Re: [Qemu-devel] E5-2620v2 - emulation stop error

[Radim Krčmář]

2015-03-30 22:32+0300, Andrey Korolyov:
> On Mon, Mar 30, 2015 at 9:56 PM, Radim Krčmář <address@hidden> wrote:
> > 2015-03-27 13:16+0300, Andrey Korolyov:
> >> On Fri, Mar 27, 2015 at 12:03 AM, Bandan Das <address@hidden> wrote:
> >> > Radim Krčmář <address@hidden> writes:
> >> >> I second Bandan -- checking that it reproduces on other machine would be
> >> >> great for sanity :)  (Although a bug in our APICv is far more likely.)
> >> >
> >> > If it's APICv related, a run without apicv enabled could give more hints.
> >> >
> >> > Your "devices not getting reset" hypothesis makes the most sense to me,
> >> > maybe the timer vector in the error message is just one part of
> >> > the whole story. Another misbehaving interrupt from the dark comes in at 
> >> > the
> >> > same time and leads to a double fault.
> >>
> >> Default trace (APICv enabled, first reboot introduced the issue):
> >> http://xdel.ru/downloads/kvm-e5v2-issue/hanged-reboot-apic-on.dat.gz
> >
> > The relevant part is here,
> > prefixed with "qemu-system-x86-4180  [002]   697.111550:"
> >
> >   kvm_exit:             reason CR_ACCESS rip 0xd272 info 0 0
> >   kvm_cr:               cr_write 0 = 0x10
> >   kvm_mmu_get_page:     existing sp gfn 0 0/4 q0 direct --- !pge !nxe root 
> > 0 sync
> >   kvm_entry:            vcpu 0
> >   kvm_emulate_insn:     f0000:d275: ea 7a d2 00 f0
> >   kvm_emulate_insn:     f0000:d27a: 2e 0f 01 1e f0 6c
> >   kvm_emulate_insn:     f0000:d280: 31 c0
> >   kvm_emulate_insn:     f0000:d282: 8e e0
> >   kvm_emulate_insn:     f0000:d284: 8e e8
> >   kvm_emulate_insn:     f0000:d286: 8e c0
> >   kvm_emulate_insn:     f0000:d288: 8e d8
> >   kvm_emulate_insn:     f0000:d28a: 8e d0
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0xd28f info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EPT_VIOLATION rip 0x8dd0 info 184 0
> >   kvm_page_fault:       address f8dd0 error_code 184
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0x8dd0 info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EPT_VIOLATION rip 0x76d6 info 184 0
> >   kvm_page_fault:       address f76d6 error_code 184
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0x76d6 info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason PENDING_INTERRUPT rip 0xd331 info 0 0
> >   kvm_inj_virq:         irq 8
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0xfea5 info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EPT_VIOLATION rip 0xfea5 info 184 0
> >   kvm_page_fault:       address ffea5 error_code 184
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0xfea5 info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EPT_VIOLATION rip 0xe990 info 184 0
> >   kvm_page_fault:       address fe990 error_code 184
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXTERNAL_INTERRUPT rip 0xe990 info 0 800000f6
> >   kvm_entry:            vcpu 0
> >   kvm_exit:             reason EXCEPTION_NMI rip 0xd334 info 0 80000b0d
> >   kvm_userspace_exit:   reason KVM_EXIT_INTERNAL_ERROR (17)
> >
> >> Trace without APICv (three reboots, just to make sure to hit the
> >> problematic condition of supposed DF, as it still have not one hundred
> >> percent reproducibility):
> >> http://xdel.ru/downloads/kvm-e5v2-issue/apic-off.dat.gz
> >
> > The trace here contains a well matching excerpt, just instead of the
> > EXCEPTION_NMI, it does
> >
> >  169.905098: kvm_exit:             reason EPT_VIOLATION rip 0xd334 info 181 > > 0
> >  169.905102: kvm_page_fault:       address feffd066 error_code 181
> >
> > and works.  Page fault says we tried to read 0xfeffd066 -- probably IOPB
> > of TSS.  (I guess it is pre-fetch for following IO instruction.)
> >
> > Nothing strikes me when looking at it, but some APICv boots don't fail,
> > so it would be interesting to compare them ... hosts's 0xf6 interrupt
> > (IRQ_WORK_VECTOR) is a possible source of races.  (We could look more
> > closely.  It is fired too often for my liking as well.)
> 
> Thanks Radim, 
> http://xdel.ru/downloads/kvm-e5v2-issue/no-fail-with-apicv.dat.gz
> 
> The related bits looks the same as with enable_apicv=0 for me.

Yeah,

 qemu-system-x86-4201  [007]   159.297337:
  kvm_exit:             reason CR_ACCESS rip 0xd272 info 0 0
  kvm_cr:               cr_write 0 = 0x10
  kvm_mmu_get_page:     existing sp gfn 0 0/4 q0 direct --- !pge !nxe root 0 
sync
  kvm_entry:            vcpu 0
  kvm_emulate_insn:     f0000:d275: ea 7a d2 00 f0
  kvm_emulate_insn:     f0000:d27a: 2e 0f 01 1e f0 6c
  kvm_emulate_insn:     f0000:d280: 31 c0
  kvm_emulate_insn:     f0000:d282: 8e e0
  kvm_emulate_insn:     f0000:d284: 8e e8
  kvm_emulate_insn:     f0000:d286: 8e c0
  kvm_emulate_insn:     f0000:d288: 8e d8
  kvm_emulate_insn:     f0000:d28a: 8e d0
  kvm_entry:            vcpu 0
  kvm_exit:             reason EPT_VIOLATION rip 0x8dd0 info 184 0
  kvm_page_fault:       address f8dd0 error_code 184
  kvm_entry:            vcpu 0
  kvm_exit:             reason EPT_VIOLATION rip 0x76d6 info 184 0
  kvm_page_fault:       address f76d6 error_code 184
  kvm_entry:            vcpu 0
  kvm_exit:             reason EPT_VIOLATION rip 0xd331 info 181 0
  kvm_page_fault:       address feffd066 error_code 181
  kvm_inj_virq:         irq 25
  kvm_entry:            vcpu 0
  kvm_exit:             reason EPT_VIOLATION rip 0xe6f2 info 184 0
  kvm_page_fault:       address fe6f2 error_code 184
  kvm_entry:            vcpu 0
  kvm_exit:             reason IO_INSTRUCTION rip 0xd202 info 700040 0
  kvm_pio:              pio_write at 0x70 size 1 count 1 val 0x8f
  kvm_userspace_exit:   reason KVM_EXIT_IO (2)

I noticed three differences from the failing version:
 1) Page fault 0xfeffd066 @ 0xd331, KVM also injects vector 0x19 there
 2) No PENDING_INTERRUPT irq 8  (not sure why it happened on APICv)
 3) No EXTERNAL_INTERRUPT vector 0xf6 (lucky race)

Chasing the culprit this way could take a long time, so a new tracepoint
that shows if 0xef is set on entry would let us guess the bug faster ...

Please provide a failing trace with the following patch:

---
diff --git a/arch/x86/kvm/trace.h b/arch/x86/kvm/trace.h
index 7c7bc8bef21f..8cf85dcaa1ee 100644
--- a/arch/x86/kvm/trace.h
+++ b/arch/x86/kvm/trace.h
@@ -742,6 +742,29 @@ TRACE_EVENT(kvm_emulate_insn,
 #define trace_kvm_emulate_insn_start(vcpu) trace_kvm_emulate_insn(vcpu, 0)
 #define trace_kvm_emulate_insn_failed(vcpu) trace_kvm_emulate_insn(vcpu, 1)
 
+TRACE_EVENT(kvm_0xef,
+       TP_PROTO(bool irr, bool isr, u32 vmcs),
+       TP_ARGS(irr, isr, vmcs),
+
+       TP_STRUCT__entry(
+               __field(bool, irr )
+               __field(bool, isr )
+               __field(u32,  vmcs)
+               ),
+
+       TP_fast_assign(
+               __entry->irr  = irr;
+               __entry->isr  = isr;
+               __entry->vmcs = vmcs;
+               ),
+
+       TP_printk("irr %s, isr %s, vmcs 0x%x",
+                 __entry->irr ? "set  " : "clear",
+                 __entry->isr ? "set  " : "clear",
+                 __entry->vmcs
+                )
+       );
+
 TRACE_EVENT(
        vcpu_match_mmio,
        TP_PROTO(gva_t gva, gpa_t gpa, bool write, bool gpa_match),
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index eee63dc33d89..254cca12c2ec 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -8129,6 +8129,13 @@ static void atomic_switch_perf_msrs(struct vcpu_vmx *vmx)
                                        msrs[i].host);
 }
 
+#define VEC_POS(v) ((v) & (32 - 1))
+#define REG_POS(v) (((v) >> 5) << 4)
+static inline int apic_test_vector(int vec, void *bitmap)
+{
+       return test_bit(VEC_POS(vec), (bitmap) + REG_POS(vec));
+}
+
 static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
 {
        struct vcpu_vmx *vmx = to_vmx(vcpu);
@@ -8143,6 +8150,10 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
        if (vmx->emulation_required)
                return;
 
+       trace_kvm_0xef(apic_test_vector(0xef, vcpu->arch.apic->regs + APIC_IRR),
+                      apic_test_vector(0xef, vcpu->arch.apic->regs + APIC_ISR),
+                      vmcs_read32(VM_ENTRY_INTR_INFO_FIELD));
+
        if (vmx->ple_window_dirty) {
                vmx->ple_window_dirty = false;
                vmcs_write32(PLE_WINDOW, vmx->ple_window);
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 32bf19ef3115..a45fa01bd354 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -7881,3 +7881,4 @@ EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_intercepts);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_write_tsc_offset);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_ple_window);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_pml_full);
+EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_0xef);