Control: tag -1 patch
Jens:
It looks like the 1.2.3-349-g315b5f5-2.2+deb7u1 version in Wheezy also needs a
patch to implement upstream commit dee463ef52d8406d0a925facfabead616f0f9dc2
like was done for the 1.2.4-0.2 version currently in Unstable and Testing.
https://github.com/mumble-voip/mumble/commit/dee463ef52d8406d0a925facfabead616f0f9dc2
The same patch used for 1.2.4-0.2 seems to apply fine. Attached.
If you get a chance, try this patch on 1.2.3-349-g315b5f5-2.2+deb7u1 just to
verify that this also fixes the problem.
I'll see if I can queue up a fix for this for the .5 Wheezy point release.
-- Chris
--
Chris Knadle
chris.kna...@coredump.us
Author: Mikkel Krautz <mik...@krautz.dk>
Applied-Upstream: commit dee463ef52d8406d0a925facfabead616f0f9dc2
Date: Sat Jun 8 14:35:34 2013 +0200
Last-Updated: 2013-11-08
Description: Fix Avahi libdns_sd deadlock in Bounjour code
bonjour: use Qt::AutoConnection for BonjourServiceResolver's QSocketNotifier slot.
Using a QueuedConnection for the slot had the unpleasant
side effect that the QSocketNotifier could have its activated()
slot invoked even though no data was waiting to be read.
In our case, this could cause a deadlock inside Avahi's
libdns_sd compatibility library.
I've settled on using Qt::AutoConnection to be consitent
with the rest of the code base. The Bonjour code should
always be invoked from the main thread, so in this case
Qt::AutoConnection will always mean Qt::DirectConnection.
Why does this happen? Qt seems to process events before
invoking queued slot invocations. If the Qt event loop
finds that the file descriptor that our QSocketNotifier
is providing notification for is ready for reading, it
queues up an invocation of the activated() slot for the
next event loop iteration (because we use a QueuedConnection).
As mentioned above, because Qt seems to poll() FDs before
invoking queued-up slots, the end result is that an
invocation of the activated() slot for a given
QSocketNotifier's file descriptor can be queued up in
the very same event loop iteration that a read() is
performed for the exact same file descriptor.
After performing the read(), the queued-up activated()
slot invocation is no longer valid, and can wreak havoc,
which in our case causes a deadlock in the Avahi libdns_sd
code.
The flow below describes the event loop iterations
in more detail:
1st event loop iteration
------------------------
* poll() is invoked; the QSocketNotifier's FD is
ready for reading.
* An invocation of the activated() slot is queued
up, to be executed next time we enter the event
loop (due to Qt::QueuedConnection).
2nd event loop iteration
------------------------
* poll() is invoked; the QSocketNotifier's FD is
_still_ ready for reading.
* An invocation of the activated() slot is again
queued up, to be executed in the 3rd iteration.
* The queued-up slot invocation from the 1st iteration
is invoked. (read() is called.)
3rd event loop iteration
------------------------
* poll is invoked(); the QSocketNotifier's FD has
nothing to read anymore. Everything was read in
the activated() slot that was invoked in the
2nd iteration.
* The queued-up slot invocation from the 2nd iteration
is invoked. This time, the read() syscall will block,
because there is nothing to read.
diff --git a/src/bonjour/BonjourServiceResolver.cpp b/src/bonjour/BonjourServiceResolver.cpp
index 7b5f632..99f2bc1 100644
--- a/src/bonjour/BonjourServiceResolver.cpp
+++ b/src/bonjour/BonjourServiceResolver.cpp
@@ -63,7 +63,7 @@ void BonjourServiceResolver::resolveBonjourRecord(const BonjourRecord &record) {
err = kDNSServiceErr_Invalid;
} else {
rr->bonjourSocket = new QSocketNotifier(sockfd, QSocketNotifier::Read, this);
- connect(rr->bonjourSocket, SIGNAL(activated(int)), this, SLOT(bonjourSocketReadyRead(int)), Qt::QueuedConnection);
+ connect(rr->bonjourSocket, SIGNAL(activated(int)), this, SLOT(bonjourSocketReadyRead(int)));
qmResolvers.insert(sockfd, rr);
return;
}