From mboxrd@z Thu Jan 1 00:00:00 1970 From: Michal Hocko Subject: Re: [RFC PATCH] mm: memcg: fix css double put in mem_cgroup_iter Date: Wed, 26 Jul 2017 15:44:51 +0200 Message-ID: <20170726134451.GR2981@dhcp22.suse.cz> References: <20170726130742.5976-1-wenwei.tww@gmail.com> Mime-Version: 1.0 Return-path: Content-Disposition: inline In-Reply-To: <20170726130742.5976-1-wenwei.tww-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> Sender: cgroups-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org List-ID: Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Wenwei Tao Cc: hannes-druUgvl0LCNAfugRpC6u6w@public.gmane.org, bsingharora-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org, kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org, yuwang.yuwang-gPhfCIXyaqCqndwCJWfcng@public.gmane.org, cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Wenwei Tao On Wed 26-07-17 21:07:42, Wenwei Tao wrote: > From: Wenwei Tao > > By removing the child cgroup while the parent cgroup is > under reclaim, we could trigger the following kernel panic > on kernel 3.10: > ------------------------------------------------------------------------ > kernel BUG at kernel/cgroup.c:893! > invalid opcode: 0000 [#1] SMP > CPU: 1 PID: 22477 Comm: kworker/1:1 Not tainted 3.10.107 #1 > Workqueue: cgroup_destroy css_dput_fn > task: ffff8817959a5780 ti: ffff8817e8886000 task.ti: ffff8817e8886000 > RIP: 0010:[] [] > cgroup_diput+0xc0/0xf0 > RSP: 0000:ffff8817e8887da0 EFLAGS: 00010246 > RAX: 0000000000000000 RBX: ffff8817a5dd5d40 RCX: dead000000000200 > RDX: 0000000000000000 RSI: ffff8817973a6910 RDI: ffff8817f54c2a00 > RBP: ffff8817e8887dc8 R08: ffff8817a5dd5dd0 R09: df9fb35794b01820 > R10: df9fb35794b01820 R11: 00007fa95b1efcda R12: ffff8817a5dd5d9c > R13: ffff8817f38b3a40 R14: ffff8817973a6910 R15: ffff8817973a6910 > FS: 0000000000000000(0000) GS:ffff88181f220000(0000) > knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 00007fa6e6234000 CR3: 000000179f19d000 CR4: 00000000000407e0 > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 > Stack: > ffff8817a5dd5d40 ffff8817a5dd5d9c ffff8817f38b3a40 ffff8817973a6910 > 0000000000000040 ffff8817e8887df8 ffffffff811b37c2 ffff8817fa23c000 > ffff8817f57dbb80 ffff88181f232ac0 ffff88181f237500 ffff8817e8887e10 > Call Trace: > [] dput+0x1a2/0x2f0 > [] cgroup_dput.isra.21+0x1c/0x30 > [] css_dput_fn+0x1d/0x20 > [] process_one_work+0x17c/0x460 > [] worker_thread+0x116/0x3b0 > [] ? manage_workers.isra.25+0x290/0x290 > [] kthread+0xc0/0xd0 > [] ? insert_kthread_work+0x40/0x40 > [] ret_from_fork+0x58/0x90 > [] ? insert_kthread_work+0x40/0x40 > Code: 41 5e 41 5f 5d c3 0f 1f 44 00 00 48 8b 7f 78 48 8b 07 a8 01 74 15 > 48 81 c7 30 01 00 00 48 c7 c6 a0 a7 0c 81 e8 b2 83 02 00 eb c8 <0f> 0b > 49 8b 4e 18 48 c7 c2 7e f1 7a 81 be 85 03 00 00 48 c7 c7 > RIP [] cgroup_diput+0xc0/0xf0 > RSP > ---[ end trace 85eeea5212c44f51 ]--- > ------------------------------------------------------------------------ > > I think there is a css double put in mem_cgroup_iter. Under reclaim, > we call mem_cgroup_iter the first time with prev == NULL, and we get > last_visited memcg from per zone's reclaim_iter then call __mem_cgroup_iter_next > try to get next alive memcg, __mem_cgroup_iter_next could return NULL > if last_visited is already the last one so we put the last_visited's > memcg css and continue to the next while loop, this time we might not > do css_tryget(&last_visited->css) if the dead_count is changed, but > we still do css_put(&last_visited->css), we put it twice, this could > trigger the BUG_ON at kernel/cgroup.c:893. Yes, I guess your are right and I suspect that this has been silently fixed by 519ebea3bf6d ("mm: memcontrol: factor out reclaim iterator loading and updating"). I think a more appropriate fix is would be. Are you able to reproduce and re-test it? --- diff --git a/mm/memcontrol.c b/mm/memcontrol.c index 437ae2cbe102..0848ec05c12a 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -1224,6 +1224,8 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, if (last_visited && last_visited != root && !css_tryget(&last_visited->css)) last_visited = NULL; + } else { + last_visited = true; } } -- Michal Hocko SUSE Labs