On 2016-04-26 18:07, Radim Krčmář wrote: > 2016-04-26 17:28+0200, Jan Kiszka: >> On 2016-04-26 16:59, Radim Krčmář wrote: >>> 2016-04-26 16:24+0200, Jan Kiszka: >>>> On 2016-04-26 13:40, Peter Xu wrote: >>>>> Currently, all the interrupts will be translated into one MSI in >>>>> vtd_generate_msi_message(), in which only 8 bits of dest_id is used >>>>> (msg.dest = irq->dest). We may possibly need to use the high 32 bits >>>>> of MSI address to store the rest dest[31:8]? Don't know whether this >>>>> would be enough though. >>>> >>>> Yes, I ran into this topic as well as I hacked up those line. Currently, >>>> KVM does not support more than 254 vCPUs, so 8 bits of those 32 are >>>> actually fine, and piggy-backing them in an MSI message works. >>>> >>>> Once KVM supports more CPUs, it has to come up with a new userspace >>>> interface to inject APIC events for more than 255 CPUs. Maybe the >>>> existing direct MSI inject with its unused flags could be "bended", >>>> maybe there are already better ideas (Radim?). >>> >>> Adding a flag to msi_msg and taking 3-4 bytes from padding to express >>> x2APIC addresses is reasonable. (It is what my prototype did. :]) >>> >>> The conceptually different idea is forcing all userspace interrupts >>> through irqfd routes, which would obsolete the ad-host inject. >> >> irqfd for userspace sources is a bit clumsy from the API POV. On the >> other hand, we need to tweak the routing API anyway to achieve the same >> address extension there, too. > > I agree, both of them should change. KVM_SIGNAL_MSI was added just > because the route-based injection in kvm_irqchip_send_msi() sucked; > maybe we could find a good solution now, but the direct interface is > already here ...
We won't get away without irq routes because we need them for sources that only issue binary signals for interrupt events (in-kernel, eventfd from userspace). In fact, those sources should not know more than that about their irqs. But we also do not want them to route everything through userspace for on-the-fly translation and reinjection to kvm. Jan