[Mesa-dev] [PATCH v4 10/10] anv: Do relocations in userspace before execbuf ioctl

[Chris Wilson]

On Mon, Nov 07, 2016 at 02:27:54PM -0800, Jason Ekstrand wrote:
> From: Kristian Høgsberg Kristensen <kristian.h.kristensen at intel.com>
> 
> Since our surface state buffer is shared by all batches, the kernel does a
> full stall and sync with the CPU between batches every time we call
> execbuf2 because it refuses to do relocations on an active buffer.  Doing
> them in userspace and passing the NO_RELOC flag to the kernel allows us to
> perform the relocations without stalling.
> 
> This improves the performance of Dota 2 by around 30% on a Sky Lake GT2.
> 
> v2 (Jason Ekstrand):
>  - Better comments (Chris Wilson)
>  - Fixed write_reloc for correct canonical form (Chris Wilson)
> 
> v3 (Jason Ekstrand):
>  - Skip relocations which aren't needed
>  - Provide an environment variable to always use the kernel
>  - More comments about correctness (Chris Wilson)
> 
> v4 (Jason Ekstrand):
>  - More comments (Chris Wilson)
> 
> v5 (Jason Ekstrand):
>  - Rebase on top of moving execbuf2 setup go QueueSubmit
> 
> Signed-off-by: Jason Ekstrand <jason at jlekstrand.net>
> ---
> @@ -1134,14 +1238,48 @@ anv_cmd_buffer_execbuf(struct anv_device *device,
>        .rsvd2 = 0,
>     };
>  
> -   /* Since surface states are shared between command buffers and we don't
> -    * know what order they will be submitted to the kernel, we don't know what
> -    * address is actually written in the surface state object at any given
> -    * time.  The only option is to set a bogus presumed offset and let
> -    * relocations do their job.
> -    */
> -   for (size_t i = 0; i < cmd_buffer->surface_relocs.num_relocs; i++)
> -      cmd_buffer->surface_relocs.relocs[i].presumed_offset = -1;
> +   if (relocate_cmd_buffer(cmd_buffer, &execbuf)) {
> +      /* If we were able to successfully relocate everything, tell the kernel
> +       * that it can skip doing relocations. The requirement for using
> +       * NO_RELOC is:
> +       *
> +       *  1) The addresses written in the objects must match the corresponding
> +       *     reloc.presumed_offset which in turn must match the corresponding
> +       *     execobject.offset.
> +       *
> +       *  2) To avoid stalling, execobject.offset should match the current
> +       *     address of that object within the active context.
> +       *
> +       * In order to satisfy all of the invariants that make userspace
> +       * relocations to be safe (see relocate_cmd_buffer()), we need to
> +       * further ensure that the addresses we use match those used by the
> +       * kernel for the most recent execbuf2.
> +       *
> +       * If we were able to successfully relocate everything, tell the kernel
> +       * that it can skip doing relocations.

Duplicate (with opening) sentence.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre