From: Stefan Hajnoczi <stefanha@redhat.com>
Date: Tue, 17 May 2022 09:27:45 +0100
Subject: [PATCH] virtio-scsi: don't waste CPU polling the event virtqueue
Content-type: text/plain

The virtio-scsi event virtqueue is not emptied by its handler function.
This is typical for rx virtqueues where the device uses buffers when
some event occurs (e.g. a packet is received, an error condition
happens, etc).

Polling non-empty virtqueues wastes CPU cycles. We are not waiting for
new buffers to become available, we are waiting for an event to occur,
so it's a misuse of CPU resources to poll for buffers.

Introduce the new virtio_queue_aio_attach_host_notifier_no_poll() API,
which is identical to virtio_queue_aio_attach_host_notifier() except
that it does not poll the virtqueue.

Before this patch the following command-line consumed 100% CPU in the
IOThread polling and calling virtio_scsi_handle_event():

  $ qemu-system-x86_64 -M accel=kvm -m 1G -cpu host \
      --object iothread,id=iothread0 \
      --device virtio-scsi-pci,iothread=iothread0 \
      --blockdev file,filename=test.img,aio=native,cache.direct=on,node-name=drive0 \
      --device scsi-hd,drive=drive0

After this patch CPU is no longer wasted.

Reported-by: Nir Soffer <nsoffer@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Tested-by: Nir Soffer <nsoffer@redhat.com>
Message-id: 20220427143541.119567-3-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
(cherry picked from commit 38738f7dbbda90fbc161757b7f4be35b52205552)
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 hw/scsi/virtio-scsi-dataplane.c |  2 +-
 hw/virtio/virtio.c              | 13 +++++++++++++
 include/hw/virtio/virtio.h      |  1 +
 3 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/hw/scsi/virtio-scsi-dataplane.c b/hw/scsi/virtio-scsi-dataplane.c
index 29575cbaf6..8bb6e6acfc 100644
--- a/hw/scsi/virtio-scsi-dataplane.c
+++ b/hw/scsi/virtio-scsi-dataplane.c
@@ -138,7 +138,7 @@ int virtio_scsi_dataplane_start(VirtIODevice *vdev)
 
     aio_context_acquire(s->ctx);
     virtio_queue_aio_attach_host_notifier(vs->ctrl_vq, s->ctx);
-    virtio_queue_aio_attach_host_notifier(vs->event_vq, s->ctx);
+    virtio_queue_aio_attach_host_notifier_no_poll(vs->event_vq, s->ctx);
 
     for (i = 0; i < vs->conf.num_queues; i++) {
         virtio_queue_aio_attach_host_notifier(vs->cmd_vqs[i], s->ctx);
diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
index 9d637e043e..67a873f54a 100644
--- a/hw/virtio/virtio.c
+++ b/hw/virtio/virtio.c
@@ -3534,6 +3534,19 @@ void virtio_queue_aio_attach_host_notifier(VirtQueue *vq, AioContext *ctx)
                                 virtio_queue_host_notifier_aio_poll_end);
 }
 
+/*
+ * Same as virtio_queue_aio_attach_host_notifier() but without polling. Use
+ * this for rx virtqueues and similar cases where the virtqueue handler
+ * function does not pop all elements. When the virtqueue is left non-empty
+ * polling consumes CPU cycles and should not be used.
+ */
+void virtio_queue_aio_attach_host_notifier_no_poll(VirtQueue *vq, AioContext *ctx)
+{
+    aio_set_event_notifier(ctx, &vq->host_notifier, true,
+                           virtio_queue_host_notifier_read,
+                           NULL, NULL);
+}
+
 void virtio_queue_aio_detach_host_notifier(VirtQueue *vq, AioContext *ctx)
 {
     aio_set_event_notifier(ctx, &vq->host_notifier, true, NULL, NULL, NULL);
diff --git a/include/hw/virtio/virtio.h b/include/hw/virtio/virtio.h
index b31c4507f5..b62a35fdca 100644
--- a/include/hw/virtio/virtio.h
+++ b/include/hw/virtio/virtio.h
@@ -317,6 +317,7 @@ EventNotifier *virtio_queue_get_host_notifier(VirtQueue *vq);
 void virtio_queue_set_host_notifier_enabled(VirtQueue *vq, bool enabled);
 void virtio_queue_host_notifier_read(EventNotifier *n);
 void virtio_queue_aio_attach_host_notifier(VirtQueue *vq, AioContext *ctx);
+void virtio_queue_aio_attach_host_notifier_no_poll(VirtQueue *vq, AioContext *ctx);
 void virtio_queue_aio_detach_host_notifier(VirtQueue *vq, AioContext *ctx);
 VirtQueue *virtio_vector_first_queue(VirtIODevice *vdev, uint16_t vector);
 VirtQueue *virtio_vector_next_queue(VirtQueue *vq);