Re: [virtio-comment] [PATCH] virtio-net: support per-queue coalescing moderation

[Heng Qi <hengqi@linux.alibaba.com>]

å 2023/2/9 äå4:52, Michael S. Tsirkin åé:
> On Wed, Feb 08, 2023 at 07:53:09PM +0200, Alvaro Karsz wrote:
> >   > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > Sent: Wednesday, February 8, 2023 9:48 AM
> > > >
> > > > On Wed, Feb 08, 2023 at 02:44:37PM +0000, Parav Pandit wrote:
> > > > > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > > > Sent: Wednesday, February 8, 2023 9:43 AM
> > > > > >
> > > > > > On Wed, Feb 08, 2023 at 02:37:55PM +0000, Parav Pandit wrote:
> > > > > > > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > > > > > Sent: Wednesday, February 8, 2023 9:18 AM
> > > > > > > >
> > > > > > > > On Wed, Feb 08, 2023 at 07:30:34PM +0800, Heng Qi wrote:
> > > > > > > > > > I see two options.
> > > > > > > > > > 1. Just have per VQ params. Software has the full knowledge
> > > > > > > > > > of in which it is
> > > > > > > > operating, and state remains at software level.
> > > > > > > > > > This effectively achieves both the mode.
> > > > > > > > > >
> > > > > > > > > > 2. Have a mode cmd,
> > > > > > > > > > Mode = (a) per device or (b) per VQ (c) disable After the
> > > > > > > > > > mode is set, driver can set per device or per VQ.
> > > > > > > > > I find this more clear.
> > > > > > > > >
> > > > > > > > > Thanks.
> > > > > > > > Rereading this I think I misunderstood the proposal.
> > > > > > > > Now we are burning memory on maintaining mode, and this
> > > > > > > > information is duplicated.
> > > > > > > It is not maintained in the pci resident memory, so it doesn't hurt.
> > > > > > >
> > > > > > > > I'd say let's just add a new command COAL_QUEUE_SET with vqn as
> > > > > > parameter.
> > > > > > > > Existing commands are simply defined as a shortcut to running
> > > > > > > > COAL_QUEUE_SET on all tx/rx queues respectively.
> > > > > > > >
> > > > > > > > Latest command dictates the parameters. To disable just set
> > > > > > > > everything to 0 (btw we should make this explicit in the spec,
> > > > > > > > but it can be
> > > > > > guessed from:
> > > > > > > > Upon reset, a device MUST initialize all coalescing parameters to 0.
> > > > > > > > )
> > > > > > > Switching between the modes (per q vs per device) implicitly is
> > > > > > > ambiguous
> > > > > > and it only means device may need to iterate.
> > > > > >
> > > > > > hmm i feel it's only ambiguous because i failed to explain in well.
> > > > > >
> > > > > > > This state is either better maintained in sw by always having per
> > > > > > > vq or have
> > > > > > clearly defined mode of what device should do.
> > > > > > > Per Q is very common even for several years old devices.
> > > > > > > Last time I counted, there were at least 15 such devices supporting it.
> > > > > > >
> > > > > > > So actual usage wise, I practically see that most implementations
> > > > > > > will end up
> > > > > > with per vq mode.
> > > > > > > I like to hear from Heng or Alvaro if they see any use of per device.
> > > > > >
> > > > > > Right so given this, most devices will be in per queue mode all the
> > > > > > time. why do you want a mode then? just keep per queue.
> > > > > > existing commands are kept around for compat but internally just
> > > > > > translate to per-queue.
> > > > > Since the space is not released, do we need to keep the compat?
> > > > It's been accepted for half a year so we can't say for sure no one built this.
> > > That is likely but we should have the ability to have the Errata/ECN to correct it, specially for unrelease spec.
> > >
> > > > The way I propose is just a bit of firmware on device that scans all queues and
> > > > copies same parameters everywhere.
> > > This scanning loop in sw appears cheaper to me than some embedded fw.
> > > But is not a lot of concern.
> > >
> > > > Seems easier than worrying about this,
> > > > and we get disabling coalescing for free which you wanted. With an extra mode
> > > > its extra logic in the device fast path. Maybe it's cheap on hardware side but in
> > > > software it's an extra branch, not free.
> > > Most performant data path wouldn't implement and read the extra mode.
> > > It is always fw that is going to program same value, or per queue valued or disable value in each Q regardless whichever way we craft the CVQ cmd.
> > >
> > > The sequence that bothers me is below.
> > > 1. driver set global params
> > > 2. few minutes later, now driver set param for Q=1
> > >
> > > On this command, a device need to decide:
> > > Should Q = 2 to N
> > > (a) either work with previous globals, or
> > > (b) because per Q was set for one queue, they rest of the queues implicitly disable it.
> > >
> > > If it is (b),
> > > When a command on Q object =1 is issued, it affects other Q objects. <- This I want to avoid.
> > > A cmd that modifies the object, should only modify that object.
> > >
> > > If it is (a), it is mixed mode operation, which is ambiguous definition.
> > >
> > > A better semantic is to define such change at device level and no extra cost in the data path.
> > I think that (a) is the way to go.
> > I don't think that we should work with operation modes at all.
> >
> > In my opinion:
> >
> > We should have 2 features:
> > VIRTIO_NET_F_PERQUEUE_NOTF_COAL and VIRTIO_NET_F_NOTF_COAL.
> >
> > VIRTIO_NET_F_PERQUEUE_NOTF_COAL sets per queue parameters, and
> > VIRTIO_NET_F_NOTF_COAL sets parameters for all queues.
> >
> > VIRTIO_NET_F_NOTF_COAL has 2 commands:
> >      VIRTIO_NET_CTRL_NOTF_COAL_RX_SET
> >      VIRTIO_NET_CTRL_NOTF_COAL_TX_SET
> >
> > VIRTIO_NET_F_PERQUEUE_NOTF_COAL has 2 commands:
> >      VIRTIO_NET_CTRL_NOTF_COAL_PER_QUEUE_TX_SET
> >      VIRTIO_NET_CTRL_NOTF_COAL_PER_QUEUE_RX_SET
> >
> > We can see VIRTIO_NET_CTRL_NOTF_COAL_RX_SET as a virtio level shortcut
> > for setting all queues with one command, exactly as intended with
> > rx_qid= 0xFFFF, and without breaking devices following the current
> > spec.
> >
> > The device's FW can decide if it stores parameters received with
> > VIRTIO_NET_CTRL_NOTF_COAL_RX_SET in a global set, or if it iterates
> > through all queues, but IMO the best way it to iterate through all
> > queues.
> >
> > Seems like a win-win situation to me.
> > We achieve the same functionality as described in the patch, but
> > without breaking devices following the current spec.
> >
> > Now, if we follow this method,
> > VIRTIO_NET_CTRL_NOTF_COAL_PER_QUEUE_RX_SET with rx_qid= 0xFFF seems
> > redundant.
> > If VIRTIO_NET_F_PERQUEUE_NOTF_COAL requires VIRTIO_NET_F_NOTF_COAL, a
> > device supporting VIRTIO_NET_F_PERQUEUE_NOTF_COAL can achieve the same
> > functionality with the VIRTIO_NET_CTRL_NOTF_COAL_RX_SET command.
> Yes. Just some comments:
>
> - I don't think we need two commands. We have RX and TX because we
>    did not have vq number previously. No we do so just pass that.
>    It's also clearer since struct name can match command exactly.

Yes, I believe I got your point yesterday.

> - Once we do that we can use a short _VQ_ instead of the wordy "PER_QUEUE".

This seems clearer and more straightforward.

> - Accordingly a well understood "vqn" instead of our own "qid" which
>    we then need to define.

Sure.

Thanks.

> - And yes no need for a reserved "qid" - it's a distinct command.