virtio: don't enable notifications during polling
commitd0435bc513e23a4961b6af20164d1c6c219eb4ea
authorStefan Hajnoczi <stefanha@redhat.com>
Mon, 9 Dec 2019 21:09:57 +0000 (9 21:09 +0000)
committerMichael S. Tsirkin <mst@redhat.com>
Sun, 5 Jan 2020 12:03:03 +0000 (5 07:03 -0500)
tree0434e1d07119ee8ea9e80143a5aa0a8c82cff41c
parent6620801f39f0cd02b30ab5ad8729c92c78ce9fd3
virtio: don't enable notifications during polling

Virtqueue notifications are not necessary during polling, so we disable
them.  This allows the guest driver to avoid MMIO vmexits.
Unfortunately the virtio-blk and virtio-scsi handler functions re-enable
notifications, defeating this optimization.

Fix virtio-blk and virtio-scsi emulation so they leave notifications
disabled.  The key thing to remember for correctness is that polling
always checks one last time after ending its loop, therefore it's safe
to lose the race when re-enabling notifications at the end of polling.

There is a measurable performance improvement of 5-10% with the null-co
block driver.  Real-life storage configurations will see a smaller
improvement because the MMIO vmexit overhead contributes less to
latency.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20191209210957.65087-1-stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
hw/block/virtio-blk.c
hw/scsi/virtio-scsi.c
hw/virtio/virtio.c
include/hw/virtio/virtio.h