On Wed, Mar 28, 2018 at 02:37:04PM +0800, Jason Wang wrote:
>
>
> On 2018年03月28日 12:01, haibinzhang(张海斌) wrote:
> > On 2018年03月27日 19:26, Jason wrote
> > On 2018年03月27日 17:12, haibinzhang wrote:
> > > > handle_tx() will delay rx for a long time when busy tx polling udp packets
> > > > with short length(ie: 1byte udp payload), because setting VHOST_NET_WEIGHT
> > > > takes into account only sent-bytes but no time.
> > > Interesting.
> > >
> > > Looking at vhost_can_busy_poll() it tries to poke pending vhost work and
> > > exit the busy loop if it found one. So I believe something block the
> > > work queuing. E.g did reverting 8241a1e466cd56e6c10472cac9c1ad4e54bc65db
> > > fix the issue?
> > "busy tx polling" means using netperf send udp packets with 1 bytes payload(total 47bytes frame lenght),
> > and handle_tx() will be busy sending packets continuously.
> >
> > > > It's not fair for handle_rx(),
> > > > so needs to limit max time of tx polling.
> > > >
> > > > ---
> > > > drivers/vhost/net.c | 3 ++-
> > > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> > > > index 8139bc70ad7d..dc9218a3a75b 100644
> > > > --- a/drivers/vhost/net.c
> > > > +++ b/drivers/vhost/net.c
> > > > @@ -473,6 +473,7 @@ static void handle_tx(struct vhost_net *net)
> > > > struct socket *sock;
> > > > struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
> > > > bool zcopy, zcopy_used;
> > > > + unsigned long start = jiffies;
> > > Checking jiffies is tricky, need to convert it to ms or whatever others.
> > >
> > > > mutex_lock(&vq->mutex);
> > > > sock = vq->private_data;
> > > > @@ -580,7 +581,7 @@ static void handle_tx(struct vhost_net *net)
> > > > else
> > > > vhost_zerocopy_signal_used(net, vq);
> > > > vhost_net_tx_packet(net);
> > > > - if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
> > > > + if (unlikely(total_len >= VHOST_NET_WEIGHT) || unlikely(jiffies - start >= 1)) {
> > > How value 1 is determined here? And we need a complete test to make sure
> > > this won't affect other use cases.
> > We just want <1ms ping latency, but actually we are not sure what value is reasonable.
> > We have some test results using netperf before this patch as follow,
> >
> > Udp payload 1byte 100bytes 1000bytes 1400bytes
> > Ping avg latency 25ms 10ms 2ms 1.5ms
> >
> > What is other testcases?
>
> Something like https://patchwork.kernel.org/patch/10151645/.
>
> Btw, you need use time_before() to properly handle jiffies overflow and I
> would also suggest you to try something like #packets limit (e.g 64).
Maybe a ring size?
> For long term, we definitely need more worker threads.
>
> Thanks
Only helps when you have spare CPUs.
> >
> > > Another thought is introduce another limit of #packets, but this need
> > > benchmark too.
> > >
> > > Thanks
> > >
> > > > vhost_poll_queue(&vq->poll);
> > > > break;
> > > > }
> > >
