Rather than running lookups in the global shared thread pool belonging
to `GTask`, run them in a private thread pool.
This is needed because the global shared thread pool is constrained to
only 14 threads. If there are 14 ongoing calls to
`g_task_run_in_thread()` from any library/code in the process, and then
one of them asks to do a DNS lookup, the lookup will block forever.
Under certain circumstances, particularly where there are a couple of
deep chains of dependent tasks running with `g_task_run_in_thread()`,
this can livelock the program.
Since `GResolver` is likely to be called as a frequent leaf call in
certain workloads, and in particular there are likely to be several
lookups requested at the same time, it makes sense to move resolver
lookups to a private thread pool.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
This will make it simpler to handle timeouts and cancellation in future,
as all the logic for working out whether to return will all be in one
place, and all the lookup-specific code is now implemented in simple
sync functions which don’t need to care about `GTask`s.
This commit introduces no functional changes, it’s just setting up for
the following commit.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
This introduces no functional changes, but will make a reorganisation of
the code simpler in the next commit.
Rather than dealing with three different closure types, this changes the
code to deal with one which is a tagged union of the three.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
The class and its header are not public, so this should not be an API or
ABI break.
This just simplifies the code a little and allows for easy extension of
the object’s private data in future commits.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Without a timeout, some lookup requests can go on forever, typically due
to bugs in underlying systems.
This can have particularly significant effects on the Happy Eyeballs
algorithm in `GSocketClient`, which relies on multiple name lookups as
its first step.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #2866
These make it a bit easier to track the ongoing resolver tasks, as the
tasks and/or their closures are not tracked in a big list somewhere.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Update a series of error messages to use `g_set_error_literal` instead
of `g_set_error`. This should prevent `format-nonliteral` compiler
issues when `-Werror` is configured:
../gio/gunixconnection.c: In function ‘g_unix_connection_receive_fd’:
../gio/gunixconnection.c:183:9: error: format not a string literal, argument types not checked [-Werror=format-nonliteral]
183 | nscm);
| ^~~~
../gio/gunixconnection.c:217:20: error: format not a string literal, argument types not checked [-Werror=format-nonliteral]
217 | nfd);
| ^~~
../gio/gunixconnection.c: In function ‘g_unix_connection_receive_credentials’:
../gio/gunixconnection.c:601:24: error: format not a string literal, argument types not checked [-Werror=format-nonliteral]
601 | nscm);
| ^~~~
This is similar to a previous change [1] made to `gunixconnection.c`.
[1]: 44b3d5d80445234041f6c59feb89645f7102c3a4
Signed-off-by: James Knight <james.d.knight@live.com>
Since commit c0ca3f99 this test is strictly depending on GDesktopAppInfo
that is not defined or available in macos, so skip the test as we do for
windows.
We could have done this at meson level too, but keeping it this way is
probably a better reminder that this should be adapted for such scenario
one day™
See: https://gitlab.gnome.org/GNOME/glib/-/jobs/2753753
The merge request !2848 added code to automatically detect the module
prefix on macOS, with a test for the Mac #define TARGET_OS_OSX. However,
older versions of the SDK (at least 10.11) don't provide this #define,
leading to build failure. If the #define is missing, fall back to
checking TARGET_OS_MAC. On newer SDKs this symbol is also true for
watchOS, etc., but in those situations TARGET_OS_OSX is available.
Various gio modules include gmodule.h that requires the
gmodule-visibility.h to be already built.
To make this easier, just provide a dependency and use it where we are
building modules that do not depend on libgio_dep (that already includes
that).
Fixes: https://gitlab.gnome.org/GNOME/glib/-/issues/2982
The generated gir file marks the size parameter as "out" by default. This is wrong in the context of a caller allocated buffer with a given size. Explicitly marking the size parameter as (in) fixes the issue.
The ref on the timeout source owned by `SendMessageData` was being
dropped just after attaching the source to the main context, leaving it
unowned in that struct. That meant the only ref on the source was held
by the `GMainContext` it was attached to.
This ref was dropped when returning `G_SOURCE_REMOVE` from
`send_message_with_reply_timeout_cb()`. Before that happens,
`send_message_data_deliver_error()` is called, which normally calls
`send_message_with_reply_cleanup()` and destroys the source.
However, if `send_message_data_deliver_error()` is called when the
message has already been delivered, calling
`send_message_with_reply_cleanup()` will be skipped. This leaves the
source pointer in `SendMessageData` dangling, which will cause problems
when `g_source_destroy()` is subsequently called on it.
I’m not sure if it’s possible in practice for this situation to occur,
but the code certainly does nothing to prevent it, and it’s easy enough
to avoid by keeping a strong ref on the source in `SendMessageData`.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #1264
It already implicitly held a strong ref on its `GTask` values, but
didn’t have a free function set so that they would be automatically
unreffed on removal from the map.
This meant that the functions handling removals from the map,
`on_worker_closed()` (via `cancel_method_on_close()`) and
`send_message_with_reply_cleanup()` had to call unref once more than
they would otherwise.
In `send_message_with_reply_cleanup()`, this behaviour depended on
whether it was called with `remove == TRUE`. If not, it was `(transfer
none)` not `(transfer full)`. This led to bugs in its callers.
For example, this led to a direct leak in `cancel_method_on_close()`, as
it needed to remove tasks from `map_method_serial_to_task`, but called
`send_message_with_reply_cleanup(remove = FALSE)` and erroneously didn’t
call unref an additional time.
Try and simplify it all by setting a `GDestroyNotify` on
`map_method_serial_to_task`’s values, and making the refcount handling
of `send_message_with_reply_cleanup()` not be conditional on its
arguments.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #1264
This is equivalent to the current behaviour, but a little clearer in its
meaning.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #1264
The ownership transfers in this code are a bit complex, so adding some
extra documentation and `g_steal_pointer()` calls should hopefully help
clarify things.
This doesn’t introduce any functional changes, just code documentation.
Another drive-by improvement in the quest for #1264.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #1264
The existing comment in the code was correct that `data` is freed when
the task callback is called, because `data` is also pointed to by the
`user_data` for the task, and that’s freed at the end of the callback.
So the existing code was correct to take a copy of `data->task` before
calling `g_task_return_*()`.
After calling `g_task_return_*()`, the existing code unreffed the task
(which is correct), but then didn’t clear the `data->task` pointer,
leaving `data->task` dangling. That could cause a use-after-free or a
double-unref.
Avoid that risk by explicitly clearing `data->task` before calling
`g_task_return_*()`.
After some testing, it turns out this doesn’t actually fix any bugs, but
it’s still a good robustness improvement.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Helps: #1264
The generated gir file marks the size parameter as "out" by default. This is wrong in the context of a caller allocated buffer with a given size. Explicitly marking the size parameter as (in) fixes the issue.
As explained in the previous commit, GNU Coreutils 9.2 changes the behaviour
of `du --bytes` to only count regular files and symlinks.
The previous commit makes the test pass with GNU Coreutils >=9.2, but the
machine running the tests may have an older version, or perhaps even a
reimplementation such as uutils. So we can't rely on the size returned by `du`
to be the consistent across systems any more.
However, the plus side of the new behaviour is that the size reported by `du`
/ `G_FILE_MEASURE_APPARENT_SIZE` is now well-defined across filesystems
(as the sum of the sizes of regular files & symlinks), so we can hardcode it.
Fixes: https://gitlab.gnome.org/GNOME/glib/-/issues/2965
Since GNU Coreutils 9.2 (commit 110bcd28386b1f47a4cd876098acb708fdcbbb25),
`du --apparent-size` (including `du --bytes`) no longer counts all kinds of
files (directories, FIFOs, etc.), but only those for which `st_size` in
`struct stat` is defined by POSIX, namely regular files and symlinks
(and also rarely supported memory objects).
This aligns the behaviour of GLib's `G_FILE_MEASURE_APPARENT_SIZE` flag
with the new GNU Coreutils `du` and correct POSIX use.
Note that this may be a breaking change for some uses.
Link: https://lists.gnu.org/archive/html/bug-coreutils/2023-03/msg00007.html
Fixes: https://gitlab.gnome.org/GNOME/glib/-/issues/2965
Warn when the boolean return isn't used, since we may not initialize
the out arguments in the FALSE case. Update all internal callers to
use the return value. They were already safe, but users outside GLib
may not be.
Previously, `-Wl,--export-dynamic` was in `Libs` key of `gmodule-2.0.pc`,
even though `-Wl` is a compiler flag, rather than a linker one.
This caused issues with API reference builds in evolution-data-server,
which passes the output of `pkg-config --libs` through `--ldflags`
argument of `gtkdoc-scan`, which are forwarded unchanged to `ld`:
ld: unrecognized option '-Wl,--export-dynamic'
Let’s move the flag to `Cflags` so that the compiler can deal with it.
https://gitlab.gnome.org/GNOME/evolution-data-server/-/issues/463
Add support for it, but don’t use it anywhere yet — this is an API
addition, but currently doesn’t cause any functional changes. It’ll be
used in the next commit.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
The `inotify_init1()` API has supported this flag for a long time
(possibly since it was first introduced, although I haven’t bothered
doing the archaeology).
This saves a syscall when first connecting to inotify.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Both `GPollableInputStream` and `GPollableOutputStream` are dynamic
interfaces, in that their implementation on a class may only be
functional if certain prerequisites are met at runtime. For example,
a `GConverterInputStream` is only pollable if its base stream is
pollable, and that’s determined at runtime rather than compile time.
As such, both interfaces have a `can_poll()` method. If that method
returns `FALSE`, the behaviour of all other methods on the interface is
undefined.
That was mentioned in the documentation for `can_poll()`, but not any of
the other documentation for the interfaces, which made it a bit hard to
find.
Mention it more widely.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Fixes: #2952
