[0/8,v3] target: fix up locking/refcounting in IO paths

Message

Mike Christie Nov. 1, 2020, 6:59 p.m. UTC

The following patches made over Martin's staging branch fix some
ref counting issues I hit while testing and improves the locking
in the IO paths. To do the latter, the patches:

1. move the sess_cmd_lock to tcm_qla2xxx since it was the only
driver using the sess_cmd_list.

2. makes the execution lock/list per cpu'ish. I just allocate
nr_cpu_ids's worth of lock/lists then make sure we complete
the cmd on the cpu it was started on.

With the patches I'm seeing a 25% improvement in IOPs for small
IO tests like:

fio  --filename=/dev/sdXYZ  --direct=1 --rw=randrw --bs=4k \
--iodepth=128  --numjobs=16

with drivers like vhost (with those other patches on the list to
fix up multiple virtqueue support) and with the included loop
patch when nr hw queues is increased.

v3:
- Fixed issue where qla2xxx's cpuid was overwritten.
- Fixed up email submit prefix to have "qla2xxx".
v2:
- Got access to qla2xxx setup and tested patch. Fixed various issues.
- Added fixes for issues found in the same code paths I was testing:
        - target: fix lun ref count handling
        - target: fix cmd_count ref leak
v1/RFC
- Initial posting.

Comments

Martin K. Petersen Nov. 5, 2020, 3:41 a.m. UTC | #1

Mike,

> The following patches made over Martin's staging branch fix some
> ref counting issues I hit while testing and improves the locking
> in the IO paths. To do the latter, the patches:

Applied to 5.11/scsi-staging, thanks!