I think I've stumbled upon a bug in DM-RAID. The primary symptom is that whenOK, I've done some more experimentation, and have figured out that adjusting the sync region size from the default (and thus adjusting the bitmap size) can temporarily work around this. If I adjust things so that the bitmap is less than 32 pages, then everything works fine, until I try to reboot, at which point the device either (in order of decreasing probability):
creating a new DM-RAID based device (using either LVM or dmsetup) in a RAID1
configuration, it very quickly claims one by one that all of the disks failed
except the first, and goes degraded. When this happens on a given system, the
disks always 'fail' in the reverse of the order of the mirror numbers. All of
the other RAID profiles work just fine. Curiously, it also only seems to
happen for 'big' devices (I haven't been able to determine exactly what the
minimum size is, but I see it 100% of the time with 32G devices, never with 16G
ones, and only intermittently with 24G).
Here's what I got from dmesg when creating a 32G LVM volume that exhibited
this issue:
[66318.401295] device-mapper: raid: Superblocks created for new array
[66318.450452] md/raid1:mdX: active with 2 out of 2 mirrors
[66318.450467] Choosing daemon_sleep default (5 sec)
[66318.450482] created bitmap (32 pages) for device mdX
[66318.450495] attempt to access beyond end of device
[66318.450501] dm-91: rw=13329, want=0, limit=8192
[66318.450506] md: super_written gets error=-5, uptodate=0
[66318.450513] md/raid1:mdX: Disk failure on dm-92, disabling device.
md/raid1:mdX: Operation continuing on 1 devices.
[66318.459815] attempt to access beyond end of device
[66318.459819] dm-89: rw=13329, want=0, limit=8192
[66318.459822] md: super_written gets error=-5, uptodate=0
[66318.492852] attempt to access beyond end of device
[66318.492862] dm-89: rw=13329, want=0, limit=8192
[66318.492868] md: super_written gets error=-5, uptodate=0
[66318.627183] mdX: bitmap file is out of date, doing full recovery
[66318.714107] mdX: bitmap initialized from disk: read 3 pages, set 65536 of 65536 bits
[66318.782045] RAID1 conf printout:
[66318.782054] --- wd:1 rd:2
[66318.782061] disk 0, wo:0, o:1, dev:dm-90
[66318.782068] disk 1, wo:1, o:0, dev:dm-92
[66318.836598] RAID1 conf printout:
[66318.836607] --- wd:1 rd:2
[66318.836614] disk 0, wo:0, o:1, dev:dm-90
And here's output for a 24G LVM volume that didn't display the issue.
[66343.407954] device-mapper: raid: Superblocks created for new array
[66343.479065] md/raid1:mdX: active with 2 out of 2 mirrors
[66343.479078] Choosing daemon_sleep default (5 sec)
[66343.479101] created bitmap (24 pages) for device mdX
[66343.629329] mdX: bitmap file is out of date, doing full recovery
[66343.677374] mdX: bitmap initialized from disk: read 2 pages, set 49152 of 49152 bits
I'm using a lightly patched version of 4.2.3
(the source can be found at https://github.com/ferroin/linux)
but none of the patches I'm using come anywhere near anything in the block layer,
let alone the DM/MD code.
I've attempted to bisect this, although it got kind of complicated. So far I've
determined that the first commit that I see this issue on is d3b178a: md: Skip cluster setup for dm-raid
Prior to that commit, I can't initialize any dm-raid devices due to the bug it fixes.
I have not tested anything prior to d51e4fe (the merge commit that pulled in the md-cluster code),
but I do distinctly remember that I did not see this issue in 3.19.
I'll be happy to provide more info if needed.
Attachment:
smime.p7s
Description: S/MIME Cryptographic Signature