On 03/31/2015 09:33 AM, Frederic Weisbecker wrote:
> On Tue, Mar 31, 2015 at 09:07:11AM -0600, Jens Axboe wrote:
>> On 03/31/2015 08:27 AM, Rik van Riel wrote:
>>> CPUs with nohz_full do not want disruption from timer interrupts,
>>> or other random system things.  This includes block mq work.
>>>
>>> There is another issue with block mq vs. realtime tasks that run
>>> 100% of the time, which is not uncommon on systems that have CPUs
>>> dedicated to real time use with isolcpus= and nohz_full=
>>>
>>> Specifically, on systems like that, a block work item may never
>>> get to run, which could lead to filesystems getting stuck forever.
>>>
>>> We can avoid both issues by not scheduling blk-mq workqueues on
>>> cpus in nohz_full mode.
>>>
>>> Question for Jens: should we try to spread out the load for
>>> currently offline and nohz CPUs across the remaining CPUs in
>>> the system, to get the full benefit of blk-mq in these situations?
>>>
>>> If so, do you have any preference on how I should implement that?
>>>
>>> Cc: Frederic Weisbecker <fweisbec@redhat.com>
>>> Cc: Ingo Molnar <mingo@kernel.org>
>>> Cc: Jens Axboe <axboe@kernel.org>
>>> Signed-off-by: Rik van Riel <riel@redhat.com>
>>> ---
>>>   block/blk-mq.c | 5 +++++
>>>   1 file changed, 5 insertions(+)
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index 4f4bea21052e..1004d6817fa4 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -21,6 +21,7 @@
>>>   #include <linux/sched/sysctl.h>
>>>   #include <linux/delay.h>
>>>   #include <linux/crash_dump.h>
>>> +#include <linux/tick.h>
>>>
>>>   #include <trace/events/block.h>
>>>
>>> @@ -1760,6 +1761,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
>>>   		if (!cpu_online(i))
>>>   			continue;
>>>
>>> +		/* Do not schedule work on nohz full dedicated CPUs. */
>>> +		if (tick_nohz_full_cpu(i))
>>> +			continue;
>>
>> Is this CPU ever going to queue IO? If yes, then it needs to be mapped. If
>> userspace never runs on it and submits IO, then we'll never run completions
>> on it nor schedule the associated workqueue. So I really don't see how it
>> doesn't already work, as-is.
>
> Well, it's fairly possible that full dynticks CPUs do IO of any sort. Is it possible
> to affine these asynchronous works to specific CPU? The usual scheme of full dynticks
> is to have CPU 0 handling any kind of housekeeping and other CPUs doing latency or performance
> sensitive works that don't want to be disturbed.

That'd be easy enough to do, that's how blk-mq handles offline CPUs as 
well. The attached patch is completely untested, but will handle offline 
or nohz CPUs in the same fashion - they will punt to hardware queue 0, 
which is mapped to CPU0 (and others, depending on the queue vs CPU ratio).

-- 
Jens Axboe