On 04/22/2010 09:19 AM, Stephen Shaw wrote:
On Thu, Apr 22, 2010 at 07:09, Daniel P. Berrange <berrange@redhat.com> wrote:
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
Thanks for taking care of this. As soon as the fix is checked in I'll rebuild libvirt and continue testing it.
I still have to go over the rules more completely, but the attached patch seems to do more or less the right thing for me. If you could give it a quick test, that would be great. Thanks, -- Chris Lalancette