Re: [PATCH 0/5][RFC] more fine grained locking for IRQ injection

[Gleb Natapov <gleb@redhat.com>]

On Mon, Jul 13, 2009 at 05:43:15PM +0300, Michael S. Tsirkin wrote:
> On Mon, Jul 13, 2009 at 05:33:12PM +0300, Gleb Natapov wrote:
> > On Mon, Jul 13, 2009 at 05:21:48PM +0300, Michael S. Tsirkin wrote:
> > > On Mon, Jul 13, 2009 at 04:58:58PM +0300, Gleb Natapov wrote:
> > > > On Mon, Jul 13, 2009 at 04:53:15PM +0300, Michael S. Tsirkin wrote:
> > > > > On Mon, Jul 13, 2009 at 04:28:09PM +0300, Gleb Natapov wrote:
> > > > > > On Mon, Jul 13, 2009 at 04:23:36PM +0300, Michael S. Tsirkin wrote:
> > > > > > > On Mon, Jul 13, 2009 at 12:12:30PM +0300, Gleb Natapov wrote:
> > > > > > > > kvm->irq_lock protects too much stuff, but still fail to protect
> > > > > > > > everything it was design to protect (see ack notifiers call in pic). I
> > > > > > > > want to make IRQ injection logic use more fine grained locking.
> > > > > > > 
> > > > > > > At least irq routing changes and notifier list changes
> > > > > > > do not seem to be involved in irq injection.
> > > > > > > So why do we want fine-grained locking there?
> > > > > > > 
> > > > > > When you have one big lock and you want to eliminate it you look at all
> > > > > > things it protects and you start introducing different locking for
> > > > > > unrelated stuff. This is what this patch series does, so I don't really
> > > > > > get you point.
> > > > > 
> > > > > But why do you want to eliminate it? What is the operation that this
> > > > > will speed up? Is there a contention issue and which operations
> > > > > contend on the lock?
> > > > > 
> > > > Interrupt injection. I want to be able to inject interrupts without the
> > > > lock at all.
> > > 
> > > You are talking about the lapic changes below, right?  OK. But how do
> > > the other changes help, specifically? I'm less familiar with ioapic, but
> > > I think irq routing and irq notifier list changes are done off data
> > > path.
> > > 
> > irq_routing is accessed on irq injection path, so making in lockless is
> > necessary to achieve the goal. Making ack notifiers lockless removes one
> > lock from EOI path.
> 
> But this is done with RCU patches, right?
> 
No, not right. Only rcu patch is meaningless since irq injection is
still done under irq_lock. Look at the whole series, not each line by
itself.

--
			Gleb.