Re: Re: [PATCH 1/5] ftrace: Synchronize variable setting with breakpoints

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

(2012/06/01 5:49), Steven Rostedt wrote:
> On Thu, 2012-05-31 at 22:40 +0200, Peter Zijlstra wrote:
>> On Thu, 2012-05-31 at 16:37 -0400, Steven Rostedt wrote:
>>> On Thu, 2012-05-31 at 22:26 +0200, Peter Zijlstra wrote:
>>>
>>>> Right, but when you loose stop-machine you could simply do 30k
>>>> kmap_atomic/kunmap_atomic's consecutively since you're not holding
>>>> anybody up.
>>>
>>> It requires 3 IPIs per update too. Thus that's 90,000 IPIs you are
>>> blasting^Wsending to all CPUs.
>>
>> Uhm, no. 
>>
> 
> ----------------------------------+
>> for_each() {			|
>>   kmap_atomic()			|
>>   frob int3			|
>>   kunmap_atomic();		|
>> }				|
>> 				|
>> sync-ipi-broadcast();		+--- Break points applied
>> 				|
>> for_each() {			|
>>   kmap_atomic();		|
>>   frob tail			|
>>   kunmap_atomic();		|
>> }				|
> ----------------------------------+
> 
> Note, for the above time, the entire kernel has breakpoints added, and
> every function is taking a hit due to it. By slowing down this process,
> the rest of the system *will* be impacted. Ideally, we want to finish it
> as quick as possible. Not to mention, the kmap_atomics will be slowed
> down by the breakpoints that are attached to them.

Hmm, why don't we have two text_poke interfaces for single and
batch? As like dyn_ftrace, since modifying massive points takes
a lot time, so we may have additional kconfig something like
CONFIG_QUICK_BATCH_TEXT_POKE which switches text area to rw while
batch text_poke.

Thank you,

-- 
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com