Re: [libvirt] seg fault when running snapshot-create

On Thu, Apr 22, 2010 at 08:54:30AM -0400, Chris Lalancette wrote: > On 04/22/2010 08:34 AM, Daniel P. Berrange wrote: > > On Wed, Apr 21, 2010 at 05:16:21PM -0400, Chris Lalancette wrote: > >> On 04/21/2010 04:34 PM, Stephen Shaw wrote: > >>> I'm getting a seg fault when running virsh snapshot-create 1, but only > >>> when virt-manager is open and connected. > >>> > >>> Here is some of the debug info I was able to come up with - > >>> http://fpaste.org/9GO6/ (bt) > >>> http://fpaste.org/7gkH/ ('thread apply all bt) > >>> > >>> * After the crash > >>> (gdb) p mon->msg > >>> $1 = (qemuMonitorMessagePtr) 0x0 > >>> > >>> > >>> nibbler:~ # libvirtd --version > >>> libvirtd (libvirt) 0.8.0 > >>> > >>> > >>> Please let me know if there is any other information you need. > >>> Stephen > >> > >> Thanks for the report.  To be perfectly honest, I can't see how what > >> happened could happen :).  But I'll take a closer look at it and see > >> if I can reproduce and see what is going on with it. > > > > I see thread locking problems in the code > > > >  - qemuDomainSnapshotCreateXML() is calling monitor commands, but has > >    not run  qemuDomainObjBeginJobWithDriver() to ensure exclusive > >    access to the monitor > > > >  - qemuDomainSnapshotDiscard has same problem > > Yep, just fixing those now.  I didn't quite understand the ObjBeginJob > before, but I think I'm understanding it now.  This is probably the source of > the problems. There's some notes about the rules in src/qemu/THREADS.txt. You must acquire locks on objects in the following order, not missing any steps.   qemud_driver             (qemudDriverLock)   virDomainObjPtr          (implicit via virDomainObjFindByXXX)   qemuMonitorPrivatePtr    (qemuDomainObjBeginJob)   qemuMonitorPtr           (implicit via qemuDomainObjEnterMonitor) Note that qemuDomainObjEnterMonitor() will release the locks on the qemud_driver & virDomainObjPtr objects, once it has acquired the lock on qemuMonitorPtr. The qemuMonitorPrivatePtr object has a condition variable that ensures continued mutual exclusion, even though the qemud_driver, virDomainObjPtr object are now unlocked. So by missing qemuDomainObjBeginJob(), the condition variable was not acquired, and mutual exclusion was not assured Regards, Daniel