On 17/05/17 22:50, Nicolai Hähnle wrote:
On 17.05.2017 14:26, Timothy Arceri wrote:
On 17/05/17 22:03, Nicolai Hähnle wrote:
On 17.05.2017 01:49, Timothy Arceri wrote:
This might work for gallium based drivers but I'm pretty sure this will
cause problems for the i965 fallback path.

What do those problems look like?

Sorry, I meant to add more details. The problem is that when the i965
driver has a cache miss for a variant this could be happening long after
link time. By this point we already have a reference and we also can't
just reallocate the shader data because uniforms etc have most likely
been set by this point.

Oh I see, thanks for the explanation.


And any ideas on how to address the issue then? There are a number of
places that assume prog->sh.data is non-NULL.

I'll need to take a closer look at the problem tomorrow. I'm not sure I
understand why it is NULL, I thought the reference happens in the GLSL
IR cache. Is it getting cleared somewhere by the tgsi pass?

As far as I understand, the gl_program object doesn't come from the cache in the fallback path. So the reference should happen in the place modified by this patch.

I see what has happened, this patch made it in but some other fallback patches that should really go along with it were left in the i965 branch.

https://cgit.freedesktop.org/~jljusten/mesa/commit/?h=i965-shader-cache&id=5d6c2450fe48e777f6c57d9b70653832afa9bccc
https://cgit.freedesktop.org/~jljusten/mesa/commit/?h=i965-shader-cache&id=b832afe2c52c946a77b213b94850a765bbaf36af
https://cgit.freedesktop.org/~jljusten/mesa/commit/?h=i965-shader-cache&id=d8e8e39cc8b83393eac1bc32c017d312fd9af082
https://cgit.freedesktop.org/~jljusten/mesa/commit/?h=i965-shader-cache&id=919ca263e258142aa8649b2b1c640368c1c4bab9

I can see two solutions:

1. Make use of the mesa_build_fallback_shader_program() helper Jordan has created (patch 2 above). This would probably allow us to skip some things in the glsl to tgsi pass on fallback.

2. Stop setting cache_fallback in st_load_tgsi_from_disk_cache(). This should be safe because unlike i965 we are always falling back at link time. We will unnecessarily free and reallocate a bunch of things here but it's probably not a huge deal.

Either way we should probably turn the env var in patch 4 above into a MESA_GLSL param so that we can force the fallback path for easy testing with piglit. I think you will find with this current patch that uniforms (among other things) will never be allocated storage on fallback because we skip that also.

For stable I think we might just want to land option 2, once we do some testing. I think option 1 would need a bit more time sitting in master before I was comfortable with it.

Thanks for spotting this, it was an oversight in testing on my behalf.

I've send out a patch for the env var and wait for your thoughts before doing anything further.

Thanks,
Tim


Cheers,
Nicolai



Cheers,
Nicolai



On 17/05/17 04:40, Nicolai Hähnle wrote:
From: Nicolai Hähnle <nicolai.haeh...@amd.com>

This reverts commit 0e9991f957e296f46cfff40a94ffba0adf2a58e1.

At least when we fallback because of TGSI, the gl_program objects are
re-allocated, and we don't in fact already have a reference to the
gl_shader_program_data.

This fixes a crash that can be reproduced as follows:

1. Run any GL program with MESA_GLSL=cache_info.
2. Note the SHA of some tgsi_tokens.
3. Remove the corresponding file from ~/.cache/mesa and re-run.

Cc: 17.1 <mesa-sta...@lists.freedesktop.org>
---
  src/compiler/glsl/linker.cpp | 3 +--
  src/mesa/main/shaderobj.c    | 3 +--
  2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/src/compiler/glsl/linker.cpp
b/src/compiler/glsl/linker.cpp
index 2e7dd2b..4776d09 100644
--- a/src/compiler/glsl/linker.cpp
+++ b/src/compiler/glsl/linker.cpp
@@ -2243,22 +2243,21 @@ link_intrastage_shaders(void *mem_ctx,
     struct gl_program *gl_prog =
        ctx->Driver.NewProgram(ctx,

_mesa_shader_stage_to_program(shader_list[0]->Stage),
                               prog->Name, false);
     if (!gl_prog) {
        prog->data->LinkStatus = linking_failure;
        _mesa_delete_linked_shader(ctx, linked);
        return NULL;
     }
  -   if (!prog->data->cache_fallback)
-      _mesa_reference_shader_program_data(ctx, &gl_prog->sh.data,
prog->data);
+   _mesa_reference_shader_program_data(ctx, &gl_prog->sh.data,
prog->data);
       /* Don't use _mesa_reference_program() just take ownership */
     linked->Program = gl_prog;
       linked->ir = new(linked) exec_list;
     clone_ir_list(mem_ctx, linked->ir, main->ir);
       link_fs_inout_layout_qualifiers(prog, linked, shader_list,
num_shaders);
     link_tcs_out_layout_qualifiers(prog, gl_prog, shader_list,
num_shaders);
     link_tes_in_layout_qualifiers(prog, gl_prog, shader_list,
num_shaders);
diff --git a/src/mesa/main/shaderobj.c b/src/mesa/main/shaderobj.c
index 8a5fa5e..f6f7bf5 100644
--- a/src/mesa/main/shaderobj.c
+++ b/src/mesa/main/shaderobj.c
@@ -427,22 +427,21 @@ _mesa_free_shader_program_data(struct gl_context
*ctx,
      /**
   * Free/delete a shader program object.
   */
  void
  _mesa_delete_shader_program(struct gl_context *ctx,
                              struct gl_shader_program *shProg)
  {
     _mesa_free_shader_program_data(ctx, shProg);
-   if (!shProg->data->cache_fallback)
-      _mesa_reference_shader_program_data(ctx, &shProg->data, NULL);
+   _mesa_reference_shader_program_data(ctx, &shProg->data, NULL);
     ralloc_free(shProg);
  }
      /**
   * Lookup a GLSL program object.
   */
  struct gl_shader_program *
  _mesa_lookup_shader_program(struct gl_context *ctx, GLuint name)
  {

_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev




_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to