Hi all, This patchset looks to reduce iget overhead in two ways: First, a previous patch conditionally set DONTCACHE on inodes during xchk_irele on the grounds that we knew better at irele time if an inode should be dropped. Unfortunately, over time that patch morphed into a call to d_mark_dontcache, which resulted in inodes being dropped even if they were referenced by the dcache. This actually caused *more* recycle overhead than if we'd simply called xfs_iget to set DONTCACHE only on misses. The second patch reduces the cost of untrusted iget for a vectored scrub call by having the scrubv code maintain a separate refcount to the inode so that the cache will always hit. If you're going to start using this code, I strongly recommend pulling from my git trees, which are linked below. This has been running on the djcloud for months with no problems. Enjoy! Comments and questions are, as always, welcome. --D kernel git tree: https://git.kernel.org/cgit/linux/kernel/git/djwong/xfs-linux.git/log/?h=reduce-scrub-iget-overhead --- Commits in this patchset: * xfs: use dontcache for grabbing inodes during scrub * xfs: only iget the file once when doing vectored scrub-by-handle --- fs/xfs/scrub/common.c | 12 +++--------- fs/xfs/scrub/iscan.c | 13 +++++++++++-- fs/xfs/scrub/scrub.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ fs/xfs/scrub/scrub.h | 7 +++++++ 4 files changed, 66 insertions(+), 11 deletions(-)
