From: Yu Kuai <yukuai3@xxxxxxxxxx> Lock contention for sq elevator is quite intense when multiple contexts are dispatching request, main reason is that a global lock is held and release for each request: while spin_lock disaptch one request spin_unlock This set support request batch dispatching to reduce the lock contention a bit. A simple test for null_blk in my VM: 32 job randwrite | elevator | bandwidth Mib/s | | -------------------------- | --------------- | | none | 813 | | deadline before this patch | 531 | | deadline after this patch | 689 | The test is one-sided, however, this is just a quick test to show this set might be meaningful. I'm sending this RFC and want to get some feedbacks before I continue work on this set. Noted: do not test bfq for this set, it's not changed yet(like patch 2 for deadline). Yu Kuai (4): elevator: introduce global lock for sq_shared elevator mq-deadline: switch to use elevator lock blk-mq-sched: refactor __blk_mq_do_dispatch_sched() blk-mq-sched: support request batch dispatching for sq elevator block/blk-mq-sched.c | 224 +++++++++++++++++++++++++++++-------------- block/blk-mq.c | 5 +- block/blk-mq.h | 21 ++++ block/elevator.c | 1 + block/elevator.h | 62 +++++++++++- block/mq-deadline.c | 60 +++++------- 6 files changed, 259 insertions(+), 114 deletions(-) -- 2.39.2
