Hi all, This is a v2 of the stale buffer locking fix for an aborted transaction. It has minor refactoring changes noted below. I've added two optional patches on top because I wanted to take a stab at a bit more cleanup in this area, but these are aesthetic changes only and can be dropped if desired. Patch 1 is the original bug fix. Patch 2 cleans up xfs_trans_brelse() to remove some cruft and use similar logic/flow as xfs_buf_item_unlock(). Patch 3 refactors the bli refcount cleanup logic that is common between the two functions. There is still open-coded bli refcount handling on unpin because there are additional considerations in that path. Stale inode buffer handling has special handling, for whatever reason. The stale inode buf handling actually looks quite similar to the non-specialized buffer handling, but further cleanup in that area would require a more in-depth audit to be sure. I find the end result a bit cleaner but I'm not wedded to it and I'm fine with just dropping patches 2-3 if the current approach is preferred. FWIW, this also raised a couple thoughts on alternative approaches, such as reworking the bli refcount such that the AIL also holds a reference, but it's not totally clear that (by itself) simplifies things in the case of a shutdown with AIL-held blis. Thoughts, reviews, flames appreciated. Brian v2: - Refactor stale bli unlock logic to more closely resemble original. - Use bool for 'freed' variable and clean up type casting. - Reorder asserts/tracepoints and remove unused stale tracepoint. - Added patches 2 and 3. v1: https://marc.info/?l=linux-xfs&m=153486008127962&w=2 Brian Foster (3): xfs: don't unlock invalidated buf on aborted tx commit xfs: clean up xfs_trans_brelse() xfs: refactor xfs_buf_log_item reference count handling fs/xfs/xfs_buf_item.c | 115 +++++++++++++++++++++-------------------- fs/xfs/xfs_buf_item.h | 1 + fs/xfs/xfs_trace.h | 1 - fs/xfs/xfs_trans_buf.c | 98 +++++++++-------------------------- 4 files changed, 84 insertions(+), 131 deletions(-) -- 2.17.1
