Two serious mistakes:
- we need to call dbus_server_free_data_slot as many times as we call
dbus_server_allocate_data_slot
- we need to delete the d pointer...
The changes to the unit tests are simply to cause the used peer
connections to be removed so they don't show up in valgrind.
Change-Id: I9fd1ada5503db9ba481806c09116874ee81f450d
Reviewed-by: Alex Blasche <alexander.blasche@theqtcompany.com>
This has been deprecated since QDBusContext was introduced (Qt 4.3). So
it's time to remove the functionality.
[ChangeLog][Important Behavior Change] QDBusConnection::sender()
(deprecated since Qt 4.3) has changed to always return an invalid
QDBusConnection. To know what connection the incoming call was received
from, use QDBusContext.
Change-Id: I355efb82c14e54ed718c8f892d8267e727b19118
Reviewed-by: Alex Blasche <alexander.blasche@theqtcompany.com>
This makes the output slightly easier to read.
Change-Id: I590b9abcb0263ae5f0580391b42e179c47569a8a
Reviewed-by: Alex Blasche <alexander.blasche@theqtcompany.com>
The loop is there because watchers may have two Watcher for the same
fd, one for read and one for write, but after we're processed the
correct one we don't need to keep looping.
This fixes a crash since it's possible that while in processing
q_dbus_watch_handle we get a watch added/remove this invalidating
the iterator and crashing
Change-Id: Icb61deae272d2f237a4c616fae598404d419af90
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The commit was 9361be58f47ec256bf920c378479a02501219c1f (2008-11-17),
referring to the race condition fix that was applied in commit
d47c05b1889bb4f06203bbc65f4660b8d0128954 (2008-10-08). The fix for the
deadlock reintroduced the race condition and the commit message noted
it.
The workaround is no longer necessary since we've fixed the original race
condition differently now (see the previous two commits).
Task-number: QTBUG-42189
Change-Id: I5a83249597a83c4d4caa2ae57964ad3cc61c1d70
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@digia.com>
Reviewed-by: Albert Astals Cid <albert.astals@canonical.com>
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@theqtcompany.com>
We don't need two anymore because they now protect the same thing: the
state of the DBusConnection. The difference existed when it was possible
for two threads to access the DBusConnection at the same time: one doing
dispatching and one doing something else. Unfortunately, even though
DBusConnection supports this, QtDBus doesn't.
From d47c05b1889bb4f06203bbc65f4660b8d0128954 (2008-10-08):
Details: if we're removing a timer or a watcher from our list,
there's a race condition: one thread (not the QDBusConnection thread)
could be asking for the removal (which causes an event to be sent),
then deletes the pointer. In the meantime, QDBusConnection will
process the timers and socket notifiers and could end up calling
lidbus-1 with deleted pointers.
That commit fixed the race condition but introduced a deadlock.
Task-number: QTBUG-42189
Change-Id: I034038f763cbad3a67398909defd31a23c27c965
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@digia.com>
Reviewed-by: Albert Astals Cid <albert.astals@canonical.com>
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@theqtcompany.com>
We lock it before dbus_connection_send_with_reply (the async version) in
QDBusConnectionPrivate::sendWithReplyAsync. We weren't locking it before
send_with_reply_and_block and we apparently should. The locking around
the dbus_connection_send function might not be necessary, but let's do
it to be safe.
The lock now needs to be recursive because we may be inside
QDBusConnectionPrivate::doDispatch.
Task-number: QTBUG-42189
Change-Id: I7b6b350909359817ea8b3f9c693bced042c9779a
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@digia.com>
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@theqtcompany.com>
It doesn't allocate memory, so cannot throw and is a lot faster
than qgetenv().
Change-Id: Ib0c0f903531a3a656919e87df8065a9c6c7a666c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
It was originally created to avoid allocating memory for the QString at
every turn, but we have QStringLiteral for that today. It has also
served a very good run by catching qatomic.h implementations that had
bad cv qualifications.
Change-Id: Id6d952b8cce363015ec2611d346b4cccedecf137
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@digia.com>
Blocking DBus calls have the potential to totally wreck user interactivity at
best (actions taking too long) and make things appear completely broken at worst
(the default timeout is 30 seconds, which is a huge amount of time, especially
if you get unfortunate and have a repeated blocking call).
Provide a warning when a call is found that takes too long, based on some preset
durations (200ms for the main thread, 500ms for other threads on the basis that
the main thread is generally more important).
Also provide configuration knobs for these environment variables, in miliseconds:
setting them to 0 will warn on all blocking DBus calls. Setting them to -1 (the
default, on release builds) will disable the warning.
[ChangeLog][QtDBus] Blocking calls that take a long time will now generate a
warning. The time taken may be tuned using the environment variables
Q_DBUS_BLOCKING_CALL_MAIN_THREAD_WARNING_MS and
Q_DBUS_BLOCKING_CALL_OTHER_THREAD_WARNING_MS. The value represents (in
milliseconds) how long before a blocking call is warned on. A value below zero
disables the warning, a value of zero will warn on all blocking calls.
Change-Id: I0ab4c34aa01670a154d794d9f2694b3235e789db
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Conflicts:
mkspecs/qnx-x86-qcc/qplatformdefs.h
src/corelib/global/qglobal.h
src/network/socket/qnativesocketengine_winrt.cpp
src/plugins/platforms/android/androidjniaccessibility.cpp
src/plugins/platforms/windows/qwindowswindow.cpp
Manually adjusted:
mkspecs/qnx-armle-v7-qcc/qplatformdefs.h
to include 9ce697f2d5
Thanks goes to Sergio for the qnx mkspecs adjustments.
Change-Id: I53b1fd6bc5bc884e5ee2c2b84975f58171a1cb8e
The trick of creating a copy is not thread-safe. I'd known this since
the moment I wrote that code, but thought "what could go wrong?".
Task-number: QTBUG-39285
Change-Id: If521d4a649c06e6a34926687e85623aa25cb4c35
Reviewed-by: David Faure <david.faure@kdab.com>
A disconnectSignal() call with a wrong signal name caused that hook
wasn't found and thus kept in QDBusConnectionPrivate::signalHooks
forever.
Change-Id: Id7cda225be7580529fc835b377636226abb229f9
Reviewed-by: Aaron McCarthy <mccarthy.aaron@gmail.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This change adds a new method to QDBusServer to allow anonymous
connections. This is part of the DBus API and was not yet possible
to use with QDBusServer. It is set in the newConnection callback
when a new client tries to connect.
Anonymous connections are enabled by default in DBus but not allowed
by default.
[ChangeLog][QtDBus][QDBusServer] Added method to QDBusServer to allow
anonymous client connections, even if the connecting client is not
authenticated as a user.
Change-Id: I984c9e634101ecd2e67bb25c8d12bb1071836fd3
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
the diff -w for this commit is empty.
Started-by: Thiago Macieira <thiago.macieira@intel.com>
Change-Id: I77bb84e71c63ce75e0709e5b94bee18e3ce6ab9e
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Unlike in regular connection to DBus server, we don't get pending call
notifies when a peer drops the connection in peer-to-peer mode.
Thus, we need to keep track of pending calls in such cases and get rid of
them in ~QDBusConnectionPrivate().
Change-Id: I83e20db0bc7b2ebf509c7fdb1382ffc7d0ede9d3
Done-with: Kalle Vahlman <kalle.vahlman@movial.com>
Reviewed-by: Daniele E. Domenichelli <daniele.domenichelli@gmail.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Partial revert of 3c6bb0ed8b and
57aed703d2.
When registering an object that was previously unregistered but not yet
garbage collected, the activeChildren count on the parent node was not
incremented, which could result in other registered objects disappearing
after a later unregisterObject.
Copying objects in the tree is not free, but it's not expensive enough
or used frequently enough to justify that error-prone logic. It's much
safer to simply remove objects immediately.
Change-Id: I3dc59c2ebd07b237518424fcd8ea7371a22d6d15
Reviewed-by: Robin Burchell <robin+qt@viroteck.net>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Fixes race between QDBusConnectionPrivate::processFinishedCall()
releasing the mutex before emitting signals (using various members of
QDBusPendingCallPrivate) and deletion of the QDBusPendingCallPrivate
object through QDBusPendingCall::d's destructor (a member of type
QExplicitlySharedDataPointer<QDBusPendingCallPrivate>) leeds to
segmentation fault with CrashTest example on slow/single core
arm cpu).
Task-number: QTBUG-27809
Change-Id: I3590d74d1cfa5816ede764b50b83a7008ec780ff
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
First step to fix race condition about deleting QDBusPendingCallPrivate.
In a multithreaded application on a slow/single core cpu the following
race (and segmentation fault) can occur:
First thread A is running:
A: QDBusPendingReply<> reply = pi->asyncCallWithArgumentList(method, argumentList);
Then when the dbus answer arrives thread B will call:
B: QDBusConnectionPrivate::processFinishedCall()
B: ...
B: locker.unlock()
and runs until here, go on with thread A:
A: reply.waitForFinished();
A: QDBusPendingCallPrivate::waitForFinished()
A: {
A: QMutexLocker locker(&mutex);
A: if (replyMessage.type() != QDBusMessage::InvalidMessage)
A: return;
which returns immediately (mutex acquired, replyMessage alread set), now
reply goes out of scope (destructor called) and QDBusPendingCall::d's
destructor of type QExplicitlySharedDataPointer<QDBusPendingCallPrivate>
deletes the reference counted object QDBusPendingCallPrivate.
Now thread B continues, still in processFinishedCall()
B: if (call->watcherHelper)
B: call->watcherHelper->emitSignals(msg, call->sentMessage);
B:
B: if (msg.type() == QDBusMessage::ErrorMessage)
B: emit connection->callWithCallbackFailed(QDBusError(msg),
B: call->sentMessage);
accessing alread deleted object QDBusPendingCallPrivate via call->...
Fixed QDBusPendingCallPrivate deletion by proper reference counting
will be done in the next commit.
Task-number: QTBUG-27809
Change-Id: I15b3f0242471b62eaafadc763fb6a33339ff2fe1
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
In QDBusConnectionPrivate::waitForFinished() pcall->pending was used
after the protection by pcall->mutex was released. A simultaneous
call to QDBusConnectionPrivate::processFinishedCall() was able
to reset pcall->pending to null before it was used for the
q_dbus_pending_call_block(pcall->pending) call.
Fixed by releasing (and setting to 0) of pcall->pending in
processFinishedCall() only in case no one is waiting yet, otherwise
release pcall->pending by the first thread waiting in waitForFinished().
There is still a race condition about deleting QDBusPendingCallPrivate
(too early) which will be fixed in the next two commits.
Task-number: QTBUG-27809
Change-Id: I040173810ad90653fe1bd1915f22d8dd70d47d8c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
If the connection failed, the DBusConnection object is null, but we
still add our QDBusConnectionPrivate to the global hash (maybe we
shouldn't). Both disconnectFromXXX functions check that they are
disconnecting a connection of the right type, but we never initialized
the type if the connection failed.
So simply make sure we initialize before handling the error state.
Task-number: QTBUG-27973
Change-Id: I96f4825ab1b71adf1b72caf4f72db41742b44a55
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@digia.com>
For instance, exporting an object with a slot like
QString complexMethod(const MyVariantMap& vars);
(even with a simple typedef QVariantMap MyVariantMap)
used to silently skip that method in the introspection.
Now it outputs:
generateInterfaceXml: Skipped method "complexMethod" : Invalid type in parameter list: MyVariantMap
Change-Id: I7964cfb63e973257ce1abe47b9625e361b2ad23f
Reviewed-by: Kevin Ottens <kevin.ottens@kdab.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This has been broken forever, just like generic signal
disconnection. It didn't use to show up before because in Qt 4,
QObject's destructor would not call disconnectNotify().
Just like in the previous commit, we need to verify whether the signal
was disconnected from the last receiver. A wildcard disconnect might
be disconnecting only from a specific receiver.
Task-number: QTBUG-29498
Change-Id: I0790128ea878fdf3ac563c99d96c6aa7d270e9a3
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@digia.com>
The dbus_watch_get_fd function was deprecated in D-Bus 1.2 (technically,
in 1.1.1, but that was a development release) because it had a bad name.
Sockets on Windows have file descriptors, but they are not shared from
the same pool as the CRT library's file descriptors.
This commit raises the minimum required version of D-Bus to 1.2. This is
the first requirement raise since this code was introduced in 2006. For
some reason, the D-Bus 1.2.0 release seems to be missing, but 1.2.1 was
released on 04-Apr-2008. That's ancient enough for all distributions
Qt 5 is supposed to run on.
Change-Id: Ia6bbc137fffbb27c77290ed3e32d3380f0ae3c54
Reviewed-by: Lorn Potter <lorn.potter@jollamobile.com>
Change-Id: I18697037db742d38874c8a95df12c189ccc51068
Reviewed-by: Frederik Gladhorn <frederik.gladhorn@digia.com>
Reviewed-by: Lorn Potter <lorn.potter@jollamobile.com>
This replaces the implementation from ac9ab9703f.
If the number of active children drops to zero, we know we can simply
delete the vector of children. We know none that might be there are
active.
If the number is not zero, but is considerably smaller than the vector
size, we can shrink the vector by reordering the elements, skipping the
inactive ones.
We use qMove, which expands to std::move on C++11, but a regular copy on
C++98.
Change-Id: I2e74446081f91fbd698425b08910fbda4746d673
Reviewed-by: David Faure (KDE) <faure@kde.org>
The new member variable activeChildren shall contain the number of
direct children that are active. This number differs from
children.count() because the vector may contain empty entries that
haven't been garbage-collected yet (obj == NULL and activeChildren ==
0).
When this count drops to zero, we know we can simply erase the vector of
children.
Change-Id: Ia20604d3fac852ea4a6e8862d934fbb936fa5e18
Reviewed-by: David Faure (KDE) <faure@kde.org>
The current implementation is a loop. We need it to be recursive so that
we can execute more operations when unwinding. This will be necessary
in the next commit.
Change-Id: Ia3c98fed0719cede0a0d92d3e343cf016ec7baf2
Reviewed-by: David Faure (KDE) <faure@kde.org>
Fixes performance issues in apps which register and deregister objects
very frequently (like nepomukstorage).
Change-Id: Ib4ce8d65868f0e26cd45f1053e4b2f4c13528cfa
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The dbus_get_version function was introduced in 1.2, so we'd need to
detect pre-1.2 by the absence of the function. But if we're going to
detect the presence or absence of any function, we might as well do it
on dbus_connection_can_send_type, which is the function we wanted anyway.
Change-Id: I6e17a3a8f1382c6a489490084f6e3f61aa5a1947
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>
Those were using an inconsistent macro compared to the function
declaration in qdbusthreaddebug_p.h.
Change-Id: I3e77ba83ceedc99cb1f957fdfe318e34ab9c9628
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
Change copyrights and license headers from Nokia to Digia
Change-Id: If1cc974286d29fd01ec6c19dd4719a67f4c3f00e
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Reviewed-by: Sergio Ahumada <sergio.ahumada@digia.com>
There's not much point in caching the result of qMetaTypeId<>,
because it's already internally memoised.
In addition, the code that initialised the static int caches wasn't
protected against concurrent access under the assumption that the
operations performed were thread-safe.
That is true for most of them, but not for the stores to the static ints,
which race against each other:
// Thread A // Thread B
r1 = initialized /*=false*/
r1 = initialized /*=false*/
r2 = qMetaTypeId<...>();
r2 = qMetaTypeId<...>();
message = r2; message = r2; // race, ditto for all other ints
To fix, turn the ints into inline functions that just call the respective
qMetaTypeId<>() function.
Change-Id: I5aa80c624872c3867232abc26ffdcde70cd54022
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
cpp files should include their own headers first (but below config.h)
Change-Id: I4115604aee3211118e2ecf604067f3559dbb9f4c
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Replace them with std::lower_bound; this allows for deprecation of
qLowerBound.
Change-Id: I536e7338eb85ea6c7c1a5bf23121292767927e0b
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Robin Burchell <robin+qt@viroteck.net>
Frederik Gladhorn
Thiago Macieira
Simon Hausmann
Albert Astals Cid
Oswald Buddenhagen
Marc Mutz
Matti Paaso
Robin Burchell
Frederik Gladhorn
David Faure
Jakub Adam
Roland Winklmeier
John Brooks
Peter Seiderer
David Faure
Sze Howe Koh
Sergio Ahumada
David Faure
Iikka Eklund
Giuseppe D'Angelo