On Thu, Sep 16, 2010 at 11:25:53AM +0200, Gleb Natapov wrote: > On Thu, Sep 16, 2010 at 11:10:55AM +0200, Michael S. Tsirkin wrote: > > On Thu, Sep 16, 2010 at 11:02:56AM +0200, Gleb Natapov wrote: > > > On Wed, Sep 15, 2010 at 08:54:17PM +0200, Michael S. Tsirkin wrote: > > > > To avoid bouncing irqfd injection out to workqueue context, > > > > we must support injecting irqs from local interrupt > > > > context. Doing this seems to only require disabling > > > > irqs locally. > > > > > > > > RFC, completely untested, x86 only. > > > > > > > > Thoughts? > > > > > > > We do not want to disable irqs for a long time and some of code paths > > > under lock involve looping over all cpus. For MSI injection path is > > > lockless and this is the only case that matters, > > > > MSI only appeared in rhel6, older guests still use level interrupts. > So they are already slow for other reasons. They are slower than MSI, but I do not want irqfd to be slower than ioctls. > > Which paths require looping over all cpus? Do PCI interrupts > > need this? > > > All interrupts need it. IOAPIC has a loop to find dst cpu for interrupt. > Pic has a loop to find cpu in virtual wire mode. > > -- > 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
