"Daniel P. Berrange" <berrange@redhat.com> wrote:
On Wed, Jan 30, 2008 at 06:25:19PM +0100, Jim Meyering wrote:
Without the following patch, this command would hang
printf 'domuuid fc4\ndomstate fc4\n' \ | ./virsh --connect test://$PWD/../docs/testnode.xml
with this stack trace:
__lll_lock_wait ... _L_lock_105 ... __pthread_mutex_lock ... virUnrefDomain (domain=0x6a8b30) at hash.c:884 virDomainFree (domain=0x6a8b30) at libvirt.c:1211 cmdDomstate (ctl=0x7fffea723390, cmd=0x6a8b10) at virsh.c:677 vshCommandRun (ctl=0x7fffea723390, cmd=0x6a8b10) at virsh.c:4033 main (argc=3, argv=0x7fffea7234e8) at virsh.c:501
The problem is that cmdDomuuid didn't call virDomainFree(dom), so cmdDomstate hangs trying to do the Unref.
I don't understand why it would hang simply because it forgot to free an object. Surely it ought to just be a memory leak not hang ? Each individual libvirt API must always have a matched lock & unlock pair in all codepaths, so a hang should only occur if an unlock is missing in some codepath.
Well maybe virDomainFree is misnamed then. It does more: The missing virDomainFree calls cause that because they are what is supposed to do the unlock. virDomainFree calls virUnrefDomain, which calls virReleaseDomain (when refs hits 0), which calls pthread_mutex_unlock.