On Mon, Nov 23, 2009 at 04:34:15PM +0100, Peter Zijlstra wrote: > On Mon, 2009-11-23 at 16:06 +0200, Gleb Natapov wrote: > > Do not preempt kernel. Just maintain counter to know if task can be rescheduled. > > Asynchronous page fault may be delivered while spinlock is held or current > > process can't be preempted for other reasons. KVM uses preempt_count() to check if preemptions is allowed and schedule other process if possible. This works > > with preemptable kernels since they maintain accurate information about > > preemptability in preempt_count. This patch make non-preemptable kernel > > maintain accurate information in preempt_count too. > > I'm thinking you're going to have to convince some people this won't > slow them down for no good. > I saw old discussions about this in mailing list archives. Usually someone wanted to use in_atomic() in driver code and this, of course, caused the resistant. In this case, I think, the use is legitimate. > Personally I always have PREEMPT=y, but other people seem to feel > strongly about not doing so. > It is possible to add one more config option to enable reliable preempt_count() without enabling preemption or make async pf be dependable on PREEMPT=y. Don't like both of this options especially first one. There are more then enough options already. -- Gleb. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
