drm/vc4: Fix races when the CS reads from render targets.
authorEric Anholt <eric@anholt.net>
Tue, 27 Sep 2016 16:03:13 +0000 (09:03 -0700)
committerEric Anholt <eric@anholt.net>
Thu, 6 Oct 2016 18:53:50 +0000 (11:53 -0700)
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")
Cc: stable@vger.kernel.org
drivers/gpu/drm/vc4/vc4_drv.h
drivers/gpu/drm/vc4/vc4_gem.c
drivers/gpu/drm/vc4/vc4_render_cl.c
drivers/gpu/drm/vc4/vc4_validate.c

index 61c1902168a1fa6e10b8629d46eb4e01ac3a43a3..7c1e4d97486fb57d9ce7002c30a23138519761ae 100644 (file)
@@ -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;
 
index 77daea6cb8668df85d70d4b760f4e5a5be41800a..47a095f392f8a7b5443227933c78b500a4c7137f 100644 (file)
@@ -467,6 +467,11 @@ vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
        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
@@ -669,6 +674,14 @@ vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
                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);
index 0f12418725e552f7378dada1d5af1244f52599ad..08886a3097577242f5c9e025fd6446d81bcc6dec 100644 (file)
@@ -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(struct vc4_exec_info *exec,
        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(struct vc4_exec_info *exec,
 
 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 vc4_exec_info *exec,
        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(struct vc4_exec_info *exec,
        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, struct vc4_exec_info *exec)
        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;
 
index 9ce1d0adf8826e893a58529328ce3f53f05150ee..26503e307438a34fe526222c8c15be158eb332a2 100644 (file)
@@ -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 *dev,
                        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 *dev,
                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;