Before this patch set, several gfs2 functions failed to reserve enough revoke entries in the journal. Some examples:
1. gfs2_dinode_dealloc failed to reserve a revoke for the dinode being deleted. 2. Any function that allocates dinodes with gfs2_alloc_blocks should reserve a revoke because alloc_blocks will premptively call trans_remove_revoke to make sure there isn't a pending revoke for the new dinode. 3. Any function that potentially will unstuff a stuffed directory needs to reserve a revoke because gfs2_unstuff_dinode calls gfs2_trans_remove_revoke for the new journaled leaf block. In addition, function gfs2_trans_remove_revoke unconditionally decrements tr->tr_num_revoke, and if not enough revokes are reserved, the value goes from 0 to 4294967295 (-1, but it's an unsigned int). This is later re-added to the system-wide revoke numbers, thereby decrementing the value (sd_log_commited_revoke) "properly," but by accident. This worked properly most of the time because one transaction would reserve space for revokes, then it would be merged with the system transaction (sdp->sd_log_tr) and it usually did not run out, because you can hold a lot of revoke entries per log descriptor block. Some of the code, such as gfs2_write_revokes, would work around this and somehow got it right most of the time. However, some jdata tests with xfstests generic/269 encountered problems when it actually ran out. This series adds needed revoke entries to the transactions that need them. So now we try to do proper accounting of revokes. Bob Peterson (6): gfs2: revoke cleanup: leaf_dealloc gfs2: revoke cleanup: alloc_dinode and gfs2_create_inode gfs2: revoke cleanup: gfs2_dinode_dealloc gfs2: revoke cleanup: gfs2_iomap_begin_write gfs2: revoke cleanup: truncate functions gfs2: revoke cleanup: gfs2_trans_remove_revoke fs/gfs2/bmap.c | 25 +++++++++++++------------ fs/gfs2/dir.c | 3 ++- fs/gfs2/inode.c | 5 +++-- fs/gfs2/super.c | 2 +- fs/gfs2/trans.c | 16 ++++++++++++++-- 5 files changed, 33 insertions(+), 18 deletions(-) -- 2.24.1
