As started in https://github.com/servo/servo/issues/42168 let's report
memory pressure for webgl objects. CanvasContexts report memory twice:
once because of underlying texture object and once by themself, but
that's okay as we also need to account for swapchain textures.
Computing exact size would be to much work and code so we report rough
estimations.
Testing: None
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Reviewable per commits:
As noted in
https://github.com/servo/servo/pull/42180#discussion_r2749861902
transmuting `&Reflector<AssociatedMemory>` to `&Reflector<()>` will
ignore AssociatedMemory information and thus it will not remove extra
associated memory. By returning `&Reflector<Self::ReflectorType>` from
`DomObject::reflector()` we will preserve this information and thus
correctly handle cases with associated memory. We also do not need
`overrideMemoryUsage` in bindings.conf anymore. 🎉
Instead of removing associated memory in drop code we should do it as
part of finalizers, otherwise we have problems in nested (inherited)
structs, where drop is run for both parent and child, but we only added
associated memory for child (on init_reflector) which already included
size of parent. The only exception here is promise, because it is RCed
and not finalized.
Testing: Tested locally that it fixes speedometer.
Fixes#42269
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
By reporting memory usage of rust objects back to SM it can trigger GC
sooner and release more memory (I hope that we could use this to avoid
OOB in webgpu in the future).
Currently we use simple `size_of::<DomStruct>() +
size_of::<Box<DomStruct>>()` but it's possible to override reported
memory with `update_memory_size(self, nbytes);` on reflector. This is
mostly useful for canvases and buffers which usually have large
associated data.
Testing: None, but I am wondering if we can connect this to
`about:memory`.
Fixes: #42168
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This moves WebGL component from IpcChannel to GenericChannel. Of note
are the following changes:
- Removal of WebGlSender and WebGlReceiver type alias. As we removed
them in a similar component earlier.
- Moving from IpcByteSender to IpcSender over serde_bytes::ByteBuf. I
think the tradeoff for easier reading is worth it.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Testing: Compilation and unit tests.
---------
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Co-authored-by: Sam <16504129+sagudev@users.noreply.github.com>
Refactors the WebGLRenderingContext to manage resource cleanup via a
dedicated `DroppableWebGLRenderingContext` struct.
Testing: This task is only a refactor, no tests added
Fixes: partially #26488
Signed-off-by: Domenico Rizzo <domenico.rizzo@gmail.com>
Changed most #[allow]s to #[expect]s, mainly for
clippy::too_many_arguments
Removed unfulfilled expectations
This is my first opensource contribution, so please let me know if
anything should
be done differently.
Testing: Refactor
Part of: #40838
---------
Signed-off-by: Tim van der Lippe <tvanderlippe@gmail.com>
Signed-off-by: TimurBora <timurborisov5561@gmail.com>
Co-authored-by: Tim van der Lippe <TimvdLippe@users.noreply.github.com>
Replace `allow` with `expect` lints for `unused`, `unsafe_code`,
`dead_code`, and `non_upper_case_globals`.
Testing: So far just check it compiled on `x86_64-linux` on NixOS. Need
to use the module `system.fontconfig.enable = true;` I think in my NixOS
config.
Part of: #40383
Searching `allow\(.*\)` for `.rs` files shows the following. for
(total_results:total_files) went from `707:386` to `675:368`, a
reduction of `32:18`.
How many files is too many files per PR? I feel like the 20-30 I have is
too big.
---------
Signed-off-by: Wayne Van Son <waynevanson@gmail.com>
In the future, servo components should depend on the generic channels in
base instead of IpcChannels to correctly optimize for multiprocess vs
non-multiprocess mode.
This reexports IpcSharedMemory as GenericSharedMemory in GenericChannel
and changes all dependencies on it.
Currently this is only a type/name change and does not change
functionality. But in the future we would want want to use non-ipc
things for the data.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Testing: This changes types and does not need testing.
---------
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Changed some allow to expects and removed the unfulfilled expectations.
Testing: Refactor
Part of: #40383
Signed-off-by: anonmiraj <nabilmalek48@gmail.com>
I used find and replace to finish the job. All this PR does is replace
all `Error::<error_name>` occurrences with `Error::<error_name>(None)`.
Testing: Refactor
Fixes: #39053
Signed-off-by: Ashwin Naren <arihant2math@gmail.com>
Many WebGL objects refer to a `WebGLRenderingContext` and rely on it for
messaging to the `WebGLThread`. This poses a problem, because WebGL
objects often need to send a message to the `WebGLThread` during their
`Drop` implementation. If the `Drop` is triggered as part of garbage
collection, references to the `WebGLRenderingContext` might be invalid,
if they were garbage collected first as part of the same harvest.
This change makes it so that all of these objects store a `WeakRef`
instead of a `Dom<>`. The `WeakRef` is only used if it can be rooted,
otherwise a `ContextLost` error is given. In cases where only messaging
is needed, a cloned `WebGLMsgSender` is used to perform messages
regardless of whether the context is garbage collected or not.
This isn't a replacement for #37622, but should make it easier to
implement as
the `WebGLMsgSender` and the `WeakRef` could be stored in the droppable
portion of the DOM object.
Testing: This fixes a use-after-free issue which is mainly detectable
via ASAN
builds. Since we do not run ASAN on CI, this is a bit hard to create
automated
tests for. I verified that this fixed the issue manually.
Fixes: #40655.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Currently, a single `surfman::Device` is used for all WebGL contexts.
This cannot work when we have multiple rendering contexts. So, extract
the surfman `Connection` and `Adapter` into a per-painter data structure
(`PainterSurfmanDetailsMap`) and store the `Device` directly in the
WebGL context.
This patch also modifies the WebXR traits so that the `Device` doesn't
need to be explictly passed into most methods.
*This is a reland of #40594* with the following changes:
- Do not remove WebGL contexts from the context map before cleaning up
the
WebXR layers. The layer cleanup process consults the map.
- When cleanup up layers, be sure to replace the WebXRBridge in the
WebGLThread data structure.
- Allow failing to the Device when processing WebXR commands. WebXR
sometimes tries to access contexts after they have been removed.
Testing: Should be covered by existing tests.
---------
Signed-off-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Currently, a single `surfman::Device` is used for all WebGL contexts.
This cannot work when we have multiple rendering contexts. So, extract
the surfman `Connection` and `Adapter` into a per-painter data structure
(`PainterSurfmanDetailsMap`) and store the `Device` directly in the
WebGL context.
This patch also modifies the WebXR traits so that the `Device` doesn't
need to be explictly passed into most methods.
Testing: Should be covered by existing tests.
---------
Signed-off-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
This is last step toward enabling the default rustc
`unsafe_op_in_unsafe_fn` warning for the script crate. It wraps the
remaining unsafe code in `unsafe {}` and removes the line disabling this
warning from `script`'s `Cargo.toml`. In addition, two variables are
renamed from `v` to something slightly more descriptive.
Testing: This should not change behavior so is covered by existing
tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
CanvasContext: follow spec for isSync and isTransformFeedback
Testing: more CanvasContext test will pass.
Signed-off-by: Taym Haddadi <haddadi.taym@gmail.com>
Removes some unneeded lints, especially `#[allow(unsafe_code)]`.
Testing: Refactor
Part of: #40383
Signed-off-by: WaterWhisperer <waterwhisperer24@qq.com>
This change makes it so that the `HTMLCanvasElement` is responsible for
managing the `ImageKey` for associated `RenderingContext`s. Only
canvases display their contents into WebRender directly, so this makes
it so that keys are not generated for `OffscreenCanvas`.
The main goal here is that `ImageKey`s are always associated with a
particular `WebView`, which isn't possible in the various canvas
backends yet. This is important because each `WebView` may soon have a
different WebRender instance entirely with its own set of `ImageKey`s.
This also allows for clearing `ImageKey`s when canvases are disconnected
from the DOM in a future change. One tricky thing here is placeholder
canvases, which are meant to be driven from workers.
It seems that the implementation isn't correct for these at the moment
as they need to be updated to the specification. Instead, what is
happening is that any existing context / image is completely lost when
converting to an `OffscreenCanvas`.
Testing: This shouldn't change observable behavior, so is covered by
existing tests.
Fixes: This is part of #40261.
---------
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Signed-off-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Servo has a lot of comments like this:
```rust
// https://example-spec.com/#do-the-thing
fn do_the_thing() {}
```
and I keep turning these into doc comments whenever I'm working close to
one of them. Doing so allows me to hover over a function call in an IDE
and open its specification without having to jump to the function
definition first. This change fixes all of these comments at once.
This was done using `find components -name '*.rs' -exec perl -i -0777
-pe 's|^([ \t]*)// (https?://.*)\n\1(fn )|\1/// <$2>\n\1$3|mg' {} +`.
Note that these comments should be doc comments even within trait `impl`
blocks, because rustdoc will use them as fallback documentation when the
method definition on the trait does not have documentation.
Testing: Comments only, no testing required
Preparation for https://github.com/servo/servo/pull/39552
---------
Signed-off-by: Simon Wülker <simon.wuelker@arcor.de>
Canvas expects input data to be in RGBA premultiplied format and
WebRender already supports RGBA and BGRA data as long as they are
premultiplied. Pre-multiplying up front allows:
- Avoiding many conversions when painting images to canvas.
- Passing the `RasterImage` IpcSharedMemory of the image instead of
creating
a new one with the premultiplied data every time we upload to
WebRender. This is a big deal for animated gifs, because before every
frame was creating a new shared memory segment.
It seems that for rasterized SVGs were were already putting
premultplied data into the `RasterImage` so it's quite likely SVGs were
being composited incorrectly.
Testing: This causes 8 tests to start passing and 2 tests to fail in the
WebGL conformance suite.
The failures are due to the fact that premultiplying alpha is lossy when
alpha is 0. In that case,
the resulting color of a blend operation might be wrong. This is
typically only a problem if you
use RGBA data as RGB data, which is pretty unusual. In the case that you
are blending with
RGBA the final color values will be 0 or close to 0 anyway. Gecko solves
this issue by having a
cacheable surface generation API that can fetch both premulitplied and
unpremulitplied data
from things like image elements. We do not have that yet, but I argue
that this change
is important anyway due to the amount that it reduces memory and file
descriptor usage
as well as the cost of copying image data so much in memory.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This splits `Snapshot` into two structs:
- `Snapshot`: A non-serializable owned image buffer that might contain
either shared memory or a `Vec<u8>`. When mutated the shared memory
version is converted into a `Vec<u8>` avoiding a copy of the data
until absolutely necessary.
- `SharedSnapshot`: A serialzable version of `Snapshot` that contains
only shared memory and is suitable for sending across IPC channels.
This version is cheaply convertible from and to owned `Snapshot`s as
long as the source is also a shared memory `Snapshot`.
In addition, it does copyless conversions of `RasterImage` into
`Snapshot`s (including for frame offsets in animated images). Finally,
there are a few minor changes to try harder not to have to do a
`transform()` operation.
Testing: This should not change observable behavior, so is covered by
existing
tests. It should come with a mild performance improvement and open up
the
opportunity for others.
Fixes: #36594.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This implements LazyDOMString (from now on DOMString) as outlined in
https://github.com/servo/servo/issues/39479.
Constructing from a *mut JSString we keep the in a
RootedTraceableBox<Heap<*mut JSString>> and transform
the string into a rust string if necessary via the `make_rust_string`
method.
Methods used in script are implemented on this string. Currently we
transform the string at all times.
But in the future more efficient implementations are possible.
We implement the safety critical sections in a separate module
DOMStringInner which allows simple constructors, `make_rust_string` and
the `bytes` method.
This method returns the new type `EncodedBytes` which contains the
reference to the underlying string in either format.
Testing: WPT tests still seem to work, so this should test this
functionality.
---------
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
This is part of the future work of implementing LazyDOMString as
outlined in issue #39479.
We use str() method or direct implementations on DOMString for these
methods. We also change some types.
This is independent of https://github.com/servo/servo/pull/39480
Signed-off-by: Narfinger Narfinger@users.noreply.github.com
Testing: This is essentially just renaming a method and a type and
should not change functionality.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
This should be the final PR for the Hash Function series that is
trivial.
Of note: I decided to transform `HashMapTracedValues<Atom,..>` to use
FxBuildHasher. This is likely not going to improve performance as Atom's
already have a unique u32 that is used as the Hash but it safes a few
bytes for the RandomState that is normally in the HashMap.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Testing: Hash function changes should not change functionality, we
slightly decrease the size and unit tests still work.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
I think this is the correct solution? GlobalScope::get_cx() doesn't take
a self parameter anymore.
Testing: WebGL tests should suffice.
Fixes: #39228
Signed-off-by: lumiscosity <averyrudelphe@gmail.com>
Any RenderingContext/OffscreenRenderingContext type has readonly
"canvas" attribute
and associated native-code DOM context objects have reference to target
DOM canvas objects.
https://html.spec.whatwg.org/multipage/canvas.html#renderingcontexthttps://html.spec.whatwg.org/multipage/canvas.html#offscreenrenderingcontext
And currently the reference to DOM canvas object is the rooting pointer
on the stack,
which leads to the circular reference problem.
The SpiderMonkey's (SM) garbage collector will not be able to free the
DOM canvas and context
objects (unreacheble from JS) because of the rooting pointer on stack
(see STACK_ROOTS).
And these objects will be stored until the associated script
runtime/thread will be terminated.
SM -> JS Roots -> DOM Canvas* (on heap) -> DOM Context (on heap)
SM -> Rust Roots -> Dom Canvas* (on stack) <- as "canvas" member field
Let's replace the rooting pointer to the traceble pointer (DomRoot ->
Dom)
in the "canvas" member field of DOM context object, which allows to
broke circular referencing problem.
Testing: No changes in existed tests
Signed-off-by: Andrei Volykhin <volykhin.andrei@huawei.com>
Co-authored-by: Andrei Volykhin <volykhin.andrei@huawei.com>
Adds epoch to each WR image op command that is sent to compositor. The
renderer now has a `FrameDelayer` data structure that is responsible for
tracking when a frame is ready to be displayed. When asking canvases to
update their rendering, they are given an optional `Epoch` which denotes
the `Document`'s canvas epoch. When all image updates for that `Epoch`
are seen in the renderer, the frame can be displayed.
Testing: Existing WPT tests
Fixes: #35733
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
Move interfaces defined by the WebGL spec to the `script/dom/webgl/
`module from `script/dom/`.
`script/dom/webgl*.rs` -> `script/dom/webgl/`
`script/dom/webgl_extensions` -> `script/dom/webgl/extensions`
`script/dom/webgl_validations` -> `script/dom/webgl/validations`
Testing: No changes, just a refactoring
Fixes (partially): #38901
Signed-off-by: Andrei Volykhin <volykhin.andrei@huawei.com>
Co-authored-by: Andrei Volykhin <volykhin.andrei@huawei.com>
