Re: [PATCH v2 22/30] KVM: s390: intercept the rpcit instruction

[Pierre Morel]

On 1/18/22 18:27, Matthew Rosato wrote:
> On 1/18/22 6:05 AM, Pierre Morel wrote:
> > On 1/14/22 21:31, Matthew Rosato wrote:
> > > For faster handling of PCI translation refreshes, intercept in KVM
> > > and call the associated handler.
> > >
> > > Signed-off-by: Matthew Rosato <mjrosato@xxxxxxxxxxxxx>
> > > ---
> > >   arch/s390/kvm/priv.c | 46 ++++++++++++++++++++++++++++++++++++++++++++
> > >   1 file changed, 46 insertions(+)
> > >
> > > diff --git a/arch/s390/kvm/priv.c b/arch/s390/kvm/priv.c
> > > index 417154b314a6..5b65c1830de2 100644
> > > --- a/arch/s390/kvm/priv.c
> > > +++ b/arch/s390/kvm/priv.c
> > > @@ -29,6 +29,7 @@
> > >   #include <asm/ap.h>
> > >   #include "gaccess.h"
> > >   #include "kvm-s390.h"
> > > +#include "pci.h"
> > >   #include "trace.h"
> > >   static int handle_ri(struct kvm_vcpu *vcpu)
> > > @@ -335,6 +336,49 @@ static int handle_rrbe(struct kvm_vcpu *vcpu)
> > >       return 0;
> > >   }
> > > +static int handle_rpcit(struct kvm_vcpu *vcpu)
> > > +{
> > > +    int reg1, reg2;
> > > +    u8 status;
> > > +    int rc;
> > > +
> > > +    if (vcpu->arch.sie_block->gpsw.mask & PSW_MASK_PSTATE)
> > > +        return kvm_s390_inject_program_int(vcpu, PGM_PRIVILEGED_OP);
> > > +
> > > +    /* If the host doesn't support PCI, it must be an emulated 
> > > device */
> > > +    if (!IS_ENABLED(CONFIG_PCI))
> > > +        return -EOPNOTSUPP;
> >
> > AFAIU this makes also sure that the following code is not compiled in 
> > case PCI is not supported.
> >
> > I am not very used to compilation options, is it true with all our 
> > compilers and options?
> > Or do we have to specify a compiler version?
> >
> > Another concern is, shouldn't we use IS_ENABLED(CONFIG_VFIO_PCI) ?
>
> Same idea as in the other thread -- What we are trying to protect 
> against here is referencing symbols that won't be linked (like 
> zpci_refresh_trans, or the aift->mdd a few lines below)
>
> It is indeed true that we should never need to handle the rpcit 
> intercept in KVM if CONFIG_VFIO_PCI=n -- but the necessary symbols/code 
> are linked at least, so we can just let the SHM logic sort this out. 
> When CONFIG_PCI=y|m, arch/s390/kvm/pci.o will be linked and so we can 
> compare the function handle against afit->mdd (check to see if the 
> device is emulated) and use this to determine whether or not to 
> immediately send to userspace -- And if CONFIG_VFIO_PCI=n, a SHM bit 
> will always be on and so we'll always go to userspace via this check.

So we agree.
But as I I said somewhere else I wonder if CONFIG_VFIO_PCI_ZDEV would 
not even be better here.

> > > +
> > > +    kvm_s390_get_regs_rre(vcpu, &reg1, &reg2);
> > > +
> > > +    /* If the device has a SHM bit on, let userspace take care of 
> > > this */
> > > +    if (((vcpu->run->s.regs.gprs[reg1] >> 32) & aift->mdd) != 0)
> > > +        return -EOPNOTSUPP;
> > > +
> > > +    rc = kvm_s390_pci_refresh_trans(vcpu, vcpu->run->s.regs.gprs[reg1],
> > > +                    vcpu->run->s.regs.gprs[reg2],
> > > +                    vcpu->run->s.regs.gprs[reg2+1],
> > > +                    &status);
> > > +
> > > +    switch (rc) {
> > > +    case 0:
> > > +        kvm_s390_set_psw_cc(vcpu, 0);
> > > +        break;
> > > +    case -EOPNOTSUPP:
> > > +        return -EOPNOTSUPP;
> > > +    default:
> > > +        vcpu->run->s.regs.gprs[reg1] &= 0xffffffff00ffffffUL;
> > > +        vcpu->run->s.regs.gprs[reg1] |= (u64) status << 24;
> > > +        if (status != 0)
> > > +            kvm_s390_set_psw_cc(vcpu, 1);
> > > +        else
> > > +            kvm_s390_set_psw_cc(vcpu, 3);
> > > +        break;
> > > +    }
> > > +
> > > +    return 0;
> > > +}
> > > +
> > >   #define SSKE_NQ 0x8
> > >   #define SSKE_MR 0x4
> > >   #define SSKE_MC 0x2
> > > @@ -1275,6 +1319,8 @@ int kvm_s390_handle_b9(struct kvm_vcpu *vcpu)
> > >           return handle_essa(vcpu);
> > >       case 0xaf:
> > >           return handle_pfmf(vcpu);
> > > +    case 0xd3:
> > > +        return handle_rpcit(vcpu);
> > >       default:
> > >           return -EOPNOTSUPP;
> > >       }

--
Pierre Morel
IBM Lab Boeblingen