Re: [PATCH v4 2/2] i386: Add support to get/set/migrate Intel Processor Trace feature

[Eduardo Habkost <ehabkost@redhat.com>]

On Mon, Oct 21, 2019 at 06:02:28AM +0000, Kang, Luwei wrote:
> > > > > f9f4cd1..097c953 100644
> > > > > --- a/target/i386/kvm.c
> > > > > +++ b/target/i386/kvm.c
> > > > > @@ -1811,6 +1811,25 @@ static int kvm_put_msrs(X86CPU *cpu, int level)
> > > > >                  kvm_msr_entry_add(cpu, MSR_MTRRphysMask(i), mask);
> > > > >              }
> > > > >          }
> > > > > +        if (env->features[FEAT_7_0_EBX] & CPUID_7_0_EBX_INTEL_PT) {
> > > > > +            int addr_num = kvm_arch_get_supported_cpuid(kvm_state,
> > > > > +                                                    0x14, 1,
> > > > > + R_EAX) & 0x7;
> > > > > +
> > > > > +            kvm_msr_entry_add(cpu, MSR_IA32_RTIT_CTL,
> > > > > +                            env->msr_rtit_ctrl);
> > > > > +            kvm_msr_entry_add(cpu, MSR_IA32_RTIT_STATUS,
> > > > > +                            env->msr_rtit_status);
> > > > > +            kvm_msr_entry_add(cpu, MSR_IA32_RTIT_OUTPUT_BASE,
> > > > > +                            env->msr_rtit_output_base);
> > > >
> > > > This causes the following crash on some hosts:
> > > >
> > > >   qemu-system-x86_64: error: failed to set MSR 0x560 to 0x0
> > > >   qemu-system-x86_64: target/i386/kvm.c:2673: kvm_put_msrs: Assertion `ret == cpu->kvm_msr_buf->nmsrs' failed.
> > > >
> > > > Checking for CPUID_7_0_EBX_INTEL_PT is not enough: KVM has
> > > > additional conditions that might prevent writing to this MSR (PT_CAP_topa_output && PT_CAP_single_range_output).  This
> > causes QEMU to crash if some of the conditions aren't met.
> > > >
> > > > Writing and reading this MSR (and the ones below) need to be conditional on KVM_GET_MSR_INDEX_LIST.
> > > >
> > >
> > > Hi Eduardo,
> > >     I found this issue can't be reproduced in upstream source code but can be reproduced on RHEL8.1. I haven't got the qemu source
> > code of RHEL8.1. But after adding some trace in KVM, I found the KVM has reported the complete Intel PT CPUID information to qemu
> > but the Intel PT CPUID (0x14) is lost when qemu setting the CPUID to KVM (cpuid level is 0xd). It looks like lost the below patch.
> > >
> > > commit f24c3a79a415042f6dc195f029a2ba7247d14cac
> > > Author: Luwei Kang <luwei.kang@intel.com>
> > > Date:   Tue Jan 29 18:52:59 2019 -0500
> > >     i386: extended the cpuid_level when Intel PT is enabled
> > >
> > >     Intel Processor Trace required CPUID[0x14] but the cpuid_level
> > >     have no change when create a kvm guest with
> > >     e.g. "-cpu qemu64,+intel-pt".
> > 
> > Thanks for the pointer.  This may avoid triggering the bug in the default configuration, but we still need to make the MSR writing
> > conditional on KVM_GET_MSR_INDEX_LIST.  Older machine-types have x-intel-pt-auto-level=off, and the user may set `level`
> > manually.
> 
> Hi Eduardo,
> Sorry for a delay reply because my mail filter. I tried with
> the Q35 machine type and default, all looks work well (With
> some old cpu type + "intel_pt" also work well).  KVM will check
> the Intel PT work mode and HW to decide if Intel PT can be
> exposed to guest, only extended the CPUID level is useless. If
> the guest doesn't support Intel PT, any MSR read or write will
> cause #GP. Please remind me if I lost something.

I understand you have tried q35 and pc, but have you tried with
older machine-type versions?

Commit f24c3a79a415 doesn't change behavior on pc-*-3.1 and
older, so it only avoids triggering the crash in the default
case.  Doesn't QEMU crash if running:
"-cpu qemu64,+intel-pt -machine pc-i440fx-3.1"?

KVM rejecting MSR writes when something is missing is correct.
QEMU trying to write the MSR when something is missing (and
crashing because of that) is a bug.

-- 
Eduardo