On 5/30/22 00:25, Jens Axboe wrote:
On 5/29/22 10:20 AM, Hao Xu wrote:
From: Hao Xu <howeyxu@xxxxxxxxxxx>
From: Hao Xu <howeyxu@xxxxxxxxxxx>
Use per list lock for cancel_hash, this removes some completion lock
invocation and remove contension between different cancel_hash entries
Interesting, do you have any numbers on this?
Also, I'd make a hash bucket struct:
struct io_hash_bucket {
spinlock_t lock;
struct hlist_head list;
};
rather than two separate structs, that'll have nicer memory locality too
and should further improve it. Could be done as a prep patch with the
old locking in place, making the end patch doing the per-bucket lock
simpler as well.
Hmm?
I've done a v2 here, also a test which issues async poll densely to
make high frequency cancel_hash[] visits. But I won't have a real box
with big number of cpu processors which is suitable for testing until
tomorrow, so I'd test it tomorrow.
https://github.com/HowHsu/linux/commits/for-5.20/io_uring_hash_lock
https://github.com/HowHsu/liburing/commit/b9fb4d20a5dfe7c7bd62fe36c37aea3b261d4499
