fanotify: unkillable hanged processes

[Dmitry Vyukov]

Hello,

The following program:

https://gist.githubusercontent.com/dvyukov/0952eeac71069b46b3fe0e28bd1a02bf/raw/396b9dcce2636cecab1a4161c15d3f066e6ef639/gistfile1.txt

if run in a parallel loop creates unkillable hanged processes:

-bash-4.3# ps afxu | grep a.out
root 4489 0.0 0.0 8868 340 ttyp1 S+ 17:19 0:00
\_ grep a.out
root 4423 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
root 4424 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
root 4425 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
root 4470 0.0 0.1 7016 2316 ttyp0 D 17:16 0:00
./stress ./a.out

This looks like a classical deadlock, but LOCKDEP is silent so
+LOCKDEP maintainers.

Most of the processes are hanged at (from /proc/pid/stack):

[<ffffffff814dd6c8>] __synchronize_srcu+0x248/0x380 kernel/rcu/srcu.c:448
[<ffffffff814dd81e>] synchronize_srcu+0x1e/0x40 kernel/rcu/srcu.c:492
[<ffffffff819448d7>] fsnotify_mark_destroy_list+0x107/0x370 fs/notify/mark.c:551
[<ffffffff8194163e>] fsnotify_destroy_group+0x1e/0xc0 fs/notify/group.c:57
[<ffffffff8194a84b>] fanotify_release+0x20b/0x2d0
fs/notify/fanotify/fanotify_user.c:392
[<ffffffff81868bbc>] __fput+0x28c/0x780 fs/file_table.c:208
[<ffffffff81869135>] ____fput+0x15/0x20 fs/file_table.c:244
[<ffffffff813ebf63>] task_work_run+0xf3/0x170 kernel/task_work.c:116
[< inline >] exit_task_work ./include/linux/task_work.h:21
[<ffffffff81394218>] do_exit+0x868/0x2e70 kernel/exit.c:828
[<ffffffff81396998>] do_group_exit+0x108/0x330 kernel/exit.c:958
[< inline >] SYSC_exit_group kernel/exit.c:969
[<ffffffff81396bdd>] SyS_exit_group+0x1d/0x20 kernel/exit.c:967
[<ffffffff86e10700>] entry_SYSCALL_64_fastpath+0x23/0xc1
arch/x86/entry/entry_64.S:208
[<ffffffffffffffff>] 0xffffffffffffffff

One process holds the srcu lock:

stress D ffffffff86dfe09a 28024 4470 1 0x00000004
ffff880064d9c440 0000000000000000 ffff8800686ab0c0 ffff88006cecc080
ffff88006d322bd8 ffff8800653f7630 ffffffff86dfe09a ffff8800653f76e0
0000000000000282 ffff88006d323568 ffff88006d323540 ffff880064d9c448
Call Trace:
[<ffffffff86dff497>] schedule+0x97/0x1c0 kernel/sched/core.c:3414
[< inline >] fanotify_get_response fs/notify/fanotify/fanotify.c:70
[<ffffffff8194a347>] fanotify_handle_event+0x537/0x830
fs/notify/fanotify/fanotify.c:233
[< inline >] send_to_group fs/notify/fsnotify.c:179
[<ffffffff8193fc8d>] fsnotify+0x73d/0x1020 fs/notify/fsnotify.c:275
[< inline >] fsnotify_perm ./include/linux/fsnotify.h:55
[<ffffffff82acefb1>] security_file_open+0x151/0x190 security/security.c:887
[<ffffffff8185a8db>] do_dentry_open+0x2ab/0xd30 fs/open.c:736
[<ffffffff8185e445>] vfs_open+0x105/0x220 fs/open.c:860
[< inline >] do_last fs/namei.c:3374
[<ffffffff81895909>] path_openat+0x12f9/0x2ab0 fs/namei.c:3497
[<ffffffff8189a7ac>] do_filp_open+0x18c/0x250 fs/namei.c:3532
[<ffffffff81873528>] do_open_execat+0xe8/0x4d0 fs/exec.c:818
[<ffffffff8187a18f>] do_execveat_common.isra.35+0x71f/0x1d80 fs/exec.c:1679
[< inline >] do_execve fs/exec.c:1783
[< inline >] SYSC_execve fs/exec.c:1864
[<ffffffff8187c262>] SyS_execve+0x42/0x50 fs/exec.c:1859
[<ffffffff810088ff>] do_syscall_64+0x1df/0x640 arch/x86/entry/common.c:288
[<ffffffff86e107c3>] entry_SYSCALL64_slow_path+0x25/0x25
arch/x86/entry/entry_64.S:249


[ 467.548485] Showing all locks held in the system:
[ 467.548981] 2 locks held by bash/4044:
[ 467.549313] #0: (&tty->ldisc_sem){.+.+.+}, at:
[<ffffffff86e0f7f7>] ldsem_down_read+0x37/0x40
[ 467.550076] #1: (&ldata->atomic_read_lock){+.+...}, at:
[<ffffffff8322ec95>] n_tty_read+0x1e5/0x1860
[ 467.550897] 3 locks held by bash/4062:
[ 467.551210] #0: (sb_writers#5){.+.+.+}, at: [<ffffffff81863c04>]
vfs_write+0x3a4/0x4e0
[ 467.551923] #1: (rcu_read_lock){......}, at: [<ffffffff83245c30>]
__handle_sysrq+0x0/0x4d0
[ 467.552655] #2: (tasklist_lock){.+.+..}, at: [<ffffffff81496fc4>]
debug_show_all_locks+0x74/0x290
[ 467.553454] 2 locks held by stress/4470:
[ 467.553754] #0: (&sig->cred_guard_mutex){+.+.+.}, at:
[<ffffffff818799b3>] prepare_bprm_creds+0x53/0x110
[ 467.554590] #1: (&fsnotify_mark_srcu){......}, at:
[<ffffffff8193f71d>] fsnotify+0x1cd/0x1020

On 0f98f121e1670eaa2a2fbb675e07d6ba7f0e146f of linux-next.

[Peter Zijlstra]

On Thu, Sep 01, 2016 at 07:53:44PM +0200, Dmitry Vyukov wrote:
> Hello,
>
> The following program:
>
> https://gist.githubusercontent.com/dvyukov/0952eeac71069b46b3fe0e28bd1a02bf/raw/396b9dcce2636cecab1a4161c15d3f066e6ef639/gistfile1.txt
>
> if run in a parallel loop creates unkillable hanged processes:
>
> -bash-4.3# ps afxu | grep a.out
> root 4489 0.0 0.0 8868 340 ttyp1 S+ 17:19 0:00
> \_ grep a.out
> root 4423 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
> root 4424 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
> root 4425 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
> root 4470 0.0 0.1 7016 2316 ttyp0 D 17:16 0:00
> ./stress ./a.out
>
> This looks like a classical deadlock, but LOCKDEP is silent so
> +LOCKDEP maintainers.

fanotify_get_response() uses wait_event(), that's an asymmetric API and
therefore one cannot include it in block-on chains. That is, while it
blocks, it doesn't know who it blocks on etc..

So not quite enough information, what is bash/4044's stack trace?
Because that is holding the fanotify srcu reference.

But it looks like a scenario where everyone is waiting for SRCU to
complete, while the task holding up SRCU completion is waiting for
something else.

But I'm not at all familiar with fanotify.

[Dmitry Vyukov]

On Thu, Sep 1, 2016 at 8:04 PM, Peter Zijlstra <pet...@infradead.org> wrote:
> On Thu, Sep 01, 2016 at 07:53:44PM +0200, Dmitry Vyukov wrote:
>> Hello,
>>
>> The following program:
>>
>> https://gist.githubusercontent.com/dvyukov/0952eeac71069b46b3fe0e28bd1a02bf/raw/396b9dcce2636cecab1a4161c15d3f066e6ef639/gistfile1.txt
>>
>> if run in a parallel loop creates unkillable hanged processes:
>>
>> -bash-4.3# ps afxu | grep a.out
>> root 4489 0.0 0.0 8868 340 ttyp1 S+ 17:19 0:00
>> \_ grep a.out
>> root 4423 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
>> root 4424 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
>> root 4425 0.0 0.0 0 0 ttyp0 D 17:16 0:00 [a.out]
>> root 4470 0.0 0.1 7016 2316 ttyp0 D 17:16 0:00
>> ./stress ./a.out
>>
>> This looks like a classical deadlock, but LOCKDEP is silent so
>> +LOCKDEP maintainers.
>
> fanotify_get_response() uses wait_event(), that's an asymmetric API and
> therefore one cannot include it in block-on chains. That is, while it
> blocks, it doesn't know who it blocks on etc..

Ah, I see. So that's deadlock on semaphores, that's not easily detectable.

4044 just holds a normal rcu_read_lock lock during sysrq processing.
I've sent the sysrq after the deadlock happened.
The bad guy is 4470, he holds fsnotify_mark_srcu.