at 11:18 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> I think that is incorrect. QEMU has guest notifier emulation for the
> non-KVM (and non-MSI-X PCI) cases. When there is no irqfd support
> available, QEMU sets up a regular eventfd and calls
> virtio_queue_guest_notifier_read() when it becomes readable.
Thanks Stefan. I finally understand why there is a `with_irqfd` parameter
for virtio_queue_set_guest_notifier_fd_handler.
But if `with_irqfd` is false, it seems OK to directly call virtio_irq(). Why
still bother using an eventfd? Is it for interrupt batching?
virtio_irq() is not thread safe so it cannot be called directly from the IOThread. Bouncing through the eventfd ensures that the virtio_irq() call happens in the QEMU main loop thread with the BQL held.
It may be cheaper to use a BH instead of an eventfd when irqfd is not available, but this is a slow path anyway. We might as well reuse the eventfd code that's already there.
Stefan