On 4/30/22 21:27, Jens Axboe wrote:
On 4/29/22 4:18 AM, Hao Xu wrote:@@ -1030,6 +1101,7 @@ static bool io_wq_work_match_item(struct io_wq_work *work, void *data) static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) { struct io_wqe_acct *acct = io_work_get_acct(wqe, work); + struct io_wqe_acct *fixed_acct; struct io_cb_cancel_data match; unsigned work_flags = work->flags; bool do_create; @@ -1044,8 +1116,14 @@ static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work) return; }+ fixed_acct = io_get_acct(wqe, !acct->index, true);+ if (fixed_acct->fixed_worker_registered && !io_wq_is_hashed(work)) { + if (io_wqe_insert_private_work(wqe, work, fixed_acct)) + return; + } +As per previous email, I was going to comment back saying "why don't we just always do hashed work on the non-fixed workers?" - but that's already what you are doing. Isn't this fine, does anything else need to get done here in terms of hashed work and fixed workers? If you need per-iowq serialization, then you don't get a fixed worker.
Hmm, seems we cannot accelerate serialized works with fixed worker. So Let's make it as it is.
