Message ID | 20200714002355.538-2-sean.j.christopherson@intel.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | nVMX: Two PCIDE related fixes | expand |
On Mon, Jul 13, 2020 at 5:23 PM Sean Christopherson <sean.j.christopherson@intel.com> wrote: > > Perform one last VMX transition to actually load the host's RIP and CR4 > at the end of test_host_addr_size(). Simply writing the VMCS doesn't > restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, > which causes spectacularly confusing explosions when other misguided > tests assume setting bit 63 in CR3 will cause a non-canonical #GP. > > Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") > Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> > Cc: Karl Heubaum <karl.heubaum@oracle.com> > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> Reviewed-by: Oliver Upton <oupton@google.com> > --- > x86/vmx_tests.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > index 29f3d0e..cb42a2d 100644 > --- a/x86/vmx_tests.c > +++ b/x86/vmx_tests.c > @@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) > vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); > vmcs_write(HOST_RIP, rip_saved); > vmcs_write(HOST_CR4, cr4_saved); > + > + /* Restore host's active RIP and CR4 values. */ > + report_prefix_pushf("restore host state"); > + test_vmx_vmlaunch(0); > + report_prefix_pop(); > } > } > > -- > 2.26.0 >
On 7/13/20 5:23 PM, Sean Christopherson wrote: > Perform one last VMX transition to actually load the host's RIP and CR4 > at the end of test_host_addr_size(). Simply writing the VMCS doesn't > restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, > which causes spectacularly confusing explosions when other misguided > tests assume setting bit 63 in CR3 will cause a non-canonical #GP. > > Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") > Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> > Cc: Karl Heubaum <karl.heubaum@oracle.com> > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > --- > x86/vmx_tests.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > index 29f3d0e..cb42a2d 100644 > --- a/x86/vmx_tests.c > +++ b/x86/vmx_tests.c > @@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) > vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); > vmcs_write(HOST_RIP, rip_saved); > vmcs_write(HOST_CR4, cr4_saved); > + > + /* Restore host's active RIP and CR4 values. */ > + report_prefix_pushf("restore host state"); > + test_vmx_vmlaunch(0); > + report_prefix_pop(); > } > } > Just for my understanding. When you say, "other misguided tests", which tests are you referring to ? In the current sequence of tests in vmx_host_state_area_test(), test_load_host_perf_global_ctrl() is the one that follows and it runs fine.
On Wed, Jul 15, 2020 at 11:34:46AM -0700, Krish Sadhukhan wrote: > > On 7/13/20 5:23 PM, Sean Christopherson wrote: > >Perform one last VMX transition to actually load the host's RIP and CR4 > >at the end of test_host_addr_size(). Simply writing the VMCS doesn't > >restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, > >which causes spectacularly confusing explosions when other misguided > >tests assume setting bit 63 in CR3 will cause a non-canonical #GP. > > > >Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") > >Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> > >Cc: Karl Heubaum <karl.heubaum@oracle.com> > >Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > >--- > > x86/vmx_tests.c | 5 +++++ > > 1 file changed, 5 insertions(+) > > > >diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > >index 29f3d0e..cb42a2d 100644 > >--- a/x86/vmx_tests.c > >+++ b/x86/vmx_tests.c > >@@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) > > vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); > > vmcs_write(HOST_RIP, rip_saved); > > vmcs_write(HOST_CR4, cr4_saved); > >+ > >+ /* Restore host's active RIP and CR4 values. */ > >+ report_prefix_pushf("restore host state"); > >+ test_vmx_vmlaunch(0); > >+ report_prefix_pop(); > > } > > } > Just for my understanding. When you say, "other misguided tests", which > tests are you referring to ? In the current sequence of tests in > vmx_host_state_area_test(), test_load_host_perf_global_ctrl() is the one > that follows and it runs fine. See test_mtf_guest() in patch 2/2. https://patchwork.kernel.org/patch/11661189/
On 7/15/20 11:48 AM, Sean Christopherson wrote: > On Wed, Jul 15, 2020 at 11:34:46AM -0700, Krish Sadhukhan wrote: >> On 7/13/20 5:23 PM, Sean Christopherson wrote: >>> Perform one last VMX transition to actually load the host's RIP and CR4 >>> at the end of test_host_addr_size(). Simply writing the VMCS doesn't >>> restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, >>> which causes spectacularly confusing explosions when other misguided >>> tests assume setting bit 63 in CR3 will cause a non-canonical #GP. >>> >>> Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") >>> Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> >>> Cc: Karl Heubaum <karl.heubaum@oracle.com> >>> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> >>> --- >>> x86/vmx_tests.c | 5 +++++ >>> 1 file changed, 5 insertions(+) >>> >>> diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c >>> index 29f3d0e..cb42a2d 100644 >>> --- a/x86/vmx_tests.c >>> +++ b/x86/vmx_tests.c >>> @@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) >>> vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); >>> vmcs_write(HOST_RIP, rip_saved); >>> vmcs_write(HOST_CR4, cr4_saved); >>> + >>> + /* Restore host's active RIP and CR4 values. */ >>> + report_prefix_pushf("restore host state"); >>> + test_vmx_vmlaunch(0); >>> + report_prefix_pop(); >>> } >>> } >> Just for my understanding. When you say, "other misguided tests", which >> tests are you referring to ? In the current sequence of tests in >> vmx_host_state_area_test(), test_load_host_perf_global_ctrl() is the one >> that follows and it runs fine. > See test_mtf_guest() in patch 2/2. https://patchwork.kernel.org/patch/11661189/ I ran the two tests as follows but couldn't reproduce it: ./x86/run x86/vmx.flat -smp 1 -cpu host,+vmx -append "vmx_host_state_area_test vmx_mtf_test" How did you run the them ?
On Wed, Jul 15, 2020 at 02:34:23PM -0700, Krish Sadhukhan wrote: > > On 7/15/20 11:48 AM, Sean Christopherson wrote: > >On Wed, Jul 15, 2020 at 11:34:46AM -0700, Krish Sadhukhan wrote: > >>On 7/13/20 5:23 PM, Sean Christopherson wrote: > >>>Perform one last VMX transition to actually load the host's RIP and CR4 > >>>at the end of test_host_addr_size(). Simply writing the VMCS doesn't > >>>restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, > >>>which causes spectacularly confusing explosions when other misguided > >>>tests assume setting bit 63 in CR3 will cause a non-canonical #GP. > >>> > >>>Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") > >>>Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> > >>>Cc: Karl Heubaum <karl.heubaum@oracle.com> > >>>Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > >>>--- > >>> x86/vmx_tests.c | 5 +++++ > >>> 1 file changed, 5 insertions(+) > >>> > >>>diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c > >>>index 29f3d0e..cb42a2d 100644 > >>>--- a/x86/vmx_tests.c > >>>+++ b/x86/vmx_tests.c > >>>@@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) > >>> vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); > >>> vmcs_write(HOST_RIP, rip_saved); > >>> vmcs_write(HOST_CR4, cr4_saved); > >>>+ > >>>+ /* Restore host's active RIP and CR4 values. */ > >>>+ report_prefix_pushf("restore host state"); > >>>+ test_vmx_vmlaunch(0); > >>>+ report_prefix_pop(); > >>> } > >>> } > >>Just for my understanding. When you say, "other misguided tests", which > >>tests are you referring to ? In the current sequence of tests in > >>vmx_host_state_area_test(), test_load_host_perf_global_ctrl() is the one > >>that follows and it runs fine. > >See test_mtf_guest() in patch 2/2. https://patchwork.kernel.org/patch/11661189/ > > I ran the two tests as follows but couldn't reproduce it: > > ./x86/run x86/vmx.flat -smp 1 -cpu host,+vmx -append > "vmx_host_state_area_test vmx_mtf_test" > > > How did you run the them ? I ran the VMX testcase from x86/unittest.cfg (below) on HSW. I eventually narrowed it down to just test_host_addr_size() and the MTF test. Note, the failure signature will change depending on whether vmx_cr_load_test() is run between those two. If it's not run, the failure is a straightforward triple fault. If it is run, for me the failure morphed into a an emulation error because the unit test was able to generate a valid translation out of CR3=0 and hit a non-existent memslot, which was all kinds of confusing. ./x86/run x86/vmx.flat -smp 1 -cpu host,+vmx -append "-exit_monitor_from_l2_test -ept_access* -vmx_smp* -vmx_vmcs_shadow_test -atomic_switch_overflow_msrs_test -vmx_init_signal_test -vmx_apic_passthrough_tpr_threshold_test"
On 7/15/20 3:22 PM, Sean Christopherson wrote: > On Wed, Jul 15, 2020 at 02:34:23PM -0700, Krish Sadhukhan wrote: >> On 7/15/20 11:48 AM, Sean Christopherson wrote: >>> On Wed, Jul 15, 2020 at 11:34:46AM -0700, Krish Sadhukhan wrote: >>>> On 7/13/20 5:23 PM, Sean Christopherson wrote: >>>>> Perform one last VMX transition to actually load the host's RIP and CR4 >>>>> at the end of test_host_addr_size(). Simply writing the VMCS doesn't >>>>> restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, >>>>> which causes spectacularly confusing explosions when other misguided >>>>> tests assume setting bit 63 in CR3 will cause a non-canonical #GP. >>>>> >>>>> Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") >>>>> Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> >>>>> Cc: Karl Heubaum <karl.heubaum@oracle.com> >>>>> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> >>>>> --- >>>>> x86/vmx_tests.c | 5 +++++ >>>>> 1 file changed, 5 insertions(+) >>>>> >>>>> diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c >>>>> index 29f3d0e..cb42a2d 100644 >>>>> --- a/x86/vmx_tests.c >>>>> +++ b/x86/vmx_tests.c >>>>> @@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) >>>>> vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); >>>>> vmcs_write(HOST_RIP, rip_saved); >>>>> vmcs_write(HOST_CR4, cr4_saved); >>>>> + >>>>> + /* Restore host's active RIP and CR4 values. */ >>>>> + report_prefix_pushf("restore host state"); >>>>> + test_vmx_vmlaunch(0); >>>>> + report_prefix_pop(); >>>>> } >>>>> } >>>> Just for my understanding. When you say, "other misguided tests", which >>>> tests are you referring to ? In the current sequence of tests in >>>> vmx_host_state_area_test(), test_load_host_perf_global_ctrl() is the one >>>> that follows and it runs fine. >>> See test_mtf_guest() in patch 2/2. https://patchwork.kernel.org/patch/11661189/ >> I ran the two tests as follows but couldn't reproduce it: >> >> ./x86/run x86/vmx.flat -smp 1 -cpu host,+vmx -append >> "vmx_host_state_area_test vmx_mtf_test" >> >> >> How did you run the them ? > I ran the VMX testcase from x86/unittest.cfg (below) on HSW. I eventually > narrowed it down to just test_host_addr_size() and the MTF test. Note, the > failure signature will change depending on whether vmx_cr_load_test() is > run between those two. If it's not run, the failure is a straightforward > triple fault. If it is run, for me the failure morphed into a an emulation > error because the unit test was able to generate a valid translation out of > CR3=0 and hit a non-existent memslot, which was all kinds of confusing. > > ./x86/run x86/vmx.flat -smp 1 -cpu host,+vmx -append "-exit_monitor_from_l2_test -ept_access* -vmx_smp* -vmx_vmcs_shadow_test -atomic_switch_overflow_msrs_test -vmx_init_signal_test -vmx_apic_passthrough_tpr_threshold_test" Thanks. I see it now, after I comment out test_load_host_perf_global_ctrl(). If any test calls enter_guest() right after test_host_addr_size(), this problem will manifest. I didn't think about this sequence of tests when adding test_host_addr_size() or any host-state-area tests for that matter.
diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c index 29f3d0e..cb42a2d 100644 --- a/x86/vmx_tests.c +++ b/x86/vmx_tests.c @@ -7673,6 +7673,11 @@ static void test_host_addr_size(void) vmcs_write(ENT_CONTROLS, entry_ctrl_saved | ENT_GUEST_64); vmcs_write(HOST_RIP, rip_saved); vmcs_write(HOST_CR4, cr4_saved); + + /* Restore host's active RIP and CR4 values. */ + report_prefix_pushf("restore host state"); + test_vmx_vmlaunch(0); + report_prefix_pop(); } }
Perform one last VMX transition to actually load the host's RIP and CR4 at the end of test_host_addr_size(). Simply writing the VMCS doesn't restore the values in hardware, e.g. as is, CR4.PCIDE can be left set, which causes spectacularly confusing explosions when other misguided tests assume setting bit 63 in CR3 will cause a non-canonical #GP. Fixes: 0786c0316ac05 ("kvm-unit-test: nVMX: Check Host Address Space Size on vmentry of nested guests") Cc: Krish Sadhukhan <krish.sadhukhan@oracle.com> Cc: Karl Heubaum <karl.heubaum@oracle.com> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> --- x86/vmx_tests.c | 5 +++++ 1 file changed, 5 insertions(+)