When using fixed size connection buffers, if either the client or the
server is sending requests faster than the other end can cope with, the
connection buffers will fill up, eventually killing the connection.
This can be a problem for example with Xwayland mapping a lot of
windows, faster than the Wayland compositor can cope with, or a
high-rate mouse flooding the Wayland client with pointer events.
To avoid the issue, resize the connection buffers dynamically when they
get full.
Both data and fd buffers are resized on demand.
The default max buffer size is controlled via the wl_display interface
while each client's connection buffer size is adjustable for finer
control.
The purpose is to explicitly have larger connection buffers for specific
clients such as Xwayland, or set a larger buffer size for the client
with pointer focus to deal with a higher input events rate.
v0: Manuel:
Dynamically resize connection buffers - Both data and fd buffers are
resized on demand.
v1: Olivier
1. Add support for unbounded buffers on the client side and growable
(yet limited) connection buffers on the server side.
2. Add the API to set the default maximum size and a limit for a given
client.
3. Add tests for growable connection buffers and adjustable limits.
v2: Additional fixes by John:
1. Fix the size calculation in ring_buffer_check_space()
2. Fix wl_connection_read() to return gracefully once it has read up to
the max buffer size, rather than returning an error.
3. If wl_connection_flush() fails with EAGAIN but the transmit
ring-buffer has space remaining (or can be expanded),
wl_connection_queue() should store the message rather than
returning an error.
4. When the receive ring-buffer is at capacity but more data is
available to be read, wl_connection_read() should attempt to
expand the ring-buffer in order to read the remaining data.
v3: Thomas Lukaszewicz <tluk@chromium.org>
Add a test for unbounded buffers
v4: Add a client API as well to force bounded buffers (unbounded
by default (Olivier)
v5: Simplify ring_buffer_ensure_space() (Sebastian)
Co-authored-by: Olivier Fourdan <ofourdan@redhat.com>
Co-authored-by: John Lindgren <john@jlindgren.net>
Co-authored-by: Sebastian Wick <sebastian@sebastianwick.net>
Signed-off-by: Manuel Stoeckl <code@mstoeckl.com>
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Signed-off-by: John Lindgren <john@jlindgren.net>
Signed-off-by: Sebastian Wick <sebastian@sebastianwick.net>
Closes: https://gitlab.freedesktop.org/wayland/wayland/-/issues/237
This appears to be what at least wlroots-based compositors and kwin do
in practice. However, it's not abundantly clear from the protocol text
what the expected behavior here is. This patch fixes that.
Signed-off-by: Isaac Freund <mail@isaacfreund.com>
This generates a header with only enum definitions. This is useful
to share enum headers between libraries and library users.
Signed-off-by: Simon Ser <contact@emersion.fr>
The symbols check only works for dynamic libraries.
When building statically, the test fails.
This is caused by the check filtering out non-dynamic symbols with nm.
This change skips the check when building only static libraries.
Signed-off-by: Jordan Williams <jordan@jwillikers.com>
If a wl_array has size zero, wl_array_for_each computes NULL + 0 to get
to the end pointer. This should be fine, and indeed it would be fine in
C++. But the C specification has a mistake here and it is actually
undefined behavior. See
https://davidben.net/2024/01/15/empty-slices.html
Clang's -fsanitize=undefined flags this. I ran into this in Chromium's
build with wayland-scanner on one of our XML files.
../../third_party/wayland/src/src/scanner.c:1853:2: runtime error: applying zero offset to null pointer
#0 0x55c979b8e02c in emit_code third_party/wayland/src/src/scanner.c:1853:2
#1 0x55c979b89323 in main third_party/wayland/src/src/scanner.c
#2 0x7f8dfdb8c6c9 in __libc_start_call_main csu/../sysdeps/nptl/libc_start_call_main.h:58:16
#3 0x7f8dfdb8c784 in __libc_start_main csu/../csu/libc-start.c:360:3
#4 0x55c979b70f39 in _start (...)
An empty XML file is sufficient to hit this case, so I've added it as a
test. To reproduce, undo the fix and include only the test, then build
with:
CC=clang CFLAGS="-fno-sanitize-recover=undefined" meson build/ -Db_sanitize=undefined -Db_lundef=false
ninja -C build test
Signed-off-by: David Benjamin <davidben@google.com>
This is less cryptic to read than letters, and allows the compiler
to check switch statements exhaustiveness.
Signed-off-by: Simon Ser <contact@emersion.fr>
There are situations in which a call into wl_client_destroy() can
result in a reentrant call into wl_client_destroy() - which
results in UAF / double free crashes.
For example, this can occur in the following scenario.
1. Server receives a message notifying it that a client has
disconnected (WL_EVENT_HANGUP [1])
2. This beings client destruction with a call to wl_client_destroy()
3. wl_client_destroy() kicks off callbacks as client-associated
resources are cleaned up and their destructors and destruction
signals are invoked.
4. These callbacks eventually lead to an explicit call to
wl_display_flush_clients() as the server attempts to flush
events to other connected clients.
5. Since the client has already begun destruction, when it is
reached in the iteration the flush fails wl_client_destroy()
is called again [2].
This patch guards against this reentrant condition by removing
the client from the display's client list when wl_client_destroy()
is first called. This prevents access / iteration over the client
after wl_client_destroy() is called.
In the example above, wl_display_flush_clients() will pass over
the client currently undergoing destruction and the reentrant
call is avoided.
[1] 8f499bf404/src/wayland-server.c (L342)
[2] 8f499bf404/src/wayland-server.c (L1512)
Signed-off-by: Thomas Lukaszewicz [thomaslukaszewicz@gmail.com](mailto:thomaslukaszewicz@gmail.com)
- wayland-egl-abi-check: try to use llvm-nm first instead of BSD nm (incompatible options)
- avoid forcing _POSIX_C_SOURCE=200809L (SOCK_CLOEXEC become available)
- epoll(7) is provided by a userspace wrapper around kqueue(2) as FreeBSD
- when using SO_PEERCRED, the struct to use is `struct sockpeercred` instead of `struct ucred` on OpenBSD
- provide a compatibility layer for count_open_fds() using sysctl(2) as FreeBSD
Signed-off-by: Sebastien Marie <semarie@online.fr>
The wl_output events should not be used anymore for guessing the
preferred scale and transform of a surface. We have explicit events
for that now.
Signed-off-by: Simon Ser <contact@emersion.fr>
The only way to attach some data to a wl_client seems to be setting up a
destroy listener and use wl_container_of. Let's make it straight forward
to attach some data.
Having an explicit destroy callback for the user data makes managing the
user data lifetime much more convenient. All other callbacks, be they
wl_resource request listeners, destroy listeners or destructors, or
wl_client destroy listeners, can assume that the wl_client user data
still exists if it was set. Otherwise making that guarantee would be
complicated.
Co-authored-by: Pekka Paalanen <pekka.paalanen@collabora.com>
Signed-off-by: Sebastian Wick <sebastian@sebastianwick.net>
Currently it is possible to iterate over client-owned resources
during client destruction that have had their associated memory
released.
This can occur when client code calls wl_client_destroy(). The
following sequence illustrates how this may occur.
1. The server initiates destruction of the connected client via
call to wl_client_destroy().
2. Resource destroy listeners / destructors are invoked and
resource memory is freed one resource at a time [1].
3. If a listener / destructor for a resource results in a call
to wl_client_for_each_resource(), the iteration will proceed
over resources that have been previously freed in step 2,
resulting in UAFs / crashes.
The issue is that resources remain in the client's object map
even after they have had their memory freed, and are removed
from the map only after each individual resource has had its
memory released.
This patch corrects this by ensuring resource destruction first
invokes listeners / destructors and then removing them from the
client's object map before releasing the associated memory.
[1] https://gitlab.freedesktop.org/wayland/wayland/-/blob/main/src/wayland-server.c?ref_type=heads#L928
Signed-off-by: Thomas Lukaszewicz thomaslukaszewicz@gmail.com
"is incompatible with the implementation in libwayland" is a common
source of confusion as evidenced by repeated discussions in IRC
channel.
Improve the wording by making clear that
- packing IDs is a protocol requirement
- there are implementations (including libwayland) that enforce it
Signed-off-by: Mikhail Gusarov <dottedmag@dottedmag.net>
Allow setting a name for an event queue. The queue is used only for
printing additional debug information.
Debug output can now show the name of the event queue an event is
dispatched from, or the event queue of a proxy when a request is made.
Signed-off-by: Derek Foreman <derek.foreman@collabora.com>
Some code paths that lead to a client error and connection termination
have no associated logging, or insufficient logging. This makes it
difficult to understand what went wrong. This commit adds or supplements
logging for all these code paths.
Signed-off-by: Erik Chen <erikchen@chromium.org>
wl_connection_write() contained an exact copy of the logic in
wl_connection_queue(). Simplify things by just calling
wl_connection_queue() from wl_connection_write().
Signed-off-by: John Lindgren <john@jlindgren.net>
For libs/cflags this is done automatically, but not for manually accessed
variables. This matches what wayland-protocols does.
Signed-off-by: Andreas Cord-Landwehr <cordlandwehr@kde.org>
Allows clients to cleanly release wl_shm objects. Useful for clients
using multiple wl_registry objects (e.g. via libraries).
Signed-off-by: Simon Ser <contact@emersion.fr>
This was hardcoded to 1 regardless of the version passed to the
callback or the version of the parent resource.
Signed-off-by: Simon Ser <contact@emersion.fr>
Since the positivity of zero is debatable, and, in some cases scale was simply
underspecified, clarify the situation.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
The cursor name spec [1] describes how cursors should be named,
and is widely used. Add aliases so that users can pass these
names to libwayland-cursor without having to add fallbacks for
X11 cursor names.
[1]: https://www.freedesktop.org/wiki/Specifications/cursor-spec/
Signed-off-by: Simon Ser <contact@emersion.fr>
This adds a command to re-generate the test data. This needs to be
done when either an XML source file or the scanner's output is
changed.
Signed-off-by: Simon Ser <contact@emersion.fr>
The spec does not describe which actions cause the compositor to assign
keyboard focus to a surface, leaving this up to the compositor.
Compositors differ in their behavior when the user clicks on a
sub-surface. Some will move the keyboard focus to the subsurface whereas
others will only ever assign the keyboard focus to toplevel surfaces.
Some applications (e.g. firefox) seem to require the second behavior.
This patch specifies that sub-surfaces never get the keyboard focus.
Signed-off-by: Julian Orth <ju.orth@gmail.com>
Don't mention when the parent surface state is applied; the parent
surface isn't necessarily a sub-surface.
Signed-off-by: Kirill Primak <vyivel@eclair.cafe>
This should be sufficient for clients to not decide to fallback to
output based logic to determine scaling/transform when compositor
doesn't send any of the v6 events.
Signed-off-by: Kirill Chibisov <contact@kchibisov.com>
The issue template is hard to notice because it's not the default.
Users have to explicitly select it from the easy-to-miss dropdown
to get the warning.
Make the template the default one, so that new users are less likely
to miss it.
Signed-off-by: Simon Ser <contact@emersion.fr>
Due to what is arguably a mistake in the C language specification,
passing NULL to memcpy and friends is undefined behavior (UB) even when
the count is 0. C additionally mistakenly leaves NULL + 0 and NULL -
NULL undefined. (C++ fixes this mistake.) These are very problematic
because (NULL, 0) is a natural representation of the empty slice.
Some details:
https://github.com/llvm/llvm-project/issues/49459https://www.imperialviolet.org/2016/06/26/nonnull.html
Unfortunately, despite how clearly this is a mistake, glibc headers and
GCC now try to exploit this specification mistake and will miscompile
code, so C projects need to workaround this. In particular, UBSan from
Clang will flag this as a bug (although Clang itself has the good sense
to never lean on this bug). We've run into a few UBSan errors in
Chromium stemming from Wayland's memcpy calls. Add runtime guards as
needed to avoid these cases.
Note: Chromium's copy of wayland has
https://gitlab.freedesktop.org/wayland/wayland/-/merge_requests/188
applied. It is possible the ring_buffer_copy UB cases are only reachable
with that MR applied, I'm not sure. But it seemed simplest to just add
the fix to wayland as-is. Then when/if that MR lands, it will pick this
up.
Signed-off-by: David Benjamin <davidben@google.com>
This can be useful for additional validation purposes when handling
proxies. This is similar to existing server side API
wl_global_get_display.
Signed-off-by: David Edmundson <david@davidedmundson.co.uk>
Newer ci-templates contains bugfixes.
While at it, stop using a GitLab YAML reference, because we only
use this value in one spot.
Signed-off-by: Simon Ser <contact@emersion.fr>
