Trouble increasing md component size

[Chris Webb <chris@arachsys.com>]

Hi. I'm having some trouble increasing the component size of arrays to match
changes in the underlying block devices. I'm running mdadm 2.6.7 against a
stock 2.6.24.4 kernel.

I'm seeing the problem in a cluster management application I'm writing, but I
can reproduce it simply using two LV block devices, 150M each:

  2# blockdev --getsz /dev/disk.4/{one,two}
  311296
  311296

I build an array that only fills two thirds of the component disks:

  2# mdadm --create /dev/md4 --level=1 --raid-disks=2 --size 102400 \
           --metadata=1.2 /dev/disk.4/{one,two}
  mdadm: largest drive (/dev/disk.4/one) exceed size (102400K) by more than 1%
  Continue creating array? y
  mdadm: array /dev/md4 started.
  2# blockdev --getsz /dev/md4
  204800

When the array is fully synced and up [UU], I try to grow the array to use
more of the component disks:

  2# mdadm --grow /dev/md4 --size max
  2# blockdev --getsz /dev/md4
  311272

and so this has worked fine.

However, if I increase the size of the underlying block devices:

  2# lvm lvresize -L 200M /dev/disk.4/one
    Extending logical volume one to 200.00 MB
    Logical volume one successfully resized
  2# lvm lvresize -L 200M /dev/disk.4/two
    Extending logical volume two to 200.00 MB
    Logical volume two successfully resized
  2# blockdev --getsz /dev/disk.4/{one,two}
  409600
  409600

and then try to grow the array:

  2# mdadm --grow /dev/md4 --size max
  2# blockdev --getsz /dev/md4
  311272
  2# mdadm --grow /dev/md4 --size 155637 just 1kB more
  mdadm: Cannot set device size for /dev/md4: No space left on device

it fails. Somehow, the change in the underlying component device sizes doesn't
seem to be noticed even though the kernel's idea (as returned by blockdev/ioctl
BLKGETSIZE[64]) has correctly changed. I can even remove and re-add both drives
in turn without it noticing the size change and allowing me to grow the array.

Trying to do something like

  echo 155637 > /sys/block/md4/md/component_size

doesn't work either. It'll accept values up as far as the size of the block
device when the array was created, but no larger.

The only recipe I've found which works is, for each drive in turn, to fail it,
remove it, zero the superblock and then re-add it. This seems to mean the md
driver sees is as a new component (e.g. different desc_nr in /proc/mdstat) and
therefore re-reads its size? This isn't at all friendly to the raid array,
though. While each device is out, any redundancy has gone. Is there some less
drastic way to persuade the array to grow?

Cheers,

Chris.