Re: [BISECTED] 2.6.35.*: horrible (exponential? >linear) slowdown to unusability (HPET)

[Artur Skawina <art.08.09@gmail.com>]

On 09/10/10 00:34, Nix wrote:
> It did: I found a system on which the fault was consistently
> reproducible. Bisected.
> 
> The horrible slowdowns some people are experiencing in 2.6.35 are *not*
> a result of bootmem interfering with the scheduler. They are a result of
> an HPET patch, specifically, this one:
> 
> commit 30a564be9d9554c168a654eddc2165869cc0d7bf
> Author: Thomas Gleixner <tglx@linutronix.de>
> Date:   Tue Apr 13 15:31:36 2010 +0200
> 
>     x86, hpet: Restrict read back to affected ATI chipsets

> On (at least) my ICH10 motherboard (a Tyan S7010), running this commit
> (or later) without hpet=verbose leads to one of two behaviours depending
> on whether or not CONFIG_NO_HZ is on. (This system is using HPET timers
> rather than the TSC even though it is a constant_tsc system, because it
> is an always-on headless server and I wanted it to spend as much time in
> C3 as possible. Why, yes, bisecting a bug on an always-on headless
> server with a dozen client systems *was* a complete pig, why do you
> ask?)

I'm seeing this too, except here it happens every couple of days of uptime,
lasts for a few minutes, and then goes away. Which made bisecting a bit
impractical... Thank you for doing it. 
HW is similar; x64 and X58/82801JI/ICH10, tsc clocksrc.
Did that printk trigger? Empirically confirming that this is the problem
could take weeks here, as it happens so rarely...

artur