On Wed, 17 Jun 2009, Gregory Haskins wrote: > Can you elaborate? I currently do not see how I could do the proposed > concept inside of irqfd while still using eventfd. Of course, that > would be possible if we fork irqfd from eventfd, and perhaps this is > what you are proposing. As previously stated I don't want to give up on > the prospect of re-using it quite yet, so bear with me. :) > > The issue with eventfd, as I see it, is that eventfd uses a > spin_lock_irqsave (by virtue of the wait-queue stuff) across the > "signal" callback (which today is implemented as a wake-up). This > spin_lock implicitly creates a non-preemptible critical section that > occurs independently of whether eventfd_signal() itself is invoked from > a sleepable context or not. > > What I strive to achieve is to remove the creation of this internal > critical section. If eventfd_signal() is called from atomic context, so > be it. We will detect this in the callback and be forced to take the > slow-path, and I am ok with that. *But*, if eventfd_signal() (or > f_ops->write(), for that matter) are called from a sleepable context > *and* eventfd doesn't introduce its own critical section (such as with > my srcu patch), we can potentially optimize within the callback by > executing serially instead of deferring (e.g. via a workqueue). Since when the scheduling (assuming it's not permanently running on another core due to high frequency work post) of a kernel thread is such a big impact that interfaces need to be redesigned for that? How much the (possible, but not certain) kernel thread context switch time weighs in the overall KVM IRQ service time? > It can! :) This is not changing from whats in mainline today (covered > above). It can/could, if the signal() function takes very accurate care of doing the magic atomic check. - Davide -- 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
