Changing the default scheduler

[Frederic Weisbecker]

(Sorry I forgot to Cc kernelnewbies, I fixed the Cc line).

On Tue, May 24, 2016 at 11:47:37AM +0000, Renato Utsch wrote:
> Hi Frederic,
> 
> Thanks, this was exactly what I was looking for! I was suspecting that all
> the processes would inherit the policy of the init process, but I didn't
> know where to look at. I'll take a look at this and see what happens.
> 
> By the way, how are the scheduler policies and the schedulers related?
> Because the RT scheduler can be used with either SCHED_FIFO or SCHED_RR.
> Are some policies linked to particular schedulers? If so, can I introduce
> new SCHED_* constants? How should I do that if I want to call
> sched_setscheduler() to change the policy to, say, SCHED_MY in an userspace
> program?

What you refer as "scheduler" is in fact a scheduler class.
So we have two different notions here: scheduler policy and scheduler class.

_ The scheduler policy is set by the user or the system to tell how to schedule a task.
  That's in the user interface level.

_ The scheduler classes handle the tasks scheduling differently on top of their scheduler policy (driven
  by the user) or on top of kernel internal needs (idle_sched_class and stop_sched_class).

  We have 4 of them, from lowest prio to highest:

     _ idle_sched_class = implement idle tasks. The idle tasks (one per CPU)
                          are set with SCHED_NORMAL but the policy of idle tasks are
			  not even checked.

     _ fair_sched_class = implement SCHED_NORMAL/SCHED_FAIR and SCHED_IDLE (not to be confused
                          with idle tasks. In fact SCHED_IDLE tasks are normal tasks that execute
			  when there is no SCHED_NORMAL tasks to run)

     _ rt_sched_class   = implement SCHED_RR and SCHED_FIFO

     _ dl_sched_class   = deadline realtime tasks (SCHED_DEADLINE)

     _ stop_sched_class = stop machine tasks. They don't refer to any policy, the stop machine
                          tasks call sched_set_stop_task() to switch to this class. Their policy
                          are SCHED_NORMAL but it's ignored.

   That's in the implementation level.

When the scheduler needs to choose the next task to run on the CPU, it iterates each class
from highest prio to lower: (stop, dl, rt, fair, idle) and calls the ->pick_next_task()
callback for each of these classes. The first one that has a task to run will have it on
the CPU. This is how class priorities are implemented.

So if you want to implement a new policy, the easiest is to create a new scheduler class
but you'll need to decide where it fits between the classes prio.

Now when people talk about re-implementing the scheduler, they usually mean the way to schedule
the SCHED_NORMAL tasks. If that's what you're interested in, I think you rather want to rewrite
kernel/sched/fair.c and do your own fair_sched_class implementation.

Now kernel/sched/fair.c does a lot more than just implement fair_sched_class so rewriting it the
way you like to experiment will be painful due to the links it has all around in the core kernel.
Perhaps it's easier to create your own policy, define its prio between idle_sched_class and
fair_sched_class then run your task under that policy on a CPU that doesn't have any SCHED_NORMAL
tasks.

You can also set the init/0 task to your policy but there is no guarantee it will propagate all
along recursively to the whole process tree as anything can override a parent back to SCHED_NORMAL
at some point.

Thanks.

> 
> Em ter, 24 de mai de 2016 às 08:39, Frederic Weisbecker <fweisbec at gmail.com>
> escreveu:
> 
> > Hi Renato,
> >
> > What you're talking about is not exactly the scheduler but the scheduler
> > policy. Basically the scheduler policy is inherited on fork(). The very
> > first task (init/0) is initialized with SCHED_NORMAL and this policy is
> > inherited recursively through all subsequent fork() calls as it is the root
> > parent task. If you want all tasks in the system to inherit a SCHED_FIFO
> > policy, you need to change the policy of init/0. This is probably feasible
> > in include/linux/init_task.h (I can't check that right now). Now since
> > every task will have a real time policy, some of them may starve others.
> > You may observe surprising behaviour,  your system could even fail to boot.
> >
> > Have fun.
> > Le 20 mai 2016 01:51, "Renato Utsch" <renatoutsch at gmail.com> a écrit :
> >
> >> Hello,
> >>
> >> I am a new developer trying to learn how to tinker with the kernel. I
> >> searched on the internet but couldn't find much info about this (and
> >> couldn't find any info up to date).
> >>
> >> My question is, how does the kernel decide which is the default scheduler
> >> that all processes start with? I can change the scheduler of a process by
> >> sched_setscheduler(), but how do I change *all* processes from using the
> >> CFS scheduler to, for example, the RR scheduler?
> >>
> >> Sorry if this is too basic, but I don't know where to search for this. If
> >> you guys could point me places where I can learn more about this, I would
> >> be grateful.
> >>
> >> _______________________________________________
> >> Kernelnewbies mailing list
> >> Kernelnewbies at kernelnewbies.org
> >> http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
> >>
> >>