[Mesa-dev] [PATCH 1/2] eg/compute: Use reference counting to handle compute memory pool.

[Nicolai Hähnle]

On 04.05.2018 08:34, Jan Vesely wrote:
> The original bug/corruption was by util_unreference_framebuffer_state,
> trying to drop reference on cbuf[0] (global AS for OCL).
> Adding reference counting to set_rat uncovered problems with acessing the pool->bo.
> 
> Also drops leaked memory from 7,4kB to 1,7Kb in single run of conformance_bruteforce trunc -w.
> 
> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> ---
>   src/gallium/drivers/r600/compute_memory_pool.c | 18 ++++++------------
>   src/gallium/drivers/r600/evergreen_compute.c   |  3 ++-
>   2 files changed, 8 insertions(+), 13 deletions(-)
> 
> diff --git a/src/gallium/drivers/r600/compute_memory_pool.c b/src/gallium/drivers/r600/compute_memory_pool.c
> index bcda155c71..155fefe54f 100644
> --- a/src/gallium/drivers/r600/compute_memory_pool.c
> +++ b/src/gallium/drivers/r600/compute_memory_pool.c
> @@ -91,10 +91,7 @@ void compute_memory_pool_delete(struct compute_memory_pool* pool)
>   {
>   	COMPUTE_DBG(pool->screen, "* compute_memory_pool_delete()\n");
>   	free(pool->shadow);
> -	if (pool->bo) {
> -		pool->screen->b.b.resource_destroy((struct pipe_screen *)
> -			pool->screen, (struct pipe_resource *)pool->bo);
> -	}
> +	pipe_resource_reference(&pool->bo, NULL);
>   	/* In theory, all of the items were freed in compute_memory_free.
>   	 * Just delete the list heads
>   	 */
> @@ -213,11 +210,9 @@ int compute_memory_grow_defrag_pool(struct compute_memory_pool *pool,
>   
>   			compute_memory_defrag(pool, src, dst, pipe);
>   
> -			pool->screen->b.b.resource_destroy(
> -					(struct pipe_screen *)pool->screen,
> -					src);
> -
> -			pool->bo = temp;
> +			pipe_resource_reference(&pool->bo, dst);
> +			/* Drop the extra reference from creation */
> +			pipe_resource_reference(&dst, NULL);

I believe you could just say

		pool->bo = dst;
		dst = NULL;

here, and save two atomic ops.


>   			pool->size_in_dw = new_size_in_dw;
>   		}
>   		else {
> @@ -230,9 +225,8 @@ int compute_memory_grow_defrag_pool(struct compute_memory_pool *pool,
>   				return -1;
>   
>   			pool->size_in_dw = new_size_in_dw;
> -			pool->screen->b.b.resource_destroy(
> -					(struct pipe_screen *)pool->screen,
> -					(struct pipe_resource *)pool->bo);
> +			pipe_resource_reference(&pool->bo, NULL);
> +			/* We already have one reference */

I find this comment confusing.

The rest looks good to me.

Cheers,
Nicolai


>   			pool->bo = r600_compute_buffer_alloc_vram(pool->screen, pool->size_in_dw * 4);
>   			compute_memory_shadow(pool, pipe, 0);
>   
> diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
> index 6bec1a333f..58626a3114 100644
> --- a/src/gallium/drivers/r600/evergreen_compute.c
> +++ b/src/gallium/drivers/r600/evergreen_compute.c
> @@ -122,7 +122,8 @@ static void evergreen_set_rat(struct r600_pipe_compute *pipe,
>   	rat_templ.u.tex.first_layer = 0;
>   	rat_templ.u.tex.last_layer = 0;
>   
> -	/* Add the RAT the list of color buffers */
> +	/* Add the RAT the list of color buffers. Drop the old buffer first. */
> +	pipe_surface_reference(&pipe->ctx->framebuffer.state.cbufs[id], NULL);
>   	pipe->ctx->framebuffer.state.cbufs[id] = pipe->ctx->b.b.create_surface(
>   		(struct pipe_context *)pipe->ctx,
>   		(struct pipe_resource *)bo, &rat_templ);
> 


-- 
Lerne, wie die Welt wirklich ist,
Aber vergiss niemals, wie sie sein sollte.