From: Boris Brezillon Date: Thu, 10 Oct 2019 13:12:30 +0000 (+0200) Subject: panfrost: Fix the DISCARD_WHOLE_RES case in transfer_map() X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=7fa5cd3ee365ed35c5799b96d7867b6ebded4fdd;p=mesa.git panfrost: Fix the DISCARD_WHOLE_RES case in transfer_map() The current implementation does not synchronize on BO readiness when DISCARD_WHOLE_RES flag is set, which can lead to misbehaviours when the resource being updated is being used by one of the pending or already flushed batches. Adding unconditional BO synchronization would do the trick, but we can sometimes optimize this path by re-allocating a new BO instead of waiting for the existing one to be ready. Reported-by: Daniel Stone Reported-by: Heinrich Fink Signed-off-by: Boris Brezillon Reviewed-by: Daniel Stone Reviewed-by: Alyssa Rosenzweig --- diff --git a/src/gallium/drivers/panfrost/pan_job.c b/src/gallium/drivers/panfrost/pan_job.c index a10d4a37528..9cff5232519 100644 --- a/src/gallium/drivers/panfrost/pan_job.c +++ b/src/gallium/drivers/panfrost/pan_job.c @@ -984,6 +984,30 @@ panfrost_flush_all_batches(struct panfrost_context *ctx, bool wait) util_dynarray_fini(&syncobjs); } +bool +panfrost_pending_batches_access_bo(struct panfrost_context *ctx, + const struct panfrost_bo *bo) +{ + struct panfrost_bo_access *access; + struct hash_entry *hentry; + + hentry = _mesa_hash_table_search(ctx->accessed_bos, bo); + access = hentry ? hentry->data : NULL; + if (!access) + return false; + + if (access->writer && access->writer->batch) + return true; + + util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *, + reader) { + if (*reader && (*reader)->batch) + return true; + } + + return false; +} + void panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx, struct panfrost_bo *bo, diff --git a/src/gallium/drivers/panfrost/pan_job.h b/src/gallium/drivers/panfrost/pan_job.h index 1da5d57aec1..2a810421caa 100644 --- a/src/gallium/drivers/panfrost/pan_job.h +++ b/src/gallium/drivers/panfrost/pan_job.h @@ -188,6 +188,10 @@ panfrost_batch_create_bo(struct panfrost_batch *batch, size_t size, void panfrost_flush_all_batches(struct panfrost_context *ctx, bool wait); +bool +panfrost_pending_batches_access_bo(struct panfrost_context *ctx, + const struct panfrost_bo *bo); + void panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx, struct panfrost_bo *bo, uint32_t flags); diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c index db0643c7305..d0619a89882 100644 --- a/src/gallium/drivers/panfrost/pan_resource.c +++ b/src/gallium/drivers/panfrost/pan_resource.c @@ -587,8 +587,41 @@ panfrost_transfer_map(struct pipe_context *pctx, /* TODO: Respect usage flags */ if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) { - /* TODO: reallocate */ - //printf("debug: Missed reallocate\n"); + /* If the BO is used by one of the pending batches or if it's + * not ready yet (still accessed by one of the already flushed + * batches), we try to allocate a new one to avoid waiting. + */ + if (panfrost_pending_batches_access_bo(ctx, bo) || + !panfrost_bo_wait(bo, 0, PAN_BO_ACCESS_RW)) { + struct panfrost_screen *screen = pan_screen(pctx->screen); + /* We want the BO to be MMAPed. */ + uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP; + struct panfrost_bo *newbo = NULL; + + /* When the BO has been imported/exported, we can't + * replace it by another one, otherwise the + * importer/exporter wouldn't see the change we're + * doing to it. + */ + if (!(bo->flags & (PAN_BO_IMPORTED | PAN_BO_EXPORTED))) + newbo = panfrost_bo_create(screen, bo->size, + flags); + + if (newbo) { + panfrost_bo_unreference(bo); + rsrc->bo = newbo; + bo = newbo; + } else { + uint32_t access = PAN_BO_ACCESS_RW; + + /* Allocation failed or was impossible, let's + * fall back on a flush+wait. + */ + panfrost_flush_batches_accessing_bo(ctx, bo, + access); + panfrost_bo_wait(bo, INT64_MAX, access); + } + } } else if ((usage & PIPE_TRANSFER_WRITE) && resource->target == PIPE_BUFFER && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) {