Bug: "corrupt leaf. slot offset bad": root subvolume unmountable, "btrfs check" crashes

[Andreas Reis <andreas.reis@gmail.com>]

Kernel 3.15.0-rc2, btrfs-progs 3.14.1

While doing some minor package updates my btrfs root partition [*] 
decided to corrupt itself. There was no system crash, although I had 
plenty of these (due to an USB-related regression) in recent weeks that 
resulted in no trouble.

First only one of a package's folders was corrupted, any access to files 
within (incl. attempts to delete) printed

btrfs: corrupt leaf, slot offset bad: block=842924032,root=1, slot=88

to dmesg (I'm actually not sure about the numbers, but that was indeed 
the error message). After moving the folder out of the way the partition 
continued to appear working as normal, one reboot also worked fine.

Now I can't boot at all (beyond loading the kernel image located on 
another partition), neither with 3,15-rc2 nor 3.14.1. Attempting to 
mount the __current/ROOT subvolume on ArchLinux's current Live-CD 
(kernel 3.13.7) prints

btrfs: device label Linux devid 1 transid 55586 /dev/sdc5
btrfs: use ssd allocation scheme
btrfs: disk space caching is enabled
btrfs: checking UUID tree
btrfs: corrupt leaf, slot offset bad: block=842924032,root=1, slot=88
btrfs: corrupt leaf, slot offset bad: block=842924032,root=1, slot=88
BTRFS error (device sdc5): Error removing orphan entry, stopping orphan 
cleanup
BTRFS critical (device sdc5): could not do orphan cleanup -22

Doing "btrfs check /dev/sdc5" merely first prints ten

free space inode generation (0) did not match free space cache 
generation ([different transids between 40010 and 55578])

to then abort with

checking fs roots
btrfs: cmds-check.c:1151: procecss_file_extent: Assertion `!(rec->ino != 
key->objectid || rec->refs > 1)' failed.

I'm reluctant to try any of "btrfs check" options (or mount with -o 
recovery) since the last three times I did this (with other partitions) 
it resulted in the partition becoming entirely trashed, while before at 
least "btrfs restore" still managed to extract some data each time.

The affected folder was one within /usr/include/qt4 (which I then moved 
to /usr/BROKEN, to successfully reinstall the package), ie. on the 
__current/ROOT subvolume.

Which seems the only subvolume affected (yet). Mounting & accessing the 
other three (__current/{var,home,opt}) still works.

[*] Organised following 
http://blog.fabio.mancinelli.me/2012/12/28/Arch_Linux_on_BTRFS.html

(Also posted on https://bugzilla.kernel.org/show_bug.cgi?id=74611 )