Re: [RFC PATCH 0/4 v2] Define killable version for access_remote_vm() and use it in fs/proc

[Yang Shi]

On 3/6/18 12:45 PM, Andrew Morton wrote:
> On Tue, 27 Feb 2018 08:25:47 +0800 Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote:
>
> > Background:
> > When running vm-scalability with large memory (> 300GB), the below hung
> > task issue happens occasionally.
> >
> > INFO: task ps:14018 blocked for more than 120 seconds.
> >         Tainted: G            E 4.9.79-009.ali3000.alios7.x86_64 #1
> >   "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >   ps              D    0 14018      1 0x00000004
> >    ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
> >    ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
> >    00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
> >   Call Trace:
> >    [<ffffffff817154d0>] ? __schedule+0x250/0x730
> >    [<ffffffff817159e6>] schedule+0x36/0x80
> >    [<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
> >    [<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
> >    [<ffffffff81717db0>] down_read+0x20/0x40
> >    [<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
> >    [<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
> >    [<ffffffff81241d87>] __vfs_read+0x37/0x150
> >    [<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
> >    [<ffffffff81242266>] vfs_read+0x96/0x130
> >    [<ffffffff812437b5>] SyS_read+0x55/0xc0
> >    [<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5
> >
> > When manipulating a large mapping, the process may hold the mmap_sem for
> > long time, so reading /proc/<pid>/cmdline may be blocked in
> > uninterruptible state for long time.
> > We already have killable version APIs for semaphore, here use down_read_killable()
> > to improve the responsiveness.
> Maybe I'm missing something, but I don't see how this solves the
> problem.  Yes, the read of /proc/pid/cmdline will be abandoned if
> someone interrupts that process.  But if nobody does that, the read
> will still just sit there for 2 minutes and the watchdog warning will
> still come out?

No, the warning will not come out since down_read_killable() puts the 
task into TASK_KILLABLE state instead of TASK_UNINTERRUPTIBLE state. The 
hung task check will skip TASK_KILLABLE tasks, please see the below code 
in (kernel/hung_task.c):

                /* use "==" to skip the TASK_KILLABLE tasks waiting on 
NFS */
                if (t->state == TASK_UNINTERRUPTIBLE)
                        check_hung_task(t, timeout);

It just mitigates the hung task warning, can't resolve the mmap_sem 
scalability issue. Furthermore, waiting on pure uninterruptible state 
for reading /proc sounds unnecessary. It doesn't wait for I/O completion.

> Where the heck are we holding mmap_sem for so long?  Can that be fixed?

The mmap_sem is held for unmapping a large map which has every single 
page mapped. This is not a issue in real production code. Just found it 
by running vm-scalability on a machine with ~600GB memory.

AFAIK, I don't see any easy fix for the mmap_sem scalability issue. I 
saw range locking patches (https://lwn.net/Articles/723648/) were 
floating around. But, it may not help too much on the case that a large 
map with every single page mapped.

Thanks,
Yang