On Wed, Mar 03, 2021 at 07:45:56AM +0100, Christoph Hellwig wrote: > On Tue, Mar 02, 2021 at 02:28:28PM -0800, Darrick J. Wong wrote: > > From: Darrick J. Wong <djwong@xxxxxxxxxx> > > > > When we're servicing an INUMBERS or BULKSTAT request, grab an empty > > transaction so that we don't hit an ABBA buffer deadlock if the inode > > btree contains a cycle. > > > > Found by fuzzing an inode btree pointer to introduce a cycle into the > > tree (xfs/365). > > > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx> > > So basically you want to piggy back on the per-trans recursion using > xfs_trans_buf_item_match? Right. > Why do we need the sb-counter for that? I was about to say that we need freeze protection to prevent fsfreeze and inumbers/bulkstat stalling each other out on the buffer LRU, but then I remembered that Brian changed freeze to wait for IO to complete without dumping the buffer cache this cycle. So, I don't think freeze protection is necessary anymore. > Can the comments be a little more clear? /* * Grab an empty transaction so that we can use the recursive locking * support to detect tree cycles instead of deadlocking. */ How does that sound? > Why don't we want that elsewhere where we're walking the btrees? I'm merely playing whackamole with tree walks here. I would guess that quotacheck will be next for evaluation. It also occurs to me that inumbers/bulkstat should probably not be calling copy_to_user() separately for every single record. --D
