On 3/5/25 10:03 AM, Jens Axboe wrote:
> On 3/5/25 9:57 AM, John Garry wrote:
>> On 04/03/2025 18:10, John Garry wrote:
>>
>> +
>>
>>> On 09/01/2025 18:15, Jens Axboe wrote:
>>>> Rather than try and have io_read/io_write turn REQ_F_REISSUE into
>>>> -EAGAIN, catch the REQ_F_REISSUE when the request is otherwise
>>>> considered as done. This is saner as we know this isn't happening
>>>> during an actual submission, and it removes the need to randomly
>>>> check REQ_F_REISSUE after read/write submission.
>>>>
>>>> If REQ_F_REISSUE is set, __io_submit_flush_completions() will skip over
>>>> this request in terms of posting a CQE, and the regular request
>>>> cleaning will ensure that it gets reissued via io-wq.
>>>>
>>>> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
>>>
>>
>> Further info, I can easily recreate this on latest block/io_uring-6.14 on real NVMe HW:
>
> Thanks, I'll take a look!
Can you give this a spin?
diff --git a/io_uring/rw.c b/io_uring/rw.c
index 9edc6baebd01..e5528cebcd06 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c
@@ -560,11 +560,10 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
if (kiocb->ki_flags & IOCB_WRITE)
io_req_end_write(req);
if (unlikely(res != req->cqe.res)) {
- if (res == -EAGAIN && io_rw_should_reissue(req)) {
+ if (res == -EAGAIN && io_rw_should_reissue(req))
req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE;
- return;
- }
- req->cqe.res = res;
+ else
+ req->cqe.res = res;
}
/* order with io_iopoll_complete() checking ->iopoll_completed */
--
Jens Axboe
