]> git.proxmox.com Git - mirror_qemu.git/commit
virtio: use defer_call() in virtio_irqfd_notify()
authorStefan Hajnoczi <stefanha@redhat.com>
Wed, 13 Sep 2023 20:00:44 +0000 (16:00 -0400)
committerKevin Wolf <kwolf@redhat.com>
Tue, 31 Oct 2023 14:42:14 +0000 (15:42 +0100)
commit84d61e5f36a73ed24742b7c7cf7b811e456dd024
tree420c313770c60a0f9aaa2487b296c6883535b4b0
parent433fcea40c31ff355f84da22a46977c2a1b596c3
virtio: use defer_call() in virtio_irqfd_notify()

virtio-blk and virtio-scsi invoke virtio_irqfd_notify() to send Used
Buffer Notifications from an IOThread. This involves an eventfd
write(2) syscall. Calling this repeatedly when completing multiple I/O
requests in a row is wasteful.

Use the defer_call() API to batch together virtio_irqfd_notify() calls
made during thread pool (aio=threads), Linux AIO (aio=native), and
io_uring (aio=io_uring) completion processing.

Behavior is unchanged for emulated devices that do not use
defer_call_begin()/defer_call_end() since defer_call() immediately
invokes the callback when called outside a
defer_call_begin()/defer_call_end() region.

fio rw=randread bs=4k iodepth=64 numjobs=8 IOPS increases by ~9% with a
single IOThread and 8 vCPUs. iodepth=1 decreases by ~1% but this could
be noise. Detailed performance data and configuration specifics are
available here:
https://gitlab.com/stefanha/virt-playbooks/-/tree/blk_io_plug-irqfd

This duplicates the BH that virtio-blk uses for batching. The next
commit will remove it.

Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-ID: <20230913200045.1024233-4-stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
block/io_uring.c
block/linux-aio.c
hw/virtio/trace-events
hw/virtio/virtio.c
util/thread-pool.c