On 23.01.24 12:12, Fiona Ebner wrote:

[...]

> I noticed poll_set_started() is not called, because
> ctx->fdmon_ops->need_wait(ctx) was true, i.e. ctx->poll_disable_cnt was
> positive (I'm using fdmon_poll). I then found this is because of the
> notifier for the event vq, being attached with
>
>> virtio_queue_aio_attach_host_notifier_no_poll(vs->event_vq, s->ctx);
> in virtio_scsi_dataplane_start(). But in virtio_scsi_drained_end() it is
> attached with virtio_queue_aio_attach_host_notifier() instead of the
> _no_poll() variant. So that might be the actual issue here?
>
>  From a quick test, I cannot see the CPU-usage-spike issue with the
> following either:
>
>> diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c
>> index 690aceec45..ba1ab8e410 100644
>> --- a/hw/scsi/virtio-scsi.c
>> +++ b/hw/scsi/virtio-scsi.c
>> @@ -1166,7 +1166,15 @@ static void virtio_scsi_drained_end(SCSIBus *bus)
>>   
>>       for (uint32_t i = 0; i < total_queues; i++) {
>>           VirtQueue *vq = virtio_get_queue(vdev, i);
>> -        virtio_queue_aio_attach_host_notifier(vq, s->ctx);
>> +        if (!virtio_queue_get_notification(vq)) {
>> +            virtio_queue_set_notification(vq, true);
>> +        }
>> +        if (vq == VIRTIO_SCSI_COMMON(s)->event_vq) {
>> +            virtio_queue_aio_attach_host_notifier_no_poll(vq, s->ctx);
>> +        } else {
>> +            virtio_queue_aio_attach_host_notifier(vq, s->ctx);
>> +        }
>> +        virtio_queue_notify(vdev, i);
>>       }
>>   }

Perfect, so we agree on trying it that way. :)

Hanna