On Tue, May 12, 2009 at 11:30:21AM -0300, Marcelo Tosatti wrote: > On Tue, May 12, 2009 at 10:13:36PM +0800, Yang, Sheng wrote: > > > > + mutex_unlock(&kvm->lock); > > > > > > assigned_dev list is protected by kvm->lock. So you could have another > > > ioctl adding to it at the same time you're searching. > > > > Oh, yes... My fault... > > > > > Could either have a separate kvm->assigned_devs_lock, to protect > > > kvm->arch.assigned_dev_head (users are ioctls that manipulate it), or > > > change the IRQ injection to use a separate spinlock, kill the workqueue > > > and call kvm_set_irq from the assigned device interrupt handler. > > > > Peferred the latter, though needs more work. But the only reason for put a > > workqueue here is because kvm->lock is a mutex? I can't believe... If so, I > > think we had made a big mistake - we have to fix all kinds of racy problem > > caused by this, but finally find it's unnecessary... > > One issue is that kvm_set_irq can take too long while interrupts are > blocked, and you'd have to disable interrupts in other contexes that > inject interrupts (say qemu->ioctl(SET_INTERRUPT)->...->), so all i can > see is a tradeoff. > > <guess mode on> > > But the interrupt injection path seems to be pretty short and efficient > to happen in host interrupt context. > > <guess mode off> > > Avi, Gleb? > > > Maybe another reason is kvm_kick_vcpu(), but have already fixed by you. > > Note you tested the spinlock_irq patch with GigE and there was no > significant performance regression right? > > > > > Continue to check the code... OK, it might take some time for bigger changes to happen. I've changed your patch to drop the lock only around cancel_work_sync. Can deadlock if someone else tries to mess with the assigned device at the same time, but the VM won't go away under it because of the vmfd reference. diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c index 4d00942..ba067db 100644 --- a/virt/kvm/kvm_main.c +++ b/virt/kvm/kvm_main.c @@ -250,7 +250,15 @@ static void deassign_host_irq(struct kvm *kvm, disable_irq_nosync(assigned_dev-> host_msix_entries[i].vector); + /* + * FIXME: kvm_assigned_dev_interrupt_work_handler can deadlock + * with cancel_work_sync, since it requires kvm->lock for irq + * injection. This is a hack, the irq code must use + * a separate lock. + */ + mutex_unlock(&kvm->lock); cancel_work_sync(&assigned_dev->interrupt_work); + mutex_lock(&kvm->lock); for (i = 0; i < assigned_dev->entries_nr; i++) free_irq(assigned_dev->host_msix_entries[i].vector, -- 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
