arch/x86/events/amd/uncore.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2013 Advanced Micro Devices, Inc.
   4  *
   5  * Author: Jacob Shin <jacob.shin@amd.com>
   6  */
   7
   8 #include <linux/perf_event.h>
   9 #include <linux/percpu.h>
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/init.h>
  13 #include <linux/cpu.h>
  14 #include <linux/cpumask.h>
  15
  16 #include <asm/cpufeature.h>
  17 #include <asm/perf_event.h>
  18 #include <asm/msr.h>
  19 #include <asm/smp.h>
  20
  21 #define NUM_COUNTERS_NB         4
  22 #define NUM_COUNTERS_L2         4
  23 #define NUM_COUNTERS_L3         6
  24 #define MAX_COUNTERS            6
  25
  26 #define RDPMC_BASE_NB           6
  27 #define RDPMC_BASE_LLC          10
  28
  29 #define COUNTER_SHIFT           16
  30
  31 #undef pr_fmt
  32 #define pr_fmt(fmt)     "amd_uncore: " fmt
  33
  34 static int num_counters_llc;
  35 static int num_counters_nb;
  36 static bool l3_mask;
  37
  38 static HLIST_HEAD(uncore_unused_list);
  39
  40 struct amd_uncore {
  41         int id;
  42         int refcnt;
  43         int cpu;
  44         int num_counters;
  45         int rdpmc_base;
  46         u32 msr_base;
  47         cpumask_t *active_mask;
  48         struct pmu *pmu;
  49         struct perf_event *events[MAX_COUNTERS];
  50         struct hlist_node node;
  51 };
  52
  53 static struct amd_uncore * __percpu *amd_uncore_nb;
  54 static struct amd_uncore * __percpu *amd_uncore_llc;
  55
  56 static struct pmu amd_nb_pmu;
  57 static struct pmu amd_llc_pmu;
  58
  59 static cpumask_t amd_nb_active_mask;
  60 static cpumask_t amd_llc_active_mask;
  61
  62 static bool is_nb_event(struct perf_event *event)
  63 {
  64         return event->pmu->type == amd_nb_pmu.type;
  65 }
  66
  67 static bool is_llc_event(struct perf_event *event)
  68 {
  69         return event->pmu->type == amd_llc_pmu.type;
  70 }
  71
  72 static struct amd_uncore *event_to_amd_uncore(struct perf_event *event)
  73 {
  74         if (is_nb_event(event) && amd_uncore_nb)
  75                 return *per_cpu_ptr(amd_uncore_nb, event->cpu);
  76         else if (is_llc_event(event) && amd_uncore_llc)
  77                 return *per_cpu_ptr(amd_uncore_llc, event->cpu);
  78
  79         return NULL;
  80 }
  81
  82 static void amd_uncore_read(struct perf_event *event)
  83 {
  84         struct hw_perf_event *hwc = &event->hw;
  85         u64 prev, new;
  86         s64 delta;
  87
  88         /*
  89          * since we do not enable counter overflow interrupts,
  90          * we do not have to worry about prev_count changing on us
  91          */
  92
  93         prev = local64_read(&hwc->prev_count);
  94         rdpmcl(hwc->event_base_rdpmc, new);
  95         local64_set(&hwc->prev_count, new);
  96         delta = (new << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
  97         delta >>= COUNTER_SHIFT;
  98         local64_add(delta, &event->count);
  99 }
 100
 101 static void amd_uncore_start(struct perf_event *event, int flags)
 102 {
 103         struct hw_perf_event *hwc = &event->hw;
 104
 105         if (flags & PERF_EF_RELOAD)
 106                 wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 107
 108         hwc->state = 0;
 109         wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
 110         perf_event_update_userpage(event);
 111 }
 112
 113 static void amd_uncore_stop(struct perf_event *event, int flags)
 114 {
 115         struct hw_perf_event *hwc = &event->hw;
 116
 117         wrmsrl(hwc->config_base, hwc->config);
 118         hwc->state |= PERF_HES_STOPPED;
 119
 120         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
 121                 amd_uncore_read(event);
 122                 hwc->state |= PERF_HES_UPTODATE;
 123         }
 124 }
 125
 126 static int amd_uncore_add(struct perf_event *event, int flags)
 127 {
 128         int i;
 129         struct amd_uncore *uncore = event_to_amd_uncore(event);
 130         struct hw_perf_event *hwc = &event->hw;
 131
 132         /* are we already assigned? */
 133         if (hwc->idx != -1 && uncore->events[hwc->idx] == event)
 134                 goto out;
 135
 136         for (i = 0; i < uncore->num_counters; i++) {
 137                 if (uncore->events[i] == event) {
 138                         hwc->idx = i;
 139                         goto out;
 140                 }
 141         }
 142
 143         /* if not, take the first available counter */
 144         hwc->idx = -1;
 145         for (i = 0; i < uncore->num_counters; i++) {
 146                 if (cmpxchg(&uncore->events[i], NULL, event) == NULL) {
 147                         hwc->idx = i;
 148                         break;
 149                 }
 150         }
 151
 152 out:
 153         if (hwc->idx == -1)
 154                 return -EBUSY;
 155
 156         hwc->config_base = uncore->msr_base + (2 * hwc->idx);
 157         hwc->event_base = uncore->msr_base + 1 + (2 * hwc->idx);
 158         hwc->event_base_rdpmc = uncore->rdpmc_base + hwc->idx;
 159         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 160
 161         if (flags & PERF_EF_START)
 162                 amd_uncore_start(event, PERF_EF_RELOAD);
 163
 164         return 0;
 165 }
 166
 167 static void amd_uncore_del(struct perf_event *event, int flags)
 168 {
 169         int i;
 170         struct amd_uncore *uncore = event_to_amd_uncore(event);
 171         struct hw_perf_event *hwc = &event->hw;
 172
 173         amd_uncore_stop(event, PERF_EF_UPDATE);
 174
 175         for (i = 0; i < uncore->num_counters; i++) {
 176                 if (cmpxchg(&uncore->events[i], event, NULL) == event)
 177                         break;
 178         }
 179
 180         hwc->idx = -1;
 181 }
 182
 183 /*
 184  * Return a full thread and slice mask until per-CPU is
 185  * properly supported.
 186  */
 187 static u64 l3_thread_slice_mask(void)
 188 {
 189         if (boot_cpu_data.x86 <= 0x18)
 190                 return AMD64_L3_SLICE_MASK | AMD64_L3_THREAD_MASK;
 191
 192         return AMD64_L3_EN_ALL_SLICES | AMD64_L3_EN_ALL_CORES |
 193                AMD64_L3_F19H_THREAD_MASK;
 194 }
 195
 196 static int amd_uncore_event_init(struct perf_event *event)
 197 {
 198         struct amd_uncore *uncore;
 199         struct hw_perf_event *hwc = &event->hw;
 200
 201         if (event->attr.type != event->pmu->type)
 202                 return -ENOENT;
 203
 204         /*
 205          * NB and Last level cache counters (MSRs) are shared across all cores
 206          * that share the same NB / Last level cache.  On family 16h and below,
 207          * Interrupts can be directed to a single target core, however, event
 208          * counts generated by processes running on other cores cannot be masked
 209          * out. So we do not support sampling and per-thread events via
 210          * CAP_NO_INTERRUPT, and we do not enable counter overflow interrupts:
 211          */
 212         hwc->config = event->attr.config & AMD64_RAW_EVENT_MASK_NB;
 213         hwc->idx = -1;
 214
 215         if (event->cpu < 0)
 216                 return -EINVAL;
 217
 218         /*
 219          * SliceMask and ThreadMask need to be set for certain L3 events.
 220          * For other events, the two fields do not affect the count.
 221          */
 222         if (l3_mask && is_llc_event(event))
 223                 hwc->config |= l3_thread_slice_mask();
 224
 225         uncore = event_to_amd_uncore(event);
 226         if (!uncore)
 227                 return -ENODEV;
 228
 229         /*
 230          * since request can come in to any of the shared cores, we will remap
 231          * to a single common cpu.
 232          */
 233         event->cpu = uncore->cpu;
 234
 235         return 0;
 236 }
 237
 238 static ssize_t amd_uncore_attr_show_cpumask(struct device *dev,
 239                                             struct device_attribute *attr,
 240                                             char *buf)
 241 {
 242         cpumask_t *active_mask;
 243         struct pmu *pmu = dev_get_drvdata(dev);
 244
 245         if (pmu->type == amd_nb_pmu.type)
 246                 active_mask = &amd_nb_active_mask;
 247         else if (pmu->type == amd_llc_pmu.type)
 248                 active_mask = &amd_llc_active_mask;
 249         else
 250                 return 0;
 251
 252         return cpumap_print_to_pagebuf(true, buf, active_mask);
 253 }
 254 static DEVICE_ATTR(cpumask, S_IRUGO, amd_uncore_attr_show_cpumask, NULL);
 255
 256 static struct attribute *amd_uncore_attrs[] = {
 257         &dev_attr_cpumask.attr,
 258         NULL,
 259 };
 260
 261 static struct attribute_group amd_uncore_attr_group = {
 262         .attrs = amd_uncore_attrs,
 263 };
 264
 265 #define DEFINE_UNCORE_FORMAT_ATTR(_var, _name, _format)                 \
 266 static ssize_t __uncore_##_var##_show(struct kobject *kobj,             \
 267                                 struct kobj_attribute *attr,            \
 268                                 char *page)                             \
 269 {                                                                       \
 270         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 271         return sprintf(page, _format "\n");                             \
 272 }                                                                       \
 273 static struct kobj_attribute format_attr_##_var =                       \
 274         __ATTR(_name, 0444, __uncore_##_var##_show, NULL)
 275
 276 DEFINE_UNCORE_FORMAT_ATTR(event12,      event,          "config:0-7,32-35");
 277 DEFINE_UNCORE_FORMAT_ATTR(event14,      event,          "config:0-7,32-35,59-60"); /* F17h+ DF */
 278 DEFINE_UNCORE_FORMAT_ATTR(event8,       event,          "config:0-7");             /* F17h+ L3 */
 279 DEFINE_UNCORE_FORMAT_ATTR(umask,        umask,          "config:8-15");
 280
 281 static struct attribute *amd_uncore_df_format_attr[] = {
 282         &format_attr_event12.attr, /* event14 if F17h+ */
 283         &format_attr_umask.attr,
 284         NULL,
 285 };
 286
 287 static struct attribute *amd_uncore_l3_format_attr[] = {
 288         &format_attr_event12.attr, /* event8 if F17h+ */
 289         &format_attr_umask.attr,
 290         NULL,
 291 };
 292
 293 static struct attribute_group amd_uncore_df_format_group = {
 294         .name = "format",
 295         .attrs = amd_uncore_df_format_attr,
 296 };
 297
 298 static struct attribute_group amd_uncore_l3_format_group = {
 299         .name = "format",
 300         .attrs = amd_uncore_l3_format_attr,
 301 };
 302
 303 static const struct attribute_group *amd_uncore_df_attr_groups[] = {
 304         &amd_uncore_attr_group,
 305         &amd_uncore_df_format_group,
 306         NULL,
 307 };
 308
 309 static const struct attribute_group *amd_uncore_l3_attr_groups[] = {
 310         &amd_uncore_attr_group,
 311         &amd_uncore_l3_format_group,
 312         NULL,
 313 };
 314
 315 static struct pmu amd_nb_pmu = {
 316         .task_ctx_nr    = perf_invalid_context,
 317         .attr_groups    = amd_uncore_df_attr_groups,
 318         .name           = "amd_nb",
 319         .event_init     = amd_uncore_event_init,
 320         .add            = amd_uncore_add,
 321         .del            = amd_uncore_del,
 322         .start          = amd_uncore_start,
 323         .stop           = amd_uncore_stop,
 324         .read           = amd_uncore_read,
 325         .capabilities   = PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
 326 };
 327
 328 static struct pmu amd_llc_pmu = {
 329         .task_ctx_nr    = perf_invalid_context,
 330         .attr_groups    = amd_uncore_l3_attr_groups,
 331         .name           = "amd_l2",
 332         .event_init     = amd_uncore_event_init,
 333         .add            = amd_uncore_add,
 334         .del            = amd_uncore_del,
 335         .start          = amd_uncore_start,
 336         .stop           = amd_uncore_stop,
 337         .read           = amd_uncore_read,
 338         .capabilities   = PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
 339 };
 340
 341 static struct amd_uncore *amd_uncore_alloc(unsigned int cpu)
 342 {
 343         return kzalloc_node(sizeof(struct amd_uncore), GFP_KERNEL,
 344                         cpu_to_node(cpu));
 345 }
 346
 347 static int amd_uncore_cpu_up_prepare(unsigned int cpu)
 348 {
 349         struct amd_uncore *uncore_nb = NULL, *uncore_llc;
 350
 351         if (amd_uncore_nb) {
 352                 uncore_nb = amd_uncore_alloc(cpu);
 353                 if (!uncore_nb)
 354                         goto fail;
 355                 uncore_nb->cpu = cpu;
 356                 uncore_nb->num_counters = num_counters_nb;
 357                 uncore_nb->rdpmc_base = RDPMC_BASE_NB;
 358                 uncore_nb->msr_base = MSR_F15H_NB_PERF_CTL;
 359                 uncore_nb->active_mask = &amd_nb_active_mask;
 360                 uncore_nb->pmu = &amd_nb_pmu;
 361                 uncore_nb->id = -1;
 362                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore_nb;
 363         }
 364
 365         if (amd_uncore_llc) {
 366                 uncore_llc = amd_uncore_alloc(cpu);
 367                 if (!uncore_llc)
 368                         goto fail;
 369                 uncore_llc->cpu = cpu;
 370                 uncore_llc->num_counters = num_counters_llc;
 371                 uncore_llc->rdpmc_base = RDPMC_BASE_LLC;
 372                 uncore_llc->msr_base = MSR_F16H_L2I_PERF_CTL;
 373                 uncore_llc->active_mask = &amd_llc_active_mask;
 374                 uncore_llc->pmu = &amd_llc_pmu;
 375                 uncore_llc->id = -1;
 376                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore_llc;
 377         }
 378
 379         return 0;
 380
 381 fail:
 382         if (amd_uncore_nb)
 383                 *per_cpu_ptr(amd_uncore_nb, cpu) = NULL;
 384         kfree(uncore_nb);
 385         return -ENOMEM;
 386 }
 387
 388 static struct amd_uncore *
 389 amd_uncore_find_online_sibling(struct amd_uncore *this,
 390                                struct amd_uncore * __percpu *uncores)
 391 {
 392         unsigned int cpu;
 393         struct amd_uncore *that;
 394
 395         for_each_online_cpu(cpu) {
 396                 that = *per_cpu_ptr(uncores, cpu);
 397
 398                 if (!that)
 399                         continue;
 400
 401                 if (this == that)
 402                         continue;
 403
 404                 if (this->id == that->id) {
 405                         hlist_add_head(&this->node, &uncore_unused_list);
 406                         this = that;
 407                         break;
 408                 }
 409         }
 410
 411         this->refcnt++;
 412         return this;
 413 }
 414
 415 static int amd_uncore_cpu_starting(unsigned int cpu)
 416 {
 417         unsigned int eax, ebx, ecx, edx;
 418         struct amd_uncore *uncore;
 419
 420         if (amd_uncore_nb) {
 421                 uncore = *per_cpu_ptr(amd_uncore_nb, cpu);
 422                 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
 423                 uncore->id = ecx & 0xff;
 424
 425                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_nb);
 426                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
 427         }
 428
 429         if (amd_uncore_llc) {
 430                 uncore = *per_cpu_ptr(amd_uncore_llc, cpu);
 431                 uncore->id = per_cpu(cpu_llc_id, cpu);
 432
 433                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_llc);
 434                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore;
 435         }
 436
 437         return 0;
 438 }
 439
 440 static void uncore_clean_online(void)
 441 {
 442         struct amd_uncore *uncore;
 443         struct hlist_node *n;
 444
 445         hlist_for_each_entry_safe(uncore, n, &uncore_unused_list, node) {
 446                 hlist_del(&uncore->node);
 447                 kfree(uncore);
 448         }
 449 }
 450
 451 static void uncore_online(unsigned int cpu,
 452                           struct amd_uncore * __percpu *uncores)
 453 {
 454         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
 455
 456         uncore_clean_online();
 457
 458         if (cpu == uncore->cpu)
 459                 cpumask_set_cpu(cpu, uncore->active_mask);
 460 }
 461
 462 static int amd_uncore_cpu_online(unsigned int cpu)
 463 {
 464         if (amd_uncore_nb)
 465                 uncore_online(cpu, amd_uncore_nb);
 466
 467         if (amd_uncore_llc)
 468                 uncore_online(cpu, amd_uncore_llc);
 469
 470         return 0;
 471 }
 472
 473 static void uncore_down_prepare(unsigned int cpu,
 474                                 struct amd_uncore * __percpu *uncores)
 475 {
 476         unsigned int i;
 477         struct amd_uncore *this = *per_cpu_ptr(uncores, cpu);
 478
 479         if (this->cpu != cpu)
 480                 return;
 481
 482         /* this cpu is going down, migrate to a shared sibling if possible */
 483         for_each_online_cpu(i) {
 484                 struct amd_uncore *that = *per_cpu_ptr(uncores, i);
 485
 486                 if (cpu == i)
 487                         continue;
 488
 489                 if (this == that) {
 490                         perf_pmu_migrate_context(this->pmu, cpu, i);
 491                         cpumask_clear_cpu(cpu, that->active_mask);
 492                         cpumask_set_cpu(i, that->active_mask);
 493                         that->cpu = i;
 494                         break;
 495                 }
 496         }
 497 }
 498
 499 static int amd_uncore_cpu_down_prepare(unsigned int cpu)
 500 {
 501         if (amd_uncore_nb)
 502                 uncore_down_prepare(cpu, amd_uncore_nb);
 503
 504         if (amd_uncore_llc)
 505                 uncore_down_prepare(cpu, amd_uncore_llc);
 506
 507         return 0;
 508 }
 509
 510 static void uncore_dead(unsigned int cpu, struct amd_uncore * __percpu *uncores)
 511 {
 512         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
 513
 514         if (cpu == uncore->cpu)
 515                 cpumask_clear_cpu(cpu, uncore->active_mask);
 516
 517         if (!--uncore->refcnt)
 518                 kfree(uncore);
 519         *per_cpu_ptr(uncores, cpu) = NULL;
 520 }
 521
 522 static int amd_uncore_cpu_dead(unsigned int cpu)
 523 {
 524         if (amd_uncore_nb)
 525                 uncore_dead(cpu, amd_uncore_nb);
 526
 527         if (amd_uncore_llc)
 528                 uncore_dead(cpu, amd_uncore_llc);
 529
 530         return 0;
 531 }
 532
 533 static int __init amd_uncore_init(void)
 534 {
 535         struct attribute **df_attr = amd_uncore_df_format_attr;
 536         struct attribute **l3_attr = amd_uncore_l3_format_attr;
 537         int ret = -ENODEV;
 538
 539         if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD &&
 540             boot_cpu_data.x86_vendor != X86_VENDOR_HYGON)
 541                 return -ENODEV;
 542
 543         if (!boot_cpu_has(X86_FEATURE_TOPOEXT))
 544                 return -ENODEV;
 545
 546         num_counters_nb = NUM_COUNTERS_NB;
 547         num_counters_llc = NUM_COUNTERS_L2;
 548         if (boot_cpu_data.x86 >= 0x17) {
 549                 /*
 550                  * For F17h and above, the Northbridge counters are
 551                  * repurposed as Data Fabric counters. Also, L3
 552                  * counters are supported too. The PMUs are exported
 553                  * based on family as either L2 or L3 and NB or DF.
 554                  */
 555                 num_counters_llc          = NUM_COUNTERS_L3;
 556                 amd_nb_pmu.name           = "amd_df";
 557                 amd_llc_pmu.name          = "amd_l3";
 558                 l3_mask                   = true;
 559         }
 560
 561         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB)) {
 562                 if (boot_cpu_data.x86 >= 0x17)
 563                         *df_attr = &format_attr_event14.attr;
 564
 565                 amd_uncore_nb = alloc_percpu(struct amd_uncore *);
 566                 if (!amd_uncore_nb) {
 567                         ret = -ENOMEM;
 568                         goto fail_nb;
 569                 }
 570                 ret = perf_pmu_register(&amd_nb_pmu, amd_nb_pmu.name, -1);
 571                 if (ret)
 572                         goto fail_nb;
 573
 574                 pr_info("%s NB counters detected\n",
 575                         boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?
 576                                 "HYGON" : "AMD");
 577                 ret = 0;
 578         }
 579
 580         if (boot_cpu_has(X86_FEATURE_PERFCTR_LLC)) {
 581                 if (boot_cpu_data.x86 >= 0x17)
 582                         *l3_attr = &format_attr_event8.attr;
 583
 584                 amd_uncore_llc = alloc_percpu(struct amd_uncore *);
 585                 if (!amd_uncore_llc) {
 586                         ret = -ENOMEM;
 587                         goto fail_llc;
 588                 }
 589                 ret = perf_pmu_register(&amd_llc_pmu, amd_llc_pmu.name, -1);
 590                 if (ret)
 591                         goto fail_llc;
 592
 593                 pr_info("%s LLC counters detected\n",
 594                         boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?
 595                                 "HYGON" : "AMD");
 596                 ret = 0;
 597         }
 598
 599         /*
 600          * Install callbacks. Core will call them for each online cpu.
 601          */
 602         if (cpuhp_setup_state(CPUHP_PERF_X86_AMD_UNCORE_PREP,
 603                               "perf/x86/amd/uncore:prepare",
 604                               amd_uncore_cpu_up_prepare, amd_uncore_cpu_dead))
 605                 goto fail_llc;
 606
 607         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING,
 608                               "perf/x86/amd/uncore:starting",
 609                               amd_uncore_cpu_starting, NULL))
 610                 goto fail_prep;
 611         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_ONLINE,
 612                               "perf/x86/amd/uncore:online",
 613                               amd_uncore_cpu_online,
 614                               amd_uncore_cpu_down_prepare))
 615                 goto fail_start;
 616         return 0;
 617
 618 fail_start:
 619         cpuhp_remove_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING);
 620 fail_prep:
 621         cpuhp_remove_state(CPUHP_PERF_X86_AMD_UNCORE_PREP);
 622 fail_llc:
 623         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB))
 624                 perf_pmu_unregister(&amd_nb_pmu);
 625         if (amd_uncore_llc)
 626                 free_percpu(amd_uncore_llc);
 627 fail_nb:
 628         if (amd_uncore_nb)
 629                 free_percpu(amd_uncore_nb);
 630
 631         return ret;
 632 }
 633 device_initcall(amd_uncore_init);