Crash when allocating unreasonable amounts of texture memory

[nical]

Track texture allocations and crash if we go beyond a certain budget. The budget (2Gb by default) is configurable in the renderer options. I'll experiment with various thresholds in gecko (which probably need to take into account the window size) when this hits mozilla-central.

Fixes #3207.

---

This change is 

[gw3583]

Looks like some of the PF code needs updating:

error[E0308]: mismatched types
  --> webrender\src\gpu_glyph_renderer.rs:66:41
   |
66 |         device.upload_texture_immediate(&area_lut_texture, area_lut_pixels);
   |                                         ^^^^^^^^^^^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
   |
   = note: expected type `&device::gl::Texture`
              found type `&std::result::Result<device::gl::Texture, device::gl::AllocError>`
error[E0308]: mismatched types
  --> webrender\src\gpu_glyph_renderer.rs:90:13
   |
90 |             area_lut_texture,
   |             ^^^^^^^^^^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
   |
   = note: expected type `device::gl::Texture`
              found type `std::result::Result<device::gl::Texture, device::gl::AllocError>`
error[E0308]: mismatched types
   --> webrender\src\gpu_glyph_renderer.rs:129:13
    |
129 |             texture,
    |             ^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
    |
    = note: expected type `device::gl::Texture`
               found type `std::result::Result<device::gl::Texture, device::gl::AllocError>`
error[E0308]: mismatched types
   --> webrender\src\gpu_glyph_renderer.rs:191:46
    |
191 |         self.device.upload_texture_immediate(&path_info_texture, &path_info_texels);
    |                                              ^^^^^^^^^^^^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
    |
    = note: expected type `&device::gl::Texture`
               found type `&std::result::Result<device::gl::Texture, device::gl::AllocError>`
error[E0308]: mismatched types
   --> webrender\src\gpu_glyph_renderer.rs:229:60
    |
229 |         self.device.bind_texture(TextureSampler::color(1), &path_info_texture);
    |                                                            ^^^^^^^^^^^^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
    |
    = note: expected type `&device::gl::Texture`
               found type `&std::result::Result<device::gl::Texture, device::gl::AllocError>`
error[E0308]: mismatched types
   --> webrender\src\gpu_glyph_renderer.rs:234:36
    |
234 |         self.device.delete_texture(path_info_texture);
    |                                    ^^^^^^^^^^^^^^^^^ expected struct `device::gl::Texture`, found enum `std::result::Result`
    |
    = note: expected type `device::gl::Texture`
               found type `std::result::Result<device::gl::Texture, device::gl::AllocError>`
error: aborting due to 6 previous errors
For more information about this error, try `rustc --explain E0308`.

[bholley]

I think it'd be preferable to derive(Display) on MemoryReport and use that here, so that any additional fields we add will automatically get picked up here.

[nical]

I fixed the failing rawtest and added a commit that prints a message and skips the rawtests that rely on a the device pixel ratio being 1.0 when it is not the case to avoid having to comment them out locally.

[nical]

Try push: https://treeherder.mozilla.org/#/jobs?repo=try&revision=965f427b69641a989091c8f5eedb2c981e061559

[gw3583]

Generally looks good to me - should we change the unwrap calls to expect to provide a bit more context on any crashes that do occur?

[gw3583]

Are there any other places in the code we re-size / re-init textures that we need to account for the change in allocation size?

[nical]

I didn't find any. Our code to resize texture currently destroys and re-creates bigger textures (with possibly a blit from the old texture to the new one.

[bholley]

Yeah, this was a change I had to make for immutable storage. Texture re-initialization is no more.

[gw3583]

It would be neat to have both the current budget and allocated memory available as a profile counter in the debug profiler overlay. This would allow us to get a sense of total texture allocations on pages as we browse, and make sure that the counters are working correctly.

[nical]

Since the budget is there to force a crash when things are so wrong that it's best not to keep the browser alive, I don't think it is very interesting to display.
That said I think that displaying some number that we think is reasonable for an average page, or some number beyond which we believe something worth investigating (but not to the point that we'd crash the browser) would be useful, as well as displaying the current allocated texture memory.

[kvark]

That try push needs another try

[nical]

That try push needs another try

Ah, yes indeed, there we go: https://treeherder.mozilla.org/#/jobs?repo=try&revision=56a80fb4e9602775ab0530099c9c443ba43b26fa

[gw3583]

@nical Looks like this is ready to merge now, is that right?

[bholley]

Per discussion on IRC late last week, we should rebase it on top of [1] when that lands to avoid having multiple mechanisms for tracking GPU memory usage.

[1] https://bugzilla.mozilla.org/show_bug.cgi?id=1504115

[bors-servo]

☔ The latest upstream changes (presumably #3269) made this pull request unmergeable. Please resolve the merge conflicts.

[gw3583]

Looks like the bug mentioned above has landed - so I guess this needs a rebase on top of that?

[gw3583]

@nical @bholley Do we just need to get this rebased and then it's ready to land?

[bholley]

It just needs to be rebased to use the global atomic byte count rather than tracking it per-device. That should be trivial, then it can land.

[gw3583]

ping @nical - should we get this rebased and landed?

[nical]

Sorry for the delay, I need to more or less re-write the PR so I'll close it for now.