-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
Issue1: thread hit BUG_ON(JBD2_UNMOUNT) at umount
The fstrim thread added by LU-17980 can start a journal handle while the journal is being destroyed at umount:
kernel BUG at fs/jbd2/transaction.c:369! CPU: 4 PID: 1649839 Comm: fstrim-sdb 6.12.0-124.28.1.el10_1.x86_64 RIP: 0010:start_this_handle+0x226/0x500 [jbd2] Call Trace: jbd2__journal_start+0xfb/0x1f0 [jbd2] ldiskfs_try_to_trim_range+0x5ec/0x7e0 [ldiskfs] ldiskfs_trim_all_free+0x168/0x1c0 [ldiskfs] ldiskfs_mb_fstrim_thread+0x418/0x550 [ldiskfs] kthread+0xfd/0x240
The thread is started in ext4_mb_init() and stopped in ext4_mb_release(), but ext4_put_super() destroys the journal before calling ext4_mb_release(). While trimming, the thread starts a
journal handle (ldiskfs_try_to_trim_range() -> jbd2__journal_start()) to record the trimmed groups, so if it is still running when jbd2_journal_destroy() has set JBD2_UNMOUNT, start_this_handle() hits the BUG_ON. The same placement also leaks the thread when mount fails after ext4_mb_init() (the failed_mount5 path does not stop it), and leaves it running after "remount,nofstrim", since ext4_mb_release() skips the stop when the option is cleared.
issue2: thread was not started/stopped at remount:
a) "remount,fstrim" on a filesystem mounted without fstrim sets the option but never starts the thread. Queued groups are never trimmed, and the allocator keeps skipping them in the fast scan passes. b) "remount,nofstrim" leaves the thread running until umount. c) "remount,ro" leaves the thread running. Trimming then hits WARN_ON_ONCE(sb_rdonly(sb)) in ext4_journal_check_start() and gets -EROFS, after which the thread never trims again, even after "remount,rw".