Re: [PATCH 08/10] posix-cpu-timers: Migrate to use new tick dependency mask model

[Luiz Capitulino <lcapitulino@redhat.com>]

On Thu, 30 Jul 2015 02:44:45 +0200
Frederic Weisbecker <fweisbec@gmail.com> wrote:

> 
> On Wed, Jul 29, 2015 at 01:24:16PM -0400, Chris Metcalf wrote:
> > On 07/29/2015 09:23 AM, Frederic Weisbecker wrote:
> > >>At a higher level, is the posix-cpu-timers code here really providing the
> > >>>right semantics? It seems like before, the code was checking a struct
> > >>>task-specific state, and now you are setting a global state such that if ANY
> > >>>task anywhere in the system (even on housekeeping cores) has a pending posix
> > >>>cpu timer, then nothing can go into nohz_full mode.
> > >>>
> > >>>Perhaps what is needed is a task_struct->tick_dependency to go along with
> > >>>the system-wide and per-cpu flag words?
> > >That's an excellent point! Indeed the tick dependency check on posix-cpu-timers
> > >was made on task granularity before and now it's a global dependency.
> > >
> > >Which means that if any task in the system has a posix-cpu-timer enqueued, it
> > >prevents all CPUs from shutting down the tick. I need to mention that in the
> > >changelog.
> > >
> > >Now here is the rationale: I expect that nohz full users are not interested in
> > >posix cpu timers at all. The only chance for one to run without breaking the
> > >isolation is on housekeeping CPUs. So perhaps there is a corner case somewhere
> > >but I assume there isn't until somebody reports an issue.
> > >
> > >Keeping a task level dependency check means that we need to update it on context
> > >switch. Plus it's not only about task but also process. So that means two
> > >states to update on context switch and to check from interrupts. I don't think
> > >it's worth the effort if there is no user at all.
> > 
> > I really worry about this!  The vision EZchip offers our customers is
> > that they can run whatever they want on the slow path housekeeping
> > cores, i.e. random control-plane code.  Then, on the fast-path cores,
> > they run their nohz_full stuff without interruption.  Often they don't
> > even know what the hell is running on their control plane cores - SNMP
> > or random third-party crap or god knows what.  And there is a decent
> > likelihood that some posix cpu timer code might sneak in.

I share this thinking. We do the exactly same thing for KVM-RT and I
wouldn't be surprised at all if a posix timer pops up in the
housekeeping CPUs.

> I see. But note that installing a posix cpu timer ends up triggering an
> IPI to all nohz full CPUs. That's how nohz full has always behaved.
> So users running posix timers on nohz should already suffer issues anyway.

I haven't checked how this would affect us, but seems a lot less serious
then not having nohz at all.

> 
> > 
> > You mentioned needing two fields, for task and for process, but in
> > fact let's just add the one field to the one thing that needs it and
> > not worry about additional possible future needs.  And note that it's
> > the task_struct->signal where we need to add the field for posix cpu
> > timers (the signal_struct) since that's where the sharing occurs, and
> > given CLONE_SIGHAND I imagine it could be different from the general
> > "process" model anyway.
> 
> Well, posix cpu timers can be install per process (signal struct) or
> per thread (task struct).
> 
> But we can certainly simplify that with a per process flag and expand
> the thread dependency to the process scope.
> 
> Still there is the issue of telling the CPUs where a process runs when
> a posix timer is installed there. There is no process-like tsk->cpus_allowed.
> Either we send an IPI everywhere like we do now or we iterate through all
> threads in the process to OR all their cpumasks in order to send that IPI.
> 
> > 
> > In any case it seems like we don't need to do work at context switch.
> > Updates to the task's tick_dependency are just done as normal in the
> > task context via "current->signal->". When we are returning to user
> > space and we want to check the tick, again, we can just read via
> > "current->signal->".  Why would we need to copy the value around at
> > task switch time?  That's only necessary if you want to do something
> > like read/write the task tick_dependency via the cpu index, I would think.
> 
> Yeah you're right, at least the context switch should be fine.
> 
> Thanks.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>