Re: 2.6.35.*: horrible (exponential? >linear) slowdown to unusability (ACPI idle?)

[Damien Wyart <damien.wyart@free.fr>]

Hello,

* Nix <nix@esperi.org.uk> [2010-09-05 01:51]:
> This is a preliminary report: I haven't bisected this problem yet, or
> isolated it in any significant fashion.

> What I can say is this.

> In 2.6.35.3, but not in 2.6.34.1, some of my systems experience an
> increasingly severe series of stalls whenever the system is idle. The
> stalls get worse at a rate above linear: about fifty seconds after they
> start, I am waiting five to ten seconds for responses to keystrokes. It
> is very hard to get anything much done in this situation: even
> restarting is hard. 'perf top' shows nothing using the time, but the
> load average is pegged at 1.

> I can verify that this does not happen on my embedded Geode box, a
> 32-bit system without ACPI support. It does happen on my x86-64 systems,
> all of which are running 64-bit kernels with ACPI. These systems are all
> running with CONFIG_NOHZ, so I tried turning it off. The problem got
> enormously worse: the kernel lasted 6.4 seconds after initial boot and
> about two seconds after entering userspace before stalling completely.

I think this is related to what has been discussed in this thread:
http://lkml.org/lkml/2010/7/14/11

I've not seen the problem with 2.6.35.4 yet, and only once with
2.6.35.3. Zeno said it disappeared for him when he started to use
2.6.35-stable.

So if you can reproduce it quite easily, I guess bisection (even
painful) will be the best way to get (hopefuly) an idea of where the
problem might come from...

-- 
Damien Wyart