Excerpts from Peter Stuge's message of 2011-03-10 08:45:09 -0500:

Chris Mason wrote:

quoted

Which tool and which version of the tool did you use to delete the
partition?

fdisk from util-linux-2.18

Straight from util-linux, or with distro patches?

The non-working partition was deleted and the current one created
with fdisk from util-linux-2.14.2.

quoted

quoted

It's a 64GB CF card with two partitions; one 40MB ext2 and "the rest"
is btrfs. This is the current fdisk output:

Ok, going back to your original email, the block you're failing on
is probably right in the middle of the drive.

Right.

quoted

We can't be sure without looking at the mapping tree (which we
don't have),

Could we guess at where it was put?

Until you do something funky like balance the drive, there's a 1:1
mapping.  The easy way to guess is to strace btrfsck and see where it
read.

quoted

quoted

I say current, because by now I have changed the sdb2 partition
twice.

Have you ever changed the start of the partition?

No.

quoted

If the start had changed the superblock should be in the wrong
place, so the mount wouldn't have gotten this far.

Right.

quoted

quoted

In any case changing the partition table shouldn't affect the
filesystem, right? Also, I changed the partition with the filesystem
mounted, so the kernel did not start using the new partition table.

I'd have to repeat the test on this flash card to say for sure.
Deleting then recreating the partition with the FS mounted isn't
very high up on the list of things that get tested often, so my
guess is that's where the problem is.

As I understand it, fdisk writes the new partition table to disk, and
asks the kernel to re-read it from there. That ioctl failed, I expect
because the filesystem was mounted.

quoted

Right, we've got a block full with zeros where they don't belong. 
Can you run dump the block contents with gdb please?

Will do!

quoted

Ok, we talked about power offs and barriers in a different thread,
but I didn't realize you were on a CF device.  I'd want to do some
tests on this device to see how well it really reacted in power
offs, but lets do that after we pull your data some where safer.

I'm of course happy to help test anything!

Great, thanks.

-chris