Re: [PATCH v2] KVM: Extend MAX_IRQ_ROUTES to 4096 for all archs

[Cornelia Huck <cohuck@redhat.com>]

On Fri, 20 Apr 2018 21:51:13 +0800
Wanpeng Li <kernellwp@gmail.com> wrote:

> 2018-04-20 15:15 GMT+08:00 Cornelia Huck <cohuck@redhat.com>:
> > On Thu, 19 Apr 2018 17:47:28 -0700
> > Wanpeng Li <kernellwp@gmail.com> wrote:
> >  
> >> From: Wanpeng Li <wanpengli@tencent.com>
> >>
> >> Our virtual machines make use of device assignment by configuring
> >> 12 NVMe disks for high I/O performance. Each NVMe device has 129
> >> MSI-X Table entries:
> >> Capabilities: [50] MSI-X: Enable+ Count=129 Masked-Vector table: BAR=0 offset=00002000
> >> The windows virtual machines fail to boot since they will map the number of
> >> MSI-table entries that the NVMe hardware reported to the bus to msi routing
> >> table, this will exceed the 1024. This patch extends MAX_IRQ_ROUTES to 4096
> >> for all archs, in the future this might be extended again if needed.
> >>
> >> Cc: Paolo Bonzini <pbonzini@redhat.com>
> >> Cc: Radim Krčmář <rkrcmar@redhat.com>
> >> Cc: Tonny Lu <tonnylu@tencent.com>
> >> Cc: Cornelia Huck <cohuck@redhat.com>
> >> Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> >> Signed-off-by: Tonny Lu <tonnylu@tencent.com>
> >> ---
> >> v1 -> v2:
> >>  * extend MAX_IRQ_ROUTES to 4096 for all archs
> >>
> >>  include/linux/kvm_host.h | 6 ------
> >>  1 file changed, 6 deletions(-)
> >>
> >> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> >> index 6930c63..0a5c299 100644
> >> --- a/include/linux/kvm_host.h
> >> +++ b/include/linux/kvm_host.h
> >> @@ -1045,13 +1045,7 @@ static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq)
> >>
> >>  #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
> >>
> >> -#ifdef CONFIG_S390
> >>  #define KVM_MAX_IRQ_ROUTES 4096 //FIXME: we can have more than that...  
> >
> > What about /* might need extension/rework in the future */ instead of
> > the FIXME?  
> 
> Yeah, I guess the maintainers can help to fix it when applying. :)
> 
> >
> > As far as I understand, 4096 should cover most architectures and the
> > sane end of s390 configurations, but will not be enough at the scarier
> > end of s390. (I'm not sure how much it matters in practice.)
> >
> > Do we want to make this a tuneable in the future? Do some kind of
> > dynamic allocation? Not sure whether it is worth the trouble.  
> 
> I think keep as it is currently.

My main question here is how long this is enough... the number of
virtqueues per device is up to 1K from the initial 64, which makes it
possible to hit the 4K limit with fewer virtio devices than before (on
s390, each virtqueue uses a routing table entry). OTOH, we don't want
giant tables everywhere just to accommodate s390.

If the s390 maintainers tell me that nobody is doing the really insane
stuff, I'm happy as well :)