On 2020/8/7 21:49, Ming Lei wrote:
On Fri, Aug 07, 2020 at 05:05:59PM +0800, Chao Leng wrote:
nvme_stop_queues quiesce queues for all name spaces, now quiesce one by
one, if there is lots of name spaces, sync wait long time(more than 10s).
Multipath can not fail over to retry quickly, cause io pause long time.
This is not expected.
To reduce quiesce time, we introduce async mechanism for sync SRCUs
and quiesce queue.
Frankly speaking, I prefer to replace SRCU with percpu_refcount:
- percpu_refcount has much less memory footprint than SRCU, so we can simply
move percpu_refcount into request_queue, instead of adding more bytes
into each hctx by this patch
- percpu_ref_get()/percpu_ref_put() isn't slower than srcu_read_lock()/srcu_read_unlock().
- with percpu_refcount, we can remove 'srcu_idx' from hctx_lock/hctx_unlock()
IO pause long time if fail over, this is a serios problem. we need fix
it as soon as possible. SRCU is just used for blocking queue,
non blocking queue need 0 bytes. So more bytes(just 24 bytes) is not
waste.
About using per_cpu to replace SRCU, I suggest separate discussion.
Can you show the patch? This will make it easier to discuss.
Thanks,
Ming
.
