[Mesa-dev] [PATCH 1/2] r600: Fix possible endless loop in compute_memory_pool allocations.

[Bruno Jimenez]

Hi,

Thanks for catching this bug!

Reviewed-by: Bruno Jiménez <brunojimen at gmail.com>

Also, could you please send me a copy of the tests?

On Thu, 2014-06-19 at 10:21 -0400, Jan Vesely wrote:
> The important part is the change of the condition to <= 0. Otherwise the loop
> gets stuck never actually growing the pool.
> 
> The change in the aux-need calculation guarantees max 2 iterations, and
> avoids wasting memory in case a smaller item can't fit into a relatively larger
> pool.
> 
> Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> CC: Bruno Jimenez <brunojimen at gmail.com>
> ---
> 
> This fixes hang in gegl colors.xml test
> 
>  src/gallium/drivers/r600/compute_memory_pool.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/src/gallium/drivers/r600/compute_memory_pool.c b/src/gallium/drivers/r600/compute_memory_pool.c
> index ec8c470..0b6d2da6 100644
> --- a/src/gallium/drivers/r600/compute_memory_pool.c
> +++ b/src/gallium/drivers/r600/compute_memory_pool.c
> @@ -320,8 +320,11 @@ int compute_memory_finalize_pending(struct compute_memory_pool* pool,
>  			int64_t need = item->size_in_dw+2048 -
>  						(pool->size_in_dw - allocated);
>  
> -			if (need < 0) {
> -				need = pool->size_in_dw / 10;
> +			if (need <= 0) {
> +				/* There's enough free space, but it's too
> +				 * fragmented. Assume half of the item can fit
> +				 * int the last chunk */
> +				need = (item->size_in_dw / 2) + ITEM_ALIGNMENT;
>  			}
>  
>  			need = align(need, ITEM_ALIGNMENT);