Re: MD RAID1 deadlock on failed disk

[Hubert Tonneau <hubert.tonneau@fullpliant.org>]

Hi,

The configuration is:
Perc H200 controller configured with no RAID (mpt2sas driver),
2 SATA disks (sda and sdb),
Linux MD Sofware RAID1 (md0),
stock Linux 2.6.35.7 kernel.

I hotunplug the second (sdb) disk, and the result is:
. as expected, I can read sda device,
. as expected, any read to sdb device fails,
. unexpectedly, any read to md0 never returns.

No oops or thing like that in the kernel log.
I did not try the same with other kernel releases.

2.6.32.24 kernel worked fine.

Neil Brown asked for /proc/sysrq-trigger ouput,
and concluded that the problem is related to 'fw_event0'.
See his answer bellow.

Regards,
Hubert Tonneau


Neil Brown wrote:
>
> The fw_event0 process is interesting.
> It seems to be hung trying to 'sync' the drive that has just been pulled.
> If that is somehow causing some IO request from the md/raid1 to be delayed
> then that would certainly hang the array.
> 
> There is a section in the middle of the trace which is missing - presumably
> the sysrq-trigger output overflowed a buffer - that isn't uncommon.
> 
> So I cannot see all the timing clearly.
> How long after pulling the drive was this trace taken?
> 
> I suspect that you need to post this to linux-scsi@vger.kernel.org
> and ask about that fw_event0 thread - whether that should happen, whether it
> has been fixed, and whether it could delay pending IO requests.
> 
> NeilBrown