Re: [Qemu-devel] [PATCH v4 4/7] pc: fix QEMU crashing when more than ~50 memory hotplugged

[Igor Mammedov <imammedo@redhat.com>]

On Mon, 13 Jul 2015 09:55:18 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Fri, Jul 10, 2015 at 12:12:36PM +0200, Igor Mammedov wrote:
> > On Thu, 9 Jul 2015 16:46:43 +0300
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > 
> > > On Thu, Jul 09, 2015 at 03:43:01PM +0200, Paolo Bonzini wrote:
> > > > 
> > > > 
> > > > On 09/07/2015 15:06, Michael S. Tsirkin wrote:
> > > > > > QEMU asserts in vhost due to hitting vhost backend limit
> > > > > > on number of supported memory regions.
> > > > > > 
> > > > > > Describe all hotplugged memory as one continuos range
> > > > > > to vhost with linear 1:1 HVA->GPA mapping in backend.
> > > > > > 
> > > > > > Signed-off-by: Igor Mammedov <imammedo@redhat.com>
> > > > >
> > > > > Hmm - a bunch of work here to recombine MRs that memory
> > > > > listener interface breaks up.  In particular KVM could
> > > > > benefit from this too (on workloads that change the table a
> > > > > lot).  Can't we teach memory core to pass hva range as a
> > > > > single continuous range to memory listeners?
> > > > 
> > > > Memory listeners are based on memory regions, not HVA ranges.
> > > > 
> > > > Paolo
> > > 
> > > Many listeners care about HVA ranges. I know KVM and vhost do.
> > I'm not sure about KVM, it works just fine with fragmented memory
> > regions, the same will apply to vhost once module parameter to
> > increase limit is merged.
> > 
> > but changing generic memory listener interface to replace HVA mapped
> > regions with HVA container would lead to a case when listeners
> > won't see exact layout that they might need.
> 
> I don't think they care, really.
> 
> > In addition vhost itself will suffer from working with big HVA
> > since it allocates log depending on size of memory => bigger log.
> 
> Not really - it allocates the log depending on the PA range.
> Leaving unused holes doesn't reduce it's size.
if it would use HVA container instead then it will always allocate
log for max possible GPA, meaning that -m 1024,maxmem=1T will waste
a lot of memory and more so for bigger maxmem.
It's still possible to induce worst case by plugging pc-dimm at the end
of hotplug-memory area by specifying address for it explicitly.
That problem exists since memory hot-add was introduced, I've just
haven't noticed it back then.

It's perfectly fine to allocate log by last GPA as far as
memory is nearly continuous but memory hot-add makes it possible to
have sparse layout with a huge gaps between guest mapped RAM
which makes current log handling inefficient.

I wonder how hard it would be to make log_size depend on present RAM
size rather than max present GPA so it wouldn't allocate excess 
memory for log.


> 
> 
> > That's one of the reasons that in this patch HVA ranges in
> > memory map are compacted only for backend consumption,
> > QEMU's side of vhost uses exact map for internal purposes.
> > And the other reason is I don't know vhost enough to rewrite it
> > to use big HVA for everything.
> > 
> > > I guess we could create dummy MRs to fill in the holes left by
> > > memory hotplug?
> > it looks like nice thing from vhost pov but complicates other side,
> 
> What other side do you have in mind?
> 
> > hence I dislike an idea inventing dummy MRs for vhost's convenience.
memory core, but lets see what Paolo thinks about it.

> > 
> > 
> > > vhost already has logic to recombine
> > > consequitive chunks created by memory core.
> > which looks a bit complicated and I was thinking about simplifying
> > it some time in the future.
>