Patch "drm/vc4: Fix races when the CS reads from render targets." has been added to the 4.8-stable tree

[<gregkh@linuxfoundation.org>]

This is a note to let you know that I've just added the patch titled

    drm/vc4: Fix races when the CS reads from render targets.

to the 4.8-stable tree which can be found at:
    http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary

The filename of the patch is:
     drm-vc4-fix-races-when-the-cs-reads-from-render-targets.patch
and it can be found in the queue-4.8 subdirectory.

If you, or anyone else, feels it should not be added to the stable tree,
please let <stable@vger.kernel.org> know about it.


>From 7edabee06a5622190d59689a64f5e17d1c343cc3 Mon Sep 17 00:00:00 2001
From: Eric Anholt <eric@anholt.net>
Date: Tue, 27 Sep 2016 09:03:13 -0700
Subject: drm/vc4: Fix races when the CS reads from render targets.

From: Eric Anholt <eric@anholt.net>

commit 7edabee06a5622190d59689a64f5e17d1c343cc3 upstream.

With the introduction of bin/render pipelining, the previous job may
not be completed when we start binning the next one.  If the previous
job wrote our VBO, IB, or CS textures, then the binning stage might
get stale or uninitialized results.

Fixes the major rendering failure in glmark2 -b terrain.

Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: ca26d28bbaa3 ("drm/vc4: improve throughput by pipelining binning and rendering jobs")
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

---
 drivers/gpu/drm/vc4/vc4_drv.h       |   19 ++++++++++++++++++-
 drivers/gpu/drm/vc4/vc4_gem.c       |   13 +++++++++++++
 drivers/gpu/drm/vc4/vc4_render_cl.c |   21 +++++++++++++++++----
 drivers/gpu/drm/vc4/vc4_validate.c  |   17 ++++++++++++++---
 4 files changed, 62 insertions(+), 8 deletions(-)

--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -122,9 +122,16 @@ to_vc4_dev(struct drm_device *dev)
 struct vc4_bo {
 	struct drm_gem_cma_object base;
 
-	/* seqno of the last job to render to this BO. */
+	/* seqno of the last job to render using this BO. */
 	uint64_t seqno;
 
+	/* seqno of the last job to use the RCL to write to this BO.
+	 *
+	 * Note that this doesn't include binner overflow memory
+	 * writes.
+	 */
+	uint64_t write_seqno;
+
 	/* List entry for the BO's position in either
 	 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
 	 */
@@ -216,6 +223,9 @@ struct vc4_exec_info {
 	/* Sequence number for this bin/render job. */
 	uint64_t seqno;
 
+	/* Latest write_seqno of any BO that binning depends on. */
+	uint64_t bin_dep_seqno;
+
 	/* Last current addresses the hardware was processing when the
 	 * hangcheck timer checked on us.
 	 */
@@ -230,6 +240,13 @@ struct vc4_exec_info {
 	struct drm_gem_cma_object **bo;
 	uint32_t bo_count;
 
+	/* List of BOs that are being written by the RCL.  Other than
+	 * the binner temporary storage, this is all the BOs written
+	 * by the job.
+	 */
+	struct drm_gem_cma_object *rcl_write_bo[4];
+	uint32_t rcl_write_bo_count;
+
 	/* Pointers for our position in vc4->job_list */
 	struct list_head head;
 
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -471,6 +471,11 @@ vc4_update_bo_seqnos(struct vc4_exec_inf
 	list_for_each_entry(bo, &exec->unref_list, unref_head) {
 		bo->seqno = 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;
+	}
 }
 
 /* Queues a struct vc4_exec_info for execution.  If no job is
@@ -673,6 +678,14 @@ vc4_get_bcl(struct drm_device *dev, stru
 		goto fail;
 
 	ret = vc4_validate_shader_recs(dev, exec);
+	if (ret)
+		goto fail;
+
+	/* Block waiting on any previous rendering into the CS's VBO,
+	 * IB, or textures, so that pixels are actually written by the
+	 * time we try to read them.
+	 */
+	ret = vc4_wait_for_seqno(dev, exec->bin_dep_seqno, ~0ull, true);
 
 fail:
 	drm_free_large(temp);
--- a/drivers/gpu/drm/vc4/vc4_render_cl.c
+++ b/drivers/gpu/drm/vc4/vc4_render_cl.c
@@ -45,6 +45,8 @@ struct vc4_rcl_setup {
 
 	struct drm_gem_cma_object *rcl;
 	u32 next_offset;
+
+	u32 next_write_bo_index;
 };
 
 static inline void rcl_u8(struct vc4_rcl_setup *setup, u8 val)
@@ -407,6 +409,8 @@ static int vc4_rcl_msaa_surface_setup(st
 	if (!*obj)
 		return -EINVAL;
 
+	exec->rcl_write_bo[exec->rcl_write_bo_count++] = *obj;
+
 	if (surf->offset & 0xf) {
 		DRM_ERROR("MSAA write must be 16b aligned.\n");
 		return -EINVAL;
@@ -417,7 +421,8 @@ static int vc4_rcl_msaa_surface_setup(st
 
 static int vc4_rcl_surface_setup(struct vc4_exec_info *exec,
 				 struct drm_gem_cma_object **obj,
-				 struct drm_vc4_submit_rcl_surface *surf)
+				 struct drm_vc4_submit_rcl_surface *surf,
+				 bool is_write)
 {
 	uint8_t tiling = VC4_GET_FIELD(surf->bits,
 				       VC4_LOADSTORE_TILE_BUFFER_TILING);
@@ -440,6 +445,9 @@ static int vc4_rcl_surface_setup(struct
 	if (!*obj)
 		return -EINVAL;
 
+	if (is_write)
+		exec->rcl_write_bo[exec->rcl_write_bo_count++] = *obj;
+
 	if (surf->flags & VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
 		if (surf == &exec->args->zs_write) {
 			DRM_ERROR("general zs write may not be a full-res.\n");
@@ -542,6 +550,8 @@ vc4_rcl_render_config_surface_setup(stru
 	if (!*obj)
 		return -EINVAL;
 
+	exec->rcl_write_bo[exec->rcl_write_bo_count++] = *obj;
+
 	if (tiling > VC4_TILING_FORMAT_LT) {
 		DRM_ERROR("Bad tiling format\n");
 		return -EINVAL;
@@ -599,15 +609,18 @@ int vc4_get_rcl(struct drm_device *dev,
 	if (ret)
 		return ret;
 
-	ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
+	ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read,
+				    false);
 	if (ret)
 		return ret;
 
-	ret = vc4_rcl_surface_setup(exec, &setup.zs_read, &args->zs_read);
+	ret = vc4_rcl_surface_setup(exec, &setup.zs_read, &args->zs_read,
+				    false);
 	if (ret)
 		return ret;
 
-	ret = vc4_rcl_surface_setup(exec, &setup.zs_write, &args->zs_write);
+	ret = vc4_rcl_surface_setup(exec, &setup.zs_write, &args->zs_write,
+				    true);
 	if (ret)
 		return ret;
 
--- a/drivers/gpu/drm/vc4/vc4_validate.c
+++ b/drivers/gpu/drm/vc4/vc4_validate.c
@@ -267,6 +267,9 @@ validate_indexed_prim_list(VALIDATE_ARGS
 	if (!ib)
 		return -EINVAL;
 
+	exec->bin_dep_seqno = max(exec->bin_dep_seqno,
+				  to_vc4_bo(&ib->base)->write_seqno);
+
 	if (offset > ib->base.size ||
 	    (ib->base.size - offset) / index_size < length) {
 		DRM_ERROR("IB access overflow (%d + %d*%d > %zd)\n",
@@ -555,8 +558,7 @@ static bool
 reloc_tex(struct vc4_exec_info *exec,
 	  void *uniform_data_u,
 	  struct vc4_texture_sample_info *sample,
-	  uint32_t texture_handle_index)
-
+	  uint32_t texture_handle_index, bool is_cs)
 {
 	struct drm_gem_cma_object *tex;
 	uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
@@ -714,6 +716,11 @@ reloc_tex(struct vc4_exec_info *exec,
 
 	*validated_p0 = tex->paddr + p0;
 
+	if (is_cs) {
+		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
+					  to_vc4_bo(&tex->base)->write_seqno);
+	}
+
 	return true;
  fail:
 	DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0);
@@ -835,7 +842,8 @@ validate_gl_shader_rec(struct drm_device
 			if (!reloc_tex(exec,
 				       uniform_data_u,
 				       &validated_shader->texture_samples[tex],
-				       texture_handles_u[tex])) {
+				       texture_handles_u[tex],
+				       i == 2)) {
 				return -EINVAL;
 			}
 		}
@@ -867,6 +875,9 @@ validate_gl_shader_rec(struct drm_device
 		uint32_t stride = *(uint8_t *)(pkt_u + o + 5);
 		uint32_t max_index;
 
+		exec->bin_dep_seqno = max(exec->bin_dep_seqno,
+					  to_vc4_bo(&vbo->base)->write_seqno);
+
 		if (state->addr & 0x8)
 			stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
 


Patches currently in stable-queue which might be from eric@anholt.net are

queue-4.8/drm-vc4-fix-races-when-the-cs-reads-from-render-targets.patch