On Thu, Aug 03, 2023 at 04:30:28PM -0700, Darrick J. Wong wrote: > From: Darrick J. Wong <djwong@xxxxxxxxxx> > > If the fscounters scrubber notices incorrect summary counters, it's > entirely possible that scrub is simply racing with other threads that > are updating the incore counters. There isn't a good way to stabilize > percpu counters or set ourselves up to observe live updates with hooks > like we do for the quotacheck or nlinks scanners, so we instead choose > to freeze the filesystem long enough to walk the incore per-AG > structures. > > Past me thought that it was going to be commonplace to have to freeze > the filesystem to perform some kind of repair and set up a whole > separate infrastructure to freeze the filesystem in such a way that > userspace could not unfreeze while we were running. This involved > adding a mutex and freeze_super/thaw_super functions and dealing with > the fact that the VFS freeze/thaw functions can free the VFS superblock > references on return. > > This was all very overwrought, since fscounters turned out to be the > only user of scrub freezes, and it doesn't require the log to quiesce, > only the incore superblock counters. We prevent other threads from > changing the freeze level by calling freeze_super_excl with a custom > freeze cookie to keep everyone else out of the filesystem. > > The end result is that fscounters should be much more efficient. When > we're checking a busy system and we can't stabilize the counters, the > custom freeze will do less work, which should result in less downtime. > Repair should be similarly speedy, but that's in the next patch. > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx> > --- > v2: remove fscounters.h > --- > fs/xfs/scrub/fscounters.c | 188 ++++++++++++++++++++++++++++++++++++--------- > fs/xfs/scrub/scrub.c | 6 + > fs/xfs/scrub/scrub.h | 1 > fs/xfs/scrub/trace.h | 26 ++++++ > 4 files changed, 183 insertions(+), 38 deletions(-) looks good. Reviewed-by: Dave Chinner <dchinner@xxxxxxxxxx> -- Dave Chinner david@xxxxxxxxxxxxx
