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/export.h>
  17 #include <linux/kernel.h>
  18 #include <linux/of_device.h>
  19 #include <linux/perf/arm_pmu.h>
  20 #include <linux/platform_device.h>
  21 #include <linux/slab.h>
  22 #include <linux/spinlock.h>
  23 #include <linux/irq.h>
  24 #include <linux/irqdesc.h>
  25
  26 #include <asm/cputype.h>
  27 #include <asm/irq_regs.h>
  28
  29 static int
  30 armpmu_map_cache_event(const unsigned (*cache_map)
  31                                       [PERF_COUNT_HW_CACHE_MAX]
  32                                       [PERF_COUNT_HW_CACHE_OP_MAX]
  33                                       [PERF_COUNT_HW_CACHE_RESULT_MAX],
  34                        u64 config)
  35 {
  36         unsigned int cache_type, cache_op, cache_result, ret;
  37
  38         cache_type = (config >>  0) & 0xff;
  39         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
  40                 return -EINVAL;
  41
  42         cache_op = (config >>  8) & 0xff;
  43         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
  44                 return -EINVAL;
  45
  46         cache_result = (config >> 16) & 0xff;
  47         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
  48                 return -EINVAL;
  49
  50         ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
  51
  52         if (ret == CACHE_OP_UNSUPPORTED)
  53                 return -ENOENT;
  54
  55         return ret;
  56 }
  57
  58 static int
  59 armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
  60 {
  61         int mapping;
  62
  63         if (config >= PERF_COUNT_HW_MAX)
  64                 return -EINVAL;
  65
  66         mapping = (*event_map)[config];
  67         return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
  68 }
  69
  70 static int
  71 armpmu_map_raw_event(u32 raw_event_mask, u64 config)
  72 {
  73         return (int)(config & raw_event_mask);
  74 }
  75
  76 int
  77 armpmu_map_event(struct perf_event *event,
  78                  const unsigned (*event_map)[PERF_COUNT_HW_MAX],
  79                  const unsigned (*cache_map)
  80                                 [PERF_COUNT_HW_CACHE_MAX]
  81                                 [PERF_COUNT_HW_CACHE_OP_MAX]
  82                                 [PERF_COUNT_HW_CACHE_RESULT_MAX],
  83                  u32 raw_event_mask)
  84 {
  85         u64 config = event->attr.config;
  86         int type = event->attr.type;
  87
  88         if (type == event->pmu->type)
  89                 return armpmu_map_raw_event(raw_event_mask, config);
  90
  91         switch (type) {
  92         case PERF_TYPE_HARDWARE:
  93                 return armpmu_map_hw_event(event_map, config);
  94         case PERF_TYPE_HW_CACHE:
  95                 return armpmu_map_cache_event(cache_map, config);
  96         case PERF_TYPE_RAW:
  97                 return armpmu_map_raw_event(raw_event_mask, config);
  98         }
  99
 100         return -ENOENT;
 101 }
 102
 103 int armpmu_event_set_period(struct perf_event *event)
 104 {
 105         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 106         struct hw_perf_event *hwc = &event->hw;
 107         s64 left = local64_read(&hwc->period_left);
 108         s64 period = hwc->sample_period;
 109         int ret = 0;
 110
 111         if (unlikely(left <= -period)) {
 112                 left = period;
 113                 local64_set(&hwc->period_left, left);
 114                 hwc->last_period = period;
 115                 ret = 1;
 116         }
 117
 118         if (unlikely(left <= 0)) {
 119                 left += period;
 120                 local64_set(&hwc->period_left, left);
 121                 hwc->last_period = period;
 122                 ret = 1;
 123         }
 124
 125         /*
 126          * Limit the maximum period to prevent the counter value
 127          * from overtaking the one we are about to program. In
 128          * effect we are reducing max_period to account for
 129          * interrupt latency (and we are being very conservative).
 130          */
 131         if (left > (armpmu->max_period >> 1))
 132                 left = armpmu->max_period >> 1;
 133
 134         local64_set(&hwc->prev_count, (u64)-left);
 135
 136         armpmu->write_counter(event, (u64)(-left) & 0xffffffff);
 137
 138         perf_event_update_userpage(event);
 139
 140         return ret;
 141 }
 142
 143 u64 armpmu_event_update(struct perf_event *event)
 144 {
 145         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 146         struct hw_perf_event *hwc = &event->hw;
 147         u64 delta, prev_raw_count, new_raw_count;
 148
 149 again:
 150         prev_raw_count = local64_read(&hwc->prev_count);
 151         new_raw_count = armpmu->read_counter(event);
 152
 153         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 154                              new_raw_count) != prev_raw_count)
 155                 goto again;
 156
 157         delta = (new_raw_count - prev_raw_count) & armpmu->max_period;
 158
 159         local64_add(delta, &event->count);
 160         local64_sub(delta, &hwc->period_left);
 161
 162         return new_raw_count;
 163 }
 164
 165 static void
 166 armpmu_read(struct perf_event *event)
 167 {
 168         armpmu_event_update(event);
 169 }
 170
 171 static void
 172 armpmu_stop(struct perf_event *event, int flags)
 173 {
 174         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 175         struct hw_perf_event *hwc = &event->hw;
 176
 177         /*
 178          * ARM pmu always has to update the counter, so ignore
 179          * PERF_EF_UPDATE, see comments in armpmu_start().
 180          */
 181         if (!(hwc->state & PERF_HES_STOPPED)) {
 182                 armpmu->disable(event);
 183                 armpmu_event_update(event);
 184                 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
 185         }
 186 }
 187
 188 static void armpmu_start(struct perf_event *event, int flags)
 189 {
 190         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 191         struct hw_perf_event *hwc = &event->hw;
 192
 193         /*
 194          * ARM pmu always has to reprogram the period, so ignore
 195          * PERF_EF_RELOAD, see the comment below.
 196          */
 197         if (flags & PERF_EF_RELOAD)
 198                 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 199
 200         hwc->state = 0;
 201         /*
 202          * Set the period again. Some counters can't be stopped, so when we
 203          * were stopped we simply disabled the IRQ source and the counter
 204          * may have been left counting. If we don't do this step then we may
 205          * get an interrupt too soon or *way* too late if the overflow has
 206          * happened since disabling.
 207          */
 208         armpmu_event_set_period(event);
 209         armpmu->enable(event);
 210 }
 211
 212 static void
 213 armpmu_del(struct perf_event *event, int flags)
 214 {
 215         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 216         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 217         struct hw_perf_event *hwc = &event->hw;
 218         int idx = hwc->idx;
 219
 220         armpmu_stop(event, PERF_EF_UPDATE);
 221         hw_events->events[idx] = NULL;
 222         clear_bit(idx, hw_events->used_mask);
 223         if (armpmu->clear_event_idx)
 224                 armpmu->clear_event_idx(hw_events, event);
 225
 226         perf_event_update_userpage(event);
 227 }
 228
 229 static int
 230 armpmu_add(struct perf_event *event, int flags)
 231 {
 232         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 233         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 234         struct hw_perf_event *hwc = &event->hw;
 235         int idx;
 236         int err = 0;
 237
 238         /* An event following a process won't be stopped earlier */
 239         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 240                 return -ENOENT;
 241
 242         perf_pmu_disable(event->pmu);
 243
 244         /* If we don't have a space for the counter then finish early. */
 245         idx = armpmu->get_event_idx(hw_events, event);
 246         if (idx < 0) {
 247                 err = idx;
 248                 goto out;
 249         }
 250
 251         /*
 252          * If there is an event in the counter we are going to use then make
 253          * sure it is disabled.
 254          */
 255         event->hw.idx = idx;
 256         armpmu->disable(event);
 257         hw_events->events[idx] = event;
 258
 259         hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
 260         if (flags & PERF_EF_START)
 261                 armpmu_start(event, PERF_EF_RELOAD);
 262
 263         /* Propagate our changes to the userspace mapping. */
 264         perf_event_update_userpage(event);
 265
 266 out:
 267         perf_pmu_enable(event->pmu);
 268         return err;
 269 }
 270
 271 static int
 272 validate_event(struct pmu *pmu, struct pmu_hw_events *hw_events,
 273                                struct perf_event *event)
 274 {
 275         struct arm_pmu *armpmu;
 276
 277         if (is_software_event(event))
 278                 return 1;
 279
 280         /*
 281          * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
 282          * core perf code won't check that the pmu->ctx == leader->ctx
 283          * until after pmu->event_init(event).
 284          */
 285         if (event->pmu != pmu)
 286                 return 0;
 287
 288         if (event->state < PERF_EVENT_STATE_OFF)
 289                 return 1;
 290
 291         if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
 292                 return 1;
 293
 294         armpmu = to_arm_pmu(event->pmu);
 295         return armpmu->get_event_idx(hw_events, event) >= 0;
 296 }
 297
 298 static int
 299 validate_group(struct perf_event *event)
 300 {
 301         struct perf_event *sibling, *leader = event->group_leader;
 302         struct pmu_hw_events fake_pmu;
 303
 304         /*
 305          * Initialise the fake PMU. We only need to populate the
 306          * used_mask for the purposes of validation.
 307          */
 308         memset(&fake_pmu.used_mask, 0, sizeof(fake_pmu.used_mask));
 309
 310         if (!validate_event(event->pmu, &fake_pmu, leader))
 311                 return -EINVAL;
 312
 313         list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
 314                 if (!validate_event(event->pmu, &fake_pmu, sibling))
 315                         return -EINVAL;
 316         }
 317
 318         if (!validate_event(event->pmu, &fake_pmu, event))
 319                 return -EINVAL;
 320
 321         return 0;
 322 }
 323
 324 static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
 325 {
 326         struct arm_pmu *armpmu;
 327         struct platform_device *plat_device;
 328         struct arm_pmu_platdata *plat;
 329         int ret;
 330         u64 start_clock, finish_clock;
 331
 332         /*
 333          * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
 334          * the handlers expect a struct arm_pmu*. The percpu_irq framework will
 335          * do any necessary shifting, we just need to perform the first
 336          * dereference.
 337          */
 338         armpmu = *(void **)dev;
 339         plat_device = armpmu->plat_device;
 340         plat = dev_get_platdata(&plat_device->dev);
 341
 342         start_clock = sched_clock();
 343         if (plat && plat->handle_irq)
 344                 ret = plat->handle_irq(irq, armpmu, armpmu->handle_irq);
 345         else
 346                 ret = armpmu->handle_irq(irq, armpmu);
 347         finish_clock = sched_clock();
 348
 349         perf_sample_event_took(finish_clock - start_clock);
 350         return ret;
 351 }
 352
 353 static void
 354 armpmu_release_hardware(struct arm_pmu *armpmu)
 355 {
 356         armpmu->free_irq(armpmu);
 357 }
 358
 359 static int
 360 armpmu_reserve_hardware(struct arm_pmu *armpmu)
 361 {
 362         int err = armpmu->request_irq(armpmu, armpmu_dispatch_irq);
 363         if (err) {
 364                 armpmu_release_hardware(armpmu);
 365                 return err;
 366         }
 367
 368         return 0;
 369 }
 370
 371 static void
 372 hw_perf_event_destroy(struct perf_event *event)
 373 {
 374         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 375         atomic_t *active_events  = &armpmu->active_events;
 376         struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex;
 377
 378         if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) {
 379                 armpmu_release_hardware(armpmu);
 380                 mutex_unlock(pmu_reserve_mutex);
 381         }
 382 }
 383
 384 static int
 385 event_requires_mode_exclusion(struct perf_event_attr *attr)
 386 {
 387         return attr->exclude_idle || attr->exclude_user ||
 388                attr->exclude_kernel || attr->exclude_hv;
 389 }
 390
 391 static int
 392 __hw_perf_event_init(struct perf_event *event)
 393 {
 394         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 395         struct hw_perf_event *hwc = &event->hw;
 396         int mapping;
 397
 398         mapping = armpmu->map_event(event);
 399
 400         if (mapping < 0) {
 401                 pr_debug("event %x:%llx not supported\n", event->attr.type,
 402                          event->attr.config);
 403                 return mapping;
 404         }
 405
 406         /*
 407          * We don't assign an index until we actually place the event onto
 408          * hardware. Use -1 to signify that we haven't decided where to put it
 409          * yet. For SMP systems, each core has it's own PMU so we can't do any
 410          * clever allocation or constraints checking at this point.
 411          */
 412         hwc->idx                = -1;
 413         hwc->config_base        = 0;
 414         hwc->config             = 0;
 415         hwc->event_base         = 0;
 416
 417         /*
 418          * Check whether we need to exclude the counter from certain modes.
 419          */
 420         if ((!armpmu->set_event_filter ||
 421              armpmu->set_event_filter(hwc, &event->attr)) &&
 422              event_requires_mode_exclusion(&event->attr)) {
 423                 pr_debug("ARM performance counters do not support "
 424                          "mode exclusion\n");
 425                 return -EOPNOTSUPP;
 426         }
 427
 428         /*
 429          * Store the event encoding into the config_base field.
 430          */
 431         hwc->config_base            |= (unsigned long)mapping;
 432
 433         if (!is_sampling_event(event)) {
 434                 /*
 435                  * For non-sampling runs, limit the sample_period to half
 436                  * of the counter width. That way, the new counter value
 437                  * is far less likely to overtake the previous one unless
 438                  * you have some serious IRQ latency issues.
 439                  */
 440                 hwc->sample_period  = armpmu->max_period >> 1;
 441                 hwc->last_period    = hwc->sample_period;
 442                 local64_set(&hwc->period_left, hwc->sample_period);
 443         }
 444
 445         if (event->group_leader != event) {
 446                 if (validate_group(event) != 0)
 447                         return -EINVAL;
 448         }
 449
 450         return 0;
 451 }
 452
 453 static int armpmu_event_init(struct perf_event *event)
 454 {
 455         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 456         int err = 0;
 457         atomic_t *active_events = &armpmu->active_events;
 458
 459         /*
 460          * Reject CPU-affine events for CPUs that are of a different class to
 461          * that which this PMU handles. Process-following events (where
 462          * event->cpu == -1) can be migrated between CPUs, and thus we have to
 463          * reject them later (in armpmu_add) if they're scheduled on a
 464          * different class of CPU.
 465          */
 466         if (event->cpu != -1 &&
 467                 !cpumask_test_cpu(event->cpu, &armpmu->supported_cpus))
 468                 return -ENOENT;
 469
 470         /* does not support taken branch sampling */
 471         if (has_branch_stack(event))
 472                 return -EOPNOTSUPP;
 473
 474         if (armpmu->map_event(event) == -ENOENT)
 475                 return -ENOENT;
 476
 477         event->destroy = hw_perf_event_destroy;
 478
 479         if (!atomic_inc_not_zero(active_events)) {
 480                 mutex_lock(&armpmu->reserve_mutex);
 481                 if (atomic_read(active_events) == 0)
 482                         err = armpmu_reserve_hardware(armpmu);
 483
 484                 if (!err)
 485                         atomic_inc(active_events);
 486                 mutex_unlock(&armpmu->reserve_mutex);
 487         }
 488
 489         if (err)
 490                 return err;
 491
 492         err = __hw_perf_event_init(event);
 493         if (err)
 494                 hw_perf_event_destroy(event);
 495
 496         return err;
 497 }
 498
 499 static void armpmu_enable(struct pmu *pmu)
 500 {
 501         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 502         struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
 503         int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
 504
 505         /* For task-bound events we may be called on other CPUs */
 506         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 507                 return;
 508
 509         if (enabled)
 510                 armpmu->start(armpmu);
 511 }
 512
 513 static void armpmu_disable(struct pmu *pmu)
 514 {
 515         struct arm_pmu *armpmu = to_arm_pmu(pmu);
 516
 517         /* For task-bound events we may be called on other CPUs */
 518         if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
 519                 return;
 520
 521         armpmu->stop(armpmu);
 522 }
 523
 524 /*
 525  * In heterogeneous systems, events are specific to a particular
 526  * microarchitecture, and aren't suitable for another. Thus, only match CPUs of
 527  * the same microarchitecture.
 528  */
 529 static int armpmu_filter_match(struct perf_event *event)
 530 {
 531         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
 532         unsigned int cpu = smp_processor_id();
 533         return cpumask_test_cpu(cpu, &armpmu->supported_cpus);
 534 }
 535
 536 static void armpmu_init(struct arm_pmu *armpmu)
 537 {
 538         atomic_set(&armpmu->active_events, 0);
 539         mutex_init(&armpmu->reserve_mutex);
 540
 541         armpmu->pmu = (struct pmu) {
 542                 .pmu_enable     = armpmu_enable,
 543                 .pmu_disable    = armpmu_disable,
 544                 .event_init     = armpmu_event_init,
 545                 .add            = armpmu_add,
 546                 .del            = armpmu_del,
 547                 .start          = armpmu_start,
 548                 .stop           = armpmu_stop,
 549                 .read           = armpmu_read,
 550                 .filter_match   = armpmu_filter_match,
 551         };
 552 }
 553
 554 int armpmu_register(struct arm_pmu *armpmu, int type)
 555 {
 556         armpmu_init(armpmu);
 557         pr_info("enabled with %s PMU driver, %d counters available\n",
 558                         armpmu->name, armpmu->num_events);
 559         return perf_pmu_register(&armpmu->pmu, armpmu->name, type);
 560 }
 561
 562 /* Set at runtime when we know what CPU type we are. */
 563 static struct arm_pmu *__oprofile_cpu_pmu;
 564
 565 /*
 566  * Despite the names, these two functions are CPU-specific and are used
 567  * by the OProfile/perf code.
 568  */
 569 const char *perf_pmu_name(void)
 570 {
 571         if (!__oprofile_cpu_pmu)
 572                 return NULL;
 573
 574         return __oprofile_cpu_pmu->name;
 575 }
 576 EXPORT_SYMBOL_GPL(perf_pmu_name);
 577
 578 int perf_num_counters(void)
 579 {
 580         int max_events = 0;
 581
 582         if (__oprofile_cpu_pmu != NULL)
 583                 max_events = __oprofile_cpu_pmu->num_events;
 584
 585         return max_events;
 586 }
 587 EXPORT_SYMBOL_GPL(perf_num_counters);
 588
 589 static void cpu_pmu_enable_percpu_irq(void *data)
 590 {
 591         int irq = *(int *)data;
 592
 593         enable_percpu_irq(irq, IRQ_TYPE_NONE);
 594 }
 595
 596 static void cpu_pmu_disable_percpu_irq(void *data)
 597 {
 598         int irq = *(int *)data;
 599
 600         disable_percpu_irq(irq);
 601 }
 602
 603 static void cpu_pmu_free_irq(struct arm_pmu *cpu_pmu)
 604 {
 605         int i, irq, irqs;
 606         struct platform_device *pmu_device = cpu_pmu->plat_device;
 607         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
 608
 609         irqs = min(pmu_device->num_resources, num_possible_cpus());
 610
 611         irq = platform_get_irq(pmu_device, 0);
 612         if (irq >= 0 && irq_is_percpu(irq)) {
 613                 on_each_cpu(cpu_pmu_disable_percpu_irq, &irq, 1);
 614                 free_percpu_irq(irq, &hw_events->percpu_pmu);
 615         } else {
 616                 for (i = 0; i < irqs; ++i) {
 617                         int cpu = i;
 618
 619                         if (cpu_pmu->irq_affinity)
 620                                 cpu = cpu_pmu->irq_affinity[i];
 621
 622                         if (!cpumask_test_and_clear_cpu(cpu, &cpu_pmu->active_irqs))
 623                                 continue;
 624                         irq = platform_get_irq(pmu_device, i);
 625                         if (irq >= 0)
 626                                 free_irq(irq, per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 627                 }
 628         }
 629 }
 630
 631 static int cpu_pmu_request_irq(struct arm_pmu *cpu_pmu, irq_handler_t handler)
 632 {
 633         int i, err, irq, irqs;
 634         struct platform_device *pmu_device = cpu_pmu->plat_device;
 635         struct pmu_hw_events __percpu *hw_events = cpu_pmu->hw_events;
 636
 637         if (!pmu_device)
 638                 return -ENODEV;
 639
 640         irqs = min(pmu_device->num_resources, num_possible_cpus());
 641         if (irqs < 1) {
 642                 pr_warn_once("perf/ARM: No irqs for PMU defined, sampling events not supported\n");
 643                 return 0;
 644         }
 645
 646         irq = platform_get_irq(pmu_device, 0);
 647         if (irq >= 0 && irq_is_percpu(irq)) {
 648                 err = request_percpu_irq(irq, handler, "arm-pmu",
 649                                          &hw_events->percpu_pmu);
 650                 if (err) {
 651                         pr_err("unable to request IRQ%d for ARM PMU counters\n",
 652                                 irq);
 653                         return err;
 654                 }
 655                 on_each_cpu(cpu_pmu_enable_percpu_irq, &irq, 1);
 656         } else {
 657                 for (i = 0; i < irqs; ++i) {
 658                         int cpu = i;
 659
 660                         err = 0;
 661                         irq = platform_get_irq(pmu_device, i);
 662                         if (irq < 0)
 663                                 continue;
 664
 665                         if (cpu_pmu->irq_affinity)
 666                                 cpu = cpu_pmu->irq_affinity[i];
 667
 668                         /*
 669                          * If we have a single PMU interrupt that we can't shift,
 670                          * assume that we're running on a uniprocessor machine and
 671                          * continue. Otherwise, continue without this interrupt.
 672                          */
 673                         if (irq_set_affinity(irq, cpumask_of(cpu)) && irqs > 1) {
 674                                 pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
 675                                         irq, cpu);
 676                                 continue;
 677                         }
 678
 679                         err = request_irq(irq, handler,
 680                                           IRQF_NOBALANCING | IRQF_NO_THREAD, "arm-pmu",
 681                                           per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 682                         if (err) {
 683                                 pr_err("unable to request IRQ%d for ARM PMU counters\n",
 684                                         irq);
 685                                 return err;
 686                         }
 687
 688                         cpumask_set_cpu(cpu, &cpu_pmu->active_irqs);
 689                 }
 690         }
 691
 692         return 0;
 693 }
 694
 695 /*
 696  * PMU hardware loses all context when a CPU goes offline.
 697  * When a CPU is hotplugged back in, since some hardware registers are
 698  * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
 699  * junk values out of them.
 700  */
 701 static int cpu_pmu_notify(struct notifier_block *b, unsigned long action,
 702                           void *hcpu)
 703 {
 704         int cpu = (unsigned long)hcpu;
 705         struct arm_pmu *pmu = container_of(b, struct arm_pmu, hotplug_nb);
 706
 707         if ((action & ~CPU_TASKS_FROZEN) != CPU_STARTING)
 708                 return NOTIFY_DONE;
 709
 710         if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
 711                 return NOTIFY_DONE;
 712
 713         if (pmu->reset)
 714                 pmu->reset(pmu);
 715         else
 716                 return NOTIFY_DONE;
 717
 718         return NOTIFY_OK;
 719 }
 720
 721 static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
 722 {
 723         int err;
 724         int cpu;
 725         struct pmu_hw_events __percpu *cpu_hw_events;
 726
 727         cpu_hw_events = alloc_percpu(struct pmu_hw_events);
 728         if (!cpu_hw_events)
 729                 return -ENOMEM;
 730
 731         cpu_pmu->hotplug_nb.notifier_call = cpu_pmu_notify;
 732         err = register_cpu_notifier(&cpu_pmu->hotplug_nb);
 733         if (err)
 734                 goto out_hw_events;
 735
 736         for_each_possible_cpu(cpu) {
 737                 struct pmu_hw_events *events = per_cpu_ptr(cpu_hw_events, cpu);
 738                 raw_spin_lock_init(&events->pmu_lock);
 739                 events->percpu_pmu = cpu_pmu;
 740         }
 741
 742         cpu_pmu->hw_events      = cpu_hw_events;
 743         cpu_pmu->request_irq    = cpu_pmu_request_irq;
 744         cpu_pmu->free_irq       = cpu_pmu_free_irq;
 745
 746         /* Ensure the PMU has sane values out of reset. */
 747         if (cpu_pmu->reset)
 748                 on_each_cpu_mask(&cpu_pmu->supported_cpus, cpu_pmu->reset,
 749                          cpu_pmu, 1);
 750
 751         /* If no interrupts available, set the corresponding capability flag */
 752         if (!platform_get_irq(cpu_pmu->plat_device, 0))
 753                 cpu_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
 754
 755         return 0;
 756
 757 out_hw_events:
 758         free_percpu(cpu_hw_events);
 759         return err;
 760 }
 761
 762 static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
 763 {
 764         unregister_cpu_notifier(&cpu_pmu->hotplug_nb);
 765         free_percpu(cpu_pmu->hw_events);
 766 }
 767
 768 /*
 769  * CPU PMU identification and probing.
 770  */
 771 static int probe_current_pmu(struct arm_pmu *pmu,
 772                              const struct pmu_probe_info *info)
 773 {
 774         int cpu = get_cpu();
 775         unsigned int cpuid = read_cpuid_id();
 776         int ret = -ENODEV;
 777
 778         pr_info("probing PMU on CPU %d\n", cpu);
 779
 780         for (; info->init != NULL; info++) {
 781                 if ((cpuid & info->mask) != info->cpuid)
 782                         continue;
 783                 ret = info->init(pmu);
 784                 break;
 785         }
 786
 787         put_cpu();
 788         return ret;
 789 }
 790
 791 static int of_pmu_irq_cfg(struct arm_pmu *pmu)
 792 {
 793         int *irqs, i = 0;
 794         bool using_spi = false;
 795         struct platform_device *pdev = pmu->plat_device;
 796
 797         irqs = kcalloc(pdev->num_resources, sizeof(*irqs), GFP_KERNEL);
 798         if (!irqs)
 799                 return -ENOMEM;
 800
 801         do {
 802                 struct device_node *dn;
 803                 int cpu, irq;
 804
 805                 /* See if we have an affinity entry */
 806                 dn = of_parse_phandle(pdev->dev.of_node, "interrupt-affinity", i);
 807                 if (!dn)
 808                         break;
 809
 810                 /* Check the IRQ type and prohibit a mix of PPIs and SPIs */
 811                 irq = platform_get_irq(pdev, i);
 812                 if (irq >= 0) {
 813                         bool spi = !irq_is_percpu(irq);
 814
 815                         if (i > 0 && spi != using_spi) {
 816                                 pr_err("PPI/SPI IRQ type mismatch for %s!\n",
 817                                         dn->name);
 818                                 kfree(irqs);
 819                                 return -EINVAL;
 820                         }
 821
 822                         using_spi = spi;
 823                 }
 824
 825                 /* Now look up the logical CPU number */
 826                 for_each_possible_cpu(cpu)
 827                         if (dn == of_cpu_device_node_get(cpu))
 828                                 break;
 829
 830                 if (cpu >= nr_cpu_ids) {
 831                         pr_warn("Failed to find logical CPU for %s\n",
 832                                 dn->name);
 833                         of_node_put(dn);
 834                         cpumask_setall(&pmu->supported_cpus);
 835                         break;
 836                 }
 837                 of_node_put(dn);
 838
 839                 /* For SPIs, we need to track the affinity per IRQ */
 840                 if (using_spi) {
 841                         if (i >= pdev->num_resources) {
 842                                 of_node_put(dn);
 843                                 break;
 844                         }
 845
 846                         irqs[i] = cpu;
 847                 }
 848
 849                 /* Keep track of the CPUs containing this PMU type */
 850                 cpumask_set_cpu(cpu, &pmu->supported_cpus);
 851                 of_node_put(dn);
 852                 i++;
 853         } while (1);
 854
 855         /* If we didn't manage to parse anything, claim to support all CPUs */
 856         if (cpumask_weight(&pmu->supported_cpus) == 0)
 857                 cpumask_setall(&pmu->supported_cpus);
 858
 859         /* If we matched up the IRQ affinities, use them to route the SPIs */
 860         if (using_spi && i == pdev->num_resources)
 861                 pmu->irq_affinity = irqs;
 862         else
 863                 kfree(irqs);
 864
 865         return 0;
 866 }
 867
 868 int arm_pmu_device_probe(struct platform_device *pdev,
 869                          const struct of_device_id *of_table,
 870                          const struct pmu_probe_info *probe_table)
 871 {
 872         const struct of_device_id *of_id;
 873         const int (*init_fn)(struct arm_pmu *);
 874         struct device_node *node = pdev->dev.of_node;
 875         struct arm_pmu *pmu;
 876         int ret = -ENODEV;
 877
 878         pmu = kzalloc(sizeof(struct arm_pmu), GFP_KERNEL);
 879         if (!pmu) {
 880                 pr_info("failed to allocate PMU device!\n");
 881                 return -ENOMEM;
 882         }
 883
 884         if (!__oprofile_cpu_pmu)
 885                 __oprofile_cpu_pmu = pmu;
 886
 887         pmu->plat_device = pdev;
 888
 889         if (node && (of_id = of_match_node(of_table, pdev->dev.of_node))) {
 890                 init_fn = of_id->data;
 891
 892                 ret = of_pmu_irq_cfg(pmu);
 893                 if (!ret)
 894                         ret = init_fn(pmu);
 895         } else {
 896                 ret = probe_current_pmu(pmu, probe_table);
 897                 cpumask_setall(&pmu->supported_cpus);
 898         }
 899
 900         if (ret) {
 901                 pr_info("failed to probe PMU!\n");
 902                 goto out_free;
 903         }
 904
 905         ret = cpu_pmu_init(pmu);
 906         if (ret)
 907                 goto out_free;
 908
 909         ret = armpmu_register(pmu, -1);
 910         if (ret)
 911                 goto out_destroy;
 912
 913         return 0;
 914
 915 out_destroy:
 916         cpu_pmu_destroy(pmu);
 917 out_free:
 918         pr_info("failed to register PMU devices!\n");
 919         kfree(pmu);
 920         return ret;
 921 }