On Mon, Jan 02, 2023 at 05:24:24PM +0900, Tetsuo Handa wrote: > Since no lockdep annotation is used for e.g. PG_locked bit, this deadlock > cannot be detected by lockdep... lockdep, unfortunately, cannot track PG_locked. Lockdep requires that the lock is released by the acquirer, and sometimes that's true for PG_locked, but when it's used to do I/O, the PG_locked bit is released in interrupt/BH context. We could maybe fake it by pretending we release the folio lock when we submit the I/O. Then we'll have to figure out how to tell lockdep that it's OK to grab the folio lock multiple times (if within the same inode, ordered by folio->index; if in different inodes, ordered by in-memory address of those inodes), and that submitting an I/O will unlock all of the folios in that I/O. Oh, but there's cases where we only submit part of a folio in an I/O, and the lock will only be released when all of the I/Os targetting that folio have been completed. It's not impossible, but it is a lot of work and needs a lot of understanding of filesystems/mm/io.
