Re: UDP rx packet loss in a cgroup with a memory limit

[Shakeel Butt <shakeelb@google.com>]

On Wed, Aug 17, 2022 at 1:12 PM Gražvydas Ignotas <notasas@gmail.com> wrote:
>
> On Wed, Aug 17, 2022 at 9:16 PM Wei Wang <weiwan@google.com> wrote:
> >
> > On Wed, Aug 17, 2022 at 10:37 AM Shakeel Butt <shakeelb@google.com> wrote:
> > >
> > > + Eric and netdev
> > >
> > > On Wed, Aug 17, 2022 at 10:13 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
> > > >
> > > > This is most likely a regression caused by this patch:
> > > >
> > > > commit 4b1327be9fe57443295ae86fe0fcf24a18469e9f
> > > > Author: Wei Wang <weiwan@google.com>
> > > > Date:   Tue Aug 17 12:40:03 2021 -0700
> > > >
> > > >     net-memcg: pass in gfp_t mask to mem_cgroup_charge_skmem()
> > > >
> > > >     Add gfp_t mask as an input parameter to mem_cgroup_charge_skmem(),
> > > >     to give more control to the networking stack and enable it to change
> > > >     memcg charging behavior. In the future, the networking stack may decide
> > > >     to avoid oom-kills when fallbacks are more appropriate.
> > > >
> > > >     One behavior change in mem_cgroup_charge_skmem() by this patch is to
> > > >     avoid force charging by default and let the caller decide when and if
> > > >     force charging is needed through the presence or absence of
> > > >     __GFP_NOFAIL.
> > > >
> > > >     Signed-off-by: Wei Wang <weiwan@google.com>
> > > >     Reviewed-by: Shakeel Butt <shakeelb@google.com>
> > > >     Signed-off-by: David S. Miller <davem@davemloft.net>
> > > >
> > > > We never used to fail these allocations. Cgroups don't have a
> > > > kswapd-style watermark reclaimer, so the network relied on
> > > > force-charging and leaving reclaim to allocations that can block.
> > > > Now it seems network packets could just fail indefinitely.
> > > >
> > > > The changelog is a bit terse given how drastic the behavior change
> > > > is. Wei, Shakeel, can you fill in why this was changed? Can we revert
> > > > this for the time being?
> > >
> > > Does reverting the patch fix the issue? However I don't think it will.
> > >
> > > Please note that we still have the force charging as before this
> > > patch. Previously when mem_cgroup_charge_skmem() force charges, it
> > > returns false and __sk_mem_raise_allocated takes suppress_allocation
> > > code path. Based on some heuristics, it may allow it or it may
> > > uncharge and return failure.
> >
> > The force charging logic in __sk_mem_raise_allocated only gets
> > considered on tx path for STREAM socket. So it probably does not take
> > effect on UDP path. And, that logic is NOT being altered in the above
> > patch.
> > So specifically for UDP receive path, what happens in
> > __sk_mem_raise_allocated() BEFORE the above patch is:
> > - mem_cgroup_charge_skmem() gets called:
> >     - try_charge() with GFP_NOWAIT gets called and  failed
> >     - try_charge() with __GFP_NOFAIL
> >     - return false
> > - goto suppress_allocation:
> >     - mem_cgroup_uncharge_skmem() gets called
> > - return 0 (which means failure)
> >
> > AFTER the above patch, what happens in __sk_mem_raise_allocated() is:
> > - mem_cgroup_charge_skmem() gets called:
> >     - try_charge() with GFP_NOWAIT gets called and failed
> >     - return false
> > - goto suppress_allocation:
> >     - We no longer calls mem_cgroup_uncharge_skmem()
> > - return 0
> >
> > So I agree with Shakeel, that this change shouldn't alter the behavior
> > of the above call path in such a situation.
> > But do let us know if reverting this change has any effect on your test.
>
> The problem is still there (the kernel wasn't compiling after revert,
> had to adjust another seemingly unrelated callsite). It's hard to tell
> if it's better or worse since it happens so randomly.
>

Hello everyone, we have a better understanding why the patch pointed
out by Johannes might have exposed this issue. See
https://lore.kernel.org/all/20221013041833.rhifxw4gqwk4ofi2@google.com/.

To summarize, the old code was depending on a subtle interaction of
force-charge and percpu charge caches which this patch removed. The
fix I am proposing is for the network stack to be explicit of its need
(i.e. use GFP_ATOMIC) instead of depending on a subtle behavior.