aboutsummaryrefslogtreecommitdiffstats
path: root/arch
diff options
context:
space:
mode:
Diffstat (limited to 'arch')
-rw-r--r--arch/x86/include/asm/perf_event.h14
-rw-r--r--arch/x86/kernel/cpu/perf_event.h5
-rw-r--r--arch/x86/kernel/cpu/perf_event_intel.c28
3 files changed, 42 insertions, 5 deletions
diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
index f61c62f7d5d..c6998bc7545 100644
--- a/arch/x86/include/asm/perf_event.h
+++ b/arch/x86/include/asm/perf_event.h
@@ -57,6 +57,7 @@
57 (1 << (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX)) 57 (1 << (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX))
58 58
59#define ARCH_PERFMON_BRANCH_MISSES_RETIRED 6 59#define ARCH_PERFMON_BRANCH_MISSES_RETIRED 6
60#define ARCH_PERFMON_EVENTS_COUNT 7
60 61
61/* 62/*
62 * Intel "Architectural Performance Monitoring" CPUID 63 * Intel "Architectural Performance Monitoring" CPUID
@@ -72,6 +73,19 @@ union cpuid10_eax {
72 unsigned int full; 73 unsigned int full;
73}; 74};
74 75
76union cpuid10_ebx {
77 struct {
78 unsigned int no_unhalted_core_cycles:1;
79 unsigned int no_instructions_retired:1;
80 unsigned int no_unhalted_reference_cycles:1;
81 unsigned int no_llc_reference:1;
82 unsigned int no_llc_misses:1;
83 unsigned int no_branch_instruction_retired:1;
84 unsigned int no_branch_misses_retired:1;
85 } split;
86 unsigned int full;
87};
88
75union cpuid10_edx { 89union cpuid10_edx {
76 struct { 90 struct {
77 unsigned int num_counters_fixed:5; 91 unsigned int num_counters_fixed:5;
diff --git a/arch/x86/kernel/cpu/perf_event.h b/arch/x86/kernel/cpu/perf_event.h
index 51a985cbc12..f49c5c21085 100644
--- a/arch/x86/kernel/cpu/perf_event.h
+++ b/arch/x86/kernel/cpu/perf_event.h
@@ -285,6 +285,11 @@ struct x86_pmu {
285 int num_counters_fixed; 285 int num_counters_fixed;
286 int cntval_bits; 286 int cntval_bits;
287 u64 cntval_mask; 287 u64 cntval_mask;
288 union {
289 unsigned long events_maskl;
290 unsigned long events_mask[BITS_TO_LONGS(ARCH_PERFMON_EVENTS_COUNT)];
291 };
292 int events_mask_len;
288 int apic; 293 int apic;
289 u64 max_period; 294 u64 max_period;
290 struct event_constraint * 295 struct event_constraint *
diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
index 8d601b18bf9..201156b80a3 100644
--- a/arch/x86/kernel/cpu/perf_event_intel.c
+++ b/arch/x86/kernel/cpu/perf_event_intel.c
@@ -1552,13 +1552,23 @@ static void intel_sandybridge_quirks(void)
1552 x86_pmu.pebs_constraints = NULL; 1552 x86_pmu.pebs_constraints = NULL;
1553} 1553}
1554 1554
1555static const int intel_event_id_to_hw_id[] __initconst = {
1556 PERF_COUNT_HW_CPU_CYCLES,
1557 PERF_COUNT_HW_INSTRUCTIONS,
1558 PERF_COUNT_HW_BUS_CYCLES,
1559 PERF_COUNT_HW_CACHE_REFERENCES,
1560 PERF_COUNT_HW_CACHE_MISSES,
1561 PERF_COUNT_HW_BRANCH_INSTRUCTIONS,
1562 PERF_COUNT_HW_BRANCH_MISSES,
1563};
1564
1555__init int intel_pmu_init(void) 1565__init int intel_pmu_init(void)
1556{ 1566{
1557 union cpuid10_edx edx; 1567 union cpuid10_edx edx;
1558 union cpuid10_eax eax; 1568 union cpuid10_eax eax;
1569 union cpuid10_ebx ebx;
1559 unsigned int unused; 1570 unsigned int unused;
1560 unsigned int ebx; 1571 int version, bit;
1561 int version;
1562 1572
1563 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 1573 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
1564 switch (boot_cpu_data.x86) { 1574 switch (boot_cpu_data.x86) {
@@ -1574,8 +1584,8 @@ __init int intel_pmu_init(void)
1574 * Check whether the Architectural PerfMon supports 1584 * Check whether the Architectural PerfMon supports
1575 * Branch Misses Retired hw_event or not. 1585 * Branch Misses Retired hw_event or not.
1576 */ 1586 */
1577 cpuid(10, &eax.full, &ebx, &unused, &edx.full); 1587 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
1578 if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED) 1588 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT)
1579 return -ENODEV; 1589 return -ENODEV;
1580 1590
1581 version = eax.split.version_id; 1591 version = eax.split.version_id;
@@ -1651,7 +1661,7 @@ __init int intel_pmu_init(void)
1651 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 1661 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
1652 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1; 1662 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
1653 1663
1654 if (ebx & 0x40) { 1664 if (ebx.split.no_branch_misses_retired) {
1655 /* 1665 /*
1656 * Erratum AAJ80 detected, we work it around by using 1666 * Erratum AAJ80 detected, we work it around by using
1657 * the BR_MISP_EXEC.ANY event. This will over-count 1667 * the BR_MISP_EXEC.ANY event. This will over-count
@@ -1659,6 +1669,7 @@ __init int intel_pmu_init(void)
1659 * architectural event which is often completely bogus: 1669 * architectural event which is often completely bogus:
1660 */ 1670 */
1661 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 1671 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
1672 ebx.split.no_branch_misses_retired = 0;
1662 1673
1663 pr_cont("erratum AAJ80 worked around, "); 1674 pr_cont("erratum AAJ80 worked around, ");
1664 } 1675 }
@@ -1738,5 +1749,12 @@ __init int intel_pmu_init(void)
1738 break; 1749 break;
1739 } 1750 }
1740 } 1751 }
1752 x86_pmu.events_maskl = ebx.full;
1753 x86_pmu.events_mask_len = eax.split.mask_length;
1754
1755 /* disable event that reported as not presend by cpuid */
1756 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_event_id_to_hw_id))
1757 intel_perfmon_event_map[intel_event_id_to_hw_id[bit]] = 0;
1758
1741 return 0; 1759 return 0;
1742} 1760}