Re: [patch net-next] virtio_net: add support for Byte Queue Limits

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Jun 06, 2024 at 07:42:35PM +0800, Jason Xing wrote:
> On Thu, Jun 6, 2024 at 12:25 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Thu, Jun 6, 2024 at 10:59 AM Jason Xing <kerneljasonxing@gmail.com> wrote:
> > >
> > > Hello Jason,
> > >
> > > On Thu, Jun 6, 2024 at 8:21 AM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Wed, Jun 5, 2024 at 7:51 PM Heng Qi <hengqi@linux.alibaba.com> wrote:
> > > > >
> > > > > On Wed, 5 Jun 2024 13:30:51 +0200, Jiri Pirko <jiri@resnulli.us> wrote:
> > > > > > Mon, May 20, 2024 at 02:48:15PM CEST, jiri@resnulli.us wrote:
> > > > > > >Fri, May 10, 2024 at 09:11:16AM CEST, hengqi@linux.alibaba.com wrote:
> > > > > > >>On Thu,  9 May 2024 13:46:15 +0200, Jiri Pirko <jiri@resnulli.us> wrote:
> > > > > > >>> From: Jiri Pirko <jiri@nvidia.com>
> > > > > > >>>
> > > > > > >>> Add support for Byte Queue Limits (BQL).
> > > > > > >>
> > > > > > >>Historically both Jason and Michael have attempted to support BQL
> > > > > > >>for virtio-net, for example:
> > > > > > >>
> > > > > > >>https://lore.kernel.org/netdev/21384cb5-99a6-7431-1039-b356521e1bc3@redhat.com/
> > > > > > >>
> > > > > > >>These discussions focus primarily on:
> > > > > > >>
> > > > > > >>1. BQL is based on napi tx. Therefore, the transfer of statistical information
> > > > > > >>needs to rely on the judgment of use_napi. When the napi mode is switched to
> > > > > > >>orphan, some statistical information will be lost, resulting in temporary
> > > > > > >>inaccuracy in BQL.
> > > > > > >>
> > > > > > >>2. If tx dim is supported, orphan mode may be removed and tx irq will be more
> > > > > > >>reasonable. This provides good support for BQL.
> > > > > > >
> > > > > > >But when the device does not support dim, the orphan mode is still
> > > > > > >needed, isn't it?
> > > > > >
> > > > > > Heng, is my assuption correct here? Thanks!
> > > > > >
> > > > >
> > > > > Maybe, according to our cloud data, napi_tx=on works better than orphan mode in
> > > > > most scenarios. Although orphan mode performs better in specific benckmark,
> > > >
> > > > For example pktgen (I meant even if the orphan mode can break pktgen,
> > > > it can finish when there's a new packet that needs to be sent after
> > > > pktgen is completed).
> > > >
> > > > > perf of napi_tx can be enhanced through tx dim. Then, there is no reason not to
> > > > > support dim for devices that want the best performance.
> > > >
> > > > Ideally, if we can drop orphan mode, everything would be simplified.
> > >
> > > Please please don't do this. Orphan mode still has its merits. In some
> > > cases which can hardly be reproduced in production, we still choose to
> > > turn off the napi_tx mode because the delay of freeing a skb could
> > > cause lower performance in the tx path,
> >
> > Well, it's probably just a side effect and it depends on how to define
> > performance here.
> 
> Yes.
> 
> >
> > > which is, I know, surely
> > > designed on purpose.
> >
> > I don't think so and no modern NIC uses that. It breaks a lot of things.
> 
> To avoid confusion, I meant napi_tx mode can delay/slow down the speed
> in the tx path and no modern nic uses skb_orphan().

Clearly it's been designed for software NICs and when the
cost of interrupts is very high.

> I think I will have some time to test BQL in virtio_net.
> 
> Thanks,
> Jason