x86/svm: Add testing for L1 intercept bug
v3 cover x86/svm: Add testing for L1 intercept bug
If a feature is not advertised to L1, L1 intercepts for instructions controlled by this feature should be ignored. Currently, the added test fails due to a bug in nested vm exit handling where vmcb12 intercepts are checked before vmcb02 intercepts, causing the #UD exception to never be injected into L2 if the L1 intercept is set. This is fixed in [0]
The first patch just adds the missing intercepts needed for testing and restructures the vmcb_control_area struct to make adding the missing intercepts less ugly. The second patch adds the test which disables all relevant features that have available instruction intercepts, and checks that the #UD exception is correctly delivered despite the L1 intercept being set.
[0] https://lore.kernel.org/all/[email protected]/
v2 -> v3:
- Added assertions to unsupported instructions test to verify that CPU does not support features enabling tested instructions [Yosry Ahmed]
- Added some comments [Yosry Ahmed]
- Changed insn_invpcid() into assembly [Yosry Ahmed]
- Added a check to the unsupported instruction test to ensure that L1's vmcb is not modified by KVM [Yosry Ahmed]
- Redacted reviewed-by tag for 2nd patch in series. I mistakingly added that
v2: https://lore.kernel.org/all/[email protected]/
v1 -> v2:
- Added save/restore helpers for all intercepts as suggested by Yosry
- Reuse invpcid_safe() for added test as suggested by Yosry
- Include '-skinit' in unittests.cfg for added test target as pointed out by Yosry
v1: https://lore.kernel.org/all/[email protected]/
Kevin Cheng (2):
x86/svm: Add missing svm intercepts x86/svm: Add unsupported instruction intercept test
lib/x86/processor.h | 1 + x86/svm.c | 18 +++- x86/svm.h | 89 ++++++++++++++++-- x86/svm_tests.c | 224 +++++++++++++++++++++++++++++++------------- x86/unittests.cfg | 11 ++- 5 files changed, 262 insertions(+), 81 deletions(-)
--
2.52.0.351.gbe84eed79e-goog
PATCH v3 1/2 x86/svm: Add missing svm intercepts
Some intercepts are missing from the KUT svm testing. Add all missing intercepts and reorganize the svm intercept definition/setting/clearing.
--- x86/svm.c | 18 +++++-- x86/svm.h | 84 ++++++++++++++++++++++++++---- x86/svm_tests.c | 136 ++++++++++++++++++++++++------------------------ 3 files changed, 159 insertions(+), 79 deletions(-)
x86/svm.c +15 -3
@@ -137,6 +137,18 @@ static void vmcb_set_seg(struct vmcb_seg *seg, u16 selector, seg->base = base; } +void vmcb_save_intercepts (struct vmcb *vmcb,u32 *saved_intercepts) +{ +for (int i =0 ; i <MAX_INTERCEPT ; i++) + saved_intercepts[i] = vmcb->control.intercept[i]; +} + +void vmcb_restore_intercepts (struct vmcb *vmcb,u32 *saved_intercepts) +{ +for (int i =0 ; i <MAX_INTERCEPT ; i++) + vmcb->control.intercept[i] = saved_intercepts[i]; +} +inline void vmmcall (void ) { asmvolatile ("vmmcall" : : :"memory" );
@@ -193,9 +205,9 @@ void vmcb_ident(struct vmcb *vmcb) save->cr2 =read_cr2 (); save->g_pat =rdmsr (MSR_IA32_CR_PAT ); save->dbgctl =rdmsr (MSR_IA32_DEBUGCTLMSR ); - ctrl->intercept = (1ULL <<INTERCEPT_VMRUN ) | - (1ULL <<INTERCEPT_VMMCALL ) | - (1ULL <<INTERCEPT_SHUTDOWN ); +vmcb_set_intercept (INTERCEPT_VMRUN ); +vmcb_set_intercept (INTERCEPT_VMMCALL ); +vmcb_set_intercept (INTERCEPT_SHUTDOWN ); ctrl->iopm_base_pa =virt_to_phys (io_bitmap); ctrl->msrpm_base_pa =virt_to_phys (msr_bitmap);
x86/svm.h +75 -9
122 lines not on this page · raw message
x86/svm_tests.c +69 -67
423 lines not on this page · raw message
Signature · 1 line
2.52.0.351.gbe84eed79e-goog
Sean Christopherson
reviewerRe: [kvm-unit-tests PATCH v3 1/2] x86/svm: Add missing svm intercepts
Quoted 2 lines · Kevin Cheng
Some intercepts are missing from the KUT svm testing. Add all missing intercepts and reorganize the svm intercept definition/setting/clearing.
Please split this up.
1. Add helpers to set/clear bits.
2. Add helpers to save/restore intercepts.
3. Reorganize
4. Add "missing" intercepts.
Maybe #2 and #3 go together? But there's _so_ much going on here I genuinely can't tell.
Paolo Bonzini
maintainerRe: [kvm-unit-tests PATCH v3 1/2] x86/svm: Add missing svm intercepts
Quoted 7 lines · Sean Christopherson
Quoted 2 lines · Kevin Cheng
Some intercepts are missing from the KUT svm testing. Add all missing intercepts and reorganize the svm intercept definition/setting/clearing.
Please split this up.
Quoted 7 lines
1. Add helpers to set/clear bits.
2. Add helpers to save/restore intercepts.
3. Reorganize
4. Add "missing" intercepts.Maybe #2 and #3 go together? But there's _so_ much going on here I genuinely can't tell.
I agree, splitting all the way can't hurt...
Paolo
PATCH v3 2/2 x86/svm: Add unsupported instruction intercept test
In reply to Kevin ChengAdd tests that expect a nested vm exit, due to an unsupported instruction, to be handled by L0 even if L1 intercepts are set for that instruction.
The new test exercises bug fixed by:
https://lore.kernel.org/all/[email protected]/
--- lib/x86/processor.h | 1 + x86/svm.h | 5 ++- x86/svm_tests.c | 88 +++++++++++++++++++++++++++++++++++++++++++++ x86/unittests.cfg | 11 +++++- 4 files changed, 103 insertions(+), 2 deletions(-)
lib/x86/processor.h +1 -0
@@ -358,6 +358,7 @@ struct x86_cpu_feature {* Extended Leafs, a.k.a. AMD defined */ #define X86_FEATURE_SVM X86_CPU_FEATURE(0x80000001, 0, ECX, 2) +#define X86_FEATURE_SKINIT X86_CPU_FEATURE(0x80000001, 0, ECX, 12)#define X86_FEATURE_PERFCTR_CORE X86_CPU_FEATURE(0x80000001, 0, ECX, 23)#define X86_FEATURE_NX X86_CPU_FEATURE(0x80000001, 0, EDX, 20)#define X86_FEATURE_GBPAGES X86_CPU_FEATURE(0x80000001, 0, EDX, 26)
x86/svm.h +4 -1
@@ -406,7 +406,10 @@ struct __attribute__ ((__packed__)) vmcb {#define SVM_EXIT_MONITOR 0x08a#define SVM_EXIT_MWAIT 0x08b#define SVM_EXIT_MWAIT_COND 0x08c -#define SVM_EXIT_NPF 0x400 +#define SVM_EXIT_XSETBV 0x08d +#define SVM_EXIT_RDPRU 0x08e +#define SVM_EXIT_INVPCID 0x0a2 +#define SVM_EXIT_NPF 0x400#define SVM_EXIT_ERR -1
x86/svm_tests.c +88 -0
100 lines not on this page · raw message
x86/unittests.cfg +10 -1
@@ -253,11 +253,20 @@ arch = x86_64 [svm] file = svm.flat smp = 2 -test_args = "-pause_filter_test" +test_args = "-pause_filter_test -svm_unsupported_instruction_intercept_test" qemu_params = -cpu max,+svm -m 4g arch = x86_64 groups = svm +# RDPRU feature name is not defined in Qemu so it can't be excluded here. This +# should be okay though since KVM does not advertise RDPRU by default anyways. +[svm_unsupported_instruction_intercept_test] +file = svm.flat +test_args = "svm_unsupported_instruction_intercept_test" +qemu_params = -cpu max,+svm,-rdtscp,-xsave,-invpcid,-skinit +arch = x86_64 +groups = svm + [svm_pause_filter] file = svm.flat test_args = pause_filter_test
Signature · 1 line
2.52.0.351.gbe84eed79e-goog
Sean Christopherson
reviewerQuoted 6 lines · Kevin Cheng
If a feature is not advertised to L1, L1 intercepts for instructions controlled by this feature should be ignored. Currently, the added test fails due to a bug in nested vm exit handling where vmcb12 intercepts are checked before vmcb02 intercepts, causing the #UD exception to never be injected into L2 if the L1 intercept is set. This is fixed in [0]
Quoted 8 lines
The first patch just adds the missing intercepts needed for testing and restructures the
vmcb_control_areastruct to make adding the missing intercepts less ugly. The second patch adds the test which disables all relevant features that have available instruction intercepts, and checks that the #UD exception is correctly delivered despite the L1 intercept being set.[...]
Applied to kvm-x86 next, even though the SKINIT test fails on the Turin system available to me. AIUI, there's a ucode bug or something for pre-production hardware?
Anyways, I mention that because if others see failures, then we'll probably need to figure out a way to automatically skip the SKINIT test (or maybe eat the #GP and treat it as XFAIL?).
[1/2] x86/svm: Add missing svm intercepts
https://github.com/kvm-x86/kvm-unit-tests/commit/8322ddeda22b [2/2] x86/svm: Add unsupported instruction intercept test https://github.com/kvm-x86/kvm-unit-tests/commit/7586f98f809e
Quoted 20 lines · Sean Christopherson
Quoted 6 lines · Kevin Cheng
If a feature is not advertised to L1, L1 intercepts for instructions controlled by this feature should be ignored. Currently, the added test fails due to a bug in nested vm exit handling where vmcb12 intercepts are checked before vmcb02 intercepts, causing the #UD exception to never be injected into L2 if the L1 intercept is set. This is fixed in [0]
Quoted 8 lines
The first patch just adds the missing intercepts needed for testing and restructures the
vmcb_control_areastruct to make adding the missing intercepts less ugly. The second patch adds the test which disables all relevant features that have available instruction intercepts, and checks that the #UD exception is correctly delivered despite the L1 intercept being set.[...]
Applied to kvm-x86 next, even though the SKINIT test fails on the Turin system available to me. AIUI, there's a ucode bug or something for pre-production hardware?
Quoted 6 lines
Anyways, I mention that because if others see failures, then we'll probably need to figure out a way to automatically skip the SKINIT test (or maybe eat the #GP and treat it as XFAIL?).
[1/2] x86/svm: Add missing svm intercepts
https://github.com/kvm-x86/kvm-unit-tests/commit/8322ddeda22b
So I believe there is no need to split that patch anymore since you already applied it, right?
Also, I think there was a newer version of this series:
https://lore.kernel.org/kvm/[email protected]/
Quoted 5 lines
[2/2] x86/svm: Add unsupported instruction intercept test
https://github.com/kvm-x86/kvm-unit-tests/commit/7586f98f809e