On Wed, Apr 22, 2020 at 02:27:33PM -0700, Jim Mattson wrote: > On Wed, Apr 22, 2020 at 2:06 PM Sean Christopherson > <sean.j.christopherson@xxxxxxxxx> wrote: > The KVM code that deals with all of these events is really hard to > follow. I wish we could take a step back and just implement Table 6-2 > from the SDM volume 3 (augmented with the scattered information about > VMX events and their priorities relative to their nearest neighbors. > Lumping priorities 7 - 10 together (faults that we either intercepted > or synthesized in emulation), I think these are the various things we > need to check, in this order... > > 0. Is there a fault to be delivered? (In L2, is it intercepted by L1?) > 1. Is there a RESET or machine check event? > 2. Is there a trap on task switch? > 3. Is there an SMI or an INIT? > 3.5 In L2, is there an MTF VM-exit? > 4. Is there a #DB trap on the previous instruction? (In L2, is it > intercepted by L1?) > 4.3 In L2, has the VMX-preemption timer expired? > 4.6 In L2, do we need to synthesize an NMI-window VM-exit? > 5. Is there an NMI? (In L2, is it intercepted by L1?) > 5.3 In L2 do we need to synthesize an interrupt-window VM-exit? > 5.6 In L2, do we need to virtualize virtual-interrupt delivery? > 6. Is there a maskable interrupt? (In L2, is it intercepted by L1?) > 7. Now, we can enter VMX non-root mode. 100% agreed. I even tried to go down that path, multiple times, while sorting this stuff out. The big problem that isn't easily resolved is kvm_vcpu_running(), which currently calls .check_nested_events() even if KVM_REQ_EVENT isn't set. Its existence makes it annoyingly difficult to provide a unified single-pass flow for exiting and non-exiting events, e.g. we'd either have to duplicate a big pile of logic (eww) or significantly rework the event handling (scary). Having the INIT and SIPI handling buried in kvm_apic_accept_events() is also a pain, but that's less scary to change. In the long term, I absolutely think it'd be worth revamping the event handling so that it's not scattered all over tarnation, but that's something that should probably have a full kernel cycle or two of testing and performance analysis. If someone does pick up that torch, I think it'd also be worth experimenting with removing KVM_REQ_EVENT, i.e. processing events on _every_ run. IMO that would simplify the code, or at least how one reasons about the code, a great deal.
