Re: [PATCH] null_blk: add 'requeue' fault attribute

[Jens Axboe <axboe@kernel.dk>]

On 2/28/18 9:14 AM, Omar Sandoval wrote:
> On Wed, Feb 28, 2018 at 08:28:25AM -0700, Jens Axboe wrote:
>> On 2/28/18 1:51 AM, Omar Sandoval wrote:
>>> On Tue, Feb 27, 2018 at 03:34:53PM -0700, Jens Axboe wrote:
>>>> Similarly to the support we have for testing/faking timeouts for
>>>> null_blk, this adds support for triggering a requeue condition.
>>>> Considering the issues around restart we've been seeing, this should be
>>>> a useful addition to the testing arsenal to ensure that we are handling
>>>> requeue conditions correctly.
>>>>
>>>> This works for queue mode 1 (legacy request_fn based path) and 2 (blk-mq
>>>> path), as there's no good way to do requeue with a bio based driver.
>>>> This is similar to the timeout path.
>>>>
>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>>>
>>>> ---
>>>>
>>>>  null_blk.c |   55 +++++++++++++++++++++++++++++++++++++++++++------------
>>>>  1 file changed, 43 insertions(+), 12 deletions(-)
>>>>
>>>> diff --git a/drivers/block/null_blk.c b/drivers/block/null_blk.c
>>>> index 287a09611c0f..363536572e19 100644
>>>> --- a/drivers/block/null_blk.c
>>>> +++ b/drivers/block/null_blk.c
>>>
>>> [snip]
>>>
>>>> @@ -1422,10 +1440,12 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
>>>>  
>>>>  	blk_mq_start_request(bd->rq);
>>>>  
>>>> -	if (!should_timeout_request(bd->rq))
>>>> -		return null_handle_cmd(cmd);
>>>> +	if (should_requeue_request(bd->rq))
>>>> +		return BLK_STS_RESOURCE;
>>>
>>> Hm, this goes through the less interesting requeue path, add to the
>>> dispatch list and __blk_mq_requeue_request(). blk_mq_requeue_request()
>>> is the one that I wanted to test since that's the one that needs to call
>>> the scheduler hook.
>>
>> Until recently, it would have :-)
>>
>> Both of them are interesting to test, though. Most of the core stall
>> cases would have been triggered by going through the STS_RESOURCE case.
>> How about we just make it exercise both? The below patch alternates
>> between them when we have chosen to requeue.
> 
> Works for me. One idle thought, if we set this up to always requeue,
> then it won't make any progress. Maybe we should limit the number of
> times each request can be requeued so people (me) don't lock up their
> test systems? Either way,

Dunno, that gets into the "doctor it hurts when I shoot myself in the
foot" territory. Same can be said for the timeout setting. I think we
should just ignore that.

-- 
Jens Axboe