Re: Re: [PATCH] block: Add blk_rq_pos(rq) to sort rq when plushing plug-list.

From: Jianpeng Ma
Date: Mon Oct 15 2012 - 22:39:32 EST

On 2012-10-15 21:18 Shaohua Li <shli@xxxxxxxxxx> Wrote:
>2012/10/15 Shaohua Li <shli@xxxxxxxxxxxx>:
>> 2012/10/15 Jianpeng Ma <majianpeng@xxxxxxxxx>:
>>> My workload is a raid5 which had 16 disks. And used our filesystem to
>>> write using direct-io mode.
>>> I used the blktrace to find those message:
>>>
>>> 8,16 0 3570 1.083923979 2519 I W 144323176 + 24 [md127_raid5]
>>> 8,16 0 0 1.083926214 0 m N cfq2519 insert_request
>>> 8,16 0 3571 1.083926586 2519 I W 144323072 + 104 [md127_raid5]
>>> 8,16 0 0 1.083926952 0 m N cfq2519 insert_request
>>> 8,16 0 3572 1.083927180 2519 U N [md127_raid5] 2
>>> 8,16 0 0 1.083927870 0 m N cfq2519 Not idling.st->count:1
>>> 8,16 0 0 1.083928320 0 m N cfq2519 dispatch_insert
>>> 8,16 0 0 1.083928951 0 m N cfq2519 dispatched a request
>>> 8,16 0 0 1.083929443 0 m N cfq2519 activate rq,drv=1
>>> 8,16 0 3573 1.083929530 2519 D W 144323176 + 24 [md127_raid5]
>>> 8,16 0 0 1.083933883 0 m N cfq2519 Not idling.st->count:1
>>> 8,16 0 0 1.083934189 0 m N cfq2519 dispatch_insert
>>> 8,16 0 0 1.083934654 0 m N cfq2519 dispatched a request
>>> 8,16 0 0 1.083935014 0 m N cfq2519 activate rq,drv=2
>>> 8,16 0 3574 1.083935101 2519 D W 144323072 + 104 [md127_raid5]
>>> 8,16 0 3575 1.084196179 0 C W 144323176 + 24 [0]
>>> 8,16 0 0 1.084197979 0 m N cfq2519 complete rqnoidle 0
>>> 8,16 0 3576 1.084769073 0 C W 144323072 + 104 [0]
>>> ......
>>> 8,16 1 3596 1.091394357 2519 I W 144322544 + 16 [md127_raid5]
>>> 8,16 1 0 1.091396181 0 m N cfq2519 insert_request
>>> 8,16 1 3597 1.091396571 2519 I W 144322520 + 24 [md127_raid5]
>>> 8,16 1 0 1.091396934 0 m N cfq2519 insert_request
>>> 8,16 1 3598 1.091397165 2519 I W 144322488 + 32 [md127_raid5]
>>> 8,16 1 0 1.091397477 0 m N cfq2519 insert_request
>>> 8,16 1 3599 1.091397708 2519 I W 144322432 + 56 [md127_raid5]
>>> 8,16 1 0 1.091398023 0 m N cfq2519 insert_request
>>> 8,16 1 3600 1.091398284 2519 U N [md127_raid5] 4
>>> 8,16 1 0 1.091398986 0 m N cfq2519 Not idling. st->count:1
>>> 8,16 1 0 1.091399511 0 m N cfq2519 dispatch_insert
>>> 8,16 1 0 1.091400217 0 m N cfq2519 dispatched a request
>>> 8,16 1 0 1.091400688 0 m N cfq2519 activate rq,drv=1
>>> 8,16 1 3601 1.091400766 2519 D W 144322544 + 16 [md127_raid5]
>>> 8,16 1 0 1.091406151 0 m N cfq2519 Not idling.st->count:1
>>> 8,16 1 0 1.091406460 0 m N cfq2519 dispatch_insert
>>> 8,16 1 0 1.091406931 0 m N cfq2519 dispatched a request
>>> 8,16 1 0 1.091407291 0 m N cfq2519 activate rq,drv=2
>>> 8,16 1 3602 1.091407378 2519 D W 144322520 + 24 [md127_raid5]
>>> 8,16 1 0 1.091414006 0 m N cfq2519 Not idling.st->count:1
>>> 8,16 1 0 1.091414297 0 m N cfq2519 dispatch_insert
>>> 8,16 1 0 1.091414702 0 m N cfq2519 dispatched a request
>>> 8,16 1 0 1.091415047 0 m N cfq2519 activate rq, drv=3
>>> 8,16 1 3603 1.091415125 2519 D W 144322488 + 32 [md127_raid5]
>>> 8,16 1 0 1.091416469 0 m N cfq2519 Not idling.st->count:1
>>> 8,16 1 0 1.091416754 0 m N cfq2519 dispatch_insert
>>> 8,16 1 0 1.091417186 0 m N cfq2519 dispatched a request
>>> 8,16 1 0 1.091417535 0 m N cfq2519 activate rq,drv=4
>>> 8,16 1 3604 1.091417628 2519 D W 144322432 + 56 [md127_raid5]
>>> 8,16 1 3605 1.091857225 4393 C W 144322544 + 16 [0]
>>> 8,16 1 0 1.091858753 0 m N cfq2519 complete rqnoidle 0
>>> 8,16 1 3606 1.092068456 4393 C W 144322520 + 24 [0]
>>> 8,16 1 0 1.092069851 0 m N cfq2519 complete rqnoidle 0
>>> 8,16 1 3607 1.092350440 4393 C W 144322488 + 32 [0]
>>> 8,16 1 0 1.092351688 0 m N cfq2519 complete rqnoidle 0
>>> 8,16 1 3608 1.093629323 0 C W 144322432 + 56 [0]
>>> 8,16 1 0 1.093631151 0 m N cfq2519 complete rqnoidle 0
>>> 8,16 1 0 1.093631574 0 m N cfq2519 will busy wait
>>> 8,16 1 0 1.093631829 0 m N cfq schedule dispatch
>>>
>>> Because in func "elv_attempt_insert_merge", it only to try to
>>> backmerge.So the four request can't merge in theory.
>>> I trace ten minutes and count those situation, it can count 25%.
>>>
>>> With the patch,i tested and not found situation like above.
>>>
>>> Signed-off-by: Jianpeng Ma <majianpeng@xxxxxxxxx>
>>> ---
>>> block/blk-core.c | 3 ++-
>>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/block/blk-core.c b/block/blk-core.c
>>> index a33870b..3c95c4d 100644
>>> --- a/block/blk-core.c
>>> +++ b/block/blk-core.c
>>> @@ -2868,7 +2868,8 @@ static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
>>> struct request *rqa = container_of(a, struct request, queuelist);
>>> struct request *rqb = container_of(b, struct request, queuelist);
>>>
>>> - return !(rqa->q <= rqb->q);
>>> + return !(rqa->q < rqb->q ||
>>> + (rqa->q == rqb->q && blk_rq_pos(rqa) < blk_rq_pos(rqb)));
>>> }
>>
>> Does this one help too?
>> http://marc.info/?l=linux-kernel&m=132399972114668&w=2
>>
>> I thought the real problem is we don't do recursive request
>> merge. I had no objection to the patch itself, but just hope we
>> can make recursive merge work, which is more generic.
>
>Oh, wait, the 4 requests aren't merged completely in your case.
>And the requests are from one thread and plug context.
>Not the issue I mentioned. I'm wondering how this could happen.
>they should be merged in attempt_plug_merge already.
Yes, i ignore this. The real problem is why not merge in attempt_plug_mege?
I'll do further study.
Thanks!
Jianpeng韬{.n?????%?lzwm?b?Р骒r?zXЩ??{ay????j?f＂?????ア?⒎?:+v???????赙zZ+????"?!?O???v??m?鹈 n?帼Y&—