Re: [PATCH] xen/blkfront: restart request queue when there is enough persistent_gnts_c

[Roger Pau Monné]

El 12/01/15 a les 14.04, David Vrabel ha escrit:
> On 12/01/15 11:36, Roger Pau Monné wrote:
>> El 12/01/15 a les 8.09, Bob Liu ha escrit:
>>>
>>> On 01/09/2015 11:51 PM, Roger Pau Monné wrote:
>>>> El 06/01/15 a les 14.19, Bob Liu ha escrit:
>>>>> When there is no enough free grants, gnttab_alloc_grant_references()
>>>>> will fail and block request queue will stop.
>>>>> If the system is always lack of grants, blkif_restart_queue_callback() can't be
>>>>> scheduled and block request queue can't be restart(block I/O hang).
>>>>>
>>>>> But when there are former requests complete, some grants may free to
>>>>> persistent_gnts_c, we can give the request queue another chance to restart and
>>>>> avoid block hang.
>>>>>
>>>>> Reported-by: Junxiao Bi <junxiao.bi@xxxxxxxxxx>
>>>>> Signed-off-by: Bob Liu <bob.liu@xxxxxxxxxx>
>>>>> ---
>>>>>  drivers/block/xen-blkfront.c |   11 +++++++++++
>>>>>  1 file changed, 11 insertions(+)
>>>>>
>>>>> diff --git a/drivers/block/xen-blkfront.c b/drivers/block/xen-blkfront.c
>>>>> index 2236c6f..dd30f99 100644
>>>>> --- a/drivers/block/xen-blkfront.c
>>>>> +++ b/drivers/block/xen-blkfront.c
>>>>> @@ -1125,6 +1125,17 @@ static void blkif_completion(struct blk_shadow *s, struct blkfront_info *info,
>>>>>  			}
>>>>>  		}
>>>>>  	}
>>>>> +
>>>>> +	/*
>>>>> +	 * Request queue would be stopped if failed to alloc enough grants and
>>>>> +	 * won't be restarted until gnttab_free_count >= info->callback->count.
>>>>> +	 *
>>>>> +	 * But there is another case, once we have enough persistent grants we
>>>>> +	 * can try to restart the request queue instead of continue to wait for
>>>>> +	 * 'gnttab_free_count'.
>>>>> +	 */
>>>>> +	if (info->persistent_gnts_c >= info->callback.count)
>>>>> +		schedule_work(&info->work);
>>>>
>>>> I guess I'm missing something here, but blkif_completion is called by
>>>> blkif_interrupt, which in turn calls kick_pending_request_queues when
>>>> finished, which IMHO should be enough to restart the processing of requests.
>>>>
>>>
>>> You are right, sorry for the mistake.
>>>
>>> The problem we met was a xenblock I/O hang.
>>> Dumped data showed at that time info->persistent_gnt_c = 8, max_gref = 8
>>> but block request queue was still stopped.
>>> It's very hard to reproduce this issue, we only see it once.
>>>
>>> I think there might be a race condition:
>>>
>>> request A                          request B:
>>>
>>>                                    info->persistent_gnts_c < max_grefs
>>>                                    and fail to alloc enough grants
>>>
>>>
>>> ^^^^
>>> interrupt happen, blkif_complte():
>>> info->persistent_gnts_c++
>>> kick_pending_request_queues()

blkif_interrupt can never interrupt blkif_queue_request, because it's
holding a spinlock (info->io_lock). If you have seen this trace in the
wild it means something is really wrong and we are calling
blkif_queue_request without acquiring the spinlock and thus without
disabling interrupts.

>>>
>>>                                     stop block request queue
>>>                                     added to callback list
>>>
>>> If the system don't have enough grants(but have enough persistent_gnts),
>>> request queue would still hang.
>>
>> Not sure how can this happen, blkif_queue_request explicitly checks that
>> persistent_gnts_c < max_grefs before adding the callback and stopping
>> the request queue, so in your case the queue should not be blocked. Can
>> you dump the state of info->connected?
> 
> I think Bob has correctly identified a race.
> 
> After calling blk_stop_queue(), check info->persistent_gnts again and
> restart the queue if there free grefs.
> 
> David
> 
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/