| Message ID | 20210818000905.1111226-15-zixuanwang@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | x86_64 UEFI and AMD SEV/SEV-ES support | expand |
On Wed, Aug 18, 2021, Zixuan Wang wrote: > AMD SEV-ES introduces a new #VC exception that handles the communication > between guest and host. UEFI already implements a #VC handler so there > is no need to re-implement it in KVM-Unit-Tests. To reuse this #VC > handler, this commit reads UEFI's IDT, copy the #VC IDT entry into > KVM-Unit-Tests' IDT. > > In this commit, load_idt() can work and now guest crashes in > setup_page_table(), which will be fixed by follow-up commits. As a stop gap to get SEV testing of any kind enabled, I think piggybacking the vBIOS #VC handler is a great idea. But long term, kvm-unit-tests absolutely needs to have its own #VC handler. In addition to the downsides Joerg pointed out[*], relying on an external #VC introduces the possibility of test failures that are tied to the vBIOS being used. Such dependencies already exist to some extent, e.g. using a buggy QEMU or SeaBIOS could certainly introduce failures, but those components are far more mature and less likely to break in weird ways unique to a specific test. Another potential issue is that it's possible vBIOS will be enlightened to the point where it _never_ expects a #VC, e.g. does #VMGEXIT directly, and thus panics on any #VC instead of requesting the necessary emulation. Fixing the vBIOS image in the repo would mostly solve those problems, but it wouldn't solve the lack of flexibility for the #VC handler, and debugging a third party #VC handler would likely be far more difficult to debug when failures inevitably occur. So, if these shenanigans give us test coverage now instead of a few months from now, than I say go for it. But, we need clear line of sight to a "native" #VC handler, confidence that it will actually get written in a timely manner, and an easily reverted set of commits to unwind all of the UEFI stuff. [*] https://lkml.kernel.org/r/YRuURERGp8CQ1jAX@suse.de > Signed-off-by: Zixuan Wang <zixuanwang@google.com> > --- > lib/x86/amd_sev.c | 10 ++++++++++ > lib/x86/amd_sev.h | 5 +++++ > 2 files changed, 15 insertions(+) > > diff --git a/lib/x86/amd_sev.c b/lib/x86/amd_sev.c > index c2aebdf..04b6912 100644 > --- a/lib/x86/amd_sev.c > +++ b/lib/x86/amd_sev.c > @@ -46,11 +46,21 @@ EFI_STATUS setup_amd_sev(void) > > #ifdef CONFIG_AMD_SEV_ES > EFI_STATUS setup_amd_sev_es(void){ > + struct descriptor_table_ptr idtr; > + idt_entry_t *idt; > + > /* Test if SEV-ES is enabled */ > if (!(rdmsr(MSR_SEV_STATUS) & SEV_ES_ENABLED_MASK)) { > return EFI_UNSUPPORTED; > } > > + /* Copy UEFI's #VC IDT entry, so KVM-Unit-Tests can reuse it and does Nit, multiline comments should have a leading bare /*, i.e. /* * Copy .... * not have to ... > + * not have to re-implement a #VC handler > + */ > + sidt(&idtr); > + idt = (idt_entry_t *)idtr.base; > + boot_idt[SEV_ES_VC_HANDLER_VECTOR] = idt[SEV_ES_VC_HANDLER_VECTOR]; > + > return EFI_SUCCESS; > } > > diff --git a/lib/x86/amd_sev.h b/lib/x86/amd_sev.h > index 4d81cae..5ebd4a6 100644 > --- a/lib/x86/amd_sev.h > +++ b/lib/x86/amd_sev.h > @@ -36,6 +36,11 @@ > #define SEV_ENABLED_MASK 0b1 > #define SEV_ES_ENABLED_MASK 0b10 > > +/* AMD Programmer's Manual Volume 2 > + * - Section "#VC Exception" > + */ > +#define SEV_ES_VC_HANDLER_VECTOR 29 > + > EFI_STATUS setup_amd_sev(void); > #ifdef CONFIG_AMD_SEV_ES > EFI_STATUS setup_amd_sev_es(void); > -- > 2.33.0.rc1.237.g0d66db33f3-goog >
On Fri, Aug 20, 2021 at 4:50 PM Sean Christopherson <seanjc@google.com> wrote: > > On Wed, Aug 18, 2021, Zixuan Wang wrote: > > AMD SEV-ES introduces a new #VC exception that handles the communication > > between guest and host. UEFI already implements a #VC handler so there > > is no need to re-implement it in KVM-Unit-Tests. To reuse this #VC > > handler, this commit reads UEFI's IDT, copy the #VC IDT entry into > > KVM-Unit-Tests' IDT. > > > > In this commit, load_idt() can work and now guest crashes in > > setup_page_table(), which will be fixed by follow-up commits. > > As a stop gap to get SEV testing of any kind enabled, I think piggybacking the > vBIOS #VC handler is a great idea. But long term, kvm-unit-tests absolutely > needs to have its own #VC handler. > > In addition to the downsides Joerg pointed out[*], relying on an external #VC > introduces the possibility of test failures that are tied to the vBIOS being > used. Such dependencies already exist to some extent, e.g. using a buggy QEMU or > SeaBIOS could certainly introduce failures, but those components are far more > mature and less likely to break in weird ways unique to a specific test. > > Another potential issue is that it's possible vBIOS will be enlightened to the > point where it _never_ expects a #VC, e.g. does #VMGEXIT directly, and thus panics > on any #VC instead of requesting the necessary emulation. > > Fixing the vBIOS image in the repo would mostly solve those problems, but it > wouldn't solve the lack of flexibility for the #VC handler, and debugging a third > party #VC handler would likely be far more difficult to debug when failures > inevitably occur. > > So, if these shenanigans give us test coverage now instead of a few months from > now, than I say go for it. But, we need clear line of sight to a "native" #VC > handler, confidence that it will actually get written in a timely manner, and an > easily reverted set of commits to unwind all of the UEFI stuff. > > [*] https://lkml.kernel.org/r/YRuURERGp8CQ1jAX@suse.de Understood. I must admit, we didn’t have this long term perspective when drafting these patches. But after reading this feedback, we see your point. Luckily, unwinding the code to install the UEFI #VC handler is trivial. Also, we do believe that completing and submitting this patch series such that it uses the UEFI’s #VC handler is the best next step, even with the understanding that it’s not where we want to be six months to one year from now. The reason is that adding a new #VC handler is non-trivial. It seems like a separate patch set. At the same time, using the UEFI #VC handler unlocks a lot of testing (that’s totally non-existent now) and it should be trivial to plumb in the (to be written) kvm-unit-tests #VC handler. In other words, with this patch the community can start using and adding to the tests that are unlocked by the UEFI #VC handler while someone (in parallel) works on a follow-on patch set to add a #VC handler to kvm-unit-tests.
> On Fri, Aug 20, 2021 at 4:50 PM Sean Christopherson <seanjc@google.com> wrote: > > Signed-off-by: Zixuan Wang <zixuanwang@google.com> > > --- > > lib/x86/amd_sev.c | 10 ++++++++++ > > lib/x86/amd_sev.h | 5 +++++ > > 2 files changed, 15 insertions(+) > > > > diff --git a/lib/x86/amd_sev.c b/lib/x86/amd_sev.c > > index c2aebdf..04b6912 100644 > > --- a/lib/x86/amd_sev.c > > +++ b/lib/x86/amd_sev.c > > @@ -46,11 +46,21 @@ EFI_STATUS setup_amd_sev(void) > > > > #ifdef CONFIG_AMD_SEV_ES > > EFI_STATUS setup_amd_sev_es(void){ > > + struct descriptor_table_ptr idtr; > > + idt_entry_t *idt; > > + > > /* Test if SEV-ES is enabled */ > > if (!(rdmsr(MSR_SEV_STATUS) & SEV_ES_ENABLED_MASK)) { > > return EFI_UNSUPPORTED; > > } > > > > + /* Copy UEFI's #VC IDT entry, so KVM-Unit-Tests can reuse it and does > > Nit, multiline comments should have a leading bare /*, i.e. > > /* > * Copy .... > * not have to ... Thanks for pointing this out, I will fix this issue in the next version.
diff --git a/lib/x86/amd_sev.c b/lib/x86/amd_sev.c index c2aebdf..04b6912 100644 --- a/lib/x86/amd_sev.c +++ b/lib/x86/amd_sev.c @@ -46,11 +46,21 @@ EFI_STATUS setup_amd_sev(void) #ifdef CONFIG_AMD_SEV_ES EFI_STATUS setup_amd_sev_es(void){ + struct descriptor_table_ptr idtr; + idt_entry_t *idt; + /* Test if SEV-ES is enabled */ if (!(rdmsr(MSR_SEV_STATUS) & SEV_ES_ENABLED_MASK)) { return EFI_UNSUPPORTED; } + /* Copy UEFI's #VC IDT entry, so KVM-Unit-Tests can reuse it and does + * not have to re-implement a #VC handler + */ + sidt(&idtr); + idt = (idt_entry_t *)idtr.base; + boot_idt[SEV_ES_VC_HANDLER_VECTOR] = idt[SEV_ES_VC_HANDLER_VECTOR]; + return EFI_SUCCESS; } diff --git a/lib/x86/amd_sev.h b/lib/x86/amd_sev.h index 4d81cae..5ebd4a6 100644 --- a/lib/x86/amd_sev.h +++ b/lib/x86/amd_sev.h @@ -36,6 +36,11 @@ #define SEV_ENABLED_MASK 0b1 #define SEV_ES_ENABLED_MASK 0b10 +/* AMD Programmer's Manual Volume 2 + * - Section "#VC Exception" + */ +#define SEV_ES_VC_HANDLER_VECTOR 29 + EFI_STATUS setup_amd_sev(void); #ifdef CONFIG_AMD_SEV_ES EFI_STATUS setup_amd_sev_es(void);
AMD SEV-ES introduces a new #VC exception that handles the communication between guest and host. UEFI already implements a #VC handler so there is no need to re-implement it in KVM-Unit-Tests. To reuse this #VC handler, this commit reads UEFI's IDT, copy the #VC IDT entry into KVM-Unit-Tests' IDT. In this commit, load_idt() can work and now guest crashes in setup_page_table(), which will be fixed by follow-up commits. Signed-off-by: Zixuan Wang <zixuanwang@google.com> --- lib/x86/amd_sev.c | 10 ++++++++++ lib/x86/amd_sev.h | 5 +++++ 2 files changed, 15 insertions(+)