[BUG] ext4/jbd2: KASAN UAF in jbd2_fc_release_bufs after loop shrink + fast commit fsync

Hello,

I am reporting a use-after-free in jbd2_fc_release_bufs() when ext4 fast
commit fails after the loop backing store is shrunk while journal I/O is
still in flight.

Summary
=======

On ext4 with fast commits mounted on loop0:

1. Write + fsync (fast commit path).
2. Shrink loop capacity (campaign case: 32768 -> 16726 sectors).
3. fsync again.

Journal block writes fail; ext4_fc_commit falls back and ext4_fc_cleanup
calls jbd2_fc_release_bufs(), which put_bh()s buffer heads whose pages
were already torn down by the loop shrink / I/O error path. KASAN reports
use-after-free at fs/jbd2/journal.c:942 (jbd2_fc_release_bufs).

This is distinct from CVE-2022-50328 (jbd2_fc_wait_bufs UAF on the wait
path, already on 6.6). Our crash is on the release/cleanup path after
failed FC commit.

Affected
========

- Confirmed on Linux 6.6.144 KASAN (QEMU guest)
- Files: fs/jbd2/journal.c, fs/ext4/fast_commit.c
- Needs CONFIG_EXT4_FS, fast_commit feature, loop device

Crash excerpt (campaign)
========================

BUG: KASAN: use-after-free in jbd2_fc_release_bufs+0x95/0x170
  ext4_fc_cleanup
  __jbd2_fc_end_commit
  ext4_fc_commit
  ext4_sync_file
  __x64_sys_fsync

(Preceding ext4_end_bio I/O errors and loop capacity change.)

Reproducer
==========

Attached: ext4-fc-jbd2-fc-release-bufs-uaf-repro.tar.gz

VM-only (KASAN runs inside QEMU, not on the host):

   tar xzf ext4-fc-jbd2-fc-release-bufs-uaf-repro.tar.gz
   cd attachments
   docker build -t ext4-fc-jbd2-fc-release-bufs-uaf -f Dockerfile .
   mkdir -p artifacts
   docker run --rm --privileged --device=/dev/kvm --network=host \
     -v "$PWD/artifacts:/artifacts" -e OUTPUT_DIR=/artifacts \
     -e RUN_TIMEOUT_SEC=900 -e REPRO_TRIES=5 \
     ext4-fc-jbd2-fc-release-bufs-uaf

First run builds Linux 6.6.144 KASAN + guest image (cached under 
artifacts/).
Expect >=1 hit in REPRO_TRIES attempts when successful.

Impact
======

Local DoS / memory corruption on error paths when loop (or similar) backing
disappears during fast commit. Practical exploitation beyond DoS not 
evaluated.

Happy to test patches. Let me know if you need a C reproducer or 
syzkaller prog.

Thanks,
Yu Junzhe
FuzzAnything <[email protected]>