arch/x86/events/intel/uncore.c

   1 #include <linux/module.h>
   2
   3 #include <asm/cpu_device_id.h>
   4 #include <asm/intel-family.h>
   5 #include "uncore.h"
   6
   7 static struct intel_uncore_type *empty_uncore[] = { NULL, };
   8 struct intel_uncore_type **uncore_msr_uncores = empty_uncore;
   9 struct intel_uncore_type **uncore_pci_uncores = empty_uncore;
  10
  11 static bool pcidrv_registered;
  12 struct pci_driver *uncore_pci_driver;
  13 /* pci bus to socket mapping */
  14 DEFINE_RAW_SPINLOCK(pci2phy_map_lock);
  15 struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
  16 struct pci_extra_dev *uncore_extra_pci_dev;
  17 static int max_packages;
  18
  19 /* mask of cpus that collect uncore events */
  20 static cpumask_t uncore_cpu_mask;
  21
  22 /* constraint for the fixed counter */
  23 static struct event_constraint uncore_constraint_fixed =
  24         EVENT_CONSTRAINT(~0ULL, 1 << UNCORE_PMC_IDX_FIXED, ~0ULL);
  25 struct event_constraint uncore_constraint_empty =
  26         EVENT_CONSTRAINT(0, 0, 0);
  27
  28 MODULE_LICENSE("GPL");
  29
  30 static int uncore_pcibus_to_physid(struct pci_bus *bus)
  31 {
  32         struct pci2phy_map *map;
  33         int phys_id = -1;
  34
  35         raw_spin_lock(&pci2phy_map_lock);
  36         list_for_each_entry(map, &pci2phy_map_head, list) {
  37                 if (map->segment == pci_domain_nr(bus)) {
  38                         phys_id = map->pbus_to_physid[bus->number];
  39                         break;
  40                 }
  41         }
  42         raw_spin_unlock(&pci2phy_map_lock);
  43
  44         return phys_id;
  45 }
  46
  47 static void uncore_free_pcibus_map(void)
  48 {
  49         struct pci2phy_map *map, *tmp;
  50
  51         list_for_each_entry_safe(map, tmp, &pci2phy_map_head, list) {
  52                 list_del(&map->list);
  53                 kfree(map);
  54         }
  55 }
  56
  57 struct pci2phy_map *__find_pci2phy_map(int segment)
  58 {
  59         struct pci2phy_map *map, *alloc = NULL;
  60         int i;
  61
  62         lockdep_assert_held(&pci2phy_map_lock);
  63
  64 lookup:
  65         list_for_each_entry(map, &pci2phy_map_head, list) {
  66                 if (map->segment == segment)
  67                         goto end;
  68         }
  69
  70         if (!alloc) {
  71                 raw_spin_unlock(&pci2phy_map_lock);
  72                 alloc = kmalloc(sizeof(struct pci2phy_map), GFP_KERNEL);
  73                 raw_spin_lock(&pci2phy_map_lock);
  74
  75                 if (!alloc)
  76                         return NULL;
  77
  78                 goto lookup;
  79         }
  80
  81         map = alloc;
  82         alloc = NULL;
  83         map->segment = segment;
  84         for (i = 0; i < 256; i++)
  85                 map->pbus_to_physid[i] = -1;
  86         list_add_tail(&map->list, &pci2phy_map_head);
  87
  88 end:
  89         kfree(alloc);
  90         return map;
  91 }
  92
  93 ssize_t uncore_event_show(struct kobject *kobj,
  94                           struct kobj_attribute *attr, char *buf)
  95 {
  96         struct uncore_event_desc *event =
  97                 container_of(attr, struct uncore_event_desc, attr);
  98         return sprintf(buf, "%s", event->config);
  99 }
 100
 101 struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu)
 102 {
 103         return pmu->boxes[topology_logical_package_id(cpu)];
 104 }
 105
 106 u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event)
 107 {
 108         u64 count;
 109
 110         rdmsrl(event->hw.event_base, count);
 111
 112         return count;
 113 }
 114
 115 /*
 116  * generic get constraint function for shared match/mask registers.
 117  */
 118 struct event_constraint *
 119 uncore_get_constraint(struct intel_uncore_box *box, struct perf_event *event)
 120 {
 121         struct intel_uncore_extra_reg *er;
 122         struct hw_perf_event_extra *reg1 = &event->hw.extra_reg;
 123         struct hw_perf_event_extra *reg2 = &event->hw.branch_reg;
 124         unsigned long flags;
 125         bool ok = false;
 126
 127         /*
 128          * reg->alloc can be set due to existing state, so for fake box we
 129          * need to ignore this, otherwise we might fail to allocate proper
 130          * fake state for this extra reg constraint.
 131          */
 132         if (reg1->idx == EXTRA_REG_NONE ||
 133             (!uncore_box_is_fake(box) && reg1->alloc))
 134                 return NULL;
 135
 136         er = &box->shared_regs[reg1->idx];
 137         raw_spin_lock_irqsave(&er->lock, flags);
 138         if (!atomic_read(&er->ref) ||
 139             (er->config1 == reg1->config && er->config2 == reg2->config)) {
 140                 atomic_inc(&er->ref);
 141                 er->config1 = reg1->config;
 142                 er->config2 = reg2->config;
 143                 ok = true;
 144         }
 145         raw_spin_unlock_irqrestore(&er->lock, flags);
 146
 147         if (ok) {
 148                 if (!uncore_box_is_fake(box))
 149                         reg1->alloc = 1;
 150                 return NULL;
 151         }
 152
 153         return &uncore_constraint_empty;
 154 }
 155
 156 void uncore_put_constraint(struct intel_uncore_box *box, struct perf_event *event)
 157 {
 158         struct intel_uncore_extra_reg *er;
 159         struct hw_perf_event_extra *reg1 = &event->hw.extra_reg;
 160
 161         /*
 162          * Only put constraint if extra reg was actually allocated. Also
 163          * takes care of event which do not use an extra shared reg.
 164          *
 165          * Also, if this is a fake box we shouldn't touch any event state
 166          * (reg->alloc) and we don't care about leaving inconsistent box
 167          * state either since it will be thrown out.
 168          */
 169         if (uncore_box_is_fake(box) || !reg1->alloc)
 170                 return;
 171
 172         er = &box->shared_regs[reg1->idx];
 173         atomic_dec(&er->ref);
 174         reg1->alloc = 0;
 175 }
 176
 177 u64 uncore_shared_reg_config(struct intel_uncore_box *box, int idx)
 178 {
 179         struct intel_uncore_extra_reg *er;
 180         unsigned long flags;
 181         u64 config;
 182
 183         er = &box->shared_regs[idx];
 184
 185         raw_spin_lock_irqsave(&er->lock, flags);
 186         config = er->config;
 187         raw_spin_unlock_irqrestore(&er->lock, flags);
 188
 189         return config;
 190 }
 191
 192 static void uncore_assign_hw_event(struct intel_uncore_box *box,
 193                                    struct perf_event *event, int idx)
 194 {
 195         struct hw_perf_event *hwc = &event->hw;
 196
 197         hwc->idx = idx;
 198         hwc->last_tag = ++box->tags[idx];
 199
 200         if (hwc->idx == UNCORE_PMC_IDX_FIXED) {
 201                 hwc->event_base = uncore_fixed_ctr(box);
 202                 hwc->config_base = uncore_fixed_ctl(box);
 203                 return;
 204         }
 205
 206         hwc->config_base = uncore_event_ctl(box, hwc->idx);
 207         hwc->event_base  = uncore_perf_ctr(box, hwc->idx);
 208 }
 209
 210 void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *event)
 211 {
 212         u64 prev_count, new_count, delta;
 213         int shift;
 214
 215         if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
 216                 shift = 64 - uncore_fixed_ctr_bits(box);
 217         else
 218                 shift = 64 - uncore_perf_ctr_bits(box);
 219
 220         /* the hrtimer might modify the previous event value */
 221 again:
 222         prev_count = local64_read(&event->hw.prev_count);
 223         new_count = uncore_read_counter(box, event);
 224         if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
 225                 goto again;
 226
 227         delta = (new_count << shift) - (prev_count << shift);
 228         delta >>= shift;
 229
 230         local64_add(delta, &event->count);
 231 }
 232
 233 /*
 234  * The overflow interrupt is unavailable for SandyBridge-EP, is broken
 235  * for SandyBridge. So we use hrtimer to periodically poll the counter
 236  * to avoid overflow.
 237  */
 238 static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer)
 239 {
 240         struct intel_uncore_box *box;
 241         struct perf_event *event;
 242         unsigned long flags;
 243         int bit;
 244
 245         box = container_of(hrtimer, struct intel_uncore_box, hrtimer);
 246         if (!box->n_active || box->cpu != smp_processor_id())
 247                 return HRTIMER_NORESTART;
 248         /*
 249          * disable local interrupt to prevent uncore_pmu_event_start/stop
 250          * to interrupt the update process
 251          */
 252         local_irq_save(flags);
 253
 254         /*
 255          * handle boxes with an active event list as opposed to active
 256          * counters
 257          */
 258         list_for_each_entry(event, &box->active_list, active_entry) {
 259                 uncore_perf_event_update(box, event);
 260         }
 261
 262         for_each_set_bit(bit, box->active_mask, UNCORE_PMC_IDX_MAX)
 263                 uncore_perf_event_update(box, box->events[bit]);
 264
 265         local_irq_restore(flags);
 266
 267         hrtimer_forward_now(hrtimer, ns_to_ktime(box->hrtimer_duration));
 268         return HRTIMER_RESTART;
 269 }
 270
 271 void uncore_pmu_start_hrtimer(struct intel_uncore_box *box)
 272 {
 273         hrtimer_start(&box->hrtimer, ns_to_ktime(box->hrtimer_duration),
 274                       HRTIMER_MODE_REL_PINNED);
 275 }
 276
 277 void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box)
 278 {
 279         hrtimer_cancel(&box->hrtimer);
 280 }
 281
 282 static void uncore_pmu_init_hrtimer(struct intel_uncore_box *box)
 283 {
 284         hrtimer_init(&box->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 285         box->hrtimer.function = uncore_pmu_hrtimer;
 286 }
 287
 288 static struct intel_uncore_box *uncore_alloc_box(struct intel_uncore_type *type,
 289                                                  int node)
 290 {
 291         int i, size, numshared = type->num_shared_regs ;
 292         struct intel_uncore_box *box;
 293
 294         size = sizeof(*box) + numshared * sizeof(struct intel_uncore_extra_reg);
 295
 296         box = kzalloc_node(size, GFP_KERNEL, node);
 297         if (!box)
 298                 return NULL;
 299
 300         for (i = 0; i < numshared; i++)
 301                 raw_spin_lock_init(&box->shared_regs[i].lock);
 302
 303         uncore_pmu_init_hrtimer(box);
 304         box->cpu = -1;
 305         box->pci_phys_id = -1;
 306         box->pkgid = -1;
 307
 308         /* set default hrtimer timeout */
 309         box->hrtimer_duration = UNCORE_PMU_HRTIMER_INTERVAL;
 310
 311         INIT_LIST_HEAD(&box->active_list);
 312
 313         return box;
 314 }
 315
 316 /*
 317  * Using uncore_pmu_event_init pmu event_init callback
 318  * as a detection point for uncore events.
 319  */
 320 static int uncore_pmu_event_init(struct perf_event *event);
 321
 322 static bool is_box_event(struct intel_uncore_box *box, struct perf_event *event)
 323 {
 324         return &box->pmu->pmu == event->pmu;
 325 }
 326
 327 static int
 328 uncore_collect_events(struct intel_uncore_box *box, struct perf_event *leader,
 329                       bool dogrp)
 330 {
 331         struct perf_event *event;
 332         int n, max_count;
 333
 334         max_count = box->pmu->type->num_counters;
 335         if (box->pmu->type->fixed_ctl)
 336                 max_count++;
 337
 338         if (box->n_events >= max_count)
 339                 return -EINVAL;
 340
 341         n = box->n_events;
 342
 343         if (is_box_event(box, leader)) {
 344                 box->event_list[n] = leader;
 345                 n++;
 346         }
 347
 348         if (!dogrp)
 349                 return n;
 350
 351         list_for_each_entry(event, &leader->sibling_list, group_entry) {
 352                 if (!is_box_event(box, event) ||
 353                     event->state <= PERF_EVENT_STATE_OFF)
 354                         continue;
 355
 356                 if (n >= max_count)
 357                         return -EINVAL;
 358
 359                 box->event_list[n] = event;
 360                 n++;
 361         }
 362         return n;
 363 }
 364
 365 static struct event_constraint *
 366 uncore_get_event_constraint(struct intel_uncore_box *box, struct perf_event *event)
 367 {
 368         struct intel_uncore_type *type = box->pmu->type;
 369         struct event_constraint *c;
 370
 371         if (type->ops->get_constraint) {
 372                 c = type->ops->get_constraint(box, event);
 373                 if (c)
 374                         return c;
 375         }
 376
 377         if (event->attr.config == UNCORE_FIXED_EVENT)
 378                 return &uncore_constraint_fixed;
 379
 380         if (type->constraints) {
 381                 for_each_event_constraint(c, type->constraints) {
 382                         if ((event->hw.config & c->cmask) == c->code)
 383                                 return c;
 384                 }
 385         }
 386
 387         return &type->unconstrainted;
 388 }
 389
 390 static void uncore_put_event_constraint(struct intel_uncore_box *box,
 391                                         struct perf_event *event)
 392 {
 393         if (box->pmu->type->ops->put_constraint)
 394                 box->pmu->type->ops->put_constraint(box, event);
 395 }
 396
 397 static int uncore_assign_events(struct intel_uncore_box *box, int assign[], int n)
 398 {
 399         unsigned long used_mask[BITS_TO_LONGS(UNCORE_PMC_IDX_MAX)];
 400         struct event_constraint *c;
 401         int i, wmin, wmax, ret = 0;
 402         struct hw_perf_event *hwc;
 403
 404         bitmap_zero(used_mask, UNCORE_PMC_IDX_MAX);
 405
 406         for (i = 0, wmin = UNCORE_PMC_IDX_MAX, wmax = 0; i < n; i++) {
 407                 c = uncore_get_event_constraint(box, box->event_list[i]);
 408                 box->event_constraint[i] = c;
 409                 wmin = min(wmin, c->weight);
 410                 wmax = max(wmax, c->weight);
 411         }
 412
 413         /* fastpath, try to reuse previous register */
 414         for (i = 0; i < n; i++) {
 415                 hwc = &box->event_list[i]->hw;
 416                 c = box->event_constraint[i];
 417
 418                 /* never assigned */
 419                 if (hwc->idx == -1)
 420                         break;
 421
 422                 /* constraint still honored */
 423                 if (!test_bit(hwc->idx, c->idxmsk))
 424                         break;
 425
 426                 /* not already used */
 427                 if (test_bit(hwc->idx, used_mask))
 428                         break;
 429
 430                 __set_bit(hwc->idx, used_mask);
 431                 if (assign)
 432                         assign[i] = hwc->idx;
 433         }
 434         /* slow path */
 435         if (i != n)
 436                 ret = perf_assign_events(box->event_constraint, n,
 437                                          wmin, wmax, n, assign);
 438
 439         if (!assign || ret) {
 440                 for (i = 0; i < n; i++)
 441                         uncore_put_event_constraint(box, box->event_list[i]);
 442         }
 443         return ret ? -EINVAL : 0;
 444 }
 445
 446 static void uncore_pmu_event_start(struct perf_event *event, int flags)
 447 {
 448         struct intel_uncore_box *box = uncore_event_to_box(event);
 449         int idx = event->hw.idx;
 450
 451         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
 452                 return;
 453
 454         if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
 455                 return;
 456
 457         event->hw.state = 0;
 458         box->events[idx] = event;
 459         box->n_active++;
 460         __set_bit(idx, box->active_mask);
 461
 462         local64_set(&event->hw.prev_count, uncore_read_counter(box, event));
 463         uncore_enable_event(box, event);
 464
 465         if (box->n_active == 1) {
 466                 uncore_enable_box(box);
 467                 uncore_pmu_start_hrtimer(box);
 468         }
 469 }
 470
 471 static void uncore_pmu_event_stop(struct perf_event *event, int flags)
 472 {
 473         struct intel_uncore_box *box = uncore_event_to_box(event);
 474         struct hw_perf_event *hwc = &event->hw;
 475
 476         if (__test_and_clear_bit(hwc->idx, box->active_mask)) {
 477                 uncore_disable_event(box, event);
 478                 box->n_active--;
 479                 box->events[hwc->idx] = NULL;
 480                 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
 481                 hwc->state |= PERF_HES_STOPPED;
 482
 483                 if (box->n_active == 0) {
 484                         uncore_disable_box(box);
 485                         uncore_pmu_cancel_hrtimer(box);
 486                 }
 487         }
 488
 489         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
 490                 /*
 491                  * Drain the remaining delta count out of a event
 492                  * that we are disabling:
 493                  */
 494                 uncore_perf_event_update(box, event);
 495                 hwc->state |= PERF_HES_UPTODATE;
 496         }
 497 }
 498
 499 static int uncore_pmu_event_add(struct perf_event *event, int flags)
 500 {
 501         struct intel_uncore_box *box = uncore_event_to_box(event);
 502         struct hw_perf_event *hwc = &event->hw;
 503         int assign[UNCORE_PMC_IDX_MAX];
 504         int i, n, ret;
 505
 506         if (!box)
 507                 return -ENODEV;
 508
 509         ret = n = uncore_collect_events(box, event, false);
 510         if (ret < 0)
 511                 return ret;
 512
 513         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 514         if (!(flags & PERF_EF_START))
 515                 hwc->state |= PERF_HES_ARCH;
 516
 517         ret = uncore_assign_events(box, assign, n);
 518         if (ret)
 519                 return ret;
 520
 521         /* save events moving to new counters */
 522         for (i = 0; i < box->n_events; i++) {
 523                 event = box->event_list[i];
 524                 hwc = &event->hw;
 525
 526                 if (hwc->idx == assign[i] &&
 527                         hwc->last_tag == box->tags[assign[i]])
 528                         continue;
 529                 /*
 530                  * Ensure we don't accidentally enable a stopped
 531                  * counter simply because we rescheduled.
 532                  */
 533                 if (hwc->state & PERF_HES_STOPPED)
 534                         hwc->state |= PERF_HES_ARCH;
 535
 536                 uncore_pmu_event_stop(event, PERF_EF_UPDATE);
 537         }
 538
 539         /* reprogram moved events into new counters */
 540         for (i = 0; i < n; i++) {
 541                 event = box->event_list[i];
 542                 hwc = &event->hw;
 543
 544                 if (hwc->idx != assign[i] ||
 545                         hwc->last_tag != box->tags[assign[i]])
 546                         uncore_assign_hw_event(box, event, assign[i]);
 547                 else if (i < box->n_events)
 548                         continue;
 549
 550                 if (hwc->state & PERF_HES_ARCH)
 551                         continue;
 552
 553                 uncore_pmu_event_start(event, 0);
 554         }
 555         box->n_events = n;
 556
 557         return 0;
 558 }
 559
 560 static void uncore_pmu_event_del(struct perf_event *event, int flags)
 561 {
 562         struct intel_uncore_box *box = uncore_event_to_box(event);
 563         int i;
 564
 565         uncore_pmu_event_stop(event, PERF_EF_UPDATE);
 566
 567         for (i = 0; i < box->n_events; i++) {
 568                 if (event == box->event_list[i]) {
 569                         uncore_put_event_constraint(box, event);
 570
 571                         for (++i; i < box->n_events; i++)
 572                                 box->event_list[i - 1] = box->event_list[i];
 573
 574                         --box->n_events;
 575                         break;
 576                 }
 577         }
 578
 579         event->hw.idx = -1;
 580         event->hw.last_tag = ~0ULL;
 581 }
 582
 583 void uncore_pmu_event_read(struct perf_event *event)
 584 {
 585         struct intel_uncore_box *box = uncore_event_to_box(event);
 586         uncore_perf_event_update(box, event);
 587 }
 588
 589 /*
 590  * validation ensures the group can be loaded onto the
 591  * PMU if it was the only group available.
 592  */
 593 static int uncore_validate_group(struct intel_uncore_pmu *pmu,
 594                                 struct perf_event *event)
 595 {
 596         struct perf_event *leader = event->group_leader;
 597         struct intel_uncore_box *fake_box;
 598         int ret = -EINVAL, n;
 599
 600         fake_box = uncore_alloc_box(pmu->type, NUMA_NO_NODE);
 601         if (!fake_box)
 602                 return -ENOMEM;
 603
 604         fake_box->pmu = pmu;
 605         /*
 606          * the event is not yet connected with its
 607          * siblings therefore we must first collect
 608          * existing siblings, then add the new event
 609          * before we can simulate the scheduling
 610          */
 611         n = uncore_collect_events(fake_box, leader, true);
 612         if (n < 0)
 613                 goto out;
 614
 615         fake_box->n_events = n;
 616         n = uncore_collect_events(fake_box, event, false);
 617         if (n < 0)
 618                 goto out;
 619
 620         fake_box->n_events = n;
 621
 622         ret = uncore_assign_events(fake_box, NULL, n);
 623 out:
 624         kfree(fake_box);
 625         return ret;
 626 }
 627
 628 static int uncore_pmu_event_init(struct perf_event *event)
 629 {
 630         struct intel_uncore_pmu *pmu;
 631         struct intel_uncore_box *box;
 632         struct hw_perf_event *hwc = &event->hw;
 633         int ret;
 634
 635         if (event->attr.type != event->pmu->type)
 636                 return -ENOENT;
 637
 638         pmu = uncore_event_to_pmu(event);
 639         /* no device found for this pmu */
 640         if (pmu->func_id < 0)
 641                 return -ENOENT;
 642
 643         /*
 644          * Uncore PMU does measure at all privilege level all the time.
 645          * So it doesn't make sense to specify any exclude bits.
 646          */
 647         if (event->attr.exclude_user || event->attr.exclude_kernel ||
 648                         event->attr.exclude_hv || event->attr.exclude_idle)
 649                 return -EINVAL;
 650
 651         /* Sampling not supported yet */
 652         if (hwc->sample_period)
 653                 return -EINVAL;
 654
 655         /*
 656          * Place all uncore events for a particular physical package
 657          * onto a single cpu
 658          */
 659         if (event->cpu < 0)
 660                 return -EINVAL;
 661         box = uncore_pmu_to_box(pmu, event->cpu);
 662         if (!box || box->cpu < 0)
 663                 return -EINVAL;
 664         event->cpu = box->cpu;
 665         event->pmu_private = box;
 666
 667         event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG;
 668
 669         event->hw.idx = -1;
 670         event->hw.last_tag = ~0ULL;
 671         event->hw.extra_reg.idx = EXTRA_REG_NONE;
 672         event->hw.branch_reg.idx = EXTRA_REG_NONE;
 673
 674         if (event->attr.config == UNCORE_FIXED_EVENT) {
 675                 /* no fixed counter */
 676                 if (!pmu->type->fixed_ctl)
 677                         return -EINVAL;
 678                 /*
 679                  * if there is only one fixed counter, only the first pmu
 680                  * can access the fixed counter
 681                  */
 682                 if (pmu->type->single_fixed && pmu->pmu_idx > 0)
 683                         return -EINVAL;
 684
 685                 /* fixed counters have event field hardcoded to zero */
 686                 hwc->config = 0ULL;
 687         } else {
 688                 hwc->config = event->attr.config &
 689                               (pmu->type->event_mask | ((u64)pmu->type->event_mask_ext << 32));
 690                 if (pmu->type->ops->hw_config) {
 691                         ret = pmu->type->ops->hw_config(box, event);
 692                         if (ret)
 693                                 return ret;
 694                 }
 695         }
 696
 697         if (event->group_leader != event)
 698                 ret = uncore_validate_group(pmu, event);
 699         else
 700                 ret = 0;
 701
 702         return ret;
 703 }
 704
 705 static ssize_t uncore_get_attr_cpumask(struct device *dev,
 706                                 struct device_attribute *attr, char *buf)
 707 {
 708         return cpumap_print_to_pagebuf(true, buf, &uncore_cpu_mask);
 709 }
 710
 711 static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 712
 713 static struct attribute *uncore_pmu_attrs[] = {
 714         &dev_attr_cpumask.attr,
 715         NULL,
 716 };
 717
 718 static struct attribute_group uncore_pmu_attr_group = {
 719         .attrs = uncore_pmu_attrs,
 720 };
 721
 722 static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
 723 {
 724         int ret;
 725
 726         if (!pmu->type->pmu) {
 727                 pmu->pmu = (struct pmu) {
 728                         .attr_groups    = pmu->type->attr_groups,
 729                         .task_ctx_nr    = perf_invalid_context,
 730                         .event_init     = uncore_pmu_event_init,
 731                         .add            = uncore_pmu_event_add,
 732                         .del            = uncore_pmu_event_del,
 733                         .start          = uncore_pmu_event_start,
 734                         .stop           = uncore_pmu_event_stop,
 735                         .read           = uncore_pmu_event_read,
 736                         .module         = THIS_MODULE,
 737                 };
 738         } else {
 739                 pmu->pmu = *pmu->type->pmu;
 740                 pmu->pmu.attr_groups = pmu->type->attr_groups;
 741         }
 742
 743         if (pmu->type->num_boxes == 1) {
 744                 if (strlen(pmu->type->name) > 0)
 745                         sprintf(pmu->name, "uncore_%s", pmu->type->name);
 746                 else
 747                         sprintf(pmu->name, "uncore");
 748         } else {
 749                 sprintf(pmu->name, "uncore_%s_%d", pmu->type->name,
 750                         pmu->pmu_idx);
 751         }
 752
 753         ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
 754         if (!ret)
 755                 pmu->registered = true;
 756         return ret;
 757 }
 758
 759 static void uncore_pmu_unregister(struct intel_uncore_pmu *pmu)
 760 {
 761         if (!pmu->registered)
 762                 return;
 763         perf_pmu_unregister(&pmu->pmu);
 764         pmu->registered = false;
 765 }
 766
 767 static void __uncore_exit_boxes(struct intel_uncore_type *type, int cpu)
 768 {
 769         struct intel_uncore_pmu *pmu = type->pmus;
 770         struct intel_uncore_box *box;
 771         int i, pkg;
 772
 773         if (pmu) {
 774                 pkg = topology_physical_package_id(cpu);
 775                 for (i = 0; i < type->num_boxes; i++, pmu++) {
 776                         box = pmu->boxes[pkg];
 777                         if (box)
 778                                 uncore_box_exit(box);
 779                 }
 780         }
 781 }
 782
 783 static void uncore_exit_boxes(void *dummy)
 784 {
 785         struct intel_uncore_type **types;
 786
 787         for (types = uncore_msr_uncores; *types; types++)
 788                 __uncore_exit_boxes(*types++, smp_processor_id());
 789 }
 790
 791 static void uncore_free_boxes(struct intel_uncore_pmu *pmu)
 792 {
 793         int pkg;
 794
 795         for (pkg = 0; pkg < max_packages; pkg++)
 796                 kfree(pmu->boxes[pkg]);
 797         kfree(pmu->boxes);
 798 }
 799
 800 static void uncore_type_exit(struct intel_uncore_type *type)
 801 {
 802         struct intel_uncore_pmu *pmu = type->pmus;
 803         int i;
 804
 805         if (pmu) {
 806                 for (i = 0; i < type->num_boxes; i++, pmu++) {
 807                         uncore_pmu_unregister(pmu);
 808                         uncore_free_boxes(pmu);
 809                 }
 810                 kfree(type->pmus);
 811                 type->pmus = NULL;
 812         }
 813         kfree(type->events_group);
 814         type->events_group = NULL;
 815 }
 816
 817 static void uncore_types_exit(struct intel_uncore_type **types)
 818 {
 819         for (; *types; types++)
 820                 uncore_type_exit(*types);
 821 }
 822
 823 static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 824 {
 825         struct intel_uncore_pmu *pmus;
 826         struct attribute_group *attr_group;
 827         struct attribute **attrs;
 828         size_t size;
 829         int i, j;
 830
 831         pmus = kzalloc(sizeof(*pmus) * type->num_boxes, GFP_KERNEL);
 832         if (!pmus)
 833                 return -ENOMEM;
 834
 835         size = max_packages * sizeof(struct intel_uncore_box *);
 836
 837         for (i = 0; i < type->num_boxes; i++) {
 838                 pmus[i].func_id = setid ? i : -1;
 839                 pmus[i].pmu_idx = i;
 840                 pmus[i].type    = type;
 841                 pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
 842                 if (!pmus[i].boxes)
 843                         return -ENOMEM;
 844         }
 845
 846         type->pmus = pmus;
 847         type->unconstrainted = (struct event_constraint)
 848                 __EVENT_CONSTRAINT(0, (1ULL << type->num_counters) - 1,
 849                                 0, type->num_counters, 0, 0);
 850
 851         if (type->event_descs) {
 852                 for (i = 0; type->event_descs[i].attr.attr.name; i++);
 853
 854                 attr_group = kzalloc(sizeof(struct attribute *) * (i + 1) +
 855                                         sizeof(*attr_group), GFP_KERNEL);
 856                 if (!attr_group)
 857                         return -ENOMEM;
 858
 859                 attrs = (struct attribute **)(attr_group + 1);
 860                 attr_group->name = "events";
 861                 attr_group->attrs = attrs;
 862
 863                 for (j = 0; j < i; j++)
 864                         attrs[j] = &type->event_descs[j].attr.attr;
 865
 866                 type->events_group = attr_group;
 867         }
 868
 869         type->pmu_group = &uncore_pmu_attr_group;
 870         return 0;
 871 }
 872
 873 static int __init
 874 uncore_types_init(struct intel_uncore_type **types, bool setid)
 875 {
 876         int ret;
 877
 878         for (; *types; types++) {
 879                 ret = uncore_type_init(*types, setid);
 880                 if (ret)
 881                         return ret;
 882         }
 883         return 0;
 884 }
 885
 886 /*
 887  * add a pci uncore device
 888  */
 889 static int uncore_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 890 {
 891         struct intel_uncore_type *type;
 892         struct intel_uncore_pmu *pmu = NULL;
 893         struct intel_uncore_box *box;
 894         int phys_id, pkg, ret;
 895
 896         phys_id = uncore_pcibus_to_physid(pdev->bus);
 897         if (phys_id < 0)
 898                 return -ENODEV;
 899
 900         pkg = topology_phys_to_logical_pkg(phys_id);
 901         if (pkg < 0)
 902                 return -EINVAL;
 903
 904         if (UNCORE_PCI_DEV_TYPE(id->driver_data) == UNCORE_EXTRA_PCI_DEV) {
 905                 int idx = UNCORE_PCI_DEV_IDX(id->driver_data);
 906
 907                 uncore_extra_pci_dev[pkg].dev[idx] = pdev;
 908                 pci_set_drvdata(pdev, NULL);
 909                 return 0;
 910         }
 911
 912         type = uncore_pci_uncores[UNCORE_PCI_DEV_TYPE(id->driver_data)];
 913
 914         /*
 915          * Some platforms, e.g.  Knights Landing, use a common PCI device ID
 916          * for multiple instances of an uncore PMU device type. We should check
 917          * PCI slot and func to indicate the uncore box.
 918          */
 919         if (id->driver_data & ~0xffff) {
 920                 struct pci_driver *pci_drv = pdev->driver;
 921                 const struct pci_device_id *ids = pci_drv->id_table;
 922                 unsigned int devfn;
 923
 924                 while (ids && ids->vendor) {
 925                         if ((ids->vendor == pdev->vendor) &&
 926                             (ids->device == pdev->device)) {
 927                                 devfn = PCI_DEVFN(UNCORE_PCI_DEV_DEV(ids->driver_data),
 928                                                   UNCORE_PCI_DEV_FUNC(ids->driver_data));
 929                                 if (devfn == pdev->devfn) {
 930                                         pmu = &type->pmus[UNCORE_PCI_DEV_IDX(ids->driver_data)];
 931                                         break;
 932                                 }
 933                         }
 934                         ids++;
 935                 }
 936                 if (pmu == NULL)
 937                         return -ENODEV;
 938         } else {
 939                 /*
 940                  * for performance monitoring unit with multiple boxes,
 941                  * each box has a different function id.
 942                  */
 943                 pmu = &type->pmus[UNCORE_PCI_DEV_IDX(id->driver_data)];
 944         }
 945
 946         if (WARN_ON_ONCE(pmu->boxes[pkg] != NULL))
 947                 return -EINVAL;
 948
 949         box = uncore_alloc_box(type, NUMA_NO_NODE);
 950         if (!box)
 951                 return -ENOMEM;
 952
 953         if (pmu->func_id < 0)
 954                 pmu->func_id = pdev->devfn;
 955         else
 956                 WARN_ON_ONCE(pmu->func_id != pdev->devfn);
 957
 958         atomic_inc(&box->refcnt);
 959         box->pci_phys_id = phys_id;
 960         box->pkgid = pkg;
 961         box->pci_dev = pdev;
 962         box->pmu = pmu;
 963         uncore_box_init(box);
 964         pci_set_drvdata(pdev, box);
 965
 966         pmu->boxes[pkg] = box;
 967         if (atomic_inc_return(&pmu->activeboxes) > 1)
 968                 return 0;
 969
 970         /* First active box registers the pmu */
 971         ret = uncore_pmu_register(pmu);
 972         if (ret) {
 973                 pci_set_drvdata(pdev, NULL);
 974                 pmu->boxes[pkg] = NULL;
 975                 uncore_box_exit(box);
 976                 kfree(box);
 977         }
 978         return ret;
 979 }
 980
 981 static void uncore_pci_remove(struct pci_dev *pdev)
 982 {
 983         struct intel_uncore_box *box;
 984         struct intel_uncore_pmu *pmu;
 985         int i, phys_id, pkg;
 986
 987         phys_id = uncore_pcibus_to_physid(pdev->bus);
 988         pkg = topology_phys_to_logical_pkg(phys_id);
 989
 990         box = pci_get_drvdata(pdev);
 991         if (!box) {
 992                 for (i = 0; i < UNCORE_EXTRA_PCI_DEV_MAX; i++) {
 993                         if (uncore_extra_pci_dev[pkg].dev[i] == pdev) {
 994                                 uncore_extra_pci_dev[pkg].dev[i] = NULL;
 995                                 break;
 996                         }
 997                 }
 998                 WARN_ON_ONCE(i >= UNCORE_EXTRA_PCI_DEV_MAX);
 999                 return;
1000         }
1001
1002         pmu = box->pmu;
1003         if (WARN_ON_ONCE(phys_id != box->pci_phys_id))
1004                 return;
1005
1006         pci_set_drvdata(pdev, NULL);
1007         pmu->boxes[pkg] = NULL;
1008         if (atomic_dec_return(&pmu->activeboxes) == 0)
1009                 uncore_pmu_unregister(pmu);
1010         uncore_box_exit(box);
1011         kfree(box);
1012 }
1013
1014 static int __init uncore_pci_init(void)
1015 {
1016         size_t size;
1017         int ret;
1018
1019         size = max_packages * sizeof(struct pci_extra_dev);
1020         uncore_extra_pci_dev = kzalloc(size, GFP_KERNEL);
1021         if (!uncore_extra_pci_dev) {
1022                 ret = -ENOMEM;
1023                 goto err;
1024         }
1025
1026         ret = uncore_types_init(uncore_pci_uncores, false);
1027         if (ret)
1028                 goto errtype;
1029
1030         uncore_pci_driver->probe = uncore_pci_probe;
1031         uncore_pci_driver->remove = uncore_pci_remove;
1032
1033         ret = pci_register_driver(uncore_pci_driver);
1034         if (ret)
1035                 goto errtype;
1036
1037         pcidrv_registered = true;
1038         return 0;
1039
1040 errtype:
1041         uncore_types_exit(uncore_pci_uncores);
1042         kfree(uncore_extra_pci_dev);
1043         uncore_extra_pci_dev = NULL;
1044         uncore_free_pcibus_map();
1045 err:
1046         uncore_pci_uncores = empty_uncore;
1047         return ret;
1048 }
1049
1050 static void uncore_pci_exit(void)
1051 {
1052         if (pcidrv_registered) {
1053                 pcidrv_registered = false;
1054                 pci_unregister_driver(uncore_pci_driver);
1055                 uncore_types_exit(uncore_pci_uncores);
1056                 kfree(uncore_extra_pci_dev);
1057                 uncore_free_pcibus_map();
1058         }
1059 }
1060
1061 static int uncore_cpu_dying(unsigned int cpu)
1062 {
1063         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1064         struct intel_uncore_pmu *pmu;
1065         struct intel_uncore_box *box;
1066         int i, pkg;
1067
1068         pkg = topology_logical_package_id(cpu);
1069         for (; *types; types++) {
1070                 type = *types;
1071                 pmu = type->pmus;
1072                 for (i = 0; i < type->num_boxes; i++, pmu++) {
1073                         box = pmu->boxes[pkg];
1074                         if (box && atomic_dec_return(&box->refcnt) == 0)
1075                                 uncore_box_exit(box);
1076                 }
1077         }
1078         return 0;
1079 }
1080
1081 static int first_init;
1082
1083 static int uncore_cpu_starting(unsigned int cpu)
1084 {
1085         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1086         struct intel_uncore_pmu *pmu;
1087         struct intel_uncore_box *box;
1088         int i, pkg, ncpus = 1;
1089
1090         if (first_init) {
1091                 /*
1092                  * On init we get the number of online cpus in the package
1093                  * and set refcount for all of them.
1094                  */
1095                 ncpus = cpumask_weight(topology_core_cpumask(cpu));
1096         }
1097
1098         pkg = topology_logical_package_id(cpu);
1099         for (; *types; types++) {
1100                 type = *types;
1101                 pmu = type->pmus;
1102                 for (i = 0; i < type->num_boxes; i++, pmu++) {
1103                         box = pmu->boxes[pkg];
1104                         if (!box)
1105                                 continue;
1106                         /* The first cpu on a package activates the box */
1107                         if (atomic_add_return(ncpus, &box->refcnt) == ncpus)
1108                                 uncore_box_init(box);
1109                 }
1110         }
1111
1112         return 0;
1113 }
1114
1115 static int uncore_cpu_prepare(unsigned int cpu)
1116 {
1117         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1118         struct intel_uncore_pmu *pmu;
1119         struct intel_uncore_box *box;
1120         int i, pkg;
1121
1122         pkg = topology_logical_package_id(cpu);
1123         for (; *types; types++) {
1124                 type = *types;
1125                 pmu = type->pmus;
1126                 for (i = 0; i < type->num_boxes; i++, pmu++) {
1127                         if (pmu->boxes[pkg])
1128                                 continue;
1129                         /* First cpu of a package allocates the box */
1130                         box = uncore_alloc_box(type, cpu_to_node(cpu));
1131                         if (!box)
1132                                 return -ENOMEM;
1133                         box->pmu = pmu;
1134                         box->pkgid = pkg;
1135                         pmu->boxes[pkg] = box;
1136                 }
1137         }
1138         return 0;
1139 }
1140
1141 static void uncore_change_type_ctx(struct intel_uncore_type *type, int old_cpu,
1142                                    int new_cpu)
1143 {
1144         struct intel_uncore_pmu *pmu = type->pmus;
1145         struct intel_uncore_box *box;
1146         int i, pkg;
1147
1148         pkg = topology_logical_package_id(old_cpu < 0 ? new_cpu : old_cpu);
1149         for (i = 0; i < type->num_boxes; i++, pmu++) {
1150                 box = pmu->boxes[pkg];
1151                 if (!box)
1152                         continue;
1153
1154                 if (old_cpu < 0) {
1155                         WARN_ON_ONCE(box->cpu != -1);
1156                         box->cpu = new_cpu;
1157                         continue;
1158                 }
1159
1160                 WARN_ON_ONCE(box->cpu != old_cpu);
1161                 box->cpu = -1;
1162                 if (new_cpu < 0)
1163                         continue;
1164
1165                 uncore_pmu_cancel_hrtimer(box);
1166                 perf_pmu_migrate_context(&pmu->pmu, old_cpu, new_cpu);
1167                 box->cpu = new_cpu;
1168         }
1169 }
1170
1171 static void uncore_change_context(struct intel_uncore_type **uncores,
1172                                   int old_cpu, int new_cpu)
1173 {
1174         for (; *uncores; uncores++)
1175                 uncore_change_type_ctx(*uncores, old_cpu, new_cpu);
1176 }
1177
1178 static int uncore_event_cpu_offline(unsigned int cpu)
1179 {
1180         int target;
1181
1182         /* Check if exiting cpu is used for collecting uncore events */
1183         if (!cpumask_test_and_clear_cpu(cpu, &uncore_cpu_mask))
1184                 return 0;
1185
1186         /* Find a new cpu to collect uncore events */
1187         target = cpumask_any_but(topology_core_cpumask(cpu), cpu);
1188
1189         /* Migrate uncore events to the new target */
1190         if (target < nr_cpu_ids)
1191                 cpumask_set_cpu(target, &uncore_cpu_mask);
1192         else
1193                 target = -1;
1194
1195         uncore_change_context(uncore_msr_uncores, cpu, target);
1196         uncore_change_context(uncore_pci_uncores, cpu, target);
1197         return 0;
1198 }
1199
1200 static int uncore_event_cpu_online(unsigned int cpu)
1201 {
1202         int target;
1203
1204         /*
1205          * Check if there is an online cpu in the package
1206          * which collects uncore events already.
1207          */
1208         target = cpumask_any_and(&uncore_cpu_mask, topology_core_cpumask(cpu));
1209         if (target < nr_cpu_ids)
1210                 return 0;
1211
1212         cpumask_set_cpu(cpu, &uncore_cpu_mask);
1213
1214         uncore_change_context(uncore_msr_uncores, -1, cpu);
1215         uncore_change_context(uncore_pci_uncores, -1, cpu);
1216         return 0;
1217 }
1218
1219 static int __init type_pmu_register(struct intel_uncore_type *type)
1220 {
1221         int i, ret;
1222
1223         for (i = 0; i < type->num_boxes; i++) {
1224                 ret = uncore_pmu_register(&type->pmus[i]);
1225                 if (ret)
1226                         return ret;
1227         }
1228         return 0;
1229 }
1230
1231 static int __init uncore_msr_pmus_register(void)
1232 {
1233         struct intel_uncore_type **types = uncore_msr_uncores;
1234         int ret;
1235
1236         for (; *types; types++) {
1237                 ret = type_pmu_register(*types);
1238                 if (ret)
1239                         return ret;
1240         }
1241         return 0;
1242 }
1243
1244 static int __init uncore_cpu_init(void)
1245 {
1246         int ret;
1247
1248         ret = uncore_types_init(uncore_msr_uncores, true);
1249         if (ret)
1250                 goto err;
1251
1252         ret = uncore_msr_pmus_register();
1253         if (ret)
1254                 goto err;
1255         return 0;
1256 err:
1257         uncore_types_exit(uncore_msr_uncores);
1258         uncore_msr_uncores = empty_uncore;
1259         return ret;
1260 }
1261
1262 #define X86_UNCORE_MODEL_MATCH(model, init)     \
1263         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long)&init }
1264
1265 struct intel_uncore_init_fun {
1266         void    (*cpu_init)(void);
1267         int     (*pci_init)(void);
1268 };
1269
1270 static const struct intel_uncore_init_fun nhm_uncore_init __initconst = {
1271         .cpu_init = nhm_uncore_cpu_init,
1272 };
1273
1274 static const struct intel_uncore_init_fun snb_uncore_init __initconst = {
1275         .cpu_init = snb_uncore_cpu_init,
1276         .pci_init = snb_uncore_pci_init,
1277 };
1278
1279 static const struct intel_uncore_init_fun ivb_uncore_init __initconst = {
1280         .cpu_init = snb_uncore_cpu_init,
1281         .pci_init = ivb_uncore_pci_init,
1282 };
1283
1284 static const struct intel_uncore_init_fun hsw_uncore_init __initconst = {
1285         .cpu_init = snb_uncore_cpu_init,
1286         .pci_init = hsw_uncore_pci_init,
1287 };
1288
1289 static const struct intel_uncore_init_fun bdw_uncore_init __initconst = {
1290         .cpu_init = snb_uncore_cpu_init,
1291         .pci_init = bdw_uncore_pci_init,
1292 };
1293
1294 static const struct intel_uncore_init_fun snbep_uncore_init __initconst = {
1295         .cpu_init = snbep_uncore_cpu_init,
1296         .pci_init = snbep_uncore_pci_init,
1297 };
1298
1299 static const struct intel_uncore_init_fun nhmex_uncore_init __initconst = {
1300         .cpu_init = nhmex_uncore_cpu_init,
1301 };
1302
1303 static const struct intel_uncore_init_fun ivbep_uncore_init __initconst = {
1304         .cpu_init = ivbep_uncore_cpu_init,
1305         .pci_init = ivbep_uncore_pci_init,
1306 };
1307
1308 static const struct intel_uncore_init_fun hswep_uncore_init __initconst = {
1309         .cpu_init = hswep_uncore_cpu_init,
1310         .pci_init = hswep_uncore_pci_init,
1311 };
1312
1313 static const struct intel_uncore_init_fun bdx_uncore_init __initconst = {
1314         .cpu_init = bdx_uncore_cpu_init,
1315         .pci_init = bdx_uncore_pci_init,
1316 };
1317
1318 static const struct intel_uncore_init_fun knl_uncore_init __initconst = {
1319         .cpu_init = knl_uncore_cpu_init,
1320         .pci_init = knl_uncore_pci_init,
1321 };
1322
1323 static const struct intel_uncore_init_fun skl_uncore_init __initconst = {
1324         .cpu_init = skl_uncore_cpu_init,
1325         .pci_init = skl_uncore_pci_init,
1326 };
1327
1328 static const struct intel_uncore_init_fun skx_uncore_init __initconst = {
1329         .cpu_init = skx_uncore_cpu_init,
1330         .pci_init = skx_uncore_pci_init,
1331 };
1332
1333 static const struct x86_cpu_id intel_uncore_match[] __initconst = {
1334         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_NEHALEM_EP,     nhm_uncore_init),
1335         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_NEHALEM,        nhm_uncore_init),
1336         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_WESTMERE,       nhm_uncore_init),
1337         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_WESTMERE_EP,    nhm_uncore_init),
1338         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_SANDYBRIDGE,    snb_uncore_init),
1339         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_IVYBRIDGE,      ivb_uncore_init),
1340         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_HASWELL_CORE,   hsw_uncore_init),
1341         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_HASWELL_ULT,    hsw_uncore_init),
1342         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_HASWELL_GT3E,   hsw_uncore_init),
1343         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_BROADWELL_CORE, bdw_uncore_init),
1344         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_BROADWELL_GT3E, bdw_uncore_init),
1345         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_SANDYBRIDGE_X,  snbep_uncore_init),
1346         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_NEHALEM_EX,     nhmex_uncore_init),
1347         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_WESTMERE_EX,    nhmex_uncore_init),
1348         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_IVYBRIDGE_X,    ivbep_uncore_init),
1349         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_HASWELL_X,      hswep_uncore_init),
1350         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_BROADWELL_X,    bdx_uncore_init),
1351         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_BROADWELL_XEON_D, bdx_uncore_init),
1352         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_XEON_PHI_KNL,   knl_uncore_init),
1353         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_XEON_PHI_KNM,   knl_uncore_init),
1354         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_SKYLAKE_DESKTOP,skl_uncore_init),
1355         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_SKYLAKE_MOBILE, skl_uncore_init),
1356         X86_UNCORE_MODEL_MATCH(INTEL_FAM6_SKYLAKE_X,      skx_uncore_init),
1357         {},
1358 };
1359
1360 MODULE_DEVICE_TABLE(x86cpu, intel_uncore_match);
1361
1362 static int __init intel_uncore_init(void)
1363 {
1364         const struct x86_cpu_id *id;
1365         struct intel_uncore_init_fun *uncore_init;
1366         int pret = 0, cret = 0, ret;
1367
1368         id = x86_match_cpu(intel_uncore_match);
1369         if (!id)
1370                 return -ENODEV;
1371
1372         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
1373                 return -ENODEV;
1374
1375         max_packages = topology_max_packages();
1376
1377         uncore_init = (struct intel_uncore_init_fun *)id->driver_data;
1378         if (uncore_init->pci_init) {
1379                 pret = uncore_init->pci_init();
1380                 if (!pret)
1381                         pret = uncore_pci_init();
1382         }
1383
1384         if (uncore_init->cpu_init) {
1385                 uncore_init->cpu_init();
1386                 cret = uncore_cpu_init();
1387         }
1388
1389         if (cret && pret)
1390                 return -ENODEV;
1391
1392         /*
1393          * Install callbacks. Core will call them for each online cpu.
1394          *
1395          * The first online cpu of each package allocates and takes
1396          * the refcounts for all other online cpus in that package.
1397          * If msrs are not enabled no allocation is required and
1398          * uncore_cpu_prepare() is not called for each online cpu.
1399          */
1400         if (!cret) {
1401                ret = cpuhp_setup_state(CPUHP_PERF_X86_UNCORE_PREP,
1402                                        "perf/x86/intel/uncore:prepare",
1403                                        uncore_cpu_prepare, NULL);
1404                 if (ret)
1405                         goto err;
1406         } else {
1407                 cpuhp_setup_state_nocalls(CPUHP_PERF_X86_UNCORE_PREP,
1408                                           "perf/x86/intel/uncore:prepare",
1409                                           uncore_cpu_prepare, NULL);
1410         }
1411         first_init = 1;
1412         cpuhp_setup_state(CPUHP_AP_PERF_X86_UNCORE_STARTING,
1413                           "perf/x86/uncore:starting",
1414                           uncore_cpu_starting, uncore_cpu_dying);
1415         first_init = 0;
1416         cpuhp_setup_state(CPUHP_AP_PERF_X86_UNCORE_ONLINE,
1417                           "perf/x86/uncore:online",
1418                           uncore_event_cpu_online, uncore_event_cpu_offline);
1419         return 0;
1420
1421 err:
1422         /* Undo box->init_box() */
1423         on_each_cpu_mask(&uncore_cpu_mask, uncore_exit_boxes, NULL, 1);
1424         uncore_types_exit(uncore_msr_uncores);
1425         uncore_pci_exit();
1426         return ret;
1427 }
1428 module_init(intel_uncore_init);
1429
1430 static void __exit intel_uncore_exit(void)
1431 {
1432         cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_UNCORE_ONLINE);
1433         cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_UNCORE_STARTING);
1434         cpuhp_remove_state_nocalls(CPUHP_PERF_X86_UNCORE_PREP);
1435         uncore_types_exit(uncore_msr_uncores);
1436         uncore_pci_exit();
1437 }
1438 module_exit(intel_uncore_exit);