| Message ID | 1516212225-37393-2-git-send-email-pbonzini@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
2018-01-17 19:03+0100, Paolo Bonzini: > The EPT access tests require at least 41 bits of guest physical address > space. This is because the test data needs a separate PML4 entry in the > EPT table. However, many consumer-grade processors have only 36 or > 39-bit MAXPHYADDR, and the tests fail there with a page fault due to > reserved bits set in the guest (GVA->GPA) page tables. > > Add a test on MAXPHYADDR and skip the tests if they cannot run on the > system under test, and only require 40 bits to lower the requirements. > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> > --- > x86/vmx_tests.c | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > index e8c97f2..0c3b5a7 100644 > --- a/x86/vmx_tests.c > +++ b/x86/vmx_tests.c > @@ -2609,6 +2609,10 @@ static void ept_access_test_setup(void) > if (setup_ept(false)) > test_skip("EPT not supported"); > > + /* We use data->gpa = 1 << 39 so that test data has a separate pml4 entry */ > + if (cpuid_maxphyaddr() < 40) This does not help at the moment -- QEMU reports 40 even on hosts that don't have as much. It is a QEMU bug, so applied still, thanks.
----- Original Message ----- > From: "Radim Krčmář" <rkrcmar@redhat.com> > To: "Paolo Bonzini" <pbonzini@redhat.com> > Cc: kvm@vger.kernel.org > Sent: Thursday, January 25, 2018 4:11:32 PM > Subject: Re: [PATCH kvm-unit-tests] vmx: fixes for MAXPHYADDR requirement > > 2018-01-17 19:03+0100, Paolo Bonzini: > > The EPT access tests require at least 41 bits of guest physical address > > space. This is because the test data needs a separate PML4 entry in the > > EPT table. However, many consumer-grade processors have only 36 or > > 39-bit MAXPHYADDR, and the tests fail there with a page fault due to > > reserved bits set in the guest (GVA->GPA) page tables. > > > > Add a test on MAXPHYADDR and skip the tests if they cannot run on the > > system under test, and only require 40 bits to lower the requirements. > > > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> > > --- > > x86/vmx_tests.c | 6 +++++- > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > > index e8c97f2..0c3b5a7 100644 > > --- a/x86/vmx_tests.c > > +++ b/x86/vmx_tests.c > > @@ -2609,6 +2609,10 @@ static void ept_access_test_setup(void) > > if (setup_ept(false)) > > test_skip("EPT not supported"); > > > > + /* We use data->gpa = 1 << 39 so that test data has a separate pml4 entry > > */ > > + if (cpuid_maxphyaddr() < 40) > > This does not help at the moment -- QEMU reports 40 even on hosts that > don't have as much. It is a QEMU bug, so applied still, thanks. True. However, vmx.flat is run with "-cpu host,+vmx", and that should forward the host MAXPHYADDR, shouldn't it? Paolo
2018-01-25 11:36-0500, Paolo Bonzini: > ----- Original Message ----- > > From: "Radim Krčmář" <rkrcmar@redhat.com> > > To: "Paolo Bonzini" <pbonzini@redhat.com> > > Cc: kvm@vger.kernel.org > > Sent: Thursday, January 25, 2018 4:11:32 PM > > Subject: Re: [PATCH kvm-unit-tests] vmx: fixes for MAXPHYADDR requirement > > > > 2018-01-17 19:03+0100, Paolo Bonzini: > > > The EPT access tests require at least 41 bits of guest physical address > > > space. This is because the test data needs a separate PML4 entry in the > > > EPT table. However, many consumer-grade processors have only 36 or > > > 39-bit MAXPHYADDR, and the tests fail there with a page fault due to > > > reserved bits set in the guest (GVA->GPA) page tables. > > > > > > Add a test on MAXPHYADDR and skip the tests if they cannot run on the > > > system under test, and only require 40 bits to lower the requirements. > > > > > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> > > > --- > > > x86/vmx_tests.c | 6 +++++- > > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > > > index e8c97f2..0c3b5a7 100644 > > > --- a/x86/vmx_tests.c > > > +++ b/x86/vmx_tests.c > > > @@ -2609,6 +2609,10 @@ static void ept_access_test_setup(void) > > > if (setup_ept(false)) > > > test_skip("EPT not supported"); > > > > > > + /* We use data->gpa = 1 << 39 so that test data has a separate pml4 entry > > > */ > > > + if (cpuid_maxphyaddr() < 40) > > > > This does not help at the moment -- QEMU reports 40 even on hosts that > > don't have as much. It is a QEMU bug, so applied still, thanks. > > True. However, vmx.flat is run with "-cpu host,+vmx", and that should forward > the host MAXPHYADDR, shouldn't it? Would make sense for QEMU to do that, but it's not what happens. I get 40 on both 39 and 46 hosts. We can do "-cpu host,+vmx,host_phys_bits" at the cost of breaking unit tests on old QEMUs ...
----- Original Message ----- > From: "Radim Krčmář" <rkrcmar@redhat.com> > To: "Paolo Bonzini" <pbonzini@redhat.com> > Cc: kvm@vger.kernel.org > Sent: Thursday, January 25, 2018 5:45:54 PM > Subject: Re: [PATCH kvm-unit-tests] vmx: fixes for MAXPHYADDR requirement > > 2018-01-25 11:36-0500, Paolo Bonzini: > > ----- Original Message ----- > > > From: "Radim Krčmář" <rkrcmar@redhat.com> > > > To: "Paolo Bonzini" <pbonzini@redhat.com> > > > Cc: kvm@vger.kernel.org > > > Sent: Thursday, January 25, 2018 4:11:32 PM > > > Subject: Re: [PATCH kvm-unit-tests] vmx: fixes for MAXPHYADDR requirement > > > > > > 2018-01-17 19:03+0100, Paolo Bonzini: > > > > The EPT access tests require at least 41 bits of guest physical address > > > > space. This is because the test data needs a separate PML4 entry in > > > > the > > > > EPT table. However, many consumer-grade processors have only 36 or > > > > 39-bit MAXPHYADDR, and the tests fail there with a page fault due to > > > > reserved bits set in the guest (GVA->GPA) page tables. > > > > > > > > Add a test on MAXPHYADDR and skip the tests if they cannot run on the > > > > system under test, and only require 40 bits to lower the requirements. > > > > > > > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> > > > > --- > > > > x86/vmx_tests.c | 6 +++++- > > > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > > > > index e8c97f2..0c3b5a7 100644 > > > > --- a/x86/vmx_tests.c > > > > +++ b/x86/vmx_tests.c > > > > @@ -2609,6 +2609,10 @@ static void ept_access_test_setup(void) > > > > if (setup_ept(false)) > > > > test_skip("EPT not supported"); > > > > > > > > + /* We use data->gpa = 1 << 39 so that test data has a separate pml4 > > > > entry > > > > */ > > > > + if (cpuid_maxphyaddr() < 40) > > > > > > This does not help at the moment -- QEMU reports 40 even on hosts that > > > don't have as much. It is a QEMU bug, so applied still, thanks. > > > > True. However, vmx.flat is run with "-cpu host,+vmx", and that should > > forward > > the host MAXPHYADDR, shouldn't it? > > Would make sense for QEMU to do that, but it's not what happens. > I get 40 on both 39 and 46 hosts. > > We can do "-cpu host,+vmx,host_phys_bits" at the cost of breaking unit > tests on old QEMUs ... No, let's fix QEMU. Paolo
diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c index e8c97f2..0c3b5a7 100644 --- a/x86/vmx_tests.c +++ b/x86/vmx_tests.c @@ -2609,6 +2609,10 @@ static void ept_access_test_setup(void) if (setup_ept(false)) test_skip("EPT not supported"); + /* We use data->gpa = 1 << 39 so that test data has a separate pml4 entry */ + if (cpuid_maxphyaddr() < 40) + test_skip("Test needs MAXPHYADDR >= 40"); + test_set_guest(ept_access_test_guest); test_add_teardown(ept_access_test_teardown, NULL); @@ -2616,7 +2620,7 @@ static void ept_access_test_setup(void) TEST_ASSERT(data->hva); data->hpa = virt_to_phys(data->hva); - data->gpa = 1ul << 40; + data->gpa = 1ul << 39; data->gva = (void *) ALIGN((unsigned long) alloc_vpages(npages * 2), size); TEST_ASSERT(!any_present_pages(page_table, data->gva, size));
The EPT access tests require at least 41 bits of guest physical address space. This is because the test data needs a separate PML4 entry in the EPT table. However, many consumer-grade processors have only 36 or 39-bit MAXPHYADDR, and the tests fail there with a page fault due to reserved bits set in the guest (GVA->GPA) page tables. Add a test on MAXPHYADDR and skip the tests if they cannot run on the system under test, and only require 40 bits to lower the requirements. Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> --- x86/vmx_tests.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)