Re: [Xen-devel] Enabling vm_event for a guest with more VCPUs than available ring buffer slots freezes the virtual machine

[Mihai Donțu <mihai.dontu@xxxxxxxxx>]

On Tue, 7 Feb 2017 22:41:57 +0200 Razvan Cojocaru wrote:
> On 02/07/2017 10:20 PM, Tamas K Lengyel wrote:
> > On Tue, Feb 7, 2017 at 11:57 AM, Razvan Cojocaru wrote:
> >     On 02/07/2017 08:39 PM, Andrew Cooper wrote:  
> >     > On 07/02/17 18:31, Razvan Cojocaru wrote:  
> >     >> On 02/07/2017 08:15 PM, Tamas K Lengyel wrote:  
> >     >>> On Tue, Feb 7, 2017 at 9:53 AM, Razvan Cojocaru wrote:  
> >     >>>     Hello,
> >     >>>
> >     >>>     Setting, e.g. 16 VCPUs for a HVM guest, ends up blocking the  
> >     guest  
> >     >>>     completely when subscribing to vm_events, apparently because  
> >     of this  
> >     >>>     code in xen/common/vm_event.c:
> >     >>>
> >     >>>     315     /* Give this vCPU a black eye if necessary, on the  
> >     way out.  
> >     >>>     316      * See the comments above wake_blocked() for more  
> >     information  
> >     >>>     317      * on how this mechanism works to avoid waiting. */
> >     >>>     318     avail_req = vm_event_ring_available(ved);
> >     >>>     319     if( current->domain == d && avail_req < d->max_vcpus )
> >     >>>     320         vm_event_mark_and_pause(current, ved);
> >     >>>
> >     >>>     It would appear that even if the guest only has 2 online  
> >     VCPUs, the  
> >     >>>     "avail_req < d->max_vcpus" condition will pause current, and we
> >     >>>     eventually end up with all the VCPUs paused.
> >     >>>
> >     >>>     An ugly hack ("avail_req < 2") has allowed booting a guest  
> >     with many  
> >     >>>     VCPUs (max_vcpus, the guest only brings 2 VCPUs online),  
> >     however that's  
> >     >>>     just to prove that that was the culprit - a real solution to  
> >     this needs  
> >     >>>     more in-depth understading of the issue and potential  
> >     solution. That's  
> >     >>>     basically very old code (pre-2012 at least) that got moved  
> >     around into  
> >     >>>     the current shape of Xen today - please CC anyone relevant  
> >     to the  
> >     >>>     discussion that you're aware of.
> >     >>>
> >     >>>     Thoughts?
> >     >>>
> >     >>>
> >     >>> I think is a side-effect of the growth of the vm_event structure  
> >     and the  
> >     >>> fact that we have a single page ring. The check effectively sets a
> >     >>> threshold of having enough space for each vCPU to place at least one
> >     >>> more event on the ring, and if that's not the case it gets  
> >     paused. OTOH  
> >     >>> I think this would only have an effect on asynchronous events,  
> >     for all  
> >     >>> other events the vCPU is already paused. Is that the case you have? 
> >  
> >     >> No, on the contrary, all my events are synchronous (the VCPU is  
> >     paused  
> >     >> waiting for the vm_event reply).
> >     >>
> >     >> I've debugged this a bit, and the problem seems to be that
> >     >> vm_event_wake_blocked() breaks here:
> >     >>
> >     >> 150     /* We remember which vcpu last woke up to avoid scanning  
> >     always  
> >     >> linearly
> >     >> 151      * from zero and starving higher-numbered vcpus under  
> >     high load */  
> >     >> 152     if ( d->vcpu )
> >     >> 153     {
> >     >> 154         int i, j, k;
> >     >> 155
> >     >> 156         for (i = ved->last_vcpu_wake_up + 1, j = 0; j <
> >     >> d->max_vcpus; i++, j++)
> >     >> 157         {
> >     >> 158             k = i % d->max_vcpus;
> >     >> 159             v = d->vcpu[k];
> >     >> 160             if ( !v )
> >     >> 161                 continue;
> >     >> 162
> >     >> 163             if ( !(ved->blocked) || online >= avail_req )
> >     >> 164                break;
> >     >> 165
> >     >> 166             if ( test_and_clear_bit(ved->pause_flag,  
> >     &v->pause_flags) )  
> >     >> 167             {
> >     >> 168                 vcpu_unpause(v);
> >     >> 169                 online++;
> >     >> 170                 ved->blocked--;
> >     >> 171                 ved->last_vcpu_wake_up = k;
> >     >> 172             }
> >     >> 173         }
> >     >> 174     }
> >     >>
> >     >> at "if ( !(ved->blocked) || online >= avail_req )". At this point,
> >     >> nothing ever gets unblocked. It's hard to believe that this is  
> >     desired  
> >     >> behaviour, as I don't know what could possibly happen for that  
> >     condition  
> >     >> to become false once all the online VCPUs are stuck (especially  
> >     when the  
> >     >> guest has just started booting).  
> > 
> > 
> > Ah I see what happens. During boot vCPU 0 generates an event and gets
> > marked blocked because the number of vCPUs is so high. The other vCPUs
> > are still unblocked since they are idle, but this test here will still
> > be true (online >= avail_req) and thus we can never unblock vCPU0. And
> > then the boot process is hanging because vCPU0 never resumes. I would
> > argue that this test should be changed to check that there is at least 1
> > spot on the ring and only break if that is not the case anymore (ie.
> > instead of incrementing online we should be decrementing avail_req).  
> 
> That is exactly what happens. And it can't really be fixed just by
> increasing the ring buffer (although that definitely helps a lot and
> would be a smart move): no matter how large it is, we can always ask the
> domain to use more VCPUs than there are slots in the buffer.
> 
> >     >
> >     > I wouldn't bet that this logic has ever been tested.  If you  
> >     recall, the  
> >     > addition of register state into the vm_event ring made each entry far
> >     > larger, which in turns makes it more likely to hit this condition.
> >     >
> >     > However, simply fixing the logic to re-online the cpus isn't a good
> >     > solution either, as having $N vcpus paused at any one time because of
> >     > ring contention is not conducive good system performance.
> >     >
> >     > Realistically, the ring size needs to be max_cpus * sizeof(largest
> >     > vm_event) at an absolute minimum, and I guess this is now beyond 1  
> >     page?
> > 
> >     Yes, of course the reason this triggers earlier now is the growth of the
> >     request's size. Yes, using e.g. 20 VCPUs in the guest's setup will
> >     exceed a page's number of slots.
> > 
> >     And yes, ideally we should have multi-page ring buffers - however that
> >     is a long-term project that requires design changes in other parts of
> >     Xen as well (Andrew, CCd here, was recently talking about one).
> > 
> >     However, even with a one-page ring buffer, surely it's not good to end
> >     up in this situation, especially for guests such as mine, which never
> >     actually bring more than 2 VCPUs online. But even if they were to use
> >     more, blocking the guest on vm_event init is completely pointless - we
> >     might as well return some kind of error if max_vcpus > available slots.
> > 
> >     I don't follow the system performance argument. Surely completely
> >     blocking the guest is worse.
> > 
> > 
> > I also don't see the point in marking a vCPU blocked if it is already
> > paused. I think this behavior of blocking vCPUs makes only sense for
> > asynchronous events. Razvan, could you test what happens if
> > vm_event_mark_and_pause is only called if the vCPU is unpaused?  
> 
> It works for me with this change (using Xen 4.7 sources here):
> 
> @@ -318,7 +329,11 @@ void vm_event_put_request(struct domain *d,
>       * on how this mechanism works to avoid waiting. */
>      avail_req = vm_event_ring_available(ved);
>      if( current->domain == d && avail_req < d->max_vcpus )
> -        vm_event_mark_and_pause(current, ved);
> +    {
> +        if ( !atomic_read( &current->vm_event_pause_count ) )
> +            vm_event_mark_and_pause(current, ved);
> +    }

If I'm reading the code correctly, when max_vcpus is greater than the
number of slots available in the ring, a race appears that can lead to
a ring corruption (in debug mode ASSERT(free_req > 0) will trigger).

For example, when a single slot is available, two vCPUs can race to
vm_event_put_request() after both being given a green light in
__vm_event_claim_slot(), whose return depends only on
vm_event_ring_available() returning non-zero (which it can do, for both
vCPUs at the same time).

As it turns out, the bug being talked about prevented this from showing
up.

PS: 
https://xenbits.xen.org/gitweb/?p=xen.git;a=commit;h=3643a961195f76ba849a213628c1979240e6fbdd

-- 
Mihai Donțu

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel