Re: [REGRESSION] Massive virtio-net throughput drop in guest VM with Linux 6.8+

[Markus Fohrer <markus.fohrer@webked.de>]

Am Donnerstag, dem 03.04.2025 um 10:26 -0400 schrieb Willem de Bruijn:
> Michael S. Tsirkin wrote:
> > On Thu, Apr 03, 2025 at 03:51:01PM +0200, Markus Fohrer wrote:
> > > Am Donnerstag, dem 03.04.2025 um 09:04 -0400 schrieb Michael S.
> > > Tsirkin:
> > > > On Wed, Apr 02, 2025 at 11:12:07PM +0200, Markus Fohrer wrote:
> > > > > Hi,
> > > > > 
> > > > > I'm observing a significant performance regression in KVM
> > > > > guest VMs
> > > > > using virtio-net with recent Linux kernels (6.8.1+ and 6.14).
> > > > > 
> > > > > When running on a host system equipped with a Broadcom
> > > > > NetXtreme-E
> > > > > (bnxt_en) NIC and AMD EPYC CPUs, the network throughput in
> > > > > the
> > > > > guest drops to 100–200 KB/s. The same guest configuration
> > > > > performs
> > > > > normally (~100 MB/s) when using kernel 6.8.0 or when the VM
> > > > > is
> > > > > moved to a host with Intel NICs.
> > > > > 
> > > > > Test environment:
> > > > > - Host: QEMU/KVM, Linux 6.8.1 and 6.14.0
> > > > > - Guest: Linux with virtio-net interface
> > > > > - NIC: Broadcom BCM57416 (bnxt_en driver, no issues at host
> > > > > level)
> > > > > - CPU: AMD EPYC
> > > > > - Storage: virtio-scsi
> > > > > - VM network: virtio-net, virtio-scsi (no CPU or IO
> > > > > bottlenecks)
> > > > > - Traffic test: iperf3, scp, wget consistently slow in guest
> > > > > 
> > > > > This issue is not present:
> > > > > - On 6.8.0 
> > > > > - On hosts with Intel NICs (same VM config)
> > > > > 
> > > > > I have bisected the issue to the following upstream commit:
> > > > > 
> > > > >   49d14b54a527 ("virtio-net: Suppress tx timeout warning for
> > > > > small
> > > > > tx")
> > > > >   https://git.kernel.org/linus/49d14b54a527
> > > > 
> > > > Thanks a lot for the info!
> > > > 
> > > > 
> > > > both the link and commit point at:
> > > > 
> > > > commit 49d14b54a527289d09a9480f214b8c586322310a
> > > > Author: Eric Dumazet <edumazet@google.com>
> > > > Date:   Thu Sep 26 16:58:36 2024 +0000
> > > > 
> > > >     net: test for not too small csum_start in
> > > > virtio_net_hdr_to_skb()
> > > >     
> > > > 
> > > > is this what you mean?
> > > > 
> > > > I don't know which commit is "virtio-net: Suppress tx timeout
> > > > warning
> > > > for small tx"
> > > > 
> > > > 
> > > > 
> > > > > Reverting this commit restores normal network performance in
> > > > > affected guest VMs.
> > > > > 
> > > > > I’m happy to provide more data or assist with testing a
> > > > > potential
> > > > > fix.
> > > > > 
> > > > > Thanks,
> > > > > Markus Fohrer
> > > > 
> > > > 
> > > > Thanks! First I think it's worth checking what is the setup,
> > > > e.g.
> > > > which offloads are enabled.
> > > > Besides that, I'd start by seeing what's doing on. Assuming I'm
> > > > right
> > > > about
> > > > Eric's patch:
> > > > 
> > > > diff --git a/include/linux/virtio_net.h
> > > > b/include/linux/virtio_net.h
> > > > index 276ca543ef44d8..02a9f4dc594d02 100644
> > > > --- a/include/linux/virtio_net.h
> > > > +++ b/include/linux/virtio_net.h
> > > > @@ -103,8 +103,10 @@ static inline int
> > > > virtio_net_hdr_to_skb(struct
> > > > sk_buff *skb,
> > > >  
> > > >  		if (!skb_partial_csum_set(skb, start, off))
> > > >  			return -EINVAL;
> > > > +		if (skb_transport_offset(skb) < nh_min_len)
> > > > +			return -EINVAL;
> > > >  
> > > > -		nh_min_len = max_t(u32, nh_min_len,
> > > > skb_transport_offset(skb));
> > > > +		nh_min_len = skb_transport_offset(skb);
> > > >  		p_off = nh_min_len + thlen;
> > > >  		if (!pskb_may_pull(skb, p_off))
> > > >  			return -EINVAL;
> > > > 
> > > > 
> > > > sticking a printk before return -EINVAL to show the offset and
> > > > nh_min_len
> > > > would be a good 1st step. Thanks!
> > > > 
> > > 
> > > 
> > > Hi Eric,
> > > 
> > > thanks a lot for the quick response — and yes, you're absolutely
> > > right.
> > > 
> > > Apologies for the confusion: I mistakenly wrote the wrong commit
> > > description in my initial mail.
> > > 
> > > The correct commit is indeed:
> > > 
> > > commit 49d14b54a527289d09a9480f214b8c586322310a
> > > Author: Eric Dumazet <edumazet@google.com>
> > > Date:   Thu Sep 26 16:58:36 2024 +0000
> > > 
> > >     net: test for not too small csum_start in
> > > virtio_net_hdr_to_skb()
> > > 
> > > This is the one I bisected and which causes the performance
> > > regression
> > > in my environment.
> 
> This commit is introduced in v6.12.
> 
> You say 6.8 is good, but 6.8.1 is bad. This commit is not in 6.8.1.
> Nor any virtio-net related change:
> 
> $ git log --oneline linux/v6.8..linux/v6.8.1 --
> include/linux/virtio_net.h drivers/net/virtio_net.c | wc -l
> 0
> 
> Is it perhaps a 6.8.1 derived distro kernel?
> 
> That patch detects silly packets created by a fuzzer. It should not
> affect sane traffic. Not saying your analysis is wrong. We just need
> more data to understand the regression better.


To clarify: my earlier tests were based on Ubuntu-patched kernels
(e.g., 6.8.0-31 to 6.8.0-53).

I've now repeated the tests using clean mainline kernels from
kernel.org.

Download speed was measured using:
  wget -O /dev/null http://speedtest.belwue.net/10G

Results:
- Kernel 6.11: ~85 MB/s
- Kernel 6.12 and 6.14: < 200 KB/s

This confirms that the regression was introduced between v6.11 and
v6.12 in upstream.