beldzhang

1) error handling pass 2) sendzc have lower cpu usage even in sync waiting for buffer complete 3) async waiting of buffer complete is postponed because buffer management redesign required...

> Finally caught the bug and see where it came from the unexpected `ECANCELED` happened on the testing server with 2 CPU, 256KB random read and linked with send. nvme0n1...

finally found it in `io_read()`, if there is no retry, currently only passed the last ret of `io_iter_do_read` to `kiocb_done` then failed in `__io_complete_rw_common()` because `ret != cqe.res`, it should...

ok, a wild wild west concept patch that fixed this abnormal `ECANCELED`, based on 6.0-rc6 ``` --- rw-orig.c 2022-09-18 16:44:14.000000000 -0400 +++ rw.c 2022-09-18 19:53:58.725235656 -0400 @@ -706,7 +706,7 @@...

oh, there are two different issues: > 4. another short recv found in 5.15.67, cause a recv-write link broken, not found in 5.19 and 6.0-rc ... this is found recently,...

> Yes this is expected. iowait literally just means "waiting on IO", which is what the task is doing. It does NOT mean that it's busy 100% of the time,...

looks no more comments, closed.

emmm... reopen?

will test when avaliable

> https://git.kernel.dk/cgit/linux/log/?h=iowait.2 @axboe brief tested, io-wait is gone, will following up. mailing list followed also. @isilence to the end users, they are sensitive to the latency and high response time...