Re: [RFC PATCH 1/2] spapr: Report correct GTSE support via ov5

[Fabiano Rosas]

David Gibson <david@gibson.dropbear.id.au> writes:

> On Tue, Mar 08, 2022 at 10:23:59PM -0300, Fabiano Rosas wrote:
>> QEMU reports MMU support to the guest via the ibm,architecture-vec-5
>> property of the /chosen node. Byte number 26 specifies Radix Table
>> Expansions, currently only GTSE (Guest Translation Shootdown
>> Enable). This feature determines whether the tlbie instruction (and
>> others) are HV privileged.
>> 
>> Up until now, we always reported GTSE=1 to guests. Even after the
>> support for GTSE=0 was added. As part of that support, a kernel
>> command line radix_hcall_invalidate=on was introduced that overrides
>> the GTSE value received via CAS. So a guest can run with GTSE=0 and
>> use the H_RPT_INVALIDATE hcall instead of tlbie.
>> 
>> In this scenario, having GTSE always set to 1 by QEMU leads to a crash
>> when running nested KVM guests because KVM does not allow a nested
>> hypervisor to set GTSE support for its nested guests. So a nested
>> guest always uses the same value for LPCR_GTSE as its HV. Since the
>> nested HV disabled GTSE, but the L2 QEMU always reports GTSE=1, we run
>> into a crash when:
>> 
>> L1 LPCR_GTSE=0
>> L2 LPCR_GTSE=0
>> L2 CAS GTSE=1
>> 
>> The nested guest will run 'tlbie' and crash because the HW looks at
>> LPCR_GTSE, which is clear.
>> 
>> Having GTSE disabled in the L1 and enabled in the L2 is not an option
>> because the whole purpose of GTSE is to disallow access to tlbie and
>> we cannot allow L1 to spawn L2s that can access features that L1
>> itself cannot.
>> 
>> We also cannot have the guest check the LPCR bit, because LPCR is
>> HV-privileged.
>> 
>> So this patch goes through the most intuitive route which is to have
>> QEMU ask KVM about GTSE support and advertise the correct value to the
>> guest. A new KVM_CAP_PPC_GTSE capability is being added.
>> 
>> TCG continues to always enable GTSE.
>> 
>> Signed-off-by: Fabiano Rosas <farosas@linux.ibm.com>
>> ---
>>  hw/ppc/spapr.c       | 38 +++++++++++++++++++++++++++++++-------
>>  target/ppc/kvm.c     |  8 ++++++++
>>  target/ppc/kvm_ppc.h |  6 ++++++
>>  3 files changed, 45 insertions(+), 7 deletions(-)
>> 
>> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
>> index 4cc204f90d..3e95a1831f 100644
>> --- a/hw/ppc/spapr.c
>> +++ b/hw/ppc/spapr.c
>> @@ -971,7 +971,7 @@ static void 
>> spapr_dt_ov5_platform_support(SpaprMachineState *spapr, void *fdt,
>>          23, 0x00, /* XICS / XIVE mode */
>>          24, 0x00, /* Hash/Radix, filled in below. */
>>          25, 0x00, /* Hash options: Segment Tables == no, GTSE == no. */
>> -        26, 0x40, /* Radix options: GTSE == yes. */
>> +        26, 0x00, /* Radix options, filled in below. */
>>      };
>>  
>>      if (spapr->irq->xics && spapr->irq->xive) {
>> @@ -1000,10 +1000,16 @@ static void 
>> spapr_dt_ov5_platform_support(SpaprMachineState *spapr, void *fdt,
>>          } else {
>>              val[3] = 0x00; /* Hash */
>>          }
>> +
>> +        if (kvmppc_has_cap_gtse()) {
>> +            val[7] = 0x40 /* OV5_MMU_RADIX_GTSE */;
>> +        }
>>      } else {
>>          /* V3 MMU supports both hash and radix in tcg (with dynamic 
>> switching) */
>>          val[3] = 0xC0;
>> +        val[7] = 0x40 /* OV5_MMU_RADIX_GTSE */;
>>      }
>> +
>>      _FDT(fdt_setprop(fdt, chosen, "ibm,arch-vec-5-platform-support",
>>                       val, sizeof(val)));
>>  }
>> @@ -2824,14 +2830,32 @@ static void spapr_machine_init(MachineState *machine)
>>      /* Init numa_assoc_array */
>>      spapr_numa_associativity_init(spapr, machine);
>>  
>> -    if ((!kvm_enabled() || kvmppc_has_cap_mmu_radix()) &&
>> -        ppc_type_check_compat(machine->cpu_type, CPU_POWERPC_LOGICAL_3_00, 
>> 0,
>> +    if (ppc_type_check_compat(machine->cpu_type, CPU_POWERPC_LOGICAL_3_00, 
>> 0,
>>                                spapr->max_compat_pvr)) {
>> -        spapr_ovec_set(spapr->ov5, OV5_MMU_RADIX_300);
>> -        /* KVM and TCG always allow GTSE with radix... */
>> -        spapr_ovec_set(spapr->ov5, OV5_MMU_RADIX_GTSE);
>> +
>> +        /* TCG always supports Radix w/ GTSE */
>> +        if (!kvm_enabled()) {
>> +            spapr_ovec_set(spapr->ov5, OV5_MMU_RADIX_300);
>> +            spapr_ovec_set(spapr->ov5, OV5_MMU_RADIX_GTSE);
>
> Yeah, this is no good.  It's never ok to change guest visible
> behaviour depending on host properties (in this case whether it's KVM
> or not).  It messes up the invariants we need for migration, which
> require that the guest visible state depend only on the user
> configuration.
>
> The usual way to handle this is with a new capability, you can then
> change the default with the next machine version.

This particular problem is tricky. TCG cannot disable GTSE because it
does not support H_RPT_INVALIDATE. And older kernels that don't know
about the feature require GTSE.

KVM can afford to disable GTSE because we have a compatibility mechanism
(although a bit crooked): We can invert the logic for the KVM_CAP so
that the presence of KVM_CAP_PPC_GTSE_DISABLE would mean QEMU is allowed
to disable GTSE. Then:
  - older KVM + new QEMU would keep GTSE enabled;

  - older L1 guests are not affected because the host would report
    GTSE=1 with the KVM capability. By the time we decide to disable
    GTSE for L1 guests hopefully all older kernels will be out of use;

  - older nested guests:
    - if L1 runs with GTSE=1, are not affected;

    - if L1 disabled GTSE via kernel cmdline, are already broken (this
      bug). But they would go from crashing to being aborted* by QEMU
      (the guest asks for HPT in the lack of GTSE; nested KVM is radix
      only);

      * there are other broken cases which are fixed completely.

To satisfy TCG we could keep a spapr capability as ON and usually the
guest would pass cap-gtse=off when running with KVM. However this
doesn't work because this crash happens precisely because the nested
guest doesn't know that it needs to use cap-rpt-invalidate=on. Another
cap wouldn't help.

So I think the only way to have a spapr capability for this is if TCG
always defaults to ON and KVM always defaults to OFF. But then we would
be changing guest visible behaviour depending on host properties.