Re: [PATCH v3] jbd2: Fix block tag checksum verification brokenness

[adding tytso and linux-ext4 to cc]

On Mon, Jun 03, 2013 at 12:55:21AM -0700, Darrick J. Wong wrote:

On Sun, Jun 02, 2013 at 10:16:30PM -0400, George Spelvin wrote:

quoted

quoted

Fortunately metadata checksumming is still "experimental" and not in a shipping
e2fsprogs, so there should be few users affected by this.

Am I reading this patch correctly that this changes which half of
the 32-bit checksum is stored on little-endian (e.g. x86) machines?
Before, it stored the low 16 bits of cpu_to_be32(csum), meaning the
high 16 bits of csum.

...unless you're using a BE system in which case it's the low 16 bits. :)

quoted

Now, it's cpu_to_be16(csum32), which is the low 16 bits of csum32.

It so happens that I have multiple metadata_csum file systems.
(I enabled it a while ago on a machine where I wasn't sure if corruption
was RAM or disk, and have been using it on SSE4.2 machines since.)

Is there an upgrade path?  Also, what's the corresponding e2fsprogs
commit that supports this change?

The kernel patch fixes journal bogosity when moving a disk between big and
little endian systems.  e2fsprogs didn't have the brokenness, so there's no

Drat, I was writing this email too late at night.  There /is/ a related
e2fsprogs patch for this:
http://comments.gmane.org/gmane.comp.file-systems.ext4/38621

Hey Ted, any thoughts on this jbd2/e2fsprogs patch pair?  Al acked the jbd2
part.

--D

change needed.  In theory you'd only hit this if you happened to crash an x86
box with an ext4 fs, move the disk to a ppc, and try to recover it there.

The upgrade path is to umount cleanly and reboot with a patched kernel while
hoping that you don't crash while rebooting.  Sorry about the bumpy
metadata_csum ride.

--D