perf: Fix perf_event_pmu_context serialization

Syzkaller triggered a WARN in put_pmu_ctx().

  WARNING: CPU: 1 PID: 2245 at kernel/events/core.c:4925 put_pmu_ctx+0x1f0/0x278

This is because there is no locking around the access of "if
(!epc->ctx)" in find_get_pmu_context() and when it is set to NULL in
put_pmu_ctx().

The decrement of the reference count in put_pmu_ctx() also happens
outside of the spinlock, leading to the possibility of this order of
events, and the context being cleared in put_pmu_ctx(), after its
refcount is non zero:

 CPU0                                   CPU1
 find_get_pmu_context()
   if (!epc->ctx) == false
                                        put_pmu_ctx()
                                        atomic_dec_and_test(&epc->refcount) == true
                                        epc->refcount == 0
     atomic_inc(&epc->refcount);
     epc->refcount == 1
                                        list_del_init(&epc->pmu_ctx_entry);
	                                      epc->ctx = NULL;

Another issue is that WARN_ON for no active PMU events in put_pmu_ctx()
is outside of the lock. If the perf_event_pmu_context is an embedded
one, even after clearing it, it won't be deleted and can be re-used. So
the warning can trigger. For this reason it also needs to be moved
inside the lock.

The above warning is very quick to trigger on Arm by running these two
commands at the same time:

  while true; do perf record -- ls; done
  while true; do perf record -- ls; done

[peterz: atomic_dec_and_raw_lock*()]
Fixes: bd2756811766 ("perf: Rewrite core context handling")
Reported-by: syzbot+697196bc0265049822bd@syzkaller.appspotmail.com
Signed-off-by: James Clark <james.clark@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ravi Bangoria <ravi.bangoria@amd.com>
Link: https://lore.kernel.org/r/20230127143141.1782804-2-james.clark@arm.com

1 files changed, 17 insertions, 22 deletions

diff --git a/kernel/events/core.c b/kernel/events/core.c
index d56328e5080e..c4be13e50547 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -4813,19 +4813,17 @@ find_get_pmu_context(struct pmu *pmu, struct perf_event_context *ctx,
 
 		cpc = per_cpu_ptr(pmu->cpu_pmu_context, event->cpu);
 		epc = &cpc->epc;
-
+		raw_spin_lock_irq(&ctx->lock);
 		if (!epc->ctx) {
 			atomic_set(&epc->refcount, 1);
 			epc->embedded = 1;
-			raw_spin_lock_irq(&ctx->lock);
 			list_add(&epc->pmu_ctx_entry, &ctx->pmu_ctx_list);
 			epc->ctx = ctx;
-			raw_spin_unlock_irq(&ctx->lock);
 		} else {
 			WARN_ON_ONCE(epc->ctx != ctx);
 			atomic_inc(&epc->refcount);
 		}
-
+		raw_spin_unlock_irq(&ctx->lock);
 		return epc;
 	}
 
@@ -4896,33 +4894,30 @@ static void free_epc_rcu(struct rcu_head *head)
 
 static void put_pmu_ctx(struct perf_event_pmu_context *epc)
 {
+	struct perf_event_context *ctx = epc->ctx;
 	unsigned long flags;
 
-	if (!atomic_dec_and_test(&epc->refcount))
+	/*
+	 * XXX
+	 *
+	 * lockdep_assert_held(&ctx->mutex);
+	 *
+	 * can't because of the call-site in _free_event()/put_event()
+	 * which isn't always called under ctx->mutex.
+	 */
+	if (!atomic_dec_and_raw_lock_irqsave(&epc->refcount, &ctx->lock, flags))
 		return;
 
-	if (epc->ctx) {
-		struct perf_event_context *ctx = epc->ctx;
+	WARN_ON_ONCE(list_empty(&epc->pmu_ctx_entry));
 
-		/*
-		 * XXX
-		 *
-		 * lockdep_assert_held(&ctx->mutex);
-		 *
-		 * can't because of the call-site in _free_event()/put_event()
-		 * which isn't always called under ctx->mutex.
-		 */
-
-		WARN_ON_ONCE(list_empty(&epc->pmu_ctx_entry));
-		raw_spin_lock_irqsave(&ctx->lock, flags);
-		list_del_init(&epc->pmu_ctx_entry);
-		epc->ctx = NULL;
-		raw_spin_unlock_irqrestore(&ctx->lock, flags);
-	}
+	list_del_init(&epc->pmu_ctx_entry);
+	epc->ctx = NULL;
 
 	WARN_ON_ONCE(!list_empty(&epc->pinned_active));
 	WARN_ON_ONCE(!list_empty(&epc->flexible_active));
 
+	raw_spin_unlock_irqrestore(&ctx->lock, flags);
+
 	if (epc->embedded)
 		return;