drm/vc4: Expose dma-buf fences for V3D rendering.

This is needed for proper synchronization with display on another DRM
device (pl111 or tinydrm) with buffers produced by vc4 V3D.  Fixes the
new igt vc4_dmabuf_poll testcase, and rendering of one of the glmark2
desktop tests on pl111+vc4.

This doesn't yet introduce waits on another device's fences before
vc4's rendering/display, because I don't have testcases for them.

v2: Reuse dma_fence_free(), retitle commit message to clarify that
    it's not a full dma-buf fencing implementation yet.

Signed-off-by: Eric Anholt <eric@anholt.net>
Link: http://patchwork.freedesktop.org/patch/msgid/20170412191202.22740-6-eric@anholt.net
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>

1 files changed, 133 insertions, 3 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index e9c381c42139..a1a01044263c 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -463,6 +463,8 @@ vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
 	for (i = 0; i < exec->bo_count; i++) {
 		bo = to_vc4_bo(&exec->bo[i]->base);
 		bo->seqno = seqno;
+
+		reservation_object_add_shared_fence(bo->resv, exec->fence);
 	}
 
 	list_for_each_entry(bo, &exec->unref_list, unref_head) {
@@ -472,7 +474,103 @@ vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
 	for (i = 0; i < exec->rcl_write_bo_count; i++) {
 		bo = to_vc4_bo(&exec->rcl_write_bo[i]->base);
 		bo->write_seqno = seqno;
+
+		reservation_object_add_excl_fence(bo->resv, exec->fence);
+	}
+}
+
+static void
+vc4_unlock_bo_reservations(struct drm_device *dev,
+			   struct vc4_exec_info *exec,
+			   struct ww_acquire_ctx *acquire_ctx)
+{
+	int i;
+
+	for (i = 0; i < exec->bo_count; i++) {
+		struct vc4_bo *bo = to_vc4_bo(&exec->bo[i]->base);
+
+		ww_mutex_unlock(&bo->resv->lock);
 	}
+
+	ww_acquire_fini(acquire_ctx);
+}
+
+/* Takes the reservation lock on all the BOs being referenced, so that
+ * at queue submit time we can update the reservations.
+ *
+ * We don't lock the RCL the tile alloc/state BOs, or overflow memory
+ * (all of which are on exec->unref_list).  They're entirely private
+ * to vc4, so we don't attach dma-buf fences to them.
+ */
+static int
+vc4_lock_bo_reservations(struct drm_device *dev,
+			 struct vc4_exec_info *exec,
+			 struct ww_acquire_ctx *acquire_ctx)
+{
+	int contended_lock = -1;
+	int i, ret;
+	struct vc4_bo *bo;
+
+	ww_acquire_init(acquire_ctx, &reservation_ww_class);
+
+retry:
+	if (contended_lock != -1) {
+		bo = to_vc4_bo(&exec->bo[contended_lock]->base);
+		ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
+						       acquire_ctx);
+		if (ret) {
+			ww_acquire_done(acquire_ctx);
+			return ret;
+		}
+	}
+
+	for (i = 0; i < exec->bo_count; i++) {
+		if (i == contended_lock)
+			continue;
+
+		bo = to_vc4_bo(&exec->bo[i]->base);
+
+		ret = ww_mutex_lock_interruptible(&bo->resv->lock, acquire_ctx);
+		if (ret) {
+			int j;
+
+			for (j = 0; j < i; j++) {
+				bo = to_vc4_bo(&exec->bo[j]->base);
+				ww_mutex_unlock(&bo->resv->lock);
+			}
+
+			if (contended_lock != -1 && contended_lock >= i) {
+				bo = to_vc4_bo(&exec->bo[contended_lock]->base);
+
+				ww_mutex_unlock(&bo->resv->lock);
+			}
+
+			if (ret == -EDEADLK) {
+				contended_lock = i;
+				goto retry;
+			}
+
+			ww_acquire_done(acquire_ctx);
+			return ret;
+		}
+	}
+
+	ww_acquire_done(acquire_ctx);
+
+	/* Reserve space for our shared (read-only) fence references,
+	 * before we commit the CL to the hardware.
+	 */
+	for (i = 0; i < exec->bo_count; i++) {
+		bo = to_vc4_bo(&exec->bo[i]->base);
+
+		ret = reservation_object_reserve_shared(bo->resv);
+		if (ret) {
+			vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
+			return ret;
+		}
+	}
+
+	return 0;
 }
 
 /* Queues a struct vc4_exec_info for execution.  If no job is
@@ -484,19 +582,34 @@ vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
  * then bump the end address.  That's a change for a later date,
  * though.
  */
-static void
-vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
+static int
+vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
+		 struct ww_acquire_ctx *acquire_ctx)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	uint64_t seqno;
 	unsigned long irqflags;
+	struct vc4_fence *fence;
+
+	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
+	if (!fence)
+		return -ENOMEM;
+	fence->dev = dev;
 
 	spin_lock_irqsave(&vc4->job_lock, irqflags);
 
 	seqno = ++vc4->emit_seqno;
 	exec->seqno = seqno;
+
+	dma_fence_init(&fence->base, &vc4_fence_ops, &vc4->job_lock,
+		       vc4->dma_fence_context, exec->seqno);
+	fence->seqno = exec->seqno;
+	exec->fence = &fence->base;
+
 	vc4_update_bo_seqnos(exec, seqno);
 
+	vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
+
 	list_add_tail(&exec->head, &vc4->bin_job_list);
 
 	/* If no job was executing, kick ours off.  Otherwise, it'll
@@ -509,6 +622,8 @@ vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
 	}
 
 	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	return 0;
 }
 
 /**
@@ -707,6 +822,12 @@ vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	unsigned i;
 
+	/* If we got force-completed because of GPU reset rather than
+	 * through our IRQ handler, signal the fence now.
+	 */
+	if (exec->fence)
+		dma_fence_signal(exec->fence);
+
 	if (exec->bo) {
 		for (i = 0; i < exec->bo_count; i++)
 			drm_gem_object_unreference_unlocked(&exec->bo[i]->base);
@@ -874,6 +995,7 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct drm_vc4_submit_cl *args = data;
 	struct vc4_exec_info *exec;
+	struct ww_acquire_ctx acquire_ctx;
 	int ret = 0;
 
 	if ((args->flags & ~VC4_SUBMIT_CL_USE_CLEAR_COLOR) != 0) {
@@ -916,12 +1038,18 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
+	ret = vc4_lock_bo_reservations(dev, exec, &acquire_ctx);
+	if (ret)
+		goto fail;
+
 	/* Clear this out of the struct we'll be putting in the queue,
 	 * since it's part of our stack.
 	 */
 	exec->args = NULL;
 
-	vc4_queue_submit(dev, exec);
+	ret = vc4_queue_submit(dev, exec, &acquire_ctx);
+	if (ret)
+		goto fail;
 
 	/* Return the seqno for our job. */
 	args->seqno = vc4->emit_seqno;
@@ -939,6 +1067,8 @@ vc4_gem_init(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 
+	vc4->dma_fence_context = dma_fence_context_alloc(1);
+
 	INIT_LIST_HEAD(&vc4->bin_job_list);
 	INIT_LIST_HEAD(&vc4->render_job_list);
 	INIT_LIST_HEAD(&vc4->job_done_list);