bounds checks for texture-buffer copies happen too late

[jimblandy]

Methods like command_encoder_copy_texture_to_buffer do check that ImageCopyTexture::mip_level field of their source and/or destination parameters is valid for the relevant texture, by calling validate_texture_copy_range. However, those checks occur after those functions have called TextureTracker::set_single, which assumes that the mip_level is in range.

The specific call chain is:

• Global::command_encoder_copy_texture_to_buffer, or related
• TextureTracker::set_single
• wgpu_core::track::texture::insert_or_barrier_single
• wgpu_core::track::texture::insert
• ComplexTextureState::from_selector_state_iter

This last uses get_unchecked_mut on selector.mips.start.

cc @cwfitzgerald

[jimblandy]

I don't understand all the plumbing here yet, but it seems like we ought to be able to do the validation in extract_texture_selector instead?

[jimblandy]

Current plan is to fuse extract_texture_selector and validate_texture_copy_range into a single function which does all bounds checks, and call that before all resource tracking operations.

[jimblandy]

for the curious, work in progress is here: https://github.com/jimblandy/wgpu/tree/early-texture-bounds-checks

[jimblandy]

This specific case happens to not affect copy_buffer_to_texture, because that calls wgpu_core::command::transfer::handle_dst_texture_init, which calls has_copy_partial_init_tracker_coverage, which happens to check the destination mip level against the texture itself. The corresponding function for copies from textures is handle_src_texture_init, which doesn't make any such call.

[jimblandy]

@nical was working on this; may be fixed now.

[andyleiserson]

This appears to be fixed, although I have found some other validation for texture transfers that is missing, so I'm not going to close the bug just yet.

writeTexture and all three copy{Texture,Buffer}To{Texture,Buffer} methods call validate_texture_copy_range early, before calling TextureTracker::set_single. The CTS tests webgpu:api,validation,encoding,cmds,copyTextureToTexture:mipmap_level:* are passing (although I am not sure they would have caught the original problem, as long as the check happened eventually).