Re: [libvirt] [PATCH] Release VM lock before acquiring virDomainObjListPtr lock

On Tue, Apr 09, 2013 at 01:19:05PM +0200, Peter Krempa wrote: > On 04/09/13 12:26, Daniel P. Berrange wrote: >> On Tue, Apr 09, 2013 at 10:08:31AM +0100, Daniel P. Berrange wrote: >>> On Mon, Apr 08, 2013 at 09:15:28PM -0600, Eric Blake wrote: >>>> On 02/11/2013 09:46 AM, Daniel P. Berrange wrote: >>>>> From: "Daniel P. Berrange" <berrange(a)redhat.com&gt; >>>>> >>>>> When removing a VM from the virDomainObjListPtr, we must not >>>>> be holding the VM lock while acquiring the list lock. Re-order >>>>> code to ensure that we can release the VM lock early. >>>>> --- >>>>> src/conf/domain_conf.c | 3 +-- >>>>> 1 file changed, 1 insertion(+), 2 deletions(-) >>>>> >>>>> diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c >>>>> index 5e16ddf..d92e54a 100644 >>>>> --- a/src/conf/domain_conf.c >>>>> +++ b/src/conf/domain_conf.c >>>>> @@ -2115,11 +2115,10 @@ void virDomainObjListRemove(virDomainObjListPtr doms, >>>>> { >>>>> char uuidstr[VIR_UUID_STRING_BUFLEN]; >>>>> >>>>> - virObjectLock(doms); >>>>> virUUIDFormat(dom->def->uuid, uuidstr); >>>>> - >>>>> virObjectUnlock(dom); >>>>> >>>>> + virObjectLock(doms); >>>> >>>> This patch seems to be implicated in Peter's latest proof of a >>>> use-after-free data race: >>>> https://www.redhat.com/archives/libvir-list/2013-April/msg00674.html >>> >>> The caller of this method should own a reference on virDomainObjPtr. >>> This method will result in that reference being released, as the >>> dom is removed from the list. >>> >>> If any other thread experiances a use-after-free, then that thread >>> must have been missing a reference. >> >> Actually, it is this code which is at fault. I forgot the rule >> that if you unlock 'dom', then you are forbidden to access it >> again, unless you had done a virDomainObjRef on it prior to >> unlocking. Once we hold the lock on 'doms', we must also relock >> 'dom' to ensure we block any other thread from using it. So >> this method should be changed to look like >> >> void virDomainObjListRemove(virDomainObjListPtr doms, >> virDomainObjPtr dom) >> { >> char uuidstr[VIR_UUID_STRING_BUFLEN]; >> >> virUUIDFormat(dom->def->uuid, uuidstr); >> virObjectRef(dom); >> virObjectUnlock(dom); >> >> virObjectLock(doms); >> virObjectLock(dom); >> virHashRemoveEntry(doms->objs, uuidstr); >> virObjectUnlock(dom); >> virObjectUnref(dom); >> virObjectUnlock(doms); >> } > > Hmm, yeah, this seems to be the correct minimal approach. Are you > going to post the patch separately? Since you can reproduce the problem can you post & test a patch todo this.

> Anyways I still think that we should re-work this to the reference > counting approach as it would allow us to avoid having to take the > lock of the domain object and potentially having to wait with the > domain objects hash lock until an API finishes and thus blocking > other domain object lookups. It would also avoid the need to check > if the exit from a qemu job freed the domain object. We would still have to hold the lock, since the lock is there to ensure two things don't read/modify fields in virDomainObjPtr at the same time.

It would, however, mean that the {Enter/Leave}Monitor fnuctions do not need to play games with the reference count, and we could simplify the code which follows since we can assume our 'dom' is always valid and not have to set it to NULL after exiting jobs.

Daniel