[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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





On Tue, Feb 7, 2017 at 9:53 AM, Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx> 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?

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

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.