On 1/30/18 1:49 PM, Keith Busch wrote:
> On Tue, Jan 30, 2018 at 01:32:25PM -0700, Jens Axboe wrote:
>> On 1/30/18 1:30 PM, Keith Busch wrote:
>>> On Tue, Jan 30, 2018 at 08:57:49AM -0700, Jens Axboe wrote:
>>>>
>>>> Looking at the disassembly, 'n' is 2 and 'segments' is 0xffff.
>>>
>>> Is this still a problem if you don't use an IO scheduler? With deadline,
>>> I'm not finding any path to bio_attempt_discard_merge which is where the
>>> nr_phys_segments is supposed to get it set to 2. Not sure how it could
>>> becmoe 0xffff, though.
>>
>> blk_mq_make_request() -> blk_mq_sched_bio_merge() -> __blk_mq_sched_bio_merge()
>> -> blk_mq_attempt_merge() -> bio_attempt_discard_merge()
>
> That's the calls only if you don't have an elevator_queue, right? With
> deadline, it looks like it goes through this path (ftrace confirms):
>
> __blk_mq_sched_bio_merge() -> dd_bio_merge() -> blk_mq_sched_try_merge()
>
> Which doesn't have a case for ELEVATOR_DISCARD_MERGE.
>
> Relavant function_graph:
>
> 46) | blk_mq_make_request() {
> 46) 0.133 us | blk_queue_bounce();
> 46) 0.370 us | blk_queue_split();
> 46) 0.314 us | bio_integrity_prep();
> 46) 0.081 us | blk_attempt_plug_merge();
> 46) | __blk_mq_sched_bio_merge() {
> 46) | dd_bio_merge() {
> 46) 0.792 us | _raw_spin_lock();
> 46) | blk_mq_sched_try_merge() {
Yeah I guess you are right, it can't happen for mq-deadline since the
generic sched path doesn't support it.
I'll see if I can make it happen again, but I'm not too hopeful. And if
I can't, it's hard to compare with "none" to see if it makes a difference
or not.
--
Jens Axboe
