arch/x86/kernel/cpu/perf_event.c

   1 /*
   2  * Performance events x86 architecture code
   3  *
   4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
   6  *  Copyright (C) 2009 Jaswinder Singh Rajput
   7  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
   8  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   9  *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
  10  *  Copyright (C) 2009 Google, Inc., Stephane Eranian
  11  *
  12  *  For licencing details see kernel-base/COPYING
  13  */
  14
  15 #include <linux/perf_event.h>
  16 #include <linux/capability.h>
  17 #include <linux/notifier.h>
  18 #include <linux/hardirq.h>
  19 #include <linux/kprobes.h>
  20 #include <linux/module.h>
  21 #include <linux/kdebug.h>
  22 #include <linux/sched.h>
  23 #include <linux/uaccess.h>
  24 #include <linux/slab.h>
  25 #include <linux/cpu.h>
  26 #include <linux/bitops.h>
  27 #include <linux/device.h>
  28
  29 #include <asm/apic.h>
  30 #include <asm/stacktrace.h>
  31 #include <asm/nmi.h>
  32 #include <asm/smp.h>
  33 #include <asm/alternative.h>
  34 #include <asm/mmu_context.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/timer.h>
  37 #include <asm/desc.h>
  38 #include <asm/ldt.h>
  39
  40 #include "perf_event.h"
  41
  42 struct x86_pmu x86_pmu __read_mostly;
  43
  44 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
  45         .enabled = 1,
  46 };
  47
  48 struct static_key rdpmc_always_available = STATIC_KEY_INIT_FALSE;
  49
  50 u64 __read_mostly hw_cache_event_ids
  51                                 [PERF_COUNT_HW_CACHE_MAX]
  52                                 [PERF_COUNT_HW_CACHE_OP_MAX]
  53                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
  54 u64 __read_mostly hw_cache_extra_regs
  55                                 [PERF_COUNT_HW_CACHE_MAX]
  56                                 [PERF_COUNT_HW_CACHE_OP_MAX]
  57                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
  58
  59 /*
  60  * Propagate event elapsed time into the generic event.
  61  * Can only be executed on the CPU where the event is active.
  62  * Returns the delta events processed.
  63  */
  64 u64 x86_perf_event_update(struct perf_event *event)
  65 {
  66         struct hw_perf_event *hwc = &event->hw;
  67         int shift = 64 - x86_pmu.cntval_bits;
  68         u64 prev_raw_count, new_raw_count;
  69         int idx = hwc->idx;
  70         s64 delta;
  71
  72         if (idx == INTEL_PMC_IDX_FIXED_BTS)
  73                 return 0;
  74
  75         /*
  76          * Careful: an NMI might modify the previous event value.
  77          *
  78          * Our tactic to handle this is to first atomically read and
  79          * exchange a new raw count - then add that new-prev delta
  80          * count to the generic event atomically:
  81          */
  82 again:
  83         prev_raw_count = local64_read(&hwc->prev_count);
  84         rdpmcl(hwc->event_base_rdpmc, new_raw_count);
  85
  86         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
  87                                         new_raw_count) != prev_raw_count)
  88                 goto again;
  89
  90         /*
  91          * Now we have the new raw value and have updated the prev
  92          * timestamp already. We can now calculate the elapsed delta
  93          * (event-)time and add that to the generic event.
  94          *
  95          * Careful, not all hw sign-extends above the physical width
  96          * of the count.
  97          */
  98         delta = (new_raw_count << shift) - (prev_raw_count << shift);
  99         delta >>= shift;
 100
 101         local64_add(delta, &event->count);
 102         local64_sub(delta, &hwc->period_left);
 103
 104         return new_raw_count;
 105 }
 106
 107 /*
 108  * Find and validate any extra registers to set up.
 109  */
 110 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
 111 {
 112         struct hw_perf_event_extra *reg;
 113         struct extra_reg *er;
 114
 115         reg = &event->hw.extra_reg;
 116
 117         if (!x86_pmu.extra_regs)
 118                 return 0;
 119
 120         for (er = x86_pmu.extra_regs; er->msr; er++) {
 121                 if (er->event != (config & er->config_mask))
 122                         continue;
 123                 if (event->attr.config1 & ~er->valid_mask)
 124                         return -EINVAL;
 125                 /* Check if the extra msrs can be safely accessed*/
 126                 if (!er->extra_msr_access)
 127                         return -ENXIO;
 128
 129                 reg->idx = er->idx;
 130                 reg->config = event->attr.config1;
 131                 reg->reg = er->msr;
 132                 break;
 133         }
 134         return 0;
 135 }
 136
 137 static atomic_t active_events;
 138 static DEFINE_MUTEX(pmc_reserve_mutex);
 139
 140 #ifdef CONFIG_X86_LOCAL_APIC
 141
 142 static bool reserve_pmc_hardware(void)
 143 {
 144         int i;
 145
 146         for (i = 0; i < x86_pmu.num_counters; i++) {
 147                 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
 148                         goto perfctr_fail;
 149         }
 150
 151         for (i = 0; i < x86_pmu.num_counters; i++) {
 152                 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
 153                         goto eventsel_fail;
 154         }
 155
 156         return true;
 157
 158 eventsel_fail:
 159         for (i--; i >= 0; i--)
 160                 release_evntsel_nmi(x86_pmu_config_addr(i));
 161
 162         i = x86_pmu.num_counters;
 163
 164 perfctr_fail:
 165         for (i--; i >= 0; i--)
 166                 release_perfctr_nmi(x86_pmu_event_addr(i));
 167
 168         return false;
 169 }
 170
 171 static void release_pmc_hardware(void)
 172 {
 173         int i;
 174
 175         for (i = 0; i < x86_pmu.num_counters; i++) {
 176                 release_perfctr_nmi(x86_pmu_event_addr(i));
 177                 release_evntsel_nmi(x86_pmu_config_addr(i));
 178         }
 179 }
 180
 181 #else
 182
 183 static bool reserve_pmc_hardware(void) { return true; }
 184 static void release_pmc_hardware(void) {}
 185
 186 #endif
 187
 188 static bool check_hw_exists(void)
 189 {
 190         u64 val, val_fail, val_new= ~0;
 191         int i, reg, reg_fail, ret = 0;
 192         int bios_fail = 0;
 193
 194         /*
 195          * Check to see if the BIOS enabled any of the counters, if so
 196          * complain and bail.
 197          */
 198         for (i = 0; i < x86_pmu.num_counters; i++) {
 199                 reg = x86_pmu_config_addr(i);
 200                 ret = rdmsrl_safe(reg, &val);
 201                 if (ret)
 202                         goto msr_fail;
 203                 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
 204                         bios_fail = 1;
 205                         val_fail = val;
 206                         reg_fail = reg;
 207                 }
 208         }
 209
 210         if (x86_pmu.num_counters_fixed) {
 211                 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 212                 ret = rdmsrl_safe(reg, &val);
 213                 if (ret)
 214                         goto msr_fail;
 215                 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
 216                         if (val & (0x03 << i*4)) {
 217                                 bios_fail = 1;
 218                                 val_fail = val;
 219                                 reg_fail = reg;
 220                         }
 221                 }
 222         }
 223
 224         /*
 225          * Read the current value, change it and read it back to see if it
 226          * matches, this is needed to detect certain hardware emulators
 227          * (qemu/kvm) that don't trap on the MSR access and always return 0s.
 228          */
 229         reg = x86_pmu_event_addr(0);
 230         if (rdmsrl_safe(reg, &val))
 231                 goto msr_fail;
 232         val ^= 0xffffUL;
 233         ret = wrmsrl_safe(reg, val);
 234         ret |= rdmsrl_safe(reg, &val_new);
 235         if (ret || val != val_new)
 236                 goto msr_fail;
 237
 238         /*
 239          * We still allow the PMU driver to operate:
 240          */
 241         if (bios_fail) {
 242                 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
 243                 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg_fail, val_fail);
 244         }
 245
 246         return true;
 247
 248 msr_fail:
 249         printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
 250         printk("%sFailed to access perfctr msr (MSR %x is %Lx)\n",
 251                 boot_cpu_has(X86_FEATURE_HYPERVISOR) ? KERN_INFO : KERN_ERR,
 252                 reg, val_new);
 253
 254         return false;
 255 }
 256
 257 static void hw_perf_event_destroy(struct perf_event *event)
 258 {
 259         if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
 260                 release_pmc_hardware();
 261                 release_ds_buffers();
 262                 mutex_unlock(&pmc_reserve_mutex);
 263         }
 264 }
 265
 266 static inline int x86_pmu_initialized(void)
 267 {
 268         return x86_pmu.handle_irq != NULL;
 269 }
 270
 271 static inline int
 272 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
 273 {
 274         struct perf_event_attr *attr = &event->attr;
 275         unsigned int cache_type, cache_op, cache_result;
 276         u64 config, val;
 277
 278         config = attr->config;
 279
 280         cache_type = (config >>  0) & 0xff;
 281         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 282                 return -EINVAL;
 283
 284         cache_op = (config >>  8) & 0xff;
 285         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 286                 return -EINVAL;
 287
 288         cache_result = (config >> 16) & 0xff;
 289         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 290                 return -EINVAL;
 291
 292         val = hw_cache_event_ids[cache_type][cache_op][cache_result];
 293
 294         if (val == 0)
 295                 return -ENOENT;
 296
 297         if (val == -1)
 298                 return -EINVAL;
 299
 300         hwc->config |= val;
 301         attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
 302         return x86_pmu_extra_regs(val, event);
 303 }
 304
 305 int x86_setup_perfctr(struct perf_event *event)
 306 {
 307         struct perf_event_attr *attr = &event->attr;
 308         struct hw_perf_event *hwc = &event->hw;
 309         u64 config;
 310
 311         if (!is_sampling_event(event)) {
 312                 hwc->sample_period = x86_pmu.max_period;
 313                 hwc->last_period = hwc->sample_period;
 314                 local64_set(&hwc->period_left, hwc->sample_period);
 315         }
 316
 317         if (attr->type == PERF_TYPE_RAW)
 318                 return x86_pmu_extra_regs(event->attr.config, event);
 319
 320         if (attr->type == PERF_TYPE_HW_CACHE)
 321                 return set_ext_hw_attr(hwc, event);
 322
 323         if (attr->config >= x86_pmu.max_events)
 324                 return -EINVAL;
 325
 326         /*
 327          * The generic map:
 328          */
 329         config = x86_pmu.event_map(attr->config);
 330
 331         if (config == 0)
 332                 return -ENOENT;
 333
 334         if (config == -1LL)
 335                 return -EINVAL;
 336
 337         /*
 338          * Branch tracing:
 339          */
 340         if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
 341             !attr->freq && hwc->sample_period == 1) {
 342                 /* BTS is not supported by this architecture. */
 343                 if (!x86_pmu.bts_active)
 344                         return -EOPNOTSUPP;
 345
 346                 /* BTS is currently only allowed for user-mode. */
 347                 if (!attr->exclude_kernel)
 348                         return -EOPNOTSUPP;
 349         }
 350
 351         hwc->config |= config;
 352
 353         return 0;
 354 }
 355
 356 /*
 357  * check that branch_sample_type is compatible with
 358  * settings needed for precise_ip > 1 which implies
 359  * using the LBR to capture ALL taken branches at the
 360  * priv levels of the measurement
 361  */
 362 static inline int precise_br_compat(struct perf_event *event)
 363 {
 364         u64 m = event->attr.branch_sample_type;
 365         u64 b = 0;
 366
 367         /* must capture all branches */
 368         if (!(m & PERF_SAMPLE_BRANCH_ANY))
 369                 return 0;
 370
 371         m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
 372
 373         if (!event->attr.exclude_user)
 374                 b |= PERF_SAMPLE_BRANCH_USER;
 375
 376         if (!event->attr.exclude_kernel)
 377                 b |= PERF_SAMPLE_BRANCH_KERNEL;
 378
 379         /*
 380          * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
 381          */
 382
 383         return m == b;
 384 }
 385
 386 int x86_pmu_hw_config(struct perf_event *event)
 387 {
 388         if (event->attr.precise_ip) {
 389                 int precise = 0;
 390
 391                 /* Support for constant skid */
 392                 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
 393                         precise++;
 394
 395                         /* Support for IP fixup */
 396                         if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
 397                                 precise++;
 398                 }
 399
 400                 if (event->attr.precise_ip > precise)
 401                         return -EOPNOTSUPP;
 402                 /*
 403                  * check that PEBS LBR correction does not conflict with
 404                  * whatever the user is asking with attr->branch_sample_type
 405                  */
 406                 if (event->attr.precise_ip > 1 &&
 407                     x86_pmu.intel_cap.pebs_format < 2) {
 408                         u64 *br_type = &event->attr.branch_sample_type;
 409
 410                         if (has_branch_stack(event)) {
 411                                 if (!precise_br_compat(event))
 412                                         return -EOPNOTSUPP;
 413
 414                                 /* branch_sample_type is compatible */
 415
 416                         } else {
 417                                 /*
 418                                  * user did not specify  branch_sample_type
 419                                  *
 420                                  * For PEBS fixups, we capture all
 421                                  * the branches at the priv level of the
 422                                  * event.
 423                                  */
 424                                 *br_type = PERF_SAMPLE_BRANCH_ANY;
 425
 426                                 if (!event->attr.exclude_user)
 427                                         *br_type |= PERF_SAMPLE_BRANCH_USER;
 428
 429                                 if (!event->attr.exclude_kernel)
 430                                         *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
 431                         }
 432                 }
 433         }
 434
 435         /*
 436          * Generate PMC IRQs:
 437          * (keep 'enabled' bit clear for now)
 438          */
 439         event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
 440
 441         /*
 442          * Count user and OS events unless requested not to
 443          */
 444         if (!event->attr.exclude_user)
 445                 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
 446         if (!event->attr.exclude_kernel)
 447                 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
 448
 449         if (event->attr.type == PERF_TYPE_RAW)
 450                 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
 451
 452         return x86_setup_perfctr(event);
 453 }
 454
 455 /*
 456  * Setup the hardware configuration for a given attr_type
 457  */
 458 static int __x86_pmu_event_init(struct perf_event *event)
 459 {
 460         int err;
 461
 462         if (!x86_pmu_initialized())
 463                 return -ENODEV;
 464
 465         err = 0;
 466         if (!atomic_inc_not_zero(&active_events)) {
 467                 mutex_lock(&pmc_reserve_mutex);
 468                 if (atomic_read(&active_events) == 0) {
 469                         if (!reserve_pmc_hardware())
 470                                 err = -EBUSY;
 471                         else
 472                                 reserve_ds_buffers();
 473                 }
 474                 if (!err)
 475                         atomic_inc(&active_events);
 476                 mutex_unlock(&pmc_reserve_mutex);
 477         }
 478         if (err)
 479                 return err;
 480
 481         event->destroy = hw_perf_event_destroy;
 482
 483         event->hw.idx = -1;
 484         event->hw.last_cpu = -1;
 485         event->hw.last_tag = ~0ULL;
 486
 487         /* mark unused */
 488         event->hw.extra_reg.idx = EXTRA_REG_NONE;
 489         event->hw.branch_reg.idx = EXTRA_REG_NONE;
 490
 491         return x86_pmu.hw_config(event);
 492 }
 493
 494 void x86_pmu_disable_all(void)
 495 {
 496         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 497         int idx;
 498
 499         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 500                 u64 val;
 501
 502                 if (!test_bit(idx, cpuc->active_mask))
 503                         continue;
 504                 rdmsrl(x86_pmu_config_addr(idx), val);
 505                 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
 506                         continue;
 507                 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 508                 wrmsrl(x86_pmu_config_addr(idx), val);
 509         }
 510 }
 511
 512 static void x86_pmu_disable(struct pmu *pmu)
 513 {
 514         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 515
 516         if (!x86_pmu_initialized())
 517                 return;
 518
 519         if (!cpuc->enabled)
 520                 return;
 521
 522         cpuc->n_added = 0;
 523         cpuc->enabled = 0;
 524         barrier();
 525
 526         x86_pmu.disable_all();
 527 }
 528
 529 void x86_pmu_enable_all(int added)
 530 {
 531         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 532         int idx;
 533
 534         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 535                 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
 536
 537                 if (!test_bit(idx, cpuc->active_mask))
 538                         continue;
 539
 540                 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
 541         }
 542 }
 543
 544 static struct pmu pmu;
 545
 546 static inline int is_x86_event(struct perf_event *event)
 547 {
 548         return event->pmu == &pmu;
 549 }
 550
 551 /*
 552  * Event scheduler state:
 553  *
 554  * Assign events iterating over all events and counters, beginning
 555  * with events with least weights first. Keep the current iterator
 556  * state in struct sched_state.
 557  */
 558 struct sched_state {
 559         int     weight;
 560         int     event;          /* event index */
 561         int     counter;        /* counter index */
 562         int     unassigned;     /* number of events to be assigned left */
 563         unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 564 };
 565
 566 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
 567 #define SCHED_STATES_MAX        2
 568
 569 struct perf_sched {
 570         int                     max_weight;
 571         int                     max_events;
 572         struct perf_event       **events;
 573         struct sched_state      state;
 574         int                     saved_states;
 575         struct sched_state      saved[SCHED_STATES_MAX];
 576 };
 577
 578 /*
 579  * Initialize interator that runs through all events and counters.
 580  */
 581 static void perf_sched_init(struct perf_sched *sched, struct perf_event **events,
 582                             int num, int wmin, int wmax)
 583 {
 584         int idx;
 585
 586         memset(sched, 0, sizeof(*sched));
 587         sched->max_events       = num;
 588         sched->max_weight       = wmax;
 589         sched->events           = events;
 590
 591         for (idx = 0; idx < num; idx++) {
 592                 if (events[idx]->hw.constraint->weight == wmin)
 593                         break;
 594         }
 595
 596         sched->state.event      = idx;          /* start with min weight */
 597         sched->state.weight     = wmin;
 598         sched->state.unassigned = num;
 599 }
 600
 601 static void perf_sched_save_state(struct perf_sched *sched)
 602 {
 603         if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
 604                 return;
 605
 606         sched->saved[sched->saved_states] = sched->state;
 607         sched->saved_states++;
 608 }
 609
 610 static bool perf_sched_restore_state(struct perf_sched *sched)
 611 {
 612         if (!sched->saved_states)
 613                 return false;
 614
 615         sched->saved_states--;
 616         sched->state = sched->saved[sched->saved_states];
 617
 618         /* continue with next counter: */
 619         clear_bit(sched->state.counter++, sched->state.used);
 620
 621         return true;
 622 }
 623
 624 /*
 625  * Select a counter for the current event to schedule. Return true on
 626  * success.
 627  */
 628 static bool __perf_sched_find_counter(struct perf_sched *sched)
 629 {
 630         struct event_constraint *c;
 631         int idx;
 632
 633         if (!sched->state.unassigned)
 634                 return false;
 635
 636         if (sched->state.event >= sched->max_events)
 637                 return false;
 638
 639         c = sched->events[sched->state.event]->hw.constraint;
 640         /* Prefer fixed purpose counters */
 641         if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
 642                 idx = INTEL_PMC_IDX_FIXED;
 643                 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
 644                         if (!__test_and_set_bit(idx, sched->state.used))
 645                                 goto done;
 646                 }
 647         }
 648         /* Grab the first unused counter starting with idx */
 649         idx = sched->state.counter;
 650         for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
 651                 if (!__test_and_set_bit(idx, sched->state.used))
 652                         goto done;
 653         }
 654
 655         return false;
 656
 657 done:
 658         sched->state.counter = idx;
 659
 660         if (c->overlap)
 661                 perf_sched_save_state(sched);
 662
 663         return true;
 664 }
 665
 666 static bool perf_sched_find_counter(struct perf_sched *sched)
 667 {
 668         while (!__perf_sched_find_counter(sched)) {
 669                 if (!perf_sched_restore_state(sched))
 670                         return false;
 671         }
 672
 673         return true;
 674 }
 675
 676 /*
 677  * Go through all unassigned events and find the next one to schedule.
 678  * Take events with the least weight first. Return true on success.
 679  */
 680 static bool perf_sched_next_event(struct perf_sched *sched)
 681 {
 682         struct event_constraint *c;
 683
 684         if (!sched->state.unassigned || !--sched->state.unassigned)
 685                 return false;
 686
 687         do {
 688                 /* next event */
 689                 sched->state.event++;
 690                 if (sched->state.event >= sched->max_events) {
 691                         /* next weight */
 692                         sched->state.event = 0;
 693                         sched->state.weight++;
 694                         if (sched->state.weight > sched->max_weight)
 695                                 return false;
 696                 }
 697                 c = sched->events[sched->state.event]->hw.constraint;
 698         } while (c->weight != sched->state.weight);
 699
 700         sched->state.counter = 0;       /* start with first counter */
 701
 702         return true;
 703 }
 704
 705 /*
 706  * Assign a counter for each event.
 707  */
 708 int perf_assign_events(struct perf_event **events, int n,
 709                         int wmin, int wmax, int *assign)
 710 {
 711         struct perf_sched sched;
 712
 713         perf_sched_init(&sched, events, n, wmin, wmax);
 714
 715         do {
 716                 if (!perf_sched_find_counter(&sched))
 717                         break;  /* failed */
 718                 if (assign)
 719                         assign[sched.state.event] = sched.state.counter;
 720         } while (perf_sched_next_event(&sched));
 721
 722         return sched.state.unassigned;
 723 }
 724 EXPORT_SYMBOL_GPL(perf_assign_events);
 725
 726 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 727 {
 728         struct event_constraint *c;
 729         unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 730         struct perf_event *e;
 731         int i, wmin, wmax, num = 0;
 732         struct hw_perf_event *hwc;
 733
 734         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
 735
 736         for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
 737                 hwc = &cpuc->event_list[i]->hw;
 738                 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
 739                 hwc->constraint = c;
 740
 741                 wmin = min(wmin, c->weight);
 742                 wmax = max(wmax, c->weight);
 743         }
 744
 745         /*
 746          * fastpath, try to reuse previous register
 747          */
 748         for (i = 0; i < n; i++) {
 749                 hwc = &cpuc->event_list[i]->hw;
 750                 c = hwc->constraint;
 751
 752                 /* never assigned */
 753                 if (hwc->idx == -1)
 754                         break;
 755
 756                 /* constraint still honored */
 757                 if (!test_bit(hwc->idx, c->idxmsk))
 758                         break;
 759
 760                 /* not already used */
 761                 if (test_bit(hwc->idx, used_mask))
 762                         break;
 763
 764                 __set_bit(hwc->idx, used_mask);
 765                 if (assign)
 766                         assign[i] = hwc->idx;
 767         }
 768
 769         /* slow path */
 770         if (i != n)
 771                 num = perf_assign_events(cpuc->event_list, n, wmin,
 772                                          wmax, assign);
 773
 774         /*
 775          * Mark the event as committed, so we do not put_constraint()
 776          * in case new events are added and fail scheduling.
 777          */
 778         if (!num && assign) {
 779                 for (i = 0; i < n; i++) {
 780                         e = cpuc->event_list[i];
 781                         e->hw.flags |= PERF_X86_EVENT_COMMITTED;
 782                 }
 783         }
 784         /*
 785          * scheduling failed or is just a simulation,
 786          * free resources if necessary
 787          */
 788         if (!assign || num) {
 789                 for (i = 0; i < n; i++) {
 790                         e = cpuc->event_list[i];
 791                         /*
 792                          * do not put_constraint() on comitted events,
 793                          * because they are good to go
 794                          */
 795                         if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
 796                                 continue;
 797
 798                         if (x86_pmu.put_event_constraints)
 799                                 x86_pmu.put_event_constraints(cpuc, e);
 800                 }
 801         }
 802         return num ? -EINVAL : 0;
 803 }
 804
 805 /*
 806  * dogrp: true if must collect siblings events (group)
 807  * returns total number of events and error code
 808  */
 809 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
 810 {
 811         struct perf_event *event;
 812         int n, max_count;
 813
 814         max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
 815
 816         /* current number of events already accepted */
 817         n = cpuc->n_events;
 818
 819         if (is_x86_event(leader)) {
 820                 if (n >= max_count)
 821                         return -EINVAL;
 822                 cpuc->event_list[n] = leader;
 823                 n++;
 824         }
 825         if (!dogrp)
 826                 return n;
 827
 828         list_for_each_entry(event, &leader->sibling_list, group_entry) {
 829                 if (!is_x86_event(event) ||
 830                     event->state <= PERF_EVENT_STATE_OFF)
 831                         continue;
 832
 833                 if (n >= max_count)
 834                         return -EINVAL;
 835
 836                 cpuc->event_list[n] = event;
 837                 n++;
 838         }
 839         return n;
 840 }
 841
 842 static inline void x86_assign_hw_event(struct perf_event *event,
 843                                 struct cpu_hw_events *cpuc, int i)
 844 {
 845         struct hw_perf_event *hwc = &event->hw;
 846
 847         hwc->idx = cpuc->assign[i];
 848         hwc->last_cpu = smp_processor_id();
 849         hwc->last_tag = ++cpuc->tags[i];
 850
 851         if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
 852                 hwc->config_base = 0;
 853                 hwc->event_base = 0;
 854         } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
 855                 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 856                 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
 857                 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
 858         } else {
 859                 hwc->config_base = x86_pmu_config_addr(hwc->idx);
 860                 hwc->event_base  = x86_pmu_event_addr(hwc->idx);
 861                 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
 862         }
 863 }
 864
 865 static inline int match_prev_assignment(struct hw_perf_event *hwc,
 866                                         struct cpu_hw_events *cpuc,
 867                                         int i)
 868 {
 869         return hwc->idx == cpuc->assign[i] &&
 870                 hwc->last_cpu == smp_processor_id() &&
 871                 hwc->last_tag == cpuc->tags[i];
 872 }
 873
 874 static void x86_pmu_start(struct perf_event *event, int flags);
 875
 876 static void x86_pmu_enable(struct pmu *pmu)
 877 {
 878         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 879         struct perf_event *event;
 880         struct hw_perf_event *hwc;
 881         int i, added = cpuc->n_added;
 882
 883         if (!x86_pmu_initialized())
 884                 return;
 885
 886         if (cpuc->enabled)
 887                 return;
 888
 889         if (cpuc->n_added) {
 890                 int n_running = cpuc->n_events - cpuc->n_added;
 891                 /*
 892                  * apply assignment obtained either from
 893                  * hw_perf_group_sched_in() or x86_pmu_enable()
 894                  *
 895                  * step1: save events moving to new counters
 896                  */
 897                 for (i = 0; i < n_running; i++) {
 898                         event = cpuc->event_list[i];
 899                         hwc = &event->hw;
 900
 901                         /*
 902                          * we can avoid reprogramming counter if:
 903                          * - assigned same counter as last time
 904                          * - running on same CPU as last time
 905                          * - no other event has used the counter since
 906                          */
 907                         if (hwc->idx == -1 ||
 908                             match_prev_assignment(hwc, cpuc, i))
 909                                 continue;
 910
 911                         /*
 912                          * Ensure we don't accidentally enable a stopped
 913                          * counter simply because we rescheduled.
 914                          */
 915                         if (hwc->state & PERF_HES_STOPPED)
 916                                 hwc->state |= PERF_HES_ARCH;
 917
 918                         x86_pmu_stop(event, PERF_EF_UPDATE);
 919                 }
 920
 921                 /*
 922                  * step2: reprogram moved events into new counters
 923                  */
 924                 for (i = 0; i < cpuc->n_events; i++) {
 925                         event = cpuc->event_list[i];
 926                         hwc = &event->hw;
 927
 928                         if (!match_prev_assignment(hwc, cpuc, i))
 929                                 x86_assign_hw_event(event, cpuc, i);
 930                         else if (i < n_running)
 931                                 continue;
 932
 933                         if (hwc->state & PERF_HES_ARCH)
 934                                 continue;
 935
 936                         x86_pmu_start(event, PERF_EF_RELOAD);
 937                 }
 938                 cpuc->n_added = 0;
 939                 perf_events_lapic_init();
 940         }
 941
 942         cpuc->enabled = 1;
 943         barrier();
 944
 945         x86_pmu.enable_all(added);
 946 }
 947
 948 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
 949
 950 /*
 951  * Set the next IRQ period, based on the hwc->period_left value.
 952  * To be called with the event disabled in hw:
 953  */
 954 int x86_perf_event_set_period(struct perf_event *event)
 955 {
 956         struct hw_perf_event *hwc = &event->hw;
 957         s64 left = local64_read(&hwc->period_left);
 958         s64 period = hwc->sample_period;
 959         int ret = 0, idx = hwc->idx;
 960
 961         if (idx == INTEL_PMC_IDX_FIXED_BTS)
 962                 return 0;
 963
 964         /*
 965          * If we are way outside a reasonable range then just skip forward:
 966          */
 967         if (unlikely(left <= -period)) {
 968                 left = period;
 969                 local64_set(&hwc->period_left, left);
 970                 hwc->last_period = period;
 971                 ret = 1;
 972         }
 973
 974         if (unlikely(left <= 0)) {
 975                 left += period;
 976                 local64_set(&hwc->period_left, left);
 977                 hwc->last_period = period;
 978                 ret = 1;
 979         }
 980         /*
 981          * Quirk: certain CPUs dont like it if just 1 hw_event is left:
 982          */
 983         if (unlikely(left < 2))
 984                 left = 2;
 985
 986         if (left > x86_pmu.max_period)
 987                 left = x86_pmu.max_period;
 988
 989         per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
 990
 991         /*
 992          * The hw event starts counting from this event offset,
 993          * mark it to be able to extra future deltas:
 994          */
 995         local64_set(&hwc->prev_count, (u64)-left);
 996
 997         wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
 998
 999         /*
1000          * Due to erratum on certan cpu we need
1001          * a second write to be sure the register
1002          * is updated properly
1003          */
1004         if (x86_pmu.perfctr_second_write) {
1005                 wrmsrl(hwc->event_base,
1006                         (u64)(-left) & x86_pmu.cntval_mask);
1007         }
1008
1009         perf_event_update_userpage(event);
1010
1011         return ret;
1012 }
1013
1014 void x86_pmu_enable_event(struct perf_event *event)
1015 {
1016         if (__this_cpu_read(cpu_hw_events.enabled))
1017                 __x86_pmu_enable_event(&event->hw,
1018                                        ARCH_PERFMON_EVENTSEL_ENABLE);
1019 }
1020
1021 /*
1022  * Add a single event to the PMU.
1023  *
1024  * The event is added to the group of enabled events
1025  * but only if it can be scehduled with existing events.
1026  */
1027 static int x86_pmu_add(struct perf_event *event, int flags)
1028 {
1029         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1030         struct hw_perf_event *hwc;
1031         int assign[X86_PMC_IDX_MAX];
1032         int n, n0, ret;
1033
1034         hwc = &event->hw;
1035
1036         perf_pmu_disable(event->pmu);
1037         n0 = cpuc->n_events;
1038         ret = n = collect_events(cpuc, event, false);
1039         if (ret < 0)
1040                 goto out;
1041
1042         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1043         if (!(flags & PERF_EF_START))
1044                 hwc->state |= PERF_HES_ARCH;
1045
1046         /*
1047          * If group events scheduling transaction was started,
1048          * skip the schedulability test here, it will be performed
1049          * at commit time (->commit_txn) as a whole.
1050          */
1051         if (cpuc->group_flag & PERF_EVENT_TXN)
1052                 goto done_collect;
1053
1054         ret = x86_pmu.schedule_events(cpuc, n, assign);
1055         if (ret)
1056                 goto out;
1057         /*
1058          * copy new assignment, now we know it is possible
1059          * will be used by hw_perf_enable()
1060          */
1061         memcpy(cpuc->assign, assign, n*sizeof(int));
1062
1063 done_collect:
1064         /*
1065          * Commit the collect_events() state. See x86_pmu_del() and
1066          * x86_pmu_*_txn().
1067          */
1068         cpuc->n_events = n;
1069         cpuc->n_added += n - n0;
1070         cpuc->n_txn += n - n0;
1071
1072         ret = 0;
1073 out:
1074         perf_pmu_enable(event->pmu);
1075         return ret;
1076 }
1077
1078 static void x86_pmu_start(struct perf_event *event, int flags)
1079 {
1080         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1081         int idx = event->hw.idx;
1082
1083         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1084                 return;
1085
1086         if (WARN_ON_ONCE(idx == -1))
1087                 return;
1088
1089         if (flags & PERF_EF_RELOAD) {
1090                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1091                 x86_perf_event_set_period(event);
1092         }
1093
1094         event->hw.state = 0;
1095
1096         cpuc->events[idx] = event;
1097         __set_bit(idx, cpuc->active_mask);
1098         __set_bit(idx, cpuc->running);
1099         x86_pmu.enable(event);
1100         perf_event_update_userpage(event);
1101 }
1102
1103 void perf_event_print_debug(void)
1104 {
1105         u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1106         u64 pebs;
1107         struct cpu_hw_events *cpuc;
1108         unsigned long flags;
1109         int cpu, idx;
1110
1111         if (!x86_pmu.num_counters)
1112                 return;
1113
1114         local_irq_save(flags);
1115
1116         cpu = smp_processor_id();
1117         cpuc = &per_cpu(cpu_hw_events, cpu);
1118
1119         if (x86_pmu.version >= 2) {
1120                 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1121                 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1122                 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1123                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1124                 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1125
1126                 pr_info("\n");
1127                 pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1128                 pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1129                 pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1130                 pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1131                 pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1132         }
1133         pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1134
1135         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1136                 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1137                 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1138
1139                 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1140
1141                 pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1142                         cpu, idx, pmc_ctrl);
1143                 pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1144                         cpu, idx, pmc_count);
1145                 pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1146                         cpu, idx, prev_left);
1147         }
1148         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1149                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1150
1151                 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1152                         cpu, idx, pmc_count);
1153         }
1154         local_irq_restore(flags);
1155 }
1156
1157 void x86_pmu_stop(struct perf_event *event, int flags)
1158 {
1159         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1160         struct hw_perf_event *hwc = &event->hw;
1161
1162         if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1163                 x86_pmu.disable(event);
1164                 cpuc->events[hwc->idx] = NULL;
1165                 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1166                 hwc->state |= PERF_HES_STOPPED;
1167         }
1168
1169         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1170                 /*
1171                  * Drain the remaining delta count out of a event
1172                  * that we are disabling:
1173                  */
1174                 x86_perf_event_update(event);
1175                 hwc->state |= PERF_HES_UPTODATE;
1176         }
1177 }
1178
1179 static void x86_pmu_del(struct perf_event *event, int flags)
1180 {
1181         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1182         int i;
1183
1184         /*
1185          * event is descheduled
1186          */
1187         event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1188
1189         /*
1190          * If we're called during a txn, we don't need to do anything.
1191          * The events never got scheduled and ->cancel_txn will truncate
1192          * the event_list.
1193          *
1194          * XXX assumes any ->del() called during a TXN will only be on
1195          * an event added during that same TXN.
1196          */
1197         if (cpuc->group_flag & PERF_EVENT_TXN)
1198                 return;
1199
1200         /*
1201          * Not a TXN, therefore cleanup properly.
1202          */
1203         x86_pmu_stop(event, PERF_EF_UPDATE);
1204
1205         for (i = 0; i < cpuc->n_events; i++) {
1206                 if (event == cpuc->event_list[i])
1207                         break;
1208         }
1209
1210         if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1211                 return;
1212
1213         /* If we have a newly added event; make sure to decrease n_added. */
1214         if (i >= cpuc->n_events - cpuc->n_added)
1215                 --cpuc->n_added;
1216
1217         if (x86_pmu.put_event_constraints)
1218                 x86_pmu.put_event_constraints(cpuc, event);
1219
1220         /* Delete the array entry. */
1221         while (++i < cpuc->n_events)
1222                 cpuc->event_list[i-1] = cpuc->event_list[i];
1223         --cpuc->n_events;
1224
1225         perf_event_update_userpage(event);
1226 }
1227
1228 int x86_pmu_handle_irq(struct pt_regs *regs)
1229 {
1230         struct perf_sample_data data;
1231         struct cpu_hw_events *cpuc;
1232         struct perf_event *event;
1233         int idx, handled = 0;
1234         u64 val;
1235
1236         cpuc = this_cpu_ptr(&cpu_hw_events);
1237
1238         /*
1239          * Some chipsets need to unmask the LVTPC in a particular spot
1240          * inside the nmi handler.  As a result, the unmasking was pushed
1241          * into all the nmi handlers.
1242          *
1243          * This generic handler doesn't seem to have any issues where the
1244          * unmasking occurs so it was left at the top.
1245          */
1246         apic_write(APIC_LVTPC, APIC_DM_NMI);
1247
1248         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1249                 if (!test_bit(idx, cpuc->active_mask)) {
1250                         /*
1251                          * Though we deactivated the counter some cpus
1252                          * might still deliver spurious interrupts still
1253                          * in flight. Catch them:
1254                          */
1255                         if (__test_and_clear_bit(idx, cpuc->running))
1256                                 handled++;
1257                         continue;
1258                 }
1259
1260                 event = cpuc->events[idx];
1261
1262                 val = x86_perf_event_update(event);
1263                 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1264                         continue;
1265
1266                 /*
1267                  * event overflow
1268                  */
1269                 handled++;
1270                 perf_sample_data_init(&data, 0, event->hw.last_period);
1271
1272                 if (!x86_perf_event_set_period(event))
1273                         continue;
1274
1275                 if (perf_event_overflow(event, &data, regs))
1276                         x86_pmu_stop(event, 0);
1277         }
1278
1279         if (handled)
1280                 inc_irq_stat(apic_perf_irqs);
1281
1282         return handled;
1283 }
1284
1285 void perf_events_lapic_init(void)
1286 {
1287         if (!x86_pmu.apic || !x86_pmu_initialized())
1288                 return;
1289
1290         /*
1291          * Always use NMI for PMU
1292          */
1293         apic_write(APIC_LVTPC, APIC_DM_NMI);
1294 }
1295
1296 static int
1297 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1298 {
1299         u64 start_clock;
1300         u64 finish_clock;
1301         int ret;
1302
1303         if (!atomic_read(&active_events))
1304                 return NMI_DONE;
1305
1306         start_clock = sched_clock();
1307         ret = x86_pmu.handle_irq(regs);
1308         finish_clock = sched_clock();
1309
1310         perf_sample_event_took(finish_clock - start_clock);
1311
1312         return ret;
1313 }
1314 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1315
1316 struct event_constraint emptyconstraint;
1317 struct event_constraint unconstrained;
1318
1319 static int
1320 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1321 {
1322         unsigned int cpu = (long)hcpu;
1323         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1324         int ret = NOTIFY_OK;
1325
1326         switch (action & ~CPU_TASKS_FROZEN) {
1327         case CPU_UP_PREPARE:
1328                 cpuc->kfree_on_online = NULL;
1329                 if (x86_pmu.cpu_prepare)
1330                         ret = x86_pmu.cpu_prepare(cpu);
1331                 break;
1332
1333         case CPU_STARTING:
1334                 if (x86_pmu.cpu_starting)
1335                         x86_pmu.cpu_starting(cpu);
1336                 break;
1337
1338         case CPU_ONLINE:
1339                 kfree(cpuc->kfree_on_online);
1340                 break;
1341
1342         case CPU_DYING:
1343                 if (x86_pmu.cpu_dying)
1344                         x86_pmu.cpu_dying(cpu);
1345                 break;
1346
1347         case CPU_UP_CANCELED:
1348         case CPU_DEAD:
1349                 if (x86_pmu.cpu_dead)
1350                         x86_pmu.cpu_dead(cpu);
1351                 break;
1352
1353         default:
1354                 break;
1355         }
1356
1357         return ret;
1358 }
1359
1360 static void __init pmu_check_apic(void)
1361 {
1362         if (cpu_has_apic)
1363                 return;
1364
1365         x86_pmu.apic = 0;
1366         pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1367         pr_info("no hardware sampling interrupt available.\n");
1368
1369         /*
1370          * If we have a PMU initialized but no APIC
1371          * interrupts, we cannot sample hardware
1372          * events (user-space has to fall back and
1373          * sample via a hrtimer based software event):
1374          */
1375         pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1376
1377 }
1378
1379 static struct attribute_group x86_pmu_format_group = {
1380         .name = "format",
1381         .attrs = NULL,
1382 };
1383
1384 /*
1385  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1386  * out of events_attr attributes.
1387  */
1388 static void __init filter_events(struct attribute **attrs)
1389 {
1390         struct device_attribute *d;
1391         struct perf_pmu_events_attr *pmu_attr;
1392         int i, j;
1393
1394         for (i = 0; attrs[i]; i++) {
1395                 d = (struct device_attribute *)attrs[i];
1396                 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1397                 /* str trumps id */
1398                 if (pmu_attr->event_str)
1399                         continue;
1400                 if (x86_pmu.event_map(i))
1401                         continue;
1402
1403                 for (j = i; attrs[j]; j++)
1404                         attrs[j] = attrs[j + 1];
1405
1406                 /* Check the shifted attr. */
1407                 i--;
1408         }
1409 }
1410
1411 /* Merge two pointer arrays */
1412 static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1413 {
1414         struct attribute **new;
1415         int j, i;
1416
1417         for (j = 0; a[j]; j++)
1418                 ;
1419         for (i = 0; b[i]; i++)
1420                 j++;
1421         j++;
1422
1423         new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1424         if (!new)
1425                 return NULL;
1426
1427         j = 0;
1428         for (i = 0; a[i]; i++)
1429                 new[j++] = a[i];
1430         for (i = 0; b[i]; i++)
1431                 new[j++] = b[i];
1432         new[j] = NULL;
1433
1434         return new;
1435 }
1436
1437 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1438                           char *page)
1439 {
1440         struct perf_pmu_events_attr *pmu_attr = \
1441                 container_of(attr, struct perf_pmu_events_attr, attr);
1442         u64 config = x86_pmu.event_map(pmu_attr->id);
1443
1444         /* string trumps id */
1445         if (pmu_attr->event_str)
1446                 return sprintf(page, "%s", pmu_attr->event_str);
1447
1448         return x86_pmu.events_sysfs_show(page, config);
1449 }
1450
1451 EVENT_ATTR(cpu-cycles,                  CPU_CYCLES              );
1452 EVENT_ATTR(instructions,                INSTRUCTIONS            );
1453 EVENT_ATTR(cache-references,            CACHE_REFERENCES        );
1454 EVENT_ATTR(cache-misses,                CACHE_MISSES            );
1455 EVENT_ATTR(branch-instructions,         BRANCH_INSTRUCTIONS     );
1456 EVENT_ATTR(branch-misses,               BRANCH_MISSES           );
1457 EVENT_ATTR(bus-cycles,                  BUS_CYCLES              );
1458 EVENT_ATTR(stalled-cycles-frontend,     STALLED_CYCLES_FRONTEND );
1459 EVENT_ATTR(stalled-cycles-backend,      STALLED_CYCLES_BACKEND  );
1460 EVENT_ATTR(ref-cycles,                  REF_CPU_CYCLES          );
1461
1462 static struct attribute *empty_attrs;
1463
1464 static struct attribute *events_attr[] = {
1465         EVENT_PTR(CPU_CYCLES),
1466         EVENT_PTR(INSTRUCTIONS),
1467         EVENT_PTR(CACHE_REFERENCES),
1468         EVENT_PTR(CACHE_MISSES),
1469         EVENT_PTR(BRANCH_INSTRUCTIONS),
1470         EVENT_PTR(BRANCH_MISSES),
1471         EVENT_PTR(BUS_CYCLES),
1472         EVENT_PTR(STALLED_CYCLES_FRONTEND),
1473         EVENT_PTR(STALLED_CYCLES_BACKEND),
1474         EVENT_PTR(REF_CPU_CYCLES),
1475         NULL,
1476 };
1477
1478 static struct attribute_group x86_pmu_events_group = {
1479         .name = "events",
1480         .attrs = events_attr,
1481 };
1482
1483 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1484 {
1485         u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1486         u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1487         bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1488         bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1489         bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1490         bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1491         ssize_t ret;
1492
1493         /*
1494         * We have whole page size to spend and just little data
1495         * to write, so we can safely use sprintf.
1496         */
1497         ret = sprintf(page, "event=0x%02llx", event);
1498
1499         if (umask)
1500                 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1501
1502         if (edge)
1503                 ret += sprintf(page + ret, ",edge");
1504
1505         if (pc)
1506                 ret += sprintf(page + ret, ",pc");
1507
1508         if (any)
1509                 ret += sprintf(page + ret, ",any");
1510
1511         if (inv)
1512                 ret += sprintf(page + ret, ",inv");
1513
1514         if (cmask)
1515                 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1516
1517         ret += sprintf(page + ret, "\n");
1518
1519         return ret;
1520 }
1521
1522 static int __init init_hw_perf_events(void)
1523 {
1524         struct x86_pmu_quirk *quirk;
1525         int err;
1526
1527         pr_info("Performance Events: ");
1528
1529         switch (boot_cpu_data.x86_vendor) {
1530         case X86_VENDOR_INTEL:
1531                 err = intel_pmu_init();
1532                 break;
1533         case X86_VENDOR_AMD:
1534                 err = amd_pmu_init();
1535                 break;
1536         default:
1537                 err = -ENOTSUPP;
1538         }
1539         if (err != 0) {
1540                 pr_cont("no PMU driver, software events only.\n");
1541                 return 0;
1542         }
1543
1544         pmu_check_apic();
1545
1546         /* sanity check that the hardware exists or is emulated */
1547         if (!check_hw_exists())
1548                 return 0;
1549
1550         pr_cont("%s PMU driver.\n", x86_pmu.name);
1551
1552         x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1553
1554         for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1555                 quirk->func();
1556
1557         if (!x86_pmu.intel_ctrl)
1558                 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1559
1560         perf_events_lapic_init();
1561         register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1562
1563         unconstrained = (struct event_constraint)
1564                 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1565                                    0, x86_pmu.num_counters, 0, 0);
1566
1567         x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1568
1569         if (x86_pmu.event_attrs)
1570                 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1571
1572         if (!x86_pmu.events_sysfs_show)
1573                 x86_pmu_events_group.attrs = &empty_attrs;
1574         else
1575                 filter_events(x86_pmu_events_group.attrs);
1576
1577         if (x86_pmu.cpu_events) {
1578                 struct attribute **tmp;
1579
1580                 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1581                 if (!WARN_ON(!tmp))
1582                         x86_pmu_events_group.attrs = tmp;
1583         }
1584
1585         pr_info("... version:                %d\n",     x86_pmu.version);
1586         pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1587         pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1588         pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1589         pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1590         pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1591         pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1592
1593         perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1594         perf_cpu_notifier(x86_pmu_notifier);
1595
1596         return 0;
1597 }
1598 early_initcall(init_hw_perf_events);
1599
1600 static inline void x86_pmu_read(struct perf_event *event)
1601 {
1602         x86_perf_event_update(event);
1603 }
1604
1605 /*
1606  * Start group events scheduling transaction
1607  * Set the flag to make pmu::enable() not perform the
1608  * schedulability test, it will be performed at commit time
1609  */
1610 static void x86_pmu_start_txn(struct pmu *pmu)
1611 {
1612         perf_pmu_disable(pmu);
1613         __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1614         __this_cpu_write(cpu_hw_events.n_txn, 0);
1615 }
1616
1617 /*
1618  * Stop group events scheduling transaction
1619  * Clear the flag and pmu::enable() will perform the
1620  * schedulability test.
1621  */
1622 static void x86_pmu_cancel_txn(struct pmu *pmu)
1623 {
1624         __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1625         /*
1626          * Truncate collected array by the number of events added in this
1627          * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1628          */
1629         __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1630         __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1631         perf_pmu_enable(pmu);
1632 }
1633
1634 /*
1635  * Commit group events scheduling transaction
1636  * Perform the group schedulability test as a whole
1637  * Return 0 if success
1638  *
1639  * Does not cancel the transaction on failure; expects the caller to do this.
1640  */
1641 static int x86_pmu_commit_txn(struct pmu *pmu)
1642 {
1643         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1644         int assign[X86_PMC_IDX_MAX];
1645         int n, ret;
1646
1647         n = cpuc->n_events;
1648
1649         if (!x86_pmu_initialized())
1650                 return -EAGAIN;
1651
1652         ret = x86_pmu.schedule_events(cpuc, n, assign);
1653         if (ret)
1654                 return ret;
1655
1656         /*
1657          * copy new assignment, now we know it is possible
1658          * will be used by hw_perf_enable()
1659          */
1660         memcpy(cpuc->assign, assign, n*sizeof(int));
1661
1662         cpuc->group_flag &= ~PERF_EVENT_TXN;
1663         perf_pmu_enable(pmu);
1664         return 0;
1665 }
1666 /*
1667  * a fake_cpuc is used to validate event groups. Due to
1668  * the extra reg logic, we need to also allocate a fake
1669  * per_core and per_cpu structure. Otherwise, group events
1670  * using extra reg may conflict without the kernel being
1671  * able to catch this when the last event gets added to
1672  * the group.
1673  */
1674 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1675 {
1676         kfree(cpuc->shared_regs);
1677         kfree(cpuc);
1678 }
1679
1680 static struct cpu_hw_events *allocate_fake_cpuc(void)
1681 {
1682         struct cpu_hw_events *cpuc;
1683         int cpu = raw_smp_processor_id();
1684
1685         cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1686         if (!cpuc)
1687                 return ERR_PTR(-ENOMEM);
1688
1689         /* only needed, if we have extra_regs */
1690         if (x86_pmu.extra_regs) {
1691                 cpuc->shared_regs = allocate_shared_regs(cpu);
1692                 if (!cpuc->shared_regs)
1693                         goto error;
1694         }
1695         cpuc->is_fake = 1;
1696         return cpuc;
1697 error:
1698         free_fake_cpuc(cpuc);
1699         return ERR_PTR(-ENOMEM);
1700 }
1701
1702 /*
1703  * validate that we can schedule this event
1704  */
1705 static int validate_event(struct perf_event *event)
1706 {
1707         struct cpu_hw_events *fake_cpuc;
1708         struct event_constraint *c;
1709         int ret = 0;
1710
1711         fake_cpuc = allocate_fake_cpuc();
1712         if (IS_ERR(fake_cpuc))
1713                 return PTR_ERR(fake_cpuc);
1714
1715         c = x86_pmu.get_event_constraints(fake_cpuc, event);
1716
1717         if (!c || !c->weight)
1718                 ret = -EINVAL;
1719
1720         if (x86_pmu.put_event_constraints)
1721                 x86_pmu.put_event_constraints(fake_cpuc, event);
1722
1723         free_fake_cpuc(fake_cpuc);
1724
1725         return ret;
1726 }
1727
1728 /*
1729  * validate a single event group
1730  *
1731  * validation include:
1732  *      - check events are compatible which each other
1733  *      - events do not compete for the same counter
1734  *      - number of events <= number of counters
1735  *
1736  * validation ensures the group can be loaded onto the
1737  * PMU if it was the only group available.
1738  */
1739 static int validate_group(struct perf_event *event)
1740 {
1741         struct perf_event *leader = event->group_leader;
1742         struct cpu_hw_events *fake_cpuc;
1743         int ret = -EINVAL, n;
1744
1745         fake_cpuc = allocate_fake_cpuc();
1746         if (IS_ERR(fake_cpuc))
1747                 return PTR_ERR(fake_cpuc);
1748         /*
1749          * the event is not yet connected with its
1750          * siblings therefore we must first collect
1751          * existing siblings, then add the new event
1752          * before we can simulate the scheduling
1753          */
1754         n = collect_events(fake_cpuc, leader, true);
1755         if (n < 0)
1756                 goto out;
1757
1758         fake_cpuc->n_events = n;
1759         n = collect_events(fake_cpuc, event, false);
1760         if (n < 0)
1761                 goto out;
1762
1763         fake_cpuc->n_events = n;
1764
1765         ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1766
1767 out:
1768         free_fake_cpuc(fake_cpuc);
1769         return ret;
1770 }
1771
1772 static int x86_pmu_event_init(struct perf_event *event)
1773 {
1774         struct pmu *tmp;
1775         int err;
1776
1777         switch (event->attr.type) {
1778         case PERF_TYPE_RAW:
1779         case PERF_TYPE_HARDWARE:
1780         case PERF_TYPE_HW_CACHE:
1781                 break;
1782
1783         default:
1784                 return -ENOENT;
1785         }
1786
1787         err = __x86_pmu_event_init(event);
1788         if (!err) {
1789                 /*
1790                  * we temporarily connect event to its pmu
1791                  * such that validate_group() can classify
1792                  * it as an x86 event using is_x86_event()
1793                  */
1794                 tmp = event->pmu;
1795                 event->pmu = &pmu;
1796
1797                 if (event->group_leader != event)
1798                         err = validate_group(event);
1799                 else
1800                         err = validate_event(event);
1801
1802                 event->pmu = tmp;
1803         }
1804         if (err) {
1805                 if (event->destroy)
1806                         event->destroy(event);
1807         }
1808
1809         if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
1810                 event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
1811
1812         return err;
1813 }
1814
1815 static void refresh_pce(void *ignored)
1816 {
1817         if (current->mm)
1818                 load_mm_cr4(current->mm);
1819 }
1820
1821 static void x86_pmu_event_mapped(struct perf_event *event)
1822 {
1823         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1824                 return;
1825
1826         if (atomic_inc_return(&current->mm->context.perf_rdpmc_allowed) == 1)
1827                 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
1828 }
1829
1830 static void x86_pmu_event_unmapped(struct perf_event *event)
1831 {
1832         if (!current->mm)
1833                 return;
1834
1835         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1836                 return;
1837
1838         if (atomic_dec_and_test(&current->mm->context.perf_rdpmc_allowed))
1839                 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
1840 }
1841
1842 static int x86_pmu_event_idx(struct perf_event *event)
1843 {
1844         int idx = event->hw.idx;
1845
1846         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
1847                 return 0;
1848
1849         if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1850                 idx -= INTEL_PMC_IDX_FIXED;
1851                 idx |= 1 << 30;
1852         }
1853
1854         return idx + 1;
1855 }
1856
1857 static ssize_t get_attr_rdpmc(struct device *cdev,
1858                               struct device_attribute *attr,
1859                               char *buf)
1860 {
1861         return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1862 }
1863
1864 static ssize_t set_attr_rdpmc(struct device *cdev,
1865                               struct device_attribute *attr,
1866                               const char *buf, size_t count)
1867 {
1868         unsigned long val;
1869         ssize_t ret;
1870
1871         ret = kstrtoul(buf, 0, &val);
1872         if (ret)
1873                 return ret;
1874
1875         if (val > 2)
1876                 return -EINVAL;
1877
1878         if (x86_pmu.attr_rdpmc_broken)
1879                 return -ENOTSUPP;
1880
1881         if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
1882                 /*
1883                  * Changing into or out of always available, aka
1884                  * perf-event-bypassing mode.  This path is extremely slow,
1885                  * but only root can trigger it, so it's okay.
1886                  */
1887                 if (val == 2)
1888                         static_key_slow_inc(&rdpmc_always_available);
1889                 else
1890                         static_key_slow_dec(&rdpmc_always_available);
1891                 on_each_cpu(refresh_pce, NULL, 1);
1892         }
1893
1894         x86_pmu.attr_rdpmc = val;
1895
1896         return count;
1897 }
1898
1899 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1900
1901 static struct attribute *x86_pmu_attrs[] = {
1902         &dev_attr_rdpmc.attr,
1903         NULL,
1904 };
1905
1906 static struct attribute_group x86_pmu_attr_group = {
1907         .attrs = x86_pmu_attrs,
1908 };
1909
1910 static const struct attribute_group *x86_pmu_attr_groups[] = {
1911         &x86_pmu_attr_group,
1912         &x86_pmu_format_group,
1913         &x86_pmu_events_group,
1914         NULL,
1915 };
1916
1917 static void x86_pmu_flush_branch_stack(void)
1918 {
1919         if (x86_pmu.flush_branch_stack)
1920                 x86_pmu.flush_branch_stack();
1921 }
1922
1923 void perf_check_microcode(void)
1924 {
1925         if (x86_pmu.check_microcode)
1926                 x86_pmu.check_microcode();
1927 }
1928 EXPORT_SYMBOL_GPL(perf_check_microcode);
1929
1930 static struct pmu pmu = {
1931         .pmu_enable             = x86_pmu_enable,
1932         .pmu_disable            = x86_pmu_disable,
1933
1934         .attr_groups            = x86_pmu_attr_groups,
1935
1936         .event_init             = x86_pmu_event_init,
1937
1938         .event_mapped           = x86_pmu_event_mapped,
1939         .event_unmapped         = x86_pmu_event_unmapped,
1940
1941         .add                    = x86_pmu_add,
1942         .del                    = x86_pmu_del,
1943         .start                  = x86_pmu_start,
1944         .stop                   = x86_pmu_stop,
1945         .read                   = x86_pmu_read,
1946
1947         .start_txn              = x86_pmu_start_txn,
1948         .cancel_txn             = x86_pmu_cancel_txn,
1949         .commit_txn             = x86_pmu_commit_txn,
1950
1951         .event_idx              = x86_pmu_event_idx,
1952         .flush_branch_stack     = x86_pmu_flush_branch_stack,
1953 };
1954
1955 void arch_perf_update_userpage(struct perf_event *event,
1956                                struct perf_event_mmap_page *userpg, u64 now)
1957 {
1958         struct cyc2ns_data *data;
1959
1960         userpg->cap_user_time = 0;
1961         userpg->cap_user_time_zero = 0;
1962         userpg->cap_user_rdpmc =
1963                 !!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
1964         userpg->pmc_width = x86_pmu.cntval_bits;
1965
1966         if (!sched_clock_stable())
1967                 return;
1968
1969         data = cyc2ns_read_begin();
1970
1971         userpg->cap_user_time = 1;
1972         userpg->time_mult = data->cyc2ns_mul;
1973         userpg->time_shift = data->cyc2ns_shift;
1974         userpg->time_offset = data->cyc2ns_offset - now;
1975
1976         userpg->cap_user_time_zero = 1;
1977         userpg->time_zero = data->cyc2ns_offset;
1978
1979         cyc2ns_read_end(data);
1980 }
1981
1982 /*
1983  * callchain support
1984  */
1985
1986 static int backtrace_stack(void *data, char *name)
1987 {
1988         return 0;
1989 }
1990
1991 static void backtrace_address(void *data, unsigned long addr, int reliable)
1992 {
1993         struct perf_callchain_entry *entry = data;
1994
1995         perf_callchain_store(entry, addr);
1996 }
1997
1998 static const struct stacktrace_ops backtrace_ops = {
1999         .stack                  = backtrace_stack,
2000         .address                = backtrace_address,
2001         .walk_stack             = print_context_stack_bp,
2002 };
2003
2004 void
2005 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
2006 {
2007         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2008                 /* TODO: We don't support guest os callchain now */
2009                 return;
2010         }
2011
2012         perf_callchain_store(entry, regs->ip);
2013
2014         dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2015 }
2016
2017 static inline int
2018 valid_user_frame(const void __user *fp, unsigned long size)
2019 {
2020         return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2021 }
2022
2023 static unsigned long get_segment_base(unsigned int segment)
2024 {
2025         struct desc_struct *desc;
2026         int idx = segment >> 3;
2027
2028         if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2029                 if (idx > LDT_ENTRIES)
2030                         return 0;
2031
2032                 if (idx > current->active_mm->context.size)
2033                         return 0;
2034
2035                 desc = current->active_mm->context.ldt;
2036         } else {
2037                 if (idx > GDT_ENTRIES)
2038                         return 0;
2039
2040                 desc = raw_cpu_ptr(gdt_page.gdt);
2041         }
2042
2043         return get_desc_base(desc + idx);
2044 }
2045
2046 #ifdef CONFIG_COMPAT
2047
2048 #include <asm/compat.h>
2049
2050 static inline int
2051 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2052 {
2053         /* 32-bit process in 64-bit kernel. */
2054         unsigned long ss_base, cs_base;
2055         struct stack_frame_ia32 frame;
2056         const void __user *fp;
2057
2058         if (!test_thread_flag(TIF_IA32))
2059                 return 0;
2060
2061         cs_base = get_segment_base(regs->cs);
2062         ss_base = get_segment_base(regs->ss);
2063
2064         fp = compat_ptr(ss_base + regs->bp);
2065         while (entry->nr < PERF_MAX_STACK_DEPTH) {
2066                 unsigned long bytes;
2067                 frame.next_frame     = 0;
2068                 frame.return_address = 0;
2069
2070                 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2071                 if (bytes != 0)
2072                         break;
2073
2074                 if (!valid_user_frame(fp, sizeof(frame)))
2075                         break;
2076
2077                 perf_callchain_store(entry, cs_base + frame.return_address);
2078                 fp = compat_ptr(ss_base + frame.next_frame);
2079         }
2080         return 1;
2081 }
2082 #else
2083 static inline int
2084 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2085 {
2086     return 0;
2087 }
2088 #endif
2089
2090 void
2091 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2092 {
2093         struct stack_frame frame;
2094         const void __user *fp;
2095
2096         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2097                 /* TODO: We don't support guest os callchain now */
2098                 return;
2099         }
2100
2101         /*
2102          * We don't know what to do with VM86 stacks.. ignore them for now.
2103          */
2104         if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2105                 return;
2106
2107         fp = (void __user *)regs->bp;
2108
2109         perf_callchain_store(entry, regs->ip);
2110
2111         if (!current->mm)
2112                 return;
2113
2114         if (perf_callchain_user32(regs, entry))
2115                 return;
2116
2117         while (entry->nr < PERF_MAX_STACK_DEPTH) {
2118                 unsigned long bytes;
2119                 frame.next_frame             = NULL;
2120                 frame.return_address = 0;
2121
2122                 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2123                 if (bytes != 0)
2124                         break;
2125
2126                 if (!valid_user_frame(fp, sizeof(frame)))
2127                         break;
2128
2129                 perf_callchain_store(entry, frame.return_address);
2130                 fp = frame.next_frame;
2131         }
2132 }
2133
2134 /*
2135  * Deal with code segment offsets for the various execution modes:
2136  *
2137  *   VM86 - the good olde 16 bit days, where the linear address is
2138  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2139  *
2140  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2141  *          to figure out what the 32bit base address is.
2142  *
2143  *    X32 - has TIF_X32 set, but is running in x86_64
2144  *
2145  * X86_64 - CS,DS,SS,ES are all zero based.
2146  */
2147 static unsigned long code_segment_base(struct pt_regs *regs)
2148 {
2149         /*
2150          * If we are in VM86 mode, add the segment offset to convert to a
2151          * linear address.
2152          */
2153         if (regs->flags & X86_VM_MASK)
2154                 return 0x10 * regs->cs;
2155
2156         /*
2157          * For IA32 we look at the GDT/LDT segment base to convert the
2158          * effective IP to a linear address.
2159          */
2160 #ifdef CONFIG_X86_32
2161         if (user_mode(regs) && regs->cs != __USER_CS)
2162                 return get_segment_base(regs->cs);
2163 #else
2164         if (test_thread_flag(TIF_IA32)) {
2165                 if (user_mode(regs) && regs->cs != __USER32_CS)
2166                         return get_segment_base(regs->cs);
2167         }
2168 #endif
2169         return 0;
2170 }
2171
2172 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2173 {
2174         if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2175                 return perf_guest_cbs->get_guest_ip();
2176
2177         return regs->ip + code_segment_base(regs);
2178 }
2179
2180 unsigned long perf_misc_flags(struct pt_regs *regs)
2181 {
2182         int misc = 0;
2183
2184         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2185                 if (perf_guest_cbs->is_user_mode())
2186                         misc |= PERF_RECORD_MISC_GUEST_USER;
2187                 else
2188                         misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2189         } else {
2190                 if (user_mode(regs))
2191                         misc |= PERF_RECORD_MISC_USER;
2192                 else
2193                         misc |= PERF_RECORD_MISC_KERNEL;
2194         }
2195
2196         if (regs->flags & PERF_EFLAGS_EXACT)
2197                 misc |= PERF_RECORD_MISC_EXACT_IP;
2198
2199         return misc;
2200 }
2201
2202 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2203 {
2204         cap->version            = x86_pmu.version;
2205         cap->num_counters_gp    = x86_pmu.num_counters;
2206         cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2207         cap->bit_width_gp       = x86_pmu.cntval_bits;
2208         cap->bit_width_fixed    = x86_pmu.cntval_bits;
2209         cap->events_mask        = (unsigned int)x86_pmu.events_maskl;
2210         cap->events_mask_len    = x86_pmu.events_mask_len;
2211 }
2212 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);