* Fix race condition in Worker destruction
During shutdown, the main script thread calls
JS_RequestInterruptCallback(cx) for each worker thread it owns
where cx is the JSContext* created for that worker.
Although JS_RequestInterruptCallback is safe to call
from threads other than the worker thread, it is possible
the JSContext* has already been destroyed
For example, as noted in #30022, since the main thread sets
the worker's `closing` flag to true to signal termination before it
calls JS_RequestInterruptCallback, we can have a race condition
where the worker exits its event loop when `closing` flags is set
and then it (worker thread) destroys its own JSContext and JSRuntime.
When the main thread resumes, it will call
JS_RequestInterruptCallback for the worker's context, leading to
a use-after-free bug.
This patch solves this issue by improving the existing
`ContextForRequestInterrupt` abstraction used for sharing the Worker's
associated JSContext* with the parent script thread.
Instead of simply wrapping a plain `*mut JSContext`, we now wrap the
`*mut JSContext` in a nullable mutex i.e Mutex<Option<*mut JSContext>>
The mutex lock needs to be held by the parent thread when it
calls JS_RequestInterruptCallback.
Similary, before the worker destroys its JSContext, it locks and
sets the Option to None, signaling that the JSContext can no longer
be used for interrupting the worker.
This patch also fixes the issue in #30052 by enforcing the use
of ContextForRequestInterrupt abstraction which ensures the correct
JSContext is used by the main thread when Worker.terminate is called.
Fixes#30022, #30052
* Fix Worker.importScripts to handle termination
Fixing #30052 uncovered this issue in the implementation
of `importScripts` method. After the fix for #30052,
the WPT test `/workers/Worker-terminate-forever-during-evaluation.html`
started to crash because when evaluation doesn't succeed
`importScripts` always returns Error::JSFailed code to the caller,
which indicates that there is a Dom/JS exception to be thrown. However,
this is not true when the script is terminated, which causes
the generated binding layer for 'importScript` to fail
the assertion that there is a pending exception.
This patch makes `importScripts` work similar to the [logic that
evaluates the top-level script][1] of the Worker - it simply prints
`evaluate_script failed - (terminated)' if the worker is terminating
[1]: 3fea90a231/components/script/dom/workerglobalscope.rs (L434)
With direction:ltr (and we don't support direction:rtl yet), the rules
from https://drafts.csswg.org/css2/#blockwidth imply that margin-left
shouldn't resolve auto to a negative amount.
This aligns Servo with Gecko and Blink. WebKit may resolve to a negative
amount in some cases.
When descending and we have the option, don't create new
PositioningContexts just to update the static position of laid out
abspos descendants. Instead, use the new PositioningContextLength to
only update the newly added hoisted abspos boxes.
Currently, `./mach test-wpt` family of commands (`test-wpt`,
`test-wpt-android`, and `test-wpt-failure`) and `./mach update-wpt`
default to using the legacy-layout option `--layout-2013` unless
`--layout-2020` is specified.
Given that we are now using layout-2020 by default, this change updates
these commands to default to using the `--layout-2020` option instead of
`--layout-2013`.
Just use clamp_between_extremums() to resolve the inline size, and then
only call solve_inline_margins_for_in_flow_block_level() once.
There should be no change in behavior.
Type inference was incorrectly inferring that our `check_output()`
helper was returning `str` when in reality, it returns `bytes`. This
fixes the caller that was no longer decoding those bytes and fixes the
type annotation on the function.
This change makes rustup a requirement for building Servo with `./mach`
and switches to the newer `rust-toolchain.toml` format. The goal here is
to make mach builds more similar to non-mach builds.
- The new format allows listing the required components, removing some of
the complexity from our mach scripts.
- This means we must raise the required version of rustup to 1.23. The
current version is 1.26.
- We no longer wrap every call to cargo and rustc in "rustup run" calls
as both cargo and rustc will take care of installing and using all
necessary components specified in `rust-toolchain.toml` when run
inside the project directory.
No difference in behavior, just these changes:
- PlacementAmongFloats::new() initializes the top of the 1st band to the
ceiling, so that other methods can just refer to the former without
having to floor by the later.
- In fact, the 'ceiling' field becomes unnecessary, and is removed.
- top_of_placement_for_current_bands() is renamed to current_ceiling().
- try_place_once() is reorganized to reduce indentation.
- The condition 'len() > 0' becomes '!is_empty()'.
- The 1st band is now popped in place() instead of try_place_once(),
then it's easier to see why the loop will end.
Now that the new version of GStreamer fixes this issue, we can remove
the workarounds for this problem as well as all of the homebrew
bootstrapping logic.
This hasn't been updated since 2017 and homebrew installation is also
provided via a cask which downloads the latest version from the website
[^1]. I think this code is basically unused.
[^1]: 9e944ae828/Casks/servo.rb
Consumers of PlacementAmongFloats weren't handling margins properly.
They were assuming that they would either get a positive adjustment,
or zero for no-op.
However, just like the regular clearance triggered by 'clear', the
clearance added onto blocks that establish an independent FC can be
zero or negative, and the effect is different than having no clearance.
Servo is no longer completely vendored into Gecko. Instead parts of
Gecko are vendored into Servo. This change removes Python mach bootstrap
code that was written to accommodate the previous situation. It's no
longer necessary.
Embed the git hash into the servo binary using vergen instead of using
custom Python code in mach. The benefit here is ones less difference
between a normal cargo run and building via mach in addition to removing
a bunch of code.
This code was written to handle both Python 2 (which we no longer
support) and old Windows CI machines that did not have up-to-date CA
stores. I think we can remove this now.
Calculating the max-content size of a block container may add the outer
max-content sizes of multiple children. The problem is that the outer
size may be negative (due to margins), producing an incorrect result.
In particular, it could happen that the max-content size was 50-25=25,
but the min-content size would just take the maximum and be 50, which
doesn't make sense.
Therefore, this patch floors the size of all children by 0. This seems
to match Blink. Firefox and WebKit don't floor in some cases, but then
the result seems suboptimal to me. Note that there is no spec for this,
see https://github.com/w3c/csswg-drafts/issues/9120 for details.
Since #29950, unit tests were only running with the legacy layout, and
there was no way to run them for layout 2020.
This patch makes './mach test-unit' run unit tests for both.
Also doing some changes so that the layout 2020 floats.rs tests compile.
Do not do the duplicate workflow checks for main workflow jobs triggered
by 'issue_comment' GitHub Actions. This prevents try jobs from being
skipped. This issue was not detected in my testing, because the main
workflow does not run in personal repositories and 'issue_comment'
actions cannot be tested from PRs.
This is the last piece of the puzzle to turning off bors. This makes
functionality provided by bors to understand "@bors-servo try" a GitHub
Action. For now the syntax is more or less the same, but we can modify
it in the future and even add support for custom configuration options
(more specific build combinations or even passing compiler flags).
The big difference between this and what bors does is that there is no
merge commit. GitHub simply runs tests on the version of the branch that
is on a pull request. There is always the risk that tests might start
failing when a branch is rebased, but this offers a bit more control
because you can easily rebase from the PR and the merge queue will check
this as well.
* compositing: rename WindowEvent to EmbedderEvent
* winit: rename winit_event_to_{servo → embedder}_event
* winit: rename ServoEvent::Awakened to WakerEvent
* winit: document App::handle_events and rename locals
* servo: rename Servo.embedder_events to messages_for_embedder
* winit: rustdoc link to EmbedderEvent
* winit: use new name queue_embedder_events_for_winit_event
* Add initial support for css-text-3 whitespace handling
This adds initial support for whitespace handling from the CSS
specification for Layout 2020. In general, the basics are covered. Since
test output is very sensitive to whitespace handling, this change
incorporates several fixes:
1. Whitespace is collapsed according to the Phase 1 rules of the
specification, though language-specific unbreaking rules are not
handled properly yet.
2. Whitespace is mostly trimmed and positioned according to the Phase 2
rules, but full support for removing whitespace at the end of lines
is pending on a temporary data structure to hold lines under
construction.
3. Completely empty box fragments left over immediately after line
breaks are now trimmed from the fragment tree.
4. This change tries to detect when an inline formatting context
collapses through.
Fixes#29994.
Co-authored-by: Mukilan Thiyagarajan <me@mukilan.in>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
* Update test results
---------
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <me@mukilan.in>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
* Update web-platform-tests to revision b'e16853520ae70ba309108ccff87a7816f74be4dd'
* Remove new spurious result
---------
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
A block that establishes an independent formatting context is placed
next to previous floats, so we should add their sizes when computing
the intrinsic contribution of the parent block container.
Remove rayon_croissant and refactor the way that information about
floats in flows bubbles up. This simplifies the code a good deal and
lets us take advantage of some more optimized functions provided by
rayon. This removes 2 crates from the dependency tree.
In addition, this allows avoiding passing `contains_floats` up from
every box tree construction function. This makes things simpler, but
also opens up the possibility of passing more of these flags up in the
future (such as `contains_counters`).
calculate_inline_content_size_for_block_level_boxes was relying on
inline_content_sizes to get the size of each block-level box child.
For absolutely positioned boxes, this was 0x0.
That was no-op before #29887, but then it prevented adding the sizes
of a sequence of floats. Abspos should just be ignored instead of
treated as 0x0.
This patch removes inline_content_sizes, moves the logic into
calculate_inline_content_size_for_block_level_boxes (the only caller),
and handles abspos correctly.
* Place replaced and non-auto inline size independent FCs next to floats
The CSS2 specification says that replaced content and independent
formatting contexts should be placed next to floats. This change adds
support for that, but punts on support for independent formatting
contexts that have an auto inline size. With an auto inline size, we
which requires a much more complex layout algorithm.
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
* Fix issue with where last band was taken into account for inline size
* adjustment_from_floats should prevent margin collapse
* Properly handle elements with 0 height
---------
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
This code was used to test buildbox_steps.yml, but Servo no longer uses
buildbot, so this code is essentially unused. In addition, YAML +
Cython 3 is causing issues on the CI.
Fixes#30003
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Much of the code used to import WPT tests from upstream has been moved
to the WPT repository itself, so this can be reused. In addition,
simplify the workflows by merging the entire process into mach and also
directly into the GitHub workflow. This should fix WPT imports after
combining compilation of layout and legacy layout.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
When the decision job decides that a workflow should be skipped because
an identical one has already run, that workflow should be marked as
successful and not unsuccessful.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This API has been removed in the latest version of WebRender and we can
simply get this information from the compositor-side scroll tree. In
addition, this change limits the amount of data sent to the pipeline
that actually changed and gives the function in the compositor a better
name.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Including hit tests in non-hit test display list items is no longer
supported in upstream WebRender, so this change switches legacy layout
to always use explicit hit test display list items.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>