Re: [RFC][PATCH] i386 x86-64 Eliminate Local APIC timer interrupt

[Andi Kleen]

On Fri, Apr 29, 2005 at 05:26:05PM -0700, Venkatesh Pallipadi wrote:
> 
> Background: 
> Local APIC timer stops functioning when CPU is in C3 state. As a
> result the local APIC timer interrupt will fire at uncertain times, depending
> on how long we spend in C3 state. And this has two side effects
> * Idle balancing will not happen as we expect it to.
> * Kernel statistics for idle time will not be proper (as we get less LAPIC
>   interrupts when we are idle). This can result in confusing other parts of
>   kernel (like ondemand cpufreq governor) which depends on this idle stats.

This is absolutely incompatible to no timer tick in IDLE - at least
my implementation that allows to keep all CPUs indepently in idle,
not require all CPUs to be idle to turn off the timers.

I had hoped that the hardware was clever enough to keep the
APIC timers working on SMP systems.

This is very unfortunate :-(( It means you will always 
waste a lot of power on a partially idle system, because
you cannot keep individual CPUs sleeping for a longer time.

I ran into the problem on AMD systems with C2 state - in 
my initial implementation there was only the APIC timer for
any time keeping because it is cheaper to reprogram and
there are a lot of chipsets which miss HPET - and on the
BP only the APIC timer was ticking at all, which did not 
work very well. This was avoided by using
PIT/HPET on the BP again. But on SMP systems there is not
much choice on the APs.



> it simplifies things and will have other advantages:
> * Should help dynamick tick as one has to change only global timer interrupt 
>   freq with varying jiffies.

Huh? On contrary it completely kills really SMP aware dynamic tick
implementations.

I would really prefer to find some other way to solve this

Perhaps if we have enough timer sources in HPET we can use the 3-4
HPET timers on different CPUs. That would work on upto 4 socket
systems (sibling/dual core can be ignored I guess because for
power management purposes they dont really exist) 

This would be a problem for the few applications that use HPET
from user space for application timers, but I  think power saving
is more important than them or they can disable dynamic tick.

It is a big mess. I begin to envy the PPC guys who
actually have usable per CPU timers.

-Andi


-
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/