Re: [PATCH RFC v7 03/14] vhost: use batched get_vq_desc version

[Eugenio Pérez]

On Thu, 2020-06-11 at 07:30 -0400, Michael S. Tsirkin wrote:
> On Wed, Jun 10, 2020 at 06:18:32PM +0200, Eugenio Perez Martin wrote:
> > On Wed, Jun 10, 2020 at 5:13 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > On Wed, Jun 10, 2020 at 02:37:50PM +0200, Eugenio Perez Martin wrote:
> > > > > +/* This function returns a value > 0 if a descriptor was found, or 0 if none were found.
> > > > > + * A negative code is returned on error. */
> > > > > +static int fetch_descs(struct vhost_virtqueue *vq)
> > > > > +{
> > > > > +       int ret;
> > > > > +
> > > > > +       if (unlikely(vq->first_desc >= vq->ndescs)) {
> > > > > +               vq->first_desc = 0;
> > > > > +               vq->ndescs = 0;
> > > > > +       }
> > > > > +
> > > > > +       if (vq->ndescs)
> > > > > +               return 1;
> > > > > +
> > > > > +       for (ret = 1;
> > > > > +            ret > 0 && vq->ndescs <= vhost_vq_num_batch_descs(vq);
> > > > > +            ret = fetch_buf(vq))
> > > > > +               ;
> > > > 
> > > > (Expanding comment in V6):
> > > > 
> > > > We get an infinite loop this way:
> > > > * vq->ndescs == 0, so we call fetch_buf() here
> > > > * fetch_buf gets less than vhost_vq_num_batch_descs(vq); descriptors. ret = 1
> > > > * This loop calls again fetch_buf, but vq->ndescs > 0 (and avail_vq ==
> > > > last_avail_vq), so it just return 1
> > > 
> > > That's what
> > >          [PATCH RFC v7 08/14] fixup! vhost: use batched get_vq_desc version
> > > is supposed to fix.
> > > 
> > 
> > Sorry, I forgot to include that fixup.
> > 
> > With it I don't see CPU stalls, but with that version latency has
> > increased a lot and I see packet lost:
> > + ping -c 5 10.200.0.1
> > PING 10.200.0.1 (10.200.0.1) 56(84) bytes of data.
> > > From 10.200.0.2 icmp_seq=1 Destination Host Unreachable
> > > From 10.200.0.2 icmp_seq=2 Destination Host Unreachable
> > > From 10.200.0.2 icmp_seq=3 Destination Host Unreachable
> > 64 bytes from 10.200.0.1: icmp_seq=5 ttl=64 time=6848 ms
> > 
> > --- 10.200.0.1 ping statistics ---
> > 5 packets transmitted, 1 received, +3 errors, 80% packet loss, time 76ms
> > rtt min/avg/max/mdev = 6848.316/6848.316/6848.316/0.000 ms, pipe 4
> > --
> > 
> > I cannot even use netperf.
> 
> OK so that's the bug to try to find and fix I think.
> 
> 
> > If I modify with my proposed version:
> > + ping -c 5 10.200.0.1
> > PING 10.200.0.1 (10.200.0.1) 56(84) bytes of data.
> > 64 bytes from 10.200.0.1: icmp_seq=1 ttl=64 time=7.07 ms
> > 64 bytes from 10.200.0.1: icmp_seq=2 ttl=64 time=0.358 ms
> > 64 bytes from 10.200.0.1: icmp_seq=3 ttl=64 time=5.35 ms
> > 64 bytes from 10.200.0.1: icmp_seq=4 ttl=64 time=2.27 ms
> > 64 bytes from 10.200.0.1: icmp_seq=5 ttl=64 time=0.426 ms
> 
> Not sure which version this is.
> 
> > [root@localhost ~]# netperf -H 10.200.0.1 -p 12865 -l 10 -t TCP_STREAM
> > MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> > 10.200.0.1 () port 0 AF_INET
> > Recv   Send    Send
> > Socket Socket  Message  Elapsed
> > Size   Size    Size     Time     Throughput
> > bytes  bytes   bytes    secs.    10^6bits/sec
> > 
> > 131072  16384  16384    10.01    4742.36
> > [root@localhost ~]# netperf -H 10.200.0.1 -p 12865 -l 10 -t UDP_STREAM
> > MIGRATED UDP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
> > 10.200.0.1 () port 0 AF_INET
> > Socket  Message  Elapsed      Messages
> > Size    Size     Time         Okay Errors   Throughput
> > bytes   bytes    secs            #      #   10^6bits/sec
> > 
> > 212992   65507   10.00        9214      0     482.83
> > 212992           10.00        9214            482.83
> > 
> > I will compare with the non-batch version for reference, but the
> > difference between the two is noticeable. Maybe it's worth finding a
> > good value for the if() inside fetch_buf?
> > 
> > Thanks!
> > 
> 
> I don't think it's performance, I think it's a bug somewhere,
> e.g. maybe we corrupt a packet, or stall the queue, or
> something like this.
> 
> Let's do this, I will squash the fixups and post v8 so you can bisect
> and then debug cleanly.

Ok, so if we apply the patch proposed in v7 08/14 (Or the version 8 of the patchset sent), this is what happens:

1. Userland (virtio_test in my case) introduces just one buffer in vq, and it kicks
2. vhost module reaches fetch_descs, called from vhost_get_vq_desc. From there we call fetch_buf in a for loop.
3. The first time we call fetch_buf, it returns properly one buffer. However, the second time we call it, it returns 0
because vq->avail_idx == vq->last_avail_idx and vq->avail_idx == last_avail_idx code path.
4. fetch_descs assign ret = 0, so it returns 0. vhost_get_vq_desc will goto err, and it will signal no new buffer
(returning vq->num).

So to fix it and maintain the batching maybe we could return vq->ndescs in case ret == 0:

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index c0dfb5e3d2af..5993d4f34ca9 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -2315,7 +2327,8 @@ static int fetch_descs(struct vhost_virtqueue *vq)
 
 	/* On success we expect some descs */
 	BUG_ON(ret > 0 && !vq->ndescs);
-	return ret;
+	return ret ?: vq->ndescs;
 }
 
 /* Reverse the effects of fetch_descs */
--

Another possibility could be to return different codes from fetch_buf, but I find the suggested modification easier.

What do you think?

Thanks!