Re: [PATCH] mm: avoid bothering interrupted task when charge memcg in softirq

[Shakeel Butt]

On Sun, Jul 15, 2018 at 6:50 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
>
> On Sun, Jul 15, 2018 at 11:04 PM, Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> > On Sun, Jul 15, 2018 at 1:02 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >>
> >> On Sun, Jul 15, 2018 at 2:34 PM, Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> >> > On Sat, Jul 14, 2018 at 10:26 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >> >>
> >> >> On Sun, Jul 15, 2018 at 12:25 PM, Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> >> >> > On Sat, Jul 14, 2018 at 7:10 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >> >> >>
> >> >> >> On Sat, Jul 14, 2018 at 11:38 PM, Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> >> >> >> > On Sat, Jul 14, 2018 at 1:32 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >> >> >> >>
> >> >> >> >> try_charge maybe executed in packet receive path, which is in interrupt
> >> >> >> >> context.
> >> >> >> >> In this situation, the 'current' is the interrupted task, which may has
> >> >> >> >> no relation to the rx softirq, So it is nonsense to use 'current'.
> >> >> >> >>
> >> >> >> >
> >> >> >> > Have you actually seen this occurring?
> >> >> >>
> >> >> >> Hi Shakeel,
> >> >> >>
> >> >> >> I'm trying to produce this issue, but haven't find it occur yet.
> >> >> >>
> >> >> >> > I am not very familiar with the
> >> >> >> > network code but I can think of two ways try_charge() can be called
> >> >> >> > from network code. Either through kmem charging or through
> >> >> >> > mem_cgroup_charge_skmem() and both locations correctly handle
> >> >> >> > interrupt context.
> >> >> >> >
> >> >> >>
> >> >> >> Why do you say that mem_cgroup_charge_skmem() correctly hanle
> >> >> >> interrupt context ?
> >> >> >>
> >> >> >> Let me show you why mem_cgroup_charge_skmem isn't hanling interrupt
> >> >> >> context correctly.
> >> >> >>
> >> >> >> mem_cgroup_charge_skmem() is calling  try_charge() twice.
> >> >> >> The first one is with GFP_NOWAIT as the gfp_mask, and the second one
> >> >> >> is with  (GFP_NOWAIT |  __GFP_NOFAIL) as the gfp_mask.
> >> >> >>
> >> >> >> If page_counter_try_charge() failes at the first time, -ENOMEM is returned.
> >> >> >> Then mem_cgroup_charge_skmem() will call try_charge() once more with
> >> >> >> __GFP_NOFAIL set, and this time if If page_counter_try_charge() failes
> >> >> >> again the '
> >> >> >> force' label in  try_charge() will be executed and 0 is returned.
> >> >> >>
> >> >> >> No matter what, the 'current' will be used and touched, that is
> >> >> >> meaning mem_cgroup_charge_skmem() isn't hanling the interrupt context
> >> >> >> correctly.
> >> >> >>
> >> >> >
> >> >> > Hi Yafang,
> >> >> >
> >> >> > If you check mem_cgroup_charge_skmem(), the memcg passed is not
> >> >> > 'current' but is from the sock object i.e. sk->sk_memcg for which the
> >> >> > network buffer is allocated for.
> >> >> >
> >> >>
> >> >> That's correct, the memcg if from the sock object.
> >> >> But the point is, in this situation why 'current' is used in try_charge() ?
> >> >> As 'current' is not related with the memcg, which is just a interrupted task.
> >> >>
> >> >
> >> > Hmm so you mean the behavior of memcg charging in the interrupt
> >> > context depends on the state of the interrupted task.
> >>
> >> Yes.
> >>
> >> > As you have
> >> > noted, mem_cgroup_charge_skmem() tries charging again with
> >> > __GFP_NOFAIL and the charge succeeds. Basically the memcg charging by
> >> > mem_cgroup_charge_skmem() will always succeed irrespective of the
> >> > state of the interrupted task. However mem_cgroup_charge_skmem() can
> >> > return true if the interrupted task was exiting or a fatal signal is
> >> > pending or oom victim or reclaiming memory. Can you please explain why
> >> > this is bad?
> >> >
> >>
> >> Let me show you the possible issues cause by this behavoir.
> >> 1.  In mem_cgroup_oom(), some  members in 'current' is set.
> >>      That means an innocent task will be in  task_in_memcg_oom state.
> >>      But this task may be in a different memcg, I mean the memcg of
> >> the 'current' may be differenct with the sk->sk_memcg.
> >>      Then when this innocent 'current' do try_charge it will hit  "if
> >> (unlikely(task_in_memcg_oom(current)))" and  -ENOMEM is returned,
> >> While there're maybe some free memory (or some memory could be freed )
> >> in the memcg of the innocent 'task'.
> >>
> >
> > No memory will be freed as try_charge() is in interrupt context.
> >
>
> I mean when this interrupted 'current' is running, that's in process context.
> In process context it should call try_to_free_mem_cgroup_pages() to
> free some memory,
> but it will hit "if (unlikely(task_in_memcg_oom(current)))"  before as
> it is set in the interrupt context.
>
> That's an obviously issue. Do you understand ?
>

Not really. I couldn't find where current->memcg_in_oom can be set in
the interrupt context.

> >> 2.  If the interrupted task was exiting or a fatal signal is  pending
> >> or oom victim,
> >>      it will directly goto force and 0 is returned, and then
> >> mem_cgroup_charge_skmem() will return true.
> >>      But mem_cgroup_charge_skmem() maybe need to try the second time
> >> and return false.
> >>
> >> That are all unexpected behavoir.
> >>
> >
> > Yes, this is inconsistent behavior. Can you explain how this will
> > affect network traffic? Basically mem_cgroup_charge_skmem() was
> > supposed to return false but sometime based on the interrupted task,
> > mem_cgroup_charge_skmem() returns true. How is this behavior bad for
> > network traffic?
> >
>
> You could see the funtion  __sk_mem_raise_allocated().
> If mem_cgroup_charge_skmem() return false, it will goto
> suppress_allocation and uncharge skmem,
> while when mem_cgroup_charge_skmem() return true, it will charge skmem
> sucessfully.
>
> The consequence behavior is  sk_rmem_schedule may fail while it should sucess.
> And then it will call  tcp_prune_queue() and tcp collapse may take a long time.
>

Is that a good thing or bad? From what I understand with your change
if charge fails, sk_rmem_schedule will always fail. However without
your change the interrupted task's state might help sk_rmem_schedule
to pass. I am all for consistent behavior but I wanted to make sure if
that is what you are aiming for.

Anyways, from what I remember Facebook is using the cgroup-v2's tcpmem
accounting. Johannes or Roman can shed some light if they have
observed this issue in production and might have opinion on how to
solve it.

thanks,
Shakeel