* Clean up tracing instrumentation
Signed-off-by: Delan Azabani <dazabani@igalia.com>
* Set all tracing spans to trace level for now
Signed-off-by: Delan Azabani <dazabani@igalia.com>
---------
Signed-off-by: Delan Azabani <dazabani@igalia.com>
Instead of creating a `ROUTER` for each fetch, create a fetch thread
which handles all incoming and outcoming fetch requests. Now messages
involving fetches carry a "request id" which indicates which fetch is
being addressed by the message. This greatly reduces the number of file
descriptors used by fetch.
In addition, the interface for kicking off fetches is simplified when
using the `Listener` with `Document`s and the `GlobalScope`.
This does not fix all leaked file descriptors / mach ports, but greatly
eliminates the number used. Now tests can be run without limiting
procesess on modern macOS systems.
Followup work:
1. There are more instances where fetch is done using the old method.
Some of these require more changes in order to be converted to the
`FetchThread` approach.
2. Eliminate usage of IPC channels when doing redirects.
3. Also eliminate the IPC channel used for cancel handling.
4. This change opens up the possiblity of controlling the priority of
fetch requests.
Fixes#29834.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
System fonts used to be instantiated using the system font loader and
this change restores that behavior. In addition, on macOS and FreeType
platforms font data for system fonts is loaded using memory mapping. The
benefit is that system font loaders typically are able to cache fonts in
system memory (using memory mapping, for instance) and we'd like to load
them in a the way most compatible with other applications.
On my Linux system, this manages to get the overhead of loading a very
large font down from 10ms to approximately 1ms. Subsequent runs show
even less overhead. We've measured similar gains on macOS systems.
Currently, system font data must be loaded into memory manually for
canvas and this is unlikely to change even with a switch to `vello`. The
use of explicit memmory mapping should help in this case -- though it
probably won't be possible to use this properly on macOS and Windows if
we ever want to load fonts from TTCs properly.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Instead of exposing many different kinds of messages to the compositor
that are routed through the constellation, expose a single message type
which can be sent across IPC channels. In addition, this IPC channel and
the route to the crossbeam channel with the compositor is created along
with the `CompositorProxy`, simplifying what needs to be passed around
during pipeline initialization.
Previously, some image updates (from video) were sent over IPC with a
special serialization routine and some were sent via crossbeam channels
(canvas). Now all updates go over the IPC channel `IpcSharedMemory` is
used to avoid serialization penalties. This should improve performance
and reduce copies for video, but add a memory copy overhead for canvas.
This will improve in the future when canvas renders directly into a
texture.
All-in-all this is a simplification which opens the path toward having a
standard compositor API and reduces the number of duplicate messages and
proxying that had to happen in libservo.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Instead of exposing many different kinds of messages to the compositor
that are routed through the constellation, expose a single message type
which can be sent across IPC channels. In addition, this IPC channel and
the route to the crossbeam channel with the compositor is created along
with the `CompositorProxy`, simplifying what needs to be passed around
during pipeline initialization.
Previously, some image updates (from video) were sent over IPC with a
special serialization routine and some were sent via crossbeam channels
(canvas). Now all updates go over the IPC channel `IpcSharedMemory` is
used to avoid serialization penalties. This should improve performance
and reduce copies for video, but add a memory copy overhead for canvas.
This will improve in the future when canvas renders directly into a
texture.
All-in-all this is a simplification which opens the path toward having a
standard compositor API and reduces the number of duplicate messages and
proxying that had to happen in libservo.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Instead of a blocking a layout thread on the generation of WebRender
`FontKey`s and `FontInstanceKey`s, generate the keys ahead of time and
send the font data to WebRender asynchronously. This has the benefit of
allowing use of the font much more quickly in layout, though blocking
display list sending itself on the font data upload.
In order to make this work for web fonts, `FontContext` now asks the
`SystemFontService` for a `FontKey`s and `FontInstanceKey`s for new web
fonts. This should happen much more quickly as the `SystemFontService`
is only blocking in order to load system fonts into memory now. In
practice this still drops layout thread blocking to fractions of a
millisecond instead of multiple milliseconds as before.
In addition, ensure that we don't send font data or generate keys for
fonts that are used in layout but never added to display lists. This
should help to reduce memory usage and increase performance.
Performance of this change was verified by putting a microbenchmark
around `FontContext::create_font` which is what triggered font key
generation.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This is done by no longer forwarding compositor-bound messages through
SystemFontService and making `FontContext` non-generic:
- Messages from the `FontContext` to the `Compositor` no longer need to be
forwarded through the `SystemFontService`. Instead send these messages
directly through the script IPC channel to the `Compositor`.
- Instead of adding a mock `SystemFontServiceProxy`, simply implement a
mock `SystemFontService` on the other side of an IPC channel in the
`font_context` unit test. This allows making `FontContext`
non-generic, greatly simplifying the code. The extra complexity moves
into the unit test.
These changes necessitate adding a new kind of `FontIdentifier`,
`FontIdentifier::Mock` due to the fact that local fonts have
platform-specific identifiers. This avoids having to pretend like the
system font service can have web fonts -- which was always a bit of a
hack.
These two changes are combined into one PR because they both require
extensive and similar chages in the font_context unit test which
dependended on the details of both of them.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This changes modifes the way that font data is sent over IPC channels.
Instead of serializing the data or sending it via IPC byte senders, font
data is copied into shared memory and a copy of the handle is sent over
the channel.
There is also the idea of sending the file handle of the on disk data of
system fonts. This could be implemented as a further followup once there
is an abstraction in `ipc-channel` over file handles.
To accomplish this, a `FontData` abstraction is added, which also allows
caching an in-memory shared `Arc<Vec<u8>>` version of the data (neeeded
by some APIs). This could also be a place for caching font tables in the
future.
Finally, the `FontCacheThread` is renamed to the `SystemFontService`
while the proxy for this is now named `SystemFontServiceProxy`.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
Do font selection using Servo's font backend, which is shared with the
rest of layout. In addition, delay the creation of the `font-kit` font
until just before rendering with `raqote`. The idea is that when
`raqote` is no longer used, we can drop the `font-kit` dependency.
This change has the side-effect of fixing text rendering in canvas,
adding support for font fallback in canvas, and also correcting a bug in
font selection with size overrides.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
This adds support for generic font families in Servo and allows for
configuration of them as well as their default font sizes. One
interesting fix here is that now monospace default to 13px, like it does
in other browsers.
In addition to that, this exposes a new interface in Stylo which allows
setting a default style. This is quite useful for fonts, but also for
other kinds of default style settings -- like text zoom.
Fixes#8371.
Fixes#14773.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
This function doesn't exist in upstream Stylo and is fairly unecessary.
Removing it will help reduce the difference between downstream Stylo and
upstream Stylo.
This crate only takes care of fonts now as graphics related things are
split into other crates. In addition, this exposes data structures at
the top of the crate, hiding the implementation details and making it
simpler to import them.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Mukilan Thiyagarajan <mukilan@igalia.com>
2024-06-19 20:26:19 +00:00
Renamed from components/gfx/font_context.rs (Browse further)
