arch/x86/events/amd/uncore.c

   1 /*
   2  * Copyright (C) 2013 Advanced Micro Devices, Inc.
   3  *
   4  * Author: Jacob Shin <jacob.shin@amd.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/perf_event.h>
  12 #include <linux/percpu.h>
  13 #include <linux/types.h>
  14 #include <linux/slab.h>
  15 #include <linux/init.h>
  16 #include <linux/cpu.h>
  17 #include <linux/cpumask.h>
  18
  19 #include <asm/cpufeature.h>
  20 #include <asm/perf_event.h>
  21 #include <asm/msr.h>
  22 #include <asm/smp.h>
  23
  24 #define NUM_COUNTERS_NB         4
  25 #define NUM_COUNTERS_L2         4
  26 #define MAX_COUNTERS            NUM_COUNTERS_NB
  27
  28 #define RDPMC_BASE_NB           6
  29 #define RDPMC_BASE_LLC          10
  30
  31 #define COUNTER_SHIFT           16
  32
  33 static HLIST_HEAD(uncore_unused_list);
  34
  35 struct amd_uncore {
  36         int id;
  37         int refcnt;
  38         int cpu;
  39         int num_counters;
  40         int rdpmc_base;
  41         u32 msr_base;
  42         cpumask_t *active_mask;
  43         struct pmu *pmu;
  44         struct perf_event *events[MAX_COUNTERS];
  45         struct hlist_node node;
  46 };
  47
  48 static struct amd_uncore * __percpu *amd_uncore_nb;
  49 static struct amd_uncore * __percpu *amd_uncore_llc;
  50
  51 static struct pmu amd_nb_pmu;
  52 static struct pmu amd_llc_pmu;
  53
  54 static cpumask_t amd_nb_active_mask;
  55 static cpumask_t amd_llc_active_mask;
  56
  57 static bool is_nb_event(struct perf_event *event)
  58 {
  59         return event->pmu->type == amd_nb_pmu.type;
  60 }
  61
  62 static bool is_llc_event(struct perf_event *event)
  63 {
  64         return event->pmu->type == amd_llc_pmu.type;
  65 }
  66
  67 static struct amd_uncore *event_to_amd_uncore(struct perf_event *event)
  68 {
  69         if (is_nb_event(event) && amd_uncore_nb)
  70                 return *per_cpu_ptr(amd_uncore_nb, event->cpu);
  71         else if (is_llc_event(event) && amd_uncore_llc)
  72                 return *per_cpu_ptr(amd_uncore_llc, event->cpu);
  73
  74         return NULL;
  75 }
  76
  77 static void amd_uncore_read(struct perf_event *event)
  78 {
  79         struct hw_perf_event *hwc = &event->hw;
  80         u64 prev, new;
  81         s64 delta;
  82
  83         /*
  84          * since we do not enable counter overflow interrupts,
  85          * we do not have to worry about prev_count changing on us
  86          */
  87
  88         prev = local64_read(&hwc->prev_count);
  89         rdpmcl(hwc->event_base_rdpmc, new);
  90         local64_set(&hwc->prev_count, new);
  91         delta = (new << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
  92         delta >>= COUNTER_SHIFT;
  93         local64_add(delta, &event->count);
  94 }
  95
  96 static void amd_uncore_start(struct perf_event *event, int flags)
  97 {
  98         struct hw_perf_event *hwc = &event->hw;
  99
 100         if (flags & PERF_EF_RELOAD)
 101                 wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 102
 103         hwc->state = 0;
 104         wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
 105         perf_event_update_userpage(event);
 106 }
 107
 108 static void amd_uncore_stop(struct perf_event *event, int flags)
 109 {
 110         struct hw_perf_event *hwc = &event->hw;
 111
 112         wrmsrl(hwc->config_base, hwc->config);
 113         hwc->state |= PERF_HES_STOPPED;
 114
 115         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
 116                 amd_uncore_read(event);
 117                 hwc->state |= PERF_HES_UPTODATE;
 118         }
 119 }
 120
 121 static int amd_uncore_add(struct perf_event *event, int flags)
 122 {
 123         int i;
 124         struct amd_uncore *uncore = event_to_amd_uncore(event);
 125         struct hw_perf_event *hwc = &event->hw;
 126
 127         /* are we already assigned? */
 128         if (hwc->idx != -1 && uncore->events[hwc->idx] == event)
 129                 goto out;
 130
 131         for (i = 0; i < uncore->num_counters; i++) {
 132                 if (uncore->events[i] == event) {
 133                         hwc->idx = i;
 134                         goto out;
 135                 }
 136         }
 137
 138         /* if not, take the first available counter */
 139         hwc->idx = -1;
 140         for (i = 0; i < uncore->num_counters; i++) {
 141                 if (cmpxchg(&uncore->events[i], NULL, event) == NULL) {
 142                         hwc->idx = i;
 143                         break;
 144                 }
 145         }
 146
 147 out:
 148         if (hwc->idx == -1)
 149                 return -EBUSY;
 150
 151         hwc->config_base = uncore->msr_base + (2 * hwc->idx);
 152         hwc->event_base = uncore->msr_base + 1 + (2 * hwc->idx);
 153         hwc->event_base_rdpmc = uncore->rdpmc_base + hwc->idx;
 154         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 155
 156         if (flags & PERF_EF_START)
 157                 amd_uncore_start(event, PERF_EF_RELOAD);
 158
 159         return 0;
 160 }
 161
 162 static void amd_uncore_del(struct perf_event *event, int flags)
 163 {
 164         int i;
 165         struct amd_uncore *uncore = event_to_amd_uncore(event);
 166         struct hw_perf_event *hwc = &event->hw;
 167
 168         amd_uncore_stop(event, PERF_EF_UPDATE);
 169
 170         for (i = 0; i < uncore->num_counters; i++) {
 171                 if (cmpxchg(&uncore->events[i], event, NULL) == event)
 172                         break;
 173         }
 174
 175         hwc->idx = -1;
 176 }
 177
 178 static int amd_uncore_event_init(struct perf_event *event)
 179 {
 180         struct amd_uncore *uncore;
 181         struct hw_perf_event *hwc = &event->hw;
 182
 183         if (event->attr.type != event->pmu->type)
 184                 return -ENOENT;
 185
 186         /*
 187          * NB and Last level cache counters (MSRs) are shared across all cores
 188          * that share the same NB / Last level cache. Interrupts can be directed
 189          * to a single target core, however, event counts generated by processes
 190          * running on other cores cannot be masked out. So we do not support
 191          * sampling and per-thread events.
 192          */
 193         if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
 194                 return -EINVAL;
 195
 196         /* NB and Last level cache counters do not have usr/os/guest/host bits */
 197         if (event->attr.exclude_user || event->attr.exclude_kernel ||
 198             event->attr.exclude_host || event->attr.exclude_guest)
 199                 return -EINVAL;
 200
 201         /* and we do not enable counter overflow interrupts */
 202         hwc->config = event->attr.config & AMD64_RAW_EVENT_MASK_NB;
 203         hwc->idx = -1;
 204
 205         if (event->cpu < 0)
 206                 return -EINVAL;
 207
 208         uncore = event_to_amd_uncore(event);
 209         if (!uncore)
 210                 return -ENODEV;
 211
 212         /*
 213          * since request can come in to any of the shared cores, we will remap
 214          * to a single common cpu.
 215          */
 216         event->cpu = uncore->cpu;
 217
 218         return 0;
 219 }
 220
 221 static ssize_t amd_uncore_attr_show_cpumask(struct device *dev,
 222                                             struct device_attribute *attr,
 223                                             char *buf)
 224 {
 225         cpumask_t *active_mask;
 226         struct pmu *pmu = dev_get_drvdata(dev);
 227
 228         if (pmu->type == amd_nb_pmu.type)
 229                 active_mask = &amd_nb_active_mask;
 230         else if (pmu->type == amd_llc_pmu.type)
 231                 active_mask = &amd_llc_active_mask;
 232         else
 233                 return 0;
 234
 235         return cpumap_print_to_pagebuf(true, buf, active_mask);
 236 }
 237 static DEVICE_ATTR(cpumask, S_IRUGO, amd_uncore_attr_show_cpumask, NULL);
 238
 239 static struct attribute *amd_uncore_attrs[] = {
 240         &dev_attr_cpumask.attr,
 241         NULL,
 242 };
 243
 244 static struct attribute_group amd_uncore_attr_group = {
 245         .attrs = amd_uncore_attrs,
 246 };
 247
 248 PMU_FORMAT_ATTR(event, "config:0-7,32-35");
 249 PMU_FORMAT_ATTR(umask, "config:8-15");
 250
 251 static struct attribute *amd_uncore_format_attr[] = {
 252         &format_attr_event.attr,
 253         &format_attr_umask.attr,
 254         NULL,
 255 };
 256
 257 static struct attribute_group amd_uncore_format_group = {
 258         .name = "format",
 259         .attrs = amd_uncore_format_attr,
 260 };
 261
 262 static const struct attribute_group *amd_uncore_attr_groups[] = {
 263         &amd_uncore_attr_group,
 264         &amd_uncore_format_group,
 265         NULL,
 266 };
 267
 268 static struct pmu amd_nb_pmu = {
 269         .task_ctx_nr    = perf_invalid_context,
 270         .attr_groups    = amd_uncore_attr_groups,
 271         .name           = "amd_nb",
 272         .event_init     = amd_uncore_event_init,
 273         .add            = amd_uncore_add,
 274         .del            = amd_uncore_del,
 275         .start          = amd_uncore_start,
 276         .stop           = amd_uncore_stop,
 277         .read           = amd_uncore_read,
 278 };
 279
 280 static struct pmu amd_llc_pmu = {
 281         .task_ctx_nr    = perf_invalid_context,
 282         .attr_groups    = amd_uncore_attr_groups,
 283         .name           = "amd_l2",
 284         .event_init     = amd_uncore_event_init,
 285         .add            = amd_uncore_add,
 286         .del            = amd_uncore_del,
 287         .start          = amd_uncore_start,
 288         .stop           = amd_uncore_stop,
 289         .read           = amd_uncore_read,
 290 };
 291
 292 static struct amd_uncore *amd_uncore_alloc(unsigned int cpu)
 293 {
 294         return kzalloc_node(sizeof(struct amd_uncore), GFP_KERNEL,
 295                         cpu_to_node(cpu));
 296 }
 297
 298 static int amd_uncore_cpu_up_prepare(unsigned int cpu)
 299 {
 300         struct amd_uncore *uncore_nb = NULL, *uncore_llc;
 301
 302         if (amd_uncore_nb) {
 303                 uncore_nb = amd_uncore_alloc(cpu);
 304                 if (!uncore_nb)
 305                         goto fail;
 306                 uncore_nb->cpu = cpu;
 307                 uncore_nb->num_counters = NUM_COUNTERS_NB;
 308                 uncore_nb->rdpmc_base = RDPMC_BASE_NB;
 309                 uncore_nb->msr_base = MSR_F15H_NB_PERF_CTL;
 310                 uncore_nb->active_mask = &amd_nb_active_mask;
 311                 uncore_nb->pmu = &amd_nb_pmu;
 312                 uncore_nb->id = -1;
 313                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore_nb;
 314         }
 315
 316         if (amd_uncore_llc) {
 317                 uncore_llc = amd_uncore_alloc(cpu);
 318                 if (!uncore_llc)
 319                         goto fail;
 320                 uncore_llc->cpu = cpu;
 321                 uncore_llc->num_counters = NUM_COUNTERS_L2;
 322                 uncore_llc->rdpmc_base = RDPMC_BASE_LLC;
 323                 uncore_llc->msr_base = MSR_F16H_L2I_PERF_CTL;
 324                 uncore_llc->active_mask = &amd_llc_active_mask;
 325                 uncore_llc->pmu = &amd_llc_pmu;
 326                 uncore_llc->id = -1;
 327                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore_llc;
 328         }
 329
 330         return 0;
 331
 332 fail:
 333         if (amd_uncore_nb)
 334                 *per_cpu_ptr(amd_uncore_nb, cpu) = NULL;
 335         kfree(uncore_nb);
 336         return -ENOMEM;
 337 }
 338
 339 static struct amd_uncore *
 340 amd_uncore_find_online_sibling(struct amd_uncore *this,
 341                                struct amd_uncore * __percpu *uncores)
 342 {
 343         unsigned int cpu;
 344         struct amd_uncore *that;
 345
 346         for_each_online_cpu(cpu) {
 347                 that = *per_cpu_ptr(uncores, cpu);
 348
 349                 if (!that)
 350                         continue;
 351
 352                 if (this == that)
 353                         continue;
 354
 355                 if (this->id == that->id) {
 356                         hlist_add_head(&this->node, &uncore_unused_list);
 357                         this = that;
 358                         break;
 359                 }
 360         }
 361
 362         this->refcnt++;
 363         return this;
 364 }
 365
 366 static int amd_uncore_cpu_starting(unsigned int cpu)
 367 {
 368         unsigned int eax, ebx, ecx, edx;
 369         struct amd_uncore *uncore;
 370
 371         if (amd_uncore_nb) {
 372                 uncore = *per_cpu_ptr(amd_uncore_nb, cpu);
 373                 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
 374                 uncore->id = ecx & 0xff;
 375
 376                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_nb);
 377                 *per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
 378         }
 379
 380         if (amd_uncore_llc) {
 381                 uncore = *per_cpu_ptr(amd_uncore_llc, cpu);
 382                 uncore->id = per_cpu(cpu_llc_id, cpu);
 383
 384                 uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_llc);
 385                 *per_cpu_ptr(amd_uncore_llc, cpu) = uncore;
 386         }
 387
 388         return 0;
 389 }
 390
 391 static void uncore_clean_online(void)
 392 {
 393         struct amd_uncore *uncore;
 394         struct hlist_node *n;
 395
 396         hlist_for_each_entry_safe(uncore, n, &uncore_unused_list, node) {
 397                 hlist_del(&uncore->node);
 398                 kfree(uncore);
 399         }
 400 }
 401
 402 static void uncore_online(unsigned int cpu,
 403                           struct amd_uncore * __percpu *uncores)
 404 {
 405         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
 406
 407         uncore_clean_online();
 408
 409         if (cpu == uncore->cpu)
 410                 cpumask_set_cpu(cpu, uncore->active_mask);
 411 }
 412
 413 static int amd_uncore_cpu_online(unsigned int cpu)
 414 {
 415         if (amd_uncore_nb)
 416                 uncore_online(cpu, amd_uncore_nb);
 417
 418         if (amd_uncore_llc)
 419                 uncore_online(cpu, amd_uncore_llc);
 420
 421         return 0;
 422 }
 423
 424 static void uncore_down_prepare(unsigned int cpu,
 425                                 struct amd_uncore * __percpu *uncores)
 426 {
 427         unsigned int i;
 428         struct amd_uncore *this = *per_cpu_ptr(uncores, cpu);
 429
 430         if (this->cpu != cpu)
 431                 return;
 432
 433         /* this cpu is going down, migrate to a shared sibling if possible */
 434         for_each_online_cpu(i) {
 435                 struct amd_uncore *that = *per_cpu_ptr(uncores, i);
 436
 437                 if (cpu == i)
 438                         continue;
 439
 440                 if (this == that) {
 441                         perf_pmu_migrate_context(this->pmu, cpu, i);
 442                         cpumask_clear_cpu(cpu, that->active_mask);
 443                         cpumask_set_cpu(i, that->active_mask);
 444                         that->cpu = i;
 445                         break;
 446                 }
 447         }
 448 }
 449
 450 static int amd_uncore_cpu_down_prepare(unsigned int cpu)
 451 {
 452         if (amd_uncore_nb)
 453                 uncore_down_prepare(cpu, amd_uncore_nb);
 454
 455         if (amd_uncore_llc)
 456                 uncore_down_prepare(cpu, amd_uncore_llc);
 457
 458         return 0;
 459 }
 460
 461 static void uncore_dead(unsigned int cpu, struct amd_uncore * __percpu *uncores)
 462 {
 463         struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
 464
 465         if (cpu == uncore->cpu)
 466                 cpumask_clear_cpu(cpu, uncore->active_mask);
 467
 468         if (!--uncore->refcnt)
 469                 kfree(uncore);
 470         *per_cpu_ptr(uncores, cpu) = NULL;
 471 }
 472
 473 static int amd_uncore_cpu_dead(unsigned int cpu)
 474 {
 475         if (amd_uncore_nb)
 476                 uncore_dead(cpu, amd_uncore_nb);
 477
 478         if (amd_uncore_llc)
 479                 uncore_dead(cpu, amd_uncore_llc);
 480
 481         return 0;
 482 }
 483
 484 static int __init amd_uncore_init(void)
 485 {
 486         int ret = -ENODEV;
 487
 488         if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
 489                 goto fail_nodev;
 490
 491         if (!boot_cpu_has(X86_FEATURE_TOPOEXT))
 492                 goto fail_nodev;
 493
 494         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB)) {
 495                 amd_uncore_nb = alloc_percpu(struct amd_uncore *);
 496                 if (!amd_uncore_nb) {
 497                         ret = -ENOMEM;
 498                         goto fail_nb;
 499                 }
 500                 ret = perf_pmu_register(&amd_nb_pmu, amd_nb_pmu.name, -1);
 501                 if (ret)
 502                         goto fail_nb;
 503
 504                 pr_info("perf: AMD NB counters detected\n");
 505                 ret = 0;
 506         }
 507
 508         if (boot_cpu_has(X86_FEATURE_PERFCTR_L2)) {
 509                 amd_uncore_llc = alloc_percpu(struct amd_uncore *);
 510                 if (!amd_uncore_llc) {
 511                         ret = -ENOMEM;
 512                         goto fail_llc;
 513                 }
 514                 ret = perf_pmu_register(&amd_llc_pmu, amd_llc_pmu.name, -1);
 515                 if (ret)
 516                         goto fail_llc;
 517
 518                 pr_info("perf: AMD LLC counters detected\n");
 519                 ret = 0;
 520         }
 521
 522         /*
 523          * Install callbacks. Core will call them for each online cpu.
 524          */
 525         if (cpuhp_setup_state(CPUHP_PERF_X86_AMD_UNCORE_PREP,
 526                               "PERF_X86_AMD_UNCORE_PREP",
 527                               amd_uncore_cpu_up_prepare, amd_uncore_cpu_dead))
 528                 goto fail_llc;
 529
 530         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING,
 531                               "AP_PERF_X86_AMD_UNCORE_STARTING",
 532                               amd_uncore_cpu_starting, NULL))
 533                 goto fail_prep;
 534         if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_ONLINE,
 535                               "AP_PERF_X86_AMD_UNCORE_ONLINE",
 536                               amd_uncore_cpu_online,
 537                               amd_uncore_cpu_down_prepare))
 538                 goto fail_start;
 539         return 0;
 540
 541 fail_start:
 542         cpuhp_remove_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING);
 543 fail_prep:
 544         cpuhp_remove_state(CPUHP_PERF_X86_AMD_UNCORE_PREP);
 545 fail_llc:
 546         if (boot_cpu_has(X86_FEATURE_PERFCTR_NB))
 547                 perf_pmu_unregister(&amd_nb_pmu);
 548         if (amd_uncore_llc)
 549                 free_percpu(amd_uncore_llc);
 550 fail_nb:
 551         if (amd_uncore_nb)
 552                 free_percpu(amd_uncore_nb);
 553
 554 fail_nodev:
 555         return ret;
 556 }
 557 device_initcall(amd_uncore_init);