Re: Polled I/O cannot find completions

[Jens Axboe <axboe@kernel.dk>]

CC'ing linux-block, this isn't an io_uring issue.


On 3/26/20 8:57 PM, Bijan Mottahedeh wrote:
> I'm seeing poll threads hang as I increase the number of threads in 
> polled fio tests.  I think this is because of polling on BLK_QC_T_NONE 
> cookie, which will never succeed.
> 
> A related problem however, is that the meaning of BLK_QC_T_NONE seems to 
> be ambiguous.
> 
> Specifically, the following cases return BLK_QC_T_NONE which I think 
> would be problematic for polled io:
> 
> 
> generic_make_request()
> ...
>          if (current->bio_list) {
>                  bio_list_add(&current->bio_list[0], bio);
>                  goto out;
>          }
> 
> In this case the request is delayed but should get a cookie eventually.  
> How does the caller know what the right action is in this case for a 
> polled request?  Polling would never succeed.
> 
> 
> __blk_mq_issue_directly()
> ...
>          case BLK_STS_RESOURCE:
>          case BLK_STS_DEV_RESOURCE:
>                  blk_mq_update_dispatch_busy(hctx, true);
>                  __blk_mq_requeue_request(rq);
>                  break;
> 
> In this case, cookie is not updated and would keep its default 
> BLK_QC_T_NONE value from blk_mq_make_request().  However, this request 
> will eventually be reissued, so again, how would the caller poll for the 
> completion of this request?
> 
> blk_mq_try_issue_directly()
> ...
>          ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
>          if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
>                  blk_mq_request_bypass_insert(rq, false, true);
> 
> Am I missing something here?
> 
> Incidentally, I don't see BLK_QC_T_EAGAIN used anywhere, should it be?
> 
> Thanks.
> 
> --bijan
> 
> 
> 
> 


-- 
Jens Axboe