Hi all, Currently, phase 4 of xfs_scrub uses per-AG repair item lists to schedule repair work across a thread pool. This scheme is suboptimal when most of the repairs involve a single AG because all the work gets dumped on a single pool thread. Instead, we should create a thread pool with the same number of workers as CPUs, and dispatch individual repair tickets as separate work items to maximize parallelization. However, we also need to ensure that repairs to space metadata and file metadata are kept in separate queues because file repairs generally depend on correctness of space metadata. 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. xfsprogs git tree: https://git.kernel.org/cgit/linux/kernel/git/djwong/xfsprogs-dev.git/log/?h=scrub-repair-scheduling-6.10 --- Commits in this patchset: * libfrog: enhance ptvar to support initializer functions * xfs_scrub: improve thread scheduling repair items during phase 4 * xfs_scrub: recheck entire metadata objects after corruption repairs * xfs_scrub: try to repair space metadata before file metadata --- libfrog/ptvar.c | 9 ++ libfrog/ptvar.h | 4 + scrub/counter.c | 2 scrub/descr.c | 2 scrub/phase1.c | 15 ++- scrub/phase2.c | 23 ++++- scrub/phase3.c | 106 ++++++++++++++------- scrub/phase4.c | 244 ++++++++++++++++++++++++++++++++++++------------- scrub/phase7.c | 2 scrub/read_verify.c | 2 scrub/repair.c | 172 ++++++++++++++++++++++------------- scrub/repair.h | 37 ++++++- scrub/scrub.c | 5 + scrub/scrub.h | 10 ++ scrub/scrub_private.h | 2 scrub/xfs_scrub.h | 3 - 16 files changed, 455 insertions(+), 183 deletions(-)
