6

This is a repost of a question I asked on ServerFault and was automatically deleted for lack of activity. Let's hope it will have better luck here.

After performing a "check" (by writing check to /sys/class/block/md0/md/sync_action) of a md software RAID array with redundancy on Linux, the md subsystem will update the /sys/class/block/md0/md/mismatch_cnt virtual file with the number of sectors comprised in chunks that have a mismatch.

Unfortunately, it doesn't tell which chunk or sector they are.

On a RAID 1 array with 2 drives (for instance md0 on sda and sdb), one can do:

blockdev --flushbufs /dev/md0 /dev/sda /dev/sdb
cmp -l /dev/sda /dev/sdb > sda-sdb.diff

To find out which bytes differ, and then derive the sector and additional information (like which file is affected if there's a filesystem there).

But is there a way to get the same information for other RAID levels? For instance, with this RAID10 array:

$ cat /sys/class/block/md0/md/mismatch_cnt
4608
$ cat /proc/mdstat
[...]
md0 : active raid10 sdd1[4] sdc1[3] sdb1[1] sda1[0]
      3906763776 blocks super 1.2 512K chunks 2 near-copies [4/4] [UUUU]

How would I know where the mismatches are on /dev/md0?

In this particular instance, the algorithm is not too complicated: we can use mdadm -E to find out where the data starts in the devices, and then sda1 is meant to be the same as sdb1 and sdc1 the same as sdd1, the first chunk of /dev/md0 is found in the first chunk of both sda1 and sda2, the second in the first chunk of both sdc1 and sdd1, so given the offset of a byte difference in a pair of devices, it's not two difficult to find out the corresponding offset in the md device, but I was hoping there was a way to avoid have to run down the algorithm manually

Community
  • 1
Stéphane Chazelas
  • 522,931
  • 91
  • 1,010
  • 1,501
  • this is in `man dmraid`: *`-r [devpath...]`: List all RAID devices ... sectors used and data offset into the device.... If `-D` is added to `-r` the RAID metadata gets dumped into a subdirectory named `dmraid.format_name` in files named devicename.dat. The byte offset where the metadata is located on the device is written into files named `devicename.offset`and the size of the device in sectors into files named device‐name.size.* I would suspect if it is useful it is only with that `-D` switch. That's a wild guess, though. – mikeserv Jan 20 '16 at 12:50
  • 1
    I added a line to my kernel to log the sector number during a RAID5 `check`, but now I'm not sure what to do with it: http://unix.stackexchange.com/questions/266432/xfs-on-md-raid5-how-to-find-what. RAID1/RAID10 mismatches can be spurious (esp. with swap on RAID), but you could add a similar `printk` to wherever the `atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);` happens for RAID1/RAID10. – Peter Cordes Feb 29 '16 at 05:01

0 Answers0