raid10: 6 out of 8 disks marked as stale on every restart

Hello,

I have a 8 disk RAID10 array, 6 of the disks are on an LSISAS2008 controller and 2 are on a 82801JI (ICH10 Family) SATA AHCI controller. I upgraded the kernel from 3.17.1 to 3.17.6 when the issue I am having started to occur, but reverting to an older kernel does not resolve the issue.

Restarting the machine causes the array not to start (or be visible in /proc/mdstat or any mention in kernel messages.) If I try to assemble the drives, mdraid complains that 6 out of the 8 disks (coincidentally all on the LSISAS2008 controller) are non-fresh:

root@kvm:~# mdadm --assemble /dev/md3 /dev/sde /dev/sdf /dev/sdg /dev/sdh /dev/sdi /dev/sdj /dev/sda /dev/sdb

Dec 11 21:08:25 kvm kernel: [  528.503736] md: kicking non-fresh sdi from array!
Dec 11 21:08:25 kvm kernel: [  528.503747] md: unbind<sdi>
Dec 11 21:08:25 kvm kernel: [  528.523775] md: export_rdev(sdi)
Dec 11 21:08:25 kvm kernel: [  528.523802] md: kicking non-fresh sdg from array!
Dec 11 21:08:25 kvm kernel: [  528.523809] md: unbind<sdg>
Dec 11 21:08:25 kvm kernel: [  528.531753] md: export_rdev(sdg)
Dec 11 21:08:25 kvm kernel: [  528.531780] md: kicking non-fresh sdf from array!
Dec 11 21:08:25 kvm kernel: [  528.531788] md: unbind<sdf>
Dec 11 21:08:25 kvm kernel: [  528.539749] md: export_rdev(sdf)
Dec 11 21:08:25 kvm kernel: [  528.539776] md: kicking non-fresh sdh from array!
Dec 11 21:08:25 kvm kernel: [  528.539785] md: unbind<sdh>
Dec 11 21:08:25 kvm kernel: [  528.547744] md: export_rdev(sdh)
Dec 11 21:08:25 kvm kernel: [  528.547771] md: kicking non-fresh sdj from array!
Dec 11 21:08:25 kvm kernel: [  528.547779] md: unbind<sdj>
Dec 11 21:08:25 kvm kernel: [  528.555755] md: export_rdev(sdj)
Dec 11 21:08:25 kvm kernel: [  528.555782] md: kicking non-fresh sde from array!
Dec 11 21:08:25 kvm kernel: [  528.555790] md: unbind<sde>
Dec 11 21:08:25 kvm kernel: [  528.563758] md: export_rdev(sde)
Dec 11 21:08:25 kvm kernel: [  528.565831] md/raid10:md3: not enough operational mirrors.
Dec 11 21:08:25 kvm kernel: [  528.567230] md: pers->run() failed ...

/dev/sda and /dev/sdb are the only drives not on the LSI controller. If I force the assembly with 6 out of the 8 drives the RAID array comes up:

root@kvm:~# mdadm --assemble /dev/md3 /dev/sde /dev/sdf /dev/sdg /dev/sdh /dev/sdi /dev/sdj --run

Then I add the extra drives:

root@kvm:~# mdadm --manage /dev/md3 --add /dev/sda
root@kvm:~# mdadm --manage /dev/md3 --add /dev/sdb

root@kvm:~# mdadm --detail /dev/md3
/dev/md3:
        Version : 1.0
  Creation Time : Thu Sep 12 18:43:56 2013
     Raid Level : raid10
     Array Size : 7814055936 (7452.06 GiB 8001.59 GB)
  Used Dev Size : 1953513984 (1863.02 GiB 2000.40 GB)
   Raid Devices : 8
  Total Devices : 8
    Persistence : Superblock is persistent

Update Time : Fri Dec 12 08:58:19 2014
State : active, degraded, recovering
Active Devices : 6
Working Devices : 8
Failed Devices : 0
Spare Devices : 2

Layout : near=2
Chunk Size : 512K

Rebuild Status : 76% complete

Name : kvm.taylor.kieser.ca:3
UUID : f0bc8469:9879a709:e4cc94a7:521bd273
Events : 82901

Number Major Minor RaidDevice State
0 8 128 0 active sync /dev/sdi
8 8 96 1 active sync /dev/sdg
11 8 0 2 spare rebuilding /dev/sda
3 8 112 3 active sync /dev/sdh
4 0 0 4 removed
10 8 80 5 active sync /dev/sdf
6 8 64 6 active sync /dev/sde
9 8 144 7 active sync /dev/sdj

12 8 16 - spare /dev/sdb

This occurs every time I restart the machine. Thoughts? I tried rebuilding the initramfs but this didn't resolve the issue. I'm also running bcache on this machine, but on top of the mdraid.

/etc/mdadm.conf:

# definitions of existing MD arrays
ARRAY /dev/md/0 metadata=1.0 UUID=3b174514:49f3e22e:550cf9a7:8ed93920 name=linux:0
ARRAY /dev/md/1 metadata=1.0 UUID=8e23f81d:73f9b393:addd1f7f:5ee1833a name=linux:1
ARRAY /dev/md/2 metadata=1.0 UUID=cc5a0495:b5262855:fb3cd40a:8b237162 name=kvm.taylor.kieser.ca:2
ARRAY /dev/md/3 metadata=1.0 UUID=f0bc8469:9879a709:e4cc94a7:521bd273 name=kvm.taylor.kieser.ca:3

root@kvm:~# uname -a
Linux kvm 3.17.6 #3 SMP Sun Dec 7 12:16:45 PST 2014 x86_64 x86_64 x86_64 GNU/Linux

root@kvm:~# mdadm -V
mdadm - v3.2.5 - 18th May 2012

root@kvm:~# cat /proc/mdstat
Personalities : [raid1] [linear] [multipath] [raid0] [raid6] [raid5] [raid4] [raid10]
md127 : inactive sdk[2](S)
      1465138448 blocks super 1.0

md3 : active raid10 sdb[12](S) sda[11] sdi[0] sdj[9] sde[6] sdf[10] sdh[3] sdg[8]

      7814055936 blocks super 1.0 512K chunks 2 near-copies [8/6] 
[UU_U_UUU]
      [===============>.....]  recovery = 76.6% (1498279040/1953513984) 

finish=4710.1min speed=1610K/sec

md1 : active raid1 sdd5[3] sdc5[2]

      25164672 blocks super 1.0 [2/2] [UU]

md0 : active raid1 sdd1[3] sdc1[2]

      16779136 blocks super 1.0 [2/2] [UU]

md2 : active raid1 sdd6[3] sdc6[2]

      192472960 blocks super 1.0 [2/2] [UU]

unused devices: <none>

-Peter

Curious.

What does "mdadm --examine" report for each device immediately after boot, before you try assembling anything?

Maybe also get the output just before you shut down to compare.

NeilBrown

Sadly, I had to disassemble this array and get something workable that didn't become unassembled on every restart. I did try sticking all drives on the same AHCI controller, which caused the array to be resynced on every restart instead of disassembled.

I suspect this issue is caused by bcache (I'm using the md as a backing device for bcache.) The bcache maintainer states "get the md people or someone to explain _what_ they want whatever has their device open to do on reboot."

I'm going to setup a test environment and see if I can reproduce it again.

-Peter