Re: [Qemu-devel] [PATCHv5 4/4] ppc: Rework CPU compatibility testing across migration

[Greg Kurz]

On Fri,  2 Jun 2017 13:15:07 +1000
David Gibson <address@hidden> wrote:

> Migrating between different CPU versions is a bit complicated for ppc.
> A long time ago, we ensured identical CPU versions at either end by
> checking the PVR had the same value.  However, this breaks under KVM
> HV, because we always have to use the host's PVR - it's not
> virtualized.  That would mean we couldn't migrate between hosts with
> different PVRs, even if the CPUs are close enough to compatible in
> practice (sometimes identical cores with different surrounding logic
> have different PVRs, so this happens in practice quite often).
> 
> So, we removed the PVR check, but instead checked that several flags
> indicating supported instructions matched.  This turns out to be a bad
> idea, because those instruction masks are not architected information, but
> essentially a TCG implementation detail.  So changes to qemu internal CPU
> modelling can break migration - this happened between qemu-2.6 and
> qemu-2.7.  That was addressed by 146c11f1 "target-ppc: Allow eventual
> removal of old migration mistakes".
> 
> Now, verification of CPU compatibility across a migration basically doesn't
> happen.  We simply ignore the PVR of the incoming migration, and hope the
> cpu on the destination is close enough to work.
> 
> Now that we've cleaned up handling of processor compatibility modes for
> pseries machine type, we can do better.  We allow migration if:
> 
>     * The source and destination PVRs are for the same type of CPU, as
>       determined by CPU class's pvr_match function
> OR  * When the source was in a compatibility mode, and the destination CPU
>       supports the same compatibility mode
> 
> Signed-off-by: David Gibson <address@hidden>
> Reviewed-by: Suraj Jitindar Singh <address@hidden>
> ---
>  target/ppc/machine.c | 69 
> +++++++++++++++++++++++++++++++++++++++++++++++++---
>  1 file changed, 66 insertions(+), 3 deletions(-)
> 
> diff --git a/target/ppc/machine.c b/target/ppc/machine.c
> index 6cb3a48..a29aabe 100644
> --- a/target/ppc/machine.c
> +++ b/target/ppc/machine.c
> @@ -8,6 +8,7 @@
>  #include "helper_regs.h"
>  #include "mmu-hash64.h"
>  #include "migration/cpu.h"
> +#include "qapi/error.h"
>  
>  static int cpu_load_old(QEMUFile *f, void *opaque, int version_id)
>  {
> @@ -195,6 +196,27 @@ static void cpu_pre_save(void *opaque)
>      }
>  }
>  
> +/*
> + * Determine if a given PVR is a "close enough" match to the CPU
> + * object.  For TCG and KVM PR it would probably be sufficient to
> + * require an exact PVR match.  However for KVM HV the user is
> + * restricted to a PVR exactly matching the host CPU.  The correct way
> + * to handle this is to put the guest into an architected
> + * compatibility mode.  However, to allow a more forgiving transition
> + * and migration from before this was widely done, we allow migration
> + * between sufficiently similar PVRs, as determined by the CPU class's
> + * pvr_match() hook.
> + */
> +static bool pvr_match(PowerPCCPU *cpu, uint32_t pvr)
> +{
> +    PowerPCCPUClass *pcc = POWERPC_CPU_GET_CLASS(cpu);
> +
> +    if (pvr == pcc->pvr) {
> +        return true;
> +    }
> +    return pcc->pvr_match(pcc, pvr);
> +}
> +
>  static int cpu_post_load(void *opaque, int version_id)
>  {
>      PowerPCCPU *cpu = opaque;
> @@ -203,10 +225,31 @@ static int cpu_post_load(void *opaque, int version_id)
>      target_ulong msr;
>  
>      /*
> -     * We always ignore the source PVR. The user or management
> -     * software has to take care of running QEMU in a compatible mode.
> +     * If we're operating in compat mode, we should be ok as long as
> +     * the destination supports the same compatiblity mode.
> +     *
> +     * Otherwise, however, we require that the destination has exactly
> +     * the same CPU model as the source.
>       */
> -    env->spr[SPR_PVR] = env->spr_cb[SPR_PVR].default_value;
> +
> +#if defined(TARGET_PPC64)
> +    if (cpu->compat_pvr) {
> +        Error *local_err = NULL;
> +
> +        ppc_set_compat(cpu, cpu->compat_pvr, &local_err);
> +        if (local_err) {
> +            error_report_err(local_err);
> +            error_free(local_err);
> +            return -1;
> +        }
> +    } else
> +#endif
> +    {
> +        if (!pvr_match(cpu, env->spr[SPR_PVR])) {
> +            return -1;
> +        }
> +    }
> +
>      env->lr = env->spr[SPR_LR];
>      env->ctr = env->spr[SPR_CTR];
>      cpu_write_xer(env, env->spr[SPR_XER]);
> @@ -560,6 +603,25 @@ static const VMStateDescription vmstate_tlbmas = {
>      }
>  };
>  
> +static bool compat_needed(void *opaque)
> +{
> +    PowerPCCPU *cpu = opaque;
> +
> +    assert(!(cpu->compat_pvr && !cpu->vhyp));
> +    return (cpu->compat_pvr != 0);

Since the CAS logic rework in 2.9, compat_pvr is likely to be != 0 and
the subsection is sent even for older machine types. This breaks backward
migration of such machines...

> +}
> +
> +static const VMStateDescription vmstate_compat = {
> +    .name = "cpu/compat",
> +    .version_id = 1,
> +    .minimum_version_id = 1,
> +    .needed = compat_needed,
> +    .fields = (VMStateField[]) {
> +        VMSTATE_UINT32(compat_pvr, PowerPCCPU),
> +        VMSTATE_END_OF_LIST()
> +    }
> +};
> +
>  const VMStateDescription vmstate_ppc_cpu = {
>      .name = "cpu",
>      .version_id = 5,
> @@ -613,6 +675,7 @@ const VMStateDescription vmstate_ppc_cpu = {
>          &vmstate_tlb6xx,
>          &vmstate_tlbemb,
>          &vmstate_tlbmas,
> +        &vmstate_compat,
>          NULL
>      }
>  };

pgpqWynufu00v.pgp
Description: OpenPGP digital signature