Bug#416512: Failures with hotplugging

[Goswin von Brederlow]

Package: mdadm
Version: 2.5.6-9
Severity: normal

Hi,

I have problems with the mdadm and software raid behaviour in Linux.

I have a 2.6.20.4 kernel with a Marvell SATA controler with the extra
marvell driver (mv_sata not sata_mv) for hotplug support. I have 2
disks (sda1+sdb1) in a software raid1 connected to it.

The problem comes when I hot-remove a disk (sdb) from the
hardware. The marvell controler sees the removal and removed the disk
from the system. The scsi layer gives a cache flush failure on sdb
(expected), sysfs removes /sys/block/sdb/(sdb1/) and udev removes
/dev/sdb(1). So far so good.

The raid does no mark /dev/sdb1 as failed or even removed. Only on the
next access it gets an I/O error and marks it.

'mdadm --fail /dev/md0 /dev/sdb1' tries to stat /dev/sdb1 [Manage.c:
Manage_subdevs(): 189] and gives an error. Same for --remove. The
device can not be removed without first manually recreating the device
node. Note that /sys/block/md0/md/dev-sdb1/block is a broken symlink
at this point and /sys/block/md0/md/dev-sdb1/block/dev can't be used
to find the major/minor of the device.

Reinserting the disk will cause a new device minor to be allocated and
the disk reappears as /dev/sdc. Given that the software raid still
things sdb1 is valid (unless written too) this is a good thing. It
shows that somewhere in the kernel it still knows sdb is in use.


mdadm should be able to fail/remove an unplugged disk but I don't know
where mdadm should get the device infos for it at this point.

And it would be nice if the kernel would tell the software raid that
the device was removed and set it failed automatically. But that
probably needs another bug.

MfG
	Goswin

-- System Information:
Debian Release: 4.0
  APT prefers unstable
  APT policy: (500, 'unstable')
Architecture: amd64 (x86_64)
Shell:  /bin/sh linked to /bin/bash
Kernel: Linux 2.6.20.4
Locale: LANG=C, LC_CTYPE=C (charmap=ANSI_X3.4-1968)