Message-ID: <CAEA6p_BhAh6f_kAHEoEJ38nunY=c=4WqxhJQUjT+dCSAr_rm8g@mail.gmail.com>
Date:   Wed, 17 Aug 2022 11:16:18 -0700
From:   Wei Wang <weiwan@...gle.com>
To:     Shakeel Butt <shakeelb@...gle.com>
Cc:     Johannes Weiner <hannes@...xchg.org>,
        Eric Dumazet <edumazet@...gle.com>,
        netdev <netdev@...r.kernel.org>,
        Gražvydas Ignotas <notasas@...il.com>,
        Michal Hocko <mhocko@...e.com>, Roman Gushchin <guro@...com>,
        Linux MM <linux-mm@...ck.org>,
        Cgroups <cgroups@...r.kernel.org>
Subject: Re: UDP rx packet loss in a cgroup with a memory limit

On Wed, Aug 17, 2022 at 10:37 AM Shakeel Butt <shakeelb@...gle.com> wrote:
>
> + Eric and netdev
>
> On Wed, Aug 17, 2022 at 10:13 AM Johannes Weiner <hannes@...xchg.org> wrote:
> >
> > On Wed, Aug 17, 2022 at 07:50:13PM +0300, Gražvydas Ignotas wrote:
> > > On Tue, Aug 16, 2022 at 9:52 PM Gražvydas Ignotas <notasas@...il.com> wrote:
> > > > Basically, when there is git activity in the container with a memory
> > > > limit, other processes in the same container start to suffer (very)
> > > > occasional network issues (mostly DNS lookup failures).
> > >
> > > ok I've traced this and it's failing in try_charge_memcg(), which
> > > doesn't seem to be trying too hard because it's called from irq
> > > context.
> > >
> > > Here is the backtrace:
> > >  <IRQ>
> > >  ? fib_validate_source+0xb4/0x100
> > >  ? ip_route_input_slow+0xa11/0xb70
> > >  mem_cgroup_charge_skmem+0x4b/0xf0
> > >  __sk_mem_raise_allocated+0x17f/0x3e0
> > >  __udp_enqueue_schedule_skb+0x220/0x270
> > >  udp_queue_rcv_one_skb+0x330/0x5e0
> > >  udp_unicast_rcv_skb+0x75/0x90
> > >  __udp4_lib_rcv+0x1ba/0xca0
> > >  ? ip_rcv_finish_core.constprop.0+0x63/0x490
> > >  ip_protocol_deliver_rcu+0xd6/0x230
> > >  ip_local_deliver_finish+0x73/0xa0
> > >  __netif_receive_skb_one_core+0x8b/0xa0
> > >  process_backlog+0x8e/0x120
> > >  __napi_poll+0x2c/0x160
> > >  net_rx_action+0x2a2/0x360
> > >  ? rebalance_domains+0xeb/0x3b0
> > >  __do_softirq+0xeb/0x2eb
> > >  __irq_exit_rcu+0xb9/0x110
> > >  sysvec_apic_timer_interrupt+0xa2/0xd0
> > >  </IRQ>
> > >
> > > Calling mem_cgroup_print_oom_meminfo() in such a case reveals:
> > >
> > > memory: usage 7812476kB, limit 7812500kB, failcnt 775198
> > > swap: usage 0kB, limit 0kB, failcnt 0
> > > Memory cgroup stats for
> > > /kubepods.slice/kubepods-burstable.slice/kubepods-burstable-podb8f4f0e9_fb95_4f2d_8443_e6a78f235c9a.slice/docker-9e7cad93b2e0774d49148474989b41fe6d67a5985d059d08d9d64495f1539a81.scope:
> > > anon 348016640
> > > file 7502163968
> > > kernel 146997248
> > > kernel_stack 327680
> > > pagetables 2224128
> > > percpu 0
> > > sock 4096
> > > vmalloc 0
> > > shmem 0
> > > zswap 0
> > > zswapped 0
> > > file_mapped 112041984
> > > file_dirty 1181028352
> > > file_writeback 2686976
> > > swapcached 0
> > > anon_thp 44040192
> > > file_thp 0
> > > shmem_thp 0
> > > inactive_anon 350756864
> > > active_anon 36864
> > > inactive_file 3614003200
> > > active_file 3888070656
> > > unevictable 0
> > > slab_reclaimable 143692600
> > > slab_unreclaimable 545120
> > > slab 144237720
> > > workingset_refault_anon 0
> > > workingset_refault_file 2318
> > > workingset_activate_anon 0
> > > workingset_activate_file 2318
> > > workingset_restore_anon 0
> > > workingset_restore_file 0
> > > workingset_nodereclaim 0
> > > pgfault 334152
> > > pgmajfault 1238
> > > pgrefill 3400
> > > pgscan 819608
> > > pgsteal 791005
> > > pgactivate 949122
> > > pgdeactivate 1694
> > > pglazyfree 0
> > > pglazyfreed 0
> > > zswpin 0
> > > zswpout 0
> > > thp_fault_alloc 709
> > > thp_collapse_alloc 0
> > >
> > > So it basically renders UDP inoperable because of disk cache. I hope
> > > this is not the intended behavior. Naturally booting with
> > > cgroup.memory=nosocket solves this issue.
> >
> > This is most likely a regression caused by this patch:
> >
> > commit 4b1327be9fe57443295ae86fe0fcf24a18469e9f
> > Author: Wei Wang <weiwan@...gle.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@...gle.com>
> >     Reviewed-by: Shakeel Butt <shakeelb@...gle.com>
> >     Signed-off-by: David S. Miller <davem@...emloft.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 given patch has not changed any heuristic. It has only changed
> when forced charging happens. After the path the initial call
> mem_cgroup_charge_skmem() can fail and we take suppress_allocation
> code path and if heuristics allow, we force charge with __GFP_NOFAIL.

Powered by blists - more mailing lists