On Mon, 2012-01-30 at 19:24 +0100, Jan Kiszka wrote:
> On 2012-01-30 19:06, Alex Williamson wrote:
> > We need to prioritize our matching when setting MSI-X vector
> > entries. Unused entries should only be used if we don't find
> > an exact match or else we risk duplicating entries. This was
> > causing an ENOSPC return when trying to mask and unmask MSI-X
> > vectors based on guest MSI-X table updates.
> >
> > Signed-off-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
> >
> > ---
> >
> > The faulting sequence went something like:
> >
> > Start:
> > [0] entry 0, vector A
> > [1] entry 1, vector B
> > [2] entry 2, vector C
> >
> > Set entry 1 to 0:
> > [0] entry 0, vector A
> > [1] entry 1->1, vector B->0
> > [2] entry 2, vector C
> >
> > Set entry 2 to 0:
> > [0] entry 0, vector A
> > [1] entry 1->2, vector 0->0 <- incorrectly matched
> > [2] entry 2, vector C
> >
> > Set entry 2 to C:
> > [0] entry 0, vector A
> > [1] entry 2->2, vector 0->C <- incorrectly matched again
> > [2] entry 2, vector C
> >
> > Set entry 1 to B:
> > [0] entry 0, vector A
> > [1] entry 2, vector C
> > [2] entry 2, vector C
> > -ENOSPC
> >
> > Userspace will need to test for this bug (or do we want to set
> > a feature flag?).
>
> The latter is much simpler. We did it before.
Yeah, the test is little cumbersome, I'll send a v2 with a feature flag.
> >
> > virt/kvm/assigned-dev.c | 21 +++++++++++++++------
> > 1 files changed, 15 insertions(+), 6 deletions(-)
> >
> > diff --git a/virt/kvm/assigned-dev.c b/virt/kvm/assigned-dev.c
> > index 758e3b3..798a090 100644
> > --- a/virt/kvm/assigned-dev.c
> > +++ b/virt/kvm/assigned-dev.c
> > @@ -728,7 +728,7 @@ msix_nr_out:
> > static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
> > struct kvm_assigned_msix_entry *entry)
> > {
> > - int r = 0, i;
> > + int r = 0, i, unused = -1;
> > struct kvm_assigned_dev_kernel *adev;
> >
> > mutex_lock(&kvm->lock);
> > @@ -741,17 +741,26 @@ static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
> > goto msix_entry_out;
> > }
> >
> > - for (i = 0; i < adev->entries_nr; i++)
> > - if (adev->guest_msix_entries[i].vector == 0 ||
> > - adev->guest_msix_entries[i].entry == entry->entry) {
> > + for (i = 0; i < adev->entries_nr; i++) {
> > + if (unused < 0 && !adev->guest_msix_entries[i].vector)
> > + unused = i;
> > +
> > + if (adev->guest_msix_entries[i].entry == entry->entry) {
> > adev->guest_msix_entries[i].entry = entry->entry;
> > adev->guest_msix_entries[i].vector = entry->gsi;
> > adev->host_msix_entries[i].entry = entry->entry;
> > break;
> > }
> > + }
> > +
> > if (i == adev->entries_nr) {
> > - r = -ENOSPC;
> > - goto msix_entry_out;
> > + if (unused < 0) {
> > + r = -ENOSPC;
> > + goto msix_entry_out;
> > + }
> > + adev->guest_msix_entries[unused].entry = entry->entry;
> > + adev->guest_msix_entries[unused].vector = entry->gsi;
> > + adev->host_msix_entries[unused].entry = entry->entry;
> > }
> >
> > msix_entry_out:
> >
>
> Looks good. Does it fix the masking problems?
It appears to, a 4-way rhel6 guest w/ assigned igb happily shifts it's
smp affinity around with irqbalance while running netperf in a loop.
Thanks,
Alex
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
