btrfs: give the read repair bio an end_io for submit failures
repair_one_sector() initializes the repair bbio with a NULL end_io. A
repair bio normally completes through btrfs_check_read_bio(), which hands
it to btrfs_end_repair_bio() and never calls end_io. But
btrfs_submit_chunk() can fail a bio before it reaches a device: when
btrfs_map_block() fails, when btrfs_split_bio() fails, or when
btrfs_lookup_bio_sums() fails for a data read. It then ends the bio with
btrfs_bio_end_io(), which calls bbio->end_io. For a repair bio that is a
jump to address 0 in a btrfs-endio worker. The original read never
completes, and an unmount hangs on it later:
BUG: kernel NULL pointer dereference, address:
0000000000000000
Workqueue: btrfs-endio simple_end_io_work
RIP: 0010:0x0
Call Trace:
btrfs_bio_end_io+0x1b7/0x420
btrfs_submit_bbio+0xc1e/0x1fa0
btrfs_check_read_bio+0x80e/0x1760
simple_end_io_work+0xc8/0x190
The repair bio's checksum lookup can fail where the original read's
lookup succeeded, because the two search different trees. A buffered
read of an extent from a past transaction searches the commit root
(csum_search_commit_root). The repair bbio leaves that flag clear and
searches the running transaction's csum tree. When that transaction
aborts, the cleanup frees its dirty tree blocks. A later lookup reads
such a block from disk, where it was never written, and fails with -EIO:
BTRFS error (device dm-6 state EA): bad tree block start, mirror 2 want
31490048 have
14842989077554535085
BTRFS error (device dm-6 state EA): bad tree block start, mirror 1 want
31490048 have
8810637562337132983
This was first seen right after a relocation forced the filesystem
read-only: six btrfs-endio workers oopsed within 1.5 seconds. It
reproduces with a raid1 filesystem:
- Commit a file, write more data so that the csum leaf holding the file's
tail is COWed in the running transaction, and corrupt the tail on both
mirrors.
- Abort the transaction by failing a log commit's superblock write, and
drop the page cache.
- A buffered read of the tail now oopses.
Give the repair bbio an end_io for this case. A failed submission does
not depend on the mirror: the mapping and the checksum lookup fail the
same way for every copy. So give up on the sector the way an exhausted
repair does: mark the original bio failed, drop the repair's count on
the failed_bio, and free the repair bio. The read then fails with -EIO.
Fixes: 7609afac6775 ("btrfs: handle checksum validation and repair at the storage layer")
Cc: stable@vger.kernel.org # 6.6+
Assisted-by: Claude:claude-opus-5-5