KVM: selftests: Fix hyperv_features test failure when built on Clang
PATCH 1/1 KVM: selftests: Fix hyperv_features test failure when built on Clang
hyperv_features test fails when built on Clang. It throws error:
Failed guest assert: !hcall->ud_expected || res == hcall->expect at x86_64/hyperv_features.c:90
On GCC, EAX is set to 0 before the hypercall whereas in Clang it is not, this causes EAX to have garbage value when hypercall is returned in Clang binary.
Fix by executing the guest assertion only when ud_expected is false.
--- tools/testing/selftests/kvm/x86_64/hyperv_features.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
tools/testing/selftests/kvm/x86_64/hyperv_features.c +4 -4
@@ -81,13 +81,13 @@ static void guest_hcall(vm_vaddr_t pgs_gpa, struct hcall_data *hcall) } vector =hypercall (hcall->control, input, output, &res); -if (hcall->ud_expected) +if (hcall->ud_expected) {GUEST_ASSERT_2 (vector ==UD_VECTOR , hcall->control, vector); -else + }else {GUEST_ASSERT_2 (!vector, hcall->control, vector); +GUEST_ASSERT_2 (res == hcall->expect, hcall->expect, res); + } -GUEST_ASSERT_2 (!hcall->ud_expected || res == hcall->expect, - hcall->expect, res);GUEST_DONE (); }
Signature · 1 line
2.37.3.968.ga6b4b080e4-goog
Quoted 12 lines · Vipin Sharma
hyperv_featurestest fails when built on Clang. It throws error:Failed guest assert: !hcall->ud_expected || res == hcall->expect at x86_64/hyperv_features.c:90On GCC, EAX is set to 0 before the hypercall whereas in Clang it is not, this causes EAX to have garbage value when hypercall is returned in Clang binary.
Fix by executing the guest assertion only when
ud_expectedis false.
Quoted 7 lines
Fixes cc5851c6be86 ("KVM: selftests: Use exception fixup for #UD/#GP Hyper-V MSR/hcall tests")Signed-off-by Vipin Sharma <[email protected]>Suggested-by Sean Christopherson <[email protected]>--- tools/testing/selftests/kvm/x86_64/hyperv_features.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
In case Sean doesn't point it out, be wary of starting a shortlog with "Fix." You may later regret it.
Also, I think the "clang" part is a red herring. You are fixing a latent bug in the code.
Sean Christopherson
reviewerQuoted 4 lines · Jim Mattson
Quoted 3 lines · Vipin Sharma
hyperv_featurestest fails when built on Clang. It throws error:
Quoted 8 lines
Quoted 8 lines
Failed guest assert: !hcall->ud_expected || res == hcall->expect at x86_64/hyperv_features.c:90On GCC, EAX is set to 0 before the hypercall whereas in Clang it is not, this causes EAX to have garbage value when hypercall is returned in Clang binary.
Fix by executing the guest assertion only when
ud_expectedis false.
TL;DR: please rewrite the changelog to explain the actual bug (checking the result when the hypercall is expected to fault) and the fix, and only mention the gcc vs. clang behavior as a footnote.
As Jim pointed out, the bug has nothing to do with clang. Ha, figured out why gcc passes; it uses RAX as the scratch reg that the asm blob loads into R8, i.e. loads RAX with @output_address. So ignore my earlier suggestion of:
*hv_status = -EFAULT,
even better is to do:
tools/testing/selftests/kvm/x86_64/hyperv_features.c +2 -1
@@ -26,7 +26,8 @@ static inline uint8_t hypercall(u64 control, vm_vaddr_t input_address, : "=a" (*hv_status), "+c" (control), "+d" (input_address), KVM_ASM_SAFE_OUTPUTS(vector) - : [output_address] "r"(output_address) + : [output_address] "r"(output_address), + "a"(-EFAULT) : "cc", "memory", "r8", KVM_ASM_SAFE_CLOBBERS);return vector; }
so that there is zero chance of getting a false positive due to the compiler (but not KVM) modifying RAX.
Anyways, this bug is not clang's fault, with above patch it fails as "expected".
==== Test Assertion Failure ====
x86_64/hyperv_features.c:622: false
pid=283847 tid=283847 errno=4 - Interrupted system call
1 0x0000000000402842: guest_test_hcalls_access at hyperv_features.c:622
2 (inlined by) main at hyperv_features.c:642
3 0x00007f23fc513082: ?? ??:0
4 0x0000000000402ebd: _start at ??:? Failed guest assert: !hcall->ud_expected || res == hcall->expect at x86_64/hyperv_features.c:90
arg1 = 0, arg2 = fffffff2
Quoted 10 lines
Quoted 7 lines
Fixes cc5851c6be86 ("KVM: selftests: Use exception fixup for #UD/#GP Hyper-V MSR/hcall tests")Signed-off-by Vipin Sharma <[email protected]>Suggested-by Sean Christopherson <[email protected]>--- tools/testing/selftests/kvm/x86_64/hyperv_features.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)In case Sean doesn't point it out, be wary of starting a shortlog with "Fix." You may later regret it.
Heh, I think our record is "Really, really fix xyz" for a shortlog.
Quoted 2 lines
Also, I think the "clang" part is a red herring. You are fixing a latent bug in the code.
Ya, it's definitely a good idea to call out why a bug was missed, but clang is not to blame here, at all.
Quoted 13 lines · Sean Christopherson
Quoted 11 lines · Jim Mattson
Quoted 10 lines · Vipin Sharma
hyperv_featurestest fails when built on Clang. It throws error:Failed guest assert: !hcall->ud_expected || res == hcall->expect at x86_64/hyperv_features.c:90On GCC, EAX is set to 0 before the hypercall whereas in Clang it is not, this causes EAX to have garbage value when hypercall is returned in Clang binary.
Quoted 5 lines
Quoted 1 line
Quoted 1 line
Fix by executing the guest assertion only when
ud_expectedis false.TL;DR: please rewrite the changelog to explain the actual bug (checking the result when the hypercall is expected to fault) and the fix, and only mention the gcc vs. clang behavior as a footnote.
On it.
Quoted 24 lines
As Jim pointed out, the bug has nothing to do with clang. Ha, figured out why gcc passes; it uses RAX as the scratch reg that the asm blob loads into R8, i.e. loads RAX with @output_address. So ignore my earlier suggestion of:
*hv_status=-EFAULT,even better is to do:
tools/testing/selftests/kvm/x86_64/hyperv_features.c +2
-1@@ -26,7 +26,8 @@ static inline uint8_t hypercall(u64 control, vm_vaddr_t input_address, : "=a" (*hv_status), "+c" (control), "+d" (input_address), KVM_ASM_SAFE_OUTPUTS(vector) - : [output_address] "r"(output_address) + : [output_address] "r"(output_address), + "a"(-EFAULT) : "cc", "memory", "r8", KVM_ASM_SAFE_CLOBBERS);return vector; }
Quoted 3 lines
so that there is zero chance of getting a false positive due to the compiler (but not KVM) modifying RAX.
Taking it.
Quoted 5 lines
Anyways, this bug is not clang's fault, with above patch it fails as "expected".
==== Test Assertion Failure ====
x86_64/hyperv_features.c:622: false
pid=283847 tid=283847 errno=4 - Interrupted system call
Quoted 18 lines
1 0x0000000000402842: guest_test_hcalls_access at hyperv_features.c:622 2 (inlined by) main at hyperv_features.c:642 3 0x00007f23fc513082: ?? ??:0 4 0x0000000000402ebd: _start at ??:?Failed guest assert: !hcall->ud_expected || res ==
hcall->expectatx86_64/hyperv_features.c:90
arg1 = 0, arg2 =fffffff2Quoted 10 lines
Quoted 7 lines
Fixes cc5851c6be86 ("KVM: selftests: Use exception fixup for #UD/#GP Hyper-V MSR/hcall tests")Signed-off-by Vipin Sharma <[email protected]>Suggested-by Sean Christopherson <[email protected]>--- tools/testing/selftests/kvm/x86_64/hyperv_features.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)In case Sean doesn't point it out, be wary of starting a shortlog with "Fix." You may later regret it.
I am doing it already! :D
Quoted 8 lines
Heh, I think our record is "Really, really fix xyz" for a shortlog.
Quoted 2 lines
Also, I think the "clang" part is a red herring. You are fixing a latent bug in the code.
Ya, it's definitely a good idea to call out why a bug was missed, but clang is not to blame here, at all.
I agree, my understanding was not complete for this bug. I will send out a v2 with an updated shortlog. Thanks, Jim and Sean, for pointing it out.