1#include <linux/perf_event.h> 2 3enum perf_msr_id { 4 PERF_MSR_TSC = 0, 5 PERF_MSR_APERF = 1, 6 PERF_MSR_MPERF = 2, 7 PERF_MSR_PPERF = 3, 8 PERF_MSR_SMI = 4, 9 10 PERF_MSR_EVENT_MAX, 11}; 12 13static bool test_aperfmperf(int idx) 14{ 15 return boot_cpu_has(X86_FEATURE_APERFMPERF); 16} 17 18static bool test_intel(int idx) 19{ 20 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL || 21 boot_cpu_data.x86 != 6) 22 return false; 23 24 switch (boot_cpu_data.x86_model) { 25 case 30: /* 45nm Nehalem */ 26 case 26: /* 45nm Nehalem-EP */ 27 case 46: /* 45nm Nehalem-EX */ 28 29 case 37: /* 32nm Westmere */ 30 case 44: /* 32nm Westmere-EP */ 31 case 47: /* 32nm Westmere-EX */ 32 33 case 42: /* 32nm SandyBridge */ 34 case 45: /* 32nm SandyBridge-E/EN/EP */ 35 36 case 58: /* 22nm IvyBridge */ 37 case 62: /* 22nm IvyBridge-EP/EX */ 38 39 case 60: /* 22nm Haswell Core */ 40 case 63: /* 22nm Haswell Server */ 41 case 69: /* 22nm Haswell ULT */ 42 case 70: /* 22nm Haswell + GT3e (Intel Iris Pro graphics) */ 43 44 case 61: /* 14nm Broadwell Core-M */ 45 case 86: /* 14nm Broadwell Xeon D */ 46 case 71: /* 14nm Broadwell + GT3e (Intel Iris Pro graphics) */ 47 case 79: /* 14nm Broadwell Server */ 48 49 case 55: /* 22nm Atom "Silvermont" */ 50 case 77: /* 22nm Atom "Silvermont Avoton/Rangely" */ 51 case 76: /* 14nm Atom "Airmont" */ 52 if (idx == PERF_MSR_SMI) 53 return true; 54 break; 55 56 case 78: /* 14nm Skylake Mobile */ 57 case 94: /* 14nm Skylake Desktop */ 58 if (idx == PERF_MSR_SMI || idx == PERF_MSR_PPERF) 59 return true; 60 break; 61 } 62 63 return false; 64} 65 66struct perf_msr { 67 u64 msr; 68 struct perf_pmu_events_attr *attr; 69 bool (*test)(int idx); 70}; 71 72PMU_EVENT_ATTR_STRING(tsc, evattr_tsc, "event=0x00"); 73PMU_EVENT_ATTR_STRING(aperf, evattr_aperf, "event=0x01"); 74PMU_EVENT_ATTR_STRING(mperf, evattr_mperf, "event=0x02"); 75PMU_EVENT_ATTR_STRING(pperf, evattr_pperf, "event=0x03"); 76PMU_EVENT_ATTR_STRING(smi, evattr_smi, "event=0x04"); 77 78static struct perf_msr msr[] = { 79 [PERF_MSR_TSC] = { 0, &evattr_tsc, NULL, }, 80 [PERF_MSR_APERF] = { MSR_IA32_APERF, &evattr_aperf, test_aperfmperf, }, 81 [PERF_MSR_MPERF] = { MSR_IA32_MPERF, &evattr_mperf, test_aperfmperf, }, 82 [PERF_MSR_PPERF] = { MSR_PPERF, &evattr_pperf, test_intel, }, 83 [PERF_MSR_SMI] = { MSR_SMI_COUNT, &evattr_smi, test_intel, }, 84}; 85 86static struct attribute *events_attrs[PERF_MSR_EVENT_MAX + 1] = { 87 NULL, 88}; 89 90static struct attribute_group events_attr_group = { 91 .name = "events", 92 .attrs = events_attrs, 93}; 94 95PMU_FORMAT_ATTR(event, "config:0-63"); 96static struct attribute *format_attrs[] = { 97 &format_attr_event.attr, 98 NULL, 99}; 100static struct attribute_group format_attr_group = { 101 .name = "format", 102 .attrs = format_attrs, 103}; 104 105static const struct attribute_group *attr_groups[] = { 106 &events_attr_group, 107 &format_attr_group, 108 NULL, 109}; 110 111static int msr_event_init(struct perf_event *event) 112{ 113 u64 cfg = event->attr.config; 114 115 if (event->attr.type != event->pmu->type) 116 return -ENOENT; 117 118 if (cfg >= PERF_MSR_EVENT_MAX) 119 return -EINVAL; 120 121 /* unsupported modes and filters */ 122 if (event->attr.exclude_user || 123 event->attr.exclude_kernel || 124 event->attr.exclude_hv || 125 event->attr.exclude_idle || 126 event->attr.exclude_host || 127 event->attr.exclude_guest || 128 event->attr.sample_period) /* no sampling */ 129 return -EINVAL; 130 131 if (!msr[cfg].attr) 132 return -EINVAL; 133 134 event->hw.idx = -1; 135 event->hw.event_base = msr[cfg].msr; 136 event->hw.config = cfg; 137 138 return 0; 139} 140 141static inline u64 msr_read_counter(struct perf_event *event) 142{ 143 u64 now; 144 145 if (event->hw.event_base) 146 rdmsrl(event->hw.event_base, now); 147 else 148 rdtscll(now); 149 150 return now; 151} 152static void msr_event_update(struct perf_event *event) 153{ 154 u64 prev, now; 155 s64 delta; 156 157 /* Careful, an NMI might modify the previous event value. */ 158again: 159 prev = local64_read(&event->hw.prev_count); 160 now = msr_read_counter(event); 161 162 if (local64_cmpxchg(&event->hw.prev_count, prev, now) != prev) 163 goto again; 164 165 delta = now - prev; 166 if (unlikely(event->hw.event_base == MSR_SMI_COUNT)) 167 delta = sign_extend64(delta, 31); 168 169 local64_add(now - prev, &event->count); 170} 171 172static void msr_event_start(struct perf_event *event, int flags) 173{ 174 u64 now; 175 176 now = msr_read_counter(event); 177 local64_set(&event->hw.prev_count, now); 178} 179 180static void msr_event_stop(struct perf_event *event, int flags) 181{ 182 msr_event_update(event); 183} 184 185static void msr_event_del(struct perf_event *event, int flags) 186{ 187 msr_event_stop(event, PERF_EF_UPDATE); 188} 189 190static int msr_event_add(struct perf_event *event, int flags) 191{ 192 if (flags & PERF_EF_START) 193 msr_event_start(event, flags); 194 195 return 0; 196} 197 198static struct pmu pmu_msr = { 199 .task_ctx_nr = perf_sw_context, 200 .attr_groups = attr_groups, 201 .event_init = msr_event_init, 202 .add = msr_event_add, 203 .del = msr_event_del, 204 .start = msr_event_start, 205 .stop = msr_event_stop, 206 .read = msr_event_update, 207 .capabilities = PERF_PMU_CAP_NO_INTERRUPT, 208}; 209 210static int __init msr_init(void) 211{ 212 int i, j = 0; 213 214 if (!boot_cpu_has(X86_FEATURE_TSC)) { 215 pr_cont("no MSR PMU driver.\n"); 216 return 0; 217 } 218 219 /* Probe the MSRs. */ 220 for (i = PERF_MSR_TSC + 1; i < PERF_MSR_EVENT_MAX; i++) { 221 u64 val; 222 223 /* 224 * Virt sucks arse; you cannot tell if a R/O MSR is present :/ 225 */ 226 if (!msr[i].test(i) || rdmsrl_safe(msr[i].msr, &val)) 227 msr[i].attr = NULL; 228 } 229 230 /* List remaining MSRs in the sysfs attrs. */ 231 for (i = 0; i < PERF_MSR_EVENT_MAX; i++) { 232 if (msr[i].attr) 233 events_attrs[j++] = &msr[i].attr->attr.attr; 234 } 235 events_attrs[j] = NULL; 236 237 perf_pmu_register(&pmu_msr, "msr", -1); 238 239 return 0; 240} 241device_initcall(msr_init); 242