Date:	Mon, 21 Jan 2013 18:09:39 -0600
From:	Matt Sealey <matt@...esi-usa.com>
To:	Russell King - ARM Linux <linux@....linux.org.uk>
Cc:	John Stultz <john.stultz@...aro.org>,
	Arnd Bergmann <arnd@...db.de>,
	Linux ARM Kernel ML <linux-arm-kernel@...ts.infradead.org>,
	LKML <linux-kernel@...r.kernel.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Ingo Molnar <mingo@...hat.com>
Subject: Re: One of these things (CONFIG_HZ) is not like the others..

Okay so the final resolution of this is;

1) That the arch/arm/Kconfig HZ block is suffering from some cruft

I think we could all be fairly confident that Exynos4 or S5P does not
require HZ=200 - in theory, it has no such timer restrictions like
EBSA110 (the docs I have show a perfectly capable 32-bit timer with a
double-digits MHz input clock, since these are multimedia-class SoCs
it'd be seriously f**ked up if they didn't).

But while some of the entries on this line may be cargo-cult
programming, the original addition on top of EBSA110 *may* be one of
your "unreported" responsiveness issues.

We could just let some Samsung employees complain when Android 6.x
starts to get laggy with a 3.8 kernel because we forced their HZ=100.

What I would do is predicate a fixed, obvious default on
ARCH_MULTIPLATFORM so that it would get the benefit of a default HZ
that you agree with. It wouldn't CHANGE anything, but it makes it look
less funky, since the non-multiplatform settings would be somewhere
else (it either needs more comments or an if - either way - otherwise
it's potentially confusing);

if ARCH_MULTIPLATFORM
config HZ
   int
   default 100
else
   # old config HZ block here
endif

2) We need to add config SCHED_HRTICK as a copy and paste from
kernel/Kconfig.hz since.. well, I still don't understand exactly what
the true effect would be, but I assume since Arnd is concerned and
John's explanation rings true that it really should be enabled on ARM
systems with the exact same dependencies as kernel/Kconfig.hz.

Or not.. I see it as an oddity until I understand if we really care
about it, but the code seems to be fairly important to the scheduler
and also enabled by default almost everywhere else, which means only
people with really freakish SMP architectures with no ability to use
GENERIC_SMP_HELPERS have ever run these code paths besides ARM. That
kind of leaves ARM in the doghouse.. who knows what weirdo scheduler
reactions are related to it not being enabled. Maybe when it is, HZ
*would* need to be allowed to be bumped when using this code path?
Matt Sealey <matt@...esi-usa.com>
Product Development Analyst, Genesi USA, Inc.


On Mon, Jan 21, 2013 at 5:49 PM, Russell King - ARM Linux
<linux@....linux.org.uk> wrote:
> On Mon, Jan 21, 2013 at 05:23:33PM -0600, Matt Sealey wrote:
>> On Mon, Jan 21, 2013 at 4:42 PM, Russell King - ARM Linux
>> <linux@....linux.org.uk> wrote:
>> > On Mon, Jan 21, 2013 at 04:20:14PM -0600, Matt Sealey wrote:
>> >> I am sorry it sounded if I was being high and mighty about not being
>> >> able to select my own HZ (or being forced by Exynos to be 200 or by
>> >> not being able to test an Exynos board, forced to default to 100). My
>> >> real "grievance" here is we got a configuration item for the scheduler
>> >> which is being left out of ARM configurations which *can* use high
>> >> resolution timers, but I don't know if this is a real problem or not,
>> >> hence asking about it, and that HZ=100 is the ARM default whether we
>> >> might be able to select that or not.. which seems low.
>> >
>> > Well, I have a versatile platform here.  It's the inteligence behind
>> > the power control system for booting the boards on the nightly tests
>> > (currently disabled because I'm waiting for my main server to lock up
>> > again, and I need to use one of the serial ports for that.)
>> >
>> > The point is, it talks via I2C to a load of power monitors to read
>> > samples out.  It does this at sub-100Hz intervals.  Yet the kernel is
>> > built with HZ=100.  NO_HZ=y and highres timers are enabled... works
>> > fine.
>> >
>> > So, no, HZ=100 is not a limit in that scenario.  With NO_HZ=y and
>> > highres timers, it all works with epoll() - you get the interval that
>> > you're after.  I've verified this with calls to gettimeofday() and
>> > the POSIX clocks.
>>
>> Okay.
>>
>> So, can you read this (it's short):
>>
>> http://ck.kolivas.org/patches/bfs/bfs-configuration-faq.txt
>>
>> And please tell me if he's batshit crazy and I should completely
>> ignore any scheduler discussion that isn't ARM-specific, or maybe..
>> and I can almost guarantee this, he doesn't have an ARM platform so
>> he's just delightfully ill-informed about anything but his quad-core
>> x86?
>
> Well... my x86 laptop is... HZ=1000, NO_HZ, HIGH_RES enabled, ondemand...
> doesn't really fit into any of those categories given there.  I'd suggest
> that what's given there is a suggestion/opinion based on behaviours
> observed on x86 platforms.
>
> Whether it's appropriate for other architectures is not really a proven
> point - is it worth running ARM at 1000Hz when the load from running at
> 100Hz is measurable as a definite error in loops_per_jiffy calibration?
> Remember - the load from the interrupt handler at 1000Hz is 10x the load
> at 100Hz.
>
> Do you want to spend more cycles per second on the possibly multi-layer
> IRQ servicing and timer servicing?
>
> And what about the interrupt latency issue that we've hit several times
> already with devices taking longer than 10ms to service their peripherals
> because the driver doesn't make use of delayed works/tasklets/etc.
>
> The lack of reasonable device DMA too has an impact for many drivers - the
> CPU has to spend more time in interrupt handlers (which are now run to the
> exclusion of any other interrupt in the system) performing PIO - or in the
> case of those systems which _do_ have DMA, they may end up having to do
> cache maintanence over large cache ranges from IRQ context which x86
> doesn't have to do.
>
> There's many factors here, and the choice of what the right HZ is for a
> platform is not as clear cut as one may think.  Given all the additional
> overheads we have on ARM because of the lack of memory coherency, the
> generally bad DMA support, etc, I think what we currently have is still
> right as an architecture default - 100Hz.
>
>> I did test it.. whatever you define last, sticks, and it's down to the
>> order they're parsed in the tree - luckily, arch/arm/Kconfig is
>> sourced first, which sources the mach/plat stuff way down at  the
>> bottom. As long as you have your "default" set somewhere, any further
>> default just has to be sourced or added later in *one* of the
>> Kconfigs, same as building any C file with "gcc -E" and spitting it
>> out.
>>
>> Someone, at the end of it all, has to set some default, and as long as
>> the one you want is the last one, everything is shiny.
>
> Actually, we're both wrong.  There seems to be two things which
> inflence it, and it basically comes down to this:
>
> - the value a particular symbol has comes from the _first_ declaration
>   which a value is assigned to a symbol.
>
> So:
>
> config HZ
>         int
>         default 300
>
> config HZ
>         int
>         default 100 if OPT1
>         default 200 if OPT2
>         default 400
>
> takes on the value of 300 no matter what combination of OPT1 and OPT2
> are enabled.
>
> config HZ
>         int
>         default 100 if OPT1
>         default 200 if OPT2
>         default 400
>
> config HZ
>         int
>         default 300
>
> never takes the value 300, but 100, 200 or 400.
>
> config HZ
>         int
>         default 100 if OPT1
>         default 200 if OPT2
>
> config HZ
>         int
>         default 300
>
> Will now take 100, 200, or 300 depending on which of OPT1/OPT2 are enabled.
>
> So, we _can_ use kernel/Kconfig.hz, but it's not very nice at all: we will
> be presenting users with configutation options for the HZ value which will
> be _silently_ ignored by Kconfig if we have a platform which overrides this.
>
> Probably fine if you think that Kconfig is a developers tool and you edit
> the configuration files (and therefore you expect them to know what they're
> doing, and how this stuff works), but not if you think that Kconfig users
> should be presented with meaningful options when configuring their kernel.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists