Hi,
On Sat, Nov 23, 2002 at 08:18:26PM +0000, Jeff Schaller wrote:
> Boy, it'd help if I actually attached the log, eh?
Yes. :)
Content-Description: kern.log
> Nov 19 06:58:47 debian kernel: attempt to access beyond end of device
> Nov 19 06:58:47 debian kernel: 09:00: rw=1, want=39121480, limit=39121408
Corruption on disk. Undiagnosable without more info --- it could be
hardware or software.
> Nov 19 06:58:47 debian kernel: Assertion failure in __journal_remove_journal_head() at journal.c:1732: "buffer_jbd(bh)"
> Nov 19 06:58:47 debian kernel: kernel BUG at journal.c:1732!
That's a core driver layer bug which we found recently, but it's
too close to 2.4.20 to include the fix. Basically, on an
out-of-bounds IO the ll_rw_block code was clearing most of the bits in
the buffer_head state, leading to the above assert failure when ext3
found its critical metadata had been corrupted.
Patch is below, I'll send it to Marcelo for 2.4.21-pre.
--Stephen
--- linux-2.4-ext3merge/drivers/block/ll_rw_blk.c.=K0026=.orig Mon Nov 25 15:03:18 2002
+++ linux-2.4-ext3merge/drivers/block/ll_rw_blk.c Mon Nov 25 15:04:00 2002
@@ -1129,7 +1129,7 @@
if (maxsector < count || maxsector - count < sector) {
/* Yecch */
- bh->b_state &= (1 << BH_Lock) | (1 << BH_Mapped);
+ bh->b_state &= ~(1 << BH_Dirty);
/* This may well happen - the kernel calls bread()
without checking the size of the device, e.g.,
@@ -1140,7 +1140,6 @@
kdevname(bh->b_rdev), rw,
(sector + count)>>1, minorsize);
- /* Yecch again */
bh->b_end_io(bh, 0);
return;
}
