Re: [PATCH v4 mptcp-next 1/4] mptcp: never fetch fwd memory from the subflow

[Paolo Abeni <pabeni@redhat.com>]

On Tue, 2022-06-21 at 15:46 -0700, Mat Martineau wrote:
> On Tue, 21 Jun 2022, Mat Martineau wrote:
> 
> > On Tue, 21 Jun 2022, Paolo Abeni wrote:
> > 
> > > The memory accounting is broken in such exceptional code
> > > path, and after commit 4890b686f408 ("net: keep sk->sk_forward_alloc
> > > as small as possible") we can't find much help there.
> > > 
> > > Drop the broken code.
> > > 
> > > Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> > > ---
> > 
> > Thanks Paolo, v4 looks good.
> > 
> 
> One more question Paolo: given the SCTP regression that was bisected to 
> the fwd alloc change:
> 
> https://lore.kernel.org/netdev/20220619150456.GB34471@xsang-OptiPlex-9020/
> 
> is there any need to wait before upstreaming this series for net-next? 

I'm guild-guessing 4890b686f408 could cause regressions (for TCP and as
as consequence for everything else using it) when cgroups are in use,
because after that patch the cgroup memory accounting hooks are called
~twice per packet (one to increase, one to decrease the cgroup
accounted memory), instead that almost never before that patch.

In any case we can't do much to counter such regression. 

I did some benchmarking on top of this patches, and I could not measure
relevant changes (but no cgroups enabled).

I'm not 110% sure that the regressions reported by LTP are relevant.
Once upon time I hit a quite large one, and it was actually due to test
instability.

I agree staging this changes for a bit will not hurt/will be for the
better.

Cheers,

Paolo