On Mon, Jan 11, 2021 at 02:25:59PM -0800, Sean Christopherson wrote:
> On Sun, Jan 10, 2021, Yang Weijiang wrote:
> > When the application is tested on a machine with 52bit-physical-address, the
> > synthesized 52bit GPA triggers EPT(4-Level) fast_page_fault infinitely.
>
> That doesn't sound right, KVM should use 5-level EPT if guest maxpa > 48.
> Hmm, unless the CPU doesn't support 5-level EPT, but I didn't think such CPUs
> (maxpa=52 w/o 5-level EPT) existed? Ah, but it would be possible with nested
> VMX, and initial KVM 5-level support didn't allow nested 5-level EPT. Any
> chance you're running this test in a VM with 5-level EPT disabled, but maxpa=52?
>
Hi, Sean,
Thanks for the reply!
I use default settings of the unit-test + 5.2.0 QEMU + 5.10 kernel, in
this case, QEMU uses cpu->phys_bits==40, so the guest's PA=40bit and
LA=57bit, hence 5-level EPT is not enabled. My physical machine is PA=52
and LA=57 as can checked from cpuid:
cpuid -1r -l 0x80000008 -s 0
CPU:
0x80000008 0x00: eax=0x00003934 ...
There're two other ways to w/a this issue: 1) change the QEMU params to
to extra_params = -cpu host,host-phys-bits, so guest's PA=52 and LA=57,
this will enable 5-level EPT, meanwhile, it escapes the problematic GPA
by adding AC_*_BIT51_MASK in invalid_mask.
2) add allow_smaller_maxphyaddr=1 to kvm-intel module.
the perf trace looks like this:
12481.879 qemu-system-x8/27004 kvm:kvm_page_fault:address 8000002000000
error_code 181
> > On the other hand, there's no reserved bits in 51:max_physical_address on
> > machines with 52bit-physical-address.
> >
> > Signed-off-by: Yang Weijiang <weijiang.yang@xxxxxxxxx>
> > ---
> > x86/access.c | 20 +++++++++++---------
> > 1 file changed, 11 insertions(+), 9 deletions(-)
> >
> > diff --git a/x86/access.c b/x86/access.c
> > index 7dc9eb6..bec1c4d 100644
> > --- a/x86/access.c
> > +++ b/x86/access.c
> > @@ -15,6 +15,7 @@ static _Bool verbose = false;
> > typedef unsigned long pt_element_t;
> > static int invalid_mask;
> > static int page_table_levels;
> > +static int max_phyaddr;
> >
> > #define PT_BASE_ADDR_MASK ((pt_element_t)((((pt_element_t)1 << 36) - 1) & PAGE_MASK))
> > #define PT_PSE_BASE_ADDR_MASK (PT_BASE_ADDR_MASK & ~(1ull << 21))
> > @@ -394,9 +395,10 @@ static void ac_emulate_access(ac_test_t *at, unsigned flags)
> > if (!F(AC_PDE_ACCESSED))
> > at->ignore_pde = PT_ACCESSED_MASK;
> >
> > - pde_valid = F(AC_PDE_PRESENT)
> > - && !F(AC_PDE_BIT51) && !F(AC_PDE_BIT36) && !F(AC_PDE_BIT13)
> > + pde_valid = F(AC_PDE_PRESENT) && !F(AC_PDE_BIT36) && !F(AC_PDE_BIT13)
> > && !(F(AC_PDE_NX) && !F(AC_CPU_EFER_NX));
> > + if (max_phyaddr < 52)
> > + pde_valid &= !F(AC_PDE_BIT51);
> >
> > if (!pde_valid) {
> > at->expected_fault = 1;
> > @@ -420,9 +422,10 @@ static void ac_emulate_access(ac_test_t *at, unsigned flags)
> >
> > at->expected_pde |= PT_ACCESSED_MASK;
> >
> > - pte_valid = F(AC_PTE_PRESENT)
> > - && !F(AC_PTE_BIT51) && !F(AC_PTE_BIT36)
> > + pte_valid = F(AC_PTE_PRESENT) && !F(AC_PTE_BIT36)
> > && !(F(AC_PTE_NX) && !F(AC_CPU_EFER_NX));
> > + if (max_phyaddr < 52)
> > + pte_valid &= !F(AC_PTE_BIT51);
>
> This _should_ be unnecessary. As below, AC_*_BIT51_MASK will be set in
> invalid_mask, and so ac_test_bump_one() will skip tests that try to set bit 51.
>
These code is to avoid some "unexpected access" messages on some platforms
if below change is added.
> > if (!pte_valid) {
> > at->expected_fault = 1;
> > @@ -964,13 +967,11 @@ static int ac_test_run(void)
> > shadow_cr4 = read_cr4();
> > shadow_efer = rdmsr(MSR_EFER);
> >
> > - if (cpuid_maxphyaddr() >= 52) {
> > - invalid_mask |= AC_PDE_BIT51_MASK;
> > - invalid_mask |= AC_PTE_BIT51_MASK;
> > - }
> > - if (cpuid_maxphyaddr() >= 37) {
> > + if (max_phyaddr >= 37 && max_phyaddr < 52) {
> > invalid_mask |= AC_PDE_BIT36_MASK;
> > invalid_mask |= AC_PTE_BIT36_MASK;
> > + invalid_mask |= AC_PDE_BIT51_MASK;
> > + invalid_mask |= AC_PTE_BIT51_MASK;
> > }
>
> This change is incorrect. "invalid_mask" is misleading in this context as it
> means "bits that can't be tested because they're legal". So setting the bit 51
> flags in invalid_mask if 'maxpa >= 52' is correct, as it states those tests are
> "invalid" because setting bit 51 will not fault.
Maybe I misunderstood the purpose of this test, so I skipped the
non-fault case, e.g., when maxpa >=52. In guest PA=40 case, AC_*_BIT51_MASK
bits are not added here, then when they're set in guest page_table
entries, does it expect a fault? What's the expected result on a real 52bit
platform?
> All that being said, it's also entirely possible I'm misreading this test, I've
> done it many times before :-)
>
I found it's hard to pass all the enclosed tests on various platforms :-)
> > if (this_cpu_has(X86_FEATURE_PKU)) {
> > @@ -1038,6 +1039,7 @@ int main(void)
> > int r;
> >
> > printf("starting test\n\n");
> > + max_phyaddr = cpuid_maxphyaddr();
> > page_table_levels = 4;
> > r = ac_test_run();
> >
> > --
> > 2.17.2
> >
