There is a lock contention on device_lock in read_one_chunk().
device_lock is taken to sync conf->active_aligned_reads and
conf->quiesce.
read_one_chunk() takes the lock, then waits for quiesce=0 (resumed)
before incrementing active_aligned_reads.
raid5_quiesce() takes the lock, sets quiesce=2 (in-progress), then waits
for active_aligned_reads to be zero before setting quiesce=1
(suspended).

Introduce a fast (lockless) path in read_one_chunk(): activate aligned
read without taking device_lock.  In case quiesce starts while
activating the aligned-read in fast path, deactivate it and revert to
old behavior (take device_lock and wait for quiesce to finish).

Add smp store/load in raid5_quiesce()/read_one_chunk() respectively to
gaurantee that read_one_chunk() does not miss an ongoing quiesce.

My setups:
1. 8 local nvme drives (each up to 250k iops).
2. 8 ram disks (brd).

Each setup with raid6 (6+2), 1024 io threads on a 96 cpu-cores (48 per
socket) system. Record both iops and cpu spent on this contention with
rand-read-4k. Record bw with sequential-read-128k.  Note: in most cases
cpu is still busy but due to "new" bottlenecks.

nvme:
              | iops           | cpu  | bw
-----------------------------------------------
without patch | 1.6M           | ~50% | 5.5GB/s
with patch    | 2M (throttled) | 0%   | 16GB/s (throttled)

ram (brd):
              | iops           | cpu  | bw
-----------------------------------------------
without patch | 2M             | ~80% | 24GB/s
with patch    | 4M             | 0%   | 55GB/s

CC: Song Liu <song@kernel.org>
CC: Neil Brown <neilb@suse.de>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Gal Ofri <gal.ofri@storing.io>
Signed-off-by: Song Liu <song@kernel.org>