Re: [PATCH V3 2/4] blk-mq: implement queue quiesce via percpu_ref for BLK_MQ_F_BLOCKING

[Hannes Reinecke <hare@suse.de>]

On 9/8/20 10:15 AM, Ming Lei wrote:
> In case of BLK_MQ_F_BLOCKING, blk-mq uses SRCU to mark read critical
> section during dispatching request, then request queue quiesce is based on
> SRCU. What we want to get is low cost added in fast path.
> 
> With percpu-ref, it is cleaner and simpler & enough for implementing queue
> quiesce. The main requirement is to make sure all read sections to observe
> QUEUE_FLAG_QUIESCED once blk_mq_quiesce_queue() returns.
> 
> Also it becomes much easier to add interface of async queue quiesce.
> 
> Meantime memory footprint can be reduced with per-request-queue percpu-ref.
> 
> From implementation viewpoint, in fast path, not see percpu_ref is
> slower than SRCU, and srcu tree(default option in most distributions)
> could be slower since memory barrier is required in both lock & unlock,
> and rcu_read_lock()/rcu_read_unlock() should be much cheap than
> smp_mb().
> 
> 1) percpu_ref just hold the rcu_read_lock, then run a check &
>    increase/decrease on the percpu variable:
> 
>    rcu_read_lock()
>    if (__ref_is_percpu(ref, &percpu_count))
> 	this_cpu_inc(*percpu_count);
>    rcu_read_unlock()
> 
> 2) srcu tree:
>         idx = READ_ONCE(ssp->srcu_idx) & 0x1;
>         this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
>         smp_mb(); /* B */  /* Avoid leaking the critical section. */
> 
> Also from my test on null_blk(blocking), not observe percpu-ref performs
> worse than srcu, see the following test:
> 
> 1) test steps:
> 
> rmmod null_blk > /dev/null 2>&1
> modprobe null_blk nr_devices=1 submit_queues=1 blocking=1
> fio --bs=4k --size=512G  --rw=randread --norandommap --direct=1 --ioengine=libaio \
> 	--iodepth=64 --runtime=60 --group_reporting=1  --name=nullb0 \
> 	--filename=/dev/nullb0 --numjobs=32
> 
> test machine: HP DL380, 16 cpu cores, 2 threads per core, dual
> sockets/numa, Intel(R) Xeon(R) Silver 4110 CPU @ 2.10GHz
> 
> 2) test result:
> - srcu quiesce: 6063K IOPS
> - percpu-ref quiesce: 6113K IOPS
> 
> Signed-off-by: Ming Lei <ming.lei@redhat.com>
> Cc: Sagi Grimberg <sagi@grimberg.me>
> Cc: Bart Van Assche <bvanassche@acm.org>
> Cc: Johannes Thumshirn <Johannes.Thumshirn@wdc.com>
> Cc: Chao Leng <lengchao@huawei.com>
> ---
>  block/blk-mq-sysfs.c   |   2 -
>  block/blk-mq.c         | 130 +++++++++++++++++++++--------------------
>  block/blk-sysfs.c      |   6 +-
>  include/linux/blk-mq.h |   8 ---
>  include/linux/blkdev.h |   4 ++
>  5 files changed, 77 insertions(+), 73 deletions(-)
> 
Reviewed-by: Hannes Reinecke <hare@suse.de>

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		           Kernel Storage Architect
hare@suse.de			                  +49 911 74053 688
SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), GF: Felix Imendörffer