kvm-unit-test fail for split irqchip

From: Wanpeng Li
Date: Tue Sep 13 2016 - 07:06:52 EST


Add -kernel_irqchip=split
./x86-run x86/eventinj.flat

qemu-system-x86_64 -enable-kvm -machine kernel_irqchip=split -cpu host
-device pc-testdev -device isa-debug-exit,iobase=0xf4,iosize=0x4 -vnc
none -serial stdio -device pci-testdev -kernel x86/eventinj.flat
enabling apic
paging enabled
cr0 = 80010011
cr3 = 7fff000
cr4 = 20
Sending vec 33 and 62 and mask one with TPR
irq1 running
irq1 running
After 33/62 TPR test
FAIL: TPR
irq0 running
irq0 running

Both irq1 and irq0 are executing twice.

qemu-system-x86-22794 [001] d..2 34591.708476: kvm_entry: vcpu 0
qemu-system-x86-22794 [001] ...1 34591.708478: kvm_exit: reason
MSR_WRITE rip 0x401f33 info 0 0
qemu-system-x86-22794 [001] ...1 34591.708478: kvm_apic: apic_write
APIC_EOI = 0x0
qemu-system-x86-22794 [001] ...1 34591.708479: kvm_eoi: apicid 0 vector 62
qemu-system-x86-22794 [001] ...1 34591.708479: kvm_msr: msr_write 80b = 0x0
qemu-system-x86-22794 [001] d..2 34591.708480: kvm_entry: vcpu 0
qemu-system-x86-22794 [001] ...1 34591.708482: kvm_exit: reason
PENDING_INTERRUPT rip 0x401f35 info 0 0
qemu-system-x86-22794 [001] ...1 34591.708482: kvm_userspace_exit:
reason KVM_EXIT_IRQ_WINDOW_OPEN (7)
qemu-system-x86-22794 [001] ...1 34591.708491: kvm_inj_virq: irq 62
qemu-system-x86-22794 [001] d..2 34591.708492: kvm_entry: vcpu 0
qemu-system-x86-22794 [001] ...1 34591.708493: kvm_exit: reason
IO_INSTRUCTION rip 0x4016ec info 3fd0008 0

>From the trace we can see there is an interrupt window exit after the
first interrupt EOI(irq 62), and the same irq(62) is injected
duplicately after the interrupt window.

The bug can disappear if kernel_irqchip is on or -x2apic, the virtual
x2apic mode will not be set due to commit (8d14695f9542 x86, apicv:
add virtual x2apic support), so that tpr shadow in the x2apic doesn't
work and wrmsr TPR register will trigger vmexit, and then kvmvapic
will be used to optimize flexpriority=N or AMD. The
report_trp_access() which is called in kvm_lapic_reg_write() will
trigger a userspace exit.

TPR report access callbacks in qemu, kvm_handle_tpr_access() ->
apic_handle_tpr_access_report() -> vapic_report_tpr_access() ->
cpu_synchronize_state() will get apic register states from kvm.

Later, kvm_arch_pre_run -> cpu_get_pic_interrupt(if there is a pic
interrupt) -> apic_get_interrupt, it is a pic interrupt, however it
gets the stale irq from apic register sync by report tpr access and
KVM_INTERRUPT the second duplicate interrupt.

Paolo pointed out it is not the TPR associated bug, and we should
figure out why there is an interrupt window exit after the first EOI.

Regards,
Wanpeng Li