1 files changed, 147 insertions, 178 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 2f24a6aa13bf..8048a4c04b47 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -42,50 +42,14 @@
 #include "amdgpu_reset.h"
 
 /*
- * Fences mark an event in the GPUs pipeline and are used
- * for GPU/CPU synchronization.  When the fence is written,
- * it is expected that all buffers associated with that fence
- * are no longer in use by the associated ring on the GPU and
- * that the relevant GPU caches have been flushed.
- */
-
-struct amdgpu_fence {
-	struct dma_fence base;
-
-	/* RB, DMA, etc. */
-	struct amdgpu_ring		*ring;
-	ktime_t				start_timestamp;
-};
-
-static struct kmem_cache *amdgpu_fence_slab;
-
-int amdgpu_fence_slab_init(void)
-{
-	amdgpu_fence_slab = KMEM_CACHE(amdgpu_fence, SLAB_HWCACHE_ALIGN);
-	if (!amdgpu_fence_slab)
-		return -ENOMEM;
-	return 0;
-}
-
-void amdgpu_fence_slab_fini(void)
-{
-	rcu_barrier();
-	kmem_cache_destroy(amdgpu_fence_slab);
-}
-/*
  * Cast helper
  */
 static const struct dma_fence_ops amdgpu_fence_ops;
-static const struct dma_fence_ops amdgpu_job_fence_ops;
 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
 {
 	struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
 
-	if (__f->base.ops == &amdgpu_fence_ops ||
-	    __f->base.ops == &amdgpu_job_fence_ops)
-		return __f;
-
-	return NULL;
+	return __f;
 }
 
 /**
@@ -129,57 +93,30 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
  * amdgpu_fence_emit - emit a fence on the requested ring
  *
  * @ring: ring the fence is associated with
- * @f: resulting fence object
- * @job: job the fence is embedded in
+ * @af: amdgpu fence input
  * @flags: flags to pass into the subordinate .emit_fence() call
  *
  * Emits a fence command on the requested ring (all asics).
- * Returns 0 on success, -ENOMEM on failure.
  */
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amdgpu_job *job,
-		      unsigned int flags)
+void amdgpu_fence_emit(struct amdgpu_ring *ring, struct amdgpu_fence *af,
+		       unsigned int flags)
 {
 	struct amdgpu_device *adev = ring->adev;
 	struct dma_fence *fence;
-	struct amdgpu_fence *am_fence;
 	struct dma_fence __rcu **ptr;
 	uint32_t seq;
-	int r;
 
-	if (job == NULL) {
-		/* create a sperate hw fence */
-		am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_ATOMIC);
-		if (am_fence == NULL)
-			return -ENOMEM;
-		fence = &am_fence->base;
-		am_fence->ring = ring;
-	} else {
-		/* take use of job-embedded fence */
-		fence = &job->hw_fence;
-	}
+	fence = &af->base;
+	af->ring = ring;
 
 	seq = ++ring->fence_drv.sync_seq;
-	if (job && job->job_run_counter) {
-		/* reinit seq for resubmitted jobs */
-		fence->seqno = seq;
-		/* TO be inline with external fence creation and other drivers */
-		dma_fence_get(fence);
-	} else {
-		if (job) {
-			dma_fence_init(fence, &amdgpu_job_fence_ops,
-				       &ring->fence_drv.lock,
-				       adev->fence_context + ring->idx, seq);
-			/* Against remove in amdgpu_job_{free, free_cb} */
-			dma_fence_get(fence);
-		} else {
-			dma_fence_init(fence, &amdgpu_fence_ops,
-				       &ring->fence_drv.lock,
-				       adev->fence_context + ring->idx, seq);
-		}
-	}
+	dma_fence_init(fence, &amdgpu_fence_ops,
+		       &ring->fence_drv.lock,
+		       adev->fence_context + ring->idx, seq);
 
 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
 			       seq, flags | AMDGPU_FENCE_FLAG_INT);
+
 	pm_runtime_get_noresume(adev_to_drm(adev)->dev);
 	ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
 	if (unlikely(rcu_dereference_protected(*ptr, 1))) {
@@ -190,10 +127,13 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
 		rcu_read_unlock();
 
 		if (old) {
-			r = dma_fence_wait(old, false);
+			/*
+			 * dma_fence_wait(old, false) is not interruptible.
+			 * It will not return an error in this case.
+			 * So we can safely ignore the return value.
+			 */
+			dma_fence_wait(old, false);
 			dma_fence_put(old);
-			if (r)
-				return r;
 		}
 	}
 
@@ -203,10 +143,6 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
 	 * emitting the fence would mess up the hardware ring buffer.
 	 */
 	rcu_assign_pointer(*ptr, dma_fence_get(fence));
-
-	*f = fence;
-
-	return 0;
 }
 
 /**
@@ -280,7 +216,7 @@ bool amdgpu_fence_process(struct amdgpu_ring *ring)
 
 	} while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
 
-	if (del_timer(&ring->fence_drv.fallback_timer) &&
+	if (timer_delete(&ring->fence_drv.fallback_timer) &&
 	    seq != ring->fence_drv.sync_seq)
 		amdgpu_fence_schedule_fallback(ring);
 
@@ -306,7 +242,6 @@ bool amdgpu_fence_process(struct amdgpu_ring *ring)
 
 		dma_fence_signal(fence);
 		dma_fence_put(fence);
-		pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
 		pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 	} while (last_seq != seq);
 
@@ -322,11 +257,13 @@ bool amdgpu_fence_process(struct amdgpu_ring *ring)
  */
 static void amdgpu_fence_fallback(struct timer_list *t)
 {
-	struct amdgpu_ring *ring = from_timer(ring, t,
-					      fence_drv.fallback_timer);
+	struct amdgpu_ring *ring = timer_container_of(ring, t,
+						      fence_drv.fallback_timer);
 
 	if (amdgpu_fence_process(ring))
-		DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
+		dev_warn(ring->adev->dev,
+			 "Fence fallback timer expired on ring %s\n",
+			 ring->name);
 }
 
 /**
@@ -501,7 +438,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
  * @ring: ring to init the fence driver on
  *
  * Init the fence driver for the requested ring (all asics).
- * Helper function for amdgpu_fence_driver_init().
+ * Helper function for amdgpu_fence_driver_sw_init().
  */
 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
 {
@@ -618,7 +555,7 @@ void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
 			amdgpu_irq_put(adev, ring->fence_drv.irq_src,
 				       ring->fence_drv.irq_type);
 
-		del_timer_sync(&ring->fence_drv.fallback_timer);
+		timer_delete_sync(&ring->fence_drv.fallback_timer);
 	}
 }
 
@@ -698,36 +635,6 @@ void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
 }
 
 /**
- * amdgpu_fence_driver_clear_job_fences - clear job embedded fences of ring
- *
- * @ring: fence of the ring to be cleared
- *
- */
-void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
-{
-	int i;
-	struct dma_fence *old, **ptr;
-
-	for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
-		ptr = &ring->fence_drv.fences[i];
-		old = rcu_dereference_protected(*ptr, 1);
-		if (old && old->ops == &amdgpu_job_fence_ops) {
-			struct amdgpu_job *job;
-
-			/* For non-scheduler bad job, i.e. failed ib test, we need to signal
-			 * it right here or we won't be able to track them in fence_drv
-			 * and they will remain unsignaled during sa_bo free.
-			 */
-			job = container_of(old, struct amdgpu_job, hw_fence);
-			if (!job->base.s_fence && !dma_fence_is_signaled(old))
-				dma_fence_signal(old);
-			RCU_INIT_POINTER(*ptr, NULL);
-			dma_fence_put(old);
-		}
-	}
-}
-
-/**
  * amdgpu_fence_driver_set_error - set error code on fences
  * @ring: the ring which contains the fences
  * @error: the error code to set
@@ -764,6 +671,128 @@ void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
 	amdgpu_fence_process(ring);
 }
 
+
+/*
+ * Kernel queue reset handling
+ *
+ * The driver can reset individual queues for most engines, but those queues
+ * may contain work from multiple contexts.  Resetting the queue will reset
+ * lose all of that state.  In order to minimize the collateral damage, the
+ * driver will save the ring contents which are not associated with the guilty
+ * context prior to resetting the queue.  After resetting the queue the queue
+ * contents from the other contexts is re-emitted to the rings so that it can
+ * be processed by the engine.  To handle this, we save the queue's write
+ * pointer (wptr) in the fences associated with each context.  If we get a
+ * queue timeout, we can then use the wptrs from the fences to determine
+ * which data needs to be saved out of the queue's ring buffer.
+ */
+
+/**
+ * amdgpu_ring_set_fence_errors_and_reemit - Set dma_fence errors and reemit
+ *
+ * @ring: the ring to operate on
+ * @guilty_fence: fence of the ring to update
+ *
+ */
+void amdgpu_ring_set_fence_errors_and_reemit(struct amdgpu_ring *ring,
+					     struct amdgpu_fence *guilty_fence)
+{
+	struct dma_fence *unprocessed;
+	struct dma_fence __rcu **ptr;
+	struct amdgpu_fence *fence;
+	unsigned long flags;
+	u32 seq, last_seq;
+	unsigned int i;
+	bool is_guilty_fence;
+	bool is_guilty_context;
+
+	last_seq = amdgpu_fence_read(ring) & ring->fence_drv.num_fences_mask;
+	seq = ring->fence_drv.sync_seq & ring->fence_drv.num_fences_mask;
+
+	ring->reemit = true;
+	amdgpu_ring_alloc(ring, ring->ring_backup_entries_to_copy);
+	spin_lock_irqsave(&ring->fence_drv.lock, flags);
+	do {
+		last_seq++;
+		last_seq &= ring->fence_drv.num_fences_mask;
+
+		ptr = &ring->fence_drv.fences[last_seq];
+		rcu_read_lock();
+		unprocessed = rcu_dereference(*ptr);
+
+		if (unprocessed && !dma_fence_is_signaled_locked(unprocessed)) {
+			fence = container_of(unprocessed, struct amdgpu_fence, base);
+			is_guilty_fence = fence == guilty_fence;
+			is_guilty_context = fence->context == guilty_fence->context;
+
+			/* mark all fences from the guilty context with an error */
+			if (is_guilty_fence)
+				dma_fence_set_error(&fence->base, -ETIME);
+			else if (is_guilty_context)
+				dma_fence_set_error(&fence->base, -ECANCELED);
+
+			/* reemit the packet stream and update wptrs */
+			fence->ib_wptr = ring->wptr;
+			for (i = 0; i < fence->ib_dw_size; i++) {
+				/* Skip the IB(s) for the guilty context. */
+				if (is_guilty_context &&
+				    i >= fence->skip_ib_dw_start_offset &&
+				    i < fence->skip_ib_dw_end_offset)
+					amdgpu_ring_write(ring, ring->funcs->nop);
+				else
+					amdgpu_ring_write(ring,
+							  ring->ring_backup[fence->backup_idx + i]);
+			}
+		}
+		rcu_read_unlock();
+	} while (last_seq != seq);
+	spin_unlock_irqrestore(&ring->fence_drv.lock, flags);
+	amdgpu_ring_commit(ring);
+	ring->reemit = false;
+}
+
+static void amdgpu_ring_backup_unprocessed_command(struct amdgpu_ring *ring,
+						   struct amdgpu_fence *af)
+{
+	unsigned int first_idx = af->ib_wptr & ring->buf_mask;
+	unsigned int dw_size = af->ib_dw_size;
+	unsigned int i;
+
+	af->backup_idx = ring->ring_backup_entries_to_copy;
+	/* Backup the contents of the ring buffer. */
+	for (i = first_idx; dw_size > 0; ++i, i &= ring->buf_mask, --dw_size)
+		ring->ring_backup[ring->ring_backup_entries_to_copy++] = ring->ring[i];
+}
+
+void amdgpu_ring_backup_unprocessed_commands(struct amdgpu_ring *ring,
+					     struct amdgpu_fence *guilty_fence)
+{
+	struct dma_fence *unprocessed;
+	struct dma_fence __rcu **ptr;
+	struct amdgpu_fence *fence;
+	u32 seq, last_seq;
+
+	last_seq = amdgpu_fence_read(ring) & ring->fence_drv.num_fences_mask;
+	seq = ring->fence_drv.sync_seq & ring->fence_drv.num_fences_mask;
+	ring->ring_backup_entries_to_copy = 0;
+
+	do {
+		last_seq++;
+		last_seq &= ring->fence_drv.num_fences_mask;
+
+		ptr = &ring->fence_drv.fences[last_seq];
+		rcu_read_lock();
+		unprocessed = rcu_dereference(*ptr);
+
+		if (unprocessed && !dma_fence_is_signaled(unprocessed)) {
+			fence = container_of(unprocessed, struct amdgpu_fence, base);
+
+			amdgpu_ring_backup_unprocessed_command(ring, fence);
+		}
+		rcu_read_unlock();
+	} while (last_seq != seq);
+}
+
 /*
  * Common fence implementation
  */
@@ -778,13 +807,6 @@ static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
 	return (const char *)to_amdgpu_fence(f)->ring->name;
 }
 
-static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
-{
-	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
-
-	return (const char *)to_amdgpu_ring(job->base.sched)->name;
-}
-
 /**
  * amdgpu_fence_enable_signaling - enable signalling on fence
  * @f: fence
@@ -802,23 +824,6 @@ static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
 }
 
 /**
- * amdgpu_job_fence_enable_signaling - enable signalling on job fence
- * @f: fence
- *
- * This is the simliar function with amdgpu_fence_enable_signaling above, it
- * only handles the job embedded fence.
- */
-static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
-{
-	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
-
-	if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
-		amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
-
-	return true;
-}
-
-/**
  * amdgpu_fence_free - free up the fence memory
  *
  * @rcu: RCU callback head
@@ -830,22 +835,7 @@ static void amdgpu_fence_free(struct rcu_head *rcu)
 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
 
 	/* free fence_slab if it's separated fence*/
-	kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
-}
-
-/**
- * amdgpu_job_fence_free - free up the job with embedded fence
- *
- * @rcu: RCU callback head
- *
- * Free up the job with embedded fence after the RCU grace period.
- */
-static void amdgpu_job_fence_free(struct rcu_head *rcu)
-{
-	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
-
-	/* free job if fence has a parent job */
-	kfree(container_of(f, struct amdgpu_job, hw_fence));
+	kfree(to_amdgpu_fence(f));
 }
 
 /**
@@ -861,19 +851,6 @@ static void amdgpu_fence_release(struct dma_fence *f)
 	call_rcu(&f->rcu, amdgpu_fence_free);
 }
 
-/**
- * amdgpu_job_fence_release - callback that job embedded fence can be freed
- *
- * @f: fence
- *
- * This is the simliar function with amdgpu_fence_release above, it
- * only handles the job embedded fence.
- */
-static void amdgpu_job_fence_release(struct dma_fence *f)
-{
-	call_rcu(&f->rcu, amdgpu_job_fence_free);
-}
-
 static const struct dma_fence_ops amdgpu_fence_ops = {
 	.get_driver_name = amdgpu_fence_get_driver_name,
 	.get_timeline_name = amdgpu_fence_get_timeline_name,
@@ -881,13 +858,6 @@ static const struct dma_fence_ops amdgpu_fence_ops = {
 	.release = amdgpu_fence_release,
 };
 
-static const struct dma_fence_ops amdgpu_job_fence_ops = {
-	.get_driver_name = amdgpu_fence_get_driver_name,
-	.get_timeline_name = amdgpu_job_fence_get_timeline_name,
-	.enable_signaling = amdgpu_job_fence_enable_signaling,
-	.release = amdgpu_job_fence_release,
-};
-
 /*
  * Fence debugfs
  */
@@ -957,7 +927,6 @@ static int gpu_recover_get(void *data, u64 *val)
 
 	*val = atomic_read(&adev->reset_domain->reset_res);
 
-	pm_runtime_mark_last_busy(dev->dev);
 	pm_runtime_put_autosuspend(dev->dev);
 
 	return 0;