On Thu, Mar 14, 2019 at 13:22:38 +0100, Michal Privoznik wrote: > https://bugzilla.redhat.com/show_bug.cgi?id=1623389 > > If a device is detached twice from the same domain the following > race condition may happen: > > 1) The first DetachDevice() call will issue "device_del" on qemu > monitor, but since the DEVICE_DELETED event did not arrive in > time, the API ends claiming "Device detach request sent > successfully". > > 2) The second DetachDevice() therefore still find the device in > the domain and thus proceeds to detaching it again. It calls > EnterMonitor() and qemuMonitorSend() trying to issue "device_del" > command again. This gets both domain lock and monitor lock > released. > > 3) At this point, qemu sends us the DEVICE_DELETED event which is > going to be handled by the event loop which ends up calling > qemuDomainSignalDeviceRemoval() to determine who is going to > remove the device from domain definition. Whether it is the > caller that marked the device for removal or whether it is going > to be the event processing thread. > > 4) Because the device was marked for removal, > qemuDomainSignalDeviceRemoval() returns true, which means the > event is to be processed by the thread that has marked the device > for removal (and is currently still trying to issue "device_del" > command) > > 5) The thread finally issues the "device_del" command, which > fails (obviously) and therefore it calls > qemuDomainResetDeviceRemoval() to reset the device marking and > quits immediately after, NOT removing any device from the domain > definition. > > At this point, the device is still present in the domain > definition but doesn't exist in qemu anymore. Worse, there is no > way to remove it from the domain definition. > > Solution is to note down that we've seen the event and if the > second "device_del" fails, not take it as a failure but carry on > with the usual execution. > > Signed-off-by: Michal Privoznik <mprivozn@xxxxxxxxxx> > --- > src/qemu/qemu_domain.h | 1 + > src/qemu/qemu_hotplug.c | 52 +++++++++++++++++++++++++++++++++++++++-- > 2 files changed, 51 insertions(+), 2 deletions(-) > > diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h > index 9f468e5661..fb361515ba 100644 > --- a/src/qemu/qemu_domain.h > +++ b/src/qemu/qemu_domain.h > @@ -218,6 +218,7 @@ typedef qemuDomainUnpluggingDevice *qemuDomainUnpluggingDevicePtr; > struct _qemuDomainUnpluggingDevice { > const char *alias; > qemuDomainUnpluggingDeviceStatus status; > + bool eventSeen; /* True if DEVICE_DELETED event arrived. */ > }; > > > diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c > index 59a5d21a5b..6b5b37bda5 100644 > --- a/src/qemu/qemu_hotplug.c > +++ b/src/qemu/qemu_hotplug.c > @@ -67,6 +67,9 @@ VIR_LOG_INIT("qemu.qemu_hotplug"); > unsigned long long qemuDomainRemoveDeviceWaitTime = 1000ull * 5; > > > +static void > +qemuDomainResetDeviceRemoval(virDomainObjPtr vm); > + > /** > * qemuDomainDeleteDevice: > * @vm: domain object > @@ -103,8 +106,51 @@ qemuDomainDeleteDevice(virDomainObjPtr vm, > } > > if (enterMonitor && > - qemuDomainObjExitMonitor(driver, vm) < 0) > - rc = -1; > + qemuDomainObjExitMonitor(driver, vm) < 0) { > + /* Domain is no longer running. No cleanup needed. */ > + return -1; > + } > + > + if (rc < 0) { > + bool eventSeen; > + > + /* Deleting device failed. Let's check if DEVICE_DELETED > + * even arrived. If it did, we need to claim success to > + * make the caller remove device from domain XML. */ > + if (enterMonitor) { > + /* Here @vm is locked again. It's safe to access > + * private data directly. */ > + } else { > + /* Here @vm is not locked. Do some locking magic to > + * be able to access private data. It is safe to lock > + * and unlock both @mon and @vm here because: > + * a) qemuDomainObjEnterMonitor() ensures @mon is ref()'d > + * b) The API that is calling us ensures that @vm is ref()'d > + */ As I've said in review for patch 1. This spiel is not necessary and only relevant in the ZPCI extension device case which will not be the device which the event is meant for. > + virObjectUnlock(priv->mon); > + virObjectLock(vm); > + virObjectLock(priv->mon); > + } > + > + eventSeen = priv->unplug.eventSeen; Saving this into a temporary variable used only here does not make much sense. > + if (eventSeen) { > + /* The event arrived. Return success. */ > + VIR_DEBUG("Detaching of device %s failed, but event arrived", alias); > + qemuDomainResetDeviceRemoval(vm); > + rc = 0; > + } else if (rc == -2) { > + /* The device does not exist in qemu, but it still > + * exists in libvirt. Claim success to make caller > + * qemuDomainWaitForDeviceRemoval(). Otherwise if > + * domain XML is queried right after detach API the > + * device would still be there. */ > + VIR_DEBUG("Detaching of device %s failed and no event arrived", alias); > + rc = 0; How can this be considered success? Also this introduces a possible regression. The DEVICE_DELETED event should be fired only after the device was entirely unplugged. Claiming success before seeing the event can lead to another race when qemu deleted the device from the internal list so that 'device_del' does not see it any more but did not finish cleanup fully. We need to start the '*Remove' handler only after the DEVICE_DELETED event was received. > + } > + > + if (!enterMonitor) > + virObjectUnlock(vm); > + } > > return rc; > } > @@ -5186,6 +5232,7 @@ qemuDomainResetDeviceRemoval(virDomainObjPtr vm) > { > qemuDomainObjPrivatePtr priv = vm->privateData; > priv->unplug.alias = NULL; > + priv->unplug.eventSeen = false; > } > > /* Returns: > @@ -5245,6 +5292,7 @@ qemuDomainSignalDeviceRemoval(virDomainObjPtr vm, > VIR_DEBUG("Removal of device '%s' continues in waiting thread", devAlias); > qemuDomainResetDeviceRemoval(vm); > priv->unplug.status = status; > + priv->unplug.eventSeen = true; > virDomainObjBroadcast(vm); > return true; > } > -- > 2.19.2 > > -- > libvir-list mailing list > libvir-list@xxxxxxxxxx > https://www.redhat.com/mailman/listinfo/libvir-list
Attachment:
signature.asc
Description: PGP signature
-- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list
