Re: 2.6.35-rc2-git2: Reported regressions from 2.6.34

[Jens Axboe <jaxboe@fusionio.com>]

On 11/06/10 21.48, Vivek Goyal wrote:
> On Fri, Jun 11, 2010 at 09:11:49PM +0200, Jens Axboe wrote:
>> On 11/06/10 21.07, Vivek Goyal wrote:
>>> On Fri, Jun 11, 2010 at 11:18:47AM +0200, Jens Axboe wrote:
>>>> On 2010-06-11 10:55, Ingo Molnar wrote:
>>>>>>> Caused by the same blkiocg_update_io_add_stats() function. Bootlog and config 
>>>>>>> attached. Reproducible on that sha1 and with that config.
>>>>>>
>>>>>> I think I see it, the internal CFQ blkg groups are not properly
>>>>>> initialized... Will send a patch shortly.
>>>>>
>>>>> Cool - can test it with a short turnaround, the bug is easy to reproduce.
>>>>
>>>> Here's a nasty patch that should fix it. Not optimal, since we really
>>>> just want empty functions for these when cfq group scheduling is not
>>>> defined.
>>>>
>>>> CC'ing the guilty parties to come up with a better patch that does NOT
>>>> involve ifdefs in cfq-iosched.c. We want blk-cgroup.[ch] fixed up.
>>>> And trimming the CC list a bit.
>>>
>>> Jens, Ingo, I am sorry for this mess.
>>>
>>> Jens,
>>>
>>> How about introducing "block/cfq.h" and declaring additional set of wrapper
>>> functions to update blkiocg stats and make these do nothing if
>>> CFQ_GROUP_IOSCHED=n.
>>>
>>> For example, in linux-2.6/block/cfq.h, we can define functions as follows.
>>>
>>> #ifdef CONFIG_CFQ_GROUP_IOSCHED
>>> cfq_blkiocg_update_dequeue_stats () {
>>> 	blkiocg_update_dequeue_stats()
>>> }
>>> #else
>>> cfq_blkiocg_update_dequeue_stats () {}
>>> #endif
>>>
>>> Fixing it blk-cgroup.[ch] might not be best as BLK_CGROUP is set.
>>> Secondly, if there are other IO control policies later, they might
>>> want to make use of BLK_CGROUP while cfq has disabled the group io 
>>> scheduling.
>>
>> I already tried such a patch, but it's not exactly pretty. How about
>> splitting blk-cgroup.c into two parts, one that is built for
>> BLK_CGROUP and an additional one that is also built for
>> CFQ_GROUP_SCHED? Lets try and improve on the ifdef mess, not extend
>> it.
> 
> Sorry, I did not understand your suggestion. Can you please throw some more
> light on it.
> 
> blk-cgroup.c does not have any cfq specific parts. So I can't split it
> out and build part of it based on CFQ_GROUP_SCHED.

I know they are not cfq specific, but cfq is the only one that calls
them currently. If others depend on them later on, then let that other
blk-cgroup-iosched.o be built for them as well.

-- 
Jens Axboe