Re: constant array_state active after specific jobs

On Fri, Mar 24 2017, pdi wrote:

On Fri, 24 Mar 2017 16:25:35 +1100
NeilBrown [off-list ref] wrote:

quoted

On Thu, Mar 23 2017, pdi wrote:

quoted

Greetings all,

The problem in a nutshell is that an array is clean after boot,
until some specific jobs switch it to active where it remains until
reboot.

A similar problem was discussed, and solved, in 
https://www.spinics.net/lists/raid/msg46450.html. However, AFAICT,
it is not the same issue.

I would be grateful for any insights as to why this happens and/or
how to prevent it.

The relevant info follows, please let me know if anything further
might help.

Many thanks in advance.

- uname -a
  Linux hostname 4.4.38 #1 SMP Sun Dec 11 16:03:41 CST 2016 x86_64
  Intel(R) Core(TM)2 Duo CPU E8400 @ 3.00GHz GenuineIntel GNU/Linux
- mdadm -V
  mdadm - v3.3.4 - 3rd August 2015
- Desktop drives without sct/erc,
  with timeout mismatch correction as per
  https://raid.wiki.kernel.org/index.php/Timeout_Mismatch
- /dev/md9 is a raid10 array, 4 devices, far=2,
  with various dirs used as samba and nfs shares
- The array is in *constant* array_state active
- mdadm -D /dev/md9 | grep 'State :'
  State : active
- cat /sys/block/md9/md/array_state
  active
- watch -d 'grep md9 /proc/diskstats'
  remain unchanged
- uptime
  load average: 0.00, 0.00, 0.00
- cat /sys/block/md9/md/safe_mode_delay
  0.201
- echo 0.1 > /sys/block/md9/md/safe_mode_delay
  array_state remains active
- echo clean > /sys/block/md9/md/array_state
  echo: write error: Device or resource busy
- reboot (with or without prior check)
  array_state clean
- After reboot, array remains clean until some specific
  jobs put it in constant active state. Such jobs so far
  identified:
  - echo check > /sys/block/md9/md/sync_action
  - run an rsnapshot job
  - start a qemu/kvm vm
- Other jobs, like text/doc editing, multimedia playback,
  etc retain array_state clean  

This bug was introduced by
Commit: 20d0189b1012 ("block: Introduce new bio_split()")
in 3.14, and fixed by
Commit: 9b622e2bbcf0 ("raid10: increment write counter after bio is
split") in 4.8.

Maybe the latter patch should be sent to -stable ??

NeilBrown

NeilBrown, thank you for your swift and concise answer.

I gather you are referring to kernel version numbers. The described
behaviour was first noticed many months ago with kernel 2.6.37.6, and
persisted after a system upgrade and kernel 4.4.38. However, after the
upgrade two things were corrected, the timeout mismatch, and a
Current_Pending_Sector in one of the drives; which may, or may not,
explain the occurrence with the older kernel.

Is this constant active state in the data array something to worry about
and try kernel >= 4.8, or shall I let be?

The only important consequence of the constant active state is that if
your machine crashes at a moment when the array would otherwise have
been idle, then a resync will be needed after reboot.  Without the
constant active state, that resync would not have been needed.

If you have a write-intent bitmap, this is not particularly relevant.

I cannot say how important it is to you to avoid a resync after a crash,
so I don't know if you should just let it be or not.

NeilBrown