On Wed, Apr 10, 2024 at 09:56:45PM -0700, Darrick J. Wong wrote: > > Well, someone needs to own it, it's just not just ext4 but could us. > > Er... I don't understand this? ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ If we set current->journal and take a page faul we could not just recurse into ext4 but into any fs including XFS. Any everyone blindly dereferences is as only one fs can own it. > > > Alloc transaction -> lock rmap btree for repairs -> iscan filesystem to > > > find rmap records -> iget/irele. > > > > So this is just the magic empty transaction? > > No, that's the fully featured repair transaction that will eventually be > used to write/commit the new rmap tree. That seems a bit dangerous to me. I guess we rely on the code inside the transaction context to never race with unmount as lack of SB_ACTIVE will make the VFS ignore the dontcache flag.
