Hi Hannes,
Here is what I intend to queue for 5.7. I didn't hear from you on
the need to protect the object request list, the impact of image
request state_mutex on performance or the problem with queue_rq(),
so I concentrated on the tail of your series and reworked the payload
patch to avoid acquiring header_rwsem at least twice for each I/O.
I observed the same ~25-30% improvement in some rough synthetic tests
on a laptop (random 4k reads with very high queue depth in particular).
I assume you did your testing on real hardware -- it would be great if
you could share some details.
Thanks,
Ilya
Hannes Reinecke (2):
rbd: kill img_request kref
rbd: enable multiple blk-mq queues
Ilya Dryomov (3):
rbd: get rid of img_request_layered_clear()
rbd: acquire header_rwsem just once in rbd_queue_workfn()
rbd: embed image request in blk-mq pdu
drivers/block/rbd.c | 212 +++++++++++++++++---------------------------
1 file changed, 79 insertions(+), 133 deletions(-)
--
2.19.2
