On Mon, Feb 11, 2019 at 01:58:09AM -0500, Michael S. Tsirkin wrote:
> On Thu, Jan 31, 2019 at 01:16:29PM +0000, Savir, Gil wrote:
> > Hi,
> >
> >
> >
> > If VIRTIO_F_EVENT_IDX feature bit is negotiated, then Available Buffer
> > Notification Suppression mechanism used is avail event (not flags).
> >
> > The spec (both v1.0 / v1.1-draft) states that the device MAY use this mechanism
> > (Paragraph 2.4.9.2 / 2.6.10.2 respectively).
> >
> > This statement implies that the device may choose not to use this suppression
> > mechanism (even if VIRTIO_F_EVENT_IDX was negotiated).
> >
> >
> >
> > However â thereâs no way for the device to inform the driver that he is not
> > using avail_event.
> >
> > As consequence, since there will be a default value in avail_event (probably
> > 0x0), then the driver will always assume that it has to send notify âonce-per
> > ringâ.
>
>
> No I think this part is wrong, pls see below.
>
>
> > This will render performance futile, or force the device to actively update
> > avail_event.
> >
> >
> >
> > Is there a way for the device to inform the driver that he is not using
> > avail_event (and I missed it)?
> >
> >
> >
> > If yes, than my apologies for wasting your time.
> >
> > If no, then I suggest one of the following:
> >
> > Â Either, to change the âMAYâ (referred above) to âMUSTâ,
> >
>
> Thanks for the feedback!
>
> So we are talking about split queues.
>
> If avail_event is never set and remains 0, driver will send
> notifications every time index wraps around to 0, which
> would be every 2^16.
>
> This is I think expected since the spec says:
>
> The device MUST handle spurious notifications from the driver.
>
> 2^16 does not seem excessive so I don't think it will render performance
> futile.
To add to that, above is based on
2.6.10.1
Driver Requirements: Available Buffer Notification Suppression
which states:
After the driver writes a descriptor index into the available ring:
â If the idx field in the available ring (which determined where that descriptor index was placed)
was equal to avail_event, the driver MUST send a notification.
â Otherwise the driver SHOULD NOT send a notification.
Pls let the TC know whether taking the above into account you think
further clarification in the spec is necessary.
Thanks!
> > Â Or, to add way for the device to inform the driver that he is not
> > using avail_event (flag /certain reserved value in avail_event /other
> > mechanism).
> >
> >
> >
> > Thanks,
> >
> > Gil Savir
> >
> > Intel Corporation
>
> It might be a good addition to spec, probably along the lines
> of an option to send notifications on
> even suppression changes (which was proposed in the past, but
> was not included due to lack of time).
>
> --
> MST
