x86/vmexit: Add WBINVD and INVD VM-Exit latency testcases
Sean Christopherson
reviewerPATCH 1/1 x86/vmexit: Add WBINVD and INVD VM-Exit latency testcases
Add WBINVD and INVD testcase to the VM-Exit performance/latency test so that it's easy to measure latency of VM-Exits that are handled in KVM's fastpath on both Intel and AMD (INVD), and so that a direct comparison can be made to an exit with no meaningful emulation (WBINVD).
Don't create entries in x86/unittests.cfg, as running the INVD test on bare metal (or a hypervisor that emulates INVD) would likely corrupt memory (and similarly, WBINVD can have a massively negative impact on the system).
--- x86/vmexit.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
x86/vmexit.c +12 -0
@@ -36,6 +36,16 @@ static void vmcall(void) asmvolatile ("vmcall" :"+a" (a),"=b" (b),"=c" (c),"=d" (d)); } +static void wbinvd (void ) +{ + asmvolatile ("wbinvd" ); +} + +static void invd (void ) +{ + asmvolatile ("invd" ); +} +#define MSR_EFER 0xc0000080#define EFER_NX_MASK (1ull << 11)
@@ -482,6 +492,8 @@ static void toggle_cr4_pge(void)static struct test tests[] = { { cpuid_test,"cpuid" , .parallel =1 , }, { vmcall,"vmcall" , .parallel =1 , }, + { wbinvd,"wbinvd" , .parallel =1 , }, + { invd,"invd" , .parallel =1 , }, #ifdef __x86_64__ { mov_from_cr8,"mov_from_cr8" , .parallel =1 , }, { mov_to_cr8, "mov_to_cr8" , .parallel = 1, },
base-commit: af582a4ebaf7828c200dc7150aa0dbccb60b08a7
Signature · 1 line
2.52.0.rc1.455.g30608eb744-goog
Quoted 46 lines · Sean Christopherson
Add WBINVD and INVD testcase to the VM-Exit performance/latency test so that it's easy to measure latency of VM-Exits that are handled in KVM's fastpath on both Intel and AMD (INVD), and so that a direct comparison can be made to an exit with no meaningful emulation (WBINVD).
Don't create entries in x86/unittests.cfg, as running the INVD test on bare metal (or a hypervisor that emulates INVD) would likely corrupt memory (and similarly, WBINVD can have a massively negative impact on the system).
Signed-off-by Sean Christopherson <[email protected]>--- x86/vmexit.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)x86/vmexit.c +12
-0@@ -36,6 +36,16 @@ static void vmcall(void) asmvolatile ("vmcall" :"+a" (a),"=b" (b),"=c" (c),"=d" (d)); } +static void wbinvd (void ) +{ + asmvolatile ("wbinvd" ); +} + +static void invd (void ) +{ + asmvolatile ("invd" ); +} +#define MSR_EFER 0xc0000080#define EFER_NX_MASK (1ull << 11)@@ -482,6 +492,8 @@ static void toggle_cr4_pge(void)static struct test tests[] = { { cpuid_test,"cpuid" , .parallel =1 , }, { vmcall,"vmcall" , .parallel =1 , }, + { wbinvd,"wbinvd" , .parallel =1 , }, + { invd,"invd" , .parallel =1 , }, #ifdef __x86_64__ { mov_from_cr8,"mov_from_cr8" , .parallel =1 , }, { mov_to_cr8, "mov_to_cr8" , .parallel = 1, },
Quoted 1 line
base-commit:
af582a4ebaf7828c200dc7150aa0dbccb60b08a7
I have tested this on AMD EPYC-Genoa and EPYC-Turin machine on 6.18-rc5.
Both the tests work as expected.
Results:
---------------------------------------
| | wbinvd | invd |
----------------------------------------
| AMD EPYC-Genoa | 2.6k | 2.1k |
----------------------------------------
| AMD EPYC-Turin | 2.0k | 1.8k |
----------------------------------------
Feel free to add
-Manali
Quoted 11 lines · Sean Christopherson
Add WBINVD and INVD testcase to the VM-Exit performance/latency test so that it's easy to measure latency of VM-Exits that are handled in KVM's fastpath on both Intel and AMD (INVD), and so that a direct comparison can be made to an exit with no meaningful emulation (WBINVD).
Don't create entries in x86/unittests.cfg, as running the INVD test on bare metal (or a hypervisor that emulates INVD) would likely corrupt memory (and similarly, WBINVD can have a massively negative impact on the system).
[...]
Applied to kvm-x86 next, thanks!
[1/1] x86/vmexit: Add WBINVD and INVD VM-Exit latency testcases
https://github.com/kvm-x86/kvm-unit-tests/commit/70e7bc408e10