Re: [dm-devel] Re: Data corruption withraid5/dm-crypt/lvm/reiserfs on 126.96.36.199
From: Piet Delaney
Date: Thu Feb 22 2007 - 18:57:45 EST
On Mon, 2007-01-22 at 22:42 +0100, Christophe Saout wrote:
> Am Montag, den 22.01.2007, 11:56 -0800 schrieb Andrew Morton:
> > There has been a long history of similar problems when raid and dm-crypt
> > are used together. I thought a couple of months ago that we were hot on
> > the trail of a fix, but I don't think we ever got there. Perhaps
> > Christophe can comment?
> No, I think it's exactly this bug. Three month ago someone came up with
> a very reliable test case and I managed to nail down the bug.
> Readaheads that were aborted by the raid5 code (or some layer below)
> were signalled using a cleared BIO_UPTODATE bit, but no error code, and
> were missed as aborted by dm-crypt (all other layers apparently set the
> error code in this case, so this only happened with raid5) which could
> mess up the buffer cache.
> Anyway, it then turned out this bug was already "accidentally" fixed in
> 2.6.19 by RedHat in order to play nicely with make_request changes (the
> stuff to reduce stack usage with stacked block device layers), that's
> why you probably missed that it got fixed. The fix for pre-2.6.19
> kernels went into some 2.6.16.x and 188.8.131.52.
I've been trying Andrew's suggestion of doing fault injections,
currently just kmalloc() and mempool_alloc(), and just got a hang
on 2.6.12 that I'm poking around on with kgdb. I'm using dm-crypt
on a SCSI raid-1 (mirrored) root partition. I'm using your patch
that fixes the raid5 problem just to play it safe.
So far it looks like processes are waiting to be woken up by
the buffer cache once reads have completed.
> dm-devel mailing list
Piet Delaney Phone: (408) 200-5256
Blue Lane Technologies Fax: (408) 200-5299
10450 Bubb Rd.
Cupertino, Ca. 95014 Email: piet@xxxxxxxxxxxx
Description: This is a digitally signed message part