Re: [PATCH] mdadm/systemd: remove KillMode=none from service file

[Mariusz Tkaczyk <mariusz.tkaczyk@linux.intel.com>]

On Tue, 15 Feb 2022 21:34:15 +0800
Coly Li <colyli@suse.de> wrote:

> For mdadm's systemd configuration, current systemd KillMode is "none" in
> following service files,
> - mdadm-grow-continue@.service
> - mdmon@.service
> 
> This "none" mode is strongly againsted by systemd developers (see man 5
> systemd.kill for "KillMode=" section), and is considering to remove in
> future systemd version.
> 
> As systemd developer explained in disuccsion, the systemd kill process
> is,
> 1. send the signal specified by KillSignal= to the list of processes (if
>    any), TERM is the default
> 2. wait until either the target of process(es) exit or a timeout expires
> 3. if the timeout expires send the signal specified by FinalKillSignal=,
>    KILL is the default
> 
> For "control-group", all remaining processes will receive the SIGTERM
> signal (by default) and if there are still processes after a period f
> time, they will get the SIGKILL signal.
> 
> For "mixed", only the main process will receive the SIGTERM signal, and
> if there are still processes after a period of time, all remaining
> processes (including the main one) will receive the SIGKILL signal.
> 
> From the above comment, currently KillMode=control-group is a proper
> kill mode. Since control-gropu is the default kill mode, the fix can be
> simply removing KillMode=none line from the service file, then the
> default mode will take effect.

Hi All,
We are experiencing issues with IMSM metadata on RHEL8.7 and 9.1 (the patch
was picked by Redhat). There are several issues which results in hang task,
characteristic to missing mdmon:

[  619.521440] task:umount state:D stack: 0 pid: 6285 ppid: flags:0x00004084
[  619.534033] Call Trace:
[  619.539980]  __schedule+0x2d1/0x830
[  619.547056]  ? finish_wait+0x80/0x80
[  619.554261]  schedule+0x35/0xa0
[  619.560999]  md_write_start+0x14b/0x220
[  619.568492]  ? finish_wait+0x80/0x80
[  619.575649]  raid1_make_request+0x3c/0x90 [raid1]
[  619.584111]  md_handle_request+0x128/0x1b0
[  619.591891]  md_make_request+0x5b/0xb0
[  619.599235]  generic_make_request_no_check+0x202/0x330
[  619.608185]  submit_bio+0x3c/0x160
[  619.615161]  ? bio_add_page+0x42/0x50
[  619.622413]  submit_bh_wbc+0x16a/0x190
[  619.629713]  jbd2_write_superblock+0xf4/0x210 [jbd2]
[  619.638340]  jbd2_journal_update_sb_log_tail+0x65/0xc0 [jbd2]
[  619.647773]  __jbd2_update_log_tail+0x3f/0x100 [jbd2]
[  619.656374]  jbd2_cleanup_journal_tail+0x50/0x90 [jbd2]
[  619.665107]  jbd2_log_do_checkpoint+0xfa/0x400 [jbd2]
[  619.673572]  ? prepare_to_wait_event+0xa0/0x180
[  619.681344]  jbd2_journal_destroy+0x120/0x2a0 [jbd2]
[  619.689551]  ? finish_wait+0x80/0x80
[  619.696096]  ext4_put_super+0x76/0x390 [ext4]
[  619.703584]  generic_shutdown_super+0x6c/0x100
[  619.711065]  kill_block_super+0x21/0x50
[  619.717809]  deactivate_locked_super+0x34/0x70
[  619.725146]  cleanup_mnt+0x3b/0x70
[  619.731279]  task_work_run+0x8a/0xb0
[  619.737576]  exit_to_usermode_loop+0xeb/0xf0
[  619.744657]  do_syscall_64+0x198/0x1a0
[  619.751155]  entry_SYSCALL_64_after_hwframe+0x65/0xca

It can be reproduced by mounting LVM created on IMSM RAID1 array and then
reboot. I verified that reverting the patch fixes the issue.

I understand that from systemd perspective the behavior in not wanted, but
this is exactly what we need, to have working mdmon process even if systemd was
stopped. KillMode=none does the job.
I searched for alternative way to prevent systemd from stopping the mdmon unit
but I failed. I tried to change signals, so I configured unit to send SIGPIPE
(because it is ignored by mdmon)- it worked but later system hanged because
mdmon unit cannot be stopped.

I also tried to configure mdmon unit to be stopped after umount.target and I
failed too. It cannot be achieved by setting After= or Before=. The one
objection I have here is that systemd-shutdown tries to stop raid arrays later,
so it could be better to have running mdmon there.

IMO KillMode=none is desired in this case. Later, mdmon is restarted in dracut
by mdraid module.

If there is no other solution for the problem, I will need to ask Jes to revert
this patch. For now, I asked Redhat to do it.
Do you have any suggestions?

TIA,
Mariusz