pmu test bugs fix and improvements
v5 cover pmu test bugs fix and improvements
Changes:
v4 -> v5:
* Fix the building errors with configuration options "--arch=i386" or "--enable-efi' introduced by v4 patch "Add IBPB indirect jump asm blob" (Patch 15/18)
* Add patch 18/18 to print count details if test case fails
All changes pass validation on Intel Sapphire Rapids and Emerald Rapids platforms against latest kvm-x86/next code (0ce958282e66). No tests on AMD platforms since no AMD platform on hand. Any tests on AMD platform are appreciated.
History:
v4: https://lore.kernel.org/all/[email protected]/
v3: https://lore.kernel.org/lkml/[email protected]/
v2: https://lore.kernel.org/lkml/[email protected]/
v1: https://lore.kernel.org/lkml/[email protected]/
Dapeng Mi (17):
x86: pmu: Remove blank line and redundant space
x86: pmu: Refine fixed_events[] names
x86: pmu: Fix the issue that pmu_counter_t.config crosses cache line
x86: pmu: Enlarge cnt[] length to 48 in check_counters_many()
x86: pmu: Add asserts to warn inconsistent fixed events and counters
x86: pmu: Fix cycles event validation failure
x86: pmu: Use macro to replace hard-coded branches event index
x86: pmu: Use macro to replace hard-coded ref-cycles event index
x86: pmu: Use macro to replace hard-coded instructions event index
x86: pmu: Enable and disable PMCs in loop() asm blob
x86: pmu: Improve instruction and branches events verification
x86: pmu: Improve LLC misses event verification
x86: pmu: Adjust lower boundary of llc-misses event to 0 for legacy
CPUs
x86: pmu: Add IBPB indirect jump asm blob
x86: pmu: Adjust lower boundary of branch-misses event
x86: pmu: Optimize emulated instruction validation
x86: pmu: Print measured event count if test fails
Xiong Zhang (1):
x86: pmu: Remove duplicate code in pmu_init()
lib/x86/pmu.c | 5 - x86/pmu.c | 406 ++++++++++++++++++++++++++++++++++++++++---------- 2 files changed, 328 insertions(+), 83 deletions(-)
base-commit: d301d0187f5db09531a1c2c7608997cc3b0a5c7d
Signature · 1 line
2.40.1
PATCH v5 1/18 x86: pmu: Remove duplicate code in pmu_init()
From: Xiong Zhang <[email protected]>
There are totally same code in pmu_init() helper, remove the duplicate code.
--- lib/x86/pmu.c | 5 ----- 1 file changed, 5 deletions(-)
lib/x86/pmu.c +0 -5
@@ -16,11 +16,6 @@ void pmu_init(void) pmu.fixed_counter_width = (cpuid_10.d >>5 ) &0xff ; } -if (pmu.version >1 ) { - pmu.nr_fixed_counters = cpuid_10.d &0x1f ; - pmu.fixed_counter_width = (cpuid_10.d >>5 ) &0xff ; - } - pmu.nr_gp_counters = (cpuid_10.a >>8 ) &0xff ; pmu.gp_counter_width = (cpuid_10.a >>16 ) &0xff ; pmu.gp_counter_mask_length = (cpuid_10.a >>24 ) &0xff ;
Signature · 1 line
2.40.1
PATCH v5 2/18 x86: pmu: Remove blank line and redundant space
In reply to Dapeng Micode style changes.
--- x86/pmu.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
x86/pmu.c +1 -2
@@ -205,8 +205,7 @@ static noinline void __measure(pmu_counter_t *evt, uint64_t count)static bool verify_event (uint64_t count,struct pmu_event *e) {// printf("%d <= %ld <= %d\n", e->min, count, e->max); -return count >= e->min && count <= e->max; - +return count >= e->min && count <= e->max; }static bool verify_counter (pmu_counter_t *cnt)
Signature · 1 line
2.40.1
PATCH v5 3/18 x86: pmu: Refine fixed_events[] names
In reply to Dapeng MiIn SDM the fixed counter is numbered from 0 but currently the fixed_events names are numbered from 1. It would cause confusion for users. So Change the fixed_events[] names to number from 0 as well and keep identical with SDM.
--- x86/pmu.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
x86/pmu.c +3 -3
@@ -45,9 +45,9 @@ struct pmu_event { {"branches", 0x00c2, 1*N, 1.1*N}, {"branch misses", 0x00c3, 0, 0.1*N}, }, fixed_events[] = { - {"fixed 1", MSR_CORE_PERF_FIXED_CTR0, 10*N, 10.2*N}, - {"fixed 2", MSR_CORE_PERF_FIXED_CTR0 + 1, 1*N, 30*N}, - {"fixed 3", MSR_CORE_PERF_FIXED_CTR0 + 2, 0.1*N, 30*N} + {"fixed 0", MSR_CORE_PERF_FIXED_CTR0, 10*N, 10.2*N}, + {"fixed 1", MSR_CORE_PERF_FIXED_CTR0 + 1, 1*N, 30*N}, + {"fixed 2", MSR_CORE_PERF_FIXED_CTR0 + 2, 0.1*N, 30*N} };char *buf;
Signature · 1 line
2.40.1
PATCH v5 4/18 x86: pmu: Fix the issue that pmu_counter_t.config crosses cache line
In reply to Dapeng MiWhen running pmu test on SPR, the following #GP fault is reported.
Unhandled exception 13 #GP at ip 000000000040771f
error_code=0000 rflags=00010046 cs=00000008
rax=00000000004031ad rcx=0000000000000186 rdx=0000000000000000 rbx=00000000005142f0 rbp=0000000000514260 rsi=0000000000000020 rdi=0000000000000340
r8=0000000000513a65 r9=00000000000003f8 r10=000000000000000d r11=00000000ffffffff
r12=000000000043003c r13=0000000000514450 r14=000000000000000b r15=0000000000000001 cr0=0000000080010011 cr2=0000000000000000 cr3=0000000001007000 cr4=0000000000000020 cr8=0000000000000000
STACK: @40771f 40040e 400976 400aef 40148d 401da9 4001ad
FAIL pmu
It looks EVENTSEL0 MSR (0x186) is written a invalid value (0x4031ad) and cause a #GP.
Further investigation shows the #GP is caused by below code in __start_event().
rmsr(MSR_GP_EVENT_SELECTx(event_to_global_idx(evt)),
evt->config | EVNTSEL_EN);
The evt->config is correctly initialized but seems corrupted before writing to MSR.
The original pmu_counter_t layout looks as below.
typedef struct {
uint32_t ctr;
uint64_t config;
uint64_t count;
int idx;
} pmu_counter_t;Obviously the config filed crosses two cache lines. When the two cache lines are not updated simultaneously, the config value is corrupted.
Adjust pmu_counter_t fields order and ensure config field is cache-line aligned.
--- x86/pmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
x86/pmu.c +1 -1
@@ -21,9 +21,9 @@typedef struct {uint32_t ctr; +uint32_t idx;uint64_t config;uint64_t count; -int idx; }pmu_counter_t ;struct pmu_event {
Signature · 1 line
2.40.1
PATCH v5 5/18 x86: pmu: Enlarge cnt[] length to 48 in check_counters_many()
In reply to Dapeng MiConsidering there are already 8 GP counters and 4 fixed counters on latest Intel processors, like Sapphire Rapids. The original cnt[] array length 10 is definitely not enough to cover all supported PMU counters on these new processors even through currently KVM only supports 3 fixed counters at most. This would cause out of bound memory access and may trigger false alarm on PMU counter validation
It's probably more and more GP and fixed counters are introduced in the future and then directly extends the cnt[] array length to 48 once and for all. Base on the layout of IA32_PERF_GLOBAL_CTRL and IA32_PERF_GLOBAL_STATUS, 48 looks enough in near feature.
--- x86/pmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
x86/pmu.c +1 -1
@@ -255,7 +255,7 @@ static void check_fixed_counters(void)static void check_counters_many (void ) { -pmu_counter_t cnt[10 ]; +pmu_counter_t cnt[48 ];int i, n;for (i =0 , n =0 ; n < pmu.nr_gp_counters; i++) {
Signature · 1 line
2.40.1
PATCH v5 6/18 x86: pmu: Add asserts to warn inconsistent fixed events and counters
In reply to Dapeng MiCurrent PMU code deosn't check whether PMU fixed counter number is larger than pre-defined fixed events. If so, it would cause memory access out of range.
So add assert to warn this invalid case.
--- x86/pmu.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
x86/pmu.c +8 -2
@@ -113,8 +113,12 @@ static struct pmu_event* get_counter_event(pmu_counter_t *cnt)for (i =0 ; i < gp_events_size; i++)if (gp_events[i].unit_sel == (cnt->config &0xffff ))return &gp_events[i]; - }else -return &fixed_events[cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ]; + }else { +unsigned int idx = cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ; + +assert (idx <ARRAY_SIZE (fixed_events)); +return &fixed_events[idx]; + }return (void *)0 ; }
@@ -740,6 +744,8 @@ int main(int ac, char **av)printf ("Fixed counters: %d\n" , pmu.nr_fixed_counters);printf ("Fixed counter width: %d\n" , pmu.fixed_counter_width); +assert (pmu.nr_fixed_counters <=ARRAY_SIZE (fixed_events)); +apic_write (APIC_LVTPC ,PMI_VECTOR );check_counters ();
Signature · 1 line
2.40.1
Re: [Patch v5 06/18] x86: pmu: Add asserts to warn inconsistent fixed events and counters
Quoted 18 lines · Dapeng Mi
Current PMU code deosn't check whether PMU fixed counter number is larger than pre-defined fixed events. If so, it would cause memory access out of range.
So add assert to warn this invalid case.
Reviewed-by Mingwei Zhang <[email protected]>Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)x86/pmu.c +0
-0
x86/pmu.c static struct pmu_event* get_counter_event(pmu_counter_t *cnt) · line 119
Quoted 9 lines
x86/pmu.c +4
-2@@ -113,8 +113,12 @@ static struct pmu_event* get_counter_event(pmu_counter_t *cnt)for (i =0 ; i < gp_events_size; i++)if (gp_events[i].unit_sel == (cnt->config &0xffff ))return &gp_events[i]; - }else -return &fixed_events[cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ]; + }else { +unsigned int idx = cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ; + +assert (idx <ARRAY_SIZE (fixed_events));
Won't this assertion result in a failure on bare metal, for CPUs supporting fixed counter 3?
x86/pmu.c int main(int ac, char **av) · line 748
Quoted 11 lines
x86/pmu.c +4
-0+return &fixed_events[idx]; + }return (void *)0 ; }@@ -740,6 +744,8 @@ int main(int ac, char **av)printf ("Fixed counters: %d\n" , pmu.nr_fixed_counters);printf ("Fixed counter width: %d\n" , pmu.fixed_counter_width); +assert (pmu.nr_fixed_counters <=ARRAY_SIZE (fixed_events)); +
And this one as well?
Quoted 6 lines
x86/pmu.c +0
-1apic_write (APIC_LVTPC ,PMI_VECTOR );check_counters (); --2.40.1
Re: [Patch v5 06/18] x86: pmu: Add asserts to warn inconsistent fixed events and counters
Quoted 13 lines · Jim Mattson
Quoted 12 lines · Dapeng Mi
Current PMU code deosn't check whether PMU fixed counter number is larger than pre-defined fixed events. If so, it would cause memory access out of range.
So add assert to warn this invalid case.
Reviewed-by Mingwei Zhang <[email protected]>Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
Quoted 16 lines
Quoted 5 lines
x86/pmu.c +0
-0x86/pmu.c static struct pmu_event* get_counter_event(pmu_counter_t *cnt) · line 119
Quoted 9 lines
x86/pmu.c +4
-2@@ -113,8 +113,12 @@ static struct pmu_event* get_counter_event(pmu_counter_t *cnt)for (i =0 ; i < gp_events_size; i++)if (gp_events[i].unit_sel == (cnt->config &0xffff ))return &gp_events[i]; - }else -return &fixed_events[cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ]; + }else { +unsigned int idx = cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ; + +assert (idx <ARRAY_SIZE (fixed_events));Won't this assertion result in a failure on bare metal, for CPUs supporting fixed counter 3?
Yes, this is intended use. Currently KVM vPMU still doesn't support fixed counter 3. If it's supported in KVM vPMU one day but forget to add corresponding support in this pmu test, this assert would remind this.
Quoted 20 lines
x86/pmu.c int main(int ac, char **av) · line 748
Quoted 12 lines
x86/pmu.c +4
-0+return &fixed_events[idx]; + }return (void *)0 ; }@@ -740,6 +744,8 @@ int main(int ac, char **av)printf ("Fixed counters: %d\n" , pmu.nr_fixed_counters);printf ("Fixed counter width: %d\n" , pmu.fixed_counter_width); +assert (pmu.nr_fixed_counters <=ARRAY_SIZE (fixed_events)); +And this one as well?
Quoted 6 lines
x86/pmu.c +0
-1apic_write (APIC_LVTPC ,PMI_VECTOR );check_counters (); --2.40.1
Re: [Patch v5 06/18] x86: pmu: Add asserts to warn inconsistent fixed events and counters
Quoted 33 lines · Dapeng
Quoted 13 lines · Jim Mattson
Quoted 12 lines · Dapeng Mi
Current PMU code deosn't check whether PMU fixed counter number is larger than pre-defined fixed events. If so, it would cause memory access out of range.
So add assert to warn this invalid case.
Reviewed-by Mingwei Zhang <[email protected]>Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)Quoted 16 lines
Quoted 5 lines
x86/pmu.c +0
-0x86/pmu.c static struct pmu_event* get_counter_event(pmu_counter_t *cnt) · line 119
Quoted 9 lines
x86/pmu.c +4
-2@@ -113,8 +113,12 @@ static struct pmu_event* get_counter_event(pmu_counter_t *cnt)for (i =0 ; i < gp_events_size; i++)if (gp_events[i].unit_sel == (cnt->config &0xffff ))return &gp_events[i]; - }else -return &fixed_events[cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ]; + }else { +unsigned int idx = cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ; + +assert (idx <ARRAY_SIZE (fixed_events));Won't this assertion result in a failure on bare metal, for CPUs supporting fixed counter 3?
Quoted 3 lines
Yes, this is intended use. Currently KVM vPMU still doesn't support fixed counter 3. If it's supported in KVM vPMU one day but forget to add corresponding support in this pmu test, this assert would remind this.
These tests are supposed to run (and pass) on bare metal. Hence, they should not be dependent on a non-architectural quirk of the KVM implementation.
Perhaps a warning would serve as a reminder?
Quoted 21 lines
Quoted 21 lines
x86/pmu.c int main(int ac, char **av) · line 748
Quoted 13 lines
x86/pmu.c +4
-0+return &fixed_events[idx]; + }return (void *)0 ; }@@ -740,6 +744,8 @@ int main(int ac, char **av)printf ("Fixed counters: %d\n" , pmu.nr_fixed_counters);printf ("Fixed counter width: %d\n" , pmu.fixed_counter_width); +assert (pmu.nr_fixed_counters <=ARRAY_SIZE (fixed_events)); +And this one as well?
Quoted 6 lines
x86/pmu.c +0
-1apic_write (APIC_LVTPC ,PMI_VECTOR );check_counters (); --2.40.1
Re: [Patch v5 06/18] x86: pmu: Add asserts to warn inconsistent fixed events and counters
Quoted 32 lines · Jim Mattson
Quoted 31 lines · Dapeng
Quoted 29 lines · Jim Mattson
Quoted 17 lines · Dapeng Mi
Current PMU code deosn't check whether PMU fixed counter number is larger than pre-defined fixed events. If so, it would cause memory access out of range.
So add assert to warn this invalid case.
Reviewed-by Mingwei Zhang <[email protected]>Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)x86/pmu.c +0
-0x86/pmu.c static struct pmu_event* get_counter_event(pmu_counter_t *cnt) · line 119
Quoted 9 lines
x86/pmu.c +4
-2@@ -113,8 +113,12 @@ static struct pmu_event* get_counter_event(pmu_counter_t *cnt)for (i =0 ; i < gp_events_size; i++)if (gp_events[i].unit_sel == (cnt->config &0xffff ))return &gp_events[i]; - }else -return &fixed_events[cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ]; + }else { +unsigned int idx = cnt->ctr -MSR_CORE_PERF_FIXED_CTR0 ; + +assert (idx <ARRAY_SIZE (fixed_events));Won't this assertion result in a failure on bare metal, for CPUs supporting fixed counter 3?
Quoted 8 lines
Quoted 3 lines
Yes, this is intended use. Currently KVM vPMU still doesn't support fixed counter 3. If it's supported in KVM vPMU one day but forget to add corresponding support in this pmu test, this assert would remind this.
These tests are supposed to run (and pass) on bare metal. Hence, they should not be dependent on a non-architectural quirk of the KVM implementation.
Perhaps a warning would serve as a reminder?
Sounds reasonable. Would change to a warning. Thanks.
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
x86/pmu.c int main(int ac, char **av) · line 748
Quoted 12 lines
x86/pmu.c +4
-0+return &fixed_events[idx]; + }return (void *)0 ; }@@ -740,6 +744,8 @@ int main(int ac, char **av)printf ("Fixed counters: %d\n" , pmu.nr_fixed_counters);printf ("Fixed counter width: %d\n" , pmu.fixed_counter_width); +assert (pmu.nr_fixed_counters <=ARRAY_SIZE (fixed_events)); +And this one as well?
Quoted 6 lines
x86/pmu.c +0
-1apic_write (APIC_LVTPC ,PMI_VECTOR );check_counters (); --2.40.1
PATCH v5 7/18 x86: pmu: Fix cycles event validation failure
In reply to Dapeng MiWhen running pmu test on SPR, sometimes the following failure is reported.
PMU version: 2 GP counters: 8 GP counter width: 48 Mask length: 8 Fixed counters: 3 Fixed counter width: 48
1000000 <= 55109398 <= 50000000
FAIL: Intel: core cycles-0
1000000 <= 18279571 <= 50000000
PASS: Intel: core cycles-1
1000000 <= 12238092 <= 50000000
PASS: Intel: core cycles-2
1000000 <= 7981727 <= 50000000
PASS: Intel: core cycles-3
1000000 <= 6984711 <= 50000000
PASS: Intel: core cycles-4
1000000 <= 6773673 <= 50000000
PASS: Intel: core cycles-5
1000000 <= 6697842 <= 50000000
PASS: Intel: core cycles-6
1000000 <= 6747947 <= 50000000
PASS: Intel: core cycles-7
The count of the "core cycles" on first counter would exceed the upper boundary and leads to a failure, and then the "core cycles" count would drop gradually and reach a stable state.
That looks reasonable. The "core cycles" event is defined as the 1st event in xxx_gp_events[] array and it is always verified at first. when the program loop() is executed at the first time it needs to warm up the pipeline and cache, such as it has to wait for cache is filled. All these warm-up work leads to a quite large core cycles count which may exceeds the verification range.
To avoid the false positive of cycles event caused by warm-up, explicitly introduce a warm-up state before really starting verification.
--- x86/pmu.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
x86/pmu.c +16 -0
@@ -593,11 +593,27 @@ static void check_tsx_cycles(void)report_prefix_pop (); } +static void warm_up (void ) +{ +int i =8 ; + +/* +* Since cycles event is always run as the first event, there would be +* a warm-up state to warm up the cache, it leads to the measured cycles +* value may exceed the pre-defined cycles upper boundary and cause +* false positive. To avoid this, introduce an warm-up state before +* the real verification. +*/ +while (i--) +loop (); +} +static void check_counters (void ) {if (is_fep_available ())check_emulated_instr (); +warm_up ();check_gp_counters ();check_fixed_counters ();check_rdpmc ();
Signature · 1 line
2.40.1
PATCH v5 8/18 x86: pmu: Use macro to replace hard-coded branches event index
In reply to Dapeng MiCurrently the branches event index is a hard-coded number. User could add new events and cause the branches event index changes in the future, but don't notice the hard-coded event index and forget to update the event index synchronously, then the issue comes.
Thus, replace the hard-coded index to a macro.
--- x86/pmu.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-)
x86/pmu.c +18 -1
@@ -50,6 +50,22 @@ struct pmu_event { {"fixed 2", MSR_CORE_PERF_FIXED_CTR0 + 2, 0.1*N, 30*N} }; +/* +* Events index in intel_gp_events[], ensure consistent with +* intel_gp_events[]. +*/ +enum { +INTEL_BRANCHES_IDX =5 , +}; + +/* +* Events index in amd_gp_events[], ensure consistent with +* amd_gp_events[]. +*/ +enum { +AMD_BRANCHES_IDX =2 , +}; +char *buf;static struct pmu_event *gp_events;
@@ -483,7 +499,8 @@ static void check_emulated_instr(void) {uint64_t status, instr_start, brnch_start;uint64_t gp_counter_width = (1ull << pmu.gp_counter_width) -1 ; -unsigned int branch_idx = pmu.is_intel ?5 :2 ; +unsigned int branch_idx = pmu.is_intel ? +INTEL_BRANCHES_IDX :AMD_BRANCHES_IDX ;pmu_counter_t brnch_cnt = { .ctr =MSR_GP_COUNTERx (0 ),/* branch instructions */
Signature · 1 line
2.40.1
PATCH v5 9/18 x86: pmu: Use macro to replace hard-coded ref-cycles event index
In reply to Dapeng MiReplace hard-coded ref-cycles event index with macro to avoid possible mismatch issue if new event is added in the future and cause ref-cycles event index changed, but forget to update the hard-coded ref-cycles event index.
--- x86/pmu.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
x86/pmu.c +7 -3
@@ -55,6 +55,7 @@ struct pmu_event {* intel_gp_events[]. */ enum { +INTEL_REF_CYCLES_IDX =2 ,INTEL_BRANCHES_IDX =5 , };
@@ -699,7 +700,8 @@ static void set_ref_cycle_expectations(void) {pmu_counter_t cnt = { .ctr =MSR_IA32_PERFCTR0 , - .config =EVNTSEL_OS |EVNTSEL_USR | intel_gp_events[2 ].unit_sel, + .config =EVNTSEL_OS |EVNTSEL_USR | + intel_gp_events[INTEL_REF_CYCLES_IDX ].unit_sel, };uint64_t tsc_delta;uint64_t t0, t1, t2, t3;
@@ -735,8 +737,10 @@ static void set_ref_cycle_expectations(void)if (!tsc_delta)return ; - intel_gp_events[2 ].min = (intel_gp_events[2 ].min * cnt.count) / tsc_delta; - intel_gp_events[2 ].max = (intel_gp_events[2 ].max * cnt.count) / tsc_delta; + intel_gp_events[INTEL_REF_CYCLES_IDX ].min = + (intel_gp_events[INTEL_REF_CYCLES_IDX ].min * cnt.count) / tsc_delta; + intel_gp_events[INTEL_REF_CYCLES_IDX ].max = + (intel_gp_events[INTEL_REF_CYCLES_IDX ].max * cnt.count) / tsc_delta; }static void check_invalid_rdpmc_gp (void )
Signature · 1 line
2.40.1
PATCH v5 10/18 x86: pmu: Use macro to replace hard-coded instructions event index
In reply to Dapeng MiReplace hard-coded instruction event index with macro to avoid possible mismatch issue if new event is added in the future and cause instructions event index changed, but forget to update the hard-coded event index.
--- x86/pmu.c | 34 +++++++++++++++++++++++++++------- 1 file changed, 27 insertions(+), 7 deletions(-)
x86/pmu.c +27 -7
@@ -55,6 +55,7 @@ struct pmu_event {* intel_gp_events[]. */ enum { +INTEL_INSTRUCTIONS_IDX =1 ,INTEL_REF_CYCLES_IDX =2 ,INTEL_BRANCHES_IDX =5 , };
@@ -64,6 +65,7 @@ enum {* amd_gp_events[]. */ enum { +AMD_INSTRUCTIONS_IDX =1 ,AMD_BRANCHES_IDX =2 , };
@@ -319,11 +321,16 @@ static uint64_t measure_for_overflow(pmu_counter_t *cnt)static void check_counter_overflow (void ) { -uint64_t overflow_preset;int i; +uint64_t overflow_preset; +int instruction_idx = pmu.is_intel ? +INTEL_INSTRUCTIONS_IDX : +AMD_INSTRUCTIONS_IDX ; +pmu_counter_t cnt = { .ctr =MSR_GP_COUNTERx (0 ), - .config =EVNTSEL_OS |EVNTSEL_USR | gp_events[1 ].unit_sel/* instructions */ , + .config =EVNTSEL_OS |EVNTSEL_USR | + gp_events[instruction_idx].unit_sel/* instructions */ , }; overflow_preset =measure_for_overflow (&cnt);
@@ -379,13 +386,18 @@ static void check_counter_overflow(void)static void check_gp_counter_cmask (void ) { +int instruction_idx = pmu.is_intel ? +INTEL_INSTRUCTIONS_IDX : +AMD_INSTRUCTIONS_IDX ; +pmu_counter_t cnt = { .ctr =MSR_GP_COUNTERx (0 ), - .config =EVNTSEL_OS |EVNTSEL_USR | gp_events[1 ].unit_sel/* instructions */ , + .config =EVNTSEL_OS |EVNTSEL_USR | + gp_events[instruction_idx].unit_sel/* instructions */ , }; cnt.config |= (0x2 <<EVNTSEL_CMASK_SHIFT );measure_one (&cnt); -report (cnt.count < gp_events[1 ].min,"cmask" ); +report (cnt.count < gp_events[instruction_idx].min,"cmask" ); }static void do_rdpmc_fast (void *ptr)
@@ -460,9 +472,14 @@ static void check_running_counter_wrmsr(void) {uint64_t status;uint64_t count; +unsigned int instruction_idx = pmu.is_intel ? +INTEL_INSTRUCTIONS_IDX : +AMD_INSTRUCTIONS_IDX ; +pmu_counter_t evt = { .ctr =MSR_GP_COUNTERx (0 ), - .config =EVNTSEL_OS |EVNTSEL_USR | gp_events[1 ].unit_sel, + .config =EVNTSEL_OS |EVNTSEL_USR | + gp_events[instruction_idx].unit_sel, };report_prefix_push ("running counter wrmsr" );
@@ -471,7 +488,7 @@ static void check_running_counter_wrmsr(void)loop ();wrmsr (MSR_GP_COUNTERx (0 ),0 );stop_event (&evt); -report (evt.count < gp_events[1 ].min,"cntr" ); +report (evt.count < gp_events[instruction_idx].min,"cntr" );/* clear status before overflow test */ if (this_cpu_has_perf_global_status ())
@@ -502,6 +519,9 @@ static void check_emulated_instr(void)uint64_t gp_counter_width = (1ull << pmu.gp_counter_width) -1 ;unsigned int branch_idx = pmu.is_intel ?INTEL_BRANCHES_IDX :AMD_BRANCHES_IDX ; +unsigned int instruction_idx = pmu.is_intel ? +INTEL_INSTRUCTIONS_IDX : +AMD_INSTRUCTIONS_IDX ;pmu_counter_t brnch_cnt = { .ctr =MSR_GP_COUNTERx (0 ),/* branch instructions */
@@ -510,7 +530,7 @@ static void check_emulated_instr(void)pmu_counter_t instr_cnt = { .ctr =MSR_GP_COUNTERx (1 ),/* instructions */ - .config =EVNTSEL_OS |EVNTSEL_USR | gp_events[1 ].unit_sel, + .config =EVNTSEL_OS |EVNTSEL_USR | gp_events[instruction_idx].unit_sel, };report_prefix_push ("emulated instruction" );
Signature · 1 line
2.40.1
PATCH v5 11/18 x86: pmu: Enable and disable PMCs in loop() asm blob
In reply to Dapeng MiCurrently enabling PMCs, executing loop() and disabling PMCs are divided 3 separated functions. So there could be other instructions executed between enabling PMCS and running loop() or running loop() and disabling PMCs, e.g. if there are multiple counters enabled in measure_many() function, the instructions which enabling the 2nd and more counters would be counted in by the 1st counter.
So current implementation can only verify the correctness of count by an rough range rather than a precise count even for instructions and branches events. Strictly speaking, this verification is meaningless as the test could still pass even though KVM vPMU has something wrong and reports an incorrect instructions or branches count which is in the rough range.
Thus, move the PMCs enabling and disabling into the loop() asm blob and ensure only the loop asm instructions would be counted, then the instructions or branches events can be verified with an precise count instead of an rough range.
--- x86/pmu.c | 80 ++++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 65 insertions(+), 15 deletions(-)
x86/pmu.c +65 -15
@@ -19,6 +19,15 @@#define EXPECTED_INSTR 17#define EXPECTED_BRNCH 5 +#define LOOP_ASM (_wrmsr) \ + _wrmsr "\n\t" \ + "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \ + "1: mov (%1), %2; add $64, %1;\n\t" \ + "nop; nop; nop; nop; nop; nop; nop;\n\t" \ + "loop 1b;\n\t" \ + "mov %%edi, %%ecx; xor %%eax, %%eax; xor %%edx, %%edx;\n\t" \ + _wrmsr "\n\t" + typedef struct { uint32_t ctr; uint32_t idx;
@@ -74,13 +83,43 @@ char *buf;static struct pmu_event *gp_events;static unsigned int gp_events_size; -static inline void loop (void ) + +static inline void __loop (void ) +{ +unsigned long tmp, tmp2, tmp3; + + asm volatile(LOOP_ASM("nop") + : "=c"(tmp), "=r"(tmp2), "=r"(tmp3) + : "0"(N), "1"(buf)); +} + +/* +* Enable and disable counters in a whole asm blob to ensure +* no other instructions are counted in the window between +* counters enabling and really LOOP_ASM code executing. +* Thus counters can verify instructions and branches events +* against precise counts instead of a rough valid count range. +*/ +static inline void __precise_loop (u64 cntrs) {unsigned long tmp, tmp2, tmp3; +unsigned int global_ctl = pmu.msr_global_ctl; +u32 eax = cntrs & (BIT_ULL (32 ) -1 ); +u32 edx = cntrs >>32 ; - asmvolatile ("1: mov (%1), %2; add $64, %1; nop; nop; nop; nop; nop; nop; nop; loop 1b" - :"=c" (tmp),"=r" (tmp2),"=r" (tmp3):"0" (N ),"1" (buf)); + asm volatile(LOOP_ASM("wrmsr") + : "=b"(tmp), "=r"(tmp2), "=r"(tmp3) + : "a"(eax), "d"(edx), "c"(global_ctl), + "0"(N), "1"(buf) + : "edi"); +} +static inline void loop (u64 cntrs) +{ +if (!this_cpu_has_perf_global_ctrl ()) +__loop (); +else +__precise_loop (cntrs); }volatile uint64_t irq_received;
@@ -180,18 +219,17 @@ static void __start_event(pmu_counter_t *evt, uint64_t count) ctrl = (ctrl & ~(0xf << shift)) | (usrospmi << shift);wrmsr (MSR_CORE_PERF_FIXED_CTR_CTRL , ctrl); } -global_enable (evt);apic_write (APIC_LVTPC ,PMI_VECTOR ); }static void start_event (pmu_counter_t *evt) {__start_event (evt,0 ); +global_enable (evt); } -static void stop_event (pmu_counter_t *evt) +static void __stop_event (pmu_counter_t *evt) { -global_disable (evt);if (is_gp (evt)) {wrmsr (MSR_GP_EVENT_SELECTx (event_to_global_idx (evt)), evt->config & ~EVNTSEL_EN );
@@ -203,14 +241,24 @@ static void stop_event(pmu_counter_t *evt) evt->count =rdmsr (evt->ctr); } +static void stop_event (pmu_counter_t *evt) +{ +global_disable (evt); +__stop_event (evt); +} + static noinline void measure_many(pmu_counter_t *evt, int count) {int i; +u64 cntrs =0 ; + +for (i =0 ; i < count; i++) { +__start_event (&evt[i],0 ); + cntrs |=BIT_ULL (event_to_global_idx (&evt[i])); + } +loop (cntrs);for (i =0 ; i < count; i++) -start_event (&evt[i]); -loop (); -for (i =0 ; i < count; i++) -stop_event (&evt[i]); +__stop_event (&evt[i]); }static void measure_one (pmu_counter_t *evt)
@@ -220,9 +268,11 @@ static void measure_one(pmu_counter_t *evt) static noinline void __measure(pmu_counter_t *evt, uint64_t count) { +u64 cntrs =BIT_ULL (event_to_global_idx (evt)); +__start_event (evt, count); -loop (); -stop_event (evt); +loop (cntrs); +__stop_event (evt); }static bool verify_event (uint64_t count,struct pmu_event *e)
@@ -485,7 +535,7 @@ static void check_running_counter_wrmsr(void)report_prefix_push ("running counter wrmsr" );start_event (&evt); -loop (); +__loop ();wrmsr (MSR_GP_COUNTERx (0 ),0 );stop_event (&evt);report (evt.count < gp_events[instruction_idx].min,"cntr" );
@@ -502,7 +552,7 @@ static void check_running_counter_wrmsr(void)wrmsr (MSR_GP_COUNTERx (0 ), count); -loop (); +__loop ();stop_event (&evt);if (this_cpu_has_perf_global_status ()) {
@@ -643,7 +693,7 @@ static void warm_up(void)* the real verification. */ while (i--) -loop (); +loop (0 ); }static void check_counters (void )
Signature · 1 line
2.40.1
PATCH v5 12/18 x86: pmu: Improve instruction and branches events verification
In reply to Dapeng MiIf HW supports GLOBAL_CTRL MSR, enabling and disabling PMCs are moved in __precise_count_loop(). Thus, instructions and branches events can be verified against a precise count instead of a rough range.
--- x86/pmu.c | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)
x86/pmu.c +31 -0
@@ -19,6 +19,11 @@#define EXPECTED_INSTR 17#define EXPECTED_BRNCH 5 + +/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL instructions */ +#define EXTRA_INSTRNS (3 + 3) +#define LOOP_INSTRNS (N * 10 + EXTRA_INSTRNS) +#define LOOP_BRANCHES (N)#define LOOP_ASM (_wrmsr) \ _wrmsr "\n\t" \ "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \
@@ -122,6 +127,24 @@ static inline void loop(u64 cntrs)__precise_loop (cntrs); } +static void adjust_events_range (struct pmu_event *gp_events, +int instruction_idx,int branch_idx) +{ +/* +* If HW supports GLOBAL_CTRL MSR, enabling and disabling PMCs are +* moved in __precise_loop(). Thus, instructions and branches events +* can be verified against a precise count instead of a rough range. +*/ +if (this_cpu_has_perf_global_ctrl ()) { +/* instructions event */ + gp_events[instruction_idx].min =LOOP_INSTRNS ; + gp_events[instruction_idx].max =LOOP_INSTRNS ; +/* branches event */ + gp_events[branch_idx].min =LOOP_BRANCHES ; + gp_events[branch_idx].max =LOOP_BRANCHES ; + } +} +volatile uint64_t irq_received;static void cnt_overflow (isr_regs_t *regs)
@@ -823,6 +846,9 @@ static void check_invalid_rdpmc_gp(void)int main (int ac,char **av) { +int instruction_idx; +int branch_idx; +setup_vm ();handle_irq (PMI_VECTOR , cnt_overflow); buf =malloc (N *64 );
@@ -836,13 +862,18 @@ int main(int ac, char **av) } gp_events = (struct pmu_event *)intel_gp_events; gp_events_size =sizeof (intel_gp_events)/sizeof (intel_gp_events[0 ]); + instruction_idx =INTEL_INSTRUCTIONS_IDX ; + branch_idx =INTEL_BRANCHES_IDX ;report_prefix_push ("Intel" );set_ref_cycle_expectations (); }else { gp_events_size =sizeof (amd_gp_events)/sizeof (amd_gp_events[0 ]); gp_events = (struct pmu_event *)amd_gp_events; + instruction_idx =AMD_INSTRUCTIONS_IDX ; + branch_idx =AMD_BRANCHES_IDX ;report_prefix_push ("AMD" ); } +adjust_events_range (gp_events, instruction_idx, branch_idx);printf ("PMU version: %d\n" , pmu.version);printf ("GP counters: %d\n" , pmu.nr_gp_counters);
Signature · 1 line
2.40.1
Re: [Patch v5 12/18] x86: pmu: Improve instruction and branches events verification
Quoted 28 lines · Dapeng Mi
If HW supports
GLOBAL_CTRLMSR, enabling and disabling PMCs are moved in__precise_count_loop(). Thus, instructions and branches events can be verified against a precise count instead of a rough range.Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)x86/pmu.c +5
-0@@ -19,6 +19,11 @@#define EXPECTED_INSTR 17#define EXPECTED_BRNCH 5 + +/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL instructions */ +#define EXTRA_INSTRNS (3 + 3) +#define LOOP_INSTRNS (N * 10 + EXTRA_INSTRNS) +#define LOOP_BRANCHES (N)#define LOOP_ASM (_wrmsr) \ _wrmsr "\n\t" \ "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \@@ -122,6 +127,24 @@ static inline void loop(u64 cntrs)__precise_loop (cntrs); }
x86/pmu.c static inline void loop(u64 cntrs) · line 138
Quoted 10 lines
x86/pmu.c +9
-0+static void adjust_events_range (struct pmu_event *gp_events, +int instruction_idx,int branch_idx) +{ +/* +* If HW supports GLOBAL_CTRL MSR, enabling and disabling PMCs are +* moved in __precise_loop(). Thus, instructions and branches events +* can be verified against a precise count instead of a rough range. +*/ +if (this_cpu_has_perf_global_ctrl ()) {
This causes some intermittent failures on AMD processors using PerfMonV2 due to variance in counts. This probably has to do with the way instructions leading to a VM-Entry or VM-Exit are accounted when counting retired instructions and branches. Adding the following change makes all the tests pass again.
x86/pmu.c +1 -1
@@ -222,7 +222,7 @@ static void adjust_events_range(struct pmu_event *gp_events,* moved in __precise_loop(). Thus, instructions and branches events * can be verified against a precise count instead of a rough range. */ -if (this_cpu_has_perf_global_ctrl ()) { +if (pmu.is_intel &&this_cpu_has_perf_global_ctrl ()) {/* instructions event */ gp_events[instruction_idx].min =LOOP_INSTRNS ; gp_events[instruction_idx].max =LOOP_INSTRNS ;
Quoted 41 lines
x86/pmu.c +17
-0+/* instructions event */ + gp_events[instruction_idx].min =LOOP_INSTRNS ; + gp_events[instruction_idx].max =LOOP_INSTRNS ; +/* branches event */ + gp_events[branch_idx].min =LOOP_BRANCHES ; + gp_events[branch_idx].max =LOOP_BRANCHES ; + } +} +volatile uint64_t irq_received;static void cnt_overflow (isr_regs_t *regs)@@ -823,6 +846,9 @@ static void check_invalid_rdpmc_gp(void)int main (int ac,char **av) { +int instruction_idx; +int branch_idx; +setup_vm ();handle_irq (PMI_VECTOR , cnt_overflow); buf =malloc (N *64 );@@ -836,13 +862,18 @@ int main(int ac, char **av) } gp_events = (struct pmu_event *)intel_gp_events; gp_events_size =sizeof (intel_gp_events)/sizeof (intel_gp_events[0 ]); + instruction_idx =INTEL_INSTRUCTIONS_IDX ; + branch_idx =INTEL_BRANCHES_IDX ;report_prefix_push ("Intel" );set_ref_cycle_expectations (); }else { gp_events_size =sizeof (amd_gp_events)/sizeof (amd_gp_events[0 ]); gp_events = (struct pmu_event *)amd_gp_events; + instruction_idx =AMD_INSTRUCTIONS_IDX ; + branch_idx =AMD_BRANCHES_IDX ;report_prefix_push ("AMD" ); } +adjust_events_range (gp_events, instruction_idx, branch_idx);printf ("PMU version: %d\n" , pmu.version);printf ("GP counters: %d\n" , pmu.nr_gp_counters);
Re: [Patch v5 12/18] x86: pmu: Improve instruction and branches events verification
Quoted 30 lines · Sandipan Das
Quoted 29 lines · Dapeng Mi
If HW supports
GLOBAL_CTRLMSR, enabling and disabling PMCs are moved in__precise_count_loop(). Thus, instructions and branches events can be verified against a precise count instead of a rough range.Signed-off-by Dapeng Mi <[email protected]>--- x86/pmu.c | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)x86/pmu.c +5
-0@@ -19,6 +19,11 @@#define EXPECTED_INSTR 17#define EXPECTED_BRNCH 5 + +/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL instructions */ +#define EXTRA_INSTRNS (3 + 3) +#define LOOP_INSTRNS (N * 10 + EXTRA_INSTRNS) +#define LOOP_BRANCHES (N)#define LOOP_ASM (_wrmsr) \ _wrmsr "\n\t" \ "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \@@ -122,6 +127,24 @@ static inline void loop(u64 cntrs)__precise_loop (cntrs); }
Quoted 14 lines
x86/pmu.c static inline void loop(u64 cntrs) · line 138
Quoted 9 lines
x86/pmu.c +9
-0+static void adjust_events_range (struct pmu_event *gp_events, +int instruction_idx,int branch_idx) +{ +/* +* If HW supports GLOBAL_CTRL MSR, enabling and disabling PMCs are +* moved in __precise_loop(). Thus, instructions and branches events +* can be verified against a precise count instead of a rough range. +*/ +if (this_cpu_has_perf_global_ctrl ()) {This causes some intermittent failures on AMD processors using PerfMonV2 due to variance in counts. This probably has to do with the way instructions leading to a VM-Entry or VM-Exit are accounted when counting retired instructions and branches. Adding the following change makes all the tests pass again.
Thanks to verify on AMD platforms. Would add it in next version.
Quoted 58 lines
x86/pmu.c +1
-1@@ -222,7 +222,7 @@ static void adjust_events_range(struct pmu_event *gp_events,* moved in __precise_loop(). Thus, instructions and branches events * can be verified against a precise count instead of a rough range. */ -if (this_cpu_has_perf_global_ctrl ()) { +if (pmu.is_intel &&this_cpu_has_perf_global_ctrl ()) {/* instructions event */ gp_events[instruction_idx].min =LOOP_INSTRNS ; gp_events[instruction_idx].max =LOOP_INSTRNS ;Quoted 43 lines
x86/pmu.c +17
-0+/* instructions event */ + gp_events[instruction_idx].min =LOOP_INSTRNS ; + gp_events[instruction_idx].max =LOOP_INSTRNS ; +/* branches event */ + gp_events[branch_idx].min =LOOP_BRANCHES ; + gp_events[branch_idx].max =LOOP_BRANCHES ; + } +} +volatile uint64_t irq_received;static void cnt_overflow (isr_regs_t *regs)@@ -823,6 +846,9 @@ static void check_invalid_rdpmc_gp(void)int main (int ac,char **av) { +int instruction_idx; +int branch_idx; +setup_vm ();handle_irq (PMI_VECTOR , cnt_overflow); buf =malloc (N *64 );@@ -836,13 +862,18 @@ int main(int ac, char **av) } gp_events = (struct pmu_event *)intel_gp_events; gp_events_size =sizeof (intel_gp_events)/sizeof (intel_gp_events[0 ]); + instruction_idx =INTEL_INSTRUCTIONS_IDX ; + branch_idx =INTEL_BRANCHES_IDX ;report_prefix_push ("Intel" );set_ref_cycle_expectations (); }else { gp_events_size =sizeof (amd_gp_events)/sizeof (amd_gp_events[0 ]); gp_events = (struct pmu_event *)amd_gp_events; + instruction_idx =AMD_INSTRUCTIONS_IDX ; + branch_idx =AMD_BRANCHES_IDX ;report_prefix_push ("AMD" ); } +adjust_events_range (gp_events, instruction_idx, branch_idx);printf ("PMU version: %d\n" , pmu.version);printf ("GP counters: %d\n" , pmu.nr_gp_counters);
PATCH v5 13/18 x86: pmu: Improve LLC misses event verification
In reply to Dapeng MiWhen running pmu test on SPR, sometimes the following failure is reported.
1 <= 0 <= 1000000
FAIL: Intel: llc misses-4
Currently The LLC misses occurring only depends on probability. It's possible that there is no LLC misses happened in the whole loop(), especially along with processors have larger and larger cache size just like what we observed on SPR.
Thus, add clflush instruction into the loop() asm blob and ensure once LLC miss is triggered at least.
--- x86/pmu.c | 39 ++++++++++++++++++++++++++------------- 1 file changed, 26 insertions(+), 13 deletions(-)
x86/pmu.c +26 -13
@@ -20,19 +20,30 @@#define EXPECTED_BRNCH 5 -/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL instructions */ -#define EXTRA_INSTRNS (3 + 3) +/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL + clflush/mfence instructions */ +#define EXTRA_INSTRNS (3 + 3 + 2)#define LOOP_INSTRNS (N * 10 + EXTRA_INSTRNS)#define LOOP_BRANCHES (N) -#define LOOP_ASM (_wrmsr) \ +#define LOOP_ASM (_wrmsr, _clflush) \ _wrmsr "\n\t" \ "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \ + _clflush "\n\t" \ + "mfence;\n\t" \ "1: mov (%1), %2; add $64, %1;\n\t" \ "nop; nop; nop; nop; nop; nop; nop;\n\t" \ "loop 1b;\n\t" \ "mov %%edi, %%ecx; xor %%eax, %%eax; xor %%edx, %%edx;\n\t" \ _wrmsr "\n\t" +#define _loop_asm (_wrmsr, _clflush) \ +do { \ + asm volatile(LOOP_ASM(_wrmsr, _clflush) \ + : "=b"(tmp), "=r"(tmp2), "=r"(tmp3) \ + : "a"(eax), "d"(edx), "c"(global_ctl), \ + "0"(N), "1"(buf) \ + : "edi"); \ +} while (0) + typedef struct { uint32_t ctr; uint32_t idx;
@@ -88,14 +99,17 @@ char *buf;static struct pmu_event *gp_events;static unsigned int gp_events_size; -static inline void __loop (void ) {unsigned long tmp, tmp2, tmp3; +u32 global_ctl =0 ; +u32 eax =0 ; +u32 edx =0 ; - asm volatile(LOOP_ASM("nop") - : "=c"(tmp), "=r"(tmp2), "=r"(tmp3) - : "0"(N), "1"(buf)); +if (this_cpu_has (X86_FEATURE_CLFLUSH )) +_loop_asm ("nop" ,"clflush (%1)" ); +else +_loop_asm ("nop" ,"nop" ); }/*
@@ -108,15 +122,14 @@ static inline void __loop(void)static inline void __precise_loop (u64 cntrs) {unsigned long tmp, tmp2, tmp3; -unsigned int global_ctl = pmu.msr_global_ctl; +u32 global_ctl = pmu.msr_global_ctl;u32 eax = cntrs & (BIT_ULL (32 ) -1 );u32 edx = cntrs >>32 ; - asm volatile(LOOP_ASM("wrmsr") - : "=b"(tmp), "=r"(tmp2), "=r"(tmp3) - : "a"(eax), "d"(edx), "c"(global_ctl), - "0"(N), "1"(buf) - : "edi"); +if (this_cpu_has (X86_FEATURE_CLFLUSH )) +_loop_asm ("wrmsr" ,"clflush (%1)" ); +else +_loop_asm ("wrmsr" ,"nop" ); }static inline void loop (u64 cntrs)
Signature · 1 line
2.40.1
PATCH v5 14/18 x86: pmu: Adjust lower boundary of llc-misses event to 0 for legacy CPUs
In reply to Dapeng MiFor these legacy Intel CPUs without clflush/clflushopt support, there is on way to force to trigger a LLC miss and the measured llc misses is possible to be 0. Thus adjust the lower boundary of llc-misses event to 0 to avoid possible false positive.
--- x86/pmu.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
x86/pmu.c +10 -0
@@ -82,6 +82,7 @@ struct pmu_event {enum {INTEL_INSTRUCTIONS_IDX =1 ,INTEL_REF_CYCLES_IDX =2 , +INTEL_LLC_MISSES_IDX =4 ,INTEL_BRANCHES_IDX =5 , };
@@ -877,6 +878,15 @@ int main(int ac, char **av) gp_events_size =sizeof (intel_gp_events)/sizeof (intel_gp_events[0 ]); instruction_idx =INTEL_INSTRUCTIONS_IDX ; branch_idx =INTEL_BRANCHES_IDX ; + +/* +* For legacy Intel CPUS without clflush/clflushopt support, +* there is no way to force to trigger a LLC miss, thus set +* the minimum value to 0 to avoid false positives. +*/ +if (!this_cpu_has (X86_FEATURE_CLFLUSH )) + gp_events[INTEL_LLC_MISSES_IDX ].min =0 ; +report_prefix_push ("Intel" );set_ref_cycle_expectations (); }else {
Signature · 1 line
2.40.1
PATCH v5 15/18 x86: pmu: Add IBPB indirect jump asm blob
In reply to Dapeng MiCurrently the lower boundary of branch misses event is set to 0. Strictly speaking 0 shouldn't be a valid count since it can't tell us if branch misses event counter works correctly or even disabled. Whereas it's also possible and reasonable that branch misses event count is 0 especailly for such simple loop() program with advanced branch predictor.
To eliminate such ambiguity and make branch misses event verification more acccurately, an extra IBPB indirect jump asm blob is appended and IBPB command is leveraged to clear the branch target buffer and force to cause a branch miss for the indirect jump.
--- x86/pmu.c | 71 +++++++++++++++++++++++++++++++++++++++++++------------ 1 file changed, 56 insertions(+), 15 deletions(-)
x86/pmu.c +56 -15
@@ -19,25 +19,52 @@#define EXPECTED_INSTR 17#define EXPECTED_BRNCH 5 - -/* Enable GLOBAL_CTRL + disable GLOBAL_CTRL + clflush/mfence instructions */ -#define EXTRA_INSTRNS (3 + 3 + 2) +#define IBPB_JMP_INSTRNS 9 +#define IBPB_JMP_BRANCHES 2 + +#if defined(__i386__) || defined(_M_IX86)/* i386 */ +#define IBPB_JMP_ASM (_wrmsr) \ + "mov $1, %%eax; xor %%edx, %%edx;\n\t" \ + "mov $73, %%ecx;\n\t" \ + _wrmsr "\n\t" \ + "call 1f\n\t" \ + "1: pop %%eax\n\t" \ + "add $(2f-1b), %%eax\n\t" \ + "jmp *%%eax;\n\t" \ + "nop;\n\t" \ + "2: nop;\n\t" +#else /* x86_64 */ +#define IBPB_JMP_ASM (_wrmsr) \ + "mov $1, %%eax; xor %%edx, %%edx;\n\t" \ + "mov $73, %%ecx;\n\t" \ + _wrmsr "\n\t" \ + "call 1f\n\t" \ + "1: pop %%rax\n\t" \ + "add $(2f-1b), %%rax\n\t" \ + "jmp *%%rax;\n\t" \ + "nop;\n\t" \ + "2: nop;\n\t" +#endif + +/* GLOBAL_CTRL enable + disable + clflush/mfence + IBPB_JMP */ +#define EXTRA_INSTRNS (3 + 3 + 2 + IBPB_JMP_INSTRNS)#define LOOP_INSTRNS (N * 10 + EXTRA_INSTRNS) -#define LOOP_BRANCHES (N) -#define LOOP_ASM (_wrmsr, _clflush) \ - _wrmsr "\n\t" \ +#define LOOP_BRANCHES (N + IBPB_JMP_BRANCHES) +#define LOOP_ASM (_wrmsr1, _clflush, _wrmsr2) \ + _wrmsr1 "\n\t" \ "mov %%ecx, %%edi; mov %%ebx, %%ecx;\n\t" \ _clflush "\n\t" \ "mfence;\n\t" \ "1: mov (%1), %2; add $64, %1;\n\t" \ "nop; nop; nop; nop; nop; nop; nop;\n\t" \ "loop 1b;\n\t" \ + IBPB_JMP_ASM(_wrmsr2) \ "mov %%edi, %%ecx; xor %%eax, %%eax; xor %%edx, %%edx;\n\t" \ - _wrmsr "\n\t" + _wrmsr1 "\n\t" -#define _loop_asm (_wrmsr, _clflush) \ +#define _loop_asm (_wrmsr1, _clflush, _wrmsr2) \ do { \ - asm volatile(LOOP_ASM(_wrmsr, _clflush) \ + asm volatile(LOOP_ASM(_wrmsr1, _clflush, _wrmsr2) \ : "=b"(tmp), "=r"(tmp2), "=r"(tmp3) \ : "a"(eax), "d"(edx), "c"(global_ctl), \ "0"(N), "1"(buf) \
@@ -100,6 +127,12 @@ char *buf;static struct pmu_event *gp_events;static unsigned int gp_events_size; +static int has_ibpb (void ) +{ +return this_cpu_has (X86_FEATURE_SPEC_CTRL ) || +this_cpu_has (X86_FEATURE_AMD_IBPB ); +} +static inline void __loop (void ) {unsigned long tmp, tmp2, tmp3;
@@ -107,10 +140,14 @@ static inline void __loop(void)u32 eax =0 ;u32 edx =0 ; -if (this_cpu_has (X86_FEATURE_CLFLUSH )) -_loop_asm ("nop" ,"clflush (%1)" ); +if (this_cpu_has (X86_FEATURE_CLFLUSH ) &&has_ibpb ()) +_loop_asm ("nop" ,"clflush (%1)" ,"wrmsr" ); +else if (this_cpu_has (X86_FEATURE_CLFLUSH )) +_loop_asm ("nop" ,"clflush (%1)" ,"nop" ); +else if (has_ibpb ()) +_loop_asm ("nop" ,"nop" ,"wrmsr" );else -_loop_asm ("nop" ,"nop" ); +_loop_asm ("nop" ,"nop" ,"nop" ); }/*
@@ -127,10 +164,14 @@ static inline void __precise_loop(u64 cntrs)u32 eax = cntrs & (BIT_ULL (32 ) -1 );u32 edx = cntrs >>32 ; -if (this_cpu_has (X86_FEATURE_CLFLUSH )) -_loop_asm ("wrmsr" ,"clflush (%1)" ); +if (this_cpu_has (X86_FEATURE_CLFLUSH ) &&has_ibpb ()) +_loop_asm ("wrmsr" ,"clflush (%1)" ,"wrmsr" ); +else if (this_cpu_has (X86_FEATURE_CLFLUSH )) +_loop_asm ("wrmsr" ,"clflush (%1)" ,"nop" ); +else if (has_ibpb ()) +_loop_asm ("wrmsr" ,"nop" ,"wrmsr" );else -_loop_asm ("wrmsr" ,"nop" ); +_loop_asm ("wrmsr" ,"nop" ,"nop" ); }static inline void loop (u64 cntrs)
Signature · 1 line
2.40.1
PATCH v5 16/18 x86: pmu: Adjust lower boundary of branch-misses event
In reply to Dapeng MiSince the IBPB command is added to force to trigger a branch miss at least, the lower boundary of branch misses event is increased to 1 by default. For these CPUs without IBPB support, adjust dynamically the lower boundary to 0 to avoid false positive.
--- x86/pmu.c | 25 +++++++++++++++++++++---- 1 file changed, 21 insertions(+), 4 deletions(-)
x86/pmu.c +21 -4
@@ -90,12 +90,12 @@ struct pmu_event { {"llc references", 0x4f2e, 1, 2*N}, {"llc misses", 0x412e, 1, 1*N}, {"branches", 0x00c4, 1*N, 1.1*N}, - {"branch misses", 0x00c5, 0, 0.1*N}, + {"branch misses", 0x00c5, 1, 0.1*N}, }, amd_gp_events[] = { {"core cycles", 0x0076, 1*N, 50*N}, {"instructions", 0x00c0, 10*N, 10.2*N}, {"branches", 0x00c2, 1*N, 1.1*N}, - {"branch misses", 0x00c3, 0, 0.1*N}, + {"branch misses", 0x00c3, 1, 0.1*N}, }, fixed_events[] = { {"fixed 0", MSR_CORE_PERF_FIXED_CTR0, 10*N, 10.2*N}, {"fixed 1", MSR_CORE_PERF_FIXED_CTR0 + 1, 1*N, 30*N},
@@ -111,6 +111,7 @@ enum {INTEL_REF_CYCLES_IDX =2 ,INTEL_LLC_MISSES_IDX =4 ,INTEL_BRANCHES_IDX =5 , + INTEL_BRANCH_MISS_IDX = 6, };/*
@@ -120,6 +121,7 @@ enum {enum {AMD_INSTRUCTIONS_IDX =1 ,AMD_BRANCHES_IDX =2 , +AMD_BRANCH_MISS_IDX =3 , };char *buf;
@@ -183,7 +185,8 @@ static inline void loop(u64 cntrs) }static void adjust_events_range (struct pmu_event *gp_events, -int instruction_idx,int branch_idx) +int instruction_idx,int branch_idx, +int branch_miss_idx) {/* * If HW supports GLOBAL_CTRL MSR, enabling and disabling PMCs are
@@ -198,6 +201,17 @@ static void adjust_events_range(struct pmu_event *gp_events, gp_events[branch_idx].min =LOOP_BRANCHES ; gp_events[branch_idx].max =LOOP_BRANCHES ; } + +/* +* For CPUs without IBPB support, no way to force to trigger a +* branch miss and the measured branch misses is possible to be +* 0. Thus overwrite the lower boundary of branch misses event +* to 0 to avoid false positive. +*/ +if (!has_ibpb ()) { +/* branch misses event */ + gp_events[branch_miss_idx].min =0 ; + } }volatile uint64_t irq_received;
@@ -903,6 +917,7 @@ int main(int ac, char **av) {int instruction_idx;int branch_idx; +int branch_miss_idx;setup_vm ();handle_irq (PMI_VECTOR , cnt_overflow);
@@ -919,6 +934,7 @@ int main(int ac, char **av) gp_events_size =sizeof (intel_gp_events)/sizeof (intel_gp_events[0 ]); instruction_idx =INTEL_INSTRUCTIONS_IDX ; branch_idx =INTEL_BRANCHES_IDX ; + branch_miss_idx =INTEL_BRANCH_MISS_IDX ;/* * For legacy Intel CPUS without clflush/clflushopt support,
@@ -935,9 +951,10 @@ int main(int ac, char **av) gp_events = (struct pmu_event *)amd_gp_events; instruction_idx =AMD_INSTRUCTIONS_IDX ; branch_idx =AMD_BRANCHES_IDX ; + branch_miss_idx =AMD_BRANCH_MISS_IDX ;report_prefix_push ("AMD" ); } -adjust_events_range (gp_events, instruction_idx, branch_idx); +adjust_events_range (gp_events, instruction_idx, branch_idx, branch_miss_idx);printf ("PMU version: %d\n" , pmu.version);printf ("GP counters: %d\n" , pmu.nr_gp_counters);
Signature · 1 line
2.40.1
PATCH v5 17/18 x86: pmu: Optimize emulated instruction validation
In reply to Dapeng MiFor support CPUs supporting PERF_GLOBAL_CTRL MSR, the validation for emulated instruction can be improved to check against precise counts for instructions and branches events instead of a rough range.
Move enabling and disabling PERF_GLOBAL_CTRL MSR into kvm_fep_asm blob, thus instructions and branches events can be verified against precise counts.
--- x86/pmu.c | 108 ++++++++++++++++++++++++++++++++---------------------- 1 file changed, 65 insertions(+), 43 deletions(-)
x86/pmu.c +65 -43
@@ -14,11 +14,6 @@#define N 1000000 -// These values match the number of instructions and branches in the -// assembly block in check_emulated_instr(). -#define EXPECTED_INSTR 17 -#define EXPECTED_BRNCH 5 -#define IBPB_JMP_INSTRNS 9#define IBPB_JMP_BRANCHES 2
@@ -71,6 +66,40 @@ do { \ : "edi"); \ }while (0 ) +/* the number of instructions and branches of the kvm_fep_asm() blob */ +#define KVM_FEP_INSTR 22 +#define KVM_FEP_BRNCH 5 + +/* +* KVM_FEP is a magic prefix that forces emulation so +* 'KVM_FEP "jne label\n"' just counts as a single instruction. +*/ +#define kvm_fep_asm (_wrmsr) \ +do { \ + asm volatile( \ + _wrmsr "\n\t" \ + "mov %%ecx, %%edi;\n\t" \ + "mov $0x0, %%eax;\n\t" \ + "cmp $0x0, %%eax;\n\t" \ + KVM_FEP "jne 1f\n\t" \ + KVM_FEP "jne 1f\n\t" \ + KVM_FEP "jne 1f\n\t" \ + KVM_FEP "jne 1f\n\t" \ + KVM_FEP "jne 1f\n\t" \ + "mov $0xa, %%eax; cpuid;\n\t" \ + "mov $0xa, %%eax; cpuid;\n\t" \ + "mov $0xa, %%eax; cpuid;\n\t" \ + "mov $0xa, %%eax; cpuid;\n\t" \ + "mov $0xa, %%eax; cpuid;\n\t" \ + "1: mov %%edi, %%ecx; \n\t" \ + "xor %%eax, %%eax; \n\t" \ + "xor %%edx, %%edx;\n\t" \ + _wrmsr "\n\t" \ + : \ + : "a"(eax), "d"(edx), "c"(ecx) \ + : "ebx", "edi"); \ +} while (0) + typedef struct { uint32_t ctr; uint32_t idx;
@@ -657,6 +686,7 @@ static void check_running_counter_wrmsr(void)static void check_emulated_instr (void ) { +u32 eax, edx, ecx;uint64_t status, instr_start, brnch_start;uint64_t gp_counter_width = (1ull << pmu.gp_counter_width) -1 ; unsigned int branch_idx = pmu.is_intel ?
@@ -664,6 +694,7 @@ static void check_emulated_instr(void)unsigned int instruction_idx = pmu.is_intel ?INTEL_INSTRUCTIONS_IDX :AMD_INSTRUCTIONS_IDX ; +pmu_counter_t brnch_cnt = { .ctr =MSR_GP_COUNTERx (0 ),/* branch instructions */
@@ -679,55 +710,46 @@ static void check_emulated_instr(void)if (this_cpu_has_perf_global_status ())pmu_clear_global_status (); -start_event (&brnch_cnt); -start_event (&instr_cnt); +__start_event (&brnch_cnt,0 ); +__start_event (&instr_cnt,0 ); - brnch_start = -EXPECTED_BRNCH ; - instr_start = -EXPECTED_INSTR ; + brnch_start = -KVM_FEP_BRNCH ; + instr_start = -KVM_FEP_INSTR ;wrmsr (MSR_GP_COUNTERx (0 ), brnch_start & gp_counter_width);wrmsr (MSR_GP_COUNTERx (1 ), instr_start & gp_counter_width); -// KVM_FEP is a magic prefix that forces emulation so -// 'KVM_FEP "jne label\n"' just counts as a single instruction. - asmvolatile ( -"mov $0x0, %%eax\n" -"cmp $0x0, %%eax\n" -KVM_FEP "jne label\n" -KVM_FEP "jne label\n" -KVM_FEP "jne label\n" -KVM_FEP "jne label\n" -KVM_FEP "jne label\n" -"mov $0xa, %%eax\n" -"cpuid\n" -"mov $0xa, %%eax\n" -"cpuid\n" -"mov $0xa, %%eax\n" -"cpuid\n" -"mov $0xa, %%eax\n" -"cpuid\n" -"mov $0xa, %%eax\n" -"cpuid\n" -"label:\n" - : - : - :"eax" ,"ebx" ,"ecx" ,"edx" ); -if (this_cpu_has_perf_global_ctrl ()) -wrmsr (pmu.msr_global_ctl,0 ); +if (this_cpu_has_perf_global_ctrl ()) { + eax =BIT (0 ) |BIT (1 ); + ecx = pmu.msr_global_ctl; + edx =0 ; +kvm_fep_asm ("wrmsr" ); + }else { + eax = ecx = edx =0 ; +kvm_fep_asm ("nop" ); + } -stop_event (&brnch_cnt); -stop_event (&instr_cnt); +__stop_event (&brnch_cnt); +__stop_event (&instr_cnt);// Check that the end count - start count is at least the expected // number of instructions and branches. -report (instr_cnt.count - instr_start >=EXPECTED_INSTR , -"instruction count" ); -report (brnch_cnt.count - brnch_start >=EXPECTED_BRNCH , -"branch count" ); +if (this_cpu_has_perf_global_ctrl ()) { +report (instr_cnt.count - instr_start ==KVM_FEP_INSTR , +"instruction count" ); +report (brnch_cnt.count - brnch_start ==KVM_FEP_BRNCH , +"branch count" ); + }else { +report (instr_cnt.count - instr_start >=KVM_FEP_INSTR , +"instruction count" ); +report (brnch_cnt.count - brnch_start >=KVM_FEP_BRNCH , +"branch count" ); + } +if (this_cpu_has_perf_global_status ()) {// Additionally check that those counters overflowed properly. status =rdmsr (pmu.msr_global_status); -report (status &1 ,"branch counter overflow" ); -report (status &2 ,"instruction counter overflow" ); +report (status &BIT_ULL (0 ),"branch counter overflow" ); +report (status &BIT_ULL (1 ),"instruction counter overflow" ); }report_prefix_pop ();
Signature · 1 line
2.40.1
PATCH v5 18/18 x86: pmu: Print measured event count if test fails
In reply to Dapeng MiPrint the measured event count if the test case fails. This helps users quickly know why the test case fails.
--- x86/pmu.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
x86/pmu.c +6 -2
@@ -398,8 +398,12 @@ static noinline void __measure(pmu_counter_t *evt, uint64_t count)static bool verify_event (uint64_t count,struct pmu_event *e) { -// printf("%d <= %ld <= %d\n", e->min, count, e->max); -return count >= e->min && count <= e->max; +bool pass = count >= e->min && count <= e->max; + +if (!pass) +printf ("FAIL: %d <= %" PRId64" <= %d\n" , e->min, count, e->max); + +return pass; }static bool verify_counter (pmu_counter_t *cnt)
Signature · 1 line
2.40.1