 	 */
diff --git a/kernel/sched/ext/idle.c b/kernel/sched/ext/idle.c
index 6f93cc32b650..d2973fb3af6d 100644
--- a/kernel/sched/ext/idle.c
+++ b/kernel/sched/ext/idle.c
@@ -12,6 +12,7 @@
 #include "internal.h"
 #include "cid.h"
 #include "idle.h"
+#include "sub.h"
 
 /* Enable/disable built-in idle CPU selection policy */
 static DEFINE_STATIC_KEY_FALSE(scx_builtin_idle_enabled);
@@ -733,6 +734,55 @@ static void update_builtin_idle(int cpu, bool idle)
 }
 
 /*
+ * Notify schedulers of an idle transition on @cpu's cid, delivering to every
+ * sched that holds %SCX_CAP_BASE on the cid (the root holds every cap). A real
+ * transition (@do_notify) reaches all holders. A forced one (@root_renotify for
+ * the root, a sub-sched's idle_renotify marker for a sub) reaches only the owed
+ * scheds.
+ */
+static void scx_idle_notify(struct rq *rq, bool idle, bool do_notify, bool root_renotify)
+{
+	s32 cpu = cpu_of(rq);
+	s32 cid = scx_cpu_arg(cpu);
+	struct scx_sched *root = scx_root_protected_live();
+	struct scx_sched *pos;
+
+	lockdep_assert_rq_held(rq);
+
+	/* with no sub-sched, only the root can be owed a notification */
+	if (!scx_has_subs()) {
+		if ((do_notify || root_renotify) &&
+		    SCX_HAS_OP(root, update_idle) && !scx_bypassing(root, cpu))
+			SCX_CALL_OP(root, update_idle, rq, cid, idle);
+		return;
+	}
+
+	pos = scx_next_descendant_pre(NULL, root);
+	while (pos) {
+		bool forced = false;
+
+		if (unlikely(scx_missing_caps(pos, cpu, SCX_CAP_BASE))) {
+			pos = scx_skip_subtree_pre(pos, root);
+			continue;
+		}
+
+		if (!pos->level) {
+			forced = root_renotify;
+		}
+#ifdef CONFIG_EXT_SUB_SCHED
+		else if (per_cpu_ptr(pos->pcpu, cpu)->idle_renotify) {
+			per_cpu_ptr(pos->pcpu, cpu)->idle_renotify = false;
+			forced = true;
+		}
+#endif
+		if ((do_notify || forced) && SCX_HAS_OP(pos, update_idle) &&
+		    !scx_bypassing(pos, cpu))
+			SCX_CALL_OP(pos, update_idle, rq, cid, idle);
+		pos = scx_next_descendant_pre(pos, root);
+	}
+}
+
+/*
  * Update the idle state of a CPU to @idle.
  *
  * If @do_notify is true, ops.update_idle() is invoked to notify the scx
@@ -750,44 +800,39 @@ static void update_builtin_idle(int cpu, bool idle)
  */
 void __scx_update_idle(struct rq *rq, bool idle, bool do_notify)
 {
-	struct scx_sched *sch = scx_root;
 	int cpu = cpu_of(rq);
 
 	lockdep_assert_rq_held(rq);
 
 	/*
-	 * Update the idle masks:
-	 * - for real idle transitions (do_notify == true)
-	 * - for idle-to-idle transitions (indicated by the previous task
-	 *   being the idle thread, managed by pick_task_idle())
-	 *
-	 * Skip updating idle masks if the previous task is not the idle
-	 * thread, since set_next_task_idle() has already handled it when
-	 * transitioning from a task to the idle thread (calling this
-	 * function with do_notify == true).
-	 *
-	 * In this way we can avoid updating the idle masks twice,
-	 * unnecessarily.
+	 * pick_task_idle() calls here only on an idle-to-idle re-pick and the
+	 * transitions call with @do_notify, so every reaching call updates the
+	 * masks.
 	 */
 	if (static_branch_likely(&scx_builtin_idle_enabled))
-		if (do_notify || is_idle_task(rq->curr))
-			update_builtin_idle(cpu, idle);
+		update_builtin_idle(cpu, idle);
 
 	/*
-	 * Trigger ops.update_idle() only when transitioning from a task to
-	 * the idle thread and vice versa.
+	 * ops.update_idle() fires on real idle transitions, indicated by
+	 * @do_notify and managed by put_prev_task_idle()/set_next_task_idle().
+	 * An idle pick also fires it to flush a forced notify owed to a sched
+	 * that missed transitions while bypassed or on a cid it just gained.
+	 * unbypass_renotify_idle() and scx_process_sync_ecaps() arm the per-rq
+	 * gates, and scx_idle_notify() targets the owed scheds.
 	 *
-	 * Idle transitions are indicated by do_notify being set to true,
-	 * managed by put_prev_task_idle()/set_next_task_idle().
-	 *
-	 * This must come after builtin idle update so that BPF schedulers can
-	 * create interlocking between ops.update_idle() and ops.enqueue() -
+	 * This must come after the builtin idle update so that BPF schedulers
+	 * can create interlocking between ops.update_idle() and ops.enqueue() -
 	 * either enqueue() sees the idle bit or update_idle() sees the task
 	 * that enqueue() queued.
 	 */
-	if (SCX_HAS_OP(sch, update_idle) && do_notify &&
-	    !scx_bypassing(sch, cpu_of(rq)))
-		SCX_CALL_OP(sch, update_idle, rq, scx_cpu_arg(cpu_of(rq)), idle);
+	if (do_notify ||
+	    (idle && (rq->scx.flags &
+		      (SCX_RQ_SUB_IDLE_RENOTIFY | SCX_RQ_ROOT_IDLE_RENOTIFY)))) {
+		bool root_renotify = rq->scx.flags & SCX_RQ_ROOT_IDLE_RENOTIFY;
+
+		rq->scx.flags &= ~(SCX_RQ_SUB_IDLE_RENOTIFY | SCX_RQ_ROOT_IDLE_RENOTIFY);
+		scx_idle_notify(rq, idle, do_notify, root_renotify);
+	}
 }
 
 static void reset_idle_masks(struct sched_ext_ops *ops)
@@ -795,20 +840,20 @@ static void reset_idle_masks(struct sched_ext_ops *ops)
 	int node;
 
 	/*
-	 * Consider all online cpus idle. Should converge to the actual state
-	 * quickly.
+	 * Start with all CPUs marked busy. The idle masks are populated when
+	 * bypass is lifted and each idle CPU is forced through an idle re-pick.
+	 * This may temporarily omit idle CPUs but never advertises a busy CPU as
+	 * idle.
 	 */
 	if (!(ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)) {
-		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->cpu, cpu_online_mask);
-		cpumask_copy(idle_cpumask(NUMA_NO_NODE)->smt, cpu_online_mask);
+		cpumask_clear(idle_cpumask(NUMA_NO_NODE)->cpu);
+		cpumask_clear(idle_cpumask(NUMA_NO_NODE)->smt);
 		return;
 	}
 
 	for_each_node(node) {
-		const struct cpumask *node_mask = cpumask_of_node(node);
-
-		cpumask_and(idle_cpumask(node)->cpu, cpu_online_mask, node_mask);
-		cpumask_and(idle_cpumask(node)->smt, cpu_online_mask, node_mask);
+		cpumask_clear(idle_cpumask(node)->cpu);
+		cpumask_clear(idle_cpumask(node)->smt);
 	}
 }
 
@@ -1318,7 +1363,7 @@ __bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(const struct cpumask *cpus_allowed,
 /**
  * scx_bpf_pick_idle_cpu - Pick and claim an idle cpu
  * @cpus_allowed: Allowed cpumask
- * @flags: %SCX_PICK_IDLE_CPU_* flags
+ * @flags: %SCX_PICK_IDLE_* flags
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
  * Pick and claim an idle cpu in @cpus_allowed. Returns the picked idle cpu
@@ -1365,7 +1410,7 @@ __bpf_kfunc s32 scx_bpf_pick_idle_cpu(const struct cpumask *cpus_allowed,
  *			       or pick any CPU from @node
  * @cpus_allowed: Allowed cpumask
  * @node: target NUMA node
- * @flags: %SCX_PICK_IDLE_CPU_* flags
+ * @flags: %SCX_PICK_IDLE_* flags
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
@@ -1416,7 +1461,7 @@ __bpf_kfunc s32 scx_bpf_pick_any_cpu_node(const struct cpumask *cpus_allowed,
 /**
  * scx_bpf_pick_any_cpu - Pick and claim an idle cpu if available or pick any CPU
  * @cpus_allowed: Allowed cpumask
- * @flags: %SCX_PICK_IDLE_CPU_* flags
+ * @flags: %SCX_PICK_IDLE_* flags
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
  * Pick and claim an idle cpu in @cpus_allowed. If none is available, pick any
diff --git a/kernel/sched/ext/inlines.h b/kernel/sched/ext/inlines.h
new file mode 100644
index 000000000000..ed423bcc26b8
--- /dev/null
+++ b/kernel/sched/ext/inlines.h
@@ -0,0 +1,138 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
+ *
+ * Inline definitions layered on top of internal.h and cid.h.
+ *
+ * Copyright (c) 2026 Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2026 Tejun Heo <tj@kernel.org>
+ */
+#ifndef _KERNEL_SCHED_EXT_INLINES_H
+#define _KERNEL_SCHED_EXT_INLINES_H
+
+#include "internal.h"
+#include "cid.h"
+
+/* what dispatch concluded, consumed by the pick that follows */
+enum scx_dsp_verdict {
+	SCX_DSP_NONE,		/* nothing to run */
+	SCX_DSP_LOCAL,		/* local DSQ has tasks */
+	SCX_DSP_PREV,		/* keep running @prev */
+	SCX_DSP_RETRY,		/* pick helpers only: restart the pick */
+};
+
+/*
+ * One user of this function is scx_bpf_sub_dispatch() which can be called
+ * recursively as sub-sched dispatches nest. Always inline to reduce stack usage
+ * from the call frame.
+ */
+static __always_inline enum scx_dsp_verdict
+scx_dispatch_sched(struct scx_sched *sch, struct rq *rq,
+		   struct task_struct *prev, bool nested)
+{
+	struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx;
+	int nr_loops = SCX_DSP_MAX_LOOPS;
+	s32 cpu = cpu_of(rq);
+	bool prev_on_sch = (prev->sched_class == &ext_sched_class) &&
+		scx_task_on_sched(sch, prev);
+
+	if (scx_consume_global_dsq(sch, rq))
+		return SCX_DSP_LOCAL;
+
+	if (scx_bypass_dsp_enabled(sch)) {
+		/* if @sch is bypassing, only the bypass DSQs are active */
+		if (scx_bypassing(sch, cpu)) {
+			if (scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0))
+				return SCX_DSP_LOCAL;
+			return SCX_DSP_NONE;
+		}
+
+#ifdef CONFIG_EXT_SUB_SCHED
+		/*
+		 * If @sch isn't bypassing but its children are, @sch is
+		 * responsible for making forward progress for both its own
+		 * tasks that aren't bypassing and the bypassing descendants'
+		 * tasks. The following implements a simple built-in behavior -
+		 * let each CPU try to run the bypass DSQ every Nth time.
+		 *
+		 * Later, if necessary, we can add an ops flag to suppress the
+		 * auto-consumption and a kfunc to consume the bypass DSQ and,
+		 * so that the BPF scheduler can fully control scheduling of
+		 * bypassed tasks.
+		 */
+		struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
+
+		if (!(pcpu->bypass_host_seq++ % SCX_BYPASS_HOST_NTH) &&
+		    scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0)) {
+			__scx_add_event(sch, SCX_EV_SUB_BYPASS_DISPATCH, 1);
+			return SCX_DSP_LOCAL;
+		}
+#endif	/* CONFIG_EXT_SUB_SCHED */
+	}
+
+	if (unlikely(!SCX_HAS_OP(sch, dispatch)) || !scx_rq_online(rq))
+		return SCX_DSP_NONE;
+
+	dspc->rq = rq;
+
+	/*
+	 * The dispatch loop. Because scx_flush_dispatch_buf() may drop the rq
+	 * lock, the local DSQ might still end up empty after a successful
+	 * ops.dispatch(). If the local DSQ is empty even after ops.dispatch()
+	 * produced some tasks, retry. The BPF scheduler may depend on this
+	 * looping behavior to simplify its implementation.
+	 */
+	do {
+		dspc->nr_tasks = 0;
+
+#ifdef CONFIG_EXT_SUB_SCHED
+		/* stash @prev so that nested invocations can access it */
+		if (!nested)
+			rq->scx.sub_dispatch_prev = prev;
+#endif
+
+		SCX_CALL_OP(sch, dispatch, rq, scx_cpu_arg(cpu),
+			    prev_on_sch ? prev : NULL);
+
+#ifdef CONFIG_EXT_SUB_SCHED
+		if (!nested)
+			rq->scx.sub_dispatch_prev = NULL;
+#endif
+
+		scx_flush_dispatch_buf(sch, rq);
+
+		if ((prev->scx.flags & SCX_TASK_QUEUED) && prev->scx.slice)
+			return SCX_DSP_PREV;
+		if (rq->scx.local_dsq.nr)
+			return SCX_DSP_LOCAL;
+		if (scx_consume_global_dsq(sch, rq))
+			return SCX_DSP_LOCAL;
+
+		/*
+		 * ops.dispatch() can trap us in this loop by repeatedly
+		 * dispatching ineligible tasks. Break out once in a while to
+		 * allow the watchdog to run. As IRQ can't be enabled in
+		 * dispatch, we want to complete this scheduling cycle and then
+		 * start a new one. IOW, we want to call resched_curr() on the
+		 * next, most likely idle, task, not the current one. Use
+		 * __scx_bpf_kick_cpu() for deferred kicking.
+		 */
+		if (unlikely(!--nr_loops)) {
+			scx_kick_cpu(sch, cpu, 0);
+			break;
+		}
+	} while (dspc->nr_tasks);
+
+	/*
+	 * Prevent the CPU from going idle while bypassed descendants have tasks
+	 * queued. Without this fallback, bypassed tasks could stall if the host
+	 * scheduler's ops.dispatch() doesn't yield any tasks.
+	 */
+	if (scx_bypass_dsp_enabled(sch) &&
+	    scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0))
+		return SCX_DSP_LOCAL;
+
+	return SCX_DSP_NONE;
+}
+
+#endif /* _KERNEL_SCHED_EXT_INLINES_H */
diff --git a/kernel/sched/ext/internal.h b/kernel/sched/ext/internal.h
index 673059fa9d72..27bbf5e04d90 100644
--- a/kernel/sched/ext/internal.h
+++ b/kernel/sched/ext/internal.h
@@ -11,6 +11,34 @@
 #include "../sched.h"
 #include "types.h"
 
+#include <trace/events/sched_ext.h>
+
+/**
+ * scx_add_event - Increase an event counter for 'name' by 'cnt'
+ * @sch: scx_sched to account events for
+ * @name: an event name defined in struct scx_event_stats
+ * @cnt: the number of the event occurred
+ *
+ * This can be used when preemption is not disabled.
+ */
+#define scx_add_event(sch, name, cnt) do {					\
+	this_cpu_add((sch)->pcpu->event_stats.name, (cnt));			\
+	trace_sched_ext_event(#name, (cnt));					\
+} while(0)
+
+/**
+ * __scx_add_event - Increase an event counter for 'name' by 'cnt'
+ * @sch: scx_sched to account events for
+ * @name: an event name defined in struct scx_event_stats
+ * @cnt: the number of the event occurred
+ *
+ * This should be used only when preemption is disabled.
+ */
+#define __scx_add_event(sch, name, cnt) do {					\
+	__this_cpu_add((sch)->pcpu->event_stats.name, (cnt));			\
+	trace_sched_ext_event(#name, cnt);					\
+} while(0)
+
 #define SCX_OP_IDX(op)		(offsetof(struct sched_ext_ops, op) / sizeof(void (*)(void)))
 #define SCX_MOFF_IDX(moff)	((moff) / sizeof(void (*)(void)))
 
@@ -23,10 +51,13 @@ enum scx_exit_kind {
 	SCX_EXIT_UNREG_KERN,	/* kernel-initiated unregistration */
 	SCX_EXIT_SYSRQ,		/* requested by 'S' sysrq */
 	SCX_EXIT_PARENT,	/* parent exiting */
+	SCX_EXIT_PARENT_KILL,	/* killed by parent scheduler */
 
 	SCX_EXIT_ERROR = 1024,	/* runtime error, error msg contains details */
 	SCX_EXIT_ERROR_BPF,	/* ERROR but triggered through scx_bpf_error() */
 	SCX_EXIT_ERROR_STALL,	/* watchdog detected stalled runnable tasks */
+	SCX_EXIT_ERROR_REENQ,	/* task hit reenqueue limit without running */
+	SCX_EXIT_ERROR_RESCUE,	/* ejected for overloading rescue execution */
 };
 
 /*
@@ -372,8 +403,9 @@ struct sched_ext_ops {
 	 * @p: task running currently
 	 *
 	 * This operation is called every 1/HZ seconds on CPUs which are
-	 * executing an SCX task. Setting @p->scx.slice to 0 will trigger an
-	 * immediate dispatch cycle on the CPU.
+	 * executing an SCX task. Setting a slice of 0 for @p with
+	 * scx_bpf_task_set_slice() will trigger an immediate dispatch cycle on
+	 * the CPU.
 	 */
 	void (*tick)(struct task_struct *p);
 
@@ -453,7 +485,7 @@ struct sched_ext_ops {
 	 * - sleeping (%SCX_DEQ_SLEEP)
 	 * - being moved to another CPU
 	 * - being temporarily taken off the queue for an attribute change
-	 *   (%SCX_DEQ_SAVE)
+	 *   (%SCX_DEQ_SCHED_CHANGE)
 	 *
 	 * This and ->dequeue() are related but not coupled. This operation
 	 * notifies @p's state transition and may not be preceded by ->dequeue()
@@ -489,6 +521,11 @@ struct sched_ext_ops {
 	 * the BPF scheduler. Should return %true if @a should run before @b.
 	 * %false if there's no required ordering or @b should run before @a.
 	 *
+	 * In a scheduler hierarchy, a pair spanning two schedulers is ordered
+	 * by the nearest common ancestor implementing this op, so the op may be
+	 * called on tasks that the scheduler delegated to its sub-schedulers
+	 * and is not scheduling anymore. See scx_prio_less().
+	 *
 	 * If not specified, the default is ordering them according to when they
 	 * became runnable.
 	 */
@@ -615,8 +652,19 @@ struct sched_ext_ops {
 	 * @cgrp: cgroup being initialized
 	 * @args: init arguments, see the struct definition
 	 *
-	 * Either the BPF scheduler is being loaded or @cgrp created, initialize
-	 * @cgrp for sched_ext. This operation may block.
+	 * Initialize @cgrp for sched_ext, delivered to @cgrp's sched either
+	 * when the BPF scheduler is being loaded or when @cgrp is created. This
+	 * operation may block.
+	 *
+	 * Cgroup handovers also generate these ops: an enabling sub-scheduler
+	 * receives ops.cgroup_init() for every cgroup in its subtree while the
+	 * previous sched receives ops.cgroup_exit(), and disabling reverses the
+	 * two.
+	 *
+	 * When the BPF scheduler is being loaded or cgroups are being handed
+	 * over, @cgrp may already have been removed by userspace: a removed
+	 * cgroup stays schedulable until its dying tasks finish their final
+	 * context switches.
 	 *
 	 * Return 0 for success, -errno for failure. An error return while
 	 * loading will abort loading of the BPF scheduler. During cgroup
@@ -629,8 +677,13 @@ struct sched_ext_ops {
 	 * @cgroup_exit: Exit a cgroup
 	 * @cgrp: cgroup being exited
 	 *
-	 * Either the BPF scheduler is being unloaded or @cgrp destroyed, exit
-	 * @cgrp for sched_ext. This operation my block.
+	 * Exit @cgrp for sched_ext, delivered to the sched whose
+	 * ops.cgroup_init() it pairs with, either when the BPF scheduler is
+	 * being unloaded or when @cgrp is destroyed. This operation may block.
+	 *
+	 * For a destroyed @cgrp, delivery follows the last scheduling event on
+	 * it: a removed cgroup stays schedulable until its dying tasks finish
+	 * their final context switches.
 	 */
 	void (*cgroup_exit)(struct cgroup *cgrp);
 
@@ -643,6 +696,12 @@ struct sched_ext_ops {
 	 * Prepare @p for move from cgroup @from to @to. This operation may
 	 * block and can be used for allocations.
 	 *
+	 * The cgroup_move ops are delivered to @p's sched, and only for moves
+	 * that don't re-home @p. A re-homing move is reported through
+	 * ops.exit_task() and ops.init_task() instead. @from and @to can
+	 * reference cgroups the sched never received ops.cgroup_init() for, as
+	 * the cpu controller can be coarser than the sub-scheduler topology.
+	 *
 	 * Return 0 for success, -errno for failure. An error return aborts the
 	 * migration.
 	 */
@@ -678,6 +737,11 @@ struct sched_ext_ops {
 	 * @weight: new weight [1..10000]
 	 *
 	 * Update @cgrp's weight to @weight.
+	 *
+	 * Knobs of a cgroup belong to the parent, so the set_* ops are
+	 * delivered to @cgrp's parent's sched. That sched may never have seen
+	 * ops.cgroup_init() for @cgrp - at a sub-scheduler attach point, the
+	 * parent sched tracks @cgrp through ops.sub_attach() instead.
 	 */
 	void (*cgroup_set_weight)(struct cgroup *cgrp, u32 weight);
 
@@ -698,6 +762,8 @@ struct sched_ext_ops {
 	 * burst temporarily. The specific control mechanism and thus the
 	 * interpretation of @period_us and burstiness is up to the BPF
 	 * scheduler.
+	 *
+	 * Delivery follows the same rule as cgroup_set_weight().
 	 */
 	void (*cgroup_set_bandwidth)(struct cgroup *cgrp,
 				     u64 period_us, u64 quota_us, u64 burst_us);
@@ -710,6 +776,8 @@ struct sched_ext_ops {
 	 * Update @cgrp's idle state to @idle. This callback is invoked when
 	 * a cgroup transitions between idle and non-idle states, allowing the
 	 * BPF scheduler to adjust its behavior accordingly.
+	 *
+	 * Delivery follows the same rule as cgroup_set_weight().
 	 */
 	void (*cgroup_set_idle)(struct cgroup *cgrp, bool idle);
 
@@ -729,6 +797,39 @@ struct sched_ext_ops {
 	 */
 	void (*sub_detach)(struct scx_sub_detach_args *args);
 
+	/**
+	 * @sub_caps_updated: Caps on this sub-sched's shard changed
+	 * @cmask: cids whose caps changed (cmask->base identifies the shard)
+	 * @caps: SCX_CAP_* that changed
+	 *
+	 * Invoked after grant or revoke modifies caps on a shard. There can be
+	 * only one in-flight invocation per shard. @cmask and @caps coalesce
+	 * all changes since the last delivery. Direction (set vs cleared) isn't
+	 * encoded. Query current state with scx_bpf_sub_caps().
+	 *
+	 * Delivered asynchronously after the change is recorded, and may run
+	 * before it takes effect on any given cpu. Use it to track which caps
+	 * the sub-sched holds and propagate to its own children, not to decide
+	 * if a task can run on a cpu now. sub_ecaps_updated() reports that per
+	 * cpu, once it is in effect.
+	 *
+	 * May call scx_bpf_sub_grant() / scx_bpf_sub_revoke() on children.
+	 */
+	void (*sub_caps_updated)(const struct scx_cmask *cmask, u64 caps);
+
+	/**
+	 * @sub_ecaps_updated: This sub-sched's effective caps on a cid changed
+	 * @cid: the cid whose effective caps changed
+	 * @before: effective caps as of the last delivery
+	 * @after: effective caps now
+	 *
+	 * Invoked when this sub-sched's effective caps on @cid change, once the
+	 * change is in effect on the cpu. Runs in dispatch context with rq lock
+	 * held, and can perform all operations allowed in ops.dispatch()
+	 * including inserting/moving tasks.
+	 */
+	void (*sub_ecaps_updated)(s32 cid, u64 before, u64 after);
+
 	/*
 	 * All online ops must come before ops.cpu_online().
 	 */
@@ -752,10 +853,19 @@ struct sched_ext_ops {
 	void (*cpu_offline)(s32 cpu);
 
 	/*
-	 * All CPU hotplug ops must come before ops.init().
+	 * All CPU hotplug ops must come before ops.init_cids().
 	 */
 
 	/**
+	 * @init_cids: Finalize the cid layout (cid-form only)
+	 *
+	 * Runs after the default cid layout is built, before caps and shards
+	 * are finalized. A cid-form scheduler may call scx_bpf_cid_override()
+	 * here for a custom layout. Ignored for cpu-form schedulers.
+	 */
+	s32 (*init_cids)(void);
+
+	/**
 	 * @init: Initialize the BPF scheduler
 	 */
 	s32 (*init)(void);
@@ -809,8 +919,51 @@ struct sched_ext_ops {
 	u64 hotplug_seq;
 
 	/**
-	 * @cgroup_id: When >1, attach the scheduler as a sub-scheduler on the
-	 * specified cgroup.
+	 * @cid_shard_size: Target number of CIDs per shard
+	 *
+	 * Shards are contiguous CID ranges used as operation and locking
+	 * domains for sub-scheduling. Each LLC is divided into ceil(nr_cpus /
+	 * @cid_shard_size) shards, then cores are distributed across them
+	 * evenly. If one core has more logical CPUs than @cid_shard_size, its
+	 * shard will become larger than @cid_shard_size. Values above
+	 * SCX_CID_SHARD_MAX_CPUS are capped. 0 means use the default (24).
+	 */
+	u32 cid_shard_size;
+
+	/**
+	 * @rescue_bandwidth_ppt: Rescue execution bandwidth in parts per thousand
+	 *
+	 * The fraction of each CPU's time that may be consumed running tasks
+	 * from its rescue DSQ. A higher bandwidth admits and escalates rescues
+	 * faster, see @rescue_quantum_us.
+	 *
+	 * Only the root scheduler's value is used. 0 means the default of 20
+	 * (2%). May not exceed 250 (25%). %SCX_RESCUE_DISABLE disables rescue -
+	 * %SCX_ENQ_RESCUE inserts are then rejected like any other insert
+	 * lacking the caps.
+	 */
+	u32 rescue_bandwidth_ppt;
+
+	/**
+	 * @rescue_quantum_us: Rescue execution quantum in microseconds
+	 *
+	 * How much CPU time each rescue gets. Rescues run one at a time per CPU
+	 * and admissions are paced to keep rescue execution within
+	 * @rescue_bandwidth_ppt - with the defaults, one 5ms rescue every
+	 * 250ms. A crowded queue round-robins on the quantum divided across the
+	 * waiters, floored at 1ms. A stuck rescue eventually escalates to
+	 * forced execution. A larger quantum interrupts the CPU less often but
+	 * for longer and spaces rescues further apart.
+	 *
+	 * Only the root scheduler's value is used. 0 means the default (5000).
+	 * Non-zero values must be within [1000, 100000]. Values too short for
+	 * the kernel to meter are lifted silently.
+	 */
+	u32 rescue_quantum_us;
+
+	/**
+	 * @sub_cgroup_id: When >1, attach the scheduler as a sub-scheduler
+	 * on the specified cgroup.
 	 */
 	u64 sub_cgroup_id;
 
@@ -818,8 +971,9 @@ struct sched_ext_ops {
 	 * @name: BPF scheduler's name
 	 *
 	 * Must be a non-zero valid BPF object name including only isalnum(),
-	 * '_' and '.' chars. Shows up in kernel.sched_ext_ops sysctl while the
-	 * BPF scheduler is enabled.
+	 * '_' and '.' chars. Exposed via the ops file in the scheduler's sysfs
+	 * directory, /sys/kernel/sched_ext/root/ops for the root scheduler,
+	 * while the BPF scheduler is enabled.
 	 */
 	char name[SCX_OPS_NAME_LEN];
 
@@ -877,6 +1031,7 @@ struct sched_ext_ops {
  *   - cpu_online       -> cid_online
  *   - cpu_offline      -> cid_offline
  *   - dump_cpu         -> dump_cid
+ *   - cgroup_*         -> cpuctl_* (they track the cgroup cpu controller)
  *   - cpu_acquire/cpu_release  -> not present (deprecated in sched_ext_ops)
  *
  * BPF schedulers using this type cannot call cpu-form scx_bpf_* kfuncs;
@@ -900,7 +1055,7 @@ struct sched_ext_ops_cid {
 				   struct task_struct *b);
 	void (*set_weight)(struct task_struct *p, u32 weight);
 	void (*set_cmask)(struct task_struct *p,
-			   const struct scx_cmask *cmask);
+			   const struct scx_cmask *cmask__arena);
 	void (*update_idle)(s32 cid, bool idle);
 	s32 (*init_task)(struct task_struct *p,
 			  struct scx_init_task_args *args);
@@ -912,24 +1067,25 @@ struct sched_ext_ops_cid {
 	void (*dump_cid)(struct scx_dump_ctx *ctx, s32 cid, bool idle);
 	void (*dump_task)(struct scx_dump_ctx *ctx, struct task_struct *p);
 #ifdef CONFIG_EXT_GROUP_SCHED
-	s32 (*cgroup_init)(struct cgroup *cgrp,
-			    struct scx_cgroup_init_args *args);
-	void (*cgroup_exit)(struct cgroup *cgrp);
-	s32 (*cgroup_prep_move)(struct task_struct *p,
-				 struct cgroup *from, struct cgroup *to);
-	void (*cgroup_move)(struct task_struct *p,
-			     struct cgroup *from, struct cgroup *to);
-	void (*cgroup_cancel_move)(struct task_struct *p,
-				    struct cgroup *from, struct cgroup *to);
-	void (*cgroup_set_weight)(struct cgroup *cgrp, u32 weight);
-	void (*cgroup_set_bandwidth)(struct cgroup *cgrp,
-				      u64 period_us, u64 quota_us, u64 burst_us);
-	void (*cgroup_set_idle)(struct cgroup *cgrp, bool idle);
+	s32 (*cpuctl_init)(struct cgroup *cgrp, struct scx_cgroup_init_args *args);
+	void (*cpuctl_exit)(struct cgroup *cgrp);
+	s32 (*cpuctl_prep_move)(struct task_struct *p, struct cgroup *from,
+				struct cgroup *to);
+	void (*cpuctl_move)(struct task_struct *p, struct cgroup *from, struct cgroup *to);
+	void (*cpuctl_cancel_move)(struct task_struct *p, struct cgroup *from,
+				   struct cgroup *to);
+	void (*cpuctl_set_weight)(struct cgroup *cgrp, u32 weight);
+	void (*cpuctl_set_bandwidth)(struct cgroup *cgrp, u64 period_us, u64 quota_us,
+				     u64 burst_us);
+	void (*cpuctl_set_idle)(struct cgroup *cgrp, bool idle);
 #endif	/* CONFIG_EXT_GROUP_SCHED */
 	s32 (*sub_attach)(struct scx_sub_attach_args *args);
 	void (*sub_detach)(struct scx_sub_detach_args *args);
+	void (*sub_caps_updated)(const struct scx_cmask *cmask__arena, u64 caps);
+	void (*sub_ecaps_updated)(s32 cid, u64 before, u64 after);
 	void (*cid_online)(s32 cid);
 	void (*cid_offline)(s32 cid);
+	s32 (*init_cids)(void);
 	s32 (*init)(void);
 	void (*exit)(struct scx_exit_info *info);
 
@@ -939,6 +1095,9 @@ struct sched_ext_ops_cid {
 	u32 timeout_ms;
 	u32 exit_dump_len;
 	u64 hotplug_seq;
+	u32 cid_shard_size;
+	u32 rescue_bandwidth_ppt;
+	u32 rescue_quantum_us;
 	u64 sub_cgroup_id;
 	char name[SCX_OPS_NAME_LEN];
 
@@ -954,8 +1113,8 @@ enum scx_opi {
 	SCX_OPI_NORMAL_BEGIN		= 0,
 	SCX_OPI_NORMAL_END		= SCX_OP_IDX(cpu_online),
 	SCX_OPI_CPU_HOTPLUG_BEGIN	= SCX_OP_IDX(cpu_online),
-	SCX_OPI_CPU_HOTPLUG_END		= SCX_OP_IDX(init),
-	SCX_OPI_END			= SCX_OP_IDX(init),
+	SCX_OPI_CPU_HOTPLUG_END		= SCX_OP_IDX(init_cids),
+	SCX_OPI_END			= SCX_OP_IDX(init_cids),
 };
 
 /*
@@ -1001,15 +1160,13 @@ struct scx_event_stats {
 	s64		SCX_EV_REENQ_IMMED;
 
 	/*
-	 * The number of times a reenq of local DSQ caused another reenq of
-	 * local DSQ. This can happen when %SCX_ENQ_IMMED races against a higher
-	 * priority class task even if the BPF scheduler always satisfies the
-	 * prerequisites for %SCX_ENQ_IMMED at the time of enqueue. However,
-	 * that scenario is very unlikely and this count going up regularly
-	 * indicates that the BPF scheduler is handling %SCX_ENQ_REENQ
-	 * incorrectly causing recursive reenqueues.
+	 * The number of times a reenqueue (%SCX_ENQ_REENQ) led to another
+	 * reenqueue without the task running in between. This count climbing
+	 * rapidly indicates that the BPF scheduler keeps re-deciding placements
+	 * it can't honor. A single task reenqueued more than
+	 * %SCX_REENQ_MAX_REPEAT times gets its owning scheduler ejected.
 	 */
-	s64		SCX_EV_REENQ_LOCAL_REPEAT;
+	s64		SCX_EV_REENQ_REPEAT;
 
 	/*
 	 * Total number of times a task's time slice was refilled with the
@@ -1018,6 +1175,18 @@ struct scx_event_stats {
 	s64		SCX_EV_REFILL_SLICE_DFL;
 
 	/*
+	 * The number of times an out-of-band slice request exceeded the maximum
+	 * representable value and was clamped.
+	 */
+	s64		SCX_EV_SLICE_CLAMPED;
+
+	/*
+	 * The number of times a slice extension was denied because the
+	 * scheduler lacked baseline cpu access on the task's cpu.
+	 */
+	s64		SCX_EV_SLICE_DENIED;
+
+	/*
 	 * The total duration of bypass modes in nanoseconds.
 	 */
 	s64		SCX_EV_BYPASS_DURATION;
@@ -1049,8 +1218,70 @@ struct scx_event_stats {
 	 * from sub_bypass_dsq's.
 	 */
 	s64		SCX_EV_SUB_BYPASS_DISPATCH;
+
+	/*
+	 * The number of times a migration-disabled task lacking the cap for its
+	 * cid was allowed onto the local DSQ. It must run on its pinned CPU, so
+	 * it can't be rejected. The violation is counted here.
+	 */
+	s64		SCX_EV_SUB_FORCED_ADMIT;
+
+	/*
+	 * The number of times a preempting kick was refused because the
+	 * sub-sched lacked SCX_CAP_PREEMPT for a task outside its subtree. The
+	 * kick degrades to a plain reschedule.
+	 */
+	s64		SCX_EV_SUB_PREEMPT_DENIED;
+
+	/*
+	 * The number of times a kick was skipped because the sub-sched lacked
+	 * baseline access on the target cid. The preempt-part degradation of a
+	 * delivered kick is counted in SCX_EV_SUB_PREEMPT_DENIED instead.
+	 */
+	s64		SCX_EV_SUB_KICK_DENIED;
+
+	/*
+	 * The number of times a local DSQ reenq was dropped because the
+	 * sub-sched lacked baseline access on the target cid.
+	 */
+	s64		SCX_EV_SUB_REENQ_DENIED;
+
+	/*
+	 * The number of times scx_bpf_cidperf_set() was denied because the
+	 * sub-sched lacked SCX_CAP_PERF on the target cid.
+	 */
+	s64		SCX_EV_SUB_CIDPERF_DENIED;
+
+	/*
+	 * The number of times an insert carrying %SCX_ENQ_RESCUE lacked the
+	 * caps for its cid and the task entered the rescue path.
+	 */
+	s64		SCX_EV_SUB_RESCUE;
 };
 
+#define SCX_EVENTS_LIST(SCX_EVENT)					\
+	SCX_EVENT(SCX_EV_SELECT_CPU_FALLBACK);				\
+	SCX_EVENT(SCX_EV_DISPATCH_LOCAL_DSQ_OFFLINE);			\
+	SCX_EVENT(SCX_EV_DISPATCH_KEEP_LAST);				\
+	SCX_EVENT(SCX_EV_ENQ_SKIP_EXITING);				\
+	SCX_EVENT(SCX_EV_ENQ_SKIP_MIGRATION_DISABLED);			\
+	SCX_EVENT(SCX_EV_REENQ_IMMED);					\
+	SCX_EVENT(SCX_EV_REENQ_REPEAT);					\
+	SCX_EVENT(SCX_EV_REFILL_SLICE_DFL);				\
+	SCX_EVENT(SCX_EV_SLICE_CLAMPED);				\
+	SCX_EVENT(SCX_EV_SLICE_DENIED);					\
+	SCX_EVENT(SCX_EV_BYPASS_DURATION);				\
+	SCX_EVENT(SCX_EV_BYPASS_DISPATCH);				\
+	SCX_EVENT(SCX_EV_BYPASS_ACTIVATE);				\
+	SCX_EVENT(SCX_EV_INSERT_NOT_OWNED);				\
+	SCX_EVENT(SCX_EV_SUB_BYPASS_DISPATCH);				\
+	SCX_EVENT(SCX_EV_SUB_FORCED_ADMIT);				\
+	SCX_EVENT(SCX_EV_SUB_PREEMPT_DENIED);				\
+	SCX_EVENT(SCX_EV_SUB_KICK_DENIED);				\
+	SCX_EVENT(SCX_EV_SUB_REENQ_DENIED);				\
+	SCX_EVENT(SCX_EV_SUB_CIDPERF_DENIED);				\
+	SCX_EVENT(SCX_EV_SUB_RESCUE)
+
 struct scx_sched;
 
 enum scx_sched_pcpu_flags {
@@ -1062,6 +1293,8 @@ struct scx_dsp_buf_ent {
 	struct task_struct	*task;
 	unsigned long		qseq;
 	u64			dsq_id;
+	u64			slice;
+	u64			vtime;
 	u64			enq_flags;
 };
 
@@ -1075,8 +1308,6 @@ struct scx_dsp_ctx {
 struct scx_deferred_reenq_local {
 	struct list_head	node;
 	u64			flags;
-	u64			seq;
-	u32			cnt;
 };
 
 struct scx_sched_pcpu {
@@ -1084,6 +1315,49 @@ struct scx_sched_pcpu {
 	u64			flags;	/* protected by rq lock */
 
 	/*
+	 * Kick state owned by this cpu for this sched. scx_kick_cpu() records
+	 * targets here and links @to_kick_node onto the cpu's
+	 * rq->scx.sched_pcpus_to_kick. The cpu's single kick irq_work walks
+	 * that list and kicks each sched's targets on its behalf. Per-sched so
+	 * a kick stays attributed to its scheduler.
+	 */
+	cpumask_var_t		cpus_to_kick;
+	cpumask_var_t		cpus_to_kick_if_idle;
+	cpumask_var_t		cpus_to_preempt;
+	cpumask_var_t		cpus_to_wait;
+	struct list_head	to_kick_node;
+
+#ifdef CONFIG_EXT_SUB_SCHED
+	/*
+	 * pshard->caps[cap_bit] is the set of cids the sched holds that one
+	 * cap on. ecaps is its transpose: the set of SCX_CAP_* bits the sched
+	 * effectively holds on this cpu, with implied caps folded in, so that
+	 * the hot-path check is a single read.
+	 *
+	 * While pshard->caps[] under pshard->lock is the target configuration,
+	 * ecaps is the effective copy owned by the cpu. It is written under the
+	 * rq lock while processing rq->ecaps_to_sync. Can also be read with
+	 * READ_ONCE() outside rq lock.
+	 *
+	 * See queue_sync_ecaps() and scx_process_sync_ecaps().
+	 */
+	u64			ecaps;
+	struct llist_node	ecaps_to_sync_node;
+	/* owed a forced update_idle() re-notify on this cpu */
+	bool			idle_renotify;
+	/* effective caps as of the last sub_ecaps_updated() delivery */
+	u64			reported_ecaps;
+
+	/*
+	 * Decaying rescue runtime consumed on this cpu, see
+	 * scx_rescue_decay_avg(). Overload on this cpu ejects the sub with the
+	 * largest value. Accessed only under this cpu's rq lock.
+	 */
+	u64			rescue_avg;
+	u64			rescue_avg_at;	/* last decay, jiffies_64 */
+#endif
+
+	/*
 	 * The event counters are in a per-CPU variable to minimize the
 	 * accounting overhead. A system-wide view on the event counter is
 	 * constructed when requested by scx_bpf_events().
@@ -1104,6 +1378,146 @@ struct scx_sched_pnode {
 	struct scx_dispatch_q	global_dsq;
 };
 
+/*
+ * Sub-sched capability delegation.
+ *
+ * Caps are per-cid permissions parents delegate to direct children via
+ * scx_bpf_sub_grant() / scx_bpf_sub_revoke(). A child's cap set is always a
+ * subset of its parent's. A sub-sched checks its caps locally, and cross-sched
+ * communication is needed only when the delegation set itself changes.
+ *
+ * Caps are used to implement sub-sched scheduling on the enqueue path. Picking
+ * a cid for a task at a leaf depends on which cids the leaf is allowed to use.
+ * Resolving that programmatically on every enqueue would mean a cross-sched
+ * round-trip call chain, possibly retrying if the request can't be granted
+ * as-is.
+ *
+ * The dispatch path is different - it runs as top-down recursion via
+ * scx_bpf_sub_dispatch(): a sched's dispatch op invokes a child's dispatch op
+ * on the local rq, and the subtree dispatches in a single pass.
+ *
+ * Locking is per shard. cid space is split into shards, and each sub-sched has
+ * its own pshard->lock for each shard. Operations are broken up on shard
+ * boundaries. Different shards never contend. Shards are expected to be
+ * topology-aligned and likely to serve as the locality unit when cids are
+ * allocated to schedulers, so per-shard lock granularity scales naturally with
+ * the allocation pattern.
+ *
+ * ENQ_IMMED  insert an IMMED task onto the cid's local DSQ
+ *            - kick the cid's cpu (except SCX_KICK_PREEMPT)
+ *
+ * ENQ        insert any task onto the cid's local DSQ (implies ENQ_IMMED)
+ *
+ * PREEMPT    preempt any task running on the cid regardless of the owning
+ *            sched (implies ENQ). Preempting a task in the sched's own subtree
+ *            doesn't require any cap.
+ *            - SCX_ENQ_PREEMPT inserts
+ *            - SCX_KICK_PREEMPT kicks
+ *
+ * PERF       control the cid's cpu power/perf management state, currently the
+ *            cpufreq target set through scx_bpf_cidperf_set(). Hardware
+ *            control is a separate axis from queue access: PERF neither
+ *            implies nor is implied by the caps above.
+ *
+ * Implied caps apply to the holder's own use of a cid, not to delegation.
+ * scx_bpf_sub_grant() delegates literally-held caps, so a cap held only through
+ * implication is usable but cannot be re-delegated to a child. When granting a
+ * cap, it usually makes sense to delegate its implied caps explicitly alongside
+ * it.
+ */
+enum scx_cap_flags {
+	__SCX_CAP_ENQ_IMMED		= 0,
+	__SCX_CAP_ENQ			= 1,
+	__SCX_CAP_PREEMPT		= 2,
+	__SCX_CAP_PERF			= 3,
+
+	__SCX_NR_CAPS,
+	__SCX_CAP_ALL			= BIT_U64(__SCX_NR_CAPS) - 1,
+
+	SCX_CAP_ENQ_IMMED		= BIT_U64(__SCX_CAP_ENQ_IMMED),
+	SCX_CAP_ENQ			= BIT_U64(__SCX_CAP_ENQ),
+	SCX_CAP_PREEMPT			= BIT_U64(__SCX_CAP_PREEMPT),
+	SCX_CAP_PERF			= BIT_U64(__SCX_CAP_PERF),
+
+	/* alias for minimal cap to make any use of a cpu */
+	SCX_CAP_BASE			= SCX_CAP_ENQ_IMMED,
+
+	/* caps whose loss strands queued tasks, see scx_process_sync_ecaps() */
+	SCX_CAPS_REENQ_ON_LOSS		= SCX_CAP_ENQ_IMMED | SCX_CAP_ENQ,
+};
+
+#ifdef CONFIG_EXT_SUB_SCHED
+/* iterate set bits in a u64 cap mask */
+#define scx_for_each_cap_bit(cap_bit, caps)				\
+	for (u64 __caps = (caps);					\
+	     __caps && ((cap_bit) = __ffs64(__caps), true);		\
+	     __caps &= __caps - 1)
+
+/*
+ * Sub-cap update notifier.
+ *
+ * ops_cid.sub_caps_updated() notifies sub-scheds when their cap state changes
+ * so they can refresh internal state without polling scx_bpf_sub_caps() per
+ * enqueue.
+ *
+ * Three constraints shape the design:
+ *
+ *   1. Static memory. Deliveries use a fixed-size buffer, both for runtime
+ *      efficiency and so notifications can't be lost under memory pressure.
+ *
+ *   2. High-frequency updates. Grant/revoke can mutate caps in bursts, and the
+ *      notifier path must absorb that without amplifying it.
+ *
+ *   3. Recursive grant/revoke from the callback. A child receiving a
+ *      notification can call grant/revoke on its own children, which can
+ *      cascade recursively down its subtree.
+ *
+ * (1) and (2) lead to coalescing into a fixed payload. Each delivery carries a
+ * single (cmask, caps) pair covering every change since the previous one.
+ * Direction (set vs cleared) isn't encoded as it doesn't fit in the fixed-size
+ * summary. The callback queries scx_bpf_sub_caps() for current state. Only one
+ * delivery is in flight per shard. Further changes fold into the same buffer
+ * and ship as the next callback, so a shard's callbacks fire in order.
+ *
+ * (3) leads to deferred delivery. Events accumulate during grant/revoke and are
+ * delivered after the shard lock is released.
+ */
+struct scx_caps_updated {
+	raw_spinlock_t		lock;
+	u64			caps;
+	struct scx_cmask	*cmask_arena_out;
+	struct list_head	node_in_flight;
+	/* Kernel-side accumulator. Access as &cu->cmask. */
+	TRAILING_OVERLAP(struct scx_cmask, cmask, bits,
+			 u64 _bits[SCX_CMASK_NR_WORDS(SCX_CID_SHARD_MAX_CPUS)];
+	);
+};
+
+struct scx_pshard {
+	raw_spinlock_t		lock;		/* serializes caps */
+	struct scx_sched	*sch;		/* backpointer */
+	struct scx_caps_updated	caps_updated;
+
+	/*
+	 * Per-cap cmask, inline via TRAILING_OVERLAP so cmask.bits[] overlaps
+	 * the trailing _bits[] storage. Access as &caps[i].cmask. See
+	 * scx_sched_pcpu->ecaps.
+	 */
+	TRAILING_OVERLAP(struct scx_cmask, cmask, bits,
+			 u64 _bits[SCX_CMASK_NR_WORDS(SCX_CID_SHARD_MAX_CPUS)];
+	) caps[__SCX_NR_CAPS];
+
+	/*
+	 * Shard geometry captured at alloc. cmask_arena_out's own header is
+	 * bpf-writable and the live shard range can change before the
+	 * rcu-deferred free, so re-init and size cmask_arena_out from these
+	 * trusted copies instead.
+	 */
+	u32			base;
+	u32			nr_cids;
+};
+#endif
+
 struct scx_sched {
 	/*
 	 * cpu-form and cid-form ops share field offsets up to .priv (verified
@@ -1117,6 +1531,7 @@ struct scx_sched {
 		struct sched_ext_ops_cid	ops_cid;
 	};
 	bool			is_cid_type;	/* true if registered via bpf_sched_ext_ops_cid */
+	bool			dead;		/* set after ops.exit(), gates scx_prog_sched() */
 
 	/*
 	 * Arena map auto-discovered from member progs at struct_ops attach.
@@ -1125,8 +1540,7 @@ struct scx_sched {
 	 *
 	 * @arena_pool sub-allocates @arena_map. Each gen_pool chunk is added
 	 * at the kernel-side mapping address. @arena_kern_base is the start
-	 * of the arena's kern_vm range. See scx_arena_to_kaddr() and
-	 * scx_kaddr_to_arena().
+	 * of the arena's kern_vm range. See scx_arena_to_kaddr().
 	 */
 	struct bpf_map		*arena_map;
 	struct gen_pool		*arena_pool;
@@ -1135,7 +1549,7 @@ struct scx_sched {
 	/*
 	 * Per-CPU arena cmask used by scx_call_op_set_cpumask() to hand a cmask
 	 * to ops_cid.set_cmask(). The kernel writes through the stored kern_va
-	 * and hands BPF its arena pointer via scx_kaddr_to_arena().
+	 * and passes it to the callback's __arena argument.
 	 */
 	struct scx_cmask * __percpu *set_cmask_scratch;
 
@@ -1151,13 +1565,16 @@ struct scx_sched {
 	 */
 	struct rhashtable	dsq_hash;
 	struct scx_sched_pnode	**pnode;
+#ifdef CONFIG_EXT_SUB_SCHED
+	struct scx_pshard	**pshard;	/* indexed by shard_idx */
+#endif
 	struct scx_sched_pcpu __percpu *pcpu;
 
 	u64			slice_dfl;
 	u64			bypass_timestamp;
 	s32			bypass_depth;
 
-	/* bypass dispatch path enable state, see bypass_dsp_enabled() */
+	/* bypass dispatch path enable state, see scx_bypass_dsp_enabled() */
 	unsigned long		bypass_dsp_claim;
 	atomic_t		bypass_dsp_enable_depth;
 
@@ -1166,16 +1583,28 @@ struct scx_sched {
 	u32			dsp_max_batch;
 	s32			level;
 
+#ifdef CONFIG_EXT_SUB_SCHED
+	/*
+	 * pshard[] size captured at enable for the async RCU free path -
+	 * scx_nr_cid_shards may be rewritten by a later enable's
+	 * scx_cid_publish_tables() before free runs. While sch is active, use
+	 * the global.
+	 */
+	u32			nr_pshards;
+#endif
+
 	/*
 	 * Updates to the following warned bitfields can race causing RMW issues
 	 * but it doesn't really matter.
 	 */
 	bool			warned_zero_slice:1;
-	bool			warned_deprecated_rq:1;
 	bool			warned_unassoc_progs:1;
 
 	struct list_head	all;
 
+	/* unique instance id, monotonic and never reused */
+	u64			id;
+
 #ifdef CONFIG_EXT_SUB_SCHED
 	struct rhash_head	hash_node;
 
@@ -1185,6 +1614,7 @@ struct scx_sched {
 	char			*cgrp_path;
 	struct kset		*sub_kset;
 
+	bool			linked;		/* on ->children, see scx_link_sched() */
 	bool			sub_attached;
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
@@ -1203,9 +1633,11 @@ struct scx_sched {
 	struct kthread_worker	*helper;
 	struct irq_work		disable_irq_work;
 	struct kthread_work	disable_work;
+	struct irq_work		propagate_exit_irq_work; /* see scx_claim_exit() */
 	struct timer_list	bypass_lb_timer;
 	cpumask_var_t		bypass_lb_donee_cpumask;
 	cpumask_var_t		bypass_lb_resched_cpumask;
+	cpumask_var_t		stall_cpus;
 	struct rcu_work		rcu_work;
 
 	/* all ancestors including self */
@@ -1227,16 +1659,6 @@ static inline void *scx_arena_to_kaddr(struct scx_sched *sch, const void *bpf_pt
 	return (void *)(sch->arena_kern_base + (u32)(uintptr_t)bpf_ptr);
 }
 
-/**
- * scx_kaddr_to_arena - Translate a kernel arena address to its BPF form
- * @sch: scheduler whose arena hosts @kaddr
- * @kaddr: kernel-side arena address, supplied by trusted kernel code
- */
-static inline void *scx_kaddr_to_arena(struct scx_sched *sch, const void *kaddr)
-{
-	return (void *)((uintptr_t)kaddr - sch->arena_kern_base);
-}
-
 enum scx_wake_flags {
 	/* expose select WF_* flags as enums */
 	SCX_WAKE_FORK		= WF_FORK,
@@ -1279,6 +1701,17 @@ enum scx_enq_flags {
 	SCX_ENQ_IMMED		= 1LLU << 33,
 
 	/*
+	 * Only allowed on local DSQs. If the insert lacks the caps for the
+	 * target cid, divert the task to the CPU's rescue path instead of
+	 * rejecting and reenqueueing, e.g. when the task's affinity is
+	 * restricted to cids the scheduler doesn't hold. The kernel runs
+	 * rescued tasks on the target CPU. Rescue execution is guaranteed to
+	 * make forward progress and is bandwidth-limited, see the
+	 * rescue_bandwidth_ppt and rescue_quantum_us ops fields.
+	 */
+	SCX_ENQ_RESCUE		= 1LLU << 34,
+
+	/*
 	 * The task being enqueued was previously enqueued on a DSQ, but was
 	 * removed and is being re-enqueued. See SCX_TASK_REENQ_* flags to find
 	 * out why a given task is being reenqueued.
@@ -1303,6 +1736,9 @@ enum scx_enq_flags {
 	SCX_ENQ_DSQ_PRIQ	= 1LLU << 57,
 	SCX_ENQ_NESTED		= 1LLU << 58,
 	SCX_ENQ_GDSQ_FALLBACK	= 1LLU << 59,	/* fell back to global DSQ */
+	SCX_ENQ_IGNORE_CAPS	= 1LLU << 60,	/* admit to local DSQ ignoring caps */
+	SCX_ENQ_APPLY_SLICE	= 1LLU << 61,	/* apply carried slice/vtime at insertion */
+	SCX_ENQ_SLICE_DFL	= 1LLU << 62,	/* carried slice is a default refill */
 };
 
 enum scx_deq_flags {
@@ -1329,6 +1765,9 @@ enum scx_reenq_flags {
 	/* low 16bits determine which tasks should be reenqueued */
 	SCX_REENQ_ANY		= 1LLU << 0,	/* all tasks */
 
+	/* internal: kernel-issued on cap revoke, not accepted from BPF */
+	SCX_REENQ_CAP_REVOKE	= 1LLU << 1,
+
 	__SCX_REENQ_FILTER_MASK	= 0xffffLLU,
 
 	__SCX_REENQ_USER_MASK	= SCX_REENQ_ANY,
@@ -1374,6 +1813,7 @@ enum scx_kick_flags {
 enum scx_tg_flags {
 	SCX_TG_ONLINE		= 1U << 0,
 	SCX_TG_INITED		= 1U << 1,
+	SCX_TG_SUB_INIT		= 1U << 2,	/* see scx_cgroup_claim_subtree() */
 };
 
 enum scx_enable_state {
@@ -1520,6 +1960,47 @@ enum scx_ops_state {
 #define SCX_OPSS_STATE_MASK	((1LU << SCX_OPSS_QSEQ_SHIFT) - 1)
 #define SCX_OPSS_QSEQ_MASK	(~SCX_OPSS_STATE_MASK)
 
+/*
+ * SCX task iterator.
+ */
+struct scx_task_iter {
+	struct sched_ext_entity		cursor;
+	struct task_struct		*locked_task;
+	struct rq			*rq;
+	struct rq_flags			rf;
+	u32				cnt;
+	bool				list_locked;
+#ifdef CONFIG_EXT_SUB_SCHED
+	struct cgroup			*cgrp;
+	struct cgroup_subsys_state	*css_pos;
+	struct css_task_iter		css_iter;
+#endif
+};
+
+/*
+ * scx_enable() is offloaded to a dedicated system-wide RT kthread to avoid
+ * starvation. During the READY -> ENABLED task switching loop, the calling
+ * thread's sched_class gets switched from fair to ext. As fair has higher
+ * priority than ext, the calling thread can be indefinitely starved under
+ * fair-class saturation, leading to a system hang.
+ */
+struct scx_enable_cmd {
+	struct kthread_work	work;
+	union {
+		struct sched_ext_ops		*ops;
+		struct sched_ext_ops_cid	*ops_cid;
+	};
+	bool			is_cid_type;
+	struct bpf_map		*arena_map;	/* arena ref to transfer to sch */
+	int			ret;
+};
+
+/* string formatting from BPF */
+struct scx_bstr_buf {
+	u64			data[MAX_BPRINTF_VARARGS];
+	char			line[SCX_EXIT_MSG_LEN];
+};
+
 extern struct scx_sched __rcu *scx_root;
 DECLARE_PER_CPU(struct rq *, scx_locked_rq_state);
 
@@ -1540,12 +2021,159 @@ __printf(5, 0) bool scx_vexit(struct scx_sched *sch, enum scx_exit_kind kind,
 __printf(5, 6) bool __scx_exit(struct scx_sched *sch, enum scx_exit_kind kind,
 			       s64 exit_code, s32 exit_cpu, const char *fmt, ...);
 
+u32 scx_get_task_state(const struct task_struct *p);
+void scx_set_task_state(struct task_struct *p, u32 state);
+void scx_task_iter_start(struct scx_task_iter *iter, struct cgroup *cgrp);
+void scx_task_iter_unlock(struct scx_task_iter *iter);
+void scx_task_iter_stop(struct scx_task_iter *iter);
+struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter);
+bool scx_set_task_slice(struct task_struct *p, u64 slice);
+void scx_task_slice_ended(struct rq *rq, struct task_struct *p);
+void scx_task_unlink_from_dsq(struct task_struct *p, struct scx_dispatch_q *dsq);
+void scx_dispatch_dequeue(struct rq *rq, struct task_struct *p);
+void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
+			 int sticky_cpu);
+void scx_move_local_task_to_local_dsq(struct scx_sched *sch, struct task_struct *p,
+				      u64 enq_flags, struct scx_dispatch_q *src_dsq,
+				      struct rq *dst_rq);
+bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
+			    struct scx_dispatch_q *dsq, u64 enq_flags);
+bool scx_consume_global_dsq(struct scx_sched *sch, struct rq *rq);
+bool scx_rq_online(struct rq *rq);
+void scx_flush_dispatch_buf(struct scx_sched *sch, struct rq *rq);
+s32 scx_init_dsq(struct scx_dispatch_q *dsq, u64 dsq_id, struct scx_sched *sch);
+__printf(2, 3) void scx_dump_line(struct seq_buf *s, const char *fmt, ...);
+void scx_kick_cpu(struct scx_sched *sch, s32 cpu, u64 flags);
+u64 __scx_bpf_now(struct rq *rq);
+void schedule_dsq_reenq(struct scx_sched *sch, struct scx_dispatch_q *dsq,
+			u64 reenq_flags, struct rq *locked_rq);
+int __scx_init_task(struct scx_sched *sch, struct task_struct *p,
+		    struct cgroup *cgrp, bool fork);
+void scx_enable_task(struct scx_sched *sch, struct task_struct *p);
+void __scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p);
+void scx_sub_init_cancel_task(struct scx_sched *sch, struct task_struct *p);
+void scx_disable_and_exit_task(struct scx_sched *sch, struct task_struct *p);
+#if defined(CONFIG_EXT_GROUP_SCHED) || defined(CONFIG_EXT_SUB_SCHED)
+void scx_cgroup_lock(void);
+void scx_cgroup_unlock(void);
+#endif
+s32 scx_set_cmask_scratch_alloc(struct scx_sched *sch);
+void scx_disable_bypass_dsp(struct scx_sched *sch);
+void scx_bypass(struct scx_sched *sch, bool bypass);
+s32 scx_link_sched(struct scx_sched *sch);
+void scx_unlink_sched(struct scx_sched *sch);
+void scx_disable_dump(struct scx_sched *sch);
+void scx_log_sched_disable(struct scx_sched *sch);
+void scx_flush_disable_work(struct scx_sched *sch);
+struct scx_sched *scx_alloc_and_add_sched(struct scx_enable_cmd *cmd,
+					  struct cgroup *cgrp,
+					  struct scx_sched *parent);
+int scx_validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops);
+int scx_sched_sysfs_add(struct scx_sched *sch);
+bool scx_is_descendant(struct scx_sched *sch, struct scx_sched *ancestor);
+__printf(5, 0) bool scx_exit_bstr(struct scx_sched *sch, enum scx_exit_kind kind,
+				  s64 exit_code, struct scx_sched *fmt_blame,
+				  char *fmt, unsigned long long *data, u32 data__sz);
+
+extern raw_spinlock_t scx_sched_lock;
+extern struct mutex scx_enable_mutex;
+extern struct percpu_rw_semaphore scx_fork_rwsem;
+extern bool scx_cgroup_enabled;
+extern struct list_head scx_sched_all;
+#ifdef CONFIG_EXT_SUB_SCHED
+extern const struct rhashtable_params scx_sched_hash_params;
+extern struct rhashtable scx_sched_hash;
+extern struct scx_sched *scx_enabling_sub_sched;
+#endif
+
 #define scx_exit(sch, kind, exit_code, fmt, args...)				\
 	__scx_exit(sch, kind, exit_code, raw_smp_processor_id(), fmt, ##args)
 #define scx_error(sch, fmt, args...)						\
 	scx_exit((sch), SCX_EXIT_ERROR, 0, fmt, ##args)
-#define scx_verror(sch, fmt, args)						\
-	scx_vexit((sch), SCX_EXIT_ERROR, 0, raw_smp_processor_id(), fmt, args)
+
+/**
+ * scx_root_protected_live - Root sched for paths that only run while live
+ *
+ * scx_root is published before the scheduler goes live and cleared only after
+ * it is fully drained, so a path that only executes while the scheduler is live
+ * can never race an update. Return the root sched with a plain load, never
+ * %NULL.
+ */
+static inline struct scx_sched *scx_root_protected_live(void)
+{
+	return rcu_dereference_protected(scx_root, true);
+}
+
+/**
+ * scx_root_protected - Root sched for contexts that exclude its updates
+ *
+ * Both scx_root updates run under the locks checked below, so holding one
+ * excludes them. Return the root sched with a plain load, %NULL if no scheduler
+ * is loaded.
+ */
+static inline struct scx_sched *scx_root_protected(void)
+{
+	return rcu_dereference_protected(scx_root,
+					 lockdep_is_cpus_held() ||
+					 lockdep_is_held(&scx_enable_mutex));
+}
+
+static inline struct scx_dispatch_q *scx_bypass_dsq(struct scx_sched *sch, s32 cpu)
+{
+	return &per_cpu_ptr(sch->pcpu, cpu)->bypass_dsq;
+}
+
+/**
+ * scx_bypass_dsp_enabled - Check if bypass dispatch path is enabled
+ * @sch: scheduler to check
+ *
+ * When a descendant scheduler enters bypass mode, bypassed tasks are scheduled
+ * by the nearest non-bypassing ancestor, or the root scheduler if all ancestors
+ * are bypassing. In the former case, the ancestor is not itself bypassing but
+ * its bypass DSQs will be populated with bypassed tasks from descendants. Thus,
+ * the ancestor's bypass dispatch path must be active even though its own
+ * bypass_depth remains zero.
+ *
+ * This function checks bypass_dsp_enable_depth which is managed separately from
+ * bypass_depth to enable this decoupling. See enable_bypass_dsp() and
+ * scx_disable_bypass_dsp().
+ */
+static inline bool scx_bypass_dsp_enabled(struct scx_sched *sch)
+{
+	return unlikely(atomic_read(&sch->bypass_dsp_enable_depth));
+}
+
+/**
+ * scx_ops_sanitize_err - Sanitize a -errno value
+ * @sch: scx_sched to error out on error
+ * @ops_name: operation to blame on failure
+ * @err: -errno value to sanitize
+ *
+ * Verify @err is a valid -errno. If not, trigger scx_error() and return
+ * -%EPROTO. This is necessary because returning a rogue -errno up the chain can
+ * cause misbehaviors. For an example, a large negative return from
+ * ops.init_task() triggers an oops when passed up the call chain because the
+ * value fails IS_ERR() test after being encoded with ERR_PTR() and then is
+ * handled as a pointer.
+ */
+static inline int scx_ops_sanitize_err(struct scx_sched *sch, const char *ops_name, s32 err)
+{
+	if (err < 0 && err >= -MAX_ERRNO)
+		return err;
+
+	scx_error(sch, "ops.%s() returned an invalid errno %d", ops_name, err);
+	return -EPROTO;
+}
+
+static inline void scx_schedule_reenq_local(struct rq *rq, u64 reenq_flags)
+{
+	struct scx_sched *root = rcu_dereference_sched(scx_root);
+
+	if (WARN_ON_ONCE(!root))
+		return;
+
+	schedule_dsq_reenq(root, &rq->scx.local_dsq, reenq_flags, rq);
+}
 
 /*
  * Return the rq currently locked from an scx callback, or NULL if no rq is
@@ -1573,8 +2201,11 @@ static inline void update_locked_rq(struct rq *rq)
 /*
  * SCX ops can recurse via scx_bpf_sub_dispatch() - the inner call must not
  * clobber the outer's scx_locked_rq_state. Save it on entry, restore on exit.
+ *
+ * @ops is the ops table to dispatch through: ops for the cpu form, ops_cid
+ * for the cid form.
  */
-#define SCX_CALL_OP(sch, op, locked_rq, args...)				\
+#define __SCX_CALL_OP(sch, ops, op, locked_rq, args...)				\
 do {										\
 	struct rq *__prev_locked_rq;						\
 										\
@@ -1587,6 +2218,9 @@ do {										\
 		update_locked_rq(__prev_locked_rq);				\
 } while (0)
 
+#define SCX_CALL_OP(sch, op, locked_rq, args...)				\
+	__SCX_CALL_OP(sch, ops, op, locked_rq, ##args)
+
 #define SCX_CALL_OP_RET(sch, op, locked_rq, args...)				\
 ({										\
 	struct rq *__prev_locked_rq;						\
@@ -1618,17 +2252,36 @@ do {										\
  * WARN_ON_ONCE() in each macro catches a re-entry of any of the three variants
  * while a previous one is still in progress.
  */
-#define SCX_CALL_OP_TASK(sch, op, locked_rq, task, args...)			\
+#define __SCX_CALL_OP_TASK(sch, ops, op, locked_rq, task, args...)		\
 do {										\
 	WARN_ON_ONCE(current->scx.kf_tasks[0]);					\
 	current->scx.kf_tasks[0] = task;					\
-	SCX_CALL_OP((sch), op, locked_rq, task, ##args);			\
+	__SCX_CALL_OP((sch), ops, op, locked_rq, task, ##args);			\
 	current->scx.kf_tasks[0] = NULL;					\
 } while (0)
 
+/*
+ * A per-task op runs on @task's owner - WARN if @sch isn't it. Sites that must
+ * target a different scheduler call __SCX_CALL_OP_TASK() directly.
+ */
+#define SCX_CALL_OP_TASK(sch, op, locked_rq, task, args...)			\
+do {										\
+	WARN_ON_ONCE(scx_has_subs() && (sch) != scx_task_sched_rcu(task));	\
+	__SCX_CALL_OP_TASK((sch), ops, op, locked_rq, task, ##args);		\
+} while (0)
+
+/*
+ * Dispatch a task op through the cid-form ops_cid table. Only set_cmask() needs
+ * this: it takes an arena cmask address instead of a cpumask, so it cannot be
+ * invoked via its cpu-form set_cpumask() slot.
+ */
+#define SCX_CALL_CID_OP_TASK(sch, op, locked_rq, task, args...)			\
+	__SCX_CALL_OP_TASK(sch, ops_cid, op, locked_rq, task, ##args)
+
 #define SCX_CALL_OP_TASK_RET(sch, op, locked_rq, task, args...)			\
 ({										\
 	__typeof__((sch)->ops.op(task, ##args)) __ret;				\
+	WARN_ON_ONCE(scx_has_subs() && (sch) != scx_task_sched_rcu(task));	\
 	WARN_ON_ONCE(current->scx.kf_tasks[0]);					\
 	current->scx.kf_tasks[0] = task;					\
 	__ret = SCX_CALL_OP_RET((sch), op, locked_rq, task, ##args);		\
@@ -1668,6 +2321,19 @@ static inline bool scx_bypassing(struct scx_sched *sch, s32 cpu)
 }
 
 #ifdef CONFIG_EXT_SUB_SCHED
+DECLARE_STATIC_KEY_FALSE(__scx_has_subs);
+
+/**
+ * scx_has_subs - Whether any sub-scheduler exists
+ *
+ * Gates the sub-sched portions of hot paths so that a root-only system doesn't
+ * pay for them. See scx_sub_enable_workfn() and scx_sched_free_rcu_work().
+ */
+static inline bool scx_has_subs(void)
+{
+	return static_branch_unlikely(&__scx_has_subs);
+}
+
 /**
  * scx_task_sched - Find scx_sched scheduling a task
  * @p: task of interest
@@ -1717,14 +2383,20 @@ static inline bool scx_task_on_sched(struct scx_sched *sch,
 static inline struct scx_sched *scx_prog_sched(const struct bpf_prog_aux *aux)
 {
 	struct sched_ext_ops *ops;
-	struct scx_sched *root;
+	struct scx_sched *sch, *root;
 
 	ops = bpf_prog_get_assoc_struct_ops(aux);
-	if (likely(ops))
-		return rcu_dereference_all(ops->priv);
+	if (likely(ops)) {
+		sch = rcu_dereference_all(ops->priv);
+		if (sch && unlikely(READ_ONCE(sch->dead)))
+			return NULL;
+		return sch;
+	}
 
 	root = rcu_dereference_all(scx_root);
 	if (root) {
+		if (unlikely(READ_ONCE(root->dead)))
+			return NULL;
 		/*
 		 * COMPAT-v6.19: Schedulers built before sub-sched support was
 		 * introduced may have unassociated non-struct_ops programs.
@@ -1755,7 +2427,10 @@ static inline struct scx_sched *scx_parent(struct scx_sched *sch)
 	else
 		return NULL;
 }
+
 #else	/* CONFIG_EXT_SUB_SCHED */
+static inline bool scx_has_subs(void) { return false; }
+
 static inline struct scx_sched *scx_task_sched(const struct task_struct *p)
 {
 	return rcu_dereference_protected(scx_root,
@@ -1776,10 +2451,15 @@ static inline bool scx_task_on_sched(struct scx_sched *sch,
 
 static inline struct scx_sched *scx_prog_sched(const struct bpf_prog_aux *aux)
 {
-	return rcu_dereference_all(scx_root);
+	struct scx_sched *root = rcu_dereference_all(scx_root);
+
+	if (root && unlikely(READ_ONCE(root->dead)))
+		return NULL;
+	return root;
 }
 
 static inline struct scx_sched *scx_parent(struct scx_sched *sch) { return NULL; }
+
 #endif	/* CONFIG_EXT_SUB_SCHED */
 
 #endif /* _KERNEL_SCHED_EXT_INTERNAL_H */
diff --git a/kernel/sched/ext/sub.c b/kernel/sched/ext/sub.c
new file mode 100644
index 000000000000..0554448835bd
--- /dev/null
+++ b/kernel/sched/ext/sub.c
@@ -0,0 +1,2681 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
+ *
+ * Sub-scheduler hierarchy support.
+ *
+ * A sub-scheduler is an scx_sched attached to a cgroup subtree under another
+ * scx_sched. This file holds the sub-scheduler implementation: the scheduler
+ * tree walk, capability delegation, per-shard cap state and its sync, and the
+ * sub-scheduler enable/disable paths. The core dispatch/enqueue machinery it
+ * builds on lives in ext.c.
+ *
+ * Copyright (c) 2026 Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2026 Tejun Heo <tj@kernel.org>
+ */
+#include <linux/rhashtable.h>
+#include "internal.h"
+#include "cid.h"
+#include "arena.h"
+#include "sub.h"
+#include "inlines.h"
+
+#ifdef CONFIG_EXT_SUB_SCHED
+
+/*
+ * On while any sub-scheduler exists so that a root-only system doesn't pay for
+ * the sub-sched portions of hot paths. See scx_has_subs().
+ */
+DEFINE_STATIC_KEY_FALSE(__scx_has_subs);
+
+/* latched at root enable before any rescue runs */
+static s32 scx_rescue_bw_1024;
+static s64 scx_rescue_quantum_ns;
+static s64 scx_rescue_sat_delta_ns;
+static unsigned long scx_rescue_decay_halflife;
+static unsigned long scx_rescue_overload_after;
+
+/**
+ * scx_skip_subtree_pre - Skip @pos's subtree in a pre-order walk
+ * @pos: current position
+ * @root: walk root
+ *
+ * In a walk started by scx_next_descendant_pre(), continue past @pos's subtree:
+ * return @pos's next sibling, or the closest ancestor's next sibling, or NULL
+ * if @pos's subtree is the last under @root. Same locking rules.
+ */
+struct scx_sched *scx_skip_subtree_pre(struct scx_sched *pos, struct scx_sched *root)
+{
+	struct scx_sched *next;
+
+	lockdep_assert(lockdep_is_held(&scx_enable_mutex) ||
+		       lockdep_is_held(&scx_sched_lock) ||
+		       rcu_read_lock_any_held());
+
+	while (pos != root) {
+		next = list_next_or_null_rcu(&scx_parent(pos)->children, &pos->sibling,
+					     struct scx_sched, sibling);
+		if (next)
+			return next;
+		pos = scx_parent(pos);
+	}
+	return NULL;
+}
+
+/**
+ * scx_next_descendant_pre - find the next descendant for pre-order walk
+ * @pos: the current position (%NULL to initiate traversal)
+ * @root: sched whose descendants to walk
+ *
+ * To be used by scx_for_each_descendant_pre(). Find the next descendant to
+ * visit for pre-order traversal of @root's descendants. @root is included in
+ * the iteration and the first node to be visited.
+ */
+struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root)
+{
+	struct scx_sched *next;
+
+	lockdep_assert(lockdep_is_held(&scx_enable_mutex) ||
+		       lockdep_is_held(&scx_sched_lock) ||
+		       rcu_read_lock_any_held());
+
+	/* if first iteration, visit @root */
+	if (!pos)
+		return root;
+
+	/* visit the first child if exists */
+	next = list_first_or_null_rcu(&pos->children, struct scx_sched, sibling);
+	if (next)
+		return next;
+
+	/* no child, visit my or the closest ancestor's next sibling */
+	return scx_skip_subtree_pre(pos, root);
+}
+
+static struct scx_sched *scx_find_sub_sched(u64 cgroup_id)
+{
+	return rhashtable_lookup(&scx_sched_hash, &cgroup_id,
+				 scx_sched_hash_params);
+}
+
+void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch)
+{
+	rcu_assign_pointer(p->scx.sched, sch);
+}
+
+struct cgroup *sch_cgroup(struct scx_sched *sch)
+{
+	return sch->cgrp;
+}
+
+/* for each descendant of @cgrp including self, set ->scx_sched to @sch */
+void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch)
+{
+	struct cgroup *pos;
+	struct cgroup_subsys_state *css;
+
+	cgroup_for_each_live_descendant_pre(pos, css, cgrp)
+		rcu_assign_pointer(pos->scx_sched, sch);
+}
+
+static void free_pshard(struct scx_pshard *pshard)
+{
+	struct scx_caps_updated *cu;
+
+	if (!pshard)
+		return;
+	cu = &pshard->caps_updated;
+	if (cu->cmask_arena_out)
+		scx_arena_free(pshard->sch, cu->cmask_arena_out,
+			       struct_size_t(struct scx_cmask, bits,
+					     SCX_CMASK_NR_WORDS(pshard->nr_cids)));
+	kfree(pshard);
+}
+
+void scx_free_pshards(struct scx_sched *sch)
+{
+	s32 si;
+
+	if (!sch->pshard)
+		return;
+	for (si = 0; si < sch->nr_pshards; si++)
+		free_pshard(sch->pshard[si]);
+	kfree(sch->pshard);
+}
+
+static struct scx_pshard *alloc_pshard(struct scx_sched *sch, s32 shard_idx, s32 node)
+{
+	const struct scx_cid_shard *shard =
+		&rcu_dereference_protected(scx_cid_shard_ranges,
+					   lockdep_is_held(&scx_enable_mutex))[shard_idx];
+	size_t cmask_size = struct_size_t(struct scx_cmask, bits,
+					  SCX_CMASK_NR_WORDS(shard->nr_cids));
+	struct scx_pshard *pshard;
+	struct scx_caps_updated *cu;
+	s32 i;
+
+	pshard = kzalloc_node(sizeof(*pshard), GFP_KERNEL, node);
+	if (!pshard)
+		return NULL;
+
+	raw_spin_lock_init(&pshard->lock);
+	pshard->sch = sch;
+	pshard->base = shard->base_cid;
+	pshard->nr_cids = shard->nr_cids;
+
+	for (i = 0; i < __SCX_NR_CAPS; i++)
+		scx_cmask_init(&pshard->caps[i].cmask, shard->base_cid, shard->nr_cids);
+
+	cu = &pshard->caps_updated;
+	raw_spin_lock_init(&cu->lock);
+	INIT_LIST_HEAD(&cu->node_in_flight);
+	__scx_cmask_init(&cu->cmask, shard->base_cid, shard->nr_cids, SCX_CID_SHARD_MAX_CPUS);
+
+	cu->cmask_arena_out = scx_arena_alloc(sch, cmask_size);
+	if (!cu->cmask_arena_out) {
+		free_pshard(pshard);
+		return NULL;
+	}
+
+	scx_cmask_init(cu->cmask_arena_out, shard->base_cid, shard->nr_cids);
+
+	return pshard;
+}
+
+s32 scx_alloc_pshards(struct scx_sched *sch)
+{
+	struct scx_pshard **pshard;
+	s32 *shard_node;
+	s32 si;
+
+	if (!sch->is_cid_type || !sch->arena_pool)
+		return 0;
+
+	shard_node = rcu_dereference_protected(scx_shard_node,
+					       lockdep_is_held(&scx_enable_mutex));
+
+	pshard = kzalloc_objs(pshard[0], scx_nr_cid_shards, GFP_KERNEL);
+	if (!pshard)
+		return -ENOMEM;
+
+	for (si = 0; si < scx_nr_cid_shards; si++) {
+		pshard[si] = alloc_pshard(sch, si, shard_node[si]);
+		if (!pshard[si]) {
+			while (--si >= 0)
+				free_pshard(pshard[si]);
+			kfree(pshard);
+			return -ENOMEM;
+		}
+	}
+
+	sch->nr_pshards = scx_nr_cid_shards;
+	/*
+	 * Publish only after every entry is built so a reader observing
+	 * @sch->pshard never sees a partially-filled array or unpublished cid
+	 * tables. Pair the store with a barrier and an acquire load on the
+	 * read side.
+	 */
+	smp_wmb();
+	WRITE_ONCE(sch->pshard, pshard);
+	return 0;
+}
+
+/*
+ * Seed the root's caps fully. Root owns all cids on all caps at enable time.
+ * Children acquire caps via scx_bpf_sub_grant().
+ */
+void scx_init_root_caps(struct scx_sched *sch)
+{
+	s32 si, i;
+
+	for (si = 0; si < sch->nr_pshards; si++) {
+		struct scx_pshard *ps = sch->pshard[si];
+
+		for (i = 0; i < __SCX_NR_CAPS; i++)
+			scx_cmask_fill(&ps->caps[i].cmask);
+	}
+}
+
+/* unserved remainder of @rq's rescuee's admitted slice, 0 once fully served */
+static s64 scx_rescue_slice_remaining(struct rq *rq)
+{
+	s64 served = rq->scx.rescue.curr->se.sum_exec_runtime - rq->scx.rescue.exec_snap;
+
+	return max(rq->scx.rescue.slice - served, 0);
+}
+
+/*
+ * Decay @pcpu's rescue usage average in place, halving per the knob-derived
+ * halflife, see scx_rescue_set_knobs(). The timestamp advances only by whole
+ * halflives.
+ */
+static u64 scx_rescue_decay_avg(struct scx_sched_pcpu *pcpu)
+{
+	unsigned long halflife = scx_rescue_decay_halflife;
+	u64 n = div_u64(get_jiffies_64() - pcpu->rescue_avg_at, halflife);
+
+	if (n) {
+		pcpu->rescue_avg = n < 64 ? pcpu->rescue_avg >> n : 0;
+		pcpu->rescue_avg_at += n * halflife;
+	}
+	return pcpu->rescue_avg;
+}
+
+/**
+ * scx_rescue_charge - Charge the rescuee's runtime
+ * @rq: rq the rescuee is running on
+ * @delta_exec: runtime being charged
+ *
+ * Also ends the rescue once the admitted slice has been served in full. Ending
+ * on served time rather than slice exhaustion bounds both the rescue and the
+ * charging when a scheduler extends the rescuee's slice.
+ */
+void scx_rescue_charge(struct rq *rq, s64 delta_exec)
+{
+	struct scx_sched_pcpu *pcpu;
+
+	lockdep_assert_rq_held(rq);
+
+	/*
+	 * A rescue slice is bounded by one quantum and tick-driven expiry can
+	 * overshoot by up to a tick. Clamp to avoid wild over-charges on VMs.
+	 */
+	delta_exec = min_t(s64, delta_exec, scx_rescue_quantum_ns + TICK_NSEC);
+
+	rq->scx.rescue.budget -= delta_exec;
+
+	/* per-cpu usage average feeds the overload victim pick */
+	pcpu = per_cpu_ptr(scx_task_sched(rq->curr)->pcpu, cpu_of(rq));
+	pcpu->rescue_avg = scx_rescue_decay_avg(pcpu) + delta_exec;
+
+	if (!scx_rescue_slice_remaining(rq))
+		scx_task_slice_ended(rq, rq->scx.rescue.curr);
+}
+
+/**
+ * scx_rescue_end - End the rescue execution on @rq
+ * @rq: rq of interest
+ *
+ * When no rescuee is left pending, the session is over and the balance above
+ * one quantum dies with it - it would otherwise become a banked license to
+ * preempt the cid owner long after the starvation ended. While waiters remain,
+ * the accrued deficit belongs to the queue and carries into the next rescue.
+ */
+void scx_rescue_end(struct rq *rq)
+{
+	lockdep_assert_rq_held(rq);
+
+	rq->scx.rescue.curr = NULL;
+	if (list_empty(&rq->scx.rescue.dsq.list))
+		rq->scx.rescue.budget = min(rq->scx.rescue.budget, scx_rescue_quantum_ns);
+}
+
+/**
+ * scx_rescue_keep - Keep the rescue going for a preempted-out rescuee
+ * @rq: rq @p is running on
+ * @p: task under rescue whose slice is exhausted
+ *
+ * Called from put_prev_task_scx() to decide what an exhausted slice means for
+ * the rescuee. scx_rescue_charge() ends the rescue the moment the admitted
+ * slice is fully served, so arriving here with the rescue still open means @p
+ * was preempted. Restore the unserved remainder and return %true - @p stays the
+ * rescuee and the caller reinserts it at the tail of the local DSQ, behind
+ * whatever preempted the rescuee.
+ *
+ * Return %false to end the rescue instead - the slice is already fully served,
+ * @p is leaving the rq or bypass is dismantling rescues.
+ */
+bool scx_rescue_keep(struct rq *rq, struct task_struct *p)
+{
+	s64 remaining = scx_rescue_slice_remaining(rq);
+
+	lockdep_assert_rq_held(rq);
+
+	if (!remaining || !(p->scx.flags & SCX_TASK_QUEUED) ||
+	    scx_bypassing(scx_task_sched(p), cpu_of(rq)))
+		return false;
+
+	scx_set_task_slice(p, remaining);
+	return true;
+}
+
+/**
+ * scx_rescue_accrue - Accrue budget at the configured fraction of elapsed time
+ * @rq: rq of interest
+ *
+ * A session spans from the first arrival until no rescuee is left, pending or
+ * admitted. While one is active the cap is three quanta and the balance drives
+ * escalation, see scx_rescue_timerfn(). Outside a session the cap is one
+ * quantum, so an idle gap funds the next arrival's admission but never an
+ * escalation.
+ */
+static void scx_rescue_accrue(struct rq *rq)
+{
+	bool in_session = rq->scx.rescue.curr || !list_empty(&rq->scx.rescue.dsq.list);
+	s64 cap = in_session ? 3 * scx_rescue_quantum_ns : scx_rescue_quantum_ns;
+	s64 delta;
+	u64 now;
+
+	lockdep_assert_rq_held(rq);
+
+	/* not every path here holds an updated rq clock, use __scx_bpf_now() */
+	now = __scx_bpf_now(rq);
+	delta = now - rq->scx.rescue.clock;
+	rq->scx.rescue.clock = now;
+
+	/*
+	 * Avoid multiplication overflows by taking a shortcut when the gap is
+	 * large enough to fill the budget.
+	 */
+	if (delta >= scx_rescue_sat_delta_ns)
+		rq->scx.rescue.budget = cap;
+	else
+		rq->scx.rescue.budget =
+			min(cap, rq->scx.rescue.budget +
+			    ((delta * scx_rescue_bw_1024) >> SCHED_CAPACITY_SHIFT));
+}
+
+/*
+ * The slice for the next admission - the quantum divided across the stranded
+ * tasks so that a crowded queue round-robins on shorter slices.
+ */
+static s64 scx_rescue_next_slice(struct rq *rq)
+{
+	s64 min_slice = max_t(s64, SCX_RESCUE_MIN_SLICE_US * NSEC_PER_USEC, TICK_NSEC);
+	u32 depth = rq->scx.rescue.dsq.nr ?: 1;
+
+	return clamp(div_s64(scx_rescue_quantum_ns, depth), min_slice, scx_rescue_quantum_ns);
+}
+
+static void scx_rescue_timer_arm(struct rq *rq)
+{
+	struct timer_list *timer = &rq->scx.rescue.timer;
+	s64 delay = scx_rescue_quantum_ns / 4;	/* should be granular enough */
+
+	if (timer_pending(timer))
+		return;
+
+	/*
+	 * While the head waiter can't be admitted because the bucket is short
+	 * of a full quantum, stretch to the full funding delay.
+	 */
+	if (!rq->scx.rescue.curr && rq->scx.rescue.budget < scx_rescue_quantum_ns) {
+		s64 deficit = scx_rescue_quantum_ns - rq->scx.rescue.budget;
+
+		delay = max(delay,
+			    div_s64(deficit << SCHED_CAPACITY_SHIFT, scx_rescue_bw_1024));
+	}
+
+	/* +1 rounds up so the beat is due by the time the timer fires */
+	timer->expires = jiffies + nsecs_to_jiffies(delay) + 1;
+	add_timer_on(timer, cpu_of(rq));
+}
+
+/**
+ * scx_rescue_admit - Start rescuing @p on @rq
+ * @rq: rq @p is being admitted on
+ * @p: task being admitted, off any DSQ
+ * @slice: CPU time to grant
+ *
+ * The schedulers keep their normal control over @p and may preempt or reslice
+ * it. @slice is measured on served CPU time against the snapshot taken here, so
+ * neither shortens the rescue, see scx_rescue_charge() and scx_rescue_keep().
+ * Prolonged denial escalates into protected execution, see
+ * scx_rescue_timerfn().
+ */
+static void scx_rescue_admit(struct rq *rq, struct task_struct *p, s64 slice)
+{
+	lockdep_assert_rq_held(rq);
+	WARN_ON_ONCE(rq->scx.rescue.curr);
+
+	rq->scx.rescue.curr = p;
+	rq->scx.rescue.slice = slice;
+	rq->scx.rescue.exec_snap = p->se.sum_exec_runtime;
+	scx_set_task_slice(p, slice);
+	scx_rescue_timer_arm(rq);
+}
+
+/**
+ * scx_rescue_try_admit - Try to admit a freshly stranded task
+ * @rq: rq @p is being inserted on
+ * @p: stranded task being diverted to rescue
+ *
+ * One rescue at a time and earlier arrivals go first. Admission needs a full
+ * quantum of budget, spent as the rescue runs. Return %true if @p was admitted
+ * and should be inserted at the tail of @rq's local DSQ, %false if it has to
+ * park on the rescue DSQ, with the timer armed to admit it later.
+ */
+static bool scx_rescue_try_admit(struct rq *rq, struct task_struct *p)
+{
+	scx_rescue_accrue(rq);
+
+	if (!rq->scx.rescue.curr && list_empty(&rq->scx.rescue.dsq.list) &&
+	    rq->scx.rescue.budget >= scx_rescue_quantum_ns) {
+		scx_rescue_admit(rq, p, scx_rescue_quantum_ns);
+		return true;
+	}
+
+	scx_rescue_timer_arm(rq);
+	return false;
+}
+
+/**
+ * scx_rescue_check_overload - Eject the top rescue consumer on a stuck rescue
+ * @rq: rq whose rescue timer fired
+ *
+ * If the oldest waiter on @rq's rescue DSQ has been queued for too long, rescue
+ * demand on this cpu persistently exceeds the configured bandwidth. Eject the
+ * sub with the highest recent rescue consumption instead of letting the
+ * scheduler stall path blame the waiter's owner, who may just be crowded out.
+ */
+static void scx_rescue_check_overload(struct rq *rq)
+{
+	struct scx_sched *victim = NULL, *pos;
+	struct task_struct *p;
+	int cpu = cpu_of(rq);
+	u64 max_avg = 0;
+	u32 dur_ms;
+
+	lockdep_assert_rq_held(rq);
+
+	p = list_first_entry_or_null(&rq->scx.rescue.dsq.list, struct task_struct,
+				     scx.dsq_list.node);
+	if (!p)
+		return;
+
+	/* has the head waiter been queued for longer than the threshold? */
+	if (time_before(jiffies, p->scx.rescue_at + scx_rescue_overload_after))
+		return;
+
+	/*
+	 * Grace period after the last ejection on this cpu - the freed
+	 * bandwidth gets one threshold's worth of time to drain the backlog
+	 * before another sub is judged.
+	 */
+	if (time_before64(get_jiffies_64(), rq->scx.rescue.kill_at +
+			  scx_rescue_overload_after))
+		return;
+
+	list_for_each_entry_rcu(pos, &scx_sched_all, all) {
+		u64 avg = scx_rescue_decay_avg(per_cpu_ptr(pos->pcpu, cpu));
+
+		/* skip an already-exiting sub, else the ejection is wasted */
+		if (pos->level && avg > max_avg &&
+		    atomic_read(&pos->exit_kind) == SCX_EXIT_NONE) {
+			max_avg = avg;
+			victim = pos;
+		}
+	}
+	if (!victim)
+		return;
+
+	rq->scx.rescue.kill_at = get_jiffies_64();
+	dur_ms = jiffies_to_msecs(jiffies - p->scx.rescue_at);
+	__scx_exit(victim, SCX_EXIT_ERROR_RESCUE, 0, cpu,
+		   "used too much rescue CPU time (%llums) while %s[%d] waited %u.%03us to be rescued",
+		   div_u64(max_avg, NSEC_PER_MSEC), p->comm, p->pid, dur_ms / 1000,
+		   dur_ms % 1000);
+}
+
+/**
+ * scx_rescue_timerfn - Drive and pace rescue execution
+ * @timer: rq->scx.rescue.timer
+ *
+ * Runs every quarter quantum while a rescuee exists, pending or admitted, see
+ * scx_rescue_timer_arm(). The head waiter is admitted once the bucket holds a
+ * full quantum and granted its slice, see scx_rescue_next_slice(). A session
+ * whose budget accumulates over two quanta with the admitted rescuee still
+ * waiting escalates - the rescuee's remaining slice turns into protected
+ * execution and it preempts the current task. An overloaded rescue queue ejects
+ * the top consumer, see scx_rescue_check_overload().
+ */
+static void scx_rescue_timerfn(struct timer_list *timer)
+{
+	struct rq *rq = timer_container_of(rq, timer, scx.rescue.timer);
+	struct task_struct *p;
+
+	guard(rq_lock_irqsave)(rq);
+
+	p = rq->scx.rescue.curr;
+	if (!p && list_empty(&rq->scx.rescue.dsq.list))
+		return;
+
+	scx_rescue_accrue(rq);
+	scx_rescue_check_overload(rq);
+
+	if (!p) {
+		s64 slice = scx_rescue_next_slice(rq);
+
+		/* no rescue in progress */
+		if (rq->scx.rescue.budget < scx_rescue_quantum_ns)
+			goto out_arm;
+
+		/* there's enough budget to start rescuing the next one */
+		p = list_first_entry(&rq->scx.rescue.dsq.list, struct task_struct,
+				     scx.dsq_list.node);
+		scx_task_unlink_from_dsq(p, &rq->scx.rescue.dsq);
+		scx_rescue_admit(rq, p, slice);
+		scx_move_local_task_to_local_dsq(scx_task_sched(p), p, SCX_ENQ_IGNORE_CAPS,
+						 &rq->scx.rescue.dsq, rq);
+		if (sched_class_above(&ext_sched_class, rq->curr->sched_class))
+			resched_curr(rq);
+	} else if (p->scx.dsq && rq->scx.rescue.budget > 2 * scx_rescue_quantum_ns) {
+		/*
+		 * The rescuee waited for the CPU for too long. Escalate - grant
+		 * the unserved remainder, protect it from the schedulers and
+		 * preempt the current task. The slice is set before the
+		 * protection. Repeat beats only repeat the head move - the
+		 * slice write is refused on a protected task.
+		 */
+		scx_set_task_slice(p, scx_rescue_slice_remaining(rq));
+		p->scx.flags |= SCX_TASK_PROTECTED;
+		scx_task_unlink_from_dsq(p, &rq->scx.local_dsq);
+		scx_move_local_task_to_local_dsq(scx_task_sched(p), p,
+					SCX_ENQ_HEAD | SCX_ENQ_PREEMPT | SCX_ENQ_IGNORE_CAPS,
+					&rq->scx.local_dsq, rq);
+	}
+out_arm:
+	scx_rescue_timer_arm(rq);
+}
+
+/* flush out tasks waiting for rescue before a CPU goes down */
+void scx_rescue_flush(struct rq *rq)
+{
+	struct task_struct *p, *n;
+
+	lockdep_assert_rq_held(rq);
+
+	/* sched domain rebuilds call rq_offline with the CPU staying alive */
+	if (cpu_active(cpu_of(rq)))
+		return;
+
+	/* end the current rescue */
+	if (rq->scx.rescue.curr)
+		scx_task_slice_ended(rq, rq->scx.rescue.curr);
+
+	/* and flush out all pending ones */
+	list_for_each_entry_safe(p, n, &rq->scx.rescue.dsq.list, scx.dsq_list.node) {
+		scx_task_unlink_from_dsq(p, &rq->scx.rescue.dsq);
+		scx_move_local_task_to_local_dsq(scx_task_sched(p), p, SCX_ENQ_IGNORE_CAPS,
+						 &rq->scx.rescue.dsq, rq);
+	}
+
+	timer_delete(&rq->scx.rescue.timer);
+}
+
+void scx_rescue_dump(struct seq_buf *s, struct rq *rq)
+{
+	struct task_struct *p = rq->scx.rescue.curr;
+
+	scx_dump_line(s, "          rescue=%u budget=%lldus rescuing=%s[%d]",
+		      rq->scx.rescue.dsq.nr,
+		      div_s64(rq->scx.rescue.budget, NSEC_PER_USEC),
+		      p ? p->comm : "none", p ? p->pid : -1);
+}
+
+/*
+ * A scheduler whose stall watchdog is shorter than the overload threshold gets
+ * stall-killed over its parked waiters before the overload check can eject the
+ * actual top consumer. The root's knobs set the threshold, warn on any
+ * scheduler that doesn't fit it.
+ */
+static void scx_rescue_check_timeout(struct scx_sched *sch)
+{
+	if (!scx_rescue_bw_1024 || sch->watchdog_timeout > scx_rescue_overload_after)
+		return;
+
+	pr_warn("sched_ext: %s: watchdog timeout %ums <= rescue overload threshold %ums\n",
+		sch->ops.name, jiffies_to_msecs(sch->watchdog_timeout),
+		jiffies_to_msecs(scx_rescue_overload_after));
+}
+
+/* latch the rescue parameters on root scheduler enable */
+void scx_rescue_set_knobs(struct scx_sched *sch)
+{
+	s32 bw_ppt = sch->ops.rescue_bandwidth_ppt ?: SCX_RESCUE_DFL_BW_PPT;
+	s64 quantum_us = sch->ops.rescue_quantum_us ?: SCX_RESCUE_DFL_QUANTUM_US;
+	s64 period_ns;
+
+	if (sch->ops.rescue_bandwidth_ppt == SCX_RESCUE_DISABLE) {
+		scx_rescue_bw_1024 = 0;
+		return;
+	}
+
+	scx_rescue_bw_1024 = bw_ppt * SCHED_CAPACITY_SCALE / 1000;
+	scx_rescue_quantum_ns = max(quantum_us * NSEC_PER_USEC, TICK_NSEC);
+	scx_rescue_sat_delta_ns =
+		div_s64((4 * scx_rescue_quantum_ns + TICK_NSEC) << SCHED_CAPACITY_SHIFT,
+			scx_rescue_bw_1024);
+
+	/*
+	 * The overload threshold and the decay halflife scale with the funding
+	 * period - the time the bucket takes to fund one full quantum.
+	 */
+	period_ns = div_s64(scx_rescue_quantum_ns << SCHED_CAPACITY_SHIFT, scx_rescue_bw_1024);
+	scx_rescue_overload_after =
+		clamp(nsecs_to_jiffies(SCX_RESCUE_OVERLOAD_MULT * period_ns),
+		      msecs_to_jiffies(SCX_RESCUE_MIN_OVERLOAD_MS),
+		      msecs_to_jiffies(SCX_RESCUE_MAX_OVERLOAD_MS));
+	scx_rescue_decay_halflife = scx_rescue_overload_after / 4;
+
+	/* a single in-budget wait must not cross the overload trigger */
+	if (nsecs_to_jiffies(period_ns) > scx_rescue_overload_after / 2)
+		pr_warn("sched_ext: %s: rescue funding period %lldms > overload threshold %ums / 2\n",
+			sch->ops.name, div_s64(period_ns, NSEC_PER_MSEC),
+			jiffies_to_msecs(scx_rescue_overload_after));
+
+	scx_rescue_check_timeout(sch);
+}
+
+void scx_rescue_init(struct rq *rq)
+{
+	BUG_ON(scx_init_dsq(&rq->scx.rescue.dsq, SCX_DSQ_RESCUE, NULL));
+	timer_setup(&rq->scx.rescue.timer, scx_rescue_timerfn, TIMER_PINNED);
+	rq->scx.rescue.kill_at = get_jiffies_64();
+}
+
+/**
+ * scx_resolve_local_dsq - Pick the local, rescue or reject DSQ for an insert
+ * @sch: enqueuing sub-sched
+ * @rq: rq whose local DSQ @p targets
+ * @p: task being inserted
+ * @enq_flags: in/out, unhonored flags are cleared
+ *
+ * Return @rq's local DSQ if @sch holds the required caps on @rq's cid.
+ * Otherwise, return @rq's rescue DSQ if the insert carries %SCX_ENQ_RESCUE and
+ * rescue is enabled, or @rq's reject DSQ after recording the reenq reason on
+ * @p.
+ *
+ * %SCX_ENQ_IMMED, %SCX_ENQ_PREEMPT and %SCX_ENQ_HEAD are cleared when diverting
+ * to rescue or reject. %SCX_ENQ_PREEMPT is also cleared on a fallback
+ * migration-disabled admission.
+ *
+ * Bypass doesn't need special-casing as a bypassing sched's tasks are enqueued
+ * to and run by its nearest non-bypassing ancestor. If root is bypassing, it
+ * always holds all caps.
+ */
+struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *rq,
+					     struct task_struct *p, u64 *enq_flags)
+{
+	if (!scx_has_subs())
+		return &rq->scx.local_dsq;
+
+	s32 cid = __scx_cpu_to_cid(cpu_of(rq));
+	struct scx_sched *asch = rq->scx.remote_activate_sch ?: sch;
+	u64 needed = scx_caps_for_enq(*enq_flags);
+	u64 missing;
+
+	/*
+	 * On a remote activation the scheduling sched (@asch) differs from
+	 * @p's owner (@sch). Check caps against the scheduling sched.
+	 */
+	if (*enq_flags & SCX_ENQ_PREEMPT)
+		needed |= scx_caps_for_preempt(asch, rq, *enq_flags);
+	missing = scx_missing_caps(asch, cpu_of(rq), needed);
+
+	/* requirements met */
+	if (likely(!missing))
+		return &rq->scx.local_dsq;
+
+	/*
+	 * The task must run on this CPU regardless of caps: the rq is draining
+	 * offline (BPF scheduler bypassed), the task is migration-disabled, or a
+	 * migration is pending. Admit despite the missing caps and count it.
+	 * Refuse preemptions.
+	 */
+	if (unlikely(!scx_rq_online(rq) || is_migration_disabled(p) ||
+		     p->migration_pending)) {
+		__scx_add_event(sch, SCX_EV_SUB_FORCED_ADMIT, 1);
+		*enq_flags &= ~SCX_ENQ_PREEMPT;
+		return &rq->scx.local_dsq;
+	}
+
+	/*
+	 * Diverting to rescue or reject, neither of which honors IMMED, PREEMPT
+	 * or HEAD - a diversion has no priority and IMMED is not allowed on
+	 * non-local DSQs. Strip the enq and task flags along with the slice.
+	 */
+	*enq_flags &= ~(SCX_ENQ_IMMED | SCX_ENQ_PREEMPT | SCX_ENQ_HEAD |
+			SCX_ENQ_APPLY_SLICE | SCX_ENQ_SLICE_DFL);
+	p->scx.flags &= ~SCX_TASK_IMMED;
+
+	/* the enqueuer opted for rescue instead of rejection and reenqueue */
+	if ((*enq_flags & SCX_ENQ_RESCUE) && likely(scx_rescue_bw_1024)) {
+		__scx_add_event(sch, SCX_EV_SUB_RESCUE, 1);
+		if (scx_rescue_try_admit(rq, p))
+			return &rq->scx.local_dsq;
+
+		/* queueing, the overload trigger measures the wait from here */
+		p->scx.rescue_at = jiffies;
+		return &rq->scx.rescue.dsq;
+	}
+
+	p->scx.reenq_reason_caps = missing;
+	p->scx.reenq_reason_cid = cid;
+
+	return &rq->scx.reject_dsq;
+}
+
+/* @p lost the caps needed to stay on @rq's local DSQ? Record reason if so. */
+bool scx_task_reenq_on_cap_revoke(struct rq *rq, struct task_struct *p)
+{
+	u64 missing;
+
+	/* migration-disabled tasks and the rescuee are admitted capless */
+	if (is_migration_disabled(p) || p == scx_rescuee(rq))
+		return false;
+
+	missing = scx_missing_caps(scx_task_sched(p), cpu_of(rq), scx_caps_for_task(p));
+	if (likely(!missing))
+		return false;
+
+	p->scx.reenq_reason_caps = missing;
+	p->scx.reenq_reason_cid = __scx_cpu_to_cid(cpu_of(rq));
+	return true;
+}
+
+/*
+ * Drain @rq->scx.reject_dsq, reenqueueing each task so the BPF re-decides
+ * from p->scx.reenq_reason_*.
+ *
+ * A task can be re-rejected repeatedly. The reenqueue is bounded per task in
+ * scx_do_enqueue_task(), which ejects the owning sub past SCX_REENQ_MAX_REPEAT.
+ * Rejection can't happen for root.
+ */
+void scx_reenq_reject(struct rq *rq)
+{
+	LIST_HEAD(tasks);
+	struct task_struct *p, *n;
+
+	lockdep_assert_rq_held(rq);
+
+	if (!scx_has_subs() || list_empty(&rq->scx.reject_dsq.list))
+		return;
+
+	/*
+	 * Move to a private list so a task re-rejected by the
+	 * scx_do_enqueue_task() below isn't revisited this round.
+	 */
+	list_for_each_entry_safe(p, n, &rq->scx.reject_dsq.list, scx.dsq_list.node) {
+		/* migration_pending tasks should have bypassed to local DSQ */
+		if (WARN_ON_ONCE(p->migration_pending))
+			continue;
+
+		scx_dispatch_dequeue(rq, p);
+
+		if (WARN_ON_ONCE(p->scx.flags & SCX_TASK_REENQ_REASON_MASK))
+			p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
+		p->scx.flags |= SCX_TASK_REENQ_CAP;
+
+		list_add_tail(&p->scx.dsq_list.node, &tasks);
+	}
+
+	list_for_each_entry_safe(p, n, &tasks, scx.dsq_list.node) {
+		list_del_init(&p->scx.dsq_list.node);
+
+		scx_do_enqueue_task(rq, p, SCX_ENQ_REENQ, -1);
+
+		p->scx.flags &= ~SCX_TASK_REENQ_REASON_MASK;
+	}
+}
+
+/* record a caps change, see struct scx_caps_updated */
+static void caps_updated_record(struct scx_pshard *ps, const struct scx_cmask *cids, u64 caps,
+				struct list_head *to_deliver)
+{
+	struct scx_caps_updated *cu = &ps->caps_updated;
+
+	guard(raw_spinlock)(&cu->lock);
+	scx_cmask_or(&cu->cmask, cids);
+	cu->caps |= caps;
+	if (list_empty(&cu->node_in_flight))
+		list_add_tail(&cu->node_in_flight, to_deliver);
+}
+
+/* deliver queued caps_updated callbacks, see struct scx_caps_updated */
+static void caps_updated_deliver(struct list_head *to_deliver)
+{
+	struct scx_caps_updated *cu, *tmp;
+
+	list_for_each_entry_safe(cu, tmp, to_deliver, node_in_flight) {
+		struct scx_pshard *ps = container_of(cu, struct scx_pshard, caps_updated);
+		struct scx_sched *sch = ps->sch;
+
+		while (true) {
+			u64 caps = 0;
+
+			/*
+			 * During enable, has_op is set after ops.sub_attach(),
+			 * so !has_op means the op is absent or the sched isn't
+			 * live yet - e.g. caps grant from ops.sub_attach().
+			 * Either way don't consume - leave for
+			 * scx_sub_seed_caps() to deliver once live.
+			 */
+			scoped_guard (raw_spinlock, &cu->lock) {
+				if (cu->caps && SCX_HAS_OP(sch, sub_caps_updated) &&
+				    likely(!READ_ONCE(sch->aborting))) {
+					struct scx_cmask_ref ref;
+
+					caps = cu->caps;
+					scx_cmask_ref_init_kern(sch, cu->cmask_arena_out,
+								ps->base, ps->nr_cids, &ref);
+					scx_cmask_ref_copy(&ref, &cu->cmask);
+					scx_cmask_clear(&cu->cmask);
+					cu->caps = 0;
+				} else {
+					list_del_init(&cu->node_in_flight);
+				}
+			}
+			if (!caps)
+				break;
+
+			/* caps != 0 only when deliverable (has_op, above) */
+			SCX_CALL_OP(sch, sub_caps_updated, NULL, cu->cmask_arena_out, caps);
+		}
+	}
+}
+
+/*
+ * Deliver caps owed to @sch that couldn't be delivered earlier (e.g. a grant
+ * taken during its sub_attach(), before has_op was set). Called once @sch is
+ * enabled.
+ */
+static void scx_sub_seed_caps(struct scx_sched *sch)
+{
+	LIST_HEAD(to_deliver);
+	s32 si;
+
+	guard(irqsave)();
+
+	for (si = 0; si < sch->nr_pshards; si++) {
+		struct scx_pshard *ps = sch->pshard[si];
+		struct scx_caps_updated *cu = &ps->caps_updated;
+
+		scoped_guard (raw_spinlock, &cu->lock) {
+			if (cu->caps && list_empty(&cu->node_in_flight))
+				list_add_tail(&cu->node_in_flight, &to_deliver);
+		}
+	}
+	caps_updated_deliver(&to_deliver);
+}
+
+static u64 calc_effective_caps(struct scx_pshard *ps, s32 cid)
+{
+	u64 ecaps = 0;
+	u32 cap_bit;
+
+	for (cap_bit = 0; cap_bit < __SCX_NR_CAPS; cap_bit++)
+		if (scx_cmask_test(cid, &ps->caps[cap_bit].cmask))
+			ecaps |= BIT_U64(cap_bit) | scx_caps_implied(BIT_U64(cap_bit));
+	return ecaps;
+}
+
+/**
+ * queue_sync_ecaps - Queue ecaps update for a (sch, cid) pair
+ * @sch: sched to update
+ * @cid: cid to update
+ *
+ * Queue an ecaps update for @sch's @cid and kick the cpu so that it syncs in
+ * dispatch_one().
+ */
+static void queue_sync_ecaps(struct scx_sched *sch, s32 cid)
+{
+	s32 cpu = __scx_cid_to_cpu(cid);
+	struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
+
+	/*
+	 * Pairs with smp_mb() in scx_process_sync_ecaps(). Either the check
+	 * below sees the node off the list and queues it, or the in-flight sync
+	 * sees the caps[] update made before this call.
+	 */
+	smp_mb();
+
+	/* @cid's pshard->lock excludes concurrent queueing attempts */
+	if (llist_on_list(&pcpu->ecaps_to_sync_node))
+		return;
+	if (llist_add(&pcpu->ecaps_to_sync_node, &cpu_rq(cpu)->scx.ecaps_to_sync))
+		scx_kick_cpu(sch->ancestors[0], cpu, 0);
+}
+
+/* discard @rq's queued ecaps syncs */
+static void discard_queued_syncs(struct rq *rq)
+{
+	struct llist_node *pos, *tmp;
+
+	lockdep_assert_rq_held(rq);
+
+	llist_for_each_safe(pos, tmp, llist_del_all(&rq->scx.ecaps_to_sync))
+		init_llist_node(pos);
+}
+
+/**
+ * scx_process_sync_ecaps - Sync this cpu's ecaps to pshard->caps[]
+ * @rq: the cid's cpu rq
+ * @prev: @rq's previous task from the in-progress dispatch
+ *
+ * pshard->caps[] is the target configuration. pcpu->ecaps is the effective
+ * transposed copy owned by the cid's cpu and written only here under @rq's
+ * lock.
+ *
+ * A sched that newly gains baseline access here is owed an update_idle() so it
+ * learns the cid's idle state. Such a gain arms the per-rq
+ * %SCX_RQ_SUB_IDLE_RENOTIFY gate so the next idle pick delivers it.
+ */
+void scx_process_sync_ecaps(struct rq *rq, struct task_struct *prev)
+{
+	s32 cpu = cpu_of(rq);
+	s32 cid, shard;
+	struct llist_node *batch, *pos, *tmp;
+	u64 lost_all = 0;
+
+	lockdep_assert_rq_held(rq);
+
+	if (!scx_has_subs() || likely(llist_empty(&rq->scx.ecaps_to_sync)))
+		return;
+
+	/*
+	 * ecaps are zeroed while the cpu is inactive and must stay zero.
+	 * Discard queued syncs instead of processing them - the
+	 * scx_online_ecaps() reseed re-syncs every sched on activation.
+	 * cpu_active() clears before the offline zeroing and sets before the
+	 * reseed is queued, so this test can neither miss a racing sync nor
+	 * eat the reseed.
+	 */
+	if (unlikely(!cpu_active(cpu))) {
+		discard_queued_syncs(rq);
+		return;
+	}
+
+	/* @cid is valid here: the cpu is active with queued syncs */
+	cid = __scx_cpu_to_cid(cpu);
+	shard = rcu_dereference_all(scx_cid_to_shard)[cid];
+
+	batch = llist_del_all(&rq->scx.ecaps_to_sync);
+	llist_for_each_safe(pos, tmp, batch) {
+		struct scx_sched_pcpu *pcpu =
+			container_of(pos, struct scx_sched_pcpu, ecaps_to_sync_node);
+		struct scx_pshard *ps = pcpu->sch->pshard[shard];
+		u64 old, ecaps, lost, gained;
+
+		init_llist_node(pos);
+
+		/* pairs with smp_mb() in queue_sync_ecaps(), see there */
+		smp_mb();
+
+		old = READ_ONCE(pcpu->ecaps);
+		ecaps = calc_effective_caps(ps, cid);
+		WRITE_ONCE(pcpu->ecaps, ecaps);
+
+		lost = old & ~ecaps;
+		gained = ecaps & ~old;
+		lost_all |= lost;
+
+		/*
+		 * Tell the sched its effective caps on this cid changed. The
+		 * invocation is equivalent to the dispatch path and may drop
+		 * and re-acquire the rq lock temporarily while the rest of
+		 * @batch is held privately, see scx_discard_ecaps_to_sync().
+		 * The dispatch kfuncs resolve their context on the executing
+		 * cpu, which under core scheduling can differ from @rq's cpu,
+		 * so the context is set up there. The rq recorded in it keeps
+		 * the dispatches targeting @rq.
+		 */
+		if (ecaps != pcpu->reported_ecaps &&
+		    SCX_HAS_OP(pcpu->sch, sub_ecaps_updated) &&
+		    !scx_bypassing(pcpu->sch, cpu)) {
+			struct scx_dsp_ctx *dspc = &this_cpu_ptr(pcpu->sch->pcpu)->dsp_ctx;
+
+			dspc->rq = rq;
+			/* stash @prev so nested dispatches can access it */
+			rq->scx.sub_dispatch_prev = prev;
+			SCX_CALL_OP(pcpu->sch, sub_ecaps_updated, rq, scx_cpu_arg(cpu),
+				    pcpu->reported_ecaps, ecaps);
+			rq->scx.sub_dispatch_prev = NULL;
+			scx_flush_dispatch_buf(pcpu->sch, rq);
+			pcpu->reported_ecaps = ecaps;
+		}
+
+		/*
+		 * Gaining baseline access owes an update_idle() so the sched
+		 * learns the cpu's idle state. Arm the per-rq gate so the next
+		 * idle pick flushes it. Losing access drops any pending notify.
+		 */
+		if (gained & SCX_CAP_BASE) {
+			pcpu->idle_renotify = true;
+			rq->scx.flags |= SCX_RQ_SUB_IDLE_RENOTIFY;
+		} else if (lost & SCX_CAP_BASE) {
+			pcpu->idle_renotify = false;
+		}
+	}
+
+	/*
+	 * Losing a cap can strand already-queued tasks. Schedule a reenq scan
+	 * to move the now-capless ones off the local DSQ. The scan tests
+	 * against the effective caps and thus must come after the ecaps sync.
+	 */
+	if (lost_all & SCX_CAPS_REENQ_ON_LOSS)
+		scx_schedule_reenq_local(rq, SCX_REENQ_CAP_REVOKE);
+}
+
+/**
+ * scx_unbypass_replay_ecaps - Replay a bypass-suppressed ecaps notification
+ * @rq: rq of the cpu leaving bypass
+ * @sch: scheduler that just left bypass on @rq's cpu
+ *
+ * scx_process_sync_ecaps() consumes syncs while bypassing without delivering
+ * ops.sub_ecaps_updated(), leaving reported_ecaps stale. Nothing re-queues a
+ * sync when bypass lifts, so without a replay a cid that never changes again
+ * would never be notified. The attach-time initial grants are the acute case
+ * as they are consumed during the enable bypass window. Re-queue a sync for
+ * any undelivered delta so the next dispatch delivers it.
+ */
+void scx_unbypass_replay_ecaps(struct rq *rq, struct scx_sched *sch)
+{
+	s32 cpu = cpu_of(rq);
+	struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu);
+	struct scx_pshard *ps;
+	s32 cid;
+
+	lockdep_assert_rq_held(rq);
+
+	/* root holds every cap and never uses ecaps */
+	if (!sch->level)
+		return;
+
+	if (READ_ONCE(pcpu->ecaps) == pcpu->reported_ecaps)
+		return;
+
+	cid = __scx_cpu_to_cid(cpu);
+	ps = sch->pshard[rcu_dereference_all(scx_cid_to_shard)[cid]];
+
+	guard(raw_spinlock)(&ps->lock);
+	queue_sync_ecaps(sch, cid);
+}
+
+/*
+ * A cpu came back. Re-seed each sub-sched's ecaps on the cpu's cid. The sync
+ * recomputes effective caps from the pshard and fires ops.sub_ecaps_updated()
+ * only on a real change since offline.
+ */
+void scx_online_ecaps(struct rq *rq)
+{
+	struct scx_sched *root, *pos;
+	s32 cid, shard;
+
+	/*
+	 * Only a live hierarchy can have ecaps to reseed. This also keeps the
+	 * table reads below away from an enable that failed before publishing
+	 * the tables. A concurrent disable can't retire them, see
+	 * handle_hotplug().
+	 */
+	if (!scx_enabled())
+		return;
+
+	guard(rq_lock_irqsave)(rq);
+
+	root = scx_root_protected();
+	cid = __scx_cpu_to_cid(cpu_of(rq));
+	shard = rcu_dereference_all(scx_cid_to_shard)[cid];
+
+	scx_for_each_descendant_pre(pos, root) {
+		struct scx_pshard *ps;
+
+		/* root holds every cap and never uses ecaps */
+		if (!pos->level)
+			continue;
+
+		ps = pos->pshard[shard];
+		guard(raw_spinlock)(&ps->lock);
+		queue_sync_ecaps(pos, cid);
+	}
+}
+
+/*
+ * A cpu is going down. Zero each sub-sched's in-effect ecaps so cap checks
+ * treat the cpu as capless while offline. Pending and late-queued syncs are
+ * discarded at consumption by scx_process_sync_ecaps() while the cpu is
+ * inactive. Leave reported_ecaps. Ownership is unchanged, so the
+ * scx_online_ecaps() reseed reports only a genuine delta. No callback fires
+ * here.
+ */
+void scx_offline_ecaps(struct rq *rq)
+{
+	s32 cpu = cpu_of(rq);
+	struct scx_sched *root, *pos;
+
+	guard(rq_lock_irqsave)(rq);
+
+	root = scx_root_protected();
+
+	scx_for_each_descendant_pre(pos, root) {
+		/* root holds every cap and never uses ecaps */
+		if (!pos->level)
+			continue;
+
+		WRITE_ONCE(per_cpu_ptr(pos->pcpu, cpu)->ecaps, 0);
+	}
+}
+
+/*
+ * @pcpu's sched was unhashed before the grace period, so nothing re-queues its
+ * sync node. Remove the node from @rq's pending list so the pcpu can be freed.
+ */
+void scx_discard_ecaps_to_sync(s32 cpu, struct scx_sched_pcpu *pcpu)
+{
+	struct rq *rq = cpu_rq(cpu);
+	struct llist_node *head = NULL, *tail = NULL;
+	struct llist_node *pos, *tmp;
+
+	/*
+	 * llist can't unlink a single node. Take all queued nodes, drop @pcpu's
+	 * and resplice the rest. Nodes in the taken batch read as on-list
+	 * throughout, so queue_sync_ecaps() stays correct.
+	 */
+	if (llist_on_list(&pcpu->ecaps_to_sync_node)) {
+		scoped_guard (rq_lock_irqsave, rq) {
+			llist_for_each_safe(pos, tmp, llist_del_all(&rq->scx.ecaps_to_sync)) {
+				if (pos == &pcpu->ecaps_to_sync_node) {
+					init_llist_node(pos);
+				} else {
+					pos->next = head;
+					head = pos;
+					if (!tail)
+						tail = pos;
+				}
+			}
+			if (head)
+				llist_add_batch(head, tail, &rq->scx.ecaps_to_sync);
+		}
+	}
+
+	/*
+	 * An in-flight scx_process_sync_ecaps() batch may still hold the node
+	 * privately across dispatch-induced rq unlocks, reading as on-list.
+	 *
+	 * Because a bypassing sched gets no op call, init_llist_node() and all
+	 * @pcpu accesses share one contiguous lock hold, off-list under the rq
+	 * lock means @pcpu won't be accessed again.
+	 */
+	while (true) {
+		scoped_guard (rq_lock_irqsave, rq) {
+			if (!llist_on_list(&pcpu->ecaps_to_sync_node))
+				return;
+		}
+		cpu_relax();
+	}
+}
+
+/**
+ * scx_discard_stale_ecaps_syncs - Discard ecaps syncs from earlier schedulers
+ *
+ * To be called during root enable before the scheduler goes live. An earlier
+ * root's sub-sched may not have gone through its RCU free path yet (e.g. a
+ * still-open link fd defers it) and can leave queued ecaps syncs behind.
+ * Processing them would decode the dead sched's pshards with the current cid
+ * layout. Discard them instead. The backing scx_sched_pcpu's are still
+ * allocated as the free path removes ecaps_to_sync_node before freeing.
+ */
+void scx_discard_stale_ecaps_syncs(void)
+{
+	s32 cpu;
+
+	for_each_possible_cpu(cpu) {
+		struct rq *rq = cpu_rq(cpu);
+
+		guard(rq_lock_irqsave)(rq);
+		discard_queued_syncs(rq);
+	}
+}
+
+static DECLARE_WAIT_QUEUE_HEAD(scx_unlink_waitq);
+
+void drain_descendants(struct scx_sched *sch)
+{
+	/*
+	 * Child scheds that finished the critical part of disabling will take
+	 * themselves off @sch->children. Wait for it to drain. As propagation
+	 * is recursive, empty @sch->children means that all proper descendant
+	 * scheds reached unlinking stage.
+	 */
+	wait_event(scx_unlink_waitq, list_empty(&sch->children));
+}
+
+/**
+ * scx_rehome_task - Move a task to a sched it has been initialized for
+ * @to: sched taking over @p, @p's init on it already complete
+ * @p: task to re-home
+ *
+ * Exit @p from its current sched and switch it over to @to, overriding the
+ * state to %SCX_TASK_READY to account for the already completed init. A task
+ * on a non-ext class, possible under an %SCX_OPS_SWITCH_PARTIAL root, stays
+ * %READY and is enabled by switching_to_scx() if it switches over.
+ */
+static void scx_rehome_task(struct scx_sched *to, struct task_struct *p)
+{
+	lockdep_assert_held(&p->pi_lock);
+	lockdep_assert_rq_held(task_rq(p));
+
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
+		scx_disable_and_exit_task(scx_task_sched(p), p);
+		scx_set_task_state(p, SCX_TASK_INIT_BEGIN);
+		scx_set_task_state(p, SCX_TASK_INIT);
+		scx_set_task_sched(p, to);
+		scx_set_task_state(p, SCX_TASK_READY);
+		if (p->sched_class == &ext_sched_class)
+			scx_enable_task(to, p);
+	}
+}
+
+/**
+ * scx_punt_task - Hand a task to a failed sched without initialization
+ * @to: failed and bypassed sched taking custody of @p
+ * @p: task to punt
+ *
+ * Take @p off its current sched and put it on @to at %SCX_TASK_NONE. @to is
+ * dying and its teardown will re-home @p properly.
+ *
+ * Used when @to must take over @p but failed to initialize it. Bypass keeps
+ * scheduling decisions away from @to but @p can still trigger its task ops,
+ * which may confuse the BPF side. @to is dying anyway. The exit paths skip
+ * %NONE tasks (see __scx_disable_and_exit_task() and switched_from_scx()).
+ */
+static void scx_punt_task(struct scx_sched *to, struct task_struct *p)
+{
+	lockdep_assert_held(&p->pi_lock);
+	lockdep_assert_rq_held(task_rq(p));
+	WARN_ON_ONCE(!READ_ONCE(to->bypass_depth));
+
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
+		scx_disable_and_exit_task(scx_task_sched(p), p);
+		scx_set_task_sched(p, to);
+	}
+}
+
+static void scx_fail_parent(struct scx_sched *sch,
+			    struct task_struct *failed, s32 fail_code)
+{
+	struct scx_sched *parent = scx_parent(sch);
+	struct scx_task_iter sti;
+	struct task_struct *p;
+
+	scx_error(parent, "ops.init_task() failed (%d) for %s[%d] while disabling a sub-scheduler",
+		  fail_code, failed->comm, failed->pid);
+
+	/*
+	 * Once $parent is bypassed, tasks can be punted into it. This may
+	 * cause downstream failures on the BPF side but $parent is dying
+	 * anyway.
+	 */
+	scx_bypass(parent, true);
+
+	scx_task_iter_start(&sti, sch->cgrp);
+	while ((p = scx_task_iter_next_locked(&sti))) {
+		if (scx_task_on_sched(parent, p))
+			continue;
+
+		scx_punt_task(parent, p);
+	}
+	scx_task_iter_stop(&sti);
+}
+
+#ifdef CONFIG_EXT_GROUP_SCHED
+/**
+ * scx_cgroup_claim_subtree - Claim the subtree's cgroups for an enabling sub
+ * @sch: sub-scheduler being enabled
+ *
+ * Called while enabling @sch, after the subtree's cgrp->scx_sched's are pointed
+ * at @sch and before any task is claimed. This mirrors root enable's
+ * cgroups-before-tasks order. The ops.init_task() args are task_group-granular
+ * and can still reference a cgroup outside the handed-over set when the cpu
+ * controller is coarser than the sub topology or mounted on cgroup1.
+ *
+ * First init each of the parent sched's subtree cgroups on @sch, and only then
+ * exit them from the parent, so that a failed init can be unwound with the
+ * parent untouched. The both-inited transient is invisible outside
+ * scx_cgroup_lock(). %SCX_TG_SUB_INIT tracks the first pass's progress.
+ * %SCX_TG_INITED stays set throughout, except for a task_group whose
+ * ops.cgroup_init() failed on the parent (see scx_cgroup_return_subtree()):
+ * there is nothing to exit from the parent and %SCX_TG_INITED is set back with
+ * the transfer.
+ *
+ * Dying but not yet offlined task_groups are included: a removed cgroup keeps
+ * hosting scheduling events until its dying tasks finish their final context
+ * switches, so it still needs to be inited on a sched, and its offline-time
+ * ops.cgroup_exit() follows the last of those events.
+ *
+ * Return 0 on success, -errno on failure. On failure, @sch has been
+ * scx_error()'d and is left with no cgroups.
+ */
+static s32 scx_cgroup_claim_subtree(struct scx_sched *sch)
+{
+	struct cgroup *sub_cgrp = sch_cgroup(sch);
+	struct cgroup_subsys_state *ecss = cgroup_e_css(sub_cgrp, &cpu_cgrp_subsys);
+	struct scx_sched *parent = scx_parent(sch);
+	struct cgroup_subsys_state *css;
+	int ret;
+
+	css_for_each_descendant_pre(css, ecss) {
+		struct task_group *tg = css_tg(css);
+		struct scx_cgroup_init_args args = {
+			.weight = tg->scx.weight,
+			.bw_period_us = tg->scx.bw_period_us,
+			.bw_quota_us = tg->scx.bw_quota_us,
+			.bw_burst_us = tg->scx.bw_burst_us,
+		};
+
+		if (tg->scx.sched != parent ||
+		    !cgroup_is_descendant(css->cgroup, sub_cgrp))
+			continue;
+
+		if (SCX_HAS_OP(sch, cgroup_init)) {
+			ret = SCX_CALL_OP_RET(sch, cgroup_init, NULL, css->cgroup, &args);
+			if (ret) {
+				scx_error(sch, "ops.cgroup_init() failed (%d)", ret);
+				goto err;
+			}
+		}
+		tg->scx.flags |= SCX_TG_SUB_INIT;
+	}
+
+	css_for_each_descendant_post(css, ecss) {
+		struct task_group *tg = css_tg(css);
+
+		/*
+		 * SUB_INIT is pass 1's progress mark: pass 2 and the err path
+		 * must visit exactly the tgs pass 1 inited.
+		 */
+		if (!(tg->scx.flags & SCX_TG_SUB_INIT))
+			continue;
+
+		/* skip the exit if the parent's ops.cgroup_init() failed */
+		if ((tg->scx.flags & SCX_TG_INITED) && SCX_HAS_OP(parent, cgroup_exit))
+			SCX_CALL_OP(parent, cgroup_exit, NULL, css->cgroup);
+		tg->scx.sched = sch;
+		tg->scx.flags |= SCX_TG_INITED;
+		tg->scx.flags &= ~SCX_TG_SUB_INIT;
+	}
+
+	return 0;
+
+err:
+	css_for_each_descendant_post(css, ecss) {
+		struct task_group *tg = css_tg(css);
+
+		if (!(tg->scx.flags & SCX_TG_SUB_INIT))
+			continue;
+
+		if (SCX_HAS_OP(sch, cgroup_exit))
+			SCX_CALL_OP(sch, cgroup_exit, NULL, css->cgroup);
+		tg->scx.flags &= ~SCX_TG_SUB_INIT;
+	}
+	return ret;
+}
+
+/**
+ * scx_cgroup_return_subtree - Return the subtree's cgroups to the parent sched
+ * @sch: sub-scheduler being disabled
+ *
+ * Called while disabling @sch, after the subtree's cgrp->scx_sched's are reset
+ * to the parent sched and before tasks are re-homed, mirroring root disable's
+ * cgroups-before-tasks teardown order. The reverse of
+ * scx_cgroup_claim_subtree(): exit @sch's cgroups from @sch, then init them on
+ * the parent with the current tg->scx.* values, resyncing settings that changed
+ * while @sch had them.
+ *
+ * When an init on the parent fails, the parent is failed - the same policy as
+ * task re-homing. The remaining task_groups are punted: they move to the parent
+ * anyway with %SCX_TG_INITED cleared, as ops.cgroup_init() failed or never ran
+ * for them. A punted task_group gets no cgroup ops. The dying parent's own
+ * disable moves it one sched up, initing it there. Root ends the chain: root
+ * teardown drops cgroup ops entirely and the next enable's bulk init re-inits
+ * every online task_group.
+ *
+ * The task re-home that follows still delivers ops.init_task() to the dying
+ * parent, including for tasks in punted cgroups it never inited - tolerated
+ * like the downstream failures of task punting (see scx_punt_task()).
+ */
+static void scx_cgroup_return_subtree(struct scx_sched *sch)
+{
+	struct cgroup *sub_cgrp = sch_cgroup(sch);
+	struct cgroup_subsys_state *ecss = cgroup_e_css(sub_cgrp, &cpu_cgrp_subsys);
+	struct scx_sched *parent = scx_parent(sch);
+	struct cgroup_subsys_state *css;
+	bool parent_failed = false;
+	int ret;
+
+	css_for_each_descendant_post(css, ecss) {
+		struct task_group *tg = css_tg(css);
+
+		if (tg->scx.sched != sch ||
+		    !cgroup_is_descendant(css->cgroup, sub_cgrp))
+			continue;
+
+		/* skip the exit if @sch's ops.cgroup_init() failed for the tg */
+		if ((tg->scx.flags & SCX_TG_INITED) && SCX_HAS_OP(sch, cgroup_exit))
+			SCX_CALL_OP(sch, cgroup_exit, NULL, css->cgroup);
+		tg->scx.sched = parent;
+		tg->scx.flags |= SCX_TG_SUB_INIT;
+	}
+
+	css_for_each_descendant_pre(css, ecss) {
+		struct task_group *tg = css_tg(css);
+		struct scx_cgroup_init_args args = {
+			.weight = tg->scx.weight,
+			.bw_period_us = tg->scx.bw_period_us,
+			.bw_quota_us = tg->scx.bw_quota_us,
+			.bw_burst_us = tg->scx.bw_burst_us,
+		};
+
+		/* the first pass must have transferred everything */
+		WARN_ON_ONCE(tg->scx.sched == sch);
+
+		/*
+		 * SUB_INIT distinguishes the tgs pass 1 moved. The sched test
+		 * can't: a tg punted to the parent by an earlier failure would
+		 * also match.
+		 */
+		if (!(tg->scx.flags & SCX_TG_SUB_INIT))
+			continue;
+		tg->scx.flags &= ~(SCX_TG_SUB_INIT | SCX_TG_INITED);
+
+		/*
+		 * A re-init on $parent failed. The task_groups from here on are
+		 * punted: they stay on the dying $parent with INITED clear and
+		 * move onward when it disables.
+		 */
+		if (parent_failed)
+			continue;
+
+		if (SCX_HAS_OP(parent, cgroup_init)) {
+			ret = SCX_CALL_OP_RET(parent, cgroup_init, NULL, css->cgroup, &args);
+			if (ret) {
+				scx_error(parent, "ops.cgroup_init() failed (%d) while disabling a sub-scheduler",
+					  ret);
+				parent_failed = true;
+				continue;
+			}
+		}
+		tg->scx.flags |= SCX_TG_INITED;
+	}
+}
+#else
+static inline s32 scx_cgroup_claim_subtree(struct scx_sched *sch) { return 0; }
+static inline void scx_cgroup_return_subtree(struct scx_sched *sch) {}
+#endif
+
+void scx_sub_disable(struct scx_sched *sch)
+{
+	struct scx_sched *parent = scx_parent(sch);
+	struct scx_task_iter sti;
+	struct task_struct *p;
+	int ret;
+
+	/*
+	 * Guarantee forward progress and wait for descendants to be disabled.
+	 * To limit disruptions, $parent is not bypassed. Tasks are fully
+	 * prepped and then inserted back into $parent.
+	 */
+	scx_bypass(sch, true);
+	drain_descendants(sch);
+
+	/*
+	 * Here, every runnable task is guaranteed to make forward progress and
+	 * we can safely use blocking synchronization constructs. Actually
+	 * disable ops.
+	 */
+	mutex_lock(&scx_enable_mutex);
+	percpu_down_write(&scx_fork_rwsem);
+	scx_cgroup_lock();
+
+	/*
+	 * An enable that failed before scx_link_sched() succeeded never owned a
+	 * cgroup or task and won't be waited on by an ancestor's
+	 * drain_descendants(). Nothing to reparent and walking the tasks can
+	 * misbehave as the task ownership invariant (either owned by self or
+	 * parent) does not hold. ->sibling can't identify this case - an undone
+	 * link leaves it non-empty.
+	 */
+	if (!sch->linked)
+		goto dump;
+
+	set_cgroup_sched(sch_cgroup(sch), parent);
+
+	/*
+	 * Return the subtree's cgroups before re-homing tasks so that any
+	 * ops.init_task() on $parent only sees cgroups it has initialized.
+	 */
+	scx_cgroup_return_subtree(sch);
+
+	scx_task_iter_start(&sti, sch->cgrp);
+	while ((p = scx_task_iter_next_locked(&sti))) {
+		struct rq *rq;
+		struct rq_flags rf;
+
+		/* filter out duplicate visits */
+		if (scx_task_on_sched(parent, p))
+			continue;
+
+		/*
+		 * By the time control reaches here, all linked descendant
+		 * schedulers should have been disabled.
+		 */
+		WARN_ON_ONCE(!scx_task_on_sched(sch, p));
+
+		/*
+		 * @p is pinned by the iter: css_task_iter_next() takes a
+		 * reference and holds it until the next iter_next() call, so
+		 * @p->usage is guaranteed > 0.
+		 */
+		get_task_struct(p);
+
+		scx_task_iter_unlock(&sti);
+
+		/*
+		 * $p is READY or ENABLED on @sch. Initialize for $parent,
+		 * disable and exit from @sch, and then switch over to $parent.
+		 *
+		 * If a task fails to initialize for $parent, the only available
+		 * action is disabling $parent too. While this allows disabling
+		 * of a child sched to cause the parent scheduler to fail, the
+		 * failure can only originate from ops.init_task() of the
+		 * parent. A child can't directly affect the parent through its
+		 * own failures.
+		 */
+		ret = __scx_init_task(parent, p, NULL, false);
+		if (ret) {
+			scx_fail_parent(sch, p, ret);
+			put_task_struct(p);
+			break;
+		}
+
+		rq = task_rq_lock(p, &rf);
+
+		if (scx_get_task_state(p) == SCX_TASK_DEAD) {
+			/*
+			 * sched_ext_dead() raced us between __scx_init_task()
+			 * and this rq lock and ran exit_task() on @sch (the
+			 * sched @p was on at that point), not on $parent.
+			 * $parent's just-completed init is owed an exit_task()
+			 * and we issue it here.
+			 */
+			scx_sub_init_cancel_task(parent, p);
+			task_rq_unlock(rq, p, &rf);
+			put_task_struct(p);
+			continue;
+		}
+
+		scx_rehome_task(parent, p);
+
+		task_rq_unlock(rq, p, &rf);
+		put_task_struct(p);
+	}
+	scx_task_iter_stop(&sti);
+
+dump:
+	scx_disable_dump(sch);
+
+	scx_cgroup_unlock();
+	percpu_up_write(&scx_fork_rwsem);
+
+	/*
+	 * All tasks are moved off of @sch but there may still be on-going
+	 * operations (e.g. ops.select_cpu()). Drain them by flushing RCU. Use
+	 * the expedited version as ancestors may be waiting in bypass mode.
+	 * Also, tell the parent that there is no need to keep running bypass
+	 * DSQs for us.
+	 */
+	synchronize_rcu_expedited();
+	scx_disable_bypass_dsp(sch);
+
+	scx_unlink_sched(sch);
+
+	mutex_unlock(&scx_enable_mutex);
+
+	/*
+	 * @sch is now unlinked from the parent's children list. Notify and call
+	 * ops.sub_detach/exit(). Note that ops.sub_detach/exit() must be called
+	 * after unlinking and releasing all locks. See scx_claim_exit().
+	 */
+	wake_up_all(&scx_unlink_waitq);
+
+	if (parent->ops.sub_detach && sch->sub_attached) {
+		struct scx_sub_detach_args sub_detach_args = {
+			.ops = &sch->ops,
+			.cgroup_path = sch->cgrp_path,
+		};
+		SCX_CALL_OP(parent, sub_detach, NULL,
+			    &sub_detach_args);
+	}
+
+	scx_log_sched_disable(sch);
+
+	if (sch->ops.exit)
+		SCX_CALL_OP(sch, exit, NULL, sch->exit_info);
+
+	/*
+	 * @sch's non-ops programs such as timers and tracers can fire after
+	 * ops.exit(). Now that exit is complete, stop scx_prog_sched() from
+	 * resolving to @sch and drain in-flight resolvers.
+	 */
+	WRITE_ONCE(sch->dead, true);
+	synchronize_rcu();
+
+	if (sch->sub_kset)
+		kobject_del(&sch->sub_kset->kobj);
+	/* not added if enable failed before scx_sched_sysfs_add() */
+	if (sch->kobj.state_in_sysfs)
+		kobject_del(&sch->kobj);
+}
+
+/* verify that a scheduler can be attached to @cgrp and return the parent */
+static struct scx_sched *find_parent_sched(struct cgroup *cgrp)
+{
+	struct scx_sched *parent = scx_cgroup_sched(cgrp);
+	struct scx_sched *pos;
+
+	lockdep_assert_held(&scx_sched_lock);
+
+	/* can't attach twice to the same cgroup */
+	if (parent->cgrp == cgrp)
+		return ERR_PTR(-EBUSY);
+
+	/* does $parent allow sub-scheds? */
+	if (!parent->ops.sub_attach)
+		return ERR_PTR(-EOPNOTSUPP);
+
+	/* can't insert between $parent and its exiting children */
+	list_for_each_entry(pos, &parent->children, sibling)
+		if (cgroup_is_descendant(pos->cgrp, cgrp))
+			return ERR_PTR(-EBUSY);
+
+	return parent;
+}
+
+static bool assert_task_ready_or_enabled(struct task_struct *p)
+{
+	u32 state = scx_get_task_state(p);
+
+	switch (state) {
+	case SCX_TASK_READY:
+	case SCX_TASK_ENABLED:
+		return true;
+	default:
+		WARN_ONCE(true, "sched_ext: Invalid task state %d for %s[%d] during enabling sub sched",
+			  state, p->comm, p->pid);
+		return false;
+	}
+}
+
+void scx_sub_enable_workfn(struct kthread_work *work)
+{
+	struct scx_enable_cmd *cmd = container_of(work, struct scx_enable_cmd, work);
+	struct sched_ext_ops *ops = cmd->ops;
+	struct cgroup *cgrp;
+	struct scx_sched *parent, *sch;
+	struct scx_task_iter sti;
+	struct task_struct *p;
+	s32 i, ret;
+
+	mutex_lock(&scx_enable_mutex);
+
+	if (!scx_enabled()) {
+		ret = -ENODEV;
+		goto out_unlock;
+	}
+
+	/* See scx_root_enable_workfn() for the @ops->priv check. */
+	if (rcu_access_pointer(ops->priv)) {
+		ret = -EBUSY;
+		goto out_unlock;
+	}
+
+	cgrp = cgroup_get_from_id(ops->sub_cgroup_id);
+	if (IS_ERR(cgrp)) {
+		ret = PTR_ERR(cgrp);
+		goto out_unlock;
+	}
+
+	raw_spin_lock_irq(&scx_sched_lock);
+	parent = find_parent_sched(cgrp);
+	if (IS_ERR(parent)) {
+		raw_spin_unlock_irq(&scx_sched_lock);
+		ret = PTR_ERR(parent);
+		goto out_put_cgrp;
+	}
+	kobject_get(&parent->kobj);
+	raw_spin_unlock_irq(&scx_sched_lock);
+
+	/*
+	 * Flip the hot-path gates before ops->priv is published - the sub's
+	 * programs can e.g. kick cpus from that point on. The matching dec is
+	 * at the end of scx_sched_free_rcu_work().
+	 */
+	static_branch_inc(&__scx_has_subs);
+
+	/* scx_alloc_and_add_sched() consumes @cgrp whether it succeeds or not */
+	sch = scx_alloc_and_add_sched(cmd, cgrp, parent);
+	kobject_put(&parent->kobj);
+	if (IS_ERR(sch)) {
+		static_branch_dec(&__scx_has_subs);
+		ret = PTR_ERR(sch);
+		goto out_unlock;
+	}
+
+	/*
+	 * Validate before scx_link_sched() publishes @sch, so an invalid sub
+	 * never becomes visible with an unallocated pshard.
+	 */
+	ret = scx_validate_ops(sch, ops);
+	if (ret)
+		goto err_disable;
+
+	scx_rescue_check_timeout(sch);
+
+	/*
+	 * Allocate pshard[] before scx_link_sched() publishes @sch into the
+	 * parent's RCU children list. A concurrent revoke walking the tree
+	 * would otherwise dereference sch->pshard[si] while it's still NULL.
+	 * Unlike the root path, the cid shard layout is stable at this point.
+	 *
+	 * scx_alloc_pshards() skips allocation when @sch's arena pool isn't
+	 * initialized, so scx_arena_pool_init() must run first.
+	 */
+	ret = scx_arena_pool_init(sch);
+	if (ret)
+		goto err_disable;
+
+	ret = scx_alloc_pshards(sch);
+	if (ret)
+		goto err_disable;
+
+	ret = scx_link_sched(sch);
+	if (ret)
+		goto err_disable;
+
+	ret = scx_sched_sysfs_add(sch);
+	if (ret)
+		goto err_disable;
+
+	if (sch->level >= SCX_SUB_MAX_DEPTH) {
+		scx_error(sch, "max nesting depth %d violated",
+			  SCX_SUB_MAX_DEPTH);
+		ret = -EINVAL;
+		goto err_disable;
+	}
+
+	if (sch->ops.init) {
+		ret = SCX_CALL_OP_RET(sch, init, NULL);
+		if (ret) {
+			ret = scx_ops_sanitize_err(sch, "init", ret);
+			scx_error(sch, "ops.init() failed (%d)", ret);
+			goto err_disable;
+		}
+		sch->exit_info->flags |= SCX_EFLAG_INITIALIZED;
+	}
+
+	ret = scx_set_cmask_scratch_alloc(sch);
+	if (ret)
+		goto err_disable;
+
+	struct scx_sub_attach_args sub_attach_args = {
+		.ops = &sch->ops,
+		.cgroup_path = sch->cgrp_path,
+	};
+
+	ret = SCX_CALL_OP_RET(parent, sub_attach, NULL,
+			      &sub_attach_args);
+	if (ret) {
+		ret = scx_ops_sanitize_err(sch, "sub_attach", ret);
+		scx_error(sch, "parent rejected (%d)", ret);
+		goto err_disable;
+	}
+	sch->sub_attached = true;
+
+	scx_bypass(sch, true);
+
+	for (i = SCX_OPI_BEGIN; i < SCX_OPI_END; i++)
+		if (((void (**)(void))ops)[i])
+			set_bit(i, sch->has_op);
+
+	percpu_down_write(&scx_fork_rwsem);
+	scx_cgroup_lock();
+
+	/*
+	 * Set cgroup->scx_sched's and check CSS_ONLINE. Either we see
+	 * !CSS_ONLINE or scx_cgroup_lifetime_notify() sees and shoots us down.
+	 */
+	set_cgroup_sched(sch_cgroup(sch), sch);
+	if (!(cgrp->self.flags & CSS_ONLINE)) {
+		scx_error(sch, "cgroup is not online");
+		ret = -ENODEV;
+		goto err_unlock_and_disable;
+	}
+
+	/*
+	 * Take over the subtree's cgroups before any task is claimed,
+	 * mirroring root enable's cgroups-before-tasks order.
+	 */
+	ret = scx_cgroup_claim_subtree(sch);
+	if (ret)
+		goto err_unlock_and_disable;
+
+	/*
+	 * Initialize tasks for the new child $sch without exiting them for
+	 * $parent so that the tasks can always be reverted back to $parent
+	 * sched on child init failure.
+	 */
+	WARN_ON_ONCE(scx_enabling_sub_sched);
+	scx_enabling_sub_sched = sch;
+
+	scx_task_iter_start(&sti, sch->cgrp);
+	while ((p = scx_task_iter_next_locked(&sti))) {
+		struct rq *rq;
+		struct rq_flags rf;
+
+		/*
+		 * Task iteration may visit the same task twice when racing
+		 * against exiting. Use %SCX_TASK_SUB_INIT to mark tasks which
+		 * finished __scx_init_task() and skip if set.
+		 *
+		 * A task may exit and get freed between __scx_init_task()
+		 * completion and scx_enable_task(). In such cases,
+		 * scx_disable_and_exit_task() must exit the task for both the
+		 * parent and child scheds.
+		 */
+		if (p->scx.flags & SCX_TASK_SUB_INIT)
+			continue;
+
+		/* @p is pinned by the iter; see scx_sub_disable() */
+		get_task_struct(p);
+
+		if (!assert_task_ready_or_enabled(p)) {
+			ret = -EINVAL;
+			goto abort;
+		}
+
+		scx_task_iter_unlock(&sti);
+
+		/*
+		 * As $p is still on $parent, it can't be transitioned to INIT.
+		 * Let's worry about task state later. Use __scx_init_task().
+		 */
+		ret = __scx_init_task(sch, p, NULL, false);
+		if (ret)
+			goto abort;
+
+		rq = task_rq_lock(p, &rf);
+
+		if (scx_get_task_state(p) == SCX_TASK_DEAD) {
+			/*
+			 * sched_ext_dead() raced us between __scx_init_task()
+			 * and this rq lock and ran exit_task() on $parent (the
+			 * sched @p was on at that point), not on @sch. @sch's
+			 * just-completed init is owed an exit_task() and we
+			 * issue it here.
+			 */
+			scx_sub_init_cancel_task(sch, p);
+			task_rq_unlock(rq, p, &rf);
+			put_task_struct(p);
+			continue;
+		}
+
+		p->scx.flags |= SCX_TASK_SUB_INIT;
+		task_rq_unlock(rq, p, &rf);
+
+		put_task_struct(p);
+	}
+	scx_task_iter_stop(&sti);
+
+	/*
+	 * All tasks are prepped. Disable/exit tasks for $parent and enable for
+	 * the new @sch.
+	 */
+	scx_task_iter_start(&sti, sch->cgrp);
+	while ((p = scx_task_iter_next_locked(&sti))) {
+		/*
+		 * Use clearing of %SCX_TASK_SUB_INIT to detect and skip
+		 * duplicate iterations.
+		 */
+		if (!(p->scx.flags & SCX_TASK_SUB_INIT))
+			continue;
+
+		scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
+			/*
+			 * $p must be either READY or ENABLED. If ENABLED,
+			 * __scx_disabled_and_exit_task() first disables and
+			 * makes it READY. However, after exiting $p, it will
+			 * leave $p as READY.
+			 */
+			assert_task_ready_or_enabled(p);
+			__scx_disable_and_exit_task(parent, p);
+
+			/*
+			 * $p is now only initialized for @sch and READY, which
+			 * is what we want. Assign it to @sch and, if it's on
+			 * the ext class, enable. A non-ext task, possible under
+			 * an %SCX_OPS_SWITCH_PARTIAL root, stays READY and is
+			 * enabled by switching_to_scx() if it switches over.
+			 */
+			scx_set_task_sched(p, sch);
+			if (p->sched_class == &ext_sched_class)
+				scx_enable_task(sch, p);
+
+			p->scx.flags &= ~SCX_TASK_SUB_INIT;
+		}
+	}
+	scx_task_iter_stop(&sti);
+
+	scx_enabling_sub_sched = NULL;
+
+	scx_cgroup_unlock();
+	percpu_up_write(&scx_fork_rwsem);
+
+	scx_bypass(sch, false);
+
+	/* @sch is enabled; deliver any caps owed since its sub_attach() */
+	scx_sub_seed_caps(sch);
+
+	pr_info("sched_ext: BPF sub-scheduler \"%s\" enabled\n", sch->ops.name);
+	kobject_uevent(&sch->kobj, KOBJ_ADD);
+	ret = 0;
+	goto out_unlock;
+
+out_put_cgrp:
+	cgroup_put(cgrp);
+out_unlock:
+	mutex_unlock(&scx_enable_mutex);
+	cmd->ret = ret;
+	return;
+
+abort:
+	put_task_struct(p);
+	scx_task_iter_stop(&sti);
+
+	/*
+	 * Undo __scx_init_task() for tasks we marked. scx_enable_task() never
+	 * ran for @sch on them, so calling scx_disable_task() here would invoke
+	 * ops.disable() without a matching ops.enable(). scx_enabling_sub_sched
+	 * must stay set until SUB_INIT is cleared from every marked task -
+	 * scx_disable_and_exit_task() reads it when a task exits concurrently.
+	 */
+	scx_task_iter_start(&sti, sch->cgrp);
+	while ((p = scx_task_iter_next_locked(&sti))) {
+		if (p->scx.flags & SCX_TASK_SUB_INIT) {
+			scx_sub_init_cancel_task(sch, p);
+			p->scx.flags &= ~SCX_TASK_SUB_INIT;
+		}
+	}
+	scx_task_iter_stop(&sti);
+	scx_enabling_sub_sched = NULL;
+err_unlock_and_disable:
+	/* we'll soon enter disable path, keep bypass on */
+	scx_cgroup_unlock();
+	percpu_up_write(&scx_fork_rwsem);
+err_disable:
+	mutex_unlock(&scx_enable_mutex);
+	/*
+	 * Some enable failures only return an errno (e.g. -ENOMEM from an
+	 * allocation) without calling scx_error(). Record it so
+	 * scx_flush_disable_work() runs the disable and ops.exit() fires.
+	 */
+	scx_error(sch, "scx_sub_enable() failed (%d)", ret);
+	scx_flush_disable_work(sch);
+	cmd->ret = 0;
+}
+
+/**
+ * scx_cgroup_task_migrating - Prepare a task for a cgroup migration
+ * @ctx: migration being prepared
+ *
+ * A task's sched must match its cgroup's owner, so a migration that crosses a
+ * sched boundary re-homes the task once committed. Run the fallible part here,
+ * before the migration commits: initialize the task for the destination sched.
+ * A rejection fails the cgroup.procs write.
+ */
+static s32 scx_cgroup_task_migrating(struct cgroup_task_migrate_ctx *ctx)
+{
+	struct task_struct *p = ctx->task;
+	struct scx_sched *to;
+	int ret;
+
+	/*
+	 * Cleared under scx_cgroup_lock() before root disable starts tearing
+	 * down tasks. As cgroup_mutex is held, a set flag guarantees that the
+	 * teardown loop is not running concurrently.
+	 */
+	if (!scx_cgroup_enabled)
+		return NOTIFY_OK;
+
+	to = scx_cgroup_sched(ctx->dst_dcgrp);
+	if (scx_task_on_sched(to, p))
+		return NOTIFY_OK;
+
+	ret = __scx_init_task(to, p, ctx->dst_dcgrp, false);
+	if (ret)
+		return notifier_from_errno(ret);
+
+	return NOTIFY_OK;
+}
+
+/**
+ * scx_cgroup_task_migrated - Re-home a task that changed cgroups
+ * @ctx: committed migration
+ *
+ * Move the task to its new cgroup's sched, which scx_cgroup_task_migrating()
+ * already initialized it for. Can't fail.
+ *
+ * This is safe against all phases of the destination sched's destruction. A
+ * disable resets cgroup ownership to the parent and re-homes tasks in one
+ * scx_cgroup_lock() section. If that section already ran, the destination would
+ * be the parent. Otherwise, the re-home loop is still ahead and guaranteed to
+ * visit the task, now in the destination cgroup.
+ */
+static void scx_cgroup_task_migrated(struct cgroup_task_migrate_ctx *ctx)
+{
+	struct task_struct *p = ctx->task;
+	struct scx_sched *to;
+	struct rq *rq;
+	struct rq_flags rf;
+
+	if (!scx_cgroup_enabled)
+		return;
+
+	to = scx_cgroup_sched(ctx->dst_dcgrp);
+	if (scx_task_on_sched(to, p))
+		return;
+
+	rq = task_rq_lock(p, &rf);
+	scx_rehome_task(to, p);
+	task_rq_unlock(rq, p, &rf);
+}
+
+/**
+ * scx_cgroup_task_migrate_canceled - Undo migration preparation
+ * @ctx: canceled migration
+ *
+ * The migration failed after scx_cgroup_task_migrating() initialized the task
+ * for the destination sched. The task stays on its current sched in the source
+ * cgroup. Undo the destination's init.
+ */
+static void scx_cgroup_task_migrate_canceled(struct cgroup_task_migrate_ctx *ctx)
+{
+	struct task_struct *p = ctx->task;
+	struct scx_sched *to;
+	struct rq *rq;
+	struct rq_flags rf;
+
+	if (!scx_cgroup_enabled)
+		return;
+
+	to = scx_cgroup_sched(ctx->dst_dcgrp);
+	if (scx_task_on_sched(to, p))
+		return;
+
+	rq = task_rq_lock(p, &rf);
+	scx_sub_init_cancel_task(to, p);
+	task_rq_unlock(rq, p, &rf);
+}
+
+static s32 scx_cgroup_lifetime_notify(struct notifier_block *nb,
+				      unsigned long action, void *data)
+{
+	struct cgroup *cgrp = data;
+	struct cgroup *parent = cgroup_parent(cgrp);
+	struct scx_sched *sch;
+
+	if (!cgroup_on_dfl(cgrp))
+		return NOTIFY_OK;
+
+	switch (action) {
+	case CGROUP_LIFETIME_ONLINE:
+		/* inherit ->scx_sched from $parent */
+		if (parent)
+			rcu_assign_pointer(cgrp->scx_sched, scx_cgroup_sched(parent));
+		break;
+	case CGROUP_LIFETIME_OFFLINE:
+		/* if there is a sched attached, shoot it down */
+		sch = scx_cgroup_sched(cgrp);
+		if (sch && sch->cgrp == cgrp)
+			scx_exit(sch, SCX_EXIT_UNREG_KERN,
+				 SCX_ECODE_RSN_CGROUP_OFFLINE,
+				 "cgroup %llu going offline", cgroup_id(cgrp));
+		break;
+	}
+
+	return NOTIFY_OK;
+}
+
+static struct notifier_block scx_cgroup_lifetime_nb = {
+	.notifier_call = scx_cgroup_lifetime_notify,
+};
+
+static s32 scx_cgroup_task_notify(struct notifier_block *nb,
+				  unsigned long action, void *data)
+{
+	struct cgroup_task_migrate_ctx *ctx = data;
+
+	switch (action) {
+	case CGROUP_TASK_MIGRATING:
+		return scx_cgroup_task_migrating(ctx);
+	case CGROUP_TASK_MIGRATED:
+		scx_cgroup_task_migrated(ctx);
+		break;
+	case CGROUP_TASK_MIGRATE_CANCELED:
+		scx_cgroup_task_migrate_canceled(ctx);
+		break;
+	}
+
+	return NOTIFY_OK;
+}
+
+static struct notifier_block scx_cgroup_task_nb = {
+	.notifier_call = scx_cgroup_task_notify,
+};
+
+static s32 __init scx_cgroup_notifier_init(void)
+{
+	s32 ret;
+
+	ret = blocking_notifier_chain_register(&cgroup_lifetime_notifier,
+					       &scx_cgroup_lifetime_nb);
+	if (ret)
+		return ret;
+
+	return blocking_notifier_chain_register(&cgroup_task_notifier,
+						&scx_cgroup_task_nb);
+}
+core_initcall(scx_cgroup_notifier_init);
+
+static void scx_pstack_recursion(struct bpf_prog *prog, const char *op)
+{
+	struct scx_sched *sch;
+
+	guard(rcu)();
+	sch = scx_prog_sched(prog->aux);
+	if (unlikely(!sch))
+		return;
+
+	scx_error(sch, "%s recursion detected", op);
+}
+
+void scx_pstack_recursion_on_dispatch(struct bpf_prog *prog)
+{
+	scx_pstack_recursion(prog, "dispatch");
+}
+
+void scx_pstack_recursion_on_caps_updated(struct bpf_prog *prog)
+{
+	scx_pstack_recursion(prog, "sub_caps_updated");
+}
+
+__bpf_kfunc_start_defs();
+
+/**
+ * scx_bpf_sub_dispatch - Trigger dispatching on a child scheduler
+ * @cgroup_id: cgroup ID of the child scheduler to dispatch
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Allows a parent scheduler to trigger dispatching on one of its direct
+ * child schedulers. The child scheduler runs its dispatch operation to
+ * move tasks from dispatch queues to the local runqueue.
+ *
+ * Returns: true on success, false if cgroup_id is invalid, not a direct
+ * child, or caller lacks dispatch permission.
+ */
+__bpf_kfunc bool scx_bpf_sub_dispatch(u64 cgroup_id, const struct bpf_prog_aux *aux)
+{
+	struct rq *rq = scx_locked_rq();
+	struct scx_sched *parent, *child;
+
+	guard(rcu)();
+	parent = scx_prog_sched(aux);
+	if (unlikely(!parent))
+		return false;
+
+	child = scx_find_sub_sched(cgroup_id);
+
+	if (unlikely(!child))
+		return false;
+
+	if (unlikely(scx_parent(child) != parent)) {
+		scx_error(parent, "trying to dispatch a distant sub-sched on cgroup %llu",
+			  cgroup_id);
+		return false;
+	}
+
+	/*
+	 * Skip a child that does not effectively hold the base cap on this cpu:
+	 * its inserts would only be rejected. ecaps are synced at the top of
+	 * dispatch_one() before dispatch, so this reflects the in-effect state.
+	 */
+	if (scx_missing_caps(child, cpu_of(rq), SCX_CAP_BASE))
+		return false;
+
+	return scx_dispatch_sched(child, rq, rq->scx.sub_dispatch_prev, true) !=
+		SCX_DSP_NONE;
+}
+
+/* Validate common inputs. On success, *parent_out and *child_out are set. */
+static s32 sub_cap_preamble(u64 cgroup_id, u64 caps, const struct bpf_prog_aux *aux,
+			    struct scx_sched **parent_out, struct scx_sched **child_out)
+{
+	struct scx_sched *parent, *child;
+
+	parent = scx_prog_sched(aux);
+	if (unlikely(!parent))
+		return -ENODEV;
+
+	if (!scx_is_cid_type()) {
+		scx_error(parent, "sub-cap kfuncs require a cid-form scheduler");
+		return -EOPNOTSUPP;
+	}
+
+	child = scx_find_sub_sched(cgroup_id);
+	if (unlikely(!child))
+		return -ENODEV;
+
+	if (unlikely(scx_parent(child) != parent)) {
+		scx_error(parent, "%s: sub-%llu is not a direct child",
+			  parent->cgrp_path, cgroup_id);
+		return -EINVAL;
+	}
+
+	if (unlikely(caps & ~__SCX_CAP_ALL)) {
+		scx_error(parent, "invalid caps 0x%llx", caps);
+		return -EINVAL;
+	}
+
+	*parent_out = parent;
+	*child_out = child;
+	return 0;
+}
+
+/**
+ * scx_bpf_sub_grant - Grant @caps on a cmask's cids to a direct child
+ * @cgroup_id: cgroup id of the direct child sub-sched
+ * @caps: bitmask of SCX_CAP_* to grant
+ * @cmask__arena: cid cmask to grant @caps on
+ * @denied_out__arena__nullable: optional cmask accumulating refused cids
+ * @aux: implicit BPF argument
+ *
+ * A cid in @cmask__arena is granted to the child only if the parent holds every
+ * requested cap on it. Refused cids are OR'd into the denied mask when
+ * provided. Refusals outside the denied mask's range are not recorded.
+ *
+ * All-or-nothing keeps the caller-visible result binary per cid, so the denied
+ * mask is one mask to interpret rather than a per-cap matrix.
+ *
+ * Return 0 on full success, -EPERM if any cid was refused, or a negative
+ * errno on other failures.
+ */
+__bpf_kfunc s32 scx_bpf_sub_grant(u64 cgroup_id, u64 caps,
+				  const struct scx_cmask *cmask__arena,
+				  struct scx_cmask *denied_out__arena__nullable,
+				  const struct bpf_prog_aux *aux)
+{
+	struct scx_cmask_ref ref, denied_ref;
+	struct scx_sched *parent, *child;
+	bool any_denied = false;
+	LIST_HEAD(to_deliver);
+	s32 si, ret;
+
+	guard(irqsave)();
+
+	ret = sub_cap_preamble(cgroup_id, caps, aux, &parent, &child);
+	if (ret)
+		return ret;
+
+	ret = scx_cmask_ref_init(parent, cmask__arena, &ref);
+	if (ret) {
+		scx_error(parent, "invalid cmask (%d)", ret);
+		return ret;
+	}
+
+	if (denied_out__arena__nullable) {
+		ret = scx_cmask_ref_init(parent, denied_out__arena__nullable, &denied_ref);
+		if (ret) {
+			scx_error(parent, "invalid denied_out (%d)", ret);
+			return ret;
+		}
+	}
+
+	/* apply the grant one shard at a time */
+	for (si = ref.shard_first; si < ref.shard_end; si++) {
+		SCX_CMASK_DEFINE_SHARD(slice, 0, SCX_CID_SHARD_MAX_CPUS);
+		struct scx_pshard *pps = parent->pshard[si];
+		struct scx_pshard *cps = child->pshard[si];
+		u64 granted_caps = 0;
+		u32 cap_bit;
+
+		scx_cmask_ref_shard(&ref, si, slice);
+		if (scx_cmask_empty(slice))
+			continue;
+
+		SCX_CMASK_DEFINE_SHARD(granted_cids, slice->base, slice->nr_cids);
+		SCX_CMASK_DEFINE_SHARD(changed_cids, slice->base, slice->nr_cids);
+		SCX_CMASK_DEFINE_SHARD(delta, slice->base, slice->nr_cids);
+
+		scx_cmask_copy(granted_cids, slice);
+
+		scoped_guard (raw_spinlock, &pps->lock) {
+			guard(raw_spinlock_nested)(&cps->lock);
+
+			/*
+			 * Narrow granted_cids to cids the parent holds every
+			 * requested cap on. All-or-nothing per cid.
+			 */
+			scx_for_each_cap_bit(cap_bit, caps)
+				scx_cmask_and(granted_cids, &pps->caps[cap_bit].cmask);
+
+			/*
+			 * For each requested cap, fold the newly-set cids into
+			 * the child and accumulate the delta.
+			 */
+			scx_for_each_cap_bit(cap_bit, caps) {
+				struct scx_cmask *ccm = &cps->caps[cap_bit].cmask;
+
+				scx_cmask_copy(delta, granted_cids);
+				scx_cmask_andnot(delta, ccm);
+				if (scx_cmask_empty(delta))
+					continue;
+
+				scx_cmask_or(ccm, delta);
+				scx_cmask_or(changed_cids, delta);
+				granted_caps |= BIT_U64(cap_bit);
+			}
+
+			if (granted_caps) {
+				s32 cid;
+
+				caps_updated_record(cps, changed_cids, granted_caps,
+						    &to_deliver);
+				/*
+				 * The sync arms an update_idle() re-notify if
+				 * the cid gains baseline access, so the holder
+				 * learns of an already-idle cid.
+				 */
+				scx_cmask_for_each_cid(cid, changed_cids)
+					queue_sync_ecaps(child, cid);
+			}
+		}
+
+		/* record cids that didn't make it into the denied mask */
+		if (!scx_cmask_subset(slice, granted_cids)) {
+			any_denied = true;
+			if (denied_out__arena__nullable) {
+				SCX_CMASK_DEFINE_SHARD(denied, slice->base, slice->nr_cids);
+
+				scx_cmask_copy(denied, slice);
+				scx_cmask_andnot(denied, granted_cids);
+				scx_cmask_ref_or(&denied_ref, denied);
+			}
+		}
+	}
+
+	caps_updated_deliver(&to_deliver);
+
+	return any_denied ? -EPERM : 0;
+}
+
+/**
+ * scx_bpf_sub_revoke - Revoke @caps on a cmask's cids from a direct child
+ * @cgroup_id: cgroup id of the direct child sub-sched
+ * @caps: bitmask of SCX_CAP_* to revoke
+ * @cmask__arena: cid cmask to revoke @caps on
+ * @aux: implicit BPF argument
+ *
+ * Clear @caps bits on @cmask__arena from the child named by @cgroup_id and all
+ * its descendants. The origin parent's pshard lock is held across the subtree
+ * walk so a concurrent grant from the origin parent observes the revoked state.
+ */
+__bpf_kfunc void scx_bpf_sub_revoke(u64 cgroup_id, u64 caps,
+				    const struct scx_cmask *cmask__arena,
+				    const struct bpf_prog_aux *aux)
+{
+	struct scx_cmask_ref ref;
+	struct scx_sched *parent, *child, *pos;
+	LIST_HEAD(to_deliver);
+	s32 si, ret;
+
+	guard(irqsave)();
+
+	if (sub_cap_preamble(cgroup_id, caps, aux, &parent, &child))
+		return;
+
+	ret = scx_cmask_ref_init(parent, cmask__arena, &ref);
+	if (ret) {
+		scx_error(parent, "invalid cmask (%d)", ret);
+		return;
+	}
+
+	/* per-shard, walk child's subtree and clear @caps */
+	for (si = ref.shard_first; si < ref.shard_end; si++) {
+		SCX_CMASK_DEFINE_SHARD(slice, 0, SCX_CID_SHARD_MAX_CPUS);
+
+		scx_cmask_ref_shard(&ref, si, slice);
+		if (scx_cmask_empty(slice))
+			continue;
+
+		/*
+		 * Pre-order with subtree skip: a descendant that cleared
+		 * nothing means no descendant of it can hold @caps on these
+		 * cids either.
+		 */
+		guard(raw_spinlock)(&parent->pshard[si]->lock);
+		pos = scx_next_descendant_pre(NULL, child);
+		while (pos) {
+			struct scx_pshard *ps = pos->pshard[si];
+			SCX_CMASK_DEFINE_SHARD(changed_cids, slice->base, slice->nr_cids);
+			SCX_CMASK_DEFINE_SHARD(delta, slice->base, slice->nr_cids);
+			u64 revoked_caps = 0;
+			u32 cap_bit;
+
+			scoped_guard (raw_spinlock_nested, &ps->lock) {
+				/*
+				 * For each cap, clear lost cids and accumulate
+				 * the per-cap diff for notification.
+				 */
+				scx_for_each_cap_bit(cap_bit, caps) {
+					struct scx_cmask *cm = &ps->caps[cap_bit].cmask;
+
+					scx_cmask_copy(delta, cm);
+					scx_cmask_and(delta, slice);
+					if (scx_cmask_empty(delta))
+						continue;
+
+					scx_cmask_andnot(cm, delta);
+					scx_cmask_or(changed_cids, delta);
+					revoked_caps |= BIT_U64(cap_bit);
+				}
+
+				if (revoked_caps) {
+					s32 cid;
+
+					caps_updated_record(ps, changed_cids, revoked_caps,
+							    &to_deliver);
+					scx_cmask_for_each_cid(cid, changed_cids)
+						queue_sync_ecaps(pos, cid);
+				}
+			}
+
+			if (revoked_caps)
+				pos = scx_next_descendant_pre(pos, child);
+			else
+				pos = scx_skip_subtree_pre(pos, child);
+		}
+	}
+
+	caps_updated_deliver(&to_deliver);
+}
+
+/**
+ * scx_bpf_sub_caps - Read self's or a direct child's cap cmasks
+ * @cgroup_id: 0 for self, or a direct child's cgroup id
+ * @caps: one or more SCX_CAP_* bits
+ * @out__arena: cmask to receive the union of @caps within its range
+ * @aux: implicit BPF argument
+ *
+ * Read the cap cmasks granted on each cid for self (@cgroup_id 0) or a direct
+ * child - the literal granted set. A sched can read only itself or a direct
+ * child.
+ *
+ * Return 0, -ENODEV if @cgroup_id names no direct child, or -EINVAL on bad
+ * inputs.
+ */
+__bpf_kfunc s32 scx_bpf_sub_caps(u64 cgroup_id, u64 caps, struct scx_cmask *out__arena,
+				 const struct bpf_prog_aux *aux)
+{
+	struct scx_cmask_ref ref;
+	struct scx_sched *sch, *target;
+	struct scx_pshard **pshard;
+	s32 si, ret;
+
+	guard(irqsave)();
+
+	sch = scx_prog_sched(aux);
+	if (unlikely(!sch))
+		return -ENODEV;
+
+	if (!scx_is_cid_type()) {
+		scx_error(sch, "sub-cap kfuncs require a cid-form scheduler");
+		return -EOPNOTSUPP;
+	}
+
+	if (unlikely(caps & ~__SCX_CAP_ALL)) {
+		scx_error(sch, "invalid caps 0x%llx", caps);
+		return -EINVAL;
+	}
+
+	/* @cgroup_id 0 reads self, otherwise a direct child */
+	if (cgroup_id) {
+		target = scx_find_sub_sched(cgroup_id);
+		if (unlikely(!target))
+			return -ENODEV;
+		if (unlikely(scx_parent(target) != sch)) {
+			scx_error(sch, "%s: sub-%llu is not a direct child",
+				  sch->cgrp_path, cgroup_id);
+			return -EINVAL;
+		}
+	} else {
+		target = sch;
+	}
+
+	/*
+	 * The target's caps storage may not be set up yet (e.g. a self-read
+	 * during ops.init_cids()). Pairs with the publish in
+	 * scx_alloc_pshards(): a non-NULL pshard has every element set and the
+	 * acquire also orders the cid table reads below against it.
+	 */
+	pshard = smp_load_acquire(&target->pshard);
+	if (unlikely(!pshard)) {
+		scx_error(sch, "scx_bpf_sub_caps() called before caps storage is initialized");
+		return -ENODEV;
+	}
+
+	ret = scx_cmask_ref_init(sch, out__arena, &ref);
+	if (ret) {
+		scx_error(sch, "invalid out (%d)", ret);
+		return ret;
+	}
+
+	for (si = ref.shard_first; si < ref.shard_end; si++) {
+		const struct scx_cid_shard *shard =
+			&rcu_dereference_all(scx_cid_shard_ranges)[si];
+		SCX_CMASK_DEFINE_SHARD(local_out, shard->base_cid, shard->nr_cids);
+		u32 cap_bit;
+
+		scx_for_each_cap_bit(cap_bit, caps)
+			scx_cmask_or(local_out, &pshard[si]->caps[cap_bit].cmask);
+		scx_cmask_ref_copy(&ref, local_out);
+	}
+	return 0;
+}
+
+/**
+ * scx_bpf_sub_kill_bstr - Kill a direct child sub-scheduler
+ * @cgroup_id: cgroup id of the direct child to kill
+ * @fmt: reason message format string
+ * @data: format string parameters packaged using ___bpf_fill() macro
+ * @data__sz: @data len, must end in '__sz' for the verifier
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Evict a direct child sub-scheduler, disabling it with the supplied reason.
+ * The child and its subtree are torn down asynchronously through the usual
+ * disable path.
+ *
+ * Unlike scx_bpf_exit(), no exit code is taken: the child is a separate
+ * scheduler with its own exit-code semantics, so a code chosen by the parent
+ * would have no defined meaning. The reason string carries the intent.
+ *
+ * Return 0 on success or -ENODEV if @cgroup_id names no sub-scheduler, which
+ * can race with the child detaching on its own and so is not a scheduler error.
+ * Naming a sched that exists but is not a direct child aborts the parent.
+ */
+__printf(2, 0)
+__bpf_kfunc s32 scx_bpf_sub_kill_bstr(u64 cgroup_id, char *fmt,
+				      unsigned long long *data, u32 data__sz,
+				      const struct bpf_prog_aux *aux)
+{
+	struct scx_sched *parent, *child;
+
+	guard(rcu)();
+
+	parent = scx_prog_sched(aux);
+	if (unlikely(!parent))
+		return -ENODEV;
+
+	if (!scx_is_cid_type()) {
+		scx_error(parent, "sub-cap kfuncs require a cid-form scheduler");
+		return -EOPNOTSUPP;
+	}
+
+	child = scx_find_sub_sched(cgroup_id);
+	if (unlikely(!child))
+		return -ENODEV;
+
+	if (unlikely(scx_parent(child) != parent)) {
+		scx_error(parent, "%s: sub-%llu is not a direct child",
+			  parent->cgrp_path, cgroup_id);
+		return -EINVAL;
+	}
+
+	scx_exit_bstr(child, SCX_EXIT_PARENT_KILL, 0, parent, fmt, data, data__sz);
+	return 0;
+}
+
+__bpf_kfunc_end_defs();
+
+#else	/* !CONFIG_EXT_SUB_SCHED */
+
+__bpf_kfunc_start_defs();
+
+__bpf_kfunc s32 scx_bpf_sub_grant(u64 cgroup_id, u64 caps,
+				  const struct scx_cmask *cmask__arena,
+				  struct scx_cmask *denied_out__arena__nullable,
+				  const struct bpf_prog_aux *aux)
+{
+	return -EOPNOTSUPP;
+}
+
+__bpf_kfunc void scx_bpf_sub_revoke(u64 cgroup_id, u64 caps,
+				    const struct scx_cmask *cmask__arena,
+				    const struct bpf_prog_aux *aux)
+{
+}
+
+__bpf_kfunc s32 scx_bpf_sub_caps(u64 cgroup_id, u64 caps, struct scx_cmask *out__arena,
+				 const struct bpf_prog_aux *aux)
+{
+	return -EOPNOTSUPP;
+}
+
+__bpf_kfunc s32 scx_bpf_sub_kill_bstr(u64 cgroup_id, char *fmt,
+				      unsigned long long *data, u32 data__sz,
+				      const struct bpf_prog_aux *aux)
+{
+	return -EOPNOTSUPP;
+}
+
+__bpf_kfunc_end_defs();
+
+#endif	/* CONFIG_EXT_SUB_SCHED */
diff --git a/kernel/sched/ext/sub.h b/kernel/sched/ext/sub.h
new file mode 100644
index 000000000000..f7bcdfda8dd8
--- /dev/null
+++ b/kernel/sched/ext/sub.h
@@ -0,0 +1,232 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
+ *
+ * Sub-scheduler hierarchy support.
+ *
+ * Copyright (c) 2026 Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2026 Tejun Heo <tj@kernel.org>
+ */
+#ifndef _KERNEL_SCHED_EXT_SUB_H
+#define _KERNEL_SCHED_EXT_SUB_H
+
+#include "internal.h"
+
+#ifdef CONFIG_EXT_SUB_SCHED
+
+struct scx_sched *scx_skip_subtree_pre(struct scx_sched *pos, struct scx_sched *root);
+struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root);
+void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch);
+struct cgroup *sch_cgroup(struct scx_sched *sch);
+void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch);
+void scx_pstack_recursion_on_dispatch(struct bpf_prog *prog);
+void scx_pstack_recursion_on_caps_updated(struct bpf_prog *prog);
+void drain_descendants(struct scx_sched *sch);
+void scx_sub_disable(struct scx_sched *sch);
+void scx_sub_enable_workfn(struct kthread_work *work);
+bool scx_bpf_sub_dispatch(u64 cgroup_id, const struct bpf_prog_aux *aux);
+void scx_free_pshards(struct scx_sched *sch);
+s32 scx_alloc_pshards(struct scx_sched *sch);
+void scx_init_root_caps(struct scx_sched *sch);
+void scx_process_sync_ecaps(struct rq *rq, struct task_struct *prev);
+void scx_unbypass_replay_ecaps(struct rq *rq, struct scx_sched *sch);
+void scx_online_ecaps(struct rq *rq);
+void scx_offline_ecaps(struct rq *rq);
+void scx_discard_ecaps_to_sync(s32 cpu, struct scx_sched_pcpu *pcpu);
+void scx_discard_stale_ecaps_syncs(void);
+struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *rq,
+					     struct task_struct *p, u64 *enq_flags);
+bool scx_task_reenq_on_cap_revoke(struct rq *rq, struct task_struct *p);
+void scx_reenq_reject(struct rq *rq);
+void scx_rescue_charge(struct rq *rq, s64 delta_exec);
+void scx_rescue_end(struct rq *rq);
+bool scx_rescue_keep(struct rq *rq, struct task_struct *p);
+void scx_rescue_flush(struct rq *rq);
+void scx_rescue_dump(struct seq_buf *s, struct rq *rq);
+void scx_rescue_set_knobs(struct scx_sched *sch);
+void scx_rescue_init(struct rq *rq);
+
+/*
+ * cgrp->scx_sched is written by root/sub enable/disable under all of
+ * scx_enable_mutex, scx_fork_rwsem and cgroup_mutex. A new cgroup inherits the
+ * parent's sched under just cgroup_mutex but is not yet reachable by the other
+ * two lock holders. Any one of the three locks stabilizes the association.
+ */
+static inline struct scx_sched *scx_cgroup_sched(struct cgroup *cgrp)
+{
+	return rcu_dereference_check(cgrp->scx_sched,
+				     lockdep_is_held(&cgroup_mutex) ||
+				     percpu_rwsem_is_held(&scx_fork_rwsem) ||
+				     lockdep_is_held(&scx_enable_mutex));
+}
+
+static inline const char *sch_cgrp_path(struct scx_sched *sch)
+{
+	return sch->cgrp_path;
+}
+
+/* a dying sub's hot-path influence ends in scx_sched_free_rcu_work() */
+static inline void scx_dec_has_subs(struct scx_sched *sch)
+{
+	if (sch->level)
+		static_branch_dec(&__scx_has_subs);
+}
+
+#else	/* CONFIG_EXT_SUB_SCHED */
+
+static inline struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root) { return pos ? NULL : root; }
+static inline struct scx_sched *scx_skip_subtree_pre(struct scx_sched *pos, struct scx_sched *root) { return NULL; }
+static inline void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch) {}
+static inline struct cgroup *sch_cgroup(struct scx_sched *sch) { return NULL; }
+static inline const char *sch_cgrp_path(struct scx_sched *sch) { return "/"; }
+static inline void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch) {}
+static inline void drain_descendants(struct scx_sched *sch) { }
+static inline void scx_sub_disable(struct scx_sched *sch) { }
+static inline void scx_free_pshards(struct scx_sched *sch) {}
+static inline s32 scx_alloc_pshards(struct scx_sched *sch) { return 0; }
+static inline void scx_init_root_caps(struct scx_sched *sch) {}
+static inline void scx_process_sync_ecaps(struct rq *rq, struct task_struct *prev) {}
+static inline void scx_unbypass_replay_ecaps(struct rq *rq, struct scx_sched *sch) {}
+static inline void scx_online_ecaps(struct rq *rq) {}
+static inline void scx_offline_ecaps(struct rq *rq) {}
+static inline void scx_discard_ecaps_to_sync(s32 cpu, struct scx_sched_pcpu *pcpu) {}
+static inline void scx_discard_stale_ecaps_syncs(void) {}
+static inline struct scx_dispatch_q *scx_resolve_local_dsq(struct scx_sched *sch, struct rq *rq, struct task_struct *p, u64 *enq_flags) { return &rq->scx.local_dsq; }
+static inline bool scx_task_reenq_on_cap_revoke(struct rq *rq, struct task_struct *p) { return false; }
+static inline void scx_reenq_reject(struct rq *rq) {}
+static inline void scx_rescue_charge(struct rq *rq, s64 delta_exec) {}
+static inline void scx_rescue_end(struct rq *rq) {}
+static inline bool scx_rescue_keep(struct rq *rq, struct task_struct *p) { return false; }
+static inline void scx_rescue_flush(struct rq *rq) {}
+static inline void scx_rescue_dump(struct seq_buf *s, struct rq *rq) {}
+static inline void scx_rescue_set_knobs(struct scx_sched *sch) {}
+static inline void scx_rescue_init(struct rq *rq) {}
+static inline void scx_dec_has_subs(struct scx_sched *sch) {}
+
+#endif	/* CONFIG_EXT_SUB_SCHED */
+
+/**
+ * scx_for_each_descendant_pre - pre-order walk of a sched's descendants
+ * @pos: iteration cursor
+ * @root: sched to walk the descendants of
+ *
+ * Walk @root's descendants. @root is included in the iteration and the first
+ * node to be visited. Must be called with scx_enable_mutex, scx_sched_lock, or
+ * RCU read lock.
+ */
+#define scx_for_each_descendant_pre(pos, root)					\
+	for ((pos) = scx_next_descendant_pre(NULL, (root)); (pos);		\
+	     (pos) = scx_next_descendant_pre((pos), (root)))
+
+#ifdef CONFIG_EXT_SUB_SCHED
+
+/**
+ * scx_missing_caps - The caps in @needed that @sch lacks on @cpu
+ * @sch: sched to test
+ * @cpu: cpu to test on
+ * @needed: bitmask of SCX_CAP_* values
+ *
+ * Return the caps in @needed that @sch lacks for @cpu, 0 if it holds them all.
+ */
+static inline u64 scx_missing_caps(struct scx_sched *sch, s32 cpu, u64 needed)
+{
+	u64 ecaps;
+
+	/* no sub-scheds, no missing caps */
+	if (!scx_has_subs())
+		return 0;
+
+	/* root holds every cap on every cpu */
+	if (!sch->level)
+		return 0;
+
+	ecaps = READ_ONCE(per_cpu_ptr(sch->pcpu, cpu)->ecaps);
+
+	return needed & ~ecaps;
+}
+
+/*
+ * Cap semantics: which caps an action requires, and which caps a cap implies.
+ * Keep all such mappings collected here.
+ */
+
+/* map @enq_flags to the SCX_CAP_* bit required for the local-DSQ insert */
+static inline u64 scx_caps_for_enq(u64 enq_flags)
+{
+	/* a restored task must be put into the local DSQ regardless of caps */
+	if (unlikely(enq_flags & SCX_ENQ_IGNORE_CAPS))
+		return 0;
+	if (enq_flags & SCX_ENQ_IMMED)
+		return SCX_CAP_ENQ_IMMED;
+	return SCX_CAP_ENQ;
+}
+
+/* map queued @p to the SCX_CAP_* bit required to stay on its local DSQ */
+static inline u64 scx_caps_for_task(struct task_struct *p)
+{
+	if (p->scx.flags & SCX_TASK_IMMED)
+		return SCX_CAP_ENQ_IMMED;
+	return SCX_CAP_ENQ;
+}
+
+/* the cap @sch needs to preempt @rq's current task, 0 if none */
+static inline u64 scx_caps_for_preempt(struct scx_sched *sch, struct rq *rq, u64 enq_flags)
+{
+	struct task_struct *curr = rq->curr;
+
+	/* a kernel-forced placement preempts regardless of caps */
+	if (unlikely(enq_flags & SCX_ENQ_IGNORE_CAPS))
+		return 0;
+	/* a non-ext task can't be preempted by ext, own-subtree needs no cap */
+	if (curr->sched_class != &ext_sched_class ||
+	    scx_is_descendant(scx_task_sched(curr), sch))
+		return 0;
+	return SCX_CAP_PREEMPT;
+}
+
+/* caps implied by holding @cap */
+static inline u64 scx_caps_implied(u64 cap)
+{
+	switch (cap) {
+	case SCX_CAP_PREEMPT:
+		return SCX_CAP_ENQ | SCX_CAP_ENQ_IMMED;
+	case SCX_CAP_ENQ:
+		return SCX_CAP_ENQ_IMMED;
+	}
+	return 0;
+}
+
+/* may @p keep running on @rq's cpu? requires baseline cpu access */
+static inline bool scx_task_can_stay_on_cpu(struct rq *rq, struct task_struct *p)
+{
+	if (!scx_has_subs())
+		return true;
+
+	/* a migration-disabled task is let in without caps, keep it likewise */
+	if (unlikely(is_migration_disabled(p)))
+		return true;
+
+	return likely(!scx_missing_caps(scx_task_sched(p), cpu_of(rq), SCX_CAP_BASE));
+}
+
+/* the task admitted for rescue on @rq, NULL if none */
+static inline struct task_struct *scx_rescuee(struct rq *rq)
+{
+	lockdep_assert_rq_held(rq);
+
+	if (!scx_has_subs())
+		return NULL;
+
+	return rq->scx.rescue.curr;
+}
+
+#else	/* CONFIG_EXT_SUB_SCHED */
+
+static inline u64 scx_missing_caps(struct scx_sched *sch, s32 cpu, u64 needed) { return 0; }
+static inline u64 scx_caps_for_preempt(struct scx_sched *sch, struct rq *rq, u64 enq_flags) { return 0; }
+static inline bool scx_task_can_stay_on_cpu(struct rq *rq, struct task_struct *p) { return true; }
+static inline struct task_struct *scx_rescuee(struct rq *rq) { return NULL; }
+
+#endif	/* CONFIG_EXT_SUB_SCHED */
+
+#endif /* _KERNEL_SCHED_EXT_SUB_H */
diff --git a/kernel/sched/ext/types.h b/kernel/sched/ext/types.h
index bc74eafd43f1..943d8d429a2c 100644
--- a/kernel/sched/ext/types.h
+++ b/kernel/sched/ext/types.h
@@ -19,6 +19,18 @@ enum scx_consts {
 	SCX_DSP_MAX_LOOPS		= 32,
 	SCX_WATCHDOG_MAX_TIMEOUT	= 30 * HZ,
 
+	/* rescue knob defaults and limits, see scx_rescue_timerfn() */
+	SCX_RESCUE_DFL_BW_PPT		= 20,		/* parts per thousand, 2% */
+	SCX_RESCUE_MAX_BW_PPT		= 250,		/* 25% */
+	SCX_RESCUE_DISABLE		= U32_MAX,	/* disables rescue */
+	SCX_RESCUE_DFL_QUANTUM_US	= 5000,
+	SCX_RESCUE_MIN_QUANTUM_US	= 1000,
+	SCX_RESCUE_MAX_QUANTUM_US	= 100000,
+	SCX_RESCUE_MIN_SLICE_US		= 1000,		/* floor of the divided slice */
+	SCX_RESCUE_OVERLOAD_MULT	= 16,		/* overload threshold in funding periods */
+	SCX_RESCUE_MIN_OVERLOAD_MS	= 1000,
+	SCX_RESCUE_MAX_OVERLOAD_MS	= 15000,
+
 	/* per-CPU chunk size for p->scx.tid allocation, see scx_alloc_tid() */
 	SCX_TID_CHUNK			= 1024,
 
@@ -41,17 +53,22 @@ enum scx_consts {
 	SCX_BYPASS_LB_MIN_DELTA_DIV	= 4,
 	SCX_BYPASS_LB_BATCH		= 256,
 
-	SCX_REENQ_LOCAL_MAX_REPEAT	= 256,
+	SCX_REENQ_MAX_REPEAT		= 256,
 
 	SCX_SUB_MAX_DEPTH		= 4,
 };
 
 /*
- * Per-cid topology info. For each topology level (core, LLC, node), records
- * the first cid in the unit and its global index. Global indices are
- * consecutive integers assigned in cid-walk order, so e.g. core_idx ranges
- * over [0, nr_cores_at_init) with no gaps. No-topo cids have all fields set
- * to -1.
+ * Per-cid topology info. For each topology level (core, LLC, node) and shard,
+ * records the first cid in the unit and its global index. Global indices are
+ * consecutive integers assigned in cid-walk order, so e.g. core_idx ranges over
+ * [0, nr_cores_at_init) with no gaps. No-topo cids have core/LLC/node fields
+ * set to -1 but always have valid shard assignments.
+ *
+ * Shards are contiguous CID ranges used as scalable locking/work domains for
+ * sub-scheduler operations. By default each LLC becomes one shard, split into
+ * smaller shards if the LLC exceeds the target size. No-topo cids are packed
+ * into their own max-sized shards.
  *
  * @core_cid: first cid of this cid's core (smt-sibling group)
  * @core_idx: global index of that core, in [0, nr_cores_at_init)
@@ -59,6 +76,8 @@ enum scx_consts {
  * @llc_idx: global index of that LLC, in [0, nr_llcs_at_init)
  * @node_cid: first cid of this cid's NUMA node
  * @node_idx: global index of that node, in [0, nr_nodes_at_init)
+ * @shard_cid: first cid of this cid's shard
+ * @shard_idx: global index of that shard, in [0, scx_nr_cid_shards)
  */
 struct scx_cid_topo {
 	s32 core_cid;
@@ -67,6 +86,24 @@ struct scx_cid_topo {
 	s32 llc_idx;
 	s32 node_cid;
 	s32 node_idx;
+	s32 shard_cid;
+	s32 shard_idx;
+};
+
+enum scx_cid_consts {
+	SCX_CID_SHARD_SIZE_DFL		= 24,
+	SCX_CID_SHARD_MAX_CPUS		= 512,
+};
+
+/*
+ * Per-shard metadata for O(1) shard->cid-range lookup.
+ *
+ * @base_cid: first cid of the shard
+ * @nr_cids: number of cids in the shard
+ */
+struct scx_cid_shard {
+	s32			base_cid;
+	s32			nr_cids;
 };
 
 /*
@@ -91,7 +128,7 @@ struct scx_cmask {
 	u32 base;
 	u32 nr_cids;
 	u32 alloc_words;
-	u64 bits[] __counted_by(alloc_words);
+	u64 bits[];
 };
 
 /*
@@ -147,4 +184,41 @@ struct scx_cmask {
 #define SCX_CMASK_DEFINE_SHARD(NAME, BASE, NR_CIDS)				\
 	__SCX_CMASK_DEFINE(NAME, BASE, NR_CIDS, SCX_CID_SHARD_MAX_CPUS)
 
+/*
+ * scx_cmask_ref: validated reference to a BPF-arena cmask.
+ *
+ * scx_cmask_ref_init() snapshots @base/@nr_cids. The snapshot is what
+ * downstream code uses for sizing - the live header can be mutated concurrently
+ * by BPF.
+ *
+ * scx_cmask_ref_shard() reads one shard into a cmask. scx_cmask_ref_or() and
+ * scx_cmask_ref_copy() write back into the referenced arena cmask, bounded by
+ * the snapshot.
+ *
+ * Typical input use:
+ *
+ *	struct scx_cmask_ref ref;
+ *	SCX_CMASK_DEFINE(shard, 0, SCX_CID_SHARD_MAX_CPUS);
+ *	s32 idx, ret;
+ *
+ *	ret = scx_cmask_ref_init(sch, src, &ref);
+ *	if (ret < 0)
+ *		return ret;
+ *
+ *	for (idx = ref.shard_first; idx < ref.shard_end; idx++) {
+ *		scx_cmask_ref_shard(&ref, idx, shard);
+ *		if (!shard->nr_cids)
+ *			continue;
+ *		... use idx and shard ...
+ *	}
+ */
+struct scx_cmask_ref {
+	struct scx_sched	*sch;
+	struct scx_cmask	*src;
+	u32			base;
+	u32			nr_cids;
+	s32			shard_first;
+	s32			shard_end;
+};
+
 #endif /* _KERNEL_SCHED_EXT_TYPES_H */
diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
index 052435f4d3e3..eb73b65ce6c4 100644
--- a/kernel/sched/idle.c
+++ b/kernel/sched/idle.c
@@ -503,7 +503,13 @@ static void set_next_task_idle(struct rq *rq, struct task_struct *next, bool fir
 
 struct task_struct *pick_task_idle(struct rq *rq, struct rq_flags *rf)
 {
-	scx_update_idle(rq, true, false);
+	/*
+	 * Notify scx only on an idle-to-idle re-pick (the cpu was already idle).
+	 * A real task->idle transition is delivered by set_next_task_idle(), so
+	 * calling here too would duplicate it.
+	 */
+	if (scx_enabled() && is_idle_task(rq->curr))
+		scx_update_idle(rq, true, false);
 	return rq->idle;
 }
 
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 26ae13c86b69..e656c7059bf8 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -787,39 +787,60 @@ enum scx_rq_flags {
 	 */
 	SCX_RQ_ONLINE		= 1 << 0,
 	SCX_RQ_CAN_STOP_TICK	= 1 << 1,
-	SCX_RQ_BAL_KEEP		= 1 << 3, /* balance decided to keep current */
 	SCX_RQ_CLK_VALID	= 1 << 5, /* RQ clock is fresh and valid */
 	SCX_RQ_BAL_CB_PENDING	= 1 << 6, /* must queue a cb after dispatching */
+	SCX_RQ_SUB_IDLE_RENOTIFY	= 1 << 7, /* sub-scheds are owed update_idle() */
+	SCX_RQ_ROOT_IDLE_RENOTIFY	= 1 << 8, /* the root is owed update_idle() */
 
 	SCX_RQ_IN_WAKEUP	= 1 << 16,
-	SCX_RQ_IN_BALANCE	= 1 << 17,
+	SCX_RQ_IN_DISPATCH	= 1 << 17,
+};
+
+/* per-rq rescue execution state, see scx_rescue_timerfn() */
+struct scx_rq_rescue {
+	struct scx_dispatch_q	dsq;			/* stranded tasks awaiting rescue */
+	s64			budget;			/* execution token bucket, ns */
+	u64			clock;			/* last budget accrual timestamp */
+	struct task_struct	*curr;			/* task being rescued, one at a time */
+	s64			slice;			/* curr's admitted slice */
+	u64			exec_snap;		/* sum_exec_runtime at admission */
+	struct timer_list	timer;			/* paces admission and escalation */
+	u64			kill_at;		/* last ejection, init before any */
 };
 
 struct scx_rq {
 	struct scx_dispatch_q	local_dsq;
+#ifdef CONFIG_EXT_SUB_SCHED
+	struct scx_dispatch_q	reject_dsq;		/* staging for cap-rejected tasks */
+	struct scx_rq_rescue	rescue;
+#endif
 	struct list_head	runnable_list;		/* runnable tasks on this rq */
 	struct list_head	ddsp_deferred_locals;	/* deferred ddsps from enq */
 	unsigned long		ops_qseq;
-	u64			extra_enq_flags;	/* see move_task_to_local_dsq() */
+	/* both stashed across the activate_task() in move_remote_task_to_local_dsq() */
+	u64			remote_activate_enq_flags;
+	struct scx_sched	*remote_activate_sch;
 	u32			nr_running;
 	u32			cpuperf_target;		/* [0, SCHED_CAPACITY_SCALE] */
 	bool			in_select_cpu;
 	bool			cpu_released;
 	u32			flags;
 	u32			nr_immed;		/* ENQ_IMMED tasks on local_dsq */
+#ifdef CONFIG_SCHED_CORE
+	u32			lock_drop_seq;	/* nr dispatch lock releases */
+#endif
 	u64			clock;			/* current per-rq clock -- see scx_bpf_now() */
-	cpumask_var_t		cpus_to_kick;
-	cpumask_var_t		cpus_to_kick_if_idle;
-	cpumask_var_t		cpus_to_preempt;
-	cpumask_var_t		cpus_to_wait;
+#ifdef CONFIG_EXT_SUB_SCHED
+	struct llist_head	ecaps_to_sync;		/* pending ecaps syncs */
+	struct task_struct	*sub_dispatch_prev;
+#endif
 	cpumask_var_t		cpus_to_sync;
 	bool			kick_sync_pending;
 	unsigned long		kick_sync;
 
-	struct task_struct	*sub_dispatch_prev;
+	struct list_head	sched_pcpus_to_kick;	/* see kick_cpus_irq_workfn() */
 
 	raw_spinlock_t		deferred_reenq_lock;
-	u64			deferred_reenq_locals_seq;
 	struct list_head	deferred_reenq_locals;	/* scheds requesting reenq of local DSQ */
 	struct list_head	deferred_reenq_users;	/* user DSQs requesting reenq */
 	struct balance_callback	deferred_bal_cb;
@@ -1361,6 +1382,7 @@ struct rq {
 	unsigned int		core_forceidle_seq;
 	unsigned int		core_forceidle_occupation;
 	u64			core_forceidle_start;
+	unsigned int		core_pick_in_flight;
 #endif /* CONFIG_SCHED_CORE */
 
 	/* Scratch cpumask to be temporarily used under rq_lock */