On Fri, 2024-02-02 at 17:13 +0100, Eric Dumazet wrote:
> On Fri, Feb 2, 2024 at 5:07 PM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
> >
> > In very slow environments, most big TCP cases including
> > segmentation and reassembly of big TCP packets have a good
> > chance to fail: by default the TCP client uses write size
> > well below 64K. If the host is low enough autocorking is
> > unable to build real big TCP packets.
> >
> > Address the issue using much larger write operations.
> >
> > Note that is hard to observe the issue without an extremely
> > slow and/or overloaded environment; reduce the TCP transfer
> > time to allow for much easier/faster reproducibility.
> >
> > Fixes: 6bb382bcf742 ("selftests: add a selftest for big tcp")
> > Signed-off-by: Paolo Abeni <pabeni@xxxxxxxxxx>
> > ---
> > tools/testing/selftests/net/big_tcp.sh | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/testing/selftests/net/big_tcp.sh b/tools/testing/selftests/net/big_tcp.sh
> > index cde9a91c4797..2db9d15cd45f 100755
> > --- a/tools/testing/selftests/net/big_tcp.sh
> > +++ b/tools/testing/selftests/net/big_tcp.sh
> > @@ -122,7 +122,9 @@ do_netperf() {
> > local netns=$1
> >
> > [ "$NF" = "6" ] && serip=$SERVER_IP6
> > - ip net exec $netns netperf -$NF -t TCP_STREAM -H $serip 2>&1 >/dev/null
> > +
> > + # use large write to be sure to generate big tcp packets
> > + ip net exec $netns netperf -$NF -t TCP_STREAM -l 1 -H $serip -- -m 262144 2>&1 >/dev/null
> > }
>
> Interesting.
>
> I think we set tcp_wmem[1] to 262144 in our hosts. I think netperf
> default depends on tcp_wmem[1]
I haven't dug into the netperf source, but the above would be
consistent with what I observe: in my VM I see 16Kb writes and
tcp_wmem[1] is 16K.
> Reviewed-by: Eric Dumazet <edumazet@xxxxxxxxxx>
Thanks!
Paolo
