[Mesa-dev] [PATCH 00/13] mesa: Locking improvements and optimizations

[Brian Paul]

On 08/06/2015 06:10 PM, Matt Turner wrote:
> Patches 1-11 improve performance of SynMark OglBatch7 by 6.29586% +/- 0.277734%
> (n=337) and OglMultithread by 1.12564% +/- 0.424038% (n=209). I haven't
> benchmarked individual patches because I'd like to not waste all that time if I
> get review feedback that requires me to change things. :)
>
> Patches 12-13 were supposed to improve performance, but seem to make an
> existing thread-safety problem worse, so I'm not proposing them for inclusion.
>
>
> [01/13] c11/threads: Assert that mtx is non-NULL and check
> [02/13] mesa: Remove debugging code from _mesa_reference_*.
>
> [03/13] mesa: Add locking to sampler objects.
> [04/13] mesa: Add locking to programs.
>
>     These two add missing locks to sampler and program objects, which
>     I believe are supposed to be thread-safe.
>
> [05/13] mesa: Replace buffer object locks with atomic inc/dec.
> [06/13] mesa: Replace sampler object locks with atomic inc/dec.
> [07/13] mesa: Replace program locks with atomic inc/dec.
> [08/13] mesa: Replace renderbuffer object locks with atomic inc/dec.
> [09/13] mesa: Replace texture buffer object locks with atomic inc/dec.
>
>     These five replace locks around RefCount++/-- with atomic increment
>     and decrement.
>
> [10/13] hash: Add _mesa_HashRemoveLocked() function.
> [11/13] mesa: Replace uses of Shared->Mutex with hash-table mutexes
>
>     These two replace uses of ctx->Shared->Mutex with the mutexes in the
>     hash tables.

1-13 look good to me, just a minor nit on #10.

I'd be OK w/ squashing 3 & 6, and 4 & 7, but not a big deal.

For the series,
Reviewed-by: Brian Paul <brianp at vmware.com>


>
> [12/13] mesa: Remove unnecessary locking from container
> [13/13] mesa: Remove deleteFlag pattern.
>
>     *I am not proposing these for inclusion*
>
>     These two remove some "unnecessary" locking from so called "container
>     objects" that are not shared between threads by the GL. While I expected
>     them to improve performance, they actually cause double-free errors in
>     SynMark OglMultithread. Valgrind's helgrind tool shows that there are many
>     thread-safety issues in the texture code, and removing these locks seems to
>     exacerbate the problem.
>
>     Specifically, multiple threads are reading and writing to gl_texture_objects
>     without any synchronization from places like intel_finalize_mipmap_tree(),
>     gen7_update_texture_surface(), brw_populate_sampler_prog_key_data(),
>     update_sampler_state(), and _mesa_BindTexture().
>
>     Suggestions for solving this (apparently quite longstanding) problem are
>     welcome.

gl_texture_object appears all over the place.  Adding locking everywhere 
could be a nightmare but I'm not sure what other solution there is off-hand.

-Brian