Re: [PATCH 1/2 v2.0]sched: Removing unused fields from /proc/schedstat

[Ciju Rajan K <ciju@linux.vnet.ibm.com>]

Hi Satoru,


>> I think we can remove rq->sched_switch and rq->sched_switch
>> without no problem because they are meaningless. The former
>> is for old O(1) scheduler and means the number of runqueue
>> switching among active/expired queue. The latter is for
>> SD_WAKE_BALANCE flag and its logic is already gone.
>>
>> However sbe_* are for SD_BALANCE_EXEC flag and sbf_* are for
>> SD_BALANCE_FORK flag. Since both logic for them are still alive,
>> the absence of these accounting is regression in my perspective.
>> In addition, these fields would be useful for analyzing load
>> balance behavior.
>>

The sbe_* & sbf_* counters were added by the commit 
68767a0ae428801649d510d9a65bb71feed44dd1 But it was subsequently 
removed by the commit 476d139c218e44e045e4bc6d4cc02b010b343939

[ciju@ciju kernel]$ git describe 68767a0ae428801649d510d9a65bb71feed44dd1 --contains
v2.6.13-rc1~68^2~148
[ciju@ciju kernel]$ git describe 476d139c218e44e045e4bc6d4cc02b010b343939 --contains
v2.6.13-rc1~68^2~140

So.. it was introduced and removed in 2.6.13 time frame


When the counters were removed the sbe_* sbf_* variable
declarations were not removed. Hence it caused a little confusion.
So I believe these stats were not available and hence can't be 
considered as regression. 

476d139c218e44e045e4bc6d4cc02b010b343939 consolidated the fork and 
exec balance. Thereafter it became non-trivial to provide separate 
stats for fork and exec events. So if people think a consolidated  
balance-on-event is needed, it can be looked into separately. But 
that shouldn't prevent this documentation cleanup patch from 
getting in.

-Ciju