arch/x86/events/intel/uncore.c

   1 #include <asm/cpu_device_id.h>
   2 #include "uncore.h"
   3
   4 static struct intel_uncore_type *empty_uncore[] = { NULL, };
   5 struct intel_uncore_type **uncore_msr_uncores = empty_uncore;
   6 struct intel_uncore_type **uncore_pci_uncores = empty_uncore;
   7
   8 static bool pcidrv_registered;
   9 struct pci_driver *uncore_pci_driver;
  10 /* pci bus to socket mapping */
  11 DEFINE_RAW_SPINLOCK(pci2phy_map_lock);
  12 struct list_head pci2phy_map_head = LIST_HEAD_INIT(pci2phy_map_head);
  13 struct pci_extra_dev *uncore_extra_pci_dev;
  14 static int max_packages;
  15
  16 /* mask of cpus that collect uncore events */
  17 static cpumask_t uncore_cpu_mask;
  18
  19 /* constraint for the fixed counter */
  20 static struct event_constraint uncore_constraint_fixed =
  21         EVENT_CONSTRAINT(~0ULL, 1 << UNCORE_PMC_IDX_FIXED, ~0ULL);
  22 struct event_constraint uncore_constraint_empty =
  23         EVENT_CONSTRAINT(0, 0, 0);
  24
  25 MODULE_LICENSE("GPL");
  26
  27 static int uncore_pcibus_to_physid(struct pci_bus *bus)
  28 {
  29         struct pci2phy_map *map;
  30         int phys_id = -1;
  31
  32         raw_spin_lock(&pci2phy_map_lock);
  33         list_for_each_entry(map, &pci2phy_map_head, list) {
  34                 if (map->segment == pci_domain_nr(bus)) {
  35                         phys_id = map->pbus_to_physid[bus->number];
  36                         break;
  37                 }
  38         }
  39         raw_spin_unlock(&pci2phy_map_lock);
  40
  41         return phys_id;
  42 }
  43
  44 static void uncore_free_pcibus_map(void)
  45 {
  46         struct pci2phy_map *map, *tmp;
  47
  48         list_for_each_entry_safe(map, tmp, &pci2phy_map_head, list) {
  49                 list_del(&map->list);
  50                 kfree(map);
  51         }
  52 }
  53
  54 struct pci2phy_map *__find_pci2phy_map(int segment)
  55 {
  56         struct pci2phy_map *map, *alloc = NULL;
  57         int i;
  58
  59         lockdep_assert_held(&pci2phy_map_lock);
  60
  61 lookup:
  62         list_for_each_entry(map, &pci2phy_map_head, list) {
  63                 if (map->segment == segment)
  64                         goto end;
  65         }
  66
  67         if (!alloc) {
  68                 raw_spin_unlock(&pci2phy_map_lock);
  69                 alloc = kmalloc(sizeof(struct pci2phy_map), GFP_KERNEL);
  70                 raw_spin_lock(&pci2phy_map_lock);
  71
  72                 if (!alloc)
  73                         return NULL;
  74
  75                 goto lookup;
  76         }
  77
  78         map = alloc;
  79         alloc = NULL;
  80         map->segment = segment;
  81         for (i = 0; i < 256; i++)
  82                 map->pbus_to_physid[i] = -1;
  83         list_add_tail(&map->list, &pci2phy_map_head);
  84
  85 end:
  86         kfree(alloc);
  87         return map;
  88 }
  89
  90 ssize_t uncore_event_show(struct kobject *kobj,
  91                           struct kobj_attribute *attr, char *buf)
  92 {
  93         struct uncore_event_desc *event =
  94                 container_of(attr, struct uncore_event_desc, attr);
  95         return sprintf(buf, "%s", event->config);
  96 }
  97
  98 struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu)
  99 {
 100         return pmu->boxes[topology_logical_package_id(cpu)];
 101 }
 102
 103 u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event)
 104 {
 105         u64 count;
 106
 107         rdmsrl(event->hw.event_base, count);
 108
 109         return count;
 110 }
 111
 112 /*
 113  * generic get constraint function for shared match/mask registers.
 114  */
 115 struct event_constraint *
 116 uncore_get_constraint(struct intel_uncore_box *box, struct perf_event *event)
 117 {
 118         struct intel_uncore_extra_reg *er;
 119         struct hw_perf_event_extra *reg1 = &event->hw.extra_reg;
 120         struct hw_perf_event_extra *reg2 = &event->hw.branch_reg;
 121         unsigned long flags;
 122         bool ok = false;
 123
 124         /*
 125          * reg->alloc can be set due to existing state, so for fake box we
 126          * need to ignore this, otherwise we might fail to allocate proper
 127          * fake state for this extra reg constraint.
 128          */
 129         if (reg1->idx == EXTRA_REG_NONE ||
 130             (!uncore_box_is_fake(box) && reg1->alloc))
 131                 return NULL;
 132
 133         er = &box->shared_regs[reg1->idx];
 134         raw_spin_lock_irqsave(&er->lock, flags);
 135         if (!atomic_read(&er->ref) ||
 136             (er->config1 == reg1->config && er->config2 == reg2->config)) {
 137                 atomic_inc(&er->ref);
 138                 er->config1 = reg1->config;
 139                 er->config2 = reg2->config;
 140                 ok = true;
 141         }
 142         raw_spin_unlock_irqrestore(&er->lock, flags);
 143
 144         if (ok) {
 145                 if (!uncore_box_is_fake(box))
 146                         reg1->alloc = 1;
 147                 return NULL;
 148         }
 149
 150         return &uncore_constraint_empty;
 151 }
 152
 153 void uncore_put_constraint(struct intel_uncore_box *box, struct perf_event *event)
 154 {
 155         struct intel_uncore_extra_reg *er;
 156         struct hw_perf_event_extra *reg1 = &event->hw.extra_reg;
 157
 158         /*
 159          * Only put constraint if extra reg was actually allocated. Also
 160          * takes care of event which do not use an extra shared reg.
 161          *
 162          * Also, if this is a fake box we shouldn't touch any event state
 163          * (reg->alloc) and we don't care about leaving inconsistent box
 164          * state either since it will be thrown out.
 165          */
 166         if (uncore_box_is_fake(box) || !reg1->alloc)
 167                 return;
 168
 169         er = &box->shared_regs[reg1->idx];
 170         atomic_dec(&er->ref);
 171         reg1->alloc = 0;
 172 }
 173
 174 u64 uncore_shared_reg_config(struct intel_uncore_box *box, int idx)
 175 {
 176         struct intel_uncore_extra_reg *er;
 177         unsigned long flags;
 178         u64 config;
 179
 180         er = &box->shared_regs[idx];
 181
 182         raw_spin_lock_irqsave(&er->lock, flags);
 183         config = er->config;
 184         raw_spin_unlock_irqrestore(&er->lock, flags);
 185
 186         return config;
 187 }
 188
 189 static void uncore_assign_hw_event(struct intel_uncore_box *box,
 190                                    struct perf_event *event, int idx)
 191 {
 192         struct hw_perf_event *hwc = &event->hw;
 193
 194         hwc->idx = idx;
 195         hwc->last_tag = ++box->tags[idx];
 196
 197         if (hwc->idx == UNCORE_PMC_IDX_FIXED) {
 198                 hwc->event_base = uncore_fixed_ctr(box);
 199                 hwc->config_base = uncore_fixed_ctl(box);
 200                 return;
 201         }
 202
 203         hwc->config_base = uncore_event_ctl(box, hwc->idx);
 204         hwc->event_base  = uncore_perf_ctr(box, hwc->idx);
 205 }
 206
 207 void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *event)
 208 {
 209         u64 prev_count, new_count, delta;
 210         int shift;
 211
 212         if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
 213                 shift = 64 - uncore_fixed_ctr_bits(box);
 214         else
 215                 shift = 64 - uncore_perf_ctr_bits(box);
 216
 217         /* the hrtimer might modify the previous event value */
 218 again:
 219         prev_count = local64_read(&event->hw.prev_count);
 220         new_count = uncore_read_counter(box, event);
 221         if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
 222                 goto again;
 223
 224         delta = (new_count << shift) - (prev_count << shift);
 225         delta >>= shift;
 226
 227         local64_add(delta, &event->count);
 228 }
 229
 230 /*
 231  * The overflow interrupt is unavailable for SandyBridge-EP, is broken
 232  * for SandyBridge. So we use hrtimer to periodically poll the counter
 233  * to avoid overflow.
 234  */
 235 static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer)
 236 {
 237         struct intel_uncore_box *box;
 238         struct perf_event *event;
 239         unsigned long flags;
 240         int bit;
 241
 242         box = container_of(hrtimer, struct intel_uncore_box, hrtimer);
 243         if (!box->n_active || box->cpu != smp_processor_id())
 244                 return HRTIMER_NORESTART;
 245         /*
 246          * disable local interrupt to prevent uncore_pmu_event_start/stop
 247          * to interrupt the update process
 248          */
 249         local_irq_save(flags);
 250
 251         /*
 252          * handle boxes with an active event list as opposed to active
 253          * counters
 254          */
 255         list_for_each_entry(event, &box->active_list, active_entry) {
 256                 uncore_perf_event_update(box, event);
 257         }
 258
 259         for_each_set_bit(bit, box->active_mask, UNCORE_PMC_IDX_MAX)
 260                 uncore_perf_event_update(box, box->events[bit]);
 261
 262         local_irq_restore(flags);
 263
 264         hrtimer_forward_now(hrtimer, ns_to_ktime(box->hrtimer_duration));
 265         return HRTIMER_RESTART;
 266 }
 267
 268 void uncore_pmu_start_hrtimer(struct intel_uncore_box *box)
 269 {
 270         hrtimer_start(&box->hrtimer, ns_to_ktime(box->hrtimer_duration),
 271                       HRTIMER_MODE_REL_PINNED);
 272 }
 273
 274 void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box)
 275 {
 276         hrtimer_cancel(&box->hrtimer);
 277 }
 278
 279 static void uncore_pmu_init_hrtimer(struct intel_uncore_box *box)
 280 {
 281         hrtimer_init(&box->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 282         box->hrtimer.function = uncore_pmu_hrtimer;
 283 }
 284
 285 static struct intel_uncore_box *uncore_alloc_box(struct intel_uncore_type *type,
 286                                                  int node)
 287 {
 288         int i, size, numshared = type->num_shared_regs ;
 289         struct intel_uncore_box *box;
 290
 291         size = sizeof(*box) + numshared * sizeof(struct intel_uncore_extra_reg);
 292
 293         box = kzalloc_node(size, GFP_KERNEL, node);
 294         if (!box)
 295                 return NULL;
 296
 297         for (i = 0; i < numshared; i++)
 298                 raw_spin_lock_init(&box->shared_regs[i].lock);
 299
 300         uncore_pmu_init_hrtimer(box);
 301         box->cpu = -1;
 302         box->pci_phys_id = -1;
 303         box->pkgid = -1;
 304
 305         /* set default hrtimer timeout */
 306         box->hrtimer_duration = UNCORE_PMU_HRTIMER_INTERVAL;
 307
 308         INIT_LIST_HEAD(&box->active_list);
 309
 310         return box;
 311 }
 312
 313 /*
 314  * Using uncore_pmu_event_init pmu event_init callback
 315  * as a detection point for uncore events.
 316  */
 317 static int uncore_pmu_event_init(struct perf_event *event);
 318
 319 static bool is_uncore_event(struct perf_event *event)
 320 {
 321         return event->pmu->event_init == uncore_pmu_event_init;
 322 }
 323
 324 static int
 325 uncore_collect_events(struct intel_uncore_box *box, struct perf_event *leader,
 326                       bool dogrp)
 327 {
 328         struct perf_event *event;
 329         int n, max_count;
 330
 331         max_count = box->pmu->type->num_counters;
 332         if (box->pmu->type->fixed_ctl)
 333                 max_count++;
 334
 335         if (box->n_events >= max_count)
 336                 return -EINVAL;
 337
 338         n = box->n_events;
 339
 340         if (is_uncore_event(leader)) {
 341                 box->event_list[n] = leader;
 342                 n++;
 343         }
 344
 345         if (!dogrp)
 346                 return n;
 347
 348         list_for_each_entry(event, &leader->sibling_list, group_entry) {
 349                 if (!is_uncore_event(event) ||
 350                     event->state <= PERF_EVENT_STATE_OFF)
 351                         continue;
 352
 353                 if (n >= max_count)
 354                         return -EINVAL;
 355
 356                 box->event_list[n] = event;
 357                 n++;
 358         }
 359         return n;
 360 }
 361
 362 static struct event_constraint *
 363 uncore_get_event_constraint(struct intel_uncore_box *box, struct perf_event *event)
 364 {
 365         struct intel_uncore_type *type = box->pmu->type;
 366         struct event_constraint *c;
 367
 368         if (type->ops->get_constraint) {
 369                 c = type->ops->get_constraint(box, event);
 370                 if (c)
 371                         return c;
 372         }
 373
 374         if (event->attr.config == UNCORE_FIXED_EVENT)
 375                 return &uncore_constraint_fixed;
 376
 377         if (type->constraints) {
 378                 for_each_event_constraint(c, type->constraints) {
 379                         if ((event->hw.config & c->cmask) == c->code)
 380                                 return c;
 381                 }
 382         }
 383
 384         return &type->unconstrainted;
 385 }
 386
 387 static void uncore_put_event_constraint(struct intel_uncore_box *box,
 388                                         struct perf_event *event)
 389 {
 390         if (box->pmu->type->ops->put_constraint)
 391                 box->pmu->type->ops->put_constraint(box, event);
 392 }
 393
 394 static int uncore_assign_events(struct intel_uncore_box *box, int assign[], int n)
 395 {
 396         unsigned long used_mask[BITS_TO_LONGS(UNCORE_PMC_IDX_MAX)];
 397         struct event_constraint *c;
 398         int i, wmin, wmax, ret = 0;
 399         struct hw_perf_event *hwc;
 400
 401         bitmap_zero(used_mask, UNCORE_PMC_IDX_MAX);
 402
 403         for (i = 0, wmin = UNCORE_PMC_IDX_MAX, wmax = 0; i < n; i++) {
 404                 c = uncore_get_event_constraint(box, box->event_list[i]);
 405                 box->event_constraint[i] = c;
 406                 wmin = min(wmin, c->weight);
 407                 wmax = max(wmax, c->weight);
 408         }
 409
 410         /* fastpath, try to reuse previous register */
 411         for (i = 0; i < n; i++) {
 412                 hwc = &box->event_list[i]->hw;
 413                 c = box->event_constraint[i];
 414
 415                 /* never assigned */
 416                 if (hwc->idx == -1)
 417                         break;
 418
 419                 /* constraint still honored */
 420                 if (!test_bit(hwc->idx, c->idxmsk))
 421                         break;
 422
 423                 /* not already used */
 424                 if (test_bit(hwc->idx, used_mask))
 425                         break;
 426
 427                 __set_bit(hwc->idx, used_mask);
 428                 if (assign)
 429                         assign[i] = hwc->idx;
 430         }
 431         /* slow path */
 432         if (i != n)
 433                 ret = perf_assign_events(box->event_constraint, n,
 434                                          wmin, wmax, n, assign);
 435
 436         if (!assign || ret) {
 437                 for (i = 0; i < n; i++)
 438                         uncore_put_event_constraint(box, box->event_list[i]);
 439         }
 440         return ret ? -EINVAL : 0;
 441 }
 442
 443 static void uncore_pmu_event_start(struct perf_event *event, int flags)
 444 {
 445         struct intel_uncore_box *box = uncore_event_to_box(event);
 446         int idx = event->hw.idx;
 447
 448         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
 449                 return;
 450
 451         if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
 452                 return;
 453
 454         event->hw.state = 0;
 455         box->events[idx] = event;
 456         box->n_active++;
 457         __set_bit(idx, box->active_mask);
 458
 459         local64_set(&event->hw.prev_count, uncore_read_counter(box, event));
 460         uncore_enable_event(box, event);
 461
 462         if (box->n_active == 1) {
 463                 uncore_enable_box(box);
 464                 uncore_pmu_start_hrtimer(box);
 465         }
 466 }
 467
 468 static void uncore_pmu_event_stop(struct perf_event *event, int flags)
 469 {
 470         struct intel_uncore_box *box = uncore_event_to_box(event);
 471         struct hw_perf_event *hwc = &event->hw;
 472
 473         if (__test_and_clear_bit(hwc->idx, box->active_mask)) {
 474                 uncore_disable_event(box, event);
 475                 box->n_active--;
 476                 box->events[hwc->idx] = NULL;
 477                 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
 478                 hwc->state |= PERF_HES_STOPPED;
 479
 480                 if (box->n_active == 0) {
 481                         uncore_disable_box(box);
 482                         uncore_pmu_cancel_hrtimer(box);
 483                 }
 484         }
 485
 486         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
 487                 /*
 488                  * Drain the remaining delta count out of a event
 489                  * that we are disabling:
 490                  */
 491                 uncore_perf_event_update(box, event);
 492                 hwc->state |= PERF_HES_UPTODATE;
 493         }
 494 }
 495
 496 static int uncore_pmu_event_add(struct perf_event *event, int flags)
 497 {
 498         struct intel_uncore_box *box = uncore_event_to_box(event);
 499         struct hw_perf_event *hwc = &event->hw;
 500         int assign[UNCORE_PMC_IDX_MAX];
 501         int i, n, ret;
 502
 503         if (!box)
 504                 return -ENODEV;
 505
 506         ret = n = uncore_collect_events(box, event, false);
 507         if (ret < 0)
 508                 return ret;
 509
 510         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 511         if (!(flags & PERF_EF_START))
 512                 hwc->state |= PERF_HES_ARCH;
 513
 514         ret = uncore_assign_events(box, assign, n);
 515         if (ret)
 516                 return ret;
 517
 518         /* save events moving to new counters */
 519         for (i = 0; i < box->n_events; i++) {
 520                 event = box->event_list[i];
 521                 hwc = &event->hw;
 522
 523                 if (hwc->idx == assign[i] &&
 524                         hwc->last_tag == box->tags[assign[i]])
 525                         continue;
 526                 /*
 527                  * Ensure we don't accidentally enable a stopped
 528                  * counter simply because we rescheduled.
 529                  */
 530                 if (hwc->state & PERF_HES_STOPPED)
 531                         hwc->state |= PERF_HES_ARCH;
 532
 533                 uncore_pmu_event_stop(event, PERF_EF_UPDATE);
 534         }
 535
 536         /* reprogram moved events into new counters */
 537         for (i = 0; i < n; i++) {
 538                 event = box->event_list[i];
 539                 hwc = &event->hw;
 540
 541                 if (hwc->idx != assign[i] ||
 542                         hwc->last_tag != box->tags[assign[i]])
 543                         uncore_assign_hw_event(box, event, assign[i]);
 544                 else if (i < box->n_events)
 545                         continue;
 546
 547                 if (hwc->state & PERF_HES_ARCH)
 548                         continue;
 549
 550                 uncore_pmu_event_start(event, 0);
 551         }
 552         box->n_events = n;
 553
 554         return 0;
 555 }
 556
 557 static void uncore_pmu_event_del(struct perf_event *event, int flags)
 558 {
 559         struct intel_uncore_box *box = uncore_event_to_box(event);
 560         int i;
 561
 562         uncore_pmu_event_stop(event, PERF_EF_UPDATE);
 563
 564         for (i = 0; i < box->n_events; i++) {
 565                 if (event == box->event_list[i]) {
 566                         uncore_put_event_constraint(box, event);
 567
 568                         for (++i; i < box->n_events; i++)
 569                                 box->event_list[i - 1] = box->event_list[i];
 570
 571                         --box->n_events;
 572                         break;
 573                 }
 574         }
 575
 576         event->hw.idx = -1;
 577         event->hw.last_tag = ~0ULL;
 578 }
 579
 580 void uncore_pmu_event_read(struct perf_event *event)
 581 {
 582         struct intel_uncore_box *box = uncore_event_to_box(event);
 583         uncore_perf_event_update(box, event);
 584 }
 585
 586 /*
 587  * validation ensures the group can be loaded onto the
 588  * PMU if it was the only group available.
 589  */
 590 static int uncore_validate_group(struct intel_uncore_pmu *pmu,
 591                                 struct perf_event *event)
 592 {
 593         struct perf_event *leader = event->group_leader;
 594         struct intel_uncore_box *fake_box;
 595         int ret = -EINVAL, n;
 596
 597         fake_box = uncore_alloc_box(pmu->type, NUMA_NO_NODE);
 598         if (!fake_box)
 599                 return -ENOMEM;
 600
 601         fake_box->pmu = pmu;
 602         /*
 603          * the event is not yet connected with its
 604          * siblings therefore we must first collect
 605          * existing siblings, then add the new event
 606          * before we can simulate the scheduling
 607          */
 608         n = uncore_collect_events(fake_box, leader, true);
 609         if (n < 0)
 610                 goto out;
 611
 612         fake_box->n_events = n;
 613         n = uncore_collect_events(fake_box, event, false);
 614         if (n < 0)
 615                 goto out;
 616
 617         fake_box->n_events = n;
 618
 619         ret = uncore_assign_events(fake_box, NULL, n);
 620 out:
 621         kfree(fake_box);
 622         return ret;
 623 }
 624
 625 static int uncore_pmu_event_init(struct perf_event *event)
 626 {
 627         struct intel_uncore_pmu *pmu;
 628         struct intel_uncore_box *box;
 629         struct hw_perf_event *hwc = &event->hw;
 630         int ret;
 631
 632         if (event->attr.type != event->pmu->type)
 633                 return -ENOENT;
 634
 635         pmu = uncore_event_to_pmu(event);
 636         /* no device found for this pmu */
 637         if (pmu->func_id < 0)
 638                 return -ENOENT;
 639
 640         /*
 641          * Uncore PMU does measure at all privilege level all the time.
 642          * So it doesn't make sense to specify any exclude bits.
 643          */
 644         if (event->attr.exclude_user || event->attr.exclude_kernel ||
 645                         event->attr.exclude_hv || event->attr.exclude_idle)
 646                 return -EINVAL;
 647
 648         /* Sampling not supported yet */
 649         if (hwc->sample_period)
 650                 return -EINVAL;
 651
 652         /*
 653          * Place all uncore events for a particular physical package
 654          * onto a single cpu
 655          */
 656         if (event->cpu < 0)
 657                 return -EINVAL;
 658         box = uncore_pmu_to_box(pmu, event->cpu);
 659         if (!box || box->cpu < 0)
 660                 return -EINVAL;
 661         event->cpu = box->cpu;
 662         event->pmu_private = box;
 663
 664         event->hw.idx = -1;
 665         event->hw.last_tag = ~0ULL;
 666         event->hw.extra_reg.idx = EXTRA_REG_NONE;
 667         event->hw.branch_reg.idx = EXTRA_REG_NONE;
 668
 669         if (event->attr.config == UNCORE_FIXED_EVENT) {
 670                 /* no fixed counter */
 671                 if (!pmu->type->fixed_ctl)
 672                         return -EINVAL;
 673                 /*
 674                  * if there is only one fixed counter, only the first pmu
 675                  * can access the fixed counter
 676                  */
 677                 if (pmu->type->single_fixed && pmu->pmu_idx > 0)
 678                         return -EINVAL;
 679
 680                 /* fixed counters have event field hardcoded to zero */
 681                 hwc->config = 0ULL;
 682         } else {
 683                 hwc->config = event->attr.config & pmu->type->event_mask;
 684                 if (pmu->type->ops->hw_config) {
 685                         ret = pmu->type->ops->hw_config(box, event);
 686                         if (ret)
 687                                 return ret;
 688                 }
 689         }
 690
 691         if (event->group_leader != event)
 692                 ret = uncore_validate_group(pmu, event);
 693         else
 694                 ret = 0;
 695
 696         return ret;
 697 }
 698
 699 static ssize_t uncore_get_attr_cpumask(struct device *dev,
 700                                 struct device_attribute *attr, char *buf)
 701 {
 702         return cpumap_print_to_pagebuf(true, buf, &uncore_cpu_mask);
 703 }
 704
 705 static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL);
 706
 707 static struct attribute *uncore_pmu_attrs[] = {
 708         &dev_attr_cpumask.attr,
 709         NULL,
 710 };
 711
 712 static struct attribute_group uncore_pmu_attr_group = {
 713         .attrs = uncore_pmu_attrs,
 714 };
 715
 716 static int uncore_pmu_register(struct intel_uncore_pmu *pmu)
 717 {
 718         int ret;
 719
 720         if (!pmu->type->pmu) {
 721                 pmu->pmu = (struct pmu) {
 722                         .attr_groups    = pmu->type->attr_groups,
 723                         .task_ctx_nr    = perf_invalid_context,
 724                         .event_init     = uncore_pmu_event_init,
 725                         .add            = uncore_pmu_event_add,
 726                         .del            = uncore_pmu_event_del,
 727                         .start          = uncore_pmu_event_start,
 728                         .stop           = uncore_pmu_event_stop,
 729                         .read           = uncore_pmu_event_read,
 730                 };
 731         } else {
 732                 pmu->pmu = *pmu->type->pmu;
 733                 pmu->pmu.attr_groups = pmu->type->attr_groups;
 734         }
 735
 736         if (pmu->type->num_boxes == 1) {
 737                 if (strlen(pmu->type->name) > 0)
 738                         sprintf(pmu->name, "uncore_%s", pmu->type->name);
 739                 else
 740                         sprintf(pmu->name, "uncore");
 741         } else {
 742                 sprintf(pmu->name, "uncore_%s_%d", pmu->type->name,
 743                         pmu->pmu_idx);
 744         }
 745
 746         ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
 747         if (!ret)
 748                 pmu->registered = true;
 749         return ret;
 750 }
 751
 752 static void uncore_pmu_unregister(struct intel_uncore_pmu *pmu)
 753 {
 754         if (!pmu->registered)
 755                 return;
 756         perf_pmu_unregister(&pmu->pmu);
 757         pmu->registered = false;
 758 }
 759
 760 static void __uncore_exit_boxes(struct intel_uncore_type *type, int cpu)
 761 {
 762         struct intel_uncore_pmu *pmu = type->pmus;
 763         struct intel_uncore_box *box;
 764         int i, pkg;
 765
 766         if (pmu) {
 767                 pkg = topology_physical_package_id(cpu);
 768                 for (i = 0; i < type->num_boxes; i++, pmu++) {
 769                         box = pmu->boxes[pkg];
 770                         if (box)
 771                                 uncore_box_exit(box);
 772                 }
 773         }
 774 }
 775
 776 static void uncore_exit_boxes(void *dummy)
 777 {
 778         struct intel_uncore_type **types;
 779
 780         for (types = uncore_msr_uncores; *types; types++)
 781                 __uncore_exit_boxes(*types++, smp_processor_id());
 782 }
 783
 784 static void uncore_free_boxes(struct intel_uncore_pmu *pmu)
 785 {
 786         int pkg;
 787
 788         for (pkg = 0; pkg < max_packages; pkg++)
 789                 kfree(pmu->boxes[pkg]);
 790         kfree(pmu->boxes);
 791 }
 792
 793 static void uncore_type_exit(struct intel_uncore_type *type)
 794 {
 795         struct intel_uncore_pmu *pmu = type->pmus;
 796         int i;
 797
 798         if (pmu) {
 799                 for (i = 0; i < type->num_boxes; i++, pmu++) {
 800                         uncore_pmu_unregister(pmu);
 801                         uncore_free_boxes(pmu);
 802                 }
 803                 kfree(type->pmus);
 804                 type->pmus = NULL;
 805         }
 806         kfree(type->events_group);
 807         type->events_group = NULL;
 808 }
 809
 810 static void uncore_types_exit(struct intel_uncore_type **types)
 811 {
 812         for (; *types; types++)
 813                 uncore_type_exit(*types);
 814 }
 815
 816 static int __init uncore_type_init(struct intel_uncore_type *type, bool setid)
 817 {
 818         struct intel_uncore_pmu *pmus;
 819         struct attribute_group *attr_group;
 820         struct attribute **attrs;
 821         size_t size;
 822         int i, j;
 823
 824         pmus = kzalloc(sizeof(*pmus) * type->num_boxes, GFP_KERNEL);
 825         if (!pmus)
 826                 return -ENOMEM;
 827
 828         size = max_packages * sizeof(struct intel_uncore_box *);
 829
 830         for (i = 0; i < type->num_boxes; i++) {
 831                 pmus[i].func_id = setid ? i : -1;
 832                 pmus[i].pmu_idx = i;
 833                 pmus[i].type    = type;
 834                 pmus[i].boxes   = kzalloc(size, GFP_KERNEL);
 835                 if (!pmus[i].boxes)
 836                         return -ENOMEM;
 837         }
 838
 839         type->pmus = pmus;
 840         type->unconstrainted = (struct event_constraint)
 841                 __EVENT_CONSTRAINT(0, (1ULL << type->num_counters) - 1,
 842                                 0, type->num_counters, 0, 0);
 843
 844         if (type->event_descs) {
 845                 for (i = 0; type->event_descs[i].attr.attr.name; i++);
 846
 847                 attr_group = kzalloc(sizeof(struct attribute *) * (i + 1) +
 848                                         sizeof(*attr_group), GFP_KERNEL);
 849                 if (!attr_group)
 850                         return -ENOMEM;
 851
 852                 attrs = (struct attribute **)(attr_group + 1);
 853                 attr_group->name = "events";
 854                 attr_group->attrs = attrs;
 855
 856                 for (j = 0; j < i; j++)
 857                         attrs[j] = &type->event_descs[j].attr.attr;
 858
 859                 type->events_group = attr_group;
 860         }
 861
 862         type->pmu_group = &uncore_pmu_attr_group;
 863         return 0;
 864 }
 865
 866 static int __init
 867 uncore_types_init(struct intel_uncore_type **types, bool setid)
 868 {
 869         int ret;
 870
 871         for (; *types; types++) {
 872                 ret = uncore_type_init(*types, setid);
 873                 if (ret)
 874                         return ret;
 875         }
 876         return 0;
 877 }
 878
 879 /*
 880  * add a pci uncore device
 881  */
 882 static int uncore_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 883 {
 884         struct intel_uncore_type *type;
 885         struct intel_uncore_pmu *pmu;
 886         struct intel_uncore_box *box;
 887         int phys_id, pkg, ret;
 888
 889         phys_id = uncore_pcibus_to_physid(pdev->bus);
 890         if (phys_id < 0)
 891                 return -ENODEV;
 892
 893         pkg = topology_phys_to_logical_pkg(phys_id);
 894         if (pkg < 0)
 895                 return -EINVAL;
 896
 897         if (UNCORE_PCI_DEV_TYPE(id->driver_data) == UNCORE_EXTRA_PCI_DEV) {
 898                 int idx = UNCORE_PCI_DEV_IDX(id->driver_data);
 899
 900                 uncore_extra_pci_dev[pkg].dev[idx] = pdev;
 901                 pci_set_drvdata(pdev, NULL);
 902                 return 0;
 903         }
 904
 905         type = uncore_pci_uncores[UNCORE_PCI_DEV_TYPE(id->driver_data)];
 906         /*
 907          * for performance monitoring unit with multiple boxes,
 908          * each box has a different function id.
 909          */
 910         pmu = &type->pmus[UNCORE_PCI_DEV_IDX(id->driver_data)];
 911         /* Knights Landing uses a common PCI device ID for multiple instances of
 912          * an uncore PMU device type. There is only one entry per device type in
 913          * the knl_uncore_pci_ids table inspite of multiple devices present for
 914          * some device types. Hence PCI device idx would be 0 for all devices.
 915          * So increment pmu pointer to point to an unused array element.
 916          */
 917         if (boot_cpu_data.x86_model == 87) {
 918                 while (pmu->func_id >= 0)
 919                         pmu++;
 920         }
 921
 922         if (WARN_ON_ONCE(pmu->boxes[pkg] != NULL))
 923                 return -EINVAL;
 924
 925         box = uncore_alloc_box(type, NUMA_NO_NODE);
 926         if (!box)
 927                 return -ENOMEM;
 928
 929         if (pmu->func_id < 0)
 930                 pmu->func_id = pdev->devfn;
 931         else
 932                 WARN_ON_ONCE(pmu->func_id != pdev->devfn);
 933
 934         atomic_inc(&box->refcnt);
 935         box->pci_phys_id = phys_id;
 936         box->pkgid = pkg;
 937         box->pci_dev = pdev;
 938         box->pmu = pmu;
 939         uncore_box_init(box);
 940         pci_set_drvdata(pdev, box);
 941
 942         pmu->boxes[pkg] = box;
 943         if (atomic_inc_return(&pmu->activeboxes) > 1)
 944                 return 0;
 945
 946         /* First active box registers the pmu */
 947         ret = uncore_pmu_register(pmu);
 948         if (ret) {
 949                 pci_set_drvdata(pdev, NULL);
 950                 pmu->boxes[pkg] = NULL;
 951                 uncore_box_exit(box);
 952                 kfree(box);
 953         }
 954         return ret;
 955 }
 956
 957 static void uncore_pci_remove(struct pci_dev *pdev)
 958 {
 959         struct intel_uncore_box *box = pci_get_drvdata(pdev);
 960         struct intel_uncore_pmu *pmu;
 961         int i, phys_id, pkg;
 962
 963         phys_id = uncore_pcibus_to_physid(pdev->bus);
 964         pkg = topology_phys_to_logical_pkg(phys_id);
 965
 966         box = pci_get_drvdata(pdev);
 967         if (!box) {
 968                 for (i = 0; i < UNCORE_EXTRA_PCI_DEV_MAX; i++) {
 969                         if (uncore_extra_pci_dev[pkg].dev[i] == pdev) {
 970                                 uncore_extra_pci_dev[pkg].dev[i] = NULL;
 971                                 break;
 972                         }
 973                 }
 974                 WARN_ON_ONCE(i >= UNCORE_EXTRA_PCI_DEV_MAX);
 975                 return;
 976         }
 977
 978         pmu = box->pmu;
 979         if (WARN_ON_ONCE(phys_id != box->pci_phys_id))
 980                 return;
 981
 982         pci_set_drvdata(pdev, NULL);
 983         pmu->boxes[pkg] = NULL;
 984         if (atomic_dec_return(&pmu->activeboxes) == 0)
 985                 uncore_pmu_unregister(pmu);
 986         uncore_box_exit(box);
 987         kfree(box);
 988 }
 989
 990 static int __init uncore_pci_init(void)
 991 {
 992         size_t size;
 993         int ret;
 994
 995         size = max_packages * sizeof(struct pci_extra_dev);
 996         uncore_extra_pci_dev = kzalloc(size, GFP_KERNEL);
 997         if (!uncore_extra_pci_dev) {
 998                 ret = -ENOMEM;
 999                 goto err;
1000         }
1001
1002         ret = uncore_types_init(uncore_pci_uncores, false);
1003         if (ret)
1004                 goto errtype;
1005
1006         uncore_pci_driver->probe = uncore_pci_probe;
1007         uncore_pci_driver->remove = uncore_pci_remove;
1008
1009         ret = pci_register_driver(uncore_pci_driver);
1010         if (ret)
1011                 goto errtype;
1012
1013         pcidrv_registered = true;
1014         return 0;
1015
1016 errtype:
1017         uncore_types_exit(uncore_pci_uncores);
1018         kfree(uncore_extra_pci_dev);
1019         uncore_extra_pci_dev = NULL;
1020         uncore_free_pcibus_map();
1021 err:
1022         uncore_pci_uncores = empty_uncore;
1023         return ret;
1024 }
1025
1026 static void uncore_pci_exit(void)
1027 {
1028         if (pcidrv_registered) {
1029                 pcidrv_registered = false;
1030                 pci_unregister_driver(uncore_pci_driver);
1031                 uncore_types_exit(uncore_pci_uncores);
1032                 kfree(uncore_extra_pci_dev);
1033                 uncore_free_pcibus_map();
1034         }
1035 }
1036
1037 static int uncore_cpu_dying(unsigned int cpu)
1038 {
1039         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1040         struct intel_uncore_pmu *pmu;
1041         struct intel_uncore_box *box;
1042         int i, pkg;
1043
1044         pkg = topology_logical_package_id(cpu);
1045         for (; *types; types++) {
1046                 type = *types;
1047                 pmu = type->pmus;
1048                 for (i = 0; i < type->num_boxes; i++, pmu++) {
1049                         box = pmu->boxes[pkg];
1050                         if (box && atomic_dec_return(&box->refcnt) == 0)
1051                                 uncore_box_exit(box);
1052                 }
1053         }
1054         return 0;
1055 }
1056
1057 static int first_init;
1058
1059 static int uncore_cpu_starting(unsigned int cpu)
1060 {
1061         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1062         struct intel_uncore_pmu *pmu;
1063         struct intel_uncore_box *box;
1064         int i, pkg, ncpus = 1;
1065
1066         if (first_init) {
1067                 /*
1068                  * On init we get the number of online cpus in the package
1069                  * and set refcount for all of them.
1070                  */
1071                 ncpus = cpumask_weight(topology_core_cpumask(cpu));
1072         }
1073
1074         pkg = topology_logical_package_id(cpu);
1075         for (; *types; types++) {
1076                 type = *types;
1077                 pmu = type->pmus;
1078                 for (i = 0; i < type->num_boxes; i++, pmu++) {
1079                         box = pmu->boxes[pkg];
1080                         if (!box)
1081                                 continue;
1082                         /* The first cpu on a package activates the box */
1083                         if (atomic_add_return(ncpus, &box->refcnt) == ncpus)
1084                                 uncore_box_init(box);
1085                 }
1086         }
1087
1088         return 0;
1089 }
1090
1091 static int uncore_cpu_prepare(unsigned int cpu)
1092 {
1093         struct intel_uncore_type *type, **types = uncore_msr_uncores;
1094         struct intel_uncore_pmu *pmu;
1095         struct intel_uncore_box *box;
1096         int i, pkg;
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                         if (pmu->boxes[pkg])
1104                                 continue;
1105                         /* First cpu of a package allocates the box */
1106                         box = uncore_alloc_box(type, cpu_to_node(cpu));
1107                         if (!box)
1108                                 return -ENOMEM;
1109                         box->pmu = pmu;
1110                         box->pkgid = pkg;
1111                         pmu->boxes[pkg] = box;
1112                 }
1113         }
1114         return 0;
1115 }
1116
1117 static void uncore_change_type_ctx(struct intel_uncore_type *type, int old_cpu,
1118                                    int new_cpu)
1119 {
1120         struct intel_uncore_pmu *pmu = type->pmus;
1121         struct intel_uncore_box *box;
1122         int i, pkg;
1123
1124         pkg = topology_logical_package_id(old_cpu < 0 ? new_cpu : old_cpu);
1125         for (i = 0; i < type->num_boxes; i++, pmu++) {
1126                 box = pmu->boxes[pkg];
1127                 if (!box)
1128                         continue;
1129
1130                 if (old_cpu < 0) {
1131                         WARN_ON_ONCE(box->cpu != -1);
1132                         box->cpu = new_cpu;
1133                         continue;
1134                 }
1135
1136                 WARN_ON_ONCE(box->cpu != old_cpu);
1137                 box->cpu = -1;
1138                 if (new_cpu < 0)
1139                         continue;
1140
1141                 uncore_pmu_cancel_hrtimer(box);
1142                 perf_pmu_migrate_context(&pmu->pmu, old_cpu, new_cpu);
1143                 box->cpu = new_cpu;
1144         }
1145 }
1146
1147 static void uncore_change_context(struct intel_uncore_type **uncores,
1148                                   int old_cpu, int new_cpu)
1149 {
1150         for (; *uncores; uncores++)
1151                 uncore_change_type_ctx(*uncores, old_cpu, new_cpu);
1152 }
1153
1154 static int uncore_event_cpu_offline(unsigned int cpu)
1155 {
1156         int target;
1157
1158         /* Check if exiting cpu is used for collecting uncore events */
1159         if (!cpumask_test_and_clear_cpu(cpu, &uncore_cpu_mask))
1160                 return 0;
1161
1162         /* Find a new cpu to collect uncore events */
1163         target = cpumask_any_but(topology_core_cpumask(cpu), cpu);
1164
1165         /* Migrate uncore events to the new target */
1166         if (target < nr_cpu_ids)
1167                 cpumask_set_cpu(target, &uncore_cpu_mask);
1168         else
1169                 target = -1;
1170
1171         uncore_change_context(uncore_msr_uncores, cpu, target);
1172         uncore_change_context(uncore_pci_uncores, cpu, target);
1173         return 0;
1174 }
1175
1176 static int uncore_event_cpu_online(unsigned int cpu)
1177 {
1178         int target;
1179
1180         /*
1181          * Check if there is an online cpu in the package
1182          * which collects uncore events already.
1183          */
1184         target = cpumask_any_and(&uncore_cpu_mask, topology_core_cpumask(cpu));
1185         if (target < nr_cpu_ids)
1186                 return 0;
1187
1188         cpumask_set_cpu(cpu, &uncore_cpu_mask);
1189
1190         uncore_change_context(uncore_msr_uncores, -1, cpu);
1191         uncore_change_context(uncore_pci_uncores, -1, cpu);
1192         return 0;
1193 }
1194
1195 static int __init type_pmu_register(struct intel_uncore_type *type)
1196 {
1197         int i, ret;
1198
1199         for (i = 0; i < type->num_boxes; i++) {
1200                 ret = uncore_pmu_register(&type->pmus[i]);
1201                 if (ret)
1202                         return ret;
1203         }
1204         return 0;
1205 }
1206
1207 static int __init uncore_msr_pmus_register(void)
1208 {
1209         struct intel_uncore_type **types = uncore_msr_uncores;
1210         int ret;
1211
1212         for (; *types; types++) {
1213                 ret = type_pmu_register(*types);
1214                 if (ret)
1215                         return ret;
1216         }
1217         return 0;
1218 }
1219
1220 static int __init uncore_cpu_init(void)
1221 {
1222         int ret;
1223
1224         ret = uncore_types_init(uncore_msr_uncores, true);
1225         if (ret)
1226                 goto err;
1227
1228         ret = uncore_msr_pmus_register();
1229         if (ret)
1230                 goto err;
1231         return 0;
1232 err:
1233         uncore_types_exit(uncore_msr_uncores);
1234         uncore_msr_uncores = empty_uncore;
1235         return ret;
1236 }
1237
1238 #define X86_UNCORE_MODEL_MATCH(model, init)     \
1239         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long)&init }
1240
1241 struct intel_uncore_init_fun {
1242         void    (*cpu_init)(void);
1243         int     (*pci_init)(void);
1244 };
1245
1246 static const struct intel_uncore_init_fun nhm_uncore_init __initconst = {
1247         .cpu_init = nhm_uncore_cpu_init,
1248 };
1249
1250 static const struct intel_uncore_init_fun snb_uncore_init __initconst = {
1251         .cpu_init = snb_uncore_cpu_init,
1252         .pci_init = snb_uncore_pci_init,
1253 };
1254
1255 static const struct intel_uncore_init_fun ivb_uncore_init __initconst = {
1256         .cpu_init = snb_uncore_cpu_init,
1257         .pci_init = ivb_uncore_pci_init,
1258 };
1259
1260 static const struct intel_uncore_init_fun hsw_uncore_init __initconst = {
1261         .cpu_init = snb_uncore_cpu_init,
1262         .pci_init = hsw_uncore_pci_init,
1263 };
1264
1265 static const struct intel_uncore_init_fun bdw_uncore_init __initconst = {
1266         .cpu_init = snb_uncore_cpu_init,
1267         .pci_init = bdw_uncore_pci_init,
1268 };
1269
1270 static const struct intel_uncore_init_fun snbep_uncore_init __initconst = {
1271         .cpu_init = snbep_uncore_cpu_init,
1272         .pci_init = snbep_uncore_pci_init,
1273 };
1274
1275 static const struct intel_uncore_init_fun nhmex_uncore_init __initconst = {
1276         .cpu_init = nhmex_uncore_cpu_init,
1277 };
1278
1279 static const struct intel_uncore_init_fun ivbep_uncore_init __initconst = {
1280         .cpu_init = ivbep_uncore_cpu_init,
1281         .pci_init = ivbep_uncore_pci_init,
1282 };
1283
1284 static const struct intel_uncore_init_fun hswep_uncore_init __initconst = {
1285         .cpu_init = hswep_uncore_cpu_init,
1286         .pci_init = hswep_uncore_pci_init,
1287 };
1288
1289 static const struct intel_uncore_init_fun bdx_uncore_init __initconst = {
1290         .cpu_init = bdx_uncore_cpu_init,
1291         .pci_init = bdx_uncore_pci_init,
1292 };
1293
1294 static const struct intel_uncore_init_fun knl_uncore_init __initconst = {
1295         .cpu_init = knl_uncore_cpu_init,
1296         .pci_init = knl_uncore_pci_init,
1297 };
1298
1299 static const struct intel_uncore_init_fun skl_uncore_init __initconst = {
1300         .pci_init = skl_uncore_pci_init,
1301 };
1302
1303 static const struct x86_cpu_id intel_uncore_match[] __initconst = {
1304         X86_UNCORE_MODEL_MATCH(26, nhm_uncore_init),    /* Nehalem */
1305         X86_UNCORE_MODEL_MATCH(30, nhm_uncore_init),
1306         X86_UNCORE_MODEL_MATCH(37, nhm_uncore_init),    /* Westmere */
1307         X86_UNCORE_MODEL_MATCH(44, nhm_uncore_init),
1308         X86_UNCORE_MODEL_MATCH(42, snb_uncore_init),    /* Sandy Bridge */
1309         X86_UNCORE_MODEL_MATCH(58, ivb_uncore_init),    /* Ivy Bridge */
1310         X86_UNCORE_MODEL_MATCH(60, hsw_uncore_init),    /* Haswell */
1311         X86_UNCORE_MODEL_MATCH(69, hsw_uncore_init),    /* Haswell Celeron */
1312         X86_UNCORE_MODEL_MATCH(70, hsw_uncore_init),    /* Haswell */
1313         X86_UNCORE_MODEL_MATCH(61, bdw_uncore_init),    /* Broadwell */
1314         X86_UNCORE_MODEL_MATCH(71, bdw_uncore_init),    /* Broadwell */
1315         X86_UNCORE_MODEL_MATCH(45, snbep_uncore_init),  /* Sandy Bridge-EP */
1316         X86_UNCORE_MODEL_MATCH(46, nhmex_uncore_init),  /* Nehalem-EX */
1317         X86_UNCORE_MODEL_MATCH(47, nhmex_uncore_init),  /* Westmere-EX aka. Xeon E7 */
1318         X86_UNCORE_MODEL_MATCH(62, ivbep_uncore_init),  /* Ivy Bridge-EP */
1319         X86_UNCORE_MODEL_MATCH(63, hswep_uncore_init),  /* Haswell-EP */
1320         X86_UNCORE_MODEL_MATCH(79, bdx_uncore_init),    /* BDX-EP */
1321         X86_UNCORE_MODEL_MATCH(86, bdx_uncore_init),    /* BDX-DE */
1322         X86_UNCORE_MODEL_MATCH(87, knl_uncore_init),    /* Knights Landing */
1323         X86_UNCORE_MODEL_MATCH(94, skl_uncore_init),    /* SkyLake */
1324         {},
1325 };
1326
1327 MODULE_DEVICE_TABLE(x86cpu, intel_uncore_match);
1328
1329 static int __init intel_uncore_init(void)
1330 {
1331         const struct x86_cpu_id *id;
1332         struct intel_uncore_init_fun *uncore_init;
1333         int pret = 0, cret = 0, ret;
1334
1335         id = x86_match_cpu(intel_uncore_match);
1336         if (!id)
1337                 return -ENODEV;
1338
1339         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
1340                 return -ENODEV;
1341
1342         max_packages = topology_max_packages();
1343
1344         uncore_init = (struct intel_uncore_init_fun *)id->driver_data;
1345         if (uncore_init->pci_init) {
1346                 pret = uncore_init->pci_init();
1347                 if (!pret)
1348                         pret = uncore_pci_init();
1349         }
1350
1351         if (uncore_init->cpu_init) {
1352                 uncore_init->cpu_init();
1353                 cret = uncore_cpu_init();
1354         }
1355
1356         if (cret && pret)
1357                 return -ENODEV;
1358
1359         /*
1360          * Install callbacks. Core will call them for each online cpu.
1361          *
1362          * The first online cpu of each package allocates and takes
1363          * the refcounts for all other online cpus in that package.
1364          * If msrs are not enabled no allocation is required and
1365          * uncore_cpu_prepare() is not called for each online cpu.
1366          */
1367         if (!cret) {
1368                ret = cpuhp_setup_state(CPUHP_PERF_X86_UNCORE_PREP,
1369                                         "PERF_X86_UNCORE_PREP",
1370                                         uncore_cpu_prepare, NULL);
1371                 if (ret)
1372                         goto err;
1373         } else {
1374                 cpuhp_setup_state_nocalls(CPUHP_PERF_X86_UNCORE_PREP,
1375                                           "PERF_X86_UNCORE_PREP",
1376                                           uncore_cpu_prepare, NULL);
1377         }
1378         first_init = 1;
1379         cpuhp_setup_state(CPUHP_AP_PERF_X86_UNCORE_STARTING,
1380                           "AP_PERF_X86_UNCORE_STARTING",
1381                           uncore_cpu_starting, uncore_cpu_dying);
1382         first_init = 0;
1383         cpuhp_setup_state(CPUHP_AP_PERF_X86_UNCORE_ONLINE,
1384                           "AP_PERF_X86_UNCORE_ONLINE",
1385                           uncore_event_cpu_online, uncore_event_cpu_offline);
1386         return 0;
1387
1388 err:
1389         /* Undo box->init_box() */
1390         on_each_cpu_mask(&uncore_cpu_mask, uncore_exit_boxes, NULL, 1);
1391         uncore_types_exit(uncore_msr_uncores);
1392         uncore_pci_exit();
1393         return ret;
1394 }
1395 module_init(intel_uncore_init);
1396
1397 static void __exit intel_uncore_exit(void)
1398 {
1399         cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_UNCORE_ONLINE);
1400         cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_UNCORE_STARTING);
1401         cpuhp_remove_state_nocalls(CPUHP_PERF_X86_UNCORE_PREP);
1402         uncore_types_exit(uncore_msr_uncores);
1403         uncore_pci_exit();
1404 }
1405 module_exit(intel_uncore_exit);