Re: [PATCH RFC 2/2] kvm: set affinity hint for assigned device msi

[Marcelo Tosatti <mtosatti@redhat.com>]

On Mon, Oct 17, 2011 at 03:32:15PM +0200, Michael S. Tsirkin wrote:
> On Mon, Oct 17, 2011 at 08:58:59AM -0200, Marcelo Tosatti wrote:
> > On Sun, Oct 16, 2011 at 03:12:23PM +0200, Michael S. Tsirkin wrote:
> > > On Thu, Oct 13, 2011 at 11:54:50AM -0300, Marcelo Tosatti wrote:
> > > > On Tue, Oct 11, 2011 at 08:38:28PM +0200, Michael S. Tsirkin wrote:
> > > > > To forward an interrupt to a vcpu that runs on
> > > > > a host cpu different from the current one,
> > > > > we need an ipi which likely will cost us as much
> > > > > as delivering the interrupt directly to that cpu would.
> > > > > 
> > > > > Set irq affinity hint to point there, irq balancer
> > > > > can then take this into accound and balance
> > > > > interrupts accordingly.
> > > > > 
> > > > > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > > > ---
> > > > >  virt/kvm/assigned-dev.c |    8 +++++---
> > > > >  virt/kvm/irq_comm.c     |   17 ++++++++++++++++-
> > > > >  2 files changed, 21 insertions(+), 4 deletions(-)
> > > > > 
> > > > > diff --git a/virt/kvm/assigned-dev.c b/virt/kvm/assigned-dev.c
> > > > > index f89f138..b579777 100644
> > > > > --- a/virt/kvm/assigned-dev.c
> > > > > +++ b/virt/kvm/assigned-dev.c
> > > > > @@ -142,9 +142,11 @@ static void deassign_host_irq(struct kvm *kvm,
> > > > >  		for (i = 0; i < assigned_dev->entries_nr; i++)
> > > > >  			disable_irq(assigned_dev->host_msix_entries[i].vector);
> > > > >  
> > > > > -		for (i = 0; i < assigned_dev->entries_nr; i++)
> > > > > -			free_irq(assigned_dev->host_msix_entries[i].vector,
> > > > > -				 (void *)assigned_dev);
> > > > > +		for (i = 0; i < assigned_dev->entries_nr; i++) {
> > > > > +			u32 vector = assigned_dev->host_msix_entries[i].vector;
> > > > > +			irq_set_affinity_hint(vector, NULL);
> > > > > +			free_irq(vector, (void *)assigned_dev);
> > > > > +		}
> > > > >  
> > > > >  		assigned_dev->entries_nr = 0;
> > > > >  		kfree(assigned_dev->host_msix_entries);
> > > > > diff --git a/virt/kvm/irq_comm.c b/virt/kvm/irq_comm.c
> > > > > index ac8b629..68b1f7c 100644
> > > > > --- a/virt/kvm/irq_comm.c
> > > > > +++ b/virt/kvm/irq_comm.c
> > > > > @@ -22,6 +22,7 @@
> > > > >  
> > > > >  #include <linux/kvm_host.h>
> > > > >  #include <linux/slab.h>
> > > > > +#include <linux/interrupt.h>
> > > > >  #include <trace/events/kvm.h>
> > > > >  
> > > > >  #include <asm/msidef.h>
> > > > > @@ -80,6 +81,17 @@ inline static bool kvm_is_dm_lowest_prio(struct kvm_lapic_irq *irq)
> > > > >  #endif
> > > > >  }
> > > > >  
> > > > > +static void kvm_vcpu_host_irq_hint(struct kvm_vcpu *vcpu, int host_irq)
> > > > > +{
> > > > > +	const struct cpumask *mask;
> > > > > +	/* raw_smp_processor_id() is ok here: if we get preempted we can get a
> > > > > +	 * wrong value but we don't mind much. */
> > > > > +	if (host_irq >= 0 && unlikely(vcpu->cpu != raw_smp_processor_id())) {
> > > > > +		mask = get_cpu_mask(vcpu->cpu);
> > > > > +		irq_set_affinity_hint(host_irq, mask);
> > > > > +	}
> > > > > +}
> > > > 
> > > > Unsure about the internals of irq_set_affinity_hint, but AFAICS its
> > > > exported so that irqbalance in userspace can make a decision.
> > > 
> > > Yes. Pls note at the moment there's no hint so irqbalance
> > > will likely try to move the irq away from vcpu if that
> > > is doing a lot of work. My patch tries to correct that.
> > > 
> > > > If that is the case, then irqbalance update rate should be high enough
> > > > to catch up with a vcpu migrating betweens cpus (which initially does
> > > > not appear a sensible arrangement).
> > > 
> > > At least for pinned vcpus, that's almost sure to be the case :)
> > 
> > What i mean is that the frequency of a vcpu migrating between cpus
> > might be higher than what irqbalance can cope with.
> > 
> > > > The decision to have the host interrupt follow the vcpu seems a good
> > > > one, given that it saves an IPI and is potentially more cache friendly
> > > > overall.
> > > 
> > > > And AFAICS its more intelligent for the device assignment case than
> > > > anything irqbalance can come up with
> > > 
> > > Do you just propose overwriting affinity set by userspace then?
> > 
> > Yes.
> > 
> > > My concern would be to avoid breaking setups some users have,
> > > with carefully manually optimized affinity for vcpus and device irqs.
> > 
> > They can disable automatic in-kernel affinity.
> 
> This still means code needs to be changed ...
> Anyway, what's the interface for that?
> 
> > > 
> > > > (note it depends on how the APIC is
> > > > configured, your patch ignores that).
> > > 
> > > Could you clarify please? What is meant by 'it' in 'it depends'?
> > 
> > "It" means the target vcpu selection. It depends on how the guest
> > APIC is programmed.
> > 
> > > Which APIC - host or guest - do you mean, and what are possible APIC
> > > configurations to consider?
> > 
> > Guest APIC. Guest APIC programmed with round robin would break the
> > static assignment on your patch.
> 
> For round robin we might just want to disable this
> automatic affinity?

OK.

> > Configurations to consider, all common ones used for assigned devices?
> 
> I mean, besides round robin, any other modes that
> have an issue? Interrupts can also be multicast,
> I think, but we probably don't care what happens
> to affinity then, as msi interrupts are probably never
> broadcast ...

There is also lowest priority, which can be used with MSI.