drivers/perf/arm_pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #undef DEBUG
   3
   4 /*
   5  * ARM performance counter support.
   6  *
   7  * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
   8  * Copyright (C) 2010 ARM Ltd., Will Deacon <will.deacon@arm.com>
   9  *
  10  * This code is based on the sparc64 perf event code, which is in turn based
  11  * on the x86 code.
  12  */
  13 #define pr_fmt(fmt) "hw perfevents: " fmt
  14
  15 #include <linux/bitmap.h>
  16 #include <linux/cpumask.h>
  17 #include <linux/cpu_pm.h>
  18 #include <linux/export.h>
  19 #include <linux/kernel.h>
  20 #include <linux/perf/arm_pmu.h>
  21 #include <linux/slab.h>
  22 #include <linux/sched/clock.h>
  23 #include <linux/spinlock.h>
  24 #include <linux/irq.h>
  25 #include <linux/irqdesc.h>
  26
  27 #include <asm/irq_regs.h>
  28
  29 static int armpmu_count_irq_users(const int irq);
  30
  31 struct pmu_irq_ops {
  32         void (*enable_pmuirq)(unsigned int irq);
  33         void (*disable_pmuirq)(unsigned int irq);
  34         void (*free_pmuirq)(unsigned int irq, int cpu, void __percpu *devid);
  35 };
  36
  37 static void armpmu_free_pmuirq(unsigned int irq, int cpu, void __percpu *devid)
  38 {
  39         free_irq(irq, per_cpu_ptr(devid, cpu));
  40 }
  41
  42 static const struct pmu_irq_ops pmuirq_ops = {
  43         .enable_pmuirq = enable_irq,
  44         .disable_pmuirq = disable_irq_nosync,
  45         .free_pmuirq = armpmu_free_pmuirq
  46 };
  47
  48 static void armpmu_free_pmunmi(unsigned int irq, int cpu, void __percpu *devid)
  49 {
  50         free_nmi(irq, per_cpu_ptr(devid, cpu));
  51 }
  52
  53 static const struct pmu_irq_ops pmunmi_ops = {
  54         .enable_pmuirq = enable_nmi,
  55         .disable_pmuirq = disable_nmi_nosync,
  56         .free_pmuirq = armpmu_free_pmunmi
  57 };
  58
  59 static void armpmu_enable_percpu_pmuirq(unsigned int irq)
  60 {
  61         enable_percpu_irq(irq, IRQ_TYPE_NONE);
  62 }
  63
  64 static void armpmu_free_percpu_pmuirq(unsigned int irq, int cpu,
  65                                    void __percpu *devid)
  66 {
  67         if (armpmu_count_irq_users(irq) == 1)
  68                 free_percpu_irq(irq, devid);
  69 }
  70
  71 static const struct pmu_irq_ops percpu_pmuirq_ops = {
  72         .enable_pmuirq = armpmu_enable_percpu_pmuirq,
  73         .disable_pmuirq = disable_percpu_irq,
  74         .free_pmuirq = armpmu_free_percpu_pmuirq
  75 };
  76
  77 static void armpmu_enable_percpu_pmunmi(unsigned int irq)
  78 {
  79         if (!prepare_percpu_nmi(irq))
  80                 enable_percpu_nmi(irq, IRQ_TYPE_NONE);
  81 }
  82
  83 static void armpmu_disable_percpu_pmunmi(unsigned int irq)
  84 {
  85         disable_percpu_nmi(irq);
  86         teardown_percpu_nmi(irq);
  87 }
  88
  89 static void armpmu_free_percpu_pmunmi(unsigned int irq, int cpu,
  90                                       void __percpu *devid)
  91 {
  92         if (armpmu_count_irq_users(irq) == 1)
  93                 free_percpu_nmi(irq, devid);
  94 }
  95
  96 static const struct pmu_irq_ops percpu_pmunmi_ops = {
  97         .enable_pmuirq = armpmu_enable_percpu_pmunmi,
  98         .disable_pmuirq = armpmu_disable_percpu_pmunmi,
  99         .free_pmuirq = armpmu_free_percpu_pmunmi
 100 };
 101
 102 static DEFINE_PER_CPU(struct arm_pmu *, cpu_armpmu);
 103 static DEFINE_PER_CPU(int, cpu_irq);
 104 static DEFINE_PER_CPU(const struct pmu_irq_ops *, cpu_irq_ops);
 105
 106 static bool has_nmi;
 107
 108 static inline u64 arm_pmu_event_max_period(struct perf_event *event)
 109 {
 110         if (event->hw.flags & ARMPMU_EVT_64BIT)
 111                 return GENMASK_ULL(63, 0);
 112         else
 113                 return GENMASK_ULL(31, 0);
 114 }
 115
 116 static int
 117 armpmu_map_cache_event(const unsigned (*cache_map)
 118                                       [PERF_COUNT_HW_CACHE_MAX]
 119                                       [PERF_COUNT_HW_CACHE_OP_MAX]
 120                                       [PERF_COUNT_HW_CACHE_RESULT_MAX],
 121                        u64 config)
 122 {
 123         unsigned int cache_type, cache_op, cache_result, ret;
 124
 125         cache_type = (config >>  0) & 0xff;
 126         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 127                 return -EINVAL;
 128
 129         cache_op = (config >>  8) & 0xff;
 130         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 131                 return -EINVAL;
 132
 133         cache_result = (config >> 16) & 0xff;
 134         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 135                 return -EINVAL;
 136
 137         if (!cache_map)
 138                 return -ENOENT;
 139
 140         ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
 141
 142         if (ret == CACHE_OP_UNSUPPORTED)
 143                 return -ENOENT;
 144
 145         return ret;
 146 }
 147
 148 static int
 149 armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
 150 {
 151         int mapping;
 152
 153         if (config >= PERF_COUNT_HW_MAX)
 154                 return -EINVAL;
 155
 156         if (!event_map)
 157                 return -ENOENT;
 158
 159         mapping = (*event_map)[config];
 160         return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
 161 }
 162
 163 static int
 164 armpmu_map_raw_event(u32 raw_event_mask, u64 config)
 165 {
 166         return (int)(config & raw_event_mask);
 167 }
 168
 169 int
 170 armpmu_map_event(struct perf_event *event,
 171                  const unsigned (*event_map)[PERF_COUNT_HW_MAX],
 172                  const unsigned (*cache_map)
 173                                 [PERF_COUNT_HW_CACHE_MAX]
 174                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 175                                 [PERF_COUNT_HW_CACHE_RESULT_MAX],
 176                  u32 raw_event_mask)
 177 {
 178         u64 config = event->attr.config;
 179         int type = event->attr.type;
 180
 181         if (type == event->pmu->type)
 182                 return armpmu_map_raw_event(raw_event_mask, config);
 183
 184         switch (type) {
 185         case PERF_TYPE_HARDWARE:
 186                 return armpmu_map_hw_event(event_map, config);
 187         case PERF_TYPE_HW_CACHE:
 188                 return armpmu_map_cache_event(cache_map, config);
 189         case PERF_TYPE_RAW:
 190                 return armpmu_map_raw_event(raw_event_mask, config);
 191         }
 192
 193         return -ENOENT;
 194 }
 195
 196 int armpmu_event_set_period(struct perf_event *event)
 197 {
 198         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 199         struct hw_perf_event *hwc = &event->hw;
 200         s64 left = local64_read(&hwc->period_left);
 201         s64 period = hwc->sample_period;
 202         u64 max_period;
 203         int ret = 0;
 204
 205         max_period = arm_pmu_event_max_period(event);
 206         if (unlikely(left <= -period)) {
 207                 left = period;
 208                 local64_set(&hwc->period_left, left);
 209                 hwc->last_period = period;
 210                 ret = 1;
 211         }
 212
 213         if (unlikely(left <= 0)) {
 214                 left += period;
 215                 local64_set(&hwc->period_left, left);
 216                 hwc->last_period = period;
 217                 ret = 1;
 218         }
 219
 220         /*
 221          * Limit the maximum period to prevent the counter value
 222          * from overtaking the one we are about to program. In
 223          * effect we are reducing max_period to account for
 224          * interrupt latency (and we are being very conservative).
 225          */
 226         if (left > (max_period >> 1))
 227                 left = (max_period >> 1);
 228
 229         local64_set(&hwc->prev_count, (u64)-left);
 230
 231         armpmu->write_counter(event, (u64)(-left) & max_period);
 232
 233         perf_event_update_userpage(event);
 234
 235         return ret;
 236 }
 237
 238 u64 armpmu_event_update(struct perf_event *event)
 239 {
 240         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 241         struct hw_perf_event *hwc = &event->hw;
 242         u64 delta, prev_raw_count, new_raw_count;
 243         u64 max_period = arm_pmu_event_max_period(event);
 244
 245 again:
 246         prev_raw_count = local64_read(&hwc->prev_count);
 247         new_raw_count = armpmu->read_counter(event);
 248
 249         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 250                              new_raw_count) != prev_raw_count)
 251                 goto again;
 252
 253         delta = (new_raw_count - prev_raw_count) & max_period;
 254
 255         local64_add(delta, &event->count);
 256         local64_sub(delta, &hwc->period_left);
 257
 258         return new_raw_count;
 259 }
 260
 261 static void
 262 armpmu_read(struct perf_event *event)
 263 {
 264         armpmu_event_update(event);
 265 }
 266
 267 static void
 268 armpmu_stop(struct perf_event *event, int flags)
 269 {
 270         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 271         struct hw_perf_event *hwc = &event->hw;
 272
 273         /*
 274          * ARM pmu always has to update the counter, so ignore
 275          * PERF_EF_UPDATE, see comments in armpmu_start().
 276          */
 277         if (!(hwc->state & PERF_HES_STOPPED)) {
 278                 armpmu->disable(event);
 279                 armpmu_event_update(event);
 280                 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
 281         }
 282 }
 283
 284 static void armpmu_start(struct perf_event *event, int flags)
 285 {
 286         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 287         struct hw_perf_event *hwc = &event->hw;
 288
 289         /*
 290          * ARM pmu always has to reprogram the period, so ignore
 291          * PERF_EF_RELOAD, see the comment below.
 292          */
 293         if (flags & PERF_EF_RELOAD)
 294                 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 295
 296         hwc->state = 0;
 297         /*
 298          * Set the period again. Some counters can't be stopped, so when we
 299          * were stopped we simply disabled the IRQ source and the counter
 300          * may have been left counting. If we don't do this step then we may
 301          * get an interrupt too soon or *way* too late if the overflow has
 302          * happened since disabling.
 303          */
 304         armpmu_event_set_period(event);
 305         armpmu->enable(event);
 306 }
 307
 308 static void
 309 armpmu_del(struct perf_event *event, int flags)
 310 {
 311         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 312         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 313         struct hw_perf_event *hwc = &event->hw;
 314         int idx = hwc->idx;
 315
 316         armpmu_stop(event, PERF_EF_UPDATE);
 317         hw_events->events[idx] = NULL;
 318         armpmu->clear_event_idx(hw_events, event);
 319         perf_event_update_userpage(event);
 320         /* Clear the allocated counter */
 321         hwc->idx = -1;
 322 }
 323
 324 static int
 325 armpmu_add(struct perf_event *event, int flags)
 326 {
 327         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 328         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 329         struct hw_perf_event *hwc = &event->hw;
 330         int idx;
 331
 332         /* An event following a process won't be stopped earlier */
 333         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 334                 return -ENOENT;
 335
 336         /* If we don't have a space for the counter then finish early. */
 337         idx = armpmu->get_event_idx(hw_events, event);
 338         if (idx < 0)
 339                 return idx;
 340
 341         /*
 342          * If there is an event in the counter we are going to use then make
 343          * sure it is disabled.
 344          */
 345         event->hw.idx = idx;
 346         armpmu->disable(event);
 347         hw_events->events[idx] = event;
 348
 349         hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
 350         if (flags & PERF_EF_START)
 351                 armpmu_start(event, PERF_EF_RELOAD);
 352
 353         /* Propagate our changes to the userspace mapping. */
 354         perf_event_update_userpage(event);
 355
 356         return 0;
 357 }
 358
 359 static int
 360 validate_event(struct pmu *pmu, struct pmu_hw_events *hw_events,
 361                                struct perf_event *event)
 362 {
 363         struct arm_pmu *armpmu;
 364
 365         if (is_software_event(event))
 366                 return 1;
 367
 368         /*
 369          * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
 370          * core perf code won't check that the pmu->ctx == leader->ctx
 371          * until after pmu->event_init(event).
 372          */
 373         if (event->pmu != pmu)
 374                 return 0;
 375
 376         if (event->state < PERF_EVENT_STATE_OFF)
 377                 return 1;
 378
 379         if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
 380                 return 1;
 381
 382         armpmu = to_arm_pmu(event->pmu);
 383         return armpmu->get_event_idx(hw_events, event) >= 0;
 384 }
 385
 386 static int
 387 validate_group(struct perf_event *event)
 388 {
 389         struct perf_event *sibling, *leader = event->group_leader;
 390         struct pmu_hw_events fake_pmu;
 391
 392         /*
 393          * Initialise the fake PMU. We only need to populate the
 394          * used_mask for the purposes of validation.
 395          */
 396         memset(&fake_pmu.used_mask, 0, sizeof(fake_pmu.used_mask));
 397
 398         if (!validate_event(event->pmu, &fake_pmu, leader))
 399                 return -EINVAL;
 400
 401         if (event == leader)
 402                 return 0;
 403
 404         for_each_sibling_event(sibling, leader) {
 405                 if (!validate_event(event->pmu, &fake_pmu, sibling))
 406                         return -EINVAL;
 407         }
 408
 409         if (!validate_event(event->pmu, &fake_pmu, event))
 410                 return -EINVAL;
 411
 412         return 0;
 413 }
 414
 415 static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
 416 {
 417         struct arm_pmu *armpmu;
 418         int ret;
 419         u64 start_clock, finish_clock;
 420
 421         /*
 422          * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
 423          * the handlers expect a struct arm_pmu*. The percpu_irq framework will
 424          * do any necessary shifting, we just need to perform the first
 425          * dereference.
 426          */
 427         armpmu = *(void **)dev;
 428         if (WARN_ON_ONCE(!armpmu))
 429                 return IRQ_NONE;
 430
 431         start_clock = sched_clock();
 432         ret = armpmu->handle_irq(armpmu);
 433         finish_clock = sched_clock();
 434
 435         perf_sample_event_took(finish_clock - start_clock);
 436         return ret;
 437 }
 438
 439 static int
 440 __hw_perf_event_init(struct perf_event *event)
 441 {
 442         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 443         struct hw_perf_event *hwc = &event->hw;
 444         int mapping;
 445
 446         hwc->flags = 0;
 447         mapping = armpmu->map_event(event);
 448
 449         if (mapping < 0) {
 450                 pr_debug("event %x:%llx not supported\n", event->attr.type,
 451                          event->attr.config);
 452                 return mapping;
 453         }
 454
 455         /*
 456          * We don't assign an index until we actually place the event onto
 457          * hardware. Use -1 to signify that we haven't decided where to put it
 458          * yet. For SMP systems, each core has it's own PMU so we can't do any
 459          * clever allocation or constraints checking at this point.
 460          */
 461         hwc->idx                = -1;
 462         hwc->config_base        = 0;
 463         hwc->config             = 0;
 464         hwc->event_base         = 0;
 465
 466         /*
 467          * Check whether we need to exclude the counter from certain modes.
 468          */
 469         if (armpmu->set_event_filter &&
 470             armpmu->set_event_filter(hwc, &event->attr)) {
 471                 pr_debug("ARM performance counters do not support "
 472                          "mode exclusion\n");
 473                 return -EOPNOTSUPP;
 474         }
 475
 476         /*
 477          * Store the event encoding into the config_base field.
 478          */
 479         hwc->config_base            |= (unsigned long)mapping;
 480
 481         if (!is_sampling_event(event)) {
 482                 /*
 483                  * For non-sampling runs, limit the sample_period to half
 484                  * of the counter width. That way, the new counter value
 485                  * is far less likely to overtake the previous one unless
 486                  * you have some serious IRQ latency issues.
 487                  */
 488                 hwc->sample_period  = arm_pmu_event_max_period(event) >> 1;
 489                 hwc->last_period    = hwc->sample_period;
 490                 local64_set(&hwc->period_left, hwc->sample_period);
 491         }
 492
 493         return validate_group(event);
 494 }
 495
 496 static int armpmu_event_init(struct perf_event *event)
 497 {
 498         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 499
 500         /*
 501          * Reject CPU-affine events for CPUs that are of a different class to
 502          * that which this PMU handles. Process-following events (where
 503          * event->cpu == -1) can be migrated between CPUs, and thus we have to
 504          * reject them later (in armpmu_add) if they're scheduled on a
 505          * different class of CPU.
 506          */
 507         if (event->cpu != -1 &&
 508                 !cpumask_test_cpu(event->cpu, &armpmu->supported_cpus))
 509                 return -ENOENT;
 510
 511         /* does not support taken branch sampling */
 512         if (has_branch_stack(event))
 513                 return -EOPNOTSUPP;
 514
 515         if (armpmu->map_event(event) == -ENOENT)
 516                 return -ENOENT;
 517
 518         return __hw_perf_event_init(event);
 519 }
 520
 521 static void armpmu_enable(struct pmu *pmu)
 522 {
 523         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 524         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 525         int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
 526
 527         /* For task-bound events we may be called on other CPUs */
 528         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 529                 return;
 530
 531         if (enabled)
 532                 armpmu->start(armpmu);
 533 }
 534
 535 static void armpmu_disable(struct pmu *pmu)
 536 {
 537         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 538
 539         /* For task-bound events we may be called on other CPUs */
 540         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 541                 return;
 542
 543         armpmu->stop(armpmu);
 544 }
 545
 546 /*
 547  * In heterogeneous systems, events are specific to a particular
 548  * microarchitecture, and aren't suitable for another. Thus, only match CPUs of
 549  * the same microarchitecture.
 550  */
 551 static int armpmu_filter_match(struct perf_event *event)
 552 {
 553         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 554         unsigned int cpu = smp_processor_id();
 555         int ret;
 556
 557         ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus);
 558         if (ret && armpmu->filter_match)
 559                 return armpmu->filter_match(event);
 560
 561         return ret;
 562 }
 563
 564 static ssize_t armpmu_cpumask_show(struct device *dev,
 565                                    struct device_attribute *attr, char *buf)
 566 {
 567         struct arm_pmu *armpmu = to_arm_pmu(dev_get_drvdata(dev));
 568         return cpumap_print_to_pagebuf(true, buf, &armpmu->supported_cpus);
 569 }
 570
 571 static DEVICE_ATTR(cpus, S_IRUGO, armpmu_cpumask_show, NULL);
 572
 573 static struct attribute *armpmu_common_attrs[] = {
 574         &dev_attr_cpus.attr,
 575         NULL,
 576 };
 577
 578 static struct attribute_group armpmu_common_attr_group = {
 579         .attrs = armpmu_common_attrs,
 580 };
 581
 582 /* Set at runtime when we know what CPU type we are. */
 583 static struct arm_pmu *__oprofile_cpu_pmu;
 584
 585 /*
 586  * Despite the names, these two functions are CPU-specific and are used
 587  * by the OProfile/perf code.
 588  */
 589 const char *perf_pmu_name(void)
 590 {
 591         if (!__oprofile_cpu_pmu)
 592                 return NULL;
 593
 594         return __oprofile_cpu_pmu->name;
 595 }
 596 EXPORT_SYMBOL_GPL(perf_pmu_name);
 597
 598 int perf_num_counters(void)
 599 {
 600         int max_events = 0;
 601
 602         if (__oprofile_cpu_pmu != NULL)
 603                 max_events = __oprofile_cpu_pmu->num_events;
 604
 605         return max_events;
 606 }
 607 EXPORT_SYMBOL_GPL(perf_num_counters);
 608
 609 static int armpmu_count_irq_users(const int irq)
 610 {
 611         int cpu, count = 0;
 612
 613         for_each_possible_cpu(cpu) {
 614                 if (per_cpu(cpu_irq, cpu) == irq)
 615                         count++;
 616         }
 617
 618         return count;
 619 }
 620
 621 static const struct pmu_irq_ops *armpmu_find_irq_ops(int irq)
 622 {
 623         const struct pmu_irq_ops *ops = NULL;
 624         int cpu;
 625
 626         for_each_possible_cpu(cpu) {
 627                 if (per_cpu(cpu_irq, cpu) != irq)
 628                         continue;
 629
 630                 ops = per_cpu(cpu_irq_ops, cpu);
 631                 if (ops)
 632                         break;
 633         }
 634
 635         return ops;
 636 }
 637
 638 void armpmu_free_irq(int irq, int cpu)
 639 {
 640         if (per_cpu(cpu_irq, cpu) == 0)
 641                 return;
 642         if (WARN_ON(irq != per_cpu(cpu_irq, cpu)))
 643                 return;
 644
 645         per_cpu(cpu_irq_ops, cpu)->free_pmuirq(irq, cpu, &cpu_armpmu);
 646
 647         per_cpu(cpu_irq, cpu) = 0;
 648         per_cpu(cpu_irq_ops, cpu) = NULL;
 649 }
 650
 651 int armpmu_request_irq(int irq, int cpu)
 652 {
 653         int err = 0;
 654         const irq_handler_t handler = armpmu_dispatch_irq;
 655         const struct pmu_irq_ops *irq_ops;
 656
 657         if (!irq)
 658                 return 0;
 659
 660         if (!irq_is_percpu_devid(irq)) {
 661                 unsigned long irq_flags;
 662
 663                 err = irq_force_affinity(irq, cpumask_of(cpu));
 664
 665                 if (err && num_possible_cpus() > 1) {
 666                         pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
 667                                 irq, cpu);
 668                         goto err_out;
 669                 }
 670
 671                 irq_flags = IRQF_PERCPU |
 672                             IRQF_NOBALANCING |
 673                             IRQF_NO_THREAD;
 674
 675                 irq_set_status_flags(irq, IRQ_NOAUTOEN);
 676
 677                 err = request_nmi(irq, handler, irq_flags, "arm-pmu",
 678                                   per_cpu_ptr(&cpu_armpmu, cpu));
 679
 680                 /* If cannot get an NMI, get a normal interrupt */
 681                 if (err) {
 682                         err = request_irq(irq, handler, irq_flags, "arm-pmu",
 683                                           per_cpu_ptr(&cpu_armpmu, cpu));
 684                         irq_ops = &pmuirq_ops;
 685                 } else {
 686                         has_nmi = true;
 687                         irq_ops = &pmunmi_ops;
 688                 }
 689         } else if (armpmu_count_irq_users(irq) == 0) {
 690                 err = request_percpu_nmi(irq, handler, "arm-pmu", &cpu_armpmu);
 691
 692                 /* If cannot get an NMI, get a normal interrupt */
 693                 if (err) {
 694                         err = request_percpu_irq(irq, handler, "arm-pmu",
 695                                                  &cpu_armpmu);
 696                         irq_ops = &percpu_pmuirq_ops;
 697                 } else {
 698                         has_nmi= true;
 699                         irq_ops = &percpu_pmunmi_ops;
 700                 }
 701         } else {
 702                 /* Per cpudevid irq was already requested by another CPU */
 703                 irq_ops = armpmu_find_irq_ops(irq);
 704
 705                 if (WARN_ON(!irq_ops))
 706                         err = -EINVAL;
 707         }
 708
 709         if (err)
 710                 goto err_out;
 711
 712         per_cpu(cpu_irq, cpu) = irq;
 713         per_cpu(cpu_irq_ops, cpu) = irq_ops;
 714         return 0;
 715
 716 err_out:
 717         pr_err("unable to request IRQ%d for ARM PMU counters\n", irq);
 718         return err;
 719 }
 720
 721 static int armpmu_get_cpu_irq(struct arm_pmu *pmu, int cpu)
 722 {
 723         struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
 724         return per_cpu(hw_events->irq, cpu);
 725 }
 726
 727 /*
 728  * PMU hardware loses all context when a CPU goes offline.
 729  * When a CPU is hotplugged back in, since some hardware registers are
 730  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 731  * junk values out of them.
 732  */
 733 static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
 734 {
 735         struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
 736         int irq;
 737
 738         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 739                 return 0;
 740         if (pmu->reset)
 741                 pmu->reset(pmu);
 742
 743         per_cpu(cpu_armpmu, cpu) = pmu;
 744
 745         irq = armpmu_get_cpu_irq(pmu, cpu);
 746         if (irq)
 747                 per_cpu(cpu_irq_ops, cpu)->enable_pmuirq(irq);
 748
 749         return 0;
 750 }
 751
 752 static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
 753 {
 754         struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
 755         int irq;
 756
 757         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 758                 return 0;
 759
 760         irq = armpmu_get_cpu_irq(pmu, cpu);
 761         if (irq)
 762                 per_cpu(cpu_irq_ops, cpu)->disable_pmuirq(irq);
 763
 764         per_cpu(cpu_armpmu, cpu) = NULL;
 765
 766         return 0;
 767 }
 768
 769 #ifdef CONFIG_CPU_PM
 770 static void cpu_pm_pmu_setup(struct arm_pmu *armpmu, unsigned long cmd)
 771 {
 772         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 773         struct perf_event *event;
 774         int idx;
 775
 776         for (idx = 0; idx < armpmu->num_events; idx++) {
 777                 event = hw_events->events[idx];
 778                 if (!event)
 779                         continue;
 780
 781                 switch (cmd) {
 782                 case CPU_PM_ENTER:
 783                         /*
 784                          * Stop and update the counter
 785                          */
 786                         armpmu_stop(event, PERF_EF_UPDATE);
 787                         break;
 788                 case CPU_PM_EXIT:
 789                 case CPU_PM_ENTER_FAILED:
 790                          /*
 791                           * Restore and enable the counter.
 792                           * armpmu_start() indirectly calls
 793                           *
 794                           * perf_event_update_userpage()
 795                           *
 796                           * that requires RCU read locking to be functional,
 797                           * wrap the call within RCU_NONIDLE to make the
 798                           * RCU subsystem aware this cpu is not idle from
 799                           * an RCU perspective for the armpmu_start() call
 800                           * duration.
 801                           */
 802                         RCU_NONIDLE(armpmu_start(event, PERF_EF_RELOAD));
 803                         break;
 804                 default:
 805                         break;
 806                 }
 807         }
 808 }
 809
 810 static int cpu_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
 811                              void *v)
 812 {
 813         struct arm_pmu *armpmu = container_of(b, struct arm_pmu, cpu_pm_nb);
 814         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 815         int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
 816
 817         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 818                 return NOTIFY_DONE;
 819
 820         /*
 821          * Always reset the PMU registers on power-up even if
 822          * there are no events running.
 823          */
 824         if (cmd == CPU_PM_EXIT && armpmu->reset)
 825                 armpmu->reset(armpmu);
 826
 827         if (!enabled)
 828                 return NOTIFY_OK;
 829
 830         switch (cmd) {
 831         case CPU_PM_ENTER:
 832                 armpmu->stop(armpmu);
 833                 cpu_pm_pmu_setup(armpmu, cmd);
 834                 break;
 835         case CPU_PM_EXIT:
 836         case CPU_PM_ENTER_FAILED:
 837                 cpu_pm_pmu_setup(armpmu, cmd);
 838                 armpmu->start(armpmu);
 839                 break;
 840         default:
 841                 return NOTIFY_DONE;
 842         }
 843
 844         return NOTIFY_OK;
 845 }
 846
 847 static int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu)
 848 {
 849         cpu_pmu->cpu_pm_nb.notifier_call = cpu_pm_pmu_notify;
 850         return cpu_pm_register_notifier(&cpu_pmu->cpu_pm_nb);
 851 }
 852
 853 static void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu)
 854 {
 855         cpu_pm_unregister_notifier(&cpu_pmu->cpu_pm_nb);
 856 }
 857 #else
 858 static inline int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu) { return 0; }
 859 static inline void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu) { }
 860 #endif
 861
 862 static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
 863 {
 864         int err;
 865
 866         err = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_STARTING,
 867                                        &cpu_pmu->node);
 868         if (err)
 869                 goto out;
 870
 871         err = cpu_pm_pmu_register(cpu_pmu);
 872         if (err)
 873                 goto out_unregister;
 874
 875         return 0;
 876
 877 out_unregister:
 878         cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
 879                                             &cpu_pmu->node);
 880 out:
 881         return err;
 882 }
 883
 884 static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
 885 {
 886         cpu_pm_pmu_unregister(cpu_pmu);
 887         cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
 888                                             &cpu_pmu->node);
 889 }
 890
 891 static struct arm_pmu *__armpmu_alloc(gfp_t flags)
 892 {
 893         struct arm_pmu *pmu;
 894         int cpu;
 895
 896         pmu = kzalloc(sizeof(*pmu), flags);
 897         if (!pmu) {
 898                 pr_info("failed to allocate PMU device!\n");
 899                 goto out;
 900         }
 901
 902         pmu->hw_events = alloc_percpu_gfp(struct pmu_hw_events, flags);
 903         if (!pmu->hw_events) {
 904                 pr_info("failed to allocate per-cpu PMU data.\n");
 905                 goto out_free_pmu;
 906         }
 907
 908         pmu->pmu = (struct pmu) {
 909                 .pmu_enable     = armpmu_enable,
 910                 .pmu_disable    = armpmu_disable,
 911                 .event_init     = armpmu_event_init,
 912                 .add            = armpmu_add,
 913                 .del            = armpmu_del,
 914                 .start          = armpmu_start,
 915                 .stop           = armpmu_stop,
 916                 .read           = armpmu_read,
 917                 .filter_match   = armpmu_filter_match,
 918                 .attr_groups    = pmu->attr_groups,
 919                 /*
 920                  * This is a CPU PMU potentially in a heterogeneous
 921                  * configuration (e.g. big.LITTLE). This is not an uncore PMU,
 922                  * and we have taken ctx sharing into account (e.g. with our
 923                  * pmu::filter_match callback and pmu::event_init group
 924                  * validation).
 925                  */
 926                 .capabilities   = PERF_PMU_CAP_HETEROGENEOUS_CPUS,
 927         };
 928
 929         pmu->attr_groups[ARMPMU_ATTR_GROUP_COMMON] =
 930                 &armpmu_common_attr_group;
 931
 932         for_each_possible_cpu(cpu) {
 933                 struct pmu_hw_events *events;
 934
 935                 events = per_cpu_ptr(pmu->hw_events, cpu);
 936                 raw_spin_lock_init(&events->pmu_lock);
 937                 events->percpu_pmu = pmu;
 938         }
 939
 940         return pmu;
 941
 942 out_free_pmu:
 943         kfree(pmu);
 944 out:
 945         return NULL;
 946 }
 947
 948 struct arm_pmu *armpmu_alloc(void)
 949 {
 950         return __armpmu_alloc(GFP_KERNEL);
 951 }
 952
 953 struct arm_pmu *armpmu_alloc_atomic(void)
 954 {
 955         return __armpmu_alloc(GFP_ATOMIC);
 956 }
 957
 958
 959 void armpmu_free(struct arm_pmu *pmu)
 960 {
 961         free_percpu(pmu->hw_events);
 962         kfree(pmu);
 963 }
 964
 965 int armpmu_register(struct arm_pmu *pmu)
 966 {
 967         int ret;
 968
 969         ret = cpu_pmu_init(pmu);
 970         if (ret)
 971                 return ret;
 972
 973         if (!pmu->set_event_filter)
 974                 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
 975
 976         ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
 977         if (ret)
 978                 goto out_destroy;
 979
 980         if (!__oprofile_cpu_pmu)
 981                 __oprofile_cpu_pmu = pmu;
 982
 983         pr_info("enabled with %s PMU driver, %d counters available%s\n",
 984                 pmu->name, pmu->num_events,
 985                 has_nmi ? ", using NMIs" : "");
 986
 987         return 0;
 988
 989 out_destroy:
 990         cpu_pmu_destroy(pmu);
 991         return ret;
 992 }
 993
 994 static int arm_pmu_hp_init(void)
 995 {
 996         int ret;
 997
 998         ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_STARTING,
 999                                       "perf/arm/pmu:starting",
1000                                       arm_perf_starting_cpu,
1001                                       arm_perf_teardown_cpu);
1002         if (ret)
1003                 pr_err("CPU hotplug notifier for ARM PMU could not be registered: %d\n",
1004                        ret);
1005         return ret;
1006 }
1007 subsys_initcall(arm_pmu_hp_init);