Hi Jason,
the commit doesn't seem correct, because looks like you removed the
locking completely. It should still stay with locking on reply mutex.
Can you revert the last commit? Then I can take care of removing the
unnecessary code -- note that the code before your commit was already
using the reply lock, so was like it should just be, with dedicated
mutex code being inside defines which were not enabled.
Cheers,
Daniel
On 13/05/15 08:56, Jason Penton wrote:
Hey Daniel,
You are correct this is no longer needed - we actually only used this
to be able to set a flag on the transaction that we later used for
branch picking. We made a fix a while back for doing the correct
branch picking which no longer required this extra flag and therefore
no longer required the mutex. I have tested in our env and all looks good.
p.s. I have committed the cleanup.
Cheers
Jason
On Tue, 12 May 2015 at 10:34 Daniel-Constantin Mierla
<miconda(a)gmail.com <mailto:miconda@gmail.com>> wrote:
Hi Jason,
ok, would be great to sort it out properly, thanks for your time!
Cheers,
Daniel
On 12/05/15 10:10, Jason Penton wrote:
Hey Daniel,
Okay great, let me look into this. It will be great if we have
one less mutex to worry about ;) - If not required I will remove
and commit.
Cheers
Jason
On Mon, 11 May 2015 at 09:55 Daniel-Constantin Mierla
<miconda(a)gmail.com <mailto:miconda@gmail.com>> wrote:
Hi Jason,
over the weekend I pushed a patch that disables the use of
dedicated
mutex for t_continue(). It can be enabled by defining
ENABLE_ASYNC_MUTEX.
While investigated some reports of crash when removing from
time, I
found the potential of a race when t_coninue() is executed at
the time
the fr_timer for suspended transaction elapsed. The timer
process will
get the transaction out and remove it from timer under the
reply lock
and the worker doing t_continue() will get it out under the
async lock.
I looked at the commit you did when introducing the dedicated
async
mutex, the note being:
- "dedicated lock to prevent multiple invocations of
suspend on
tz (reply lock used to be used)"
Perhaps tz is tx and stands for transmission - however, the
reply lock
should be safe for this case as well. Moreover, the continue
is like the
suspended branch got a reply and transaction continues
processing, which
implies the reply lock is aquired (like execution of
failure_route,
which can also happen if fr_timer elapses before t_continue()
is executed).
Given those, I don't see anymore a reason for dedicated async
mutex.
Also, it protects to races of using two mutexes, which can
easily lead
to deadlocks (e.g., one process acquires the reply lock and
tries to get
the async lock while another one wanted first the reply lock
and later
the async lock).
For now I disabled the code with defines, as I wanted to
discuss and be
sure I haven't overlooked any issue you tried to avoid with the
dedicated mutex. Let me know what you think about.
Cheers,
Daniel
--
Daniel-Constantin Mierla
http://twitter.com/#!/miconda
<http://twitter.com/#%21/miconda> -
http://www.linkedin.com/in/miconda
Kamailio World Conference, May 27-29, 2015
Berlin, Germany -
http://www.kamailioworld.com
--
Daniel-Constantin Mierla
http://twitter.com/#!/miconda <http://twitter.com/#%21/miconda> -
http://www.linkedin.com/in/miconda
Kamailio World Conference, May 27-29, 2015
Berlin, Germany -
http://www.kamailioworld.com