On 10/21/19 6:02 AM, yangerkun wrote:
> Before 5da0fb1ab34c ("io_uring: consider the overflow of sequence for
> timeout req"). We can meet some situation like below:
>
> 1. setup
> 2. prepare 4 timeout req which expected count is 1,1,2,UINT_MAX, and the
> sequence of this 4 requests will be 1,2,4,2, this 4 requests will not
> lead the change of cached_cq_tail and sq_dropped until the timeout
> really happened. So the tail_index in io_timeout will still be 0.
> 3. based on the above and before this patch, the order of timeout_list
> will be req1->req2->req4->req3, which the right order should be
> req1->req2->req3->req4.
> 4. setup two nop requests. And the timeout requests will return
> correctly with the patch.
>
> Add this testcase to cover it.
Thanks, applied.
--
Jens Axboe
