drivers/perf/arm_pmu.c

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