This reverts commit 799f8dcd46.
This patch seems to break applications that use GTask specific
operations with GSocket. We will need to investigate a bit more
on this issue but for now we revert it and leave it for the
next major release.
Temporarily increase the ref count of a GSource to 1 while calling its
finalize() callback, so that the finalize() implementation can call
GSource methods (like g_source_set_ready_time()) without causing
critical warnings. It’s safe to call those methods at this point, as the
source has been destroyed, but nothing has been freed.
This is an indirect way of fixing a race between GCancellable and
GCancellableSource, whereby the GCancellable::cancelled callback for the
GCancellableSource is not disconnected until the GCancellableSource’s
finalize() function is called. Previously, this meant there was a window
in which the GCancellableSource’s ref count was 0, but the ::cancelled
callback was still connected, and could legitimately be called as a
result of another thread calling g_cancellable_cancel() on the
GCancellable. The callback calls g_source_set_ready_time() on the
GSource, and there’s no thread-safe way of checking whether the GSource
has been destroyed. Instead, we have to change GSource so its ref count
is only decremented to 0 inside the locked section in
g_source_unref_internal() *after* the finalize() function has been
called, and hence after the GCancellable::cancelled callback has been
disconnected. The use of g_cancellable_disconnect() ensures that the
callback disconnection is thread safe.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
https://bugzilla.gnome.org/show_bug.cgi?id=781601
Since commit e4ee307 ("Do not wake up main loop if change is from same
thread", bug 761102), GMainContext uses context->owner to decide if the
event loop is being run in the current thread. However, what really
matters is the phase in the prepare/query/poll/check/dispatch sequence.
Wakeups are only needed between the end of prepare and the end of poll,
and then only if prepare found that no sources were ready.
There is no need to take threads into account, because prepare, check
and all callers of conditional_wakeup all look at the new need_wakeup
flag inside LOCK_CONTEXT/UNLOCK_CONTEXT.
With this change, g_main_context_is_owner and g_main_context_wait are
the only functions for which acquire/release matters, just like before
commit e4ee307.
Signed-off-by: Paolo Bonzini <bonzini@gnu.org>
Some files that this script will process might have UTF-8 items in
there, which can cause problems on Python 3.x as it is more strict and
careful on unicode issues. Fix this by:
-Doing what we did before on Python 2.x
-Open the file with encoding='utf-8' on Python 3.x
mnt_table_is_fs_mounted causes unwanted automount requests due to
canonicalization of source and target. It might be replaced by
mnt_table_find_source as per the documentation in order to prevent
the automounts, but it is redundant. All mtab entries should be already
mounted and thus mnt_table_is_fs_mounted result is always true (it
basically checks that the fs from mtab is in mtab). Let's remove
the check at all.
https://bugzilla.gnome.org/show_bug.cgi?id=781867
When we are inside a sandbox, we want to use the portal
implementation, since it is the only one that has a chance
of working.
This is safe to do, since the portal implementation will
just fail initialization when loaded outside a sandbox.
The original patch really should have introduced a helper - among
other things it deserves a code comment. We're likely to make
further changes too, so it's obviously better to only do it in one
place.
See: https://bugzilla.gnome.org/show_bug.cgi?id=761102
Being able to determine that a certificate chain is invalid is not
considered an error, but success. This might not be obvious at first
due to the way the method is named and described currently. Since we
cannot change the name, let's improve the description and clarify this
aspect of its behaviour.
https://bugzilla.gnome.org/show_bug.cgi?id=780310
g_dbus_proxy_get_cached_property() and
g_dbus_proxy_get_cached_property_names() can both return NULL if the
property cache is empty. Avoid a crash if this situation arises (which
it looks like it could, from reading the code) by gracefully bailing out
on NULL return values.
Coverity issues: #1257044, #1257045https://bugzilla.gnome.org/show_bug.cgi?id=741229
socket->priv->timeout is only a guint, and the multiplication is
performed before it’s widened to gint64 to be stored in start_time
(thanks, C). This means any timeout of 50 days or more would overflow.
Fixing this bug makes me feel a real sense of self-worth.
Coverity ID: 1159478
Signed-off-by: Philip Withnall <withnall@endlessm.com>
When compiling statically against the system-provided gettext on Mac OS
X, GLib needs to be linked against CoreFramework, which provides some
functions used by gettext.
Fix this by including the necessary macro magic from upstream gettext.
https://bugzilla.gnome.org/show_bug.cgi?id=725894
This will help us break generic GType deadlocks between people using
GDBus in different threads (which is supported), not just by GType
usage in the GDBus thread.
This should fix the common cases we're seeing in the wild, although I
have some lingering concerns that if someone e.g. referenced
e.g. `G_TYPE_DBUS_AUTH_MECHANISM_SHA1` etc. we'd need to add those
too.
https://bugzilla.gnome.org/show_bug.cgi?id=674885
Now that the old logging API forwards through to the new structured
logging API, we don’t need to check the level or domain of a message
against INFO_LEVELS or G_MESSAGES_DEBUG — just pass it straight through
to the new structured logging API writer function.
https://bugzilla.gnome.org/show_bug.cgi?id=775879
If encoding with small chunks such that the call to
g_base64_encode_close() has to deal with 0 < x < 24 bits of remaining
state, the encoding code would not correctly use zeroes to pad out the
state — it would use left-over state from an earlier iteration in the
encoding process.
This resulted in invalid base-64 encodings.
Add a unit test for incremental encoding using different sized chunks
too.
Thanks to Rainier Perske for reporting and analysing the bug.
https://bugzilla.gnome.org/show_bug.cgi?id=780066
Signed-off-by: Philip Withnall <withnall@endlessm.com>
These calls cause race warnings from tsan, but are not a thread safety
problem, because we can only ever observe single bit changes: all
modifications to the GSource.flags field are done with a lock held; all
reads are of independent fields, so no intermediate state can ever be
observed. This assumes that a non-atomic read will consistently give us
an old value or a new value.
In any case, these g_source_is_destroyed() calls can happen from any
thread, and the state could be changed from another thread immediately
after the call returns; so the checks are pointless. In addition,
calling g_source_set_ready_time() or g_source_destroy() on a destroyed
source is not a problem.
https://bugzilla.gnome.org/show_bug.cgi?id=778049
For versions of GCC which support it (≥ 4), define G_MODULE_EXPORT as
__attribute__((visibility("default"))). This is normally a no-op, unless
compiling with -fvisibility=hidden, in which case it marks a symbol to
be publicly exported from the library, which is what G_MODULE_EXPORT is
for. Previously G_MODULE_EXPORT has only worked on Windows.
The compatibility check for whether the compiler supports
__attribute__((visibility)) is based on the __GNUC__ define, and is
similar to the check done in configure.ac for defining G_GNUC_INTERNAL.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
https://bugzilla.gnome.org/show_bug.cgi?id=778287
Calling the D-Bus method for the OpenURI portal "protects" the logic from
not ever having the remote method running in case the xdg-desktop-portal
process is not yet running and the caller quits quickly after the call.
This should not be a problem as the method returns immediately (regardless
of the user making a selection), but making it synchronous would prevent
situations where the OpenURI method would never be called because of D-Bus
dropping the message after the caller dies, without explicitly waiting for
a reply.
https://bugzilla.gnome.org/show_bug.cgi?id=780441
The custom build rules did not have rules for x64 Debug builds during
the refactoring. Fix this by removing all Platform and Configuration
conditions for it, since this is done on are configs in the same manner.
Noted by Ignacio Casal Quinteiro.
We currently assume that the OpenURI portal should be used
unconditionally when running inside a flatpak sandbox. While
the portal is what we usually want, there are exceptions:
Yelp is now included in the GNOME runtime to allow displaying
help without exporting the user documentation, and the sandboxed
app itself may register a scheme handler.
To account for those cases transparently, always try the normal
code path first and only fall back to calling the portal when
that fails.
https://bugzilla.gnome.org/show_bug.cgi?id=780471
