arch/x86/events/msr.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/perf_event.h>
   3 #include <linux/sysfs.h>
   4 #include <linux/nospec.h>
   5 #include <asm/intel-family.h>
   6 #include "probe.h"
   7
   8 enum perf_msr_id {
   9         PERF_MSR_TSC                    = 0,
  10         PERF_MSR_APERF                  = 1,
  11         PERF_MSR_MPERF                  = 2,
  12         PERF_MSR_PPERF                  = 3,
  13         PERF_MSR_SMI                    = 4,
  14         PERF_MSR_PTSC                   = 5,
  15         PERF_MSR_IRPERF                 = 6,
  16         PERF_MSR_THERM                  = 7,
  17         PERF_MSR_EVENT_MAX,
  18 };
  19
  20 static bool test_aperfmperf(int idx, void *data)
  21 {
  22         return boot_cpu_has(X86_FEATURE_APERFMPERF);
  23 }
  24
  25 static bool test_ptsc(int idx, void *data)
  26 {
  27         return boot_cpu_has(X86_FEATURE_PTSC);
  28 }
  29
  30 static bool test_irperf(int idx, void *data)
  31 {
  32         return boot_cpu_has(X86_FEATURE_IRPERF);
  33 }
  34
  35 static bool test_therm_status(int idx, void *data)
  36 {
  37         return boot_cpu_has(X86_FEATURE_DTHERM);
  38 }
  39
  40 static bool test_intel(int idx, void *data)
  41 {
  42         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
  43             boot_cpu_data.x86 != 6)
  44                 return false;
  45
  46         switch (boot_cpu_data.x86_model) {
  47         case INTEL_FAM6_NEHALEM:
  48         case INTEL_FAM6_NEHALEM_G:
  49         case INTEL_FAM6_NEHALEM_EP:
  50         case INTEL_FAM6_NEHALEM_EX:
  51
  52         case INTEL_FAM6_WESTMERE:
  53         case INTEL_FAM6_WESTMERE_EP:
  54         case INTEL_FAM6_WESTMERE_EX:
  55
  56         case INTEL_FAM6_SANDYBRIDGE:
  57         case INTEL_FAM6_SANDYBRIDGE_X:
  58
  59         case INTEL_FAM6_IVYBRIDGE:
  60         case INTEL_FAM6_IVYBRIDGE_X:
  61
  62         case INTEL_FAM6_HASWELL:
  63         case INTEL_FAM6_HASWELL_X:
  64         case INTEL_FAM6_HASWELL_L:
  65         case INTEL_FAM6_HASWELL_G:
  66
  67         case INTEL_FAM6_BROADWELL:
  68         case INTEL_FAM6_BROADWELL_D:
  69         case INTEL_FAM6_BROADWELL_G:
  70         case INTEL_FAM6_BROADWELL_X:
  71         case INTEL_FAM6_SAPPHIRERAPIDS_X:
  72         case INTEL_FAM6_EMERALDRAPIDS_X:
  73         case INTEL_FAM6_GRANITERAPIDS_X:
  74         case INTEL_FAM6_GRANITERAPIDS_D:
  75
  76         case INTEL_FAM6_ATOM_SILVERMONT:
  77         case INTEL_FAM6_ATOM_SILVERMONT_D:
  78         case INTEL_FAM6_ATOM_AIRMONT:
  79
  80         case INTEL_FAM6_ATOM_GOLDMONT:
  81         case INTEL_FAM6_ATOM_GOLDMONT_D:
  82         case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
  83         case INTEL_FAM6_ATOM_TREMONT_D:
  84         case INTEL_FAM6_ATOM_TREMONT:
  85         case INTEL_FAM6_ATOM_TREMONT_L:
  86
  87         case INTEL_FAM6_XEON_PHI_KNL:
  88         case INTEL_FAM6_XEON_PHI_KNM:
  89                 if (idx == PERF_MSR_SMI)
  90                         return true;
  91                 break;
  92
  93         case INTEL_FAM6_SKYLAKE_L:
  94         case INTEL_FAM6_SKYLAKE:
  95         case INTEL_FAM6_SKYLAKE_X:
  96         case INTEL_FAM6_KABYLAKE_L:
  97         case INTEL_FAM6_KABYLAKE:
  98         case INTEL_FAM6_COMETLAKE_L:
  99         case INTEL_FAM6_COMETLAKE:
 100         case INTEL_FAM6_ICELAKE_L:
 101         case INTEL_FAM6_ICELAKE:
 102         case INTEL_FAM6_ICELAKE_X:
 103         case INTEL_FAM6_ICELAKE_D:
 104         case INTEL_FAM6_TIGERLAKE_L:
 105         case INTEL_FAM6_TIGERLAKE:
 106         case INTEL_FAM6_ROCKETLAKE:
 107         case INTEL_FAM6_ALDERLAKE:
 108         case INTEL_FAM6_ALDERLAKE_L:
 109         case INTEL_FAM6_ATOM_GRACEMONT:
 110         case INTEL_FAM6_RAPTORLAKE:
 111         case INTEL_FAM6_RAPTORLAKE_P:
 112         case INTEL_FAM6_RAPTORLAKE_S:
 113         case INTEL_FAM6_METEORLAKE:
 114         case INTEL_FAM6_METEORLAKE_L:
 115                 if (idx == PERF_MSR_SMI || idx == PERF_MSR_PPERF)
 116                         return true;
 117                 break;
 118         }
 119
 120         return false;
 121 }
 122
 123 PMU_EVENT_ATTR_STRING(tsc,                              attr_tsc,               "event=0x00"    );
 124 PMU_EVENT_ATTR_STRING(aperf,                            attr_aperf,             "event=0x01"    );
 125 PMU_EVENT_ATTR_STRING(mperf,                            attr_mperf,             "event=0x02"    );
 126 PMU_EVENT_ATTR_STRING(pperf,                            attr_pperf,             "event=0x03"    );
 127 PMU_EVENT_ATTR_STRING(smi,                              attr_smi,               "event=0x04"    );
 128 PMU_EVENT_ATTR_STRING(ptsc,                             attr_ptsc,              "event=0x05"    );
 129 PMU_EVENT_ATTR_STRING(irperf,                           attr_irperf,            "event=0x06"    );
 130 PMU_EVENT_ATTR_STRING(cpu_thermal_margin,               attr_therm,             "event=0x07"    );
 131 PMU_EVENT_ATTR_STRING(cpu_thermal_margin.snapshot,      attr_therm_snap,        "1"             );
 132 PMU_EVENT_ATTR_STRING(cpu_thermal_margin.unit,          attr_therm_unit,        "C"             );
 133
 134 static unsigned long msr_mask;
 135
 136 PMU_EVENT_GROUP(events, aperf);
 137 PMU_EVENT_GROUP(events, mperf);
 138 PMU_EVENT_GROUP(events, pperf);
 139 PMU_EVENT_GROUP(events, smi);
 140 PMU_EVENT_GROUP(events, ptsc);
 141 PMU_EVENT_GROUP(events, irperf);
 142
 143 static struct attribute *attrs_therm[] = {
 144         &attr_therm.attr.attr,
 145         &attr_therm_snap.attr.attr,
 146         &attr_therm_unit.attr.attr,
 147         NULL,
 148 };
 149
 150 static struct attribute_group group_therm = {
 151         .name  = "events",
 152         .attrs = attrs_therm,
 153 };
 154
 155 static struct perf_msr msr[] = {
 156         [PERF_MSR_TSC]          = { .no_check = true,                                                           },
 157         [PERF_MSR_APERF]        = { MSR_IA32_APERF,             &group_aperf,           test_aperfmperf,        },
 158         [PERF_MSR_MPERF]        = { MSR_IA32_MPERF,             &group_mperf,           test_aperfmperf,        },
 159         [PERF_MSR_PPERF]        = { MSR_PPERF,                  &group_pperf,           test_intel,             },
 160         [PERF_MSR_SMI]          = { MSR_SMI_COUNT,              &group_smi,             test_intel,             },
 161         [PERF_MSR_PTSC]         = { MSR_F15H_PTSC,              &group_ptsc,            test_ptsc,              },
 162         [PERF_MSR_IRPERF]       = { MSR_F17H_IRPERF,            &group_irperf,          test_irperf,            },
 163         [PERF_MSR_THERM]        = { MSR_IA32_THERM_STATUS,      &group_therm,           test_therm_status,      },
 164 };
 165
 166 static struct attribute *events_attrs[] = {
 167         &attr_tsc.attr.attr,
 168         NULL,
 169 };
 170
 171 static struct attribute_group events_attr_group = {
 172         .name = "events",
 173         .attrs = events_attrs,
 174 };
 175
 176 PMU_FORMAT_ATTR(event, "config:0-63");
 177 static struct attribute *format_attrs[] = {
 178         &format_attr_event.attr,
 179         NULL,
 180 };
 181 static struct attribute_group format_attr_group = {
 182         .name = "format",
 183         .attrs = format_attrs,
 184 };
 185
 186 static const struct attribute_group *attr_groups[] = {
 187         &events_attr_group,
 188         &format_attr_group,
 189         NULL,
 190 };
 191
 192 static const struct attribute_group *attr_update[] = {
 193         &group_aperf,
 194         &group_mperf,
 195         &group_pperf,
 196         &group_smi,
 197         &group_ptsc,
 198         &group_irperf,
 199         &group_therm,
 200         NULL,
 201 };
 202
 203 static int msr_event_init(struct perf_event *event)
 204 {
 205         u64 cfg = event->attr.config;
 206
 207         if (event->attr.type != event->pmu->type)
 208                 return -ENOENT;
 209
 210         /* unsupported modes and filters */
 211         if (event->attr.sample_period) /* no sampling */
 212                 return -EINVAL;
 213
 214         if (cfg >= PERF_MSR_EVENT_MAX)
 215                 return -EINVAL;
 216
 217         cfg = array_index_nospec((unsigned long)cfg, PERF_MSR_EVENT_MAX);
 218
 219         if (!(msr_mask & (1 << cfg)))
 220                 return -EINVAL;
 221
 222         event->hw.idx           = -1;
 223         event->hw.event_base    = msr[cfg].msr;
 224         event->hw.config        = cfg;
 225
 226         return 0;
 227 }
 228
 229 static inline u64 msr_read_counter(struct perf_event *event)
 230 {
 231         u64 now;
 232
 233         if (event->hw.event_base)
 234                 rdmsrl(event->hw.event_base, now);
 235         else
 236                 now = rdtsc_ordered();
 237
 238         return now;
 239 }
 240
 241 static void msr_event_update(struct perf_event *event)
 242 {
 243         u64 prev, now;
 244         s64 delta;
 245
 246         /* Careful, an NMI might modify the previous event value: */
 247         prev = local64_read(&event->hw.prev_count);
 248         do {
 249                 now = msr_read_counter(event);
 250         } while (!local64_try_cmpxchg(&event->hw.prev_count, &prev, now));
 251
 252         delta = now - prev;
 253         if (unlikely(event->hw.event_base == MSR_SMI_COUNT)) {
 254                 delta = sign_extend64(delta, 31);
 255                 local64_add(delta, &event->count);
 256         } else if (unlikely(event->hw.event_base == MSR_IA32_THERM_STATUS)) {
 257                 /* If valid, extract digital readout, otherwise set to -1: */
 258                 now = now & (1ULL << 31) ? (now >> 16) & 0x3f :  -1;
 259                 local64_set(&event->count, now);
 260         } else {
 261                 local64_add(delta, &event->count);
 262         }
 263 }
 264
 265 static void msr_event_start(struct perf_event *event, int flags)
 266 {
 267         u64 now = msr_read_counter(event);
 268
 269         local64_set(&event->hw.prev_count, now);
 270 }
 271
 272 static void msr_event_stop(struct perf_event *event, int flags)
 273 {
 274         msr_event_update(event);
 275 }
 276
 277 static void msr_event_del(struct perf_event *event, int flags)
 278 {
 279         msr_event_stop(event, PERF_EF_UPDATE);
 280 }
 281
 282 static int msr_event_add(struct perf_event *event, int flags)
 283 {
 284         if (flags & PERF_EF_START)
 285                 msr_event_start(event, flags);
 286
 287         return 0;
 288 }
 289
 290 static struct pmu pmu_msr = {
 291         .task_ctx_nr    = perf_sw_context,
 292         .attr_groups    = attr_groups,
 293         .event_init     = msr_event_init,
 294         .add            = msr_event_add,
 295         .del            = msr_event_del,
 296         .start          = msr_event_start,
 297         .stop           = msr_event_stop,
 298         .read           = msr_event_update,
 299         .capabilities   = PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
 300         .attr_update    = attr_update,
 301 };
 302
 303 static int __init msr_init(void)
 304 {
 305         if (!boot_cpu_has(X86_FEATURE_TSC)) {
 306                 pr_cont("no MSR PMU driver.\n");
 307                 return 0;
 308         }
 309
 310         msr_mask = perf_msr_probe(msr, PERF_MSR_EVENT_MAX, true, NULL);
 311
 312         perf_pmu_register(&pmu_msr, "msr", -1);
 313
 314         return 0;
 315 }
 316 device_initcall(msr_init);