lkml.org 
[lkml]   [2017]   [Apr]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH] kernel: sched: Provide a pointer to the valid CPU mask

* Thomas Gleixner <tglx@linutronix.de> wrote:

> On Thu, 6 Apr 2017, Ingo Molnar wrote:
>
> > Sorry if this is a back and forth - I was somehow convinced that we do need to
> > frob the cpus_allowed mask to get this functionality - but in hindsight I
> > think the counter should be enough.
> >
> > I.e. just have a counter and these two APIs:
> >
> > static inline void migrate_disable(void)
> > {
> > current->migration_disabled++;
> > }
> >
> > ...
> >
> > static inline void migrate_enable(void)
> > {
> > current->migration_disabled--;
> > }
> >
> > ... and make sure the scheduler migration code plus the CPU hotplug code considers
> > the counter.
>
> We tried that some time ago, but there are a lot of places in the scheduler
> which just rely on the cpus_allowed_mask, so we have to chase all of them and
> make sure that new users do not ignore that counter. That's why we chose the
> cpus_allowed_mask approach. And I still think that's the proper thing to do.

But but ...

The number of places in the scheduler where we actually end up migrating a task is
pretty limited:

try_to_wake_up():
- main wakeup code

migrate_swap():
- active NUMA-balancing feature

move_queued_task():
- hotplug CPU-down migration
- changing the affinity mask

The wakeup and NUMA balancing case is trivial to solve: it's an optimization and
we can skip the migration if migration is disabled.

CPU hotplug and changing the affinity mask are the more complex cases, because
there migrating or not migrating is a correctness issue:

- CPU hotplug has to be aware of this anyway, regardless of whether it's solved
via a counter of the affinity mask.

- Changing the affinity mask (set_cpus_allowed()) has two main cases:
the synchronous and asynchronous case:

- synchronous is when the current task changes its own affinity mask, this
should work fine mostly out of box, as we don't call set_cpus_allowed()
from inside migration disabled regions. (We can enforce this via a
debugging check.)

- The asynchronous case is when the affinity task of some other task is
changed - this would not have an immediate effect with migration-disabled
logic, the migration would be delayed to when migration is re-enabled
again.

As for general fragility, is there any reason why a simple debugging check in
set_task_cpu() would not catch most mishaps:

WARN_ON_ONCE(p->state != TASK_RUNNING && p->migration_disabled);

... or something like that?

I.e. my point is that I think using a counter would be much simpler, yet still as
robust and maintainable. We could in fact move migrate_disable()/enable() upstream
straight away and eliminate this small fork of functionality between mainline and
-rt.

Thanks,

Ingo

\
 
 \ /
  Last update: 2017-04-06 13:04    [W:0.256 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site