Re: [RFC] Cpu-hotplug: Using the Process Freezer (try2)

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Mon, Apr 02, 2007 at 08:56:07PM +0200, Ingo Molnar wrote:
> ok. But the only real problem would be for_each_online_cpu() loops that 
> might sleep, correct? 

I would shy from saying that "that's the only problem". It could also be

	for_each_cpu_mask(some_mask)
		do_something_which_sleeps();

where some_mask is suppsoed to represent a subset of online cpus.

For ex: policy->cpus is a mask maintained by cpufreq, which it iterates thr' 
whenever changing cpu frequency. The cpus in that mask are supposed to be 
a subset of online cpus (with CPU_ONLINE/DEAD handlers in cpufreq.c adjusting 
the mask upon hotplug).

We could adopt a similar trick (get/put_each_cpu_mask) that you describe
below in those extended cases as well.

> the 10% loops that _can_ schedule would trigger the __might_sleep() 
> atomicity test in schedule()), and those would have to be converted a 
> bit more cleverly, on a case by case basis.

The real question is how do we convert over those sleeping
for_each_cpu_mask users (for ex: flush_workqueue) such that they don't
block freezer/hotplug for long periods?

One option is to probably rewrite them to understand that
online[or a derived]_map could have changed everytime they come out of a 
(un)interruptible sleep and deal with arising races appropriately. That
would mean a bit of maintenance headache unfortunately (I was hoping
freezer will lead to zero maintenance headache :)

Besides, how problematic is this in practise (that threads sleep for
extended durations in TASK_INTERRUPTIBLE state breaking
freezer/suspend/hotplug)?

Should we ignore this for the timebeing and take up later as and when
users report problems?

-- 
Regards,
vatsa