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         for_each_sibling_event(sibling, leader) {
 402                 if (!validate_event(event->pmu, &fake_pmu, sibling))
 403                         return -EINVAL;
 404         }
 405
 406         if (!validate_event(event->pmu, &fake_pmu, event))
 407                 return -EINVAL;
 408
 409         return 0;
 410 }
 411
 412 static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
 413 {
 414         struct arm_pmu *armpmu;
 415         int ret;
 416         u64 start_clock, finish_clock;
 417
 418         /*
 419          * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
 420          * the handlers expect a struct arm_pmu*. The percpu_irq framework will
 421          * do any necessary shifting, we just need to perform the first
 422          * dereference.
 423          */
 424         armpmu = *(void **)dev;
 425         if (WARN_ON_ONCE(!armpmu))
 426                 return IRQ_NONE;
 427
 428         start_clock = sched_clock();
 429         ret = armpmu->handle_irq(armpmu);
 430         finish_clock = sched_clock();
 431
 432         perf_sample_event_took(finish_clock - start_clock);
 433         return ret;
 434 }
 435
 436 static int
 437 __hw_perf_event_init(struct perf_event *event)
 438 {
 439         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 440         struct hw_perf_event *hwc = &event->hw;
 441         int mapping;
 442
 443         hwc->flags = 0;
 444         mapping = armpmu->map_event(event);
 445
 446         if (mapping < 0) {
 447                 pr_debug("event %x:%llx not supported\n", event->attr.type,
 448                          event->attr.config);
 449                 return mapping;
 450         }
 451
 452         /*
 453          * We don't assign an index until we actually place the event onto
 454          * hardware. Use -1 to signify that we haven't decided where to put it
 455          * yet. For SMP systems, each core has it's own PMU so we can't do any
 456          * clever allocation or constraints checking at this point.
 457          */
 458         hwc->idx                = -1;
 459         hwc->config_base        = 0;
 460         hwc->config             = 0;
 461         hwc->event_base         = 0;
 462
 463         /*
 464          * Check whether we need to exclude the counter from certain modes.
 465          */
 466         if (armpmu->set_event_filter &&
 467             armpmu->set_event_filter(hwc, &event->attr)) {
 468                 pr_debug("ARM performance counters do not support "
 469                          "mode exclusion\n");
 470                 return -EOPNOTSUPP;
 471         }
 472
 473         /*
 474          * Store the event encoding into the config_base field.
 475          */
 476         hwc->config_base            |= (unsigned long)mapping;
 477
 478         if (!is_sampling_event(event)) {
 479                 /*
 480                  * For non-sampling runs, limit the sample_period to half
 481                  * of the counter width. That way, the new counter value
 482                  * is far less likely to overtake the previous one unless
 483                  * you have some serious IRQ latency issues.
 484                  */
 485                 hwc->sample_period  = arm_pmu_event_max_period(event) >> 1;
 486                 hwc->last_period    = hwc->sample_period;
 487                 local64_set(&hwc->period_left, hwc->sample_period);
 488         }
 489
 490         if (event->group_leader != event) {
 491                 if (validate_group(event) != 0)
 492                         return -EINVAL;
 493         }
 494
 495         return 0;
 496 }
 497
 498 static int armpmu_event_init(struct perf_event *event)
 499 {
 500         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 501
 502         /*
 503          * Reject CPU-affine events for CPUs that are of a different class to
 504          * that which this PMU handles. Process-following events (where
 505          * event->cpu == -1) can be migrated between CPUs, and thus we have to
 506          * reject them later (in armpmu_add) if they're scheduled on a
 507          * different class of CPU.
 508          */
 509         if (event->cpu != -1 &&
 510                 !cpumask_test_cpu(event->cpu, &armpmu->supported_cpus))
 511                 return -ENOENT;
 512
 513         /* does not support taken branch sampling */
 514         if (has_branch_stack(event))
 515                 return -EOPNOTSUPP;
 516
 517         if (armpmu->map_event(event) == -ENOENT)
 518                 return -ENOENT;
 519
 520         return __hw_perf_event_init(event);
 521 }
 522
 523 static void armpmu_enable(struct pmu *pmu)
 524 {
 525         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 526         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 527         int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
 528
 529         /* For task-bound events we may be called on other CPUs */
 530         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 531                 return;
 532
 533         if (enabled)
 534                 armpmu->start(armpmu);
 535 }
 536
 537 static void armpmu_disable(struct pmu *pmu)
 538 {
 539         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 540
 541         /* For task-bound events we may be called on other CPUs */
 542         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 543                 return;
 544
 545         armpmu->stop(armpmu);
 546 }
 547
 548 /*
 549  * In heterogeneous systems, events are specific to a particular
 550  * microarchitecture, and aren't suitable for another. Thus, only match CPUs of
 551  * the same microarchitecture.
 552  */
 553 static int armpmu_filter_match(struct perf_event *event)
 554 {
 555         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 556         unsigned int cpu = smp_processor_id();
 557         int ret;
 558
 559         ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus);
 560         if (ret && armpmu->filter_match)
 561                 return armpmu->filter_match(event);
 562
 563         return ret;
 564 }
 565
 566 static ssize_t armpmu_cpumask_show(struct device *dev,
 567                                    struct device_attribute *attr, char *buf)
 568 {
 569         struct arm_pmu *armpmu = to_arm_pmu(dev_get_drvdata(dev));
 570         return cpumap_print_to_pagebuf(true, buf, &armpmu->supported_cpus);
 571 }
 572
 573 static DEVICE_ATTR(cpus, S_IRUGO, armpmu_cpumask_show, NULL);
 574
 575 static struct attribute *armpmu_common_attrs[] = {
 576         &dev_attr_cpus.attr,
 577         NULL,
 578 };
 579
 580 static struct attribute_group armpmu_common_attr_group = {
 581         .attrs = armpmu_common_attrs,
 582 };
 583
 584 /* Set at runtime when we know what CPU type we are. */
 585 static struct arm_pmu *__oprofile_cpu_pmu;
 586
 587 /*
 588  * Despite the names, these two functions are CPU-specific and are used
 589  * by the OProfile/perf code.
 590  */
 591 const char *perf_pmu_name(void)
 592 {
 593         if (!__oprofile_cpu_pmu)
 594                 return NULL;
 595
 596         return __oprofile_cpu_pmu->name;
 597 }
 598 EXPORT_SYMBOL_GPL(perf_pmu_name);
 599
 600 int perf_num_counters(void)
 601 {
 602         int max_events = 0;
 603
 604         if (__oprofile_cpu_pmu != NULL)
 605                 max_events = __oprofile_cpu_pmu->num_events;
 606
 607         return max_events;
 608 }
 609 EXPORT_SYMBOL_GPL(perf_num_counters);
 610
 611 static int armpmu_count_irq_users(const int irq)
 612 {
 613         int cpu, count = 0;
 614
 615         for_each_possible_cpu(cpu) {
 616                 if (per_cpu(cpu_irq, cpu) == irq)
 617                         count++;
 618         }
 619
 620         return count;
 621 }
 622
 623 static const struct pmu_irq_ops *armpmu_find_irq_ops(int irq)
 624 {
 625         const struct pmu_irq_ops *ops = NULL;
 626         int cpu;
 627
 628         for_each_possible_cpu(cpu) {
 629                 if (per_cpu(cpu_irq, cpu) != irq)
 630                         continue;
 631
 632                 ops = per_cpu(cpu_irq_ops, cpu);
 633                 if (ops)
 634                         break;
 635         }
 636
 637         return ops;
 638 }
 639
 640 void armpmu_free_irq(int irq, int cpu)
 641 {
 642         if (per_cpu(cpu_irq, cpu) == 0)
 643                 return;
 644         if (WARN_ON(irq != per_cpu(cpu_irq, cpu)))
 645                 return;
 646
 647         per_cpu(cpu_irq_ops, cpu)->free_pmuirq(irq, cpu, &cpu_armpmu);
 648
 649         per_cpu(cpu_irq, cpu) = 0;
 650         per_cpu(cpu_irq_ops, cpu) = NULL;
 651 }
 652
 653 int armpmu_request_irq(int irq, int cpu)
 654 {
 655         int err = 0;
 656         const irq_handler_t handler = armpmu_dispatch_irq;
 657         const struct pmu_irq_ops *irq_ops;
 658
 659         if (!irq)
 660                 return 0;
 661
 662         if (!irq_is_percpu_devid(irq)) {
 663                 unsigned long irq_flags;
 664
 665                 err = irq_force_affinity(irq, cpumask_of(cpu));
 666
 667                 if (err && num_possible_cpus() > 1) {
 668                         pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
 669                                 irq, cpu);
 670                         goto err_out;
 671                 }
 672
 673                 irq_flags = IRQF_PERCPU |
 674                             IRQF_NOBALANCING |
 675                             IRQF_NO_THREAD;
 676
 677                 irq_set_status_flags(irq, IRQ_NOAUTOEN);
 678
 679                 err = request_nmi(irq, handler, irq_flags, "arm-pmu",
 680                                   per_cpu_ptr(&cpu_armpmu, cpu));
 681
 682                 /* If cannot get an NMI, get a normal interrupt */
 683                 if (err) {
 684                         err = request_irq(irq, handler, irq_flags, "arm-pmu",
 685                                           per_cpu_ptr(&cpu_armpmu, cpu));
 686                         irq_ops = &pmuirq_ops;
 687                 } else {
 688                         has_nmi = true;
 689                         irq_ops = &pmunmi_ops;
 690                 }
 691         } else if (armpmu_count_irq_users(irq) == 0) {
 692                 err = request_percpu_nmi(irq, handler, "arm-pmu", &cpu_armpmu);
 693
 694                 /* If cannot get an NMI, get a normal interrupt */
 695                 if (err) {
 696                         err = request_percpu_irq(irq, handler, "arm-pmu",
 697                                                  &cpu_armpmu);
 698                         irq_ops = &percpu_pmuirq_ops;
 699                 } else {
 700                         has_nmi= true;
 701                         irq_ops = &percpu_pmunmi_ops;
 702                 }
 703         } else {
 704                 /* Per cpudevid irq was already requested by another CPU */
 705                 irq_ops = armpmu_find_irq_ops(irq);
 706
 707                 if (WARN_ON(!irq_ops))
 708                         err = -EINVAL;
 709         }
 710
 711         if (err)
 712                 goto err_out;
 713
 714         per_cpu(cpu_irq, cpu) = irq;
 715         per_cpu(cpu_irq_ops, cpu) = irq_ops;
 716         return 0;
 717
 718 err_out:
 719         pr_err("unable to request IRQ%d for ARM PMU counters\n", irq);
 720         return err;
 721 }
 722
 723 static int armpmu_get_cpu_irq(struct arm_pmu *pmu, int cpu)
 724 {
 725         struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
 726         return per_cpu(hw_events->irq, cpu);
 727 }
 728
 729 bool arm_pmu_irq_is_nmi(void)
 730 {
 731         return has_nmi;
 732 }
 733
 734 /*
 735  * PMU hardware loses all context when a CPU goes offline.
 736  * When a CPU is hotplugged back in, since some hardware registers are
 737  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 738  * junk values out of them.
 739  */
 740 static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
 741 {
 742         struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
 743         int irq;
 744
 745         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 746                 return 0;
 747         if (pmu->reset)
 748                 pmu->reset(pmu);
 749
 750         per_cpu(cpu_armpmu, cpu) = pmu;
 751
 752         irq = armpmu_get_cpu_irq(pmu, cpu);
 753         if (irq)
 754                 per_cpu(cpu_irq_ops, cpu)->enable_pmuirq(irq);
 755
 756         return 0;
 757 }
 758
 759 static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
 760 {
 761         struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
 762         int irq;
 763
 764         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 765                 return 0;
 766
 767         irq = armpmu_get_cpu_irq(pmu, cpu);
 768         if (irq)
 769                 per_cpu(cpu_irq_ops, cpu)->disable_pmuirq(irq);
 770
 771         per_cpu(cpu_armpmu, cpu) = NULL;
 772
 773         return 0;
 774 }
 775
 776 #ifdef CONFIG_CPU_PM
 777 static void cpu_pm_pmu_setup(struct arm_pmu *armpmu, unsigned long cmd)
 778 {
 779         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 780         struct perf_event *event;
 781         int idx;
 782
 783         for (idx = 0; idx < armpmu->num_events; idx++) {
 784                 event = hw_events->events[idx];
 785                 if (!event)
 786                         continue;
 787
 788                 switch (cmd) {
 789                 case CPU_PM_ENTER:
 790                         /*
 791                          * Stop and update the counter
 792                          */
 793                         armpmu_stop(event, PERF_EF_UPDATE);
 794                         break;
 795                 case CPU_PM_EXIT:
 796                 case CPU_PM_ENTER_FAILED:
 797                          /*
 798                           * Restore and enable the counter.
 799                           * armpmu_start() indirectly calls
 800                           *
 801                           * perf_event_update_userpage()
 802                           *
 803                           * that requires RCU read locking to be functional,
 804                           * wrap the call within RCU_NONIDLE to make the
 805                           * RCU subsystem aware this cpu is not idle from
 806                           * an RCU perspective for the armpmu_start() call
 807                           * duration.
 808                           */
 809                         RCU_NONIDLE(armpmu_start(event, PERF_EF_RELOAD));
 810                         break;
 811                 default:
 812                         break;
 813                 }
 814         }
 815 }
 816
 817 static int cpu_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
 818                              void *v)
 819 {
 820         struct arm_pmu *armpmu = container_of(b, struct arm_pmu, cpu_pm_nb);
 821         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 822         int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
 823
 824         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 825                 return NOTIFY_DONE;
 826
 827         /*
 828          * Always reset the PMU registers on power-up even if
 829          * there are no events running.
 830          */
 831         if (cmd == CPU_PM_EXIT && armpmu->reset)
 832                 armpmu->reset(armpmu);
 833
 834         if (!enabled)
 835                 return NOTIFY_OK;
 836
 837         switch (cmd) {
 838         case CPU_PM_ENTER:
 839                 armpmu->stop(armpmu);
 840                 cpu_pm_pmu_setup(armpmu, cmd);
 841                 break;
 842         case CPU_PM_EXIT:
 843         case CPU_PM_ENTER_FAILED:
 844                 cpu_pm_pmu_setup(armpmu, cmd);
 845                 armpmu->start(armpmu);
 846                 break;
 847         default:
 848                 return NOTIFY_DONE;
 849         }
 850
 851         return NOTIFY_OK;
 852 }
 853
 854 static int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu)
 855 {
 856         cpu_pmu->cpu_pm_nb.notifier_call = cpu_pm_pmu_notify;
 857         return cpu_pm_register_notifier(&cpu_pmu->cpu_pm_nb);
 858 }
 859
 860 static void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu)
 861 {
 862         cpu_pm_unregister_notifier(&cpu_pmu->cpu_pm_nb);
 863 }
 864 #else
 865 static inline int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu) { return 0; }
 866 static inline void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu) { }
 867 #endif
 868
 869 static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
 870 {
 871         int err;
 872
 873         err = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_STARTING,
 874                                        &cpu_pmu->node);
 875         if (err)
 876                 goto out;
 877
 878         err = cpu_pm_pmu_register(cpu_pmu);
 879         if (err)
 880                 goto out_unregister;
 881
 882         return 0;
 883
 884 out_unregister:
 885         cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
 886                                             &cpu_pmu->node);
 887 out:
 888         return err;
 889 }
 890
 891 static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
 892 {
 893         cpu_pm_pmu_unregister(cpu_pmu);
 894         cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
 895                                             &cpu_pmu->node);
 896 }
 897
 898 static struct arm_pmu *__armpmu_alloc(gfp_t flags)
 899 {
 900         struct arm_pmu *pmu;
 901         int cpu;
 902
 903         pmu = kzalloc(sizeof(*pmu), flags);
 904         if (!pmu) {
 905                 pr_info("failed to allocate PMU device!\n");
 906                 goto out;
 907         }
 908
 909         pmu->hw_events = alloc_percpu_gfp(struct pmu_hw_events, flags);
 910         if (!pmu->hw_events) {
 911                 pr_info("failed to allocate per-cpu PMU data.\n");
 912                 goto out_free_pmu;
 913         }
 914
 915         pmu->pmu = (struct pmu) {
 916                 .pmu_enable     = armpmu_enable,
 917                 .pmu_disable    = armpmu_disable,
 918                 .event_init     = armpmu_event_init,
 919                 .add            = armpmu_add,
 920                 .del            = armpmu_del,
 921                 .start          = armpmu_start,
 922                 .stop           = armpmu_stop,
 923                 .read           = armpmu_read,
 924                 .filter_match   = armpmu_filter_match,
 925                 .attr_groups    = pmu->attr_groups,
 926                 /*
 927                  * This is a CPU PMU potentially in a heterogeneous
 928                  * configuration (e.g. big.LITTLE). This is not an uncore PMU,
 929                  * and we have taken ctx sharing into account (e.g. with our
 930                  * pmu::filter_match callback and pmu::event_init group
 931                  * validation).
 932                  */
 933                 .capabilities   = PERF_PMU_CAP_HETEROGENEOUS_CPUS,
 934         };
 935
 936         pmu->attr_groups[ARMPMU_ATTR_GROUP_COMMON] =
 937                 &armpmu_common_attr_group;
 938
 939         for_each_possible_cpu(cpu) {
 940                 struct pmu_hw_events *events;
 941
 942                 events = per_cpu_ptr(pmu->hw_events, cpu);
 943                 raw_spin_lock_init(&events->pmu_lock);
 944                 events->percpu_pmu = pmu;
 945         }
 946
 947         return pmu;
 948
 949 out_free_pmu:
 950         kfree(pmu);
 951 out:
 952         return NULL;
 953 }
 954
 955 struct arm_pmu *armpmu_alloc(void)
 956 {
 957         return __armpmu_alloc(GFP_KERNEL);
 958 }
 959
 960 struct arm_pmu *armpmu_alloc_atomic(void)
 961 {
 962         return __armpmu_alloc(GFP_ATOMIC);
 963 }
 964
 965
 966 void armpmu_free(struct arm_pmu *pmu)
 967 {
 968         free_percpu(pmu->hw_events);
 969         kfree(pmu);
 970 }
 971
 972 int armpmu_register(struct arm_pmu *pmu)
 973 {
 974         int ret;
 975
 976         ret = cpu_pmu_init(pmu);
 977         if (ret)
 978                 return ret;
 979
 980         if (!pmu->set_event_filter)
 981                 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
 982
 983         ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
 984         if (ret)
 985                 goto out_destroy;
 986
 987         if (!__oprofile_cpu_pmu)
 988                 __oprofile_cpu_pmu = pmu;
 989
 990         pr_info("enabled with %s PMU driver, %d counters available%s\n",
 991                 pmu->name, pmu->num_events,
 992                 has_nmi ? ", using NMIs" : "");
 993
 994         return 0;
 995
 996 out_destroy:
 997         cpu_pmu_destroy(pmu);
 998         return ret;
 999 }
1000
1001 static int arm_pmu_hp_init(void)
1002 {
1003         int ret;
1004
1005         ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_STARTING,
1006                                       "perf/arm/pmu:starting",
1007                                       arm_perf_starting_cpu,
1008                                       arm_perf_teardown_cpu);
1009         if (ret)
1010                 pr_err("CPU hotplug notifier for ARM PMU could not be registered: %d\n",
1011                        ret);
1012         return ret;
1013 }
1014 subsys_initcall(arm_pmu_hp_init);