small migration thread fix

Hi,

the small patch fixes a potential problem in the migration thread for the case that the first CPU in the cpus_allowed mask of a process is offline. Please consider applying it to your trees.

Thanks!
Regards,
Erich

diff -urNp linux-2.5.55/kernel/sched.c linux-2.5.55-fix/kernel/sched.c
@@ -2108,7 +2108,7 @@ static int migration_thread(void * data)
 		spin_unlock_irqrestore(&rq->lock, flags);
 
 		p = req->task;
-		cpu_dest = __ffs(p->cpus_allowed);
+		cpu_dest = __ffs(p->cpus_allowed & cpu_online_map);
 		rq_dest = cpu_rq(cpu_dest);
 repeat:
 		cpu_src = task_cpu(p);

I'm not mingo, but I can say this looks sane. My only question is whether there are more codepaths that need this kind of check, for instance, what happens if someone does set_cpus_allowed() to a cpumask with !(task->cpumask & cpu_online_map) ?

Thanks,
Bill

The piece of code below was intended for that. I agree with Rusty's comment, BUG() is too strong for that case.

#if 0 /* FIXME: Grab cpu_lock, return error on this case. --RR */
	new_mask &= cpu_online_map;
	if (!new_mask)
		BUG();

#endif

Anyhow, changing the new_mask in this way is BAD, because the masks are inherited. So when more CPUs come online, they remain excluded from the mask of the process and it's children.

The fix suggested in the comments still has to be done...

Regards,
Erich

I don't have much to add but another ack and a "hmm, maybe something could be done". My prior comments stand. I'd be very much obliged if you provide a fix for the set_cpus_allowed() issue. I very much rely upon you now to provide scheduler fixes and optimizations for large scale and/or NUMA machines these days.

Thanks,
Bill