[PATCH 0/7] DRM fence list helpers, V3D CSD support.

[PATCH 0/7] DRM fence list helpers, V3D CSD support.

[Eric Anholt]

The DRM fence list helpers came from needing to fix implicit
synchronization on V3D (serious flickering in X11 once you had
multiple clients involved), a bug I think I noticed in panfrost, and
having recently reviewed lima and liked their model.  Compared to
lima, I chose to use xarray as the storage, as I think it cleans
things up significantly.  The last_dep thing is a little gross, but
Matthew Wilcox is thinking about giving us an xarray pop function soon
that would let us drop that.

Unfortunately, rebasing the V3D side of things to before the job
refactor was a bit more work than I was ready for, so I'm resending it
and Ccing the Broadcom folks again in the hope that they might be able
to review or ack it.  I would still accept an ack from anyone willing
to do so -- the userspace has been sitting around for a long time at
this point, and I'd love to get it merged.

Eric Anholt (7):
  drm/v3d: Switch the type of job-> to reduce casting.
  drm/v3d: Refactor job management.
  drm/v3d: Add support for compute shader dispatch.
  drm/v3d: Drop reservation of a shared slot in the dma-buf
    reservations.
  drm: Add helpers for setting up an array of dma_fence dependencies.
  drm/v3d: Add missing implicit synchronization.
  drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

 drivers/gpu/drm/drm_gem.c         |  94 ++++++
 drivers/gpu/drm/lima/lima_gem.c   |  37 +-
 drivers/gpu/drm/lima/lima_sched.c |  66 +---
 drivers/gpu/drm/lima/lima_sched.h |   6 +-
 drivers/gpu/drm/v3d/v3d_debugfs.c |  22 ++
 drivers/gpu/drm/v3d/v3d_drv.c     |  10 +-
 drivers/gpu/drm/v3d/v3d_drv.h     | 104 +++---
 drivers/gpu/drm/v3d/v3d_fence.c   |   2 +
 drivers/gpu/drm/v3d/v3d_gem.c     | 545 +++++++++++++++++++-----------
 drivers/gpu/drm/v3d/v3d_irq.c     |  24 +-
 drivers/gpu/drm/v3d/v3d_regs.h    |  73 ++++
 drivers/gpu/drm/v3d/v3d_sched.c   | 380 ++++++++++++++-------
 drivers/gpu/drm/v3d/v3d_trace.h   |  94 ++++++
 include/drm/drm_gem.h             |   5 +
 include/uapi/drm/v3d_drm.h        |  28 ++
 15 files changed, 1019 insertions(+), 471 deletions(-)

-- 
2.20.1

[PATCH 1/7] drm/v3d: Switch the type of job-> to reduce casting.

[Eric Anholt]

All consumers wanted drm_gem_object * now.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/v3d/v3d_drv.h |  4 ++--
 drivers/gpu/drm/v3d/v3d_gem.c | 42 +++++++++++++----------------------
 2 files changed, 17 insertions(+), 29 deletions(-)

diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
index 89e5b9852c4d..238486925096 100644
--- a/drivers/gpu/drm/v3d/v3d_drv.h
+++ b/drivers/gpu/drm/v3d/v3d_drv.h
@@ -195,7 +195,7 @@ struct v3d_exec_info {
 	struct kref refcount;
 
 	/* This is the array of BOs that were looked up at the start of exec. */
-	struct v3d_bo **bo;
+	struct drm_gem_object **bo;
 	u32 bo_count;
 
 	/* List of overflow BOs used in the job that need to be
@@ -223,7 +223,7 @@ struct v3d_tfu_job {
 	struct kref refcount;
 
 	/* This is the array of BOs that were looked up at the start of exec. */
-	struct v3d_bo *bo[4];
+	struct drm_gem_object *bo[4];
 };
 
 /**
diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
index 93ff8fcbe475..aa0397d12847 100644
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -194,27 +194,17 @@ v3d_invalidate_caches(struct v3d_dev *v3d)
 }
 
 static void
-v3d_attach_object_fences(struct v3d_bo **bos, int bo_count,
+v3d_attach_object_fences(struct drm_gem_object **bos, int bo_count,
 			 struct dma_fence *fence)
 {
 	int i;
 
 	for (i = 0; i < bo_count; i++) {
 		/* XXX: Use shared fences for read-only objects. */
-		reservation_object_add_excl_fence(bos[i]->base.base.resv,
-						  fence);
+		reservation_object_add_excl_fence(bos[i]->resv, fence);
 	}
 }
 
-static void
-v3d_unlock_bo_reservations(struct v3d_bo **bos,
-			   int bo_count,
-			   struct ww_acquire_ctx *acquire_ctx)
-{
-	drm_gem_unlock_reservations((struct drm_gem_object **)bos, bo_count,
-				    acquire_ctx);
-}
-
 /* Takes the reservation lock on all the BOs being referenced, so that
  * at queue submit time we can update the reservations.
  *
@@ -223,14 +213,13 @@ v3d_unlock_bo_reservations(struct v3d_bo **bos,
  * to v3d, so we don't attach dma-buf fences to them.
  */
 static int
-v3d_lock_bo_reservations(struct v3d_bo **bos,
+v3d_lock_bo_reservations(struct drm_gem_object **bos,
 			 int bo_count,
 			 struct ww_acquire_ctx *acquire_ctx)
 {
 	int i, ret;
 
-	ret = drm_gem_lock_reservations((struct drm_gem_object **)bos,
-					bo_count, acquire_ctx);
+	ret = drm_gem_lock_reservations(bos, bo_count, acquire_ctx);
 	if (ret)
 		return ret;
 
@@ -238,11 +227,10 @@ v3d_lock_bo_reservations(struct v3d_bo **bos,
 	 * before we commit the CL to the hardware.
 	 */
 	for (i = 0; i < bo_count; i++) {
-		ret = reservation_object_reserve_shared(bos[i]->base.base.resv,
-							1);
+		ret = reservation_object_reserve_shared(bos[i]->resv, 1);
 		if (ret) {
-			v3d_unlock_bo_reservations(bos, bo_count,
-						   acquire_ctx);
+			drm_gem_unlock_reservations(bos, bo_count,
+						    acquire_ctx);
 			return ret;
 		}
 	}
@@ -319,7 +307,7 @@ v3d_cl_lookup_bos(struct drm_device *dev,
 			goto fail;
 		}
 		drm_gem_object_get(bo);
-		exec->bo[i] = to_v3d_bo(bo);
+		exec->bo[i] = bo;
 	}
 	spin_unlock(&file_priv->table_lock);
 
@@ -347,7 +335,7 @@ v3d_exec_cleanup(struct kref *ref)
 	dma_fence_put(exec->render_done_fence);
 
 	for (i = 0; i < exec->bo_count; i++)
-		drm_gem_object_put_unlocked(&exec->bo[i]->base.base);
+		drm_gem_object_put_unlocked(exec->bo[i]);
 	kvfree(exec->bo);
 
 	list_for_each_entry_safe(bo, save, &exec->unref_list, unref_head) {
@@ -378,7 +366,7 @@ v3d_tfu_job_cleanup(struct kref *ref)
 
 	for (i = 0; i < ARRAY_SIZE(job->bo); i++) {
 		if (job->bo[i])
-			drm_gem_object_put_unlocked(&job->bo[i]->base.base);
+			drm_gem_object_put_unlocked(job->bo[i]);
 	}
 
 	pm_runtime_mark_last_busy(v3d->dev);
@@ -532,7 +520,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	v3d_attach_object_fences(exec->bo, exec->bo_count,
 				 exec->render_done_fence);
 
-	v3d_unlock_bo_reservations(exec->bo, exec->bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx);
 
 	/* Update the return sync object for the */
 	sync_out = drm_syncobj_find(file_priv, args->out_sync);
@@ -547,7 +535,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	v3d_unlock_bo_reservations(exec->bo, exec->bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx);
 fail:
 	v3d_exec_put(exec);
 
@@ -616,7 +604,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 			goto fail;
 		}
 		drm_gem_object_get(bo);
-		job->bo[bo_count] = to_v3d_bo(bo);
+		job->bo[bo_count] = bo;
 	}
 	spin_unlock(&file_priv->table_lock);
 
@@ -639,7 +627,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 
 	v3d_attach_object_fences(job->bo, bo_count, sched_done_fence);
 
-	v3d_unlock_bo_reservations(job->bo, bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx);
 
 	/* Update the return sync object */
 	sync_out = drm_syncobj_find(file_priv, args->out_sync);
@@ -655,7 +643,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	v3d_unlock_bo_reservations(job->bo, bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx);
 fail:
 	v3d_tfu_job_put(job);
 
-- 
2.20.1

[PATCH 2/7] drm/v3d: Refactor job management.

[Eric Anholt]

The CL submission had two jobs embedded in an exec struct.  When I
added TFU support, I had to replicate some of the exec stuff and some
of the job stuff.  As I went to add CSD, it became clear that actually
what was in exec should just be in the two CL jobs, and it would let
us share a lot more code between the 4 queues.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/v3d/v3d_drv.h   |  77 +++----
 drivers/gpu/drm/v3d/v3d_gem.c   | 362 ++++++++++++++++----------------
 drivers/gpu/drm/v3d/v3d_irq.c   |   8 +-
 drivers/gpu/drm/v3d/v3d_sched.c | 259 +++++++++++++----------
 4 files changed, 376 insertions(+), 330 deletions(-)

diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
index 238486925096..aa21a70e6879 100644
--- a/drivers/gpu/drm/v3d/v3d_drv.h
+++ b/drivers/gpu/drm/v3d/v3d_drv.h
@@ -73,8 +73,8 @@ struct v3d_dev {
 
 	struct work_struct overflow_mem_work;
 
-	struct v3d_exec_info *bin_job;
-	struct v3d_exec_info *render_job;
+	struct v3d_bin_job *bin_job;
+	struct v3d_render_job *render_job;
 	struct v3d_tfu_job *tfu_job;
 
 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
@@ -123,7 +123,7 @@ struct v3d_bo {
 	struct drm_mm_node node;
 
 	/* List entry for the BO's position in
-	 * v3d_exec_info->unref_list
+	 * v3d_render_job->unref_list
 	 */
 	struct list_head unref_head;
 };
@@ -163,7 +163,15 @@ to_v3d_fence(struct dma_fence *fence)
 struct v3d_job {
 	struct drm_sched_job base;
 
-	struct v3d_exec_info *exec;
+	struct kref refcount;
+
+	struct v3d_dev *v3d;
+
+	/* This is the array of BOs that were looked up at the start
+	 * of submission.
+	 */
+	struct drm_gem_object **bo;
+	u32 bo_count;
 
 	/* An optional fence userspace can pass in for the job to depend on. */
 	struct dma_fence *in_fence;
@@ -171,59 +179,53 @@ struct v3d_job {
 	/* v3d fence to be signaled by IRQ handler when the job is complete. */
 	struct dma_fence *irq_fence;
 
+	/* scheduler fence for when the job is considered complete and
+	 * the BO reservations can be released.
+	 */
+	struct dma_fence *done_fence;
+
+	/* Callback for the freeing of the job on refcount going to 0. */
+	void (*free)(struct kref *ref);
+};
+
+struct v3d_bin_job {
+	struct v3d_job base;
+
 	/* GPU virtual addresses of the start/end of the CL job. */
 	u32 start, end;
 
 	u32 timedout_ctca, timedout_ctra;
-};
 
-struct v3d_exec_info {
-	struct v3d_dev *v3d;
+	/* Corresponding render job, for attaching our overflow memory. */
+	struct v3d_render_job *render;
+
+	/* Submitted tile memory allocation start/size, tile state. */
+	u32 qma, qms, qts;
+};
 
-	struct v3d_job bin, render;
+struct v3d_render_job {
+	struct v3d_job base;
 
-	/* Fence for when the scheduler considers the binner to be
-	 * done, for render to depend on.
+	/* Optional fence for the binner, to depend on before starting
+	 * our job.
 	 */
 	struct dma_fence *bin_done_fence;
 
-	/* Fence for when the scheduler considers the render to be
-	 * done, for when the BOs reservations should be complete.
-	 */
-	struct dma_fence *render_done_fence;
-
-	struct kref refcount;
+	/* GPU virtual addresses of the start/end of the CL job. */
+	u32 start, end;
 
-	/* This is the array of BOs that were looked up at the start of exec. */
-	struct drm_gem_object **bo;
-	u32 bo_count;
+	u32 timedout_ctca, timedout_ctra;
 
 	/* List of overflow BOs used in the job that need to be
 	 * released once the job is complete.
 	 */
 	struct list_head unref_list;
-
-	/* Submitted tile memory allocation start/size, tile state. */
-	u32 qma, qms, qts;
 };
 
 struct v3d_tfu_job {
-	struct drm_sched_job base;
+	struct v3d_job base;
 
 	struct drm_v3d_submit_tfu args;
-
-	/* An optional fence userspace can pass in for the job to depend on. */
-	struct dma_fence *in_fence;
-
-	/* v3d fence to be signaled by IRQ handler when the job is complete. */
-	struct dma_fence *irq_fence;
-
-	struct v3d_dev *v3d;
-
-	struct kref refcount;
-
-	/* This is the array of BOs that were looked up at the start of exec. */
-	struct drm_gem_object *bo[4];
 };
 
 /**
@@ -289,8 +291,7 @@ int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 			 struct drm_file *file_priv);
 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv);
-void v3d_exec_put(struct v3d_exec_info *exec);
-void v3d_tfu_job_put(struct v3d_tfu_job *exec);
+void v3d_job_put(struct v3d_job *job);
 void v3d_reset(struct v3d_dev *v3d);
 void v3d_invalidate_caches(struct v3d_dev *v3d);
 
diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
index aa0397d12847..dd4679a536c9 100644
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -193,18 +193,6 @@ v3d_invalidate_caches(struct v3d_dev *v3d)
 	v3d_invalidate_slices(v3d, 0);
 }
 
-static void
-v3d_attach_object_fences(struct drm_gem_object **bos, int bo_count,
-			 struct dma_fence *fence)
-{
-	int i;
-
-	for (i = 0; i < bo_count; i++) {
-		/* XXX: Use shared fences for read-only objects. */
-		reservation_object_add_excl_fence(bos[i]->resv, fence);
-	}
-}
-
 /* Takes the reservation lock on all the BOs being referenced, so that
  * at queue submit time we can update the reservations.
  *
@@ -239,11 +227,11 @@ v3d_lock_bo_reservations(struct drm_gem_object **bos,
 }
 
 /**
- * v3d_cl_lookup_bos() - Sets up exec->bo[] with the GEM objects
+ * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
  * referenced by the job.
  * @dev: DRM device
  * @file_priv: DRM file for this fd
- * @exec: V3D job being set up
+ * @job: V3D job being set up
  *
  * The command validator needs to reference BOs by their index within
  * the submitted job's BO list.  This does the validation of the job's
@@ -253,18 +241,19 @@ v3d_lock_bo_reservations(struct drm_gem_object **bos,
  * failure, because that will happen at v3d_exec_cleanup() time.
  */
 static int
-v3d_cl_lookup_bos(struct drm_device *dev,
-		  struct drm_file *file_priv,
-		  struct drm_v3d_submit_cl *args,
-		  struct v3d_exec_info *exec)
+v3d_lookup_bos(struct drm_device *dev,
+	       struct drm_file *file_priv,
+	       struct v3d_job *job,
+	       u64 bo_handles,
+	       u32 bo_count)
 {
 	u32 *handles;
 	int ret = 0;
 	int i;
 
-	exec->bo_count = args->bo_handle_count;
+	job->bo_count = bo_count;
 
-	if (!exec->bo_count) {
+	if (!job->bo_count) {
 		/* See comment on bo_index for why we have to check
 		 * this.
 		 */
@@ -272,15 +261,15 @@ v3d_cl_lookup_bos(struct drm_device *dev,
 		return -EINVAL;
 	}
 
-	exec->bo = kvmalloc_array(exec->bo_count,
-				  sizeof(struct drm_gem_cma_object *),
-				  GFP_KERNEL | __GFP_ZERO);
-	if (!exec->bo) {
+	job->bo = kvmalloc_array(job->bo_count,
+				 sizeof(struct drm_gem_cma_object *),
+				 GFP_KERNEL | __GFP_ZERO);
+	if (!job->bo) {
 		DRM_DEBUG("Failed to allocate validated BO pointers\n");
 		return -ENOMEM;
 	}
 
-	handles = kvmalloc_array(exec->bo_count, sizeof(u32), GFP_KERNEL);
+	handles = kvmalloc_array(job->bo_count, sizeof(u32), GFP_KERNEL);
 	if (!handles) {
 		ret = -ENOMEM;
 		DRM_DEBUG("Failed to allocate incoming GEM handles\n");
@@ -288,15 +277,15 @@ v3d_cl_lookup_bos(struct drm_device *dev,
 	}
 
 	if (copy_from_user(handles,
-			   (void __user *)(uintptr_t)args->bo_handles,
-			   exec->bo_count * sizeof(u32))) {
+			   (void __user *)(uintptr_t)bo_handles,
+			   job->bo_count * sizeof(u32))) {
 		ret = -EFAULT;
 		DRM_DEBUG("Failed to copy in GEM handles\n");
 		goto fail;
 	}
 
 	spin_lock(&file_priv->table_lock);
-	for (i = 0; i < exec->bo_count; i++) {
+	for (i = 0; i < job->bo_count; i++) {
 		struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
 						     handles[i]);
 		if (!bo) {
@@ -307,7 +296,7 @@ v3d_cl_lookup_bos(struct drm_device *dev,
 			goto fail;
 		}
 		drm_gem_object_get(bo);
-		exec->bo[i] = bo;
+		job->bo[i] = bo;
 	}
 	spin_unlock(&file_priv->table_lock);
 
@@ -317,67 +306,44 @@ v3d_cl_lookup_bos(struct drm_device *dev,
 }
 
 static void
-v3d_exec_cleanup(struct kref *ref)
+v3d_job_free(struct kref *ref)
 {
-	struct v3d_exec_info *exec = container_of(ref, struct v3d_exec_info,
-						  refcount);
-	struct v3d_dev *v3d = exec->v3d;
-	unsigned int i;
-	struct v3d_bo *bo, *save;
-
-	dma_fence_put(exec->bin.in_fence);
-	dma_fence_put(exec->render.in_fence);
-
-	dma_fence_put(exec->bin.irq_fence);
-	dma_fence_put(exec->render.irq_fence);
-
-	dma_fence_put(exec->bin_done_fence);
-	dma_fence_put(exec->render_done_fence);
-
-	for (i = 0; i < exec->bo_count; i++)
-		drm_gem_object_put_unlocked(exec->bo[i]);
-	kvfree(exec->bo);
+	struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
+	int i;
 
-	list_for_each_entry_safe(bo, save, &exec->unref_list, unref_head) {
-		drm_gem_object_put_unlocked(&bo->base.base);
+	for (i = 0; i < job->bo_count; i++) {
+		if (job->bo[i])
+			drm_gem_object_put_unlocked(job->bo[i]);
 	}
+	kvfree(job->bo);
 
-	pm_runtime_mark_last_busy(v3d->dev);
-	pm_runtime_put_autosuspend(v3d->dev);
+	dma_fence_put(job->in_fence);
+	dma_fence_put(job->irq_fence);
+	dma_fence_put(job->done_fence);
 
-	kfree(exec);
-}
+	pm_runtime_mark_last_busy(job->v3d->dev);
+	pm_runtime_put_autosuspend(job->v3d->dev);
 
-void v3d_exec_put(struct v3d_exec_info *exec)
-{
-	kref_put(&exec->refcount, v3d_exec_cleanup);
+	kfree(job);
 }
 
 static void
-v3d_tfu_job_cleanup(struct kref *ref)
+v3d_render_job_free(struct kref *ref)
 {
-	struct v3d_tfu_job *job = container_of(ref, struct v3d_tfu_job,
-					       refcount);
-	struct v3d_dev *v3d = job->v3d;
-	unsigned int i;
-
-	dma_fence_put(job->in_fence);
-	dma_fence_put(job->irq_fence);
+	struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
+						  base.refcount);
+	struct v3d_bo *bo, *save;
 
-	for (i = 0; i < ARRAY_SIZE(job->bo); i++) {
-		if (job->bo[i])
-			drm_gem_object_put_unlocked(job->bo[i]);
+	list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
+		drm_gem_object_put_unlocked(&bo->base.base);
 	}
 
-	pm_runtime_mark_last_busy(v3d->dev);
-	pm_runtime_put_autosuspend(v3d->dev);
-
-	kfree(job);
+	v3d_job_free(ref);
 }
 
-void v3d_tfu_job_put(struct v3d_tfu_job *job)
+void v3d_job_put(struct v3d_job *job)
 {
-	kref_put(&job->refcount, v3d_tfu_job_cleanup);
+	kref_put(&job->refcount, job->free);
 }
 
 int
@@ -413,6 +379,77 @@ v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 	return ret;
 }
 
+static int
+v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
+	     struct v3d_job *job, void (*free)(struct kref *ref),
+	     u32 in_sync)
+{
+	int ret;
+
+	job->v3d = v3d;
+	job->free = free;
+
+	ret = pm_runtime_get_sync(v3d->dev);
+	if (ret < 0)
+		return ret;
+
+	ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &job->in_fence);
+	if (ret == -EINVAL) {
+		pm_runtime_put_autosuspend(v3d->dev);
+		return ret;
+	}
+
+	kref_init(&job->refcount);
+
+	return 0;
+}
+
+static int
+v3d_push_job(struct v3d_file_priv *v3d_priv,
+	     struct v3d_job *job, enum v3d_queue queue)
+{
+	int ret;
+
+	ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
+				 v3d_priv);
+	if (ret)
+		return ret;
+
+	job->done_fence = dma_fence_get(&job->base.s_fence->finished);
+
+	/* put by scheduler job completion */
+	kref_get(&job->refcount);
+
+	drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[queue]);
+
+	return 0;
+}
+
+static void
+v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
+					 struct v3d_job *job,
+					 struct ww_acquire_ctx *acquire_ctx,
+					 u32 out_sync)
+{
+	struct drm_syncobj *sync_out;
+	int i;
+
+	for (i = 0; i < job->bo_count; i++) {
+		/* XXX: Use shared fences for read-only objects. */
+		reservation_object_add_excl_fence(job->bo[i]->resv,
+						  job->done_fence);
+	}
+
+	drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
+
+	/* Update the return sync object for the job */
+	sync_out = drm_syncobj_find(file_priv, out_sync);
+	if (sync_out) {
+		drm_syncobj_replace_fence(sync_out, job->done_fence);
+		drm_syncobj_put(sync_out);
+	}
+}
+
 /**
  * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
  * @dev: DRM device
@@ -432,9 +469,9 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	struct v3d_dev *v3d = to_v3d_dev(dev);
 	struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 	struct drm_v3d_submit_cl *args = data;
-	struct v3d_exec_info *exec;
+	struct v3d_bin_job *bin = NULL;
+	struct v3d_render_job *render;
 	struct ww_acquire_ctx acquire_ctx;
-	struct drm_syncobj *sync_out;
 	int ret = 0;
 
 	trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
@@ -444,100 +481,83 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 		return -EINVAL;
 	}
 
-	exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
-	if (!exec)
+	render = kcalloc(1, sizeof(*render), GFP_KERNEL);
+	if (!render)
 		return -ENOMEM;
 
-	ret = pm_runtime_get_sync(v3d->dev);
-	if (ret < 0) {
-		kfree(exec);
+	render->start = args->rcl_start;
+	render->end = args->rcl_end;
+	INIT_LIST_HEAD(&render->unref_list);
+
+	ret = v3d_job_init(v3d, file_priv, &render->base,
+			   v3d_render_job_free, args->in_sync_rcl);
+	if (ret) {
+		kfree(render);
 		return ret;
 	}
 
-	kref_init(&exec->refcount);
+	if (args->bcl_start != args->bcl_end) {
+		bin = kcalloc(1, sizeof(*bin), GFP_KERNEL);
+		if (!bin)
+			return -ENOMEM;
 
-	ret = drm_syncobj_find_fence(file_priv, args->in_sync_bcl,
-				     0, 0, &exec->bin.in_fence);
-	if (ret == -EINVAL)
-		goto fail;
+		ret = v3d_job_init(v3d, file_priv, &bin->base,
+				   v3d_job_free, args->in_sync_bcl);
+		if (ret) {
+			v3d_job_put(&render->base);
+			return ret;
+		}
 
-	ret = drm_syncobj_find_fence(file_priv, args->in_sync_rcl,
-				     0, 0, &exec->render.in_fence);
-	if (ret == -EINVAL)
-		goto fail;
+		bin->start = args->bcl_start;
+		bin->end = args->bcl_end;
+		bin->qma = args->qma;
+		bin->qms = args->qms;
+		bin->qts = args->qts;
+		bin->render = render;
+	}
 
-	exec->qma = args->qma;
-	exec->qms = args->qms;
-	exec->qts = args->qts;
-	exec->bin.exec = exec;
-	exec->bin.start = args->bcl_start;
-	exec->bin.end = args->bcl_end;
-	exec->render.exec = exec;
-	exec->render.start = args->rcl_start;
-	exec->render.end = args->rcl_end;
-	exec->v3d = v3d;
-	INIT_LIST_HEAD(&exec->unref_list);
-
-	ret = v3d_cl_lookup_bos(dev, file_priv, args, exec);
+	ret = v3d_lookup_bos(dev, file_priv, &render->base,
+			     args->bo_handles, args->bo_handle_count);
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(exec->bo, exec->bo_count,
+	ret = v3d_lock_bo_reservations(render->base.bo, render->base.bo_count,
 				       &acquire_ctx);
 	if (ret)
 		goto fail;
 
 	mutex_lock(&v3d->sched_lock);
-	if (exec->bin.start != exec->bin.end) {
-		ret = drm_sched_job_init(&exec->bin.base,
-					 &v3d_priv->sched_entity[V3D_BIN],
-					 v3d_priv);
+	if (bin) {
+		ret = v3d_push_job(v3d_priv, &bin->base, V3D_BIN);
 		if (ret)
 			goto fail_unreserve;
 
-		exec->bin_done_fence =
-			dma_fence_get(&exec->bin.base.s_fence->finished);
-
-		kref_get(&exec->refcount); /* put by scheduler job completion */
-		drm_sched_entity_push_job(&exec->bin.base,
-					  &v3d_priv->sched_entity[V3D_BIN]);
+		render->bin_done_fence = dma_fence_get(bin->base.done_fence);
 	}
 
-	ret = drm_sched_job_init(&exec->render.base,
-				 &v3d_priv->sched_entity[V3D_RENDER],
-				 v3d_priv);
+	ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
 	if (ret)
 		goto fail_unreserve;
-
-	exec->render_done_fence =
-		dma_fence_get(&exec->render.base.s_fence->finished);
-
-	kref_get(&exec->refcount); /* put by scheduler job completion */
-	drm_sched_entity_push_job(&exec->render.base,
-				  &v3d_priv->sched_entity[V3D_RENDER]);
 	mutex_unlock(&v3d->sched_lock);
 
-	v3d_attach_object_fences(exec->bo, exec->bo_count,
-				 exec->render_done_fence);
-
-	drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx);
+	v3d_attach_fences_and_unlock_reservation(file_priv,
+						 &render->base, &acquire_ctx,
+						 args->out_sync);
 
-	/* Update the return sync object for the */
-	sync_out = drm_syncobj_find(file_priv, args->out_sync);
-	if (sync_out) {
-		drm_syncobj_replace_fence(sync_out, exec->render_done_fence);
-		drm_syncobj_put(sync_out);
-	}
-
-	v3d_exec_put(exec);
+	if (bin)
+		v3d_job_put(&bin->base);
+	v3d_job_put(&render->base);
 
 	return 0;
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	drm_gem_unlock_reservations(exec->bo, exec->bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(render->base.bo,
+				    render->base.bo_count, &acquire_ctx);
 fail:
-	v3d_exec_put(exec);
+	if (bin)
+		v3d_job_put(&bin->base);
+	v3d_job_put(&render->base);
 
 	return ret;
 }
@@ -560,10 +580,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 	struct drm_v3d_submit_tfu *args = data;
 	struct v3d_tfu_job *job;
 	struct ww_acquire_ctx acquire_ctx;
-	struct drm_syncobj *sync_out;
-	struct dma_fence *sched_done_fence;
 	int ret = 0;
-	int bo_count;
 
 	trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
 
@@ -571,81 +588,66 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 	if (!job)
 		return -ENOMEM;
 
-	ret = pm_runtime_get_sync(v3d->dev);
-	if (ret < 0) {
+	ret = v3d_job_init(v3d, file_priv, &job->base,
+			   v3d_job_free, args->in_sync);
+	if (ret) {
 		kfree(job);
 		return ret;
 	}
 
-	kref_init(&job->refcount);
-
-	ret = drm_syncobj_find_fence(file_priv, args->in_sync,
-				     0, 0, &job->in_fence);
-	if (ret == -EINVAL)
-		goto fail;
-
+	job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
+			       sizeof(*job->base.bo), GFP_KERNEL);
 	job->args = *args;
-	job->v3d = v3d;
 
 	spin_lock(&file_priv->table_lock);
-	for (bo_count = 0; bo_count < ARRAY_SIZE(job->bo); bo_count++) {
+	for (job->base.bo_count = 0;
+	     job->base.bo_count < ARRAY_SIZE(args->bo_handles);
+	     job->base.bo_count++) {
 		struct drm_gem_object *bo;
 
-		if (!args->bo_handles[bo_count])
+		if (!args->bo_handles[job->base.bo_count])
 			break;
 
 		bo = idr_find(&file_priv->object_idr,
-			      args->bo_handles[bo_count]);
+			      args->bo_handles[job->base.bo_count]);
 		if (!bo) {
 			DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
-				  bo_count, args->bo_handles[bo_count]);
+				  job->base.bo_count,
+				  args->bo_handles[job->base.bo_count]);
 			ret = -ENOENT;
 			spin_unlock(&file_priv->table_lock);
 			goto fail;
 		}
 		drm_gem_object_get(bo);
-		job->bo[bo_count] = bo;
+		job->base.bo[job->base.bo_count] = bo;
 	}
 	spin_unlock(&file_priv->table_lock);
 
-	ret = v3d_lock_bo_reservations(job->bo, bo_count, &acquire_ctx);
+	ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count,
+				       &acquire_ctx);
 	if (ret)
 		goto fail;
 
 	mutex_lock(&v3d->sched_lock);
-	ret = drm_sched_job_init(&job->base,
-				 &v3d_priv->sched_entity[V3D_TFU],
-				 v3d_priv);
+	ret = v3d_push_job(v3d_priv, &job->base, V3D_TFU);
 	if (ret)
 		goto fail_unreserve;
-
-	sched_done_fence = dma_fence_get(&job->base.s_fence->finished);
-
-	kref_get(&job->refcount); /* put by scheduler job completion */
-	drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[V3D_TFU]);
 	mutex_unlock(&v3d->sched_lock);
 
-	v3d_attach_object_fences(job->bo, bo_count, sched_done_fence);
-
-	drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx);
-
-	/* Update the return sync object */
-	sync_out = drm_syncobj_find(file_priv, args->out_sync);
-	if (sync_out) {
-		drm_syncobj_replace_fence(sync_out, sched_done_fence);
-		drm_syncobj_put(sync_out);
-	}
-	dma_fence_put(sched_done_fence);
+	v3d_attach_fences_and_unlock_reservation(file_priv,
+						 &job->base, &acquire_ctx,
+						 args->out_sync);
 
-	v3d_tfu_job_put(job);
+	v3d_job_put(&job->base);
 
 	return 0;
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	drm_gem_unlock_reservations(job->bo, bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(job->base.bo, job->base.bo_count,
+				    &acquire_ctx);
 fail:
-	v3d_tfu_job_put(job);
+	v3d_job_put(&job->base);
 
 	return ret;
 }
@@ -703,7 +705,7 @@ v3d_gem_destroy(struct drm_device *dev)
 
 	v3d_sched_fini(v3d);
 
-	/* Waiting for exec to finish would need to be done before
+	/* Waiting for jobs to finish would need to be done before
 	 * unregistering V3D.
 	 */
 	WARN_ON(v3d->bin_job);
diff --git a/drivers/gpu/drm/v3d/v3d_irq.c b/drivers/gpu/drm/v3d/v3d_irq.c
index aa0a180ae700..ce373ffd6380 100644
--- a/drivers/gpu/drm/v3d/v3d_irq.c
+++ b/drivers/gpu/drm/v3d/v3d_irq.c
@@ -62,7 +62,7 @@ v3d_overflow_mem_work(struct work_struct *work)
 	}
 
 	drm_gem_object_get(obj);
-	list_add_tail(&bo->unref_head, &v3d->bin_job->unref_list);
+	list_add_tail(&bo->unref_head, &v3d->bin_job->render->unref_list);
 	spin_unlock_irqrestore(&v3d->job_lock, irqflags);
 
 	V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << PAGE_SHIFT);
@@ -96,7 +96,7 @@ v3d_irq(int irq, void *arg)
 
 	if (intsts & V3D_INT_FLDONE) {
 		struct v3d_fence *fence =
-			to_v3d_fence(v3d->bin_job->bin.irq_fence);
+			to_v3d_fence(v3d->bin_job->base.irq_fence);
 
 		trace_v3d_bcl_irq(&v3d->drm, fence->seqno);
 		dma_fence_signal(&fence->base);
@@ -105,7 +105,7 @@ v3d_irq(int irq, void *arg)
 
 	if (intsts & V3D_INT_FRDONE) {
 		struct v3d_fence *fence =
-			to_v3d_fence(v3d->render_job->render.irq_fence);
+			to_v3d_fence(v3d->render_job->base.irq_fence);
 
 		trace_v3d_rcl_irq(&v3d->drm, fence->seqno);
 		dma_fence_signal(&fence->base);
@@ -141,7 +141,7 @@ v3d_hub_irq(int irq, void *arg)
 
 	if (intsts & V3D_HUB_INT_TFUC) {
 		struct v3d_fence *fence =
-			to_v3d_fence(v3d->tfu_job->irq_fence);
+			to_v3d_fence(v3d->tfu_job->base.irq_fence);
 
 		trace_v3d_tfu_irq(&v3d->drm, fence->seqno);
 		dma_fence_signal(&fence->base);
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
index e740f3b99aa5..884f00f5edf4 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -30,43 +30,44 @@ to_v3d_job(struct drm_sched_job *sched_job)
 	return container_of(sched_job, struct v3d_job, base);
 }
 
-static struct v3d_tfu_job *
-to_tfu_job(struct drm_sched_job *sched_job)
+static struct v3d_bin_job *
+to_bin_job(struct drm_sched_job *sched_job)
 {
-	return container_of(sched_job, struct v3d_tfu_job, base);
+	return container_of(sched_job, struct v3d_bin_job, base.base);
 }
 
-static void
-v3d_job_free(struct drm_sched_job *sched_job)
+static struct v3d_render_job *
+to_render_job(struct drm_sched_job *sched_job)
 {
-	struct v3d_job *job = to_v3d_job(sched_job);
-
-	drm_sched_job_cleanup(sched_job);
+	return container_of(sched_job, struct v3d_render_job, base.base);
+}
 
-	v3d_exec_put(job->exec);
+static struct v3d_tfu_job *
+to_tfu_job(struct drm_sched_job *sched_job)
+{
+	return container_of(sched_job, struct v3d_tfu_job, base.base);
 }
 
 static void
-v3d_tfu_job_free(struct drm_sched_job *sched_job)
+v3d_job_free(struct drm_sched_job *sched_job)
 {
-	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+	struct v3d_job *job = to_v3d_job(sched_job);
 
 	drm_sched_job_cleanup(sched_job);
-
-	v3d_tfu_job_put(job);
+	v3d_job_put(job);
 }
 
 /**
- * Returns the fences that the bin or render job depends on, one by one.
- * v3d_job_run() won't be called until all of them have been signaled.
+ * Returns the fences that the job depends on, one by one.
+ *
+ * If placed in the scheduler's .dependency method, the corresponding
+ * .run_job won't be called until all of them have been signaled.
  */
 static struct dma_fence *
 v3d_job_dependency(struct drm_sched_job *sched_job,
 		   struct drm_sched_entity *s_entity)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
-	struct v3d_exec_info *exec = job->exec;
-	enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
 	struct dma_fence *fence;
 
 	fence = job->in_fence;
@@ -75,113 +76,132 @@ v3d_job_dependency(struct drm_sched_job *sched_job,
 		return fence;
 	}
 
-	if (q == V3D_RENDER) {
-		/* If we had a bin job, the render job definitely depends on
-		 * it. We first have to wait for bin to be scheduled, so that
-		 * its done_fence is created.
-		 */
-		fence = exec->bin_done_fence;
-		if (fence) {
-			exec->bin_done_fence = NULL;
-			return fence;
-		}
-	}
-
 	/* XXX: Wait on a fence for switching the GMP if necessary,
 	 * and then do so.
 	 */
 
-	return fence;
+	return NULL;
 }
 
 /**
- * Returns the fences that the TFU job depends on, one by one.
- * v3d_tfu_job_run() won't be called until all of them have been
- * signaled.
+ * Returns the fences that the render job depends on, one by one.
+ * v3d_job_run() won't be called until all of them have been signaled.
  */
 static struct dma_fence *
-v3d_tfu_job_dependency(struct drm_sched_job *sched_job,
-		       struct drm_sched_entity *s_entity)
+v3d_render_job_dependency(struct drm_sched_job *sched_job,
+			  struct drm_sched_entity *s_entity)
 {
-	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+	struct v3d_render_job *job = to_render_job(sched_job);
 	struct dma_fence *fence;
 
-	fence = job->in_fence;
+	fence = v3d_job_dependency(sched_job, s_entity);
+	if (fence)
+		return fence;
+
+	/* If we had a bin job, the render job definitely depends on
+	 * it. We first have to wait for bin to be scheduled, so that
+	 * its done_fence is created.
+	 */
+	fence = job->bin_done_fence;
 	if (fence) {
-		job->in_fence = NULL;
+		job->bin_done_fence = NULL;
 		return fence;
 	}
 
-	return NULL;
+	return fence;
 }
 
-static struct dma_fence *v3d_job_run(struct drm_sched_job *sched_job)
+static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job)
 {
-	struct v3d_job *job = to_v3d_job(sched_job);
-	struct v3d_exec_info *exec = job->exec;
-	enum v3d_queue q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
-	struct v3d_dev *v3d = exec->v3d;
+	struct v3d_bin_job *job = to_bin_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
 	struct drm_device *dev = &v3d->drm;
 	struct dma_fence *fence;
 	unsigned long irqflags;
 
-	if (unlikely(job->base.s_fence->finished.error))
+	if (unlikely(job->base.base.s_fence->finished.error))
 		return NULL;
 
 	/* Lock required around bin_job update vs
 	 * v3d_overflow_mem_work().
 	 */
 	spin_lock_irqsave(&v3d->job_lock, irqflags);
-	if (q == V3D_BIN) {
-		v3d->bin_job = job->exec;
-
-		/* Clear out the overflow allocation, so we don't
-		 * reuse the overflow attached to a previous job.
-		 */
-		V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
-	} else {
-		v3d->render_job = job->exec;
-	}
+	v3d->bin_job = job;
+	/* Clear out the overflow allocation, so we don't
+	 * reuse the overflow attached to a previous job.
+	 */
+	V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
 	spin_unlock_irqrestore(&v3d->job_lock, irqflags);
 
-	/* Can we avoid this flush when q==RENDER?  We need to be
-	 * careful of scheduling, though -- imagine job0 rendering to
-	 * texture and job1 reading, and them being executed as bin0,
-	 * bin1, render0, render1, so that render1's flush at bin time
+	v3d_invalidate_caches(v3d);
+
+	fence = v3d_fence_create(v3d, V3D_BIN);
+	if (IS_ERR(fence))
+		return NULL;
+
+	if (job->base.irq_fence)
+		dma_fence_put(job->base.irq_fence);
+	job->base.irq_fence = dma_fence_get(fence);
+
+	trace_v3d_submit_cl(dev, false, to_v3d_fence(fence)->seqno,
+			    job->start, job->end);
+
+	/* Set the current and end address of the control list.
+	 * Writing the end register is what starts the job.
+	 */
+	if (job->qma) {
+		V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, job->qma);
+		V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, job->qms);
+	}
+	if (job->qts) {
+		V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
+			       V3D_CLE_CT0QTS_ENABLE |
+			       job->qts);
+	}
+	V3D_CORE_WRITE(0, V3D_CLE_CT0QBA, job->start);
+	V3D_CORE_WRITE(0, V3D_CLE_CT0QEA, job->end);
+
+	return fence;
+}
+
+static struct dma_fence *v3d_render_job_run(struct drm_sched_job *sched_job)
+{
+	struct v3d_render_job *job = to_render_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
+	struct drm_device *dev = &v3d->drm;
+	struct dma_fence *fence;
+
+	if (unlikely(job->base.base.s_fence->finished.error))
+		return NULL;
+
+	v3d->render_job = job;
+
+	/* Can we avoid this flush?  We need to be careful of
+	 * scheduling, though -- imagine job0 rendering to texture and
+	 * job1 reading, and them being executed as bin0, bin1,
+	 * render0, render1, so that render1's flush at bin time
 	 * wasn't enough.
 	 */
 	v3d_invalidate_caches(v3d);
 
-	fence = v3d_fence_create(v3d, q);
+	fence = v3d_fence_create(v3d, V3D_RENDER);
 	if (IS_ERR(fence))
 		return NULL;
 
-	if (job->irq_fence)
-		dma_fence_put(job->irq_fence);
-	job->irq_fence = dma_fence_get(fence);
+	if (job->base.irq_fence)
+		dma_fence_put(job->base.irq_fence);
+	job->base.irq_fence = dma_fence_get(fence);
 
-	trace_v3d_submit_cl(dev, q == V3D_RENDER, to_v3d_fence(fence)->seqno,
+	trace_v3d_submit_cl(dev, true, to_v3d_fence(fence)->seqno,
 			    job->start, job->end);
 
-	if (q == V3D_BIN) {
-		if (exec->qma) {
-			V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, exec->qma);
-			V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, exec->qms);
-		}
-		if (exec->qts) {
-			V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
-				       V3D_CLE_CT0QTS_ENABLE |
-				       exec->qts);
-		}
-	} else {
-		/* XXX: Set the QCFG */
-	}
+	/* XXX: Set the QCFG */
 
 	/* Set the current and end address of the control list.
 	 * Writing the end register is what starts the job.
 	 */
-	V3D_CORE_WRITE(0, V3D_CLE_CTNQBA(q), job->start);
-	V3D_CORE_WRITE(0, V3D_CLE_CTNQEA(q), job->end);
+	V3D_CORE_WRITE(0, V3D_CLE_CT1QBA, job->start);
+	V3D_CORE_WRITE(0, V3D_CLE_CT1QEA, job->end);
 
 	return fence;
 }
@@ -190,7 +210,7 @@ static struct dma_fence *
 v3d_tfu_job_run(struct drm_sched_job *sched_job)
 {
 	struct v3d_tfu_job *job = to_tfu_job(sched_job);
-	struct v3d_dev *v3d = job->v3d;
+	struct v3d_dev *v3d = job->base.v3d;
 	struct drm_device *dev = &v3d->drm;
 	struct dma_fence *fence;
 
@@ -199,9 +219,9 @@ v3d_tfu_job_run(struct drm_sched_job *sched_job)
 		return NULL;
 
 	v3d->tfu_job = job;
-	if (job->irq_fence)
-		dma_fence_put(job->irq_fence);
-	job->irq_fence = dma_fence_get(fence);
+	if (job->base.irq_fence)
+		dma_fence_put(job->base.irq_fence);
+	job->base.irq_fence = dma_fence_get(fence);
 
 	trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno);
 
@@ -251,51 +271,74 @@ v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
 	mutex_unlock(&v3d->reset_lock);
 }
 
+/* If the current address or return address have changed, then the GPU
+ * has probably made progress and we should delay the reset.  This
+ * could fail if the GPU got in an infinite loop in the CL, but that
+ * is pretty unlikely outside of an i-g-t testcase.
+ */
 static void
-v3d_job_timedout(struct drm_sched_job *sched_job)
+v3d_cl_job_timedout(struct drm_sched_job *sched_job, enum v3d_queue q,
+		      u32 *timedout_ctca, u32 *timedout_ctra)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
-	struct v3d_exec_info *exec = job->exec;
-	struct v3d_dev *v3d = exec->v3d;
-	enum v3d_queue job_q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
-	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(job_q));
-	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(job_q));
-
-	/* If the current address or return address have changed, then
-	 * the GPU has probably made progress and we should delay the
-	 * reset.  This could fail if the GPU got in an infinite loop
-	 * in the CL, but that is pretty unlikely outside of an i-g-t
-	 * testcase.
-	 */
-	if (job->timedout_ctca != ctca || job->timedout_ctra != ctra) {
-		job->timedout_ctca = ctca;
-		job->timedout_ctra = ctra;
+	struct v3d_dev *v3d = job->v3d;
+	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(q));
+	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(q));
+
+	if (*timedout_ctca != ctca || *timedout_ctra != ctra) {
+		*timedout_ctca = ctca;
+		*timedout_ctra = ctra;
 		return;
 	}
 
 	v3d_gpu_reset_for_timeout(v3d, sched_job);
 }
 
+static void
+v3d_bin_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_bin_job *job = to_bin_job(sched_job);
+
+	v3d_cl_job_timedout(sched_job, V3D_BIN,
+			    &job->timedout_ctca, &job->timedout_ctra);
+}
+
+static void
+v3d_render_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_render_job *job = to_render_job(sched_job);
+
+	v3d_cl_job_timedout(sched_job, V3D_RENDER,
+			    &job->timedout_ctca, &job->timedout_ctra);
+}
+
 static void
 v3d_tfu_job_timedout(struct drm_sched_job *sched_job)
 {
-	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+	struct v3d_job *job = to_v3d_job(sched_job);
 
 	v3d_gpu_reset_for_timeout(job->v3d, sched_job);
 }
 
-static const struct drm_sched_backend_ops v3d_sched_ops = {
+static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
 	.dependency = v3d_job_dependency,
-	.run_job = v3d_job_run,
-	.timedout_job = v3d_job_timedout,
-	.free_job = v3d_job_free
+	.run_job = v3d_bin_job_run,
+	.timedout_job = v3d_bin_job_timedout,
+	.free_job = v3d_job_free,
+};
+
+static const struct drm_sched_backend_ops v3d_render_sched_ops = {
+	.dependency = v3d_render_job_dependency,
+	.run_job = v3d_render_job_run,
+	.timedout_job = v3d_render_job_timedout,
+	.free_job = v3d_job_free,
 };
 
 static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
-	.dependency = v3d_tfu_job_dependency,
+	.dependency = v3d_job_dependency,
 	.run_job = v3d_tfu_job_run,
 	.timedout_job = v3d_tfu_job_timedout,
-	.free_job = v3d_tfu_job_free
+	.free_job = v3d_job_free,
 };
 
 int
@@ -307,7 +350,7 @@ v3d_sched_init(struct v3d_dev *v3d)
 	int ret;
 
 	ret = drm_sched_init(&v3d->queue[V3D_BIN].sched,
-			     &v3d_sched_ops,
+			     &v3d_bin_sched_ops,
 			     hw_jobs_limit, job_hang_limit,
 			     msecs_to_jiffies(hang_limit_ms),
 			     "v3d_bin");
@@ -317,7 +360,7 @@ v3d_sched_init(struct v3d_dev *v3d)
 	}
 
 	ret = drm_sched_init(&v3d->queue[V3D_RENDER].sched,
-			     &v3d_sched_ops,
+			     &v3d_render_sched_ops,
 			     hw_jobs_limit, job_hang_limit,
 			     msecs_to_jiffies(hang_limit_ms),
 			     "v3d_render");
-- 
2.20.1

[PATCH 3/7] drm/v3d: Add support for compute shader dispatch.

[Eric Anholt]

The compute shader dispatch interface is pretty simple -- just pass in
the regs that userspace has passed us, with no CLs to run.  However,
with no CL to run it means that we need to do manual cache flushing of
the L2 after the HW execution completes (for SSBO, atomic, and
image_load_store writes that are the output of compute shaders).

This doesn't yet expose the L2 cache's ability to have a region of the
address space not write back to memory (which could be used for
shared_var storage).

So far, the Mesa side has been tested on V3D v4.2 simpenrose (passing
the ES31 tests), and on the kernel side on 7278 (failing atomic
compswap tests in a way that doesn't reproduce on simpenrose).

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/v3d/v3d_debugfs.c |  22 +++++
 drivers/gpu/drm/v3d/v3d_drv.c     |  10 +-
 drivers/gpu/drm/v3d/v3d_drv.h     |  28 +++++-
 drivers/gpu/drm/v3d/v3d_fence.c   |   2 +
 drivers/gpu/drm/v3d/v3d_gem.c     | 155 +++++++++++++++++++++++++++++-
 drivers/gpu/drm/v3d/v3d_irq.c     |  16 ++-
 drivers/gpu/drm/v3d/v3d_regs.h    |  73 ++++++++++++++
 drivers/gpu/drm/v3d/v3d_sched.c   | 121 +++++++++++++++++++++--
 drivers/gpu/drm/v3d/v3d_trace.h   |  94 ++++++++++++++++++
 include/uapi/drm/v3d_drm.h        |  28 ++++++
 10 files changed, 530 insertions(+), 19 deletions(-)

diff --git a/drivers/gpu/drm/v3d/v3d_debugfs.c b/drivers/gpu/drm/v3d/v3d_debugfs.c
index a24af2d2f574..a2dc4262955e 100644
--- a/drivers/gpu/drm/v3d/v3d_debugfs.c
+++ b/drivers/gpu/drm/v3d/v3d_debugfs.c
@@ -58,6 +58,17 @@ static const struct v3d_reg_def v3d_core_reg_defs[] = {
 	REGDEF(V3D_GMP_VIO_ADDR),
 };
 
+static const struct v3d_reg_def v3d_csd_reg_defs[] = {
+	REGDEF(V3D_CSD_STATUS),
+	REGDEF(V3D_CSD_CURRENT_CFG0),
+	REGDEF(V3D_CSD_CURRENT_CFG1),
+	REGDEF(V3D_CSD_CURRENT_CFG2),
+	REGDEF(V3D_CSD_CURRENT_CFG3),
+	REGDEF(V3D_CSD_CURRENT_CFG4),
+	REGDEF(V3D_CSD_CURRENT_CFG5),
+	REGDEF(V3D_CSD_CURRENT_CFG6),
+};
+
 static int v3d_v3d_debugfs_regs(struct seq_file *m, void *unused)
 {
 	struct drm_info_node *node = (struct drm_info_node *)m->private;
@@ -89,6 +100,17 @@ static int v3d_v3d_debugfs_regs(struct seq_file *m, void *unused)
 				   V3D_CORE_READ(core,
 						 v3d_core_reg_defs[i].reg));
 		}
+
+		if (v3d_has_csd(v3d)) {
+			for (i = 0; i < ARRAY_SIZE(v3d_csd_reg_defs); i++) {
+				seq_printf(m, "core %d %s (0x%04x): 0x%08x\n",
+					   core,
+					   v3d_csd_reg_defs[i].name,
+					   v3d_csd_reg_defs[i].reg,
+					   V3D_CORE_READ(core,
+							 v3d_csd_reg_defs[i].reg));
+			}
+		}
 	}
 
 	return 0;
diff --git a/drivers/gpu/drm/v3d/v3d_drv.c b/drivers/gpu/drm/v3d/v3d_drv.c
index 4afb63cc4f57..07dc47d4cb09 100644
--- a/drivers/gpu/drm/v3d/v3d_drv.c
+++ b/drivers/gpu/drm/v3d/v3d_drv.c
@@ -7,9 +7,9 @@
  * This driver supports the Broadcom V3D 3.3 and 4.1 OpenGL ES GPUs.
  * For V3D 2.x support, see the VC4 driver.
  *
- * Currently only single-core rendering using the binner and renderer,
- * along with TFU (texture formatting unit) rendering is supported.
- * V3D 4.x's CSD (compute shader dispatch) is not yet supported.
+ * The V3D GPU includes a tiled render (composed of a bin and render
+ * pipelines), the TFU (texture formatting unit), and the CSD (compute
+ * shader dispatch).
  */
 
 #include <linux/clk.h>
@@ -193,6 +193,9 @@ static int v3d_get_param_ioctl(struct drm_device *dev, void *data,
 	case DRM_V3D_PARAM_SUPPORTS_TFU:
 		args->value = 1;
 		return 0;
+	case DRM_V3D_PARAM_SUPPORTS_CSD:
+		args->value = v3d_has_csd(v3d);
+		return 0;
 	default:
 		DRM_DEBUG("Unknown parameter %d\n", args->param);
 		return -EINVAL;
@@ -252,6 +255,7 @@ static const struct drm_ioctl_desc v3d_drm_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(V3D_GET_PARAM, v3d_get_param_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(V3D_GET_BO_OFFSET, v3d_get_bo_offset_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_TFU, v3d_submit_tfu_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
+	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CSD, v3d_submit_csd_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
 };
 
 static int v3d_dumb_create(struct drm_file *file_priv,
diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
index aa21a70e6879..c4ef640effc2 100644
--- a/drivers/gpu/drm/v3d/v3d_drv.h
+++ b/drivers/gpu/drm/v3d/v3d_drv.h
@@ -17,9 +17,11 @@ enum v3d_queue {
 	V3D_BIN,
 	V3D_RENDER,
 	V3D_TFU,
+	V3D_CSD,
+	V3D_CACHE_CLEAN,
 };
 
-#define V3D_MAX_QUEUES (V3D_TFU + 1)
+#define V3D_MAX_QUEUES (V3D_CACHE_CLEAN + 1)
 
 struct v3d_queue_state {
 	struct drm_gpu_scheduler sched;
@@ -76,6 +78,7 @@ struct v3d_dev {
 	struct v3d_bin_job *bin_job;
 	struct v3d_render_job *render_job;
 	struct v3d_tfu_job *tfu_job;
+	struct v3d_csd_job *csd_job;
 
 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
 
@@ -98,6 +101,12 @@ struct v3d_dev {
 	 */
 	struct mutex sched_lock;
 
+	/* Lock taken during a cache clean and when initiating an L2
+	 * flush, to keep L2 flushes from interfering with the
+	 * synchronous L2 cleans.
+	 */
+	struct mutex cache_clean_lock;
+
 	struct {
 		u32 num_allocated;
 		u32 pages_allocated;
@@ -110,6 +119,12 @@ to_v3d_dev(struct drm_device *dev)
 	return (struct v3d_dev *)dev->dev_private;
 }
 
+static inline bool
+v3d_has_csd(struct v3d_dev *v3d)
+{
+	return v3d->ver >= 41;
+}
+
 /* The per-fd struct, which tracks the MMU mappings. */
 struct v3d_file_priv {
 	struct v3d_dev *v3d;
@@ -228,6 +243,14 @@ struct v3d_tfu_job {
 	struct drm_v3d_submit_tfu args;
 };
 
+struct v3d_csd_job {
+	struct v3d_job base;
+
+	u32 timedout_batches;
+
+	struct drm_v3d_submit_csd args;
+};
+
 /**
  * _wait_for - magic (register) wait macro
  *
@@ -289,11 +312,14 @@ int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *file_priv);
 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 			 struct drm_file *file_priv);
+int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
+			 struct drm_file *file_priv);
 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv);
 void v3d_job_put(struct v3d_job *job);
 void v3d_reset(struct v3d_dev *v3d);
 void v3d_invalidate_caches(struct v3d_dev *v3d);
+void v3d_clean_caches(struct v3d_dev *v3d);
 
 /* v3d_irq.c */
 int v3d_irq_init(struct v3d_dev *v3d);
diff --git a/drivers/gpu/drm/v3d/v3d_fence.c b/drivers/gpu/drm/v3d/v3d_fence.c
index b0a2a1ae2eb1..89840ed212c0 100644
--- a/drivers/gpu/drm/v3d/v3d_fence.c
+++ b/drivers/gpu/drm/v3d/v3d_fence.c
@@ -36,6 +36,8 @@ static const char *v3d_fence_get_timeline_name(struct dma_fence *fence)
 		return "v3d-render";
 	case V3D_TFU:
 		return "v3d-tfu";
+	case V3D_CSD:
+		return "v3d-csd";
 	default:
 		return NULL;
 	}
diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
index dd4679a536c9..b78ffc526885 100644
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -162,10 +162,52 @@ v3d_flush_l2t(struct v3d_dev *v3d, int core)
 	/* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
 	 * need to wait for completion before dispatching the job --
 	 * L2T accesses will be stalled until the flush has completed.
+	 * However, we do need to make sure we don't try to trigger a
+	 * new flush while the L2_CLEAN queue is trying to
+	 * synchronously clean after a job.
 	 */
+	mutex_lock(&v3d->cache_clean_lock);
 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 		       V3D_L2TCACTL_L2TFLS |
 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
+	mutex_unlock(&v3d->cache_clean_lock);
+}
+
+/* Cleans texture L1 and L2 cachelines (writing back dirty data).
+ *
+ * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
+ * executed, we need to make sure that the clean is done before
+ * signaling job completion.  So, we synchronously wait before
+ * returning, and we make sure that L2 invalidates don't happen in the
+ * meantime to confuse our are-we-done checks.
+ */
+void
+v3d_clean_caches(struct v3d_dev *v3d)
+{
+	struct drm_device *dev = &v3d->drm;
+	int core = 0;
+
+	trace_v3d_cache_clean_begin(dev);
+
+	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
+	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+		       V3D_L2TCACTL_L2TFLS), 100)) {
+		DRM_ERROR("Timeout waiting for L1T write combiner flush\n");
+	}
+
+	mutex_lock(&v3d->cache_clean_lock);
+	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
+		       V3D_L2TCACTL_L2TFLS |
+		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
+
+	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+		       V3D_L2TCACTL_L2TFLS), 100)) {
+		DRM_ERROR("Timeout waiting for L2T clean\n");
+	}
+
+	mutex_unlock(&v3d->cache_clean_lock);
+
+	trace_v3d_cache_clean_end(dev);
 }
 
 /* Invalidates the slice caches.  These are read-only caches. */
@@ -429,7 +471,8 @@ static void
 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
 					 struct v3d_job *job,
 					 struct ww_acquire_ctx *acquire_ctx,
-					 u32 out_sync)
+					 u32 out_sync,
+					 struct dma_fence *done_fence)
 {
 	struct drm_syncobj *sync_out;
 	int i;
@@ -445,7 +488,7 @@ v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
 	/* Update the return sync object for the job */
 	sync_out = drm_syncobj_find(file_priv, out_sync);
 	if (sync_out) {
-		drm_syncobj_replace_fence(sync_out, job->done_fence);
+		drm_syncobj_replace_fence(sync_out, done_fence);
 		drm_syncobj_put(sync_out);
 	}
 }
@@ -541,8 +584,10 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	mutex_unlock(&v3d->sched_lock);
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
-						 &render->base, &acquire_ctx,
-						 args->out_sync);
+						 &render->base,
+						 &acquire_ctx,
+						 args->out_sync,
+						 render->base.done_fence);
 
 	if (bin)
 		v3d_job_put(&bin->base);
@@ -636,9 +681,107 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
 						 &job->base, &acquire_ctx,
-						 args->out_sync);
+						 args->out_sync,
+						 job->base.done_fence);
+
+	v3d_job_put(&job->base);
+
+	return 0;
+
+fail_unreserve:
+	mutex_unlock(&v3d->sched_lock);
+	drm_gem_unlock_reservations(job->base.bo, job->base.bo_count,
+				    &acquire_ctx);
+fail:
+	v3d_job_put(&job->base);
+
+	return ret;
+}
+
+/**
+ * v3d_submit_csd_ioctl() - Submits a CSD (texture formatting) job to the V3D.
+ * @dev: DRM device
+ * @data: ioctl argument
+ * @file_priv: DRM file for this fd
+ *
+ * Userspace provides the register setup for the CSD, which we don't
+ * need to validate since the CSD is behind the MMU.
+ */
+int
+v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
+		     struct drm_file *file_priv)
+{
+	struct v3d_dev *v3d = to_v3d_dev(dev);
+	struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
+	struct drm_v3d_submit_csd *args = data;
+	struct v3d_csd_job *job;
+	struct v3d_job *clean_job;
+	struct ww_acquire_ctx acquire_ctx;
+	int ret;
+
+	trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
+
+	if (!v3d_has_csd(v3d)) {
+		DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
+		return -EINVAL;
+	}
+
+	job = kcalloc(1, sizeof(*job), GFP_KERNEL);
+	if (!job)
+		return -ENOMEM;
+
+	ret = v3d_job_init(v3d, file_priv, &job->base,
+			   v3d_job_free, args->in_sync);
+	if (ret) {
+		kfree(job);
+		return ret;
+	}
+
+	clean_job = kcalloc(1, sizeof(*job), GFP_KERNEL);
+	if (!clean_job) {
+		v3d_job_put(&job->base);
+		kfree(job);
+		return -ENOMEM;
+	}
+
+	ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
+	if (ret) {
+		v3d_job_put(&job->base);
+		kfree(clean_job);
+		return ret;
+	}
+
+	job->args = *args;
+
+	ret = v3d_lookup_bos(dev, file_priv, &job->base,
+			     args->bo_handles, args->bo_handle_count);
+	if (ret)
+		goto fail;
+
+	ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count,
+				       &acquire_ctx);
+	if (ret)
+		goto fail;
+
+	mutex_lock(&v3d->sched_lock);
+	ret = v3d_push_job(v3d_priv, &job->base, V3D_CSD);
+	if (ret)
+		goto fail_unreserve;
+
+	clean_job->in_fence = dma_fence_get(job->base.done_fence);
+	ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
+	if (ret)
+		goto fail_unreserve;
+	mutex_unlock(&v3d->sched_lock);
+
+	v3d_attach_fences_and_unlock_reservation(file_priv,
+						 &job->base,
+						 &acquire_ctx,
+						 args->out_sync,
+						 clean_job->done_fence);
 
 	v3d_job_put(&job->base);
+	v3d_job_put(clean_job);
 
 	return 0;
 
@@ -648,6 +791,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 				    &acquire_ctx);
 fail:
 	v3d_job_put(&job->base);
+	v3d_job_put(clean_job);
 
 	return ret;
 }
@@ -667,6 +811,7 @@ v3d_gem_init(struct drm_device *dev)
 	mutex_init(&v3d->bo_lock);
 	mutex_init(&v3d->reset_lock);
 	mutex_init(&v3d->sched_lock);
+	mutex_init(&v3d->cache_clean_lock);
 
 	/* Note: We don't allocate address 0.  Various bits of HW
 	 * treat 0 as special, such as the occlusion query counters
diff --git a/drivers/gpu/drm/v3d/v3d_irq.c b/drivers/gpu/drm/v3d/v3d_irq.c
index ce373ffd6380..fac3c542860b 100644
--- a/drivers/gpu/drm/v3d/v3d_irq.c
+++ b/drivers/gpu/drm/v3d/v3d_irq.c
@@ -4,9 +4,9 @@
 /**
  * DOC: Interrupt management for the V3D engine
  *
- * When we take a bin, render, or TFU done interrupt, we need to
- * signal the fence for that job so that the scheduler can queue up
- * the next one and unblock any waiters.
+ * When we take a bin, render, TFU done, or CSD done interrupt, we
+ * need to signal the fence for that job so that the scheduler can
+ * queue up the next one and unblock any waiters.
  *
  * When we take the binner out of memory interrupt, we need to
  * allocate some new memory and pass it to the binner so that the
@@ -20,6 +20,7 @@
 #define V3D_CORE_IRQS ((u32)(V3D_INT_OUTOMEM |	\
 			     V3D_INT_FLDONE |	\
 			     V3D_INT_FRDONE |	\
+			     V3D_INT_CSDDONE |	\
 			     V3D_INT_GMPV))
 
 #define V3D_HUB_IRQS ((u32)(V3D_HUB_INT_MMU_WRV |	\
@@ -112,6 +113,15 @@ v3d_irq(int irq, void *arg)
 		status = IRQ_HANDLED;
 	}
 
+	if (intsts & V3D_INT_CSDDONE) {
+		struct v3d_fence *fence =
+			to_v3d_fence(v3d->csd_job->base.irq_fence);
+
+		trace_v3d_csd_irq(&v3d->drm, fence->seqno);
+		dma_fence_signal(&fence->base);
+		status = IRQ_HANDLED;
+	}
+
 	/* We shouldn't be triggering these if we have GMP in
 	 * always-allowed mode.
 	 */
diff --git a/drivers/gpu/drm/v3d/v3d_regs.h b/drivers/gpu/drm/v3d/v3d_regs.h
index 8e88af237610..9a8ff0ce648e 100644
--- a/drivers/gpu/drm/v3d/v3d_regs.h
+++ b/drivers/gpu/drm/v3d/v3d_regs.h
@@ -238,8 +238,11 @@
 #define V3D_CTL_L2TCACTL                               0x00030
 # define V3D_L2TCACTL_TMUWCF                           BIT(8)
 # define V3D_L2TCACTL_L2T_NO_WM                        BIT(4)
+/* Invalidates cache lines. */
 # define V3D_L2TCACTL_FLM_FLUSH                        0
+/* Removes cachelines without writing dirty lines back. */
 # define V3D_L2TCACTL_FLM_CLEAR                        1
+/* Writes out dirty cachelines and marks them clean, but doesn't invalidate. */
 # define V3D_L2TCACTL_FLM_CLEAN                        2
 # define V3D_L2TCACTL_FLM_MASK                         V3D_MASK(2, 1)
 # define V3D_L2TCACTL_FLM_SHIFT                        1
@@ -255,6 +258,8 @@
 #define V3D_CTL_INT_MSK_CLR                            0x00064
 # define V3D_INT_QPU_MASK                              V3D_MASK(27, 16)
 # define V3D_INT_QPU_SHIFT                             16
+# define V3D_INT_CSDDONE                               BIT(7)
+# define V3D_INT_PCTR                                  BIT(6)
 # define V3D_INT_GMPV                                  BIT(5)
 # define V3D_INT_TRFB                                  BIT(4)
 # define V3D_INT_SPILLUSE                              BIT(3)
@@ -374,4 +379,72 @@
 #define V3D_GMP_PRESERVE_LOAD                          0x00818
 #define V3D_GMP_VALID_LINES                            0x00820
 
+#define V3D_CSD_STATUS                                 0x00900
+# define V3D_CSD_STATUS_NUM_COMPLETED_MASK             V3D_MASK(11, 4)
+# define V3D_CSD_STATUS_NUM_COMPLETED_SHIFT            4
+# define V3D_CSD_STATUS_NUM_ACTIVE_MASK                V3D_MASK(3, 2)
+# define V3D_CSD_STATUS_NUM_ACTIVE_SHIFT               2
+# define V3D_CSD_STATUS_HAVE_CURRENT_DISPATCH          BIT(1)
+# define V3D_CSD_STATUS_HAVE_QUEUED_DISPATCH           BIT(0)
+
+#define V3D_CSD_QUEUED_CFG0                            0x00904
+# define V3D_CSD_QUEUED_CFG0_NUM_WGS_X_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG0_NUM_WGS_X_SHIFT           16
+# define V3D_CSD_QUEUED_CFG0_WG_X_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG0_WG_X_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG1                            0x00908
+# define V3D_CSD_QUEUED_CFG1_NUM_WGS_Y_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG1_NUM_WGS_Y_SHIFT           16
+# define V3D_CSD_QUEUED_CFG1_WG_Y_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG1_WG_Y_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG2                            0x0090c
+# define V3D_CSD_QUEUED_CFG2_NUM_WGS_Z_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG2_NUM_WGS_Z_SHIFT           16
+# define V3D_CSD_QUEUED_CFG2_WG_Z_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG2_WG_Z_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG3                            0x00910
+# define V3D_CSD_QUEUED_CFG3_OVERLAP_WITH_PREV         BIT(26)
+# define V3D_CSD_QUEUED_CFG3_MAX_SG_ID_MASK            V3D_MASK(25, 20)
+# define V3D_CSD_QUEUED_CFG3_MAX_SG_ID_SHIFT           20
+# define V3D_CSD_QUEUED_CFG3_BATCHES_PER_SG_M1_MASK    V3D_MASK(19, 12)
+# define V3D_CSD_QUEUED_CFG3_BATCHES_PER_SG_M1_SHIFT   12
+# define V3D_CSD_QUEUED_CFG3_WGS_PER_SG_MASK           V3D_MASK(11, 8)
+# define V3D_CSD_QUEUED_CFG3_WGS_PER_SG_SHIFT          8
+# define V3D_CSD_QUEUED_CFG3_WG_SIZE_MASK              V3D_MASK(7, 0)
+# define V3D_CSD_QUEUED_CFG3_WG_SIZE_SHIFT             0
+
+/* Number of batches, minus 1 */
+#define V3D_CSD_QUEUED_CFG4                            0x00914
+
+/* Shader address, pnan, singleseg, threading, like a shader record. */
+#define V3D_CSD_QUEUED_CFG5                            0x00918
+
+/* Uniforms address (4 byte aligned) */
+#define V3D_CSD_QUEUED_CFG6                            0x0091c
+
+#define V3D_CSD_CURRENT_CFG0                          0x00920
+#define V3D_CSD_CURRENT_CFG1                          0x00924
+#define V3D_CSD_CURRENT_CFG2                          0x00928
+#define V3D_CSD_CURRENT_CFG3                          0x0092c
+#define V3D_CSD_CURRENT_CFG4                          0x00930
+#define V3D_CSD_CURRENT_CFG5                          0x00934
+#define V3D_CSD_CURRENT_CFG6                          0x00938
+
+#define V3D_CSD_CURRENT_ID0                            0x0093c
+# define V3D_CSD_CURRENT_ID0_WG_X_MASK                 V3D_MASK(31, 16)
+# define V3D_CSD_CURRENT_ID0_WG_X_SHIFT                16
+# define V3D_CSD_CURRENT_ID0_WG_IN_SG_MASK             V3D_MASK(11, 8)
+# define V3D_CSD_CURRENT_ID0_WG_IN_SG_SHIFT            8
+# define V3D_CSD_CURRENT_ID0_L_IDX_MASK                V3D_MASK(7, 0)
+# define V3D_CSD_CURRENT_ID0_L_IDX_SHIFT               0
+
+#define V3D_CSD_CURRENT_ID1                            0x00940
+# define V3D_CSD_CURRENT_ID0_WG_Z_MASK                 V3D_MASK(31, 16)
+# define V3D_CSD_CURRENT_ID0_WG_Z_SHIFT                16
+# define V3D_CSD_CURRENT_ID0_WG_Y_MASK                 V3D_MASK(15, 0)
+# define V3D_CSD_CURRENT_ID0_WG_Y_SHIFT                0
+
 #endif /* V3D_REGS_H */
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
index 884f00f5edf4..9cb3e1b3a024 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -48,6 +48,12 @@ to_tfu_job(struct drm_sched_job *sched_job)
 	return container_of(sched_job, struct v3d_tfu_job, base.base);
 }
 
+static struct v3d_csd_job *
+to_csd_job(struct drm_sched_job *sched_job)
+{
+	return container_of(sched_job, struct v3d_csd_job, base.base);
+}
+
 static void
 v3d_job_free(struct drm_sched_job *sched_job)
 {
@@ -243,6 +249,48 @@ v3d_tfu_job_run(struct drm_sched_job *sched_job)
 	return fence;
 }
 
+static struct dma_fence *
+v3d_csd_job_run(struct drm_sched_job *sched_job)
+{
+	struct v3d_csd_job *job = to_csd_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
+	struct drm_device *dev = &v3d->drm;
+	struct dma_fence *fence;
+	int i;
+
+	v3d->csd_job = job;
+
+	v3d_invalidate_caches(v3d);
+
+	fence = v3d_fence_create(v3d, V3D_CSD);
+	if (IS_ERR(fence))
+		return NULL;
+
+	if (job->base.irq_fence)
+		dma_fence_put(job->base.irq_fence);
+	job->base.irq_fence = dma_fence_get(fence);
+
+	trace_v3d_submit_csd(dev, to_v3d_fence(fence)->seqno);
+
+	for (i = 1; i <= 6; i++)
+		V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0 + 4 * i, job->args.cfg[i]);
+	/* CFG0 write kicks off the job. */
+	V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0, job->args.cfg[0]);
+
+	return fence;
+}
+
+static struct dma_fence *
+v3d_cache_clean_job_run(struct drm_sched_job *sched_job)
+{
+	struct v3d_job *job = to_v3d_job(sched_job);
+	struct v3d_dev *v3d = job->v3d;
+
+	v3d_clean_caches(v3d);
+
+	return NULL;
+}
+
 static void
 v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
 {
@@ -313,13 +361,31 @@ v3d_render_job_timedout(struct drm_sched_job *sched_job)
 }
 
 static void
-v3d_tfu_job_timedout(struct drm_sched_job *sched_job)
+v3d_generic_job_timedout(struct drm_sched_job *sched_job)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
 
 	v3d_gpu_reset_for_timeout(job->v3d, sched_job);
 }
 
+static void
+v3d_csd_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_csd_job *job = to_csd_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
+	u32 batches = V3D_CORE_READ(0, V3D_CSD_CURRENT_CFG4);
+
+	/* If we've made progress, skip reset and let the timer get
+	 * rearmed.
+	 */
+	if (job->timedout_batches != batches) {
+		job->timedout_batches = batches;
+		return;
+	}
+
+	v3d_gpu_reset_for_timeout(v3d, sched_job);
+}
+
 static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
 	.dependency = v3d_job_dependency,
 	.run_job = v3d_bin_job_run,
@@ -337,10 +403,24 @@ static const struct drm_sched_backend_ops v3d_render_sched_ops = {
 static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
 	.dependency = v3d_job_dependency,
 	.run_job = v3d_tfu_job_run,
-	.timedout_job = v3d_tfu_job_timedout,
+	.timedout_job = v3d_generic_job_timedout,
 	.free_job = v3d_job_free,
 };
 
+static const struct drm_sched_backend_ops v3d_csd_sched_ops = {
+	.dependency = v3d_job_dependency,
+	.run_job = v3d_csd_job_run,
+	.timedout_job = v3d_csd_job_timedout,
+	.free_job = v3d_job_free
+};
+
+static const struct drm_sched_backend_ops v3d_cache_clean_sched_ops = {
+	.dependency = v3d_job_dependency,
+	.run_job = v3d_cache_clean_job_run,
+	.timedout_job = v3d_generic_job_timedout,
+	.free_job = v3d_job_free
+};
+
 int
 v3d_sched_init(struct v3d_dev *v3d)
 {
@@ -367,7 +447,7 @@ v3d_sched_init(struct v3d_dev *v3d)
 	if (ret) {
 		dev_err(v3d->dev, "Failed to create render scheduler: %d.",
 			ret);
-		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
+		v3d_sched_fini(v3d);
 		return ret;
 	}
 
@@ -379,11 +459,36 @@ v3d_sched_init(struct v3d_dev *v3d)
 	if (ret) {
 		dev_err(v3d->dev, "Failed to create TFU scheduler: %d.",
 			ret);
-		drm_sched_fini(&v3d->queue[V3D_RENDER].sched);
-		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
+		v3d_sched_fini(v3d);
 		return ret;
 	}
 
+	if (v3d_has_csd(v3d)) {
+		ret = drm_sched_init(&v3d->queue[V3D_CSD].sched,
+				     &v3d_csd_sched_ops,
+				     hw_jobs_limit, job_hang_limit,
+				     msecs_to_jiffies(hang_limit_ms),
+				     "v3d_csd");
+		if (ret) {
+			dev_err(v3d->dev, "Failed to create CSD scheduler: %d.",
+				ret);
+			v3d_sched_fini(v3d);
+			return ret;
+		}
+
+		ret = drm_sched_init(&v3d->queue[V3D_CACHE_CLEAN].sched,
+				     &v3d_cache_clean_sched_ops,
+				     hw_jobs_limit, job_hang_limit,
+				     msecs_to_jiffies(hang_limit_ms),
+				     "v3d_cache_clean");
+		if (ret) {
+			dev_err(v3d->dev, "Failed to create CACHE_CLEAN scheduler: %d.",
+				ret);
+			v3d_sched_fini(v3d);
+			return ret;
+		}
+	}
+
 	return 0;
 }
 
@@ -392,6 +497,8 @@ v3d_sched_fini(struct v3d_dev *v3d)
 {
 	enum v3d_queue q;
 
-	for (q = 0; q < V3D_MAX_QUEUES; q++)
-		drm_sched_fini(&v3d->queue[q].sched);
+	for (q = 0; q < V3D_MAX_QUEUES; q++) {
+		if (v3d->queue[q].sched.ready)
+			drm_sched_fini(&v3d->queue[q].sched);
+	}
 }
diff --git a/drivers/gpu/drm/v3d/v3d_trace.h b/drivers/gpu/drm/v3d/v3d_trace.h
index edd984afa33f..7aa8dc356e54 100644
--- a/drivers/gpu/drm/v3d/v3d_trace.h
+++ b/drivers/gpu/drm/v3d/v3d_trace.h
@@ -124,6 +124,26 @@ TRACE_EVENT(v3d_tfu_irq,
 		      __entry->seqno)
 );
 
+TRACE_EVENT(v3d_csd_irq,
+	    TP_PROTO(struct drm_device *dev,
+		     uint64_t seqno),
+	    TP_ARGS(dev, seqno),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u64, seqno)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->seqno = seqno;
+			   ),
+
+	    TP_printk("dev=%u, seqno=%llu",
+		      __entry->dev,
+		      __entry->seqno)
+);
+
 TRACE_EVENT(v3d_submit_tfu_ioctl,
 	    TP_PROTO(struct drm_device *dev, u32 iia),
 	    TP_ARGS(dev, iia),
@@ -163,6 +183,80 @@ TRACE_EVENT(v3d_submit_tfu,
 		      __entry->seqno)
 );
 
+TRACE_EVENT(v3d_submit_csd_ioctl,
+	    TP_PROTO(struct drm_device *dev, u32 cfg5, u32 cfg6),
+	    TP_ARGS(dev, cfg5, cfg6),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u32, cfg5)
+			     __field(u32, cfg6)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->cfg5 = cfg5;
+			   __entry->cfg6 = cfg6;
+			   ),
+
+	    TP_printk("dev=%u, CFG5 0x%08x, CFG6 0x%08x",
+		      __entry->dev,
+		      __entry->cfg5,
+		      __entry->cfg6)
+);
+
+TRACE_EVENT(v3d_submit_csd,
+	    TP_PROTO(struct drm_device *dev,
+		     uint64_t seqno),
+	    TP_ARGS(dev, seqno),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u64, seqno)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->seqno = seqno;
+			   ),
+
+	    TP_printk("dev=%u, seqno=%llu",
+		      __entry->dev,
+		      __entry->seqno)
+);
+
+TRACE_EVENT(v3d_cache_clean_begin,
+	    TP_PROTO(struct drm_device *dev),
+	    TP_ARGS(dev),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   ),
+
+	    TP_printk("dev=%u",
+		      __entry->dev)
+);
+
+TRACE_EVENT(v3d_cache_clean_end,
+	    TP_PROTO(struct drm_device *dev),
+	    TP_ARGS(dev),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   ),
+
+	    TP_printk("dev=%u",
+		      __entry->dev)
+);
+
 TRACE_EVENT(v3d_reset_begin,
 	    TP_PROTO(struct drm_device *dev),
 	    TP_ARGS(dev),
diff --git a/include/uapi/drm/v3d_drm.h b/include/uapi/drm/v3d_drm.h
index ea70669d2138..58fbe48c91e9 100644
--- a/include/uapi/drm/v3d_drm.h
+++ b/include/uapi/drm/v3d_drm.h
@@ -37,6 +37,7 @@ extern "C" {
 #define DRM_V3D_GET_PARAM                         0x04
 #define DRM_V3D_GET_BO_OFFSET                     0x05
 #define DRM_V3D_SUBMIT_TFU                        0x06
+#define DRM_V3D_SUBMIT_CSD                        0x07
 
 #define DRM_IOCTL_V3D_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CL, struct drm_v3d_submit_cl)
 #define DRM_IOCTL_V3D_WAIT_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_WAIT_BO, struct drm_v3d_wait_bo)
@@ -45,6 +46,7 @@ extern "C" {
 #define DRM_IOCTL_V3D_GET_PARAM           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_PARAM, struct drm_v3d_get_param)
 #define DRM_IOCTL_V3D_GET_BO_OFFSET       DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_BO_OFFSET, struct drm_v3d_get_bo_offset)
 #define DRM_IOCTL_V3D_SUBMIT_TFU          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_TFU, struct drm_v3d_submit_tfu)
+#define DRM_IOCTL_V3D_SUBMIT_CSD          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CSD, struct drm_v3d_submit_csd)
 
 /**
  * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
@@ -190,6 +192,7 @@ enum drm_v3d_param {
 	DRM_V3D_PARAM_V3D_CORE0_IDENT1,
 	DRM_V3D_PARAM_V3D_CORE0_IDENT2,
 	DRM_V3D_PARAM_SUPPORTS_TFU,
+	DRM_V3D_PARAM_SUPPORTS_CSD,
 };
 
 struct drm_v3d_get_param {
@@ -230,6 +233,31 @@ struct drm_v3d_submit_tfu {
 	__u32 out_sync;
 };
 
+/* Submits a compute shader for dispatch.  This job will block on any
+ * previous compute shaders submitted on this fd, and any other
+ * synchronization must be performed with in_sync/out_sync.
+ */
+struct drm_v3d_submit_csd {
+	__u32 cfg[7];
+	__u32 coef[4];
+
+	/* Pointer to a u32 array of the BOs that are referenced by the job.
+	 */
+	__u64 bo_handles;
+
+	/* Number of BO handles passed in (size is that times 4). */
+	__u32 bo_handle_count;
+
+	/* sync object to block on before running the CSD job.  Each
+	 * CSD job will execute in the order submitted to its FD.
+	 * Synchronization against rendering/TFU jobs or CSD from
+	 * other fds requires using sync objects.
+	 */
+	__u32 in_sync;
+	/* Sync object to signal when the CSD job is done. */
+	__u32 out_sync;
+};
+
 #if defined(__cplusplus)
 }
 #endif
-- 
2.20.1

[PATCH 4/7] drm/v3d: Drop reservation of a shared slot in the dma-buf reservations.

[Eric Anholt]

We only set the excl (possible-writing) fence pointer and never add a
shared (read-only) fence.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/v3d/v3d_gem.c | 12 ------------
 1 file changed, 12 deletions(-)

diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
index b78ffc526885..ba3188b0d613 100644
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -253,18 +253,6 @@ v3d_lock_bo_reservations(struct drm_gem_object **bos,
 	if (ret)
 		return ret;
 
-	/* Reserve space for our shared (read-only) fence references,
-	 * before we commit the CL to the hardware.
-	 */
-	for (i = 0; i < bo_count; i++) {
-		ret = reservation_object_reserve_shared(bos[i]->resv, 1);
-		if (ret) {
-			drm_gem_unlock_reservations(bos, bo_count,
-						    acquire_ctx);
-			return ret;
-		}
-	}
-
 	return 0;
 }
 
-- 
2.20.1

[PATCH 5/7] drm: Add helpers for setting up an array of dma_fence dependencies.

[Eric Anholt]

I needed to add implicit dependency support for v3d, and Rob Herring
has been working on it for panfrost, and I had recently looked at the
lima implementation so I think this will be a good intersection of
what we all want and simplify our scheduler implementations.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/drm_gem.c | 94 +++++++++++++++++++++++++++++++++++++++
 include/drm/drm_gem.h     |  5 +++
 2 files changed, 99 insertions(+)

diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index 388b3742e562..7b3c73135e49 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -1311,3 +1311,97 @@ drm_gem_unlock_reservations(struct drm_gem_object **objs, int count,
 	ww_acquire_fini(acquire_ctx);
 }
 EXPORT_SYMBOL(drm_gem_unlock_reservations);
+
+/**
+ * drm_gem_fence_array_add - Adds the fence to an array of fences to be
+ * waited on, deduplicating fences from the same context.
+ *
+ * @fence_array array of dma_fence * for the job to block on.
+ * @fence the dma_fence to add to the list of dependencies.
+ *
+ * Returns:
+ * 0 on success, or an error on failing to expand the array.
+ */
+int drm_gem_fence_array_add(struct xarray *fence_array,
+			    struct dma_fence *fence)
+{
+	struct dma_fence *entry;
+	unsigned long index;
+	u32 id = 0;
+	int ret;
+
+	if (!fence)
+		return 0;
+
+	/* Deduplicate if we already depend on a fence from the same context.
+	 * This lets the size of the array of deps scale with the number of
+	 * engines involved, rather than the number of BOs.
+	 */
+	xa_for_each(fence_array, index, entry) {
+		if (entry->context != fence->context)
+			continue;
+
+		if (dma_fence_is_later(fence, entry)) {
+			dma_fence_put(entry);
+			xa_store(fence_array, index, fence, GFP_KERNEL);
+		} else {
+			dma_fence_put(fence);
+		}
+		return 0;
+	}
+
+	ret = xa_alloc(fence_array, &id, UINT_MAX, fence, GFP_KERNEL);
+	if (ret != 0)
+		dma_fence_put(fence);
+
+	return ret;
+}
+EXPORT_SYMBOL(drm_gem_fence_array_add);
+
+/**
+ * drm_gem_fence_array_add_implicit - Adds the implicit dependencies tracked
+ * in the GEM object's reservation object to an array of dma_fences for use in
+ * scheduling a rendering job.
+ *
+ * This should be called after drm_gem_lock_reservations() on your array of
+ * GEM objects used in the job but before updating the reservations with your
+ * own fences.
+ *
+ * @fence_array array of dma_fence * for the job to block on.
+ * @obj the gem object to add new dependencies from.
+ * @write whether the job might write the object (so we need to depend on
+ * shared fences in the reservation object).
+ */
+int drm_gem_fence_array_add_implicit(struct xarray *fence_array,
+				     struct drm_gem_object *obj,
+				     bool write)
+{
+	int ret;
+	struct dma_fence **fences;
+	unsigned i, fence_count;
+
+	if (!write) {
+		struct dma_fence *fence =
+			reservation_object_get_excl_rcu(obj->resv);
+
+		return drm_gem_fence_array_add(fence_array, fence);
+	}
+
+	ret = reservation_object_get_fences_rcu(obj->resv, NULL,
+						&fence_count, &fences);
+	if (ret || !fence_count)
+		return ret;
+
+	for (i = 0; i < fence_count; i++) {
+		ret = drm_gem_fence_array_add(fence_array, fences[i]);
+		if (ret)
+			break;
+	}
+
+	for (; i < fence_count; i++)
+		dma_fence_put(fences[i]);
+	kfree(fences);
+	return ret;
+}
+
+EXPORT_SYMBOL(drm_gem_fence_array_add_implicit);
diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h
index 2955aaab3dca..e957753d3f94 100644
--- a/include/drm/drm_gem.h
+++ b/include/drm/drm_gem.h
@@ -388,6 +388,11 @@ int drm_gem_lock_reservations(struct drm_gem_object **objs, int count,
 			      struct ww_acquire_ctx *acquire_ctx);
 void drm_gem_unlock_reservations(struct drm_gem_object **objs, int count,
 				 struct ww_acquire_ctx *acquire_ctx);
+int drm_gem_fence_array_add(struct xarray *fence_array,
+			    struct dma_fence *fence);
+int drm_gem_fence_array_add_implicit(struct xarray *fence_array,
+				     struct drm_gem_object *obj,
+				     bool write);
 int drm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 			    u32 handle, u64 *offset);
 int drm_gem_dumb_destroy(struct drm_file *file,
-- 
2.20.1

[PATCH 6/7] drm/v3d: Add missing implicit synchronization.

[Eric Anholt]

It is the expectation of existing userspace (X11 + Mesa, in
particular) that jobs submitted to the kernel against a shared BO will
get implicitly synchronized by their submission order.  If we want to
allow clever userspace to disable implicit synchronization, we should
do that under its own submit flag (as amdgpu and lima do).

Note that we currently only implicitly sync for the rendering pass,
not binning -- if you texture-from-pixmap in the binning vertex shader
(vertex coordinate generation), you'll miss out on synchronization.

Fixes flickering when multiple clients are running in parallel,
particularly GL apps and compositors.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/v3d/v3d_drv.h   | 11 +++---
 drivers/gpu/drm/v3d/v3d_gem.c   | 62 ++++++++++++++++++++++++---------
 drivers/gpu/drm/v3d/v3d_sched.c | 40 +++------------------
 3 files changed, 53 insertions(+), 60 deletions(-)

diff --git a/drivers/gpu/drm/v3d/v3d_drv.h b/drivers/gpu/drm/v3d/v3d_drv.h
index c4ef640effc2..6d31a6a5a08e 100644
--- a/drivers/gpu/drm/v3d/v3d_drv.h
+++ b/drivers/gpu/drm/v3d/v3d_drv.h
@@ -188,8 +188,10 @@ struct v3d_job {
 	struct drm_gem_object **bo;
 	u32 bo_count;
 
-	/* An optional fence userspace can pass in for the job to depend on. */
-	struct dma_fence *in_fence;
+	/* Array of struct dma_fence * to block on before submitting this job.
+	 */
+	struct xarray deps;
+	unsigned long last_dep;
 
 	/* v3d fence to be signaled by IRQ handler when the job is complete. */
 	struct dma_fence *irq_fence;
@@ -221,11 +223,6 @@ struct v3d_bin_job {
 struct v3d_render_job {
 	struct v3d_job base;
 
-	/* Optional fence for the binner, to depend on before starting
-	 * our job.
-	 */
-	struct dma_fence *bin_done_fence;
-
 	/* GPU virtual addresses of the start/end of the CL job. */
 	u32 start, end;
 
diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
index ba3188b0d613..cd92850684c1 100644
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -243,16 +243,25 @@ v3d_invalidate_caches(struct v3d_dev *v3d)
  * to v3d, so we don't attach dma-buf fences to them.
  */
 static int
-v3d_lock_bo_reservations(struct drm_gem_object **bos,
-			 int bo_count,
+v3d_lock_bo_reservations(struct v3d_job *job,
 			 struct ww_acquire_ctx *acquire_ctx)
 {
 	int i, ret;
 
-	ret = drm_gem_lock_reservations(bos, bo_count, acquire_ctx);
+	ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
 	if (ret)
 		return ret;
 
+	for (i = 0; i < job->bo_count; i++) {
+		ret = drm_gem_fence_array_add_implicit(&job->deps,
+						       job->bo[i], true);
+		if (ret) {
+			drm_gem_unlock_reservations(job->bo, job->bo_count,
+						    acquire_ctx);
+			return ret;
+		}
+	}
+
 	return 0;
 }
 
@@ -339,6 +348,8 @@ static void
 v3d_job_free(struct kref *ref)
 {
 	struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
+	unsigned long index;
+	struct dma_fence *fence;
 	int i;
 
 	for (i = 0; i < job->bo_count; i++) {
@@ -347,7 +358,11 @@ v3d_job_free(struct kref *ref)
 	}
 	kvfree(job->bo);
 
-	dma_fence_put(job->in_fence);
+	xa_for_each(&job->deps, index, fence) {
+		dma_fence_put(fence);
+	}
+	xa_destroy(&job->deps);
+
 	dma_fence_put(job->irq_fence);
 	dma_fence_put(job->done_fence);
 
@@ -414,6 +429,7 @@ v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
 	     struct v3d_job *job, void (*free)(struct kref *ref),
 	     u32 in_sync)
 {
+	struct dma_fence *in_fence = NULL;
 	int ret;
 
 	job->v3d = v3d;
@@ -423,15 +439,23 @@ v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
 	if (ret < 0)
 		return ret;
 
-	ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &job->in_fence);
-	if (ret == -EINVAL) {
-		pm_runtime_put_autosuspend(v3d->dev);
-		return ret;
-	}
+	xa_init_flags(&job->deps, XA_FLAGS_ALLOC);
+
+	ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &in_fence);
+	if (ret == -EINVAL)
+		goto fail;
+
+	ret = drm_gem_fence_array_add(&job->deps, in_fence);
+	if (ret)
+		goto fail;
 
 	kref_init(&job->refcount);
 
 	return 0;
+fail:
+	xa_destroy(&job->deps);
+	pm_runtime_put_autosuspend(v3d->dev);
+	return ret;
 }
 
 static int
@@ -552,8 +576,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(render->base.bo, render->base.bo_count,
-				       &acquire_ctx);
+	ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
 	if (ret)
 		goto fail;
 
@@ -563,7 +586,10 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 		if (ret)
 			goto fail_unreserve;
 
-		render->bin_done_fence = dma_fence_get(bin->base.done_fence);
+		ret = drm_gem_fence_array_add(&render->base.deps,
+					      dma_fence_get(bin->base.done_fence));
+		if (ret)
+			goto fail_unreserve;
 	}
 
 	ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
@@ -656,8 +682,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 	}
 	spin_unlock(&file_priv->table_lock);
 
-	ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count,
-				       &acquire_ctx);
+	ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
 	if (ret)
 		goto fail;
 
@@ -746,8 +771,7 @@ v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count,
-				       &acquire_ctx);
+	ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
 	if (ret)
 		goto fail;
 
@@ -756,7 +780,11 @@ v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail_unreserve;
 
-	clean_job->in_fence = dma_fence_get(job->base.done_fence);
+	ret = drm_gem_fence_array_add(&clean_job->deps,
+				      job->base.done_fence);
+	if (ret)
+		goto fail_unreserve;
+
 	ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
 	if (ret)
 		goto fail_unreserve;
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
index 9cb3e1b3a024..d688f238b2b9 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -74,47 +74,15 @@ v3d_job_dependency(struct drm_sched_job *sched_job,
 		   struct drm_sched_entity *s_entity)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
-	struct dma_fence *fence;
-
-	fence = job->in_fence;
-	if (fence) {
-		job->in_fence = NULL;
-		return fence;
-	}
 
 	/* XXX: Wait on a fence for switching the GMP if necessary,
 	 * and then do so.
 	 */
 
-	return NULL;
-}
+	if (!xa_empty(&job->deps))
+		return xa_erase(&job->deps, job->last_dep++);
 
-/**
- * Returns the fences that the render job depends on, one by one.
- * v3d_job_run() won't be called until all of them have been signaled.
- */
-static struct dma_fence *
-v3d_render_job_dependency(struct drm_sched_job *sched_job,
-			  struct drm_sched_entity *s_entity)
-{
-	struct v3d_render_job *job = to_render_job(sched_job);
-	struct dma_fence *fence;
-
-	fence = v3d_job_dependency(sched_job, s_entity);
-	if (fence)
-		return fence;
-
-	/* If we had a bin job, the render job definitely depends on
-	 * it. We first have to wait for bin to be scheduled, so that
-	 * its done_fence is created.
-	 */
-	fence = job->bin_done_fence;
-	if (fence) {
-		job->bin_done_fence = NULL;
-		return fence;
-	}
-
-	return fence;
+	return NULL;
 }
 
 static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job)
@@ -394,7 +362,7 @@ static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
 };
 
 static const struct drm_sched_backend_ops v3d_render_sched_ops = {
-	.dependency = v3d_render_job_dependency,
+	.dependency = v3d_job_dependency,
 	.run_job = v3d_render_job_run,
 	.timedout_job = v3d_render_job_timedout,
 	.free_job = v3d_job_free,
-- 
2.20.1

[PATCH 7/7] drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

[Eric Anholt]

I haven't tested this, but it's a pretty direct port of what I did for
v3d.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/lima/lima_gem.c   | 37 +----------------
 drivers/gpu/drm/lima/lima_sched.c | 66 ++++++-------------------------
 drivers/gpu/drm/lima/lima_sched.h |  6 +--
 3 files changed, 16 insertions(+), 93 deletions(-)

diff --git a/drivers/gpu/drm/lima/lima_gem.c b/drivers/gpu/drm/lima/lima_gem.c
index 2d3cf96f6c58..8f80286c80b4 100644
--- a/drivers/gpu/drm/lima/lima_gem.c
+++ b/drivers/gpu/drm/lima/lima_gem.c
@@ -144,40 +144,7 @@ static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
 	if (explicit)
 		return 0;
 
-	/* implicit sync use bo fence in resv obj */
-	if (write) {
-		unsigned nr_fences;
-		struct dma_fence **fences;
-		int i;
-
-		err = reservation_object_get_fences_rcu(
-			bo->gem.resv, NULL, &nr_fences, &fences);
-		if (err || !nr_fences)
-			return err;
-
-		for (i = 0; i < nr_fences; i++) {
-			err = lima_sched_task_add_dep(task, fences[i]);
-			if (err)
-				break;
-		}
-
-		/* for error case free remaining fences */
-		for ( ; i < nr_fences; i++)
-			dma_fence_put(fences[i]);
-
-		kfree(fences);
-	} else {
-		struct dma_fence *fence;
-
-		fence = reservation_object_get_excl_rcu(bo->gem.resv);
-		if (fence) {
-			err = lima_sched_task_add_dep(task, fence);
-			if (err)
-				dma_fence_put(fence);
-		}
-	}
-
-	return err;
+	return drm_gem_fence_array_add_implicit(&task->deps, &bo->gem, write);
 }
 
 static int lima_gem_lock_bos(struct lima_bo **bos, u32 nr_bos,
@@ -250,7 +217,7 @@ static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
 		if (err)
 			return err;
 
-		err = lima_sched_task_add_dep(submit->task, fence);
+		err = drm_gem_fence_array_add(&submit->task->deps, fence);
 		if (err) {
 			dma_fence_put(fence);
 			return err;
diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c
index 97bd9c1deb87..e253d031fb3d 100644
--- a/drivers/gpu/drm/lima/lima_sched.c
+++ b/drivers/gpu/drm/lima/lima_sched.c
@@ -3,6 +3,7 @@
 
 #include <linux/kthread.h>
 #include <linux/slab.h>
+#include <linux/xarray.h>
 
 #include "lima_drv.h"
 #include "lima_sched.h"
@@ -126,19 +127,24 @@ int lima_sched_task_init(struct lima_sched_task *task,
 
 	task->num_bos = num_bos;
 	task->vm = lima_vm_get(vm);
+
+	xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
+
 	return 0;
 }
 
 void lima_sched_task_fini(struct lima_sched_task *task)
 {
+	struct dma_fence *fence;
+	unsigned long index;
 	int i;
 
 	drm_sched_job_cleanup(&task->base);
 
-	for (i = 0; i < task->num_dep; i++)
-		dma_fence_put(task->dep[i]);
-
-	kfree(task->dep);
+	xa_for_each(&task->deps, index, fence) {
+		dma_fence_put(fence);
+	}
+	xa_destroy(&task->deps);
 
 	if (task->bos) {
 		for (i = 0; i < task->num_bos; i++)
@@ -149,42 +155,6 @@ void lima_sched_task_fini(struct lima_sched_task *task)
 	lima_vm_put(task->vm);
 }
 
-int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence)
-{
-	int i, new_dep = 4;
-
-	/* same context's fence is definitly earlier then this task */
-	if (fence->context == task->base.s_fence->finished.context) {
-		dma_fence_put(fence);
-		return 0;
-	}
-
-	if (task->dep && task->num_dep == task->max_dep)
-		new_dep = task->max_dep * 2;
-
-	if (task->max_dep < new_dep) {
-		void *dep = krealloc(task->dep, sizeof(*task->dep) * new_dep, GFP_KERNEL);
-
-		if (!dep)
-			return -ENOMEM;
-
-		task->max_dep = new_dep;
-		task->dep = dep;
-	}
-
-	for (i = 0; i < task->num_dep; i++) {
-		if (task->dep[i]->context == fence->context &&
-		    dma_fence_is_later(fence, task->dep[i])) {
-			dma_fence_put(task->dep[i]);
-			task->dep[i] = fence;
-			return 0;
-		}
-	}
-
-	task->dep[task->num_dep++] = fence;
-	return 0;
-}
-
 int lima_sched_context_init(struct lima_sched_pipe *pipe,
 			    struct lima_sched_context *context,
 			    atomic_t *guilty)
@@ -213,21 +183,9 @@ static struct dma_fence *lima_sched_dependency(struct drm_sched_job *job,
 					       struct drm_sched_entity *entity)
 {
 	struct lima_sched_task *task = to_lima_task(job);
-	int i;
-
-	for (i = 0; i < task->num_dep; i++) {
-		struct dma_fence *fence = task->dep[i];
-
-		if (!task->dep[i])
-			continue;
-
-		task->dep[i] = NULL;
 
-		if (!dma_fence_is_signaled(fence))
-			return fence;
-
-		dma_fence_put(fence);
-	}
+	if (!xa_empty(&task->deps))
+		return xa_erase(&task->deps, task->last_dep++);
 
 	return NULL;
 }
diff --git a/drivers/gpu/drm/lima/lima_sched.h b/drivers/gpu/drm/lima/lima_sched.h
index b017cfa7e327..928af91c1118 100644
--- a/drivers/gpu/drm/lima/lima_sched.h
+++ b/drivers/gpu/drm/lima/lima_sched.h
@@ -14,9 +14,8 @@ struct lima_sched_task {
 	struct lima_vm *vm;
 	void *frame;
 
-	struct dma_fence **dep;
-	int num_dep;
-	int max_dep;
+	struct xarray deps;
+	unsigned long last_dep;
 
 	struct lima_bo **bos;
 	int num_bos;
@@ -78,7 +77,6 @@ int lima_sched_task_init(struct lima_sched_task *task,
 			 struct lima_bo **bos, int num_bos,
 			 struct lima_vm *vm);
 void lima_sched_task_fini(struct lima_sched_task *task);
-int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence);
 
 int lima_sched_context_init(struct lima_sched_pipe *pipe,
 			    struct lima_sched_context *context,
-- 
2.20.1

Re: [PATCH 7/7] drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

[Qiang Yu]

On Tue, Apr 2, 2019 at 6:26 AM Eric Anholt <eric@anholt.net> wrote:
>
> I haven't tested this, but it's a pretty direct port of what I did for
> v3d.
>
> Signed-off-by: Eric Anholt <eric@anholt.net>
> ---
>  drivers/gpu/drm/lima/lima_gem.c   | 37 +----------------
>  drivers/gpu/drm/lima/lima_sched.c | 66 ++++++-------------------------
>  drivers/gpu/drm/lima/lima_sched.h |  6 +--
>  3 files changed, 16 insertions(+), 93 deletions(-)
>
> diff --git a/drivers/gpu/drm/lima/lima_gem.c b/drivers/gpu/drm/lima/lima_gem.c
> index 2d3cf96f6c58..8f80286c80b4 100644
> --- a/drivers/gpu/drm/lima/lima_gem.c
> +++ b/drivers/gpu/drm/lima/lima_gem.c
> @@ -144,40 +144,7 @@ static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
>         if (explicit)
>                 return 0;
>
> -       /* implicit sync use bo fence in resv obj */
> -       if (write) {
> -               unsigned nr_fences;
> -               struct dma_fence **fences;
> -               int i;
> -
> -               err = reservation_object_get_fences_rcu(
> -                       bo->gem.resv, NULL, &nr_fences, &fences);
> -               if (err || !nr_fences)
> -                       return err;
> -
> -               for (i = 0; i < nr_fences; i++) {
> -                       err = lima_sched_task_add_dep(task, fences[i]);
> -                       if (err)
> -                               break;
> -               }
> -
> -               /* for error case free remaining fences */
> -               for ( ; i < nr_fences; i++)
> -                       dma_fence_put(fences[i]);
> -
> -               kfree(fences);
> -       } else {
> -               struct dma_fence *fence;
> -
> -               fence = reservation_object_get_excl_rcu(bo->gem.resv);
> -               if (fence) {
> -                       err = lima_sched_task_add_dep(task, fence);
> -                       if (err)
> -                               dma_fence_put(fence);
> -               }
> -       }
> -
> -       return err;
> +       return drm_gem_fence_array_add_implicit(&task->deps, &bo->gem, write);
>  }
>
>  static int lima_gem_lock_bos(struct lima_bo **bos, u32 nr_bos,
> @@ -250,7 +217,7 @@ static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
>                 if (err)
>                         return err;
>
> -               err = lima_sched_task_add_dep(submit->task, fence);
> +               err = drm_gem_fence_array_add(&submit->task->deps, fence);
>                 if (err) {
>                         dma_fence_put(fence);
>                         return err;
> diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c
> index 97bd9c1deb87..e253d031fb3d 100644
> --- a/drivers/gpu/drm/lima/lima_sched.c
> +++ b/drivers/gpu/drm/lima/lima_sched.c
> @@ -3,6 +3,7 @@
>
>  #include <linux/kthread.h>
>  #include <linux/slab.h>
> +#include <linux/xarray.h>
>
>  #include "lima_drv.h"
>  #include "lima_sched.h"
> @@ -126,19 +127,24 @@ int lima_sched_task_init(struct lima_sched_task *task,
>
>         task->num_bos = num_bos;
>         task->vm = lima_vm_get(vm);
> +
> +       xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
> +
>         return 0;
>  }
>
>  void lima_sched_task_fini(struct lima_sched_task *task)
>  {
> +       struct dma_fence *fence;
> +       unsigned long index;
>         int i;
>
>         drm_sched_job_cleanup(&task->base);
>
> -       for (i = 0; i < task->num_dep; i++)
> -               dma_fence_put(task->dep[i]);
> -
> -       kfree(task->dep);
> +       xa_for_each(&task->deps, index, fence) {
> +               dma_fence_put(fence);
> +       }
> +       xa_destroy(&task->deps);
>
>         if (task->bos) {
>                 for (i = 0; i < task->num_bos; i++)
> @@ -149,42 +155,6 @@ void lima_sched_task_fini(struct lima_sched_task *task)
>         lima_vm_put(task->vm);
>  }
>
> -int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence)
> -{
> -       int i, new_dep = 4;
> -
> -       /* same context's fence is definitly earlier then this task */
> -       if (fence->context == task->base.s_fence->finished.context) {
> -               dma_fence_put(fence);
> -               return 0;
> -       }

Seems you dropped this check in the drm_gem_fence_array_add, no bug if we
don't have this, but redundant fence will be added in the deps array. Maybe we
can add a context parameter to drm_gem_fence_array_add and
drm_gem_fence_array_add_implicit to filter out fences from same
drm_sched_entity.

Regards,
Qiang


> -
> -       if (task->dep && task->num_dep == task->max_dep)
> -               new_dep = task->max_dep * 2;
> -
> -       if (task->max_dep < new_dep) {
> -               void *dep = krealloc(task->dep, sizeof(*task->dep) * new_dep, GFP_KERNEL);
> -
> -               if (!dep)
> -                       return -ENOMEM;
> -
> -               task->max_dep = new_dep;
> -               task->dep = dep;
> -       }
> -
> -       for (i = 0; i < task->num_dep; i++) {
> -               if (task->dep[i]->context == fence->context &&
> -                   dma_fence_is_later(fence, task->dep[i])) {
> -                       dma_fence_put(task->dep[i]);
> -                       task->dep[i] = fence;
> -                       return 0;
> -               }
> -       }
> -
> -       task->dep[task->num_dep++] = fence;
> -       return 0;
> -}
> -
>  int lima_sched_context_init(struct lima_sched_pipe *pipe,
>                             struct lima_sched_context *context,
>                             atomic_t *guilty)
> @@ -213,21 +183,9 @@ static struct dma_fence *lima_sched_dependency(struct drm_sched_job *job,
>                                                struct drm_sched_entity *entity)
>  {
>         struct lima_sched_task *task = to_lima_task(job);
> -       int i;
> -
> -       for (i = 0; i < task->num_dep; i++) {
> -               struct dma_fence *fence = task->dep[i];
> -
> -               if (!task->dep[i])
> -                       continue;
> -
> -               task->dep[i] = NULL;
>
> -               if (!dma_fence_is_signaled(fence))
> -                       return fence;
> -
> -               dma_fence_put(fence);
> -       }
> +       if (!xa_empty(&task->deps))
> +               return xa_erase(&task->deps, task->last_dep++);
>
>         return NULL;
>  }
> diff --git a/drivers/gpu/drm/lima/lima_sched.h b/drivers/gpu/drm/lima/lima_sched.h
> index b017cfa7e327..928af91c1118 100644
> --- a/drivers/gpu/drm/lima/lima_sched.h
> +++ b/drivers/gpu/drm/lima/lima_sched.h
> @@ -14,9 +14,8 @@ struct lima_sched_task {
>         struct lima_vm *vm;
>         void *frame;
>
> -       struct dma_fence **dep;
> -       int num_dep;
> -       int max_dep;
> +       struct xarray deps;
> +       unsigned long last_dep;
>
>         struct lima_bo **bos;
>         int num_bos;
> @@ -78,7 +77,6 @@ int lima_sched_task_init(struct lima_sched_task *task,
>                          struct lima_bo **bos, int num_bos,
>                          struct lima_vm *vm);
>  void lima_sched_task_fini(struct lima_sched_task *task);
> -int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence);
>
>  int lima_sched_context_init(struct lima_sched_pipe *pipe,
>                             struct lima_sched_context *context,
> --
> 2.20.1
>

Re: [PATCH 7/7] drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

[Eric Anholt]

[-- Attachment #1: Type: text/plain, Size: 4877 bytes --]

Qiang Yu <yuq825@gmail.com> writes:

> On Tue, Apr 2, 2019 at 6:26 AM Eric Anholt <eric@anholt.net> wrote:
>>
>> I haven't tested this, but it's a pretty direct port of what I did for
>> v3d.
>>
>> Signed-off-by: Eric Anholt <eric@anholt.net>
>> ---
>>  drivers/gpu/drm/lima/lima_gem.c   | 37 +----------------
>>  drivers/gpu/drm/lima/lima_sched.c | 66 ++++++-------------------------
>>  drivers/gpu/drm/lima/lima_sched.h |  6 +--
>>  3 files changed, 16 insertions(+), 93 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/lima/lima_gem.c b/drivers/gpu/drm/lima/lima_gem.c
>> index 2d3cf96f6c58..8f80286c80b4 100644
>> --- a/drivers/gpu/drm/lima/lima_gem.c
>> +++ b/drivers/gpu/drm/lima/lima_gem.c
>> @@ -144,40 +144,7 @@ static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
>>         if (explicit)
>>                 return 0;
>>
>> -       /* implicit sync use bo fence in resv obj */
>> -       if (write) {
>> -               unsigned nr_fences;
>> -               struct dma_fence **fences;
>> -               int i;
>> -
>> -               err = reservation_object_get_fences_rcu(
>> -                       bo->gem.resv, NULL, &nr_fences, &fences);
>> -               if (err || !nr_fences)
>> -                       return err;
>> -
>> -               for (i = 0; i < nr_fences; i++) {
>> -                       err = lima_sched_task_add_dep(task, fences[i]);
>> -                       if (err)
>> -                               break;
>> -               }
>> -
>> -               /* for error case free remaining fences */
>> -               for ( ; i < nr_fences; i++)
>> -                       dma_fence_put(fences[i]);
>> -
>> -               kfree(fences);
>> -       } else {
>> -               struct dma_fence *fence;
>> -
>> -               fence = reservation_object_get_excl_rcu(bo->gem.resv);
>> -               if (fence) {
>> -                       err = lima_sched_task_add_dep(task, fence);
>> -                       if (err)
>> -                               dma_fence_put(fence);
>> -               }
>> -       }
>> -
>> -       return err;
>> +       return drm_gem_fence_array_add_implicit(&task->deps, &bo->gem, write);
>>  }
>>
>>  static int lima_gem_lock_bos(struct lima_bo **bos, u32 nr_bos,
>> @@ -250,7 +217,7 @@ static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
>>                 if (err)
>>                         return err;
>>
>> -               err = lima_sched_task_add_dep(submit->task, fence);
>> +               err = drm_gem_fence_array_add(&submit->task->deps, fence);
>>                 if (err) {
>>                         dma_fence_put(fence);
>>                         return err;
>> diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c
>> index 97bd9c1deb87..e253d031fb3d 100644
>> --- a/drivers/gpu/drm/lima/lima_sched.c
>> +++ b/drivers/gpu/drm/lima/lima_sched.c
>> @@ -3,6 +3,7 @@
>>
>>  #include <linux/kthread.h>
>>  #include <linux/slab.h>
>> +#include <linux/xarray.h>
>>
>>  #include "lima_drv.h"
>>  #include "lima_sched.h"
>> @@ -126,19 +127,24 @@ int lima_sched_task_init(struct lima_sched_task *task,
>>
>>         task->num_bos = num_bos;
>>         task->vm = lima_vm_get(vm);
>> +
>> +       xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
>> +
>>         return 0;
>>  }
>>
>>  void lima_sched_task_fini(struct lima_sched_task *task)
>>  {
>> +       struct dma_fence *fence;
>> +       unsigned long index;
>>         int i;
>>
>>         drm_sched_job_cleanup(&task->base);
>>
>> -       for (i = 0; i < task->num_dep; i++)
>> -               dma_fence_put(task->dep[i]);
>> -
>> -       kfree(task->dep);
>> +       xa_for_each(&task->deps, index, fence) {
>> +               dma_fence_put(fence);
>> +       }
>> +       xa_destroy(&task->deps);
>>
>>         if (task->bos) {
>>                 for (i = 0; i < task->num_bos; i++)
>> @@ -149,42 +155,6 @@ void lima_sched_task_fini(struct lima_sched_task *task)
>>         lima_vm_put(task->vm);
>>  }
>>
>> -int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence)
>> -{
>> -       int i, new_dep = 4;
>> -
>> -       /* same context's fence is definitly earlier then this task */
>> -       if (fence->context == task->base.s_fence->finished.context) {
>> -               dma_fence_put(fence);
>> -               return 0;
>> -       }
>
> Seems you dropped this check in the drm_gem_fence_array_add, no bug if we
> don't have this, but redundant fence will be added in the deps array. Maybe we
> can add a context parameter to drm_gem_fence_array_add and
> drm_gem_fence_array_add_implicit to filter out fences from same
> drm_sched_entity.

Does this feel important to you?  To me, one extra slot in the array and
a trip through the top of drm_sched_entity_add_dependency_cb() doesn't
like it's feel worth making the API clumsier.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

Re: [PATCH 7/7] drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

[Qiang Yu]

Indeed not that important, so patch 5&7 is:
Reviewed-and-tested-by: Qiang Yu <yuq825@gmail.com>

Regards,
Qiang

On Wed, Apr 3, 2019 at 12:57 AM Eric Anholt <eric@anholt.net> wrote:
>
> Qiang Yu <yuq825@gmail.com> writes:
>
> > On Tue, Apr 2, 2019 at 6:26 AM Eric Anholt <eric@anholt.net> wrote:
> >>
> >> I haven't tested this, but it's a pretty direct port of what I did for
> >> v3d.
> >>
> >> Signed-off-by: Eric Anholt <eric@anholt.net>
> >> ---
> >>  drivers/gpu/drm/lima/lima_gem.c   | 37 +----------------
> >>  drivers/gpu/drm/lima/lima_sched.c | 66 ++++++-------------------------
> >>  drivers/gpu/drm/lima/lima_sched.h |  6 +--
> >>  3 files changed, 16 insertions(+), 93 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/lima/lima_gem.c b/drivers/gpu/drm/lima/lima_gem.c
> >> index 2d3cf96f6c58..8f80286c80b4 100644
> >> --- a/drivers/gpu/drm/lima/lima_gem.c
> >> +++ b/drivers/gpu/drm/lima/lima_gem.c
> >> @@ -144,40 +144,7 @@ static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
> >>         if (explicit)
> >>                 return 0;
> >>
> >> -       /* implicit sync use bo fence in resv obj */
> >> -       if (write) {
> >> -               unsigned nr_fences;
> >> -               struct dma_fence **fences;
> >> -               int i;
> >> -
> >> -               err = reservation_object_get_fences_rcu(
> >> -                       bo->gem.resv, NULL, &nr_fences, &fences);
> >> -               if (err || !nr_fences)
> >> -                       return err;
> >> -
> >> -               for (i = 0; i < nr_fences; i++) {
> >> -                       err = lima_sched_task_add_dep(task, fences[i]);
> >> -                       if (err)
> >> -                               break;
> >> -               }
> >> -
> >> -               /* for error case free remaining fences */
> >> -               for ( ; i < nr_fences; i++)
> >> -                       dma_fence_put(fences[i]);
> >> -
> >> -               kfree(fences);
> >> -       } else {
> >> -               struct dma_fence *fence;
> >> -
> >> -               fence = reservation_object_get_excl_rcu(bo->gem.resv);
> >> -               if (fence) {
> >> -                       err = lima_sched_task_add_dep(task, fence);
> >> -                       if (err)
> >> -                               dma_fence_put(fence);
> >> -               }
> >> -       }
> >> -
> >> -       return err;
> >> +       return drm_gem_fence_array_add_implicit(&task->deps, &bo->gem, write);
> >>  }
> >>
> >>  static int lima_gem_lock_bos(struct lima_bo **bos, u32 nr_bos,
> >> @@ -250,7 +217,7 @@ static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
> >>                 if (err)
> >>                         return err;
> >>
> >> -               err = lima_sched_task_add_dep(submit->task, fence);
> >> +               err = drm_gem_fence_array_add(&submit->task->deps, fence);
> >>                 if (err) {
> >>                         dma_fence_put(fence);
> >>                         return err;
> >> diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c
> >> index 97bd9c1deb87..e253d031fb3d 100644
> >> --- a/drivers/gpu/drm/lima/lima_sched.c
> >> +++ b/drivers/gpu/drm/lima/lima_sched.c
> >> @@ -3,6 +3,7 @@
> >>
> >>  #include <linux/kthread.h>
> >>  #include <linux/slab.h>
> >> +#include <linux/xarray.h>
> >>
> >>  #include "lima_drv.h"
> >>  #include "lima_sched.h"
> >> @@ -126,19 +127,24 @@ int lima_sched_task_init(struct lima_sched_task *task,
> >>
> >>         task->num_bos = num_bos;
> >>         task->vm = lima_vm_get(vm);
> >> +
> >> +       xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
> >> +
> >>         return 0;
> >>  }
> >>
> >>  void lima_sched_task_fini(struct lima_sched_task *task)
> >>  {
> >> +       struct dma_fence *fence;
> >> +       unsigned long index;
> >>         int i;
> >>
> >>         drm_sched_job_cleanup(&task->base);
> >>
> >> -       for (i = 0; i < task->num_dep; i++)
> >> -               dma_fence_put(task->dep[i]);
> >> -
> >> -       kfree(task->dep);
> >> +       xa_for_each(&task->deps, index, fence) {
> >> +               dma_fence_put(fence);
> >> +       }
> >> +       xa_destroy(&task->deps);
> >>
> >>         if (task->bos) {
> >>                 for (i = 0; i < task->num_bos; i++)
> >> @@ -149,42 +155,6 @@ void lima_sched_task_fini(struct lima_sched_task *task)
> >>         lima_vm_put(task->vm);
> >>  }
> >>
> >> -int lima_sched_task_add_dep(struct lima_sched_task *task, struct dma_fence *fence)
> >> -{
> >> -       int i, new_dep = 4;
> >> -
> >> -       /* same context's fence is definitly earlier then this task */
> >> -       if (fence->context == task->base.s_fence->finished.context) {
> >> -               dma_fence_put(fence);
> >> -               return 0;
> >> -       }
> >
> > Seems you dropped this check in the drm_gem_fence_array_add, no bug if we
> > don't have this, but redundant fence will be added in the deps array. Maybe we
> > can add a context parameter to drm_gem_fence_array_add and
> > drm_gem_fence_array_add_implicit to filter out fences from same
> > drm_sched_entity.
>
> Does this feel important to you?  To me, one extra slot in the array and
> a trip through the top of drm_sched_entity_add_dependency_cb() doesn't
> like it's feel worth making the API clumsier.

Re: [PATCH 7/7] drm/lima: Use the drm_gem_fence_array_add helpers for our deps.

[Eric Anholt]

[-- Attachment #1: Type: text/plain, Size: 183 bytes --]

Qiang Yu <yuq825@gmail.com> writes:

> Indeed not that important, so patch 5&7 is:
> Reviewed-and-tested-by: Qiang Yu <yuq825@gmail.com>

Merged these two.  Thanks for trying it out!

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]