Re: [PATCH 1/2] mm: memcg: switch to css_tryget() in get_mem_cgroup_from_mm()

[Roman Gushchin]

On Thu, Nov 07, 2019 at 06:02:00PM +0100, Michal Hocko wrote:
> On Thu 07-11-19 16:42:41, Roman Gushchin wrote:
> > On Thu, Nov 07, 2019 at 01:21:25PM +0100, Michal Hocko wrote:
> > > On Wed 06-11-19 14:51:30, Roman Gushchin wrote:
> > > > We've encountered a rcu stall in get_mem_cgroup_from_mm():
> > > > 
> > > >  rcu: INFO: rcu_sched self-detected stall on CPU
> > > >  rcu: 33-....: (21000 ticks this GP) idle=6c6/1/0x4000000000000002 softirq=35441/35441 fqs=5017
> > > >  (t=21031 jiffies g=324821 q=95837) NMI backtrace for cpu 33
> > > >  <...>
> > > >  RIP: 0010:get_mem_cgroup_from_mm+0x2f/0x90
> > > >  <...>
> > > >  __memcg_kmem_charge+0x55/0x140
> > > >  __alloc_pages_nodemask+0x267/0x320
> > > >  pipe_write+0x1ad/0x400
> > > >  new_sync_write+0x127/0x1c0
> > > >  __kernel_write+0x4f/0xf0
> > > >  dump_emit+0x91/0xc0
> > > >  writenote+0xa0/0xc0
> > > >  elf_core_dump+0x11af/0x1430
> > > >  do_coredump+0xc65/0xee0
> > > >  ? unix_stream_sendmsg+0x37d/0x3b0
> > > >  get_signal+0x132/0x7c0
> > > >  do_signal+0x36/0x640
> > > >  ? recalc_sigpending+0x17/0x50
> > > >  exit_to_usermode_loop+0x61/0xd0
> > > >  do_syscall_64+0xd4/0x100
> > > >  entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > > 
> > > > The problem is caused by an exiting task which is associated with
> > > > an offline memcg.
> > > 
> > > Hmm, how can we have a task in an offline memcg? I thought that any
> > > existing task will prevent cgroup removal from proceeding. Is this some
> > > sort of race where the task managed to disassociate from the cgroup
> > > while there is still a binding to a memcg existing? What am I missing?
> > 
> > It's an exiting task with the PF_EXITING flag set and it's in their late stages
> > of life.
> 
> This is a signal delivery path AFAIU (get_signal) and the coredumping
> happens before do_exit. My understanding is that that unlinking
> happens from cgroup_exit. So either I am misreading the backtrace or
> there is some other way to leave cgroups or there is something more
> going on.

Yeah, you're right. I have no better explanation for this and the similar,
mentioned in the commit bsd accounting issue, than some very rare race condition
that allows cgroups to be offlined with a task inside.

I'll think more about it.

Thanks, it's a really good question.

> 
> JFTR I am not really disputing the patch but I simply do not understand
> how the problem really happened.
> 
> -- 
> Michal Hocko
> SUSE Labs