Re: [Qemu-devel] [PATCH] vhost: fix a migration failed because of vhost region merge

[Dr. David Alan Gilbert]

* Michael S. Tsirkin (address@hidden) wrote:
> On Thu, Jul 20, 2017 at 06:22:15PM +0100, Dr. David Alan Gilbert wrote:
> > * Michael S. Tsirkin (address@hidden) wrote:
> > > On Wed, Jul 19, 2017 at 03:24:27PM +0200, Igor Mammedov wrote:
> > > > On Wed, 19 Jul 2017 12:46:13 +0100
> > > > "Dr. David Alan Gilbert" <address@hidden> wrote:
> > > > 
> > > > > * Igor Mammedov (address@hidden) wrote:
> > > > > > On Wed, 19 Jul 2017 23:17:32 +0800
> > > > > > Peng Hao <address@hidden> wrote:
> > > > > >   
> > > > > > > When a guest that has several hotplugged dimms is migrated, in
> > > > > > > destination host it will fail to resume. Because vhost regions of
> > > > > > > several dimms in source host are merged and in the restore stage
> > > > > > > in destination host it computes whether more than vhost slot limit
> > > > > > > before merging vhost regions of several dimms.  
> > > > > > could you provide a bit more detailed description of the problem
> > > > > > including command line+used device_add commands on source and
> > > > > > command line on destination?  
> > > > > 
> > > > > (ccing in Marc Andre and Maxime)
> > > > > 
> > > > > Hmm, I'd like to understade the situation where you get merging 
> > > > > between
> > > > > RAMBlocks; that complicates some stuff for postcopy.
> > > > and probably inconsistent merging breaks vhost as well
> > > > 
> > > > merging might happen if regions are adjacent or overlap
> > > > but for that to happen merged regions must have equal
> > > > distance between their GPA:HVA pairs, so that following
> > > > translation would work:
> > > > 
> > > > if gva in regionX[gva_start, len, hva_start]
> > > >    hva = hva_start + gva - gva_start
> > > > 
> > > > while GVA of regions is under QEMU control and deterministic
> > > > HVA is not, so in migration case merging might happen on source
> > > > side but not on destination, resulting in different memory maps.
> > > > 
> > > > Maybe Michael might know details why migration works in vhost usecase,
> > > > but I don't see vhost sending any vmstate data.
> > > 
> > > We aren't merging ramblocks at all.
> > > When we are passing blocks A and B to vhost, if we see that
> > > 
> > > hvaB=hvaA + lenA
> > > gpaB=gpaA + lenA
> > > 
> > > then we can improve performance a bit by passing a single
> > > chunk to vhost: hvaA,gpaA,lena+lenB
> > 
> > OK, but that means that a region can incorporate multiple
> > RAMBlocks though? Hmm that's not fun on postcopy.
> > 
> > > so it does not affect migration normally.
> > 
> > Well, why? What's required - if the region sizes/lengths/orders
> > are different on the source and destination does it matter - if
> > it does then that means we have a problem, since that heuristic
> > is non-deterministic.
> > 
> > Dave
> 
> It doesn't matter normally. But there's a limit
> on number of regions that vhost can support.
> pc_dimm_memory_plug tries to guess this number early
> because it wants to fail hotplug requests gracefully.

Why doesn't it matter - can't this heuristic potentially trigger on a
source but not a destination even in the case of something like NUMA
without hotplug?

> See
> commit 3fad87881e55aaff659408dcf25fa204f89a7896
> Author: Igor Mammedov <address@hidden>
> Date:   Tue Oct 6 10:37:28 2015 +0200
> 
>     pc-dimm: add vhost slots limit check before commiting to hotplug
> 
> It might make sense to limit this to hotplug, though
> runstate check seems like a wrong way to do this -
> VM could be stopped e.g. through QMP at the time.

My other problem is this optimisation makes it tricky for postcopy;
one-region -> one-ramblock is easy, but there's a potential for problems
if different RAMBlock's within the region have different
characteristics.

Dave

> > > 
> > > > 
> > > > > 
> > > > > > > 
> > > > > > > Signed-off-by: Peng Hao <address@hidden>
> > > > > > > Signed-off-by: Wang Yechao <address@hidden>
> > > > > > > ---
> > > > > > >  hw/mem/pc-dimm.c | 2 +-
> > > > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > > > 
> > > > > > > diff --git a/hw/mem/pc-dimm.c b/hw/mem/pc-dimm.c
> > > > > > > index ea67b46..bb0fa08 100644
> > > > > > > --- a/hw/mem/pc-dimm.c
> > > > > > > +++ b/hw/mem/pc-dimm.c
> > > > > > > @@ -101,7 +101,7 @@ void pc_dimm_memory_plug(DeviceState *dev, 
> > > > > > > MemoryHotplugState *hpms,
> > > > > > >          goto out;
> > > > > > >      }
> > > > > > >  
> > > > > > > -    if (!vhost_has_free_slot()) {
> > > > > > > +    if (!vhost_has_free_slot() && runstate_is_running()) {
> > > > > > >          error_setg(&local_err, "a used vhost backend has no free"
> > > > > > >                                 " memory slots left");
> > > > > > >          goto out;  
> > > > > 
> > > > > Even this produces the wrong error message in this case,
> > > > > it also makes me think if the existing code should undo a lot of
> > > > > the object_property_set's that happen.
> > > > > 
> > > > > Dave
> > > > > > 
> > > > > >   
> > > > > --
> > > > > Dr. David Alan Gilbert / address@hidden / Manchester, UK
> > --
> > Dr. David Alan Gilbert / address@hidden / Manchester, UK
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK