[libvirt] [PATCH RFC v2] qemu: fix deadlock when waiting in non async jobs

8 Oct 2018

On 17.10.2018 23:04, John Ferlan wrote:
> 
> 
> On 10/16/18 3:22 AM, Nikolay Shirokovskiy wrote:
>>
>>
>> On 16.10.2018 03:00, John Ferlan wrote:
>>>
>>>
>>> On 10/8/18 4:10 AM, Nikolay Shirokovskiy wrote:
>>>> Block job abort operation can not handle properly qemu crashes when waiting for
>>>> abort/pivot completion. Deadlock scenario is next:
>>>>
>>>> - qemuDomainBlockJobAbort waits for pivot/abort completion
>>>> - qemu crashes, then qemuProcessBeginStopJob broadcasts for VM condition and
>>>>   then waits for job condition (taken by qemuDomainBlockJobAbort)
>>>> - qemuDomainBlockJobAbort awakes but nothing really changed, VM is still
>>>>   active (vm->def->id != -1) so thread starts waiting for completion again.
>>>>   Now two threads are in deadlock.
>>>>
>>>> First let's remove broadcast in qemuProcessBeginStopJob. It is simply wrong
>>>> because it is not set any condition before broadcast so that awaked threads can
>>>> not detect any changes. Crashing domain during async job will continue to be
>>>> handled properly because destroy job can run concurrently with async job and
>>>> destroy job calls qemuProcessStop which sets vm->def->id to -1 and broadcasts.
>>>
>>> Hmm... Although blockjobs are not my area of expertise, I do seem to
>>> have a knack for reading and commenting on patches with these edge
>>> conditions.
>>>
>>> At first, taken alone this made it seem like separate patches are
>>> required, but maybe not depending on the relationship described above.
>>> As an aside, for this paragraph hunk you could call out commit 4d0c535a3
>>> where this is/was introduced. Beyond the refactor, the broadcast was
>>> added; however, it seems it was done so on purpose since the broadcast
>>> would seemingly allowing something to be awoken.
>>>
>>> Beyond that - take away the scenario you describing where QEMU crashes.
>>> In the normal path, if you remove the broadcast, then do things work
>>> properly?
>>
>> As far as I can see. In all jobs where we we wait on vm condition we
>> check misc state variables after that so if state is not changed than
>> broadcasting will not help. (The only exception is migration and derivatives
>> with qemu not supporting events but in this case we use sleeps and
>> do not wait).
>>
> 
> To be clear, you are referencing virDomainObjWait[Until] callers that
> aren't found within qemu_migration.c.

Not exactly. I mean look at any code that waits with virDomainObjWait[Until].
That removed broadcast won't help them to finish waiting because the only
action they take after awake is checking state variable. They don't for
example send monitor commands in which case they would get "monitor closed" error
and finishi waiting.

> 
> The two qemu_hotplug.c examples are waiting for QEMU events related to
> tray eject or device removal, but are limited in their duration. If they
> don't get the event in the specified time they have their means to
> signify the timeout.
> 
> The two qemu_driver.c examples are possibly waiting forever. One waits
> for an external event to signify a memory dump is complete via
> qemuProcessHandleDumpCompleted or the job aborted. The other waits for
> the blockjob to be completed when qemuBlockJobEventProcess clear the
> flag. Both should also theoretically fail when the domain or qemu dies;
> however, since both use virDomainObjWait which when properly tickled by
> broadcast will call virDomainObjIsActive to compare vm->def->id != -1.
> 
> So the contention (for both I think) is that because the = -1 is not
> done when QEMU is killed we're stuck. So rather than wait on something
> that won't happen - use the EOF event as a way to force exit once a
> broadcast happens via a qemu_domain specific qemuDomainObjWait.
> 
> Is that a "fair summary"?
> 

Yes. But this approach is taken only for non async jobs. Async jobs are
ok because they can run concurrently with destroy job.

>>>
>>> Since a block job would set @priv->blockjob when a block job starts and
>>> is cleared during qemuBlockJobEventProcess processing when status is
>>> VIR_DOMAIN_BLOCK_JOB_COMPLETED, VIR_DOMAIN_BLOCK_JOB_FAILED, or
>>> VIR_DOMAIN_BLOCK_JOB_CANCELED.
>>>
>>> What about setting a @priv->blockjobAbort when the abort starts. Then
>>> perhaps processMonitorEOFEvent or qemuProcessHandleMonitorEOF can handle
>>> that properly so that we don't deadlock.
>>
>> But how we can handle it the other way? I see no other option now besides
>> setting some state variable and signalling after that to help non async
>> job to finish and then let EOF handler proceed. (However check suggestions after --- )
>>
> 
> OK, so perhaps that's just a rename of your @monEOF, but specific to the
> example, but based on what I typed above would seemingly not be enough,
> so let's stick with monEOF...
> 
>> Nikolay
>>
>>>
>>> Perhaps or hopefully, Jirka or Peter will comment too with this bump.
>>>
>>> John
>>>
>>>>
>>>> Second let's introduce flag that EOF is received and broadcast after that.
>>>> Now non async jobs can check this flag in wait loop.
>>>>
>>>> Signed-off-by: Nikolay Shirokovskiy <nshirokovskiy@virtuozzo.com>
>>>>
>>>> ---
>>>>
>>>> Diff from v1:
> 
> 
> Just making sure - this RFC v2 comes from a series in Apr/May of this year:
> 
> https://www.redhat.com/archives/libvir-list/2018-April/msg01752.html
> 
> w/ review dialog spilling into May:
> https://www.redhat.com/archives/libvir-list/2018-May/msg00126.html
> 
> based on the series :
> 
> https://www.redhat.com/archives/libvir-list/2018-April/msg01713.html
> 
> that you SNACK'd.
> 
> An awful long time to remember context!  When you reference earlier
> patches, please try to remember to place a link in the cover - it makes
> it easier to find. Even if it's only a couple of days.  Weeks and months
> from now someone may reference the series and want to peruse the history
> of the previous review comments.

I will.

> 
> 
>>>>
>>>> - patches 1 and 2 are already merged
>>>> - don't bother with reporting monitor EOF reason to user as most of
>>>>   time it is simply "unexpected eof" (this implies dropping patch 3)
>>>> - drop patch 5 as we now always report "domain is being stopped"
>>>>   in qemuDomainObjWait
>>>> - don't signal on monitor error for simplicity (otherwise we need to report
>>>>   something more elaborate that "domain is being stopped" as we don't
>>>>   kill domain on monitor errors. On the other hand I guess monitor
>>>>   error is rare case to handle it right now)
>>>> - keep virDomainObjWait for async jobs
>>>>
>>>> It's a bit uneven that for async jobs domain is destroyed concurrently and for
>>>> non async jobs it will be actually destroyed after job get completed.  Also if
>>>> non async job needs issuing commands to qemu on cleanup then we will send these
>>>> commands in vain polluting logs etc because qemu process in not running at this
>>>> moment but typical check (virDomainObjIsActive) will think it is still running.
>>>>
>>>> Domain is destroyed (qemuProcessStop) in a job due to patches [1] and [2].
>>>> However AFAIU it is not neccessary. If qemuProcessStop does not drop VM lock
>>>> then we don't need extra job to make qemuProcessStop and main job not
>>>> interleave. And we can drop the lock now only in qemuDomainObjBeginNestedJob in
>>>> qemuProcessStop which is introduced in [2]. AFAIU we can fix issues mentioned in
>>>> [2] the other way for example like it is done for qemu agent - we save agent
>>>> monitor reference on stack for entering/exiting agent monitor.
>>>>
>>>> So I wonder can we instead of this fix remove job for qemuProcessStop and run
>>>> destroying domain cuncurrently for non async jobs too.
>>>>
>>>> [1]
>>>> commit 8c9ff9960b29d4703a99efdd1cadcf6f48799cc0
>>>> Author: Jiri Denemark <jdenemar@redhat.com>
>>>> Date:   Thu Feb 11 15:32:48 2016 +0100
>>>>
>>>>     qemu: Process monitor EOF in a job
>>>>
>>>> [2]
>>>> commit 81f50cb92d16643bcd749e3ab5b404b8b7cec643
>>>> Author: Jiri Denemark <jdenemar@redhat.com>
>>>> Date:   Thu Feb 11 11:20:28 2016 +0100
>>>>
>>>>     qemu: Avoid calling qemuProcessStop without a job
>>>>
>>>>  src/qemu/qemu_domain.c  | 39 +++++++++++++++++++++++++++++++++++++++
>>>>  src/qemu/qemu_domain.h  |  4 ++++
>>>>  src/qemu/qemu_driver.c  |  2 +-
>>>>  src/qemu/qemu_hotplug.c |  4 ++--
>>>>  src/qemu/qemu_process.c |  9 +++++----
>>>>  5 files changed, 51 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/src/qemu/qemu_domain.c b/src/qemu/qemu_domain.c
>>>> index 939b2a3..aead72b 100644
>>>> --- a/src/qemu/qemu_domain.c
>>>> +++ b/src/qemu/qemu_domain.c
>>>> @@ -13534,3 +13534,42 @@ qemuDomainRunningReasonToResumeEvent(virDomainRunningReason reason)
>>>>  
>>>>      return VIR_DOMAIN_EVENT_RESUMED_UNPAUSED;
>>>>  }
>>>> +
>>>> +
>>>> +/**
>>>> + * Waits for domain condition to be triggered for a specific period of time.
>>>> + * if @until is 0 then waits indefinetely.
> 
> *indefinitely
> 
>>>> + *
>>>> + * Returns:
>>>> + *  -1 on error
>>>> + *   0 on success
>>>> + *   1 on timeout
>>>> + */
>>>> +int
>>>> +qemuDomainObjWait(virDomainObjPtr vm, unsigned long long until)
> 
> Each argument on it's own line.
> 
>>>> +{
>>>> +    qemuDomainObjPrivatePtr priv = vm->privateData;
>>>> +    int rc;
>>>> +
>>>> +    if (until)
>>>> +        rc = virCondWaitUntil(&vm->cond, &vm->parent.lock, until);
>>>> +    else
>>>> +        rc = virCondWait(&vm->cond, &vm->parent.lock);
>>>> +
>>>> +    if (rc < 0) {
>>>> +        if (until && errno == ETIMEDOUT)
>>>> +            return 1;
>>>> +
>>>> +        virReportSystemError(errno, "%s",
>>>> +                             _("failed to wait for domain condition"));
>>>> +        return -1;
>>>> +    }
>>>> +
>>>> +    if (priv->monEOF) {
>>>> +        virReportError(VIR_ERR_OPERATION_FAILED, "%s",
>>>> +                             _("domain is being stopped"));
> 
> "monitor has been closed"
> 

Well I thought user is unaware of concept of monitor. We do have commands
like qemu-monitor-command but they are targeted for devs/devops I guess.
Nevertheless original message a bit confusing too (being stopped???) :)

>>>> +        return -1;
>>>> +    }
> 
> No chance that vm->def->id could be set to -1 during any of this, right?
>  Perhaps it doesn't hurt to check virDomainObjIsActive too. Paranoia,
> you know...
> 

I'm afraid this can become dead code that we will resist to touch
as we don't understand why it exists) I see no possibilities why
we miss monEOF but hit virDomainObjIsActive now. Anyway before
this patch we would get deadlock anyway so it won't get any worse
if we don't add the check you suggest. If we really find situations
when it will be useful then we will do it immediately.

>>>> +
>>>> +    return 0;
>>>> +}
>>>> diff --git a/src/qemu/qemu_domain.h b/src/qemu/qemu_domain.h
>>>> index 2f8a1bf..36ab294 100644
>>>> --- a/src/qemu/qemu_domain.h
>>>> +++ b/src/qemu/qemu_domain.h
>>>> @@ -281,6 +281,7 @@ struct _qemuDomainObjPrivate {
>>>>      virDomainChrSourceDefPtr monConfig;
>>>>      bool monJSON;
>>>>      bool monError;
>>>> +    bool monEOF;
>>>>      unsigned long long monStart;
>>>>  
>>>>      qemuAgentPtr agent;
>>>> @@ -1085,4 +1086,7 @@ void qemuDomainStorageIdReset(qemuDomainObjPrivatePtr priv);
>>>>  virDomainEventResumedDetailType
>>>>  qemuDomainRunningReasonToResumeEvent(virDomainRunningReason reason);
>>>>  
>>>> +int
>>>> +qemuDomainObjWait(virDomainObjPtr vm, unsigned long long until);
>>>> +
>>>>  #endif /* __QEMU_DOMAIN_H__ */
> 
> All of the above can be it's own patch to "qemu: Introduce
> qemuDomainObjWait" with a commit message to describe why this would be
> used instead of the virDomainObjWait especially considering the QEMU
> events and the inability to get the exit criteria set by
> virDomainObjIsActive of vm->def->id != -1.
>>>>> diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
>>>> index b238309..f4250da 100644
>>>> --- a/src/qemu/qemu_driver.c
>>>> +++ b/src/qemu/qemu_driver.c
>>>> @@ -17142,7 +17142,7 @@ qemuDomainBlockJobAbort(virDomainPtr dom,
>>>>          qemuDomainDiskPrivatePtr diskPriv = QEMU_DOMAIN_DISK_PRIVATE(disk);
>>>>          qemuBlockJobUpdate(vm, QEMU_ASYNC_JOB_NONE, disk, NULL);
>>>>          while (diskPriv->blockjob) {
>>>> -            if (virDomainObjWait(vm) < 0) {
>>>> +            if (qemuDomainObjWait(vm, 0) < 0) {
>>>>                  ret = -1;
>>>>                  goto endjob;
>>>>              }
>>>> diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
>>>> index 4558a3c..8189629 100644
>>>> --- a/src/qemu/qemu_hotplug.c
>>>> +++ b/src/qemu/qemu_hotplug.c
>>>> @@ -165,7 +165,7 @@ qemuHotplugWaitForTrayEject(virDomainObjPtr vm,
>>>>          return -1;
>>>>  
>>>>      while (disk->tray_status != VIR_DOMAIN_DISK_TRAY_OPEN) {
>>>> -        if ((rc = virDomainObjWaitUntil(vm, now + CHANGE_MEDIA_TIMEOUT)) < 0)
>>>> +        if ((rc = qemuDomainObjWait(vm, now + CHANGE_MEDIA_TIMEOUT)) < 0)
>>>>              return -1;
>>>>  
>>>>          if (rc > 0) {
>>>> @@ -5002,7 +5002,7 @@ qemuDomainWaitForDeviceRemoval(virDomainObjPtr vm)
>>>>      until += qemuDomainRemoveDeviceWaitTime;
>>>>  
>>>>      while (priv->unplug.alias) {
>>>> -        if ((rc = virDomainObjWaitUntil(vm, until)) == 1)
>>>> +        if ((rc = qemuDomainObjWait(vm, until)) == 1)
>>>>              return 0;
>>>>  
>>>>          if (rc < 0) {
>>>> diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c
>>>> index 29b0ba1..dd03269 100644
>>>> --- a/src/qemu/qemu_process.c
>>>> +++ b/src/qemu/qemu_process.c
>>>> @@ -290,9 +290,12 @@ qemuProcessHandleMonitorEOF(qemuMonitorPtr mon,
>>>>  
>>>>      virObjectLock(vm);
>>>>  
>>>> +    priv = vm->privateData;
> 
> We could just initialize this at the top since we don't even check !vm
> first anyway.
> 
>>>> +    priv->monEOF = true;
>>>> +    virDomainObjBroadcast(vm);
> 
> So why Broadcast here? Why not just let processMonitorEOFEvent and the
> call to qemuProcessBeginStopJob handle this?

Because qemuProcessBeginStopJob can not begin job, it is occupied by waiter.

> 
> Also, let's place this after the following DEBUG message - nice to keep
> those up as high as possible.
> 
>>>> +
>>>>      VIR_DEBUG("Received EOF on %p '%s'", vm, vm->def->name);
>>>>  
>>>> -    priv = vm->privateData;
>>>>      if (priv->beingDestroyed) {
>>>>          VIR_DEBUG("Domain is being destroyed, EOF is expected");
>>>>          goto cleanup;
>>>> @@ -5996,6 +5999,7 @@ qemuProcessPrepareDomain(virQEMUDriverPtr driver,
>>>>  
>>>>      priv->monJSON = true;
>>>>      priv->monError = false;
>>>> +    priv->monEOF = false;
>>>>      priv->monStart = 0;
>>>>      priv->gotShutdown = false;
>>>>      priv->runningReason = VIR_DOMAIN_RUNNING_UNKNOWN;
>>>> @@ -6965,9 +6969,6 @@ qemuProcessBeginStopJob(virQEMUDriverPtr driver,
>>>>      if (qemuProcessKill(vm, killFlags) < 0)
>>>>          goto cleanup;
>>>>  
>>>> -    /* Wake up anything waiting on domain condition */
>>>> -    virDomainObjBroadcast(vm);
>>>> -
> 
> Since this is called by more than just processMonitorEOFEvent, I would
> think removing it could cause issues for qemuProcessAutoDestroy and
> qemuDomainDestroyFlags. What would cause them to have their possibly
> blocked blockjob or external memory dump to be notified of this event?
> 

Removing won't hurt. Autodestroy and destroy both first kill qemu and
then try to acquire job. If async job is active then they just aquire
destroy job immediately because async job let destroy run cuncurrently.
If non async job is active then we get EOF first which after this patch let
not async job finish and then destroy/autodestroy can proceed with stopping
the domain. The removed broadcast just makes spurious wakeup of waiter
I think.

I would also like to discuss the alternative approach described just before
patch diff stats. It looks more appopriate to me but we need a comment
from Jiri I guess to follow it.

Nikolay

> 
>>>>      if (qemuDomainObjBeginJob(driver, vm, job) < 0)
>>>>          goto cleanup;
>>>>  
>>>>

Nikolay Shirokovskiy

John Ferlan

Nikolay Shirokovskiy

John Ferlan

Nikolay Shirokovskiy

Nikolay Shirokovskiy

tags

participants (2)