Re: current linux-2.6.git: cpusets completely broken

[Linus Torvalds <torvalds@linux-foundation.org>]

On Sat, 12 Jul 2008, Max Krasnyansky wrote:
> 
> My vote goes for Dmitry's patch. The one with the full switch() statement.
> Your simplified version with if() is correct (I think) but the switch() is
> more explicit about what events are being processed.

Well, I still haven't seen a combined patch+signoff+good explanation, so I 
can't really commit it.

> The cpu_active_map thing seems like an overkill. In a sense that we should not
> try to add a new map for every such case. Granter this migration case may be
> special enough to warrant the new map but in general I think it's not the
> right way to go.

Note how cpu_active_map has nothing to do with cpusets per se, and 
everything to do with the fact that CPU migration currently seems to be 
fundamentally flawed in the presense of a CPU hotunplug.

Can somebody tell me why some _other_ random wakeup cannot cause the same 
kind of migration at an inopportune time?

The fact is, Dmitry's patch fixed _one_ particular wakeup from happening 
(that just happened to be *guaranteed* to happen when it shouldn't!), but 
as far as I can tell, it's a totally generic problem, with any

	try_to_wake_up() -> load-balancer

chain being able to trigger it by causing a migration to a CPU that we 
are in the process of turning off.

IOW, I don't think that my patch is overkill at all. I think it fixes the 
real bug there.

(It's also true that the cpusets code calls rebuild_sched_domains() way 
too much, but that's a _stupidity_ issue, not the cause of the bug per se, 
if I follow the code!)

			Linus