Re: buggy emulate_int_real

[Avi Kivity <avi@redhat.com>]

On 04/12/2011 04:57 PM, Serge E. Hallyn wrote:
> Quoting Avi Kivity (avi@redhat.com):
> >  On 04/12/2011 10:53 AM, Serge E. Hallyn wrote:
> >  >Quoting Avi Kivity (avi@redhat.com):
> >  >>   On 04/09/2011 12:09 AM, Serge E. Hallyn wrote:
> >  >>   >Hi,
> >  >>   >
> >  >>   >at https://bugs.launchpad.net/ubuntu/+source/qemu-kvm/+bug/747090, it was
> >  >>   >found that emulate_int_real() sometimes pushes the wrong eip when doing a
> >  >>   >int.  Whereas with non-kvm qemu we push the next instruction after the
> >  >>   >int, with kvm we push the addr of the instruction itself.
> >  >>   >
> >  >>
> >  >>
> >  >>   The code says:
> >  >>
> >  >>       c->src.val = c->eip;
> >  >>       emulate_push(ctxt, ops);
> >  >>       rc = writeback(ctxt, ops);
> >  >>       if (rc != X86EMUL_CONTINUE)
> >  >>           return rc;
> >  >>
> >  >>   which appears to be the address of the next instruction from my
> >  >>   reading of the code (see how insn_fetch() increments c->eip).
> >  >
> >  >Nevertheless removing commits
> >  >
> >  >	a92601bb707f6f49fd5563ef3d09928e70cc222e
> >  >	63995653ade16deacaea5b49ceaf6376314593ac
> >  >	6e154e56b4d7a6a28c54f0984e13d3f8defc4755
> >  >
> >  >changes the eip value being pushed.  If you look at
> >  >a92601bb707f6f49fd5563ef3d09928e70cc222e, you see:
> >  >
> >  >          if (vmx->rmode.vm86_active) {
> >  >-               vmx->rmode.irq.pending = true;
> >  >-               vmx->rmode.irq.vector = nr;
> >  >-               vmx->rmode.irq.rip = kvm_rip_read(vcpu);
> >  >-               if (kvm_exception_is_soft(nr))
> >  >-                       vmx->rmode.irq.rip +=
> >  >-                               vmx->vcpu.arch.event_exit_inst_len;
> >  >-               intr_info |= INTR_TYPE_SOFT_INTR;
> >  >-               vmcs_write32(VM_ENTRY_INTR_INFO_FIELD, intr_info);
> >  >-               vmcs_write32(VM_ENTRY_INSTRUCTION_LEN, 1);
> >  >-               kvm_rip_write(vcpu, vmx->rmode.irq.rip - 1);
> >  >+               if (kvm_inject_realmode_interrupt(vcpu, nr) != EMULATE_DONE)
> >  >+                       kvm_make_request(KVM_REQ_TRIPLE_FAULT, vcpu);
> >  >                  return;
> >  >          }
> >  >
> >  >but kvm_inject_realmode_interrupt() does not appear to increment
> >  >vmx->rmode.irq.rip anywhere, as the code being replaced does.
> >
> >  Ah, I see now.  There are two cases, hard interrupt and soft
> >  interrupts.  I guess hard interrupts are handled fine, and the
> >  failing case is
> >
> >    guest executes INTn instruction in guest mode
> >    vmx intercepts a page fault (say due to access to the IDT or the stack)
> >    kvm notes that a soft interrupt was in progress (vmx_complete_interrupts)
> >    kvm handles the exception
> >    reinject the interrupt while reentering the guest
> >
> >  so we do need something like
> >
> >     if (soft)
> >         vcpu->arch.emulate_ctxt.eip += inst_len;
> >
> >  in kvm_inject_realmode_interrupt().
>
> But not always, so perhaps it would be better to still do this at
> the callers?

Possibly.

> For instance vmx_inject_nmi() didn't do it at all,
> and the other two callers only did it for a soft interrupt.

NMIs are always hard.

Perhaps have a new helper for this.

-- 
error compiling committee.c: too many arguments to function