Re: [RFC PATCH 07/30] cputime: Convert kcpustat to nsecs

From: Frederic Weisbecker
Date: Mon Dec 01 2014 - 14:59:52 EST


On Mon, Dec 01, 2014 at 06:27:38PM +0100, Martin Schwidefsky wrote:
> On Mon, 1 Dec 2014 18:15:36 +0100 (CET)
> Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> > On Mon, 1 Dec 2014, Martin Schwidefsky wrote:
> > > On Mon, 1 Dec 2014 17:10:34 +0100
> > > Frederic Weisbecker <fweisbec@xxxxxxxxx> wrote:
> > >
> > > > Speaking about the degradation in s390:
> > > >
> > > > s390 is really a special case. And it would be a shame if we prevent from a
> > > > real core cleanup just for this special case especially as it's fairly possible
> > > > to keep a specific treatment for s390 in order not to impact its performances
> > > > and time precision. We could simply accumulate the cputime in per-cpu values:
> > > >
> > > > struct s390_cputime {
> > > > cputime_t user, sys, softirq, hardirq, steal;
> > > > }
> > > >
> > > > DEFINE_PER_CPU(struct s390_cputime, s390_cputime);
> > > >
> > > > Then on irq entry/exit, just add the accumulated time to the relevant buffer
> > > > and account for real (through any account_...time() functions) only on tick
> > > > and task switch. There the costly operations (unit conversion and call to
> > > > account_...._time() functions) are deferred to a rarer yet periodic enough
> > > > event. This is what s390 does already for user/system time and kernel
> > > > boundaries.
> > > >
> > > > This way we should even improve the situation compared to what we have
> > > > upstream. It's going to be faster because calling the accounting functions
> > > > can be costlier than simple per-cpu ops. And also we keep the cputime_t
> > > > granularity. For archs like s390 which have a granularity higher than nsecs,
> > > > we can have:
> > > >
> > > > u64 cputime_to_nsecs(cputime_t time, u64 *rem);
> > > >
> > > > And to avoid remainder losses, we can do that from the tick:
> > > >
> > > > delta_cputime = this_cpu_read(s390_cputime.hardirq);
> > > > delta_nsec = cputime_to_nsecs(delta_cputime, &rem);
> > > > account_system_time(delta_nsec, HARDIRQ_OFFSET);
> > > > this_cpu_write(s390_cputime.hardirq, rem);
> > > >
> > > > Although I doubt that remainders below one nsec lost each tick matter that much.
> > > > But if it does, it's fairly possible to handle like above.
> > >
> > > To make that work we would have to move some of the logic from account_system_time
> > > to the architecture code. The decision if a system time delta is guest time,
> > > irq time, softirq time or simply system time is currently done in
> > > kernel/sched/cputime.c.
> > >
> > > As the conversion + the accounting is delayed to a regular tick we would have
> > > to split the accounting code into decision functions which bucket a system time
> > > delta should go to and introduce new function to account to the different buckets.
> > >
> > > Instead of a single account_system_time we would have account_guest_time,
> > > account_system_time, account_system_time_irq and account_system_time_softirq.
> > >
> > > In principle not a bad idea, that would make the interrupt path for s390 faster
> > > as we would not have to call account_system_time, only the decision function
> > > which could be an inline function.
> >
> > Why make this s390 specific?
> >
> > We can decouple the accounting from the time accumulation for all
> > architectures.
> >
> > struct cputime_record {
> > u64 user, sys, softirq, hardirq, steal;
> > };
> >
> > DEFINE_PER_CPU(struct cputime_record, cputime_record);
> >
> > Now let account_xxx_time() just work on that per cpu data
> > structures. That would just accumulate the deltas based on whatever
> > the architecture uses as a cputime source with whatever resolution it
> > provides.
> >
> > Then we collect that accumulated results for the various buckets on a
> > regular base and convert them to nano seconds. This is not even
> > required to be at the tick, it could be done by some async worker and
> > on idle enter/exit.
>
> And leave the decision making in kernel/sched/cputime.c. Yes, that is good.
> This would make the arch and the account_xxx_time() function care about
> cputime_t and all other common code would use nano-seconds. With the added
> benefit that I do not have to change the low level code too much ;-)

Yes that sounds really good. Besides, this whole machinery can also benefit for
CONFIG_VIRT_CPU_ACCOUNTING_GEN which is the same context entry/exit based
cputime accounting, just it's based on context tracking.

Note the current differences between those two CONFIG_VIRT_CPU_ACCOUNTING
flavours:

_ CONFIG_VIRT_CPU_ACCOUNTING_NATIVE:
* user/kernel boundaries: accumulate
* irq boundaries: account
* context_switch: account
* tick: account (pending user time if any)
* task_cputime(): direct access

_ CONFIG_VIRT_CPU_ACCOUNTING_GEN:
* user/kernel boundaries: account
* irq boundaries: account
* context_switch: account
* tick: ignore
* task_cputime(): direct access + accumulate

More details about task_cputime(): tsk->[us]time are fetched through task_cputime(),
in NATIVE tsk->[us]time are periodically accounted (thanks to the tick)
so task_cputime() simply return the fields as is.
In GEN the tick maybe off so task_cputime() returns tsk->[us]time + pending
accumulated time.

I'm recalling that because Thomas suggests that we don't _have_ to account
the accumulated time from the tick and indeed, this can be done from calls
to task_cputime() like we do for GEN. Of course this comes at the cost of
overhead on any access to utime and stime fields of a task_struct. Thus
the pros and cons must be carefully considered between tick overhead
and task_cputime() overhead, I'd personally be cautious and flush from
the tick at least as a first step.

That is:

_ CONFIG_VIRT_CPU_ACCOUNTING (GEN || NATIVE)
* user/kernel boundaries: accumulate
* irq boundaries: accumulate
* context_switch: account
* tick: account on NATIVE
* task_cputime: return accumulated on GEN

I can take care of that as a preclude before the conversion of cputime to nsecs.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/