[Libvirt-ci] Still Failing: libvirt/libvirt#1718 (master - 7882c6e)

[Travis CI]

Build Update for libvirt/libvirt
-------------------------------------

Build: #1718
Status: Still Failing

Duration: 19 mins and 18 secs
Commit: 7882c6e (master)
Author: Mark Asselstine
Message: lxc_monitor: Avoid AB / BA lock race

A deadlock situation can occur when autostarting a LXC domain 'guest'
due to two threads attempting to take opposing locks while holding
opposing locks (AB BA problem). Thread A takes and holds the 'vm' lock
while attempting to take the 'client' lock, meanwhile, thread B takes
and holds the 'client' lock while attempting to take the 'vm' lock.

The potential for this can be seen as follows:

Thread A:
virLXCProcessAutostartDomain (takes vm lock)
 --> virLXCProcessStart
  --> virLXCProcessConnectMonitor
   --> virLXCMonitorNew
    --> virNetClientSetCloseCallback (wants client lock)

Thread B:
virNetClientIncomingEvent (takes client lock)
 --> virNetClientIOHandleInput
  --> virNetClientCallDispatch
   --> virNetClientCallDispatchMessage
    --> virNetClientProgramDispatch
     --> virLXCMonitorHandleEventInit
      --> virLXCProcessMonitorInitNotify (wants vm lock)

Since these threads are scheduled independently and are preemptible it
is possible for the deadlock scenario to occur where each thread locks
their first lock but both will fail to get their second lock and just
spin forever. You get something like:

virLXCProcessAutostartDomain (takes vm lock)
 --> virLXCProcessStart
  --> virLXCProcessConnectMonitor
   --> virLXCMonitorNew
<...>
virNetClientIncomingEvent (takes client lock)
 --> virNetClientIOHandleInput
  --> virNetClientCallDispatch
   --> virNetClientCallDispatchMessage
    --> virNetClientProgramDispatch
     --> virLXCMonitorHandleEventInit
      --> virLXCProcessMonitorInitNotify (wants vm lock but spins)
<...>
    --> virNetClientSetCloseCallback (wants client lock but spins)

Neither thread ever gets the lock it needs to be able to continue
while holding the lock that the other thread needs.

The actual window for preemption which can cause this deadlock is
rather small, between the calls to virNetClientProgramNew() and
execution of virNetClientSetCloseCallback(), both in
virLXCMonitorNew(). But it can be seen in real world use that this
small window is enough.

By moving the call to virNetClientSetCloseCallback() ahead of
virNetClientProgramNew() we can close any possible chance of the
deadlock taking place. There should be no other implications to the
move since the close callback (in the unlikely event was called) will
spin on the vm lock. The remaining work that takes place between the
old call location of virNetClientSetCloseCallback() and the new
location is unaffected by the move.

Signed-off-by: Mark Asselstine <mark.asselstine at windriver.com>
Signed-off-by: Michal Privoznik <mprivozn at redhat.com>

View the changeset: https://github.com/libvirt/libvirt/compare/65ba48d26745...7882c6eca53f

View the full build log and details: https://travis-ci.org/libvirt/libvirt/builds/433006111?utm_medium=notification&utm_source=email

--

You can unsubscribe from build emails from the libvirt/libvirt repository going to https://travis-ci.org/account/preferences/unsubscribe?repository=4872032&utm_medium=notification&utm_source=email.
Or unsubscribe from *all* email updating your settings at https://travis-ci.org/account/preferences/unsubscribe?utm_medium=notification&utm_source=email.
Or configure specific recipients for build notifications in your .travis.yml file. See https://docs.travis-ci.com/user/notifications.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/libvirt-ci/attachments/20180925/8b9a3c15/attachment.htm>