On Thu, 2022-10-20 at 19:02 +0000, Sean Christopherson wrote:
> On Thu, Oct 20, 2022, Maxim Levitsky wrote:
> > This adds minimum amout of code to support tests that
> > run SVM on more that one vCPU.
>
> s/that/than
>
> >
> > Signed-off-by: Maxim Levitsky <mlevitsk@xxxxxxxxxx>
> > ---
> > lib/x86/svm_lib.c | 9 +
> > lib/x86/svm_lib.h | 10 +
> > x86/svm.c | 37 ++-
> > x86/svm.h | 5 +-
> > x86/svm_npt.c | 44 ++--
> > x86/svm_tests.c | 615 +++++++++++++++++++++++-----------------------
> > 6 files changed, 362 insertions(+), 358 deletions(-)
> >
> > diff --git a/lib/x86/svm_lib.c b/lib/x86/svm_lib.c
> > index 2b067c65..1152c497 100644
> > --- a/lib/x86/svm_lib.c
> > +++ b/lib/x86/svm_lib.c
> > @@ -157,3 +157,12 @@ void vmcb_ident(struct vmcb *vmcb)
> > ctrl->tlb_ctl = TLB_CONTROL_FLUSH_ALL_ASID;
> > }
> > }
> > +
> > +void svm_vcpu_init(struct svm_vcpu *vcpu)
> > +{
> > + vcpu->vmcb = alloc_page();
> > + vmcb_ident(vcpu->vmcb);
> > + memset(&vcpu->regs, 0, sizeof(vcpu->regs));
> > + vcpu->stack = alloc_pages(4) + (PAGE_SIZE << 4);
> > + vcpu->vmcb->save.rsp = (ulong)(vcpu->stack);
> > +}
> > diff --git a/lib/x86/svm_lib.h b/lib/x86/svm_lib.h
> > index 59db26de..c6957dba 100644
> > --- a/lib/x86/svm_lib.h
> > +++ b/lib/x86/svm_lib.h
> > @@ -89,6 +89,16 @@ struct svm_extra_regs
> > u64 r15;
> > };
> >
> > +
> > +struct svm_vcpu
> > +{
> > + struct vmcb *vmcb;
> > + struct svm_extra_regs regs;
> > + void *stack;
> > +};
> > +
> > +void svm_vcpu_init(struct svm_vcpu *vcpu);
> > +
> > #define SWAP_GPRS(reg) \
> > "xchg %%rcx, 0x08(%%" reg ")\n\t" \
> > "xchg %%rdx, 0x10(%%" reg ")\n\t" \
> > diff --git a/x86/svm.c b/x86/svm.c
> > index 9484a6d1..7aa3ebd2 100644
> > --- a/x86/svm.c
> > +++ b/x86/svm.c
> > @@ -16,7 +16,7 @@
> > #include "apic.h"
> > #include "svm_lib.h"
> >
> > -struct vmcb *vmcb;
> > +struct svm_vcpu vcpu0;
>
> It's not strictly vCPU0, e.g. svm_init_intercept_test() deliberately runs on
> vCPU2, presumably to avoid running on the BSP?
>
> Since this is churning a lot of code anyways, why not clean this all up and have
> run_svm_tests() dynamically allocate state instead of relying on global data?
Makes sense.
Best regards,
Maxim Levitsky
>
