More information with regard to RAID 1 usage and
XFS. It appears that xfs_alloc_lookup is trying
to copy a NULL reference (0x0...) during a
sys_write call. I seem to only be able to reproduce
this for a RAID 1 (not for RAID 0 or RAID 5)
using UDMA IDE drives.
kdb started and the stack traced down to a NULL
reference for xfs_alloc_lookup. It does not
happen every time, but it is very repeatable as
samba (via smbtorture) tests (via samba dbench)
cause the oops.
A different kdb oops happened from:
_page_buf_page_apply <- kio_cluster_write
and a NULL reference. The one was prefaced
by a kernel message of a 0-order allocation
error (page_buf.c 1369 error).
I applied Rajagopal Ananthanarayanan's
prepare_write.patch to check its effect, but I
have seen no change. The result is the same with
the kernel reporting:
__alloc_pages: 0-order allocation failed.
In addition, the 2.4.2-pre3 patch with the kernel
CVS tree up to date.
From: Steve Lord
To: Joakim Bodin
Cc: Steve Lord; Linux-XFS Mailing List
Sent: 2/14/01 1:40 PM
Subject: Re: For those people with IDE problems with XFS
Joakim Bodin wrote:
> Steve Lord wrote:
> > If you are using the development tree and ide is blowing chunks for
> > can you try the following patch, it removes some of the code added
> > the kio patch, and backports some changes from 2.4.2-pre3 which fix
> > problems. No guarantees on this one but for me it appears to be
> > here.
> > I think you can expect to see 2.4.2-pre3 in a day or so, it may be a
> > 2.4.2 by then.
> > Steve
> I just want to report that this seems to fix all the lockup problems
> I've been having since xfs cvs moved to 2.4.1. It survives big
> and compiles that locked it up previously. This is a P2-300, 128MB RAM
> with ide disk using udma33.
> Joakim Bodin
Thanks for the feedback,
We are fixing up some issues with kiobufs and multmode ide, once this is
done I will either push this code into the 2.4.1 tree, or bring the tree
up to 2.4.2-pre3. I was hoping to see a 2.4.2 come out by now and base
tree on that but we are still waiting.