Re: [PATCH RFC] sched: Micro optimization in pick_next_task() and in check_preempt_curr()

[Kirill Tkhai]

On 19.12.2019 17:25, Kirill Tkhai wrote:
> On 19.12.2019 17:02, Peter Zijlstra wrote:
>> On Thu, Dec 19, 2019 at 02:12:42PM +0100, Peter Zijlstra wrote:
>>> On Thu, Dec 19, 2019 at 03:39:14PM +0300, Kirill Tkhai wrote:
>>>> In kernel/sched/Makefile files, describing different sched classes, already
>>>> go in the order from the lowest priority class to the highest priority class:
>>>>
>>>> idle.o fair.o rt.o deadline.o stop_task.o
>>>>
>>>> The documentation of GNU linker says, that section appears in the order
>>>> they are seen during link time (see [1]):
>>>>
>>>>> Normally, the linker will place files and sections matched by wildcards
>>>>> in the order in which they are seen during the link. You can change this
>>>>> by using the SORT keyword, which appears before a wildcard pattern
>>>>> in parentheses (e.g., SORT(.text*)).
>>>>
>>>> So, we may expect const variables from idle.o will go before ro variables
>>>> from fair.o in RO_DATA section, while ro variables from fair.o will go
>>>> before ro variables from rt.o, etc.
>>>>
>>>> (Also, it looks like the linking order is already used in kernel, e.g.
>>>>  in drivers/md/Makefile)
>>>>
>>>> Thus, we may introduce an optimization based on xxx_sched_class addresses
>>>> in these two hot scheduler functions: pick_next_task() and check_preempt_curr().
>>>>
>>>> One more result of the patch is that size of object file becomes a little
>>>> less (excluding added BUG_ON(), which goes in __init section):
>>>>
>>>> $size kernel/sched/core.o
>>>>          text     data      bss	    dec	    hex	filename
>>>> before:  66446    18957	    676	  86079	  1503f	kernel/sched/core.o
>>>> after:   66398    18957	    676	  86031	  1500f	kernel/sched/core.o
>>>
>>> Does LTO preserve this behaviour? I've never quite dared do this exact
>>> optimization.
>>
>> Also, ld.lld seems a popular option.
> 
> I asked on their IRC. Oh, it looks like no way is for this.
> 
> About the link: https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/4/html/Using_ld_the_GNU_Linker/sections.html
> 
> (17:19:25) nbjoerg: but it is not guarenteed behavior
> (17:19:50) nbjoerg: if for some strange reason you really need to enforce relative orders of global objects, put them in consecutively named sections

Introduction of sched_class::id instead of this patch's approach does not have a big sense,
since this will help in check_preempt_curr() only. And this requires too many new lines of code.