arch/x86/events/core.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
   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 atomic_t pmc_refcount;
 139 static DEFINE_MUTEX(pmc_reserve_mutex);
 140
 141 #ifdef CONFIG_X86_LOCAL_APIC
 142
 143 static bool reserve_pmc_hardware(void)
 144 {
 145         int i;
 146
 147         for (i = 0; i < x86_pmu.num_counters; i++) {
 148                 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
 149                         goto perfctr_fail;
 150         }
 151
 152         for (i = 0; i < x86_pmu.num_counters; i++) {
 153                 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
 154                         goto eventsel_fail;
 155         }
 156
 157         return true;
 158
 159 eventsel_fail:
 160         for (i--; i >= 0; i--)
 161                 release_evntsel_nmi(x86_pmu_config_addr(i));
 162
 163         i = x86_pmu.num_counters;
 164
 165 perfctr_fail:
 166         for (i--; i >= 0; i--)
 167                 release_perfctr_nmi(x86_pmu_event_addr(i));
 168
 169         return false;
 170 }
 171
 172 static void release_pmc_hardware(void)
 173 {
 174         int i;
 175
 176         for (i = 0; i < x86_pmu.num_counters; i++) {
 177                 release_perfctr_nmi(x86_pmu_event_addr(i));
 178                 release_evntsel_nmi(x86_pmu_config_addr(i));
 179         }
 180 }
 181
 182 #else
 183
 184 static bool reserve_pmc_hardware(void) { return true; }
 185 static void release_pmc_hardware(void) {}
 186
 187 #endif
 188
 189 static bool check_hw_exists(void)
 190 {
 191         u64 val, val_fail, val_new= ~0;
 192         int i, reg, reg_fail, ret = 0;
 193         int bios_fail = 0;
 194         int reg_safe = -1;
 195
 196         /*
 197          * Check to see if the BIOS enabled any of the counters, if so
 198          * complain and bail.
 199          */
 200         for (i = 0; i < x86_pmu.num_counters; i++) {
 201                 reg = x86_pmu_config_addr(i);
 202                 ret = rdmsrl_safe(reg, &val);
 203                 if (ret)
 204                         goto msr_fail;
 205                 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
 206                         bios_fail = 1;
 207                         val_fail = val;
 208                         reg_fail = reg;
 209                 } else {
 210                         reg_safe = i;
 211                 }
 212         }
 213
 214         if (x86_pmu.num_counters_fixed) {
 215                 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 216                 ret = rdmsrl_safe(reg, &val);
 217                 if (ret)
 218                         goto msr_fail;
 219                 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
 220                         if (val & (0x03 << i*4)) {
 221                                 bios_fail = 1;
 222                                 val_fail = val;
 223                                 reg_fail = reg;
 224                         }
 225                 }
 226         }
 227
 228         /*
 229          * If all the counters are enabled, the below test will always
 230          * fail.  The tools will also become useless in this scenario.
 231          * Just fail and disable the hardware counters.
 232          */
 233
 234         if (reg_safe == -1) {
 235                 reg = reg_safe;
 236                 goto msr_fail;
 237         }
 238
 239         /*
 240          * Read the current value, change it and read it back to see if it
 241          * matches, this is needed to detect certain hardware emulators
 242          * (qemu/kvm) that don't trap on the MSR access and always return 0s.
 243          */
 244         reg = x86_pmu_event_addr(reg_safe);
 245         if (rdmsrl_safe(reg, &val))
 246                 goto msr_fail;
 247         val ^= 0xffffUL;
 248         ret = wrmsrl_safe(reg, val);
 249         ret |= rdmsrl_safe(reg, &val_new);
 250         if (ret || val != val_new)
 251                 goto msr_fail;
 252
 253         /*
 254          * We still allow the PMU driver to operate:
 255          */
 256         if (bios_fail) {
 257                 pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
 258                 pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
 259                               reg_fail, val_fail);
 260         }
 261
 262         return true;
 263
 264 msr_fail:
 265         pr_cont("Broken PMU hardware detected, using software events only.\n");
 266         printk("%sFailed to access perfctr msr (MSR %x is %Lx)\n",
 267                 boot_cpu_has(X86_FEATURE_HYPERVISOR) ? KERN_INFO : KERN_ERR,
 268                 reg, val_new);
 269
 270         return false;
 271 }
 272
 273 static void hw_perf_event_destroy(struct perf_event *event)
 274 {
 275         x86_release_hardware();
 276         atomic_dec(&active_events);
 277 }
 278
 279 void hw_perf_lbr_event_destroy(struct perf_event *event)
 280 {
 281         hw_perf_event_destroy(event);
 282
 283         /* undo the lbr/bts event accounting */
 284         x86_del_exclusive(x86_lbr_exclusive_lbr);
 285 }
 286
 287 static inline int x86_pmu_initialized(void)
 288 {
 289         return x86_pmu.handle_irq != NULL;
 290 }
 291
 292 static inline int
 293 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
 294 {
 295         struct perf_event_attr *attr = &event->attr;
 296         unsigned int cache_type, cache_op, cache_result;
 297         u64 config, val;
 298
 299         config = attr->config;
 300
 301         cache_type = (config >>  0) & 0xff;
 302         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 303                 return -EINVAL;
 304
 305         cache_op = (config >>  8) & 0xff;
 306         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 307                 return -EINVAL;
 308
 309         cache_result = (config >> 16) & 0xff;
 310         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 311                 return -EINVAL;
 312
 313         val = hw_cache_event_ids[cache_type][cache_op][cache_result];
 314
 315         if (val == 0)
 316                 return -ENOENT;
 317
 318         if (val == -1)
 319                 return -EINVAL;
 320
 321         hwc->config |= val;
 322         attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
 323         return x86_pmu_extra_regs(val, event);
 324 }
 325
 326 int x86_reserve_hardware(void)
 327 {
 328         int err = 0;
 329
 330         if (!atomic_inc_not_zero(&pmc_refcount)) {
 331                 mutex_lock(&pmc_reserve_mutex);
 332                 if (atomic_read(&pmc_refcount) == 0) {
 333                         if (!reserve_pmc_hardware())
 334                                 err = -EBUSY;
 335                         else
 336                                 reserve_ds_buffers();
 337                 }
 338                 if (!err)
 339                         atomic_inc(&pmc_refcount);
 340                 mutex_unlock(&pmc_reserve_mutex);
 341         }
 342
 343         return err;
 344 }
 345
 346 void x86_release_hardware(void)
 347 {
 348         if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
 349                 release_pmc_hardware();
 350                 release_ds_buffers();
 351                 mutex_unlock(&pmc_reserve_mutex);
 352         }
 353 }
 354
 355 /*
 356  * Check if we can create event of a certain type (that no conflicting events
 357  * are present).
 358  */
 359 int x86_add_exclusive(unsigned int what)
 360 {
 361         int i;
 362
 363         if (x86_pmu.lbr_pt_coexist)
 364                 return 0;
 365
 366         if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
 367                 mutex_lock(&pmc_reserve_mutex);
 368                 for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
 369                         if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
 370                                 goto fail_unlock;
 371                 }
 372                 atomic_inc(&x86_pmu.lbr_exclusive[what]);
 373                 mutex_unlock(&pmc_reserve_mutex);
 374         }
 375
 376         atomic_inc(&active_events);
 377         return 0;
 378
 379 fail_unlock:
 380         mutex_unlock(&pmc_reserve_mutex);
 381         return -EBUSY;
 382 }
 383
 384 void x86_del_exclusive(unsigned int what)
 385 {
 386         if (x86_pmu.lbr_pt_coexist)
 387                 return;
 388
 389         atomic_dec(&x86_pmu.lbr_exclusive[what]);
 390         atomic_dec(&active_events);
 391 }
 392
 393 int x86_setup_perfctr(struct perf_event *event)
 394 {
 395         struct perf_event_attr *attr = &event->attr;
 396         struct hw_perf_event *hwc = &event->hw;
 397         u64 config;
 398
 399         if (!is_sampling_event(event)) {
 400                 hwc->sample_period = x86_pmu.max_period;
 401                 hwc->last_period = hwc->sample_period;
 402                 local64_set(&hwc->period_left, hwc->sample_period);
 403         }
 404
 405         if (attr->type == PERF_TYPE_RAW)
 406                 return x86_pmu_extra_regs(event->attr.config, event);
 407
 408         if (attr->type == PERF_TYPE_HW_CACHE)
 409                 return set_ext_hw_attr(hwc, event);
 410
 411         if (attr->config >= x86_pmu.max_events)
 412                 return -EINVAL;
 413
 414         /*
 415          * The generic map:
 416          */
 417         config = x86_pmu.event_map(attr->config);
 418
 419         if (config == 0)
 420                 return -ENOENT;
 421
 422         if (config == -1LL)
 423                 return -EINVAL;
 424
 425         /*
 426          * Branch tracing:
 427          */
 428         if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
 429             !attr->freq && hwc->sample_period == 1) {
 430                 /* BTS is not supported by this architecture. */
 431                 if (!x86_pmu.bts_active)
 432                         return -EOPNOTSUPP;
 433
 434                 /* BTS is currently only allowed for user-mode. */
 435                 if (!attr->exclude_kernel)
 436                         return -EOPNOTSUPP;
 437
 438                 /* disallow bts if conflicting events are present */
 439                 if (x86_add_exclusive(x86_lbr_exclusive_lbr))
 440                         return -EBUSY;
 441
 442                 event->destroy = hw_perf_lbr_event_destroy;
 443         }
 444
 445         hwc->config |= config;
 446
 447         return 0;
 448 }
 449
 450 /*
 451  * check that branch_sample_type is compatible with
 452  * settings needed for precise_ip > 1 which implies
 453  * using the LBR to capture ALL taken branches at the
 454  * priv levels of the measurement
 455  */
 456 static inline int precise_br_compat(struct perf_event *event)
 457 {
 458         u64 m = event->attr.branch_sample_type;
 459         u64 b = 0;
 460
 461         /* must capture all branches */
 462         if (!(m & PERF_SAMPLE_BRANCH_ANY))
 463                 return 0;
 464
 465         m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
 466
 467         if (!event->attr.exclude_user)
 468                 b |= PERF_SAMPLE_BRANCH_USER;
 469
 470         if (!event->attr.exclude_kernel)
 471                 b |= PERF_SAMPLE_BRANCH_KERNEL;
 472
 473         /*
 474          * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
 475          */
 476
 477         return m == b;
 478 }
 479
 480 int x86_pmu_hw_config(struct perf_event *event)
 481 {
 482         if (event->attr.precise_ip) {
 483                 int precise = 0;
 484
 485                 /* Support for constant skid */
 486                 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
 487                         precise++;
 488
 489                         /* Support for IP fixup */
 490                         if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
 491                                 precise++;
 492
 493                         if (x86_pmu.pebs_prec_dist)
 494                                 precise++;
 495                 }
 496
 497                 if (event->attr.precise_ip > precise)
 498                         return -EOPNOTSUPP;
 499         }
 500         /*
 501          * check that PEBS LBR correction does not conflict with
 502          * whatever the user is asking with attr->branch_sample_type
 503          */
 504         if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
 505                 u64 *br_type = &event->attr.branch_sample_type;
 506
 507                 if (has_branch_stack(event)) {
 508                         if (!precise_br_compat(event))
 509                                 return -EOPNOTSUPP;
 510
 511                         /* branch_sample_type is compatible */
 512
 513                 } else {
 514                         /*
 515                          * user did not specify  branch_sample_type
 516                          *
 517                          * For PEBS fixups, we capture all
 518                          * the branches at the priv level of the
 519                          * event.
 520                          */
 521                         *br_type = PERF_SAMPLE_BRANCH_ANY;
 522
 523                         if (!event->attr.exclude_user)
 524                                 *br_type |= PERF_SAMPLE_BRANCH_USER;
 525
 526                         if (!event->attr.exclude_kernel)
 527                                 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
 528                 }
 529         }
 530
 531         if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
 532                 event->attach_state |= PERF_ATTACH_TASK_DATA;
 533
 534         /*
 535          * Generate PMC IRQs:
 536          * (keep 'enabled' bit clear for now)
 537          */
 538         event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
 539
 540         /*
 541          * Count user and OS events unless requested not to
 542          */
 543         if (!event->attr.exclude_user)
 544                 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
 545         if (!event->attr.exclude_kernel)
 546                 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
 547
 548         if (event->attr.type == PERF_TYPE_RAW)
 549                 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
 550
 551         if (event->attr.sample_period && x86_pmu.limit_period) {
 552                 if (x86_pmu.limit_period(event, event->attr.sample_period) >
 553                                 event->attr.sample_period)
 554                         return -EINVAL;
 555         }
 556
 557         return x86_setup_perfctr(event);
 558 }
 559
 560 /*
 561  * Setup the hardware configuration for a given attr_type
 562  */
 563 static int __x86_pmu_event_init(struct perf_event *event)
 564 {
 565         int err;
 566
 567         if (!x86_pmu_initialized())
 568                 return -ENODEV;
 569
 570         err = x86_reserve_hardware();
 571         if (err)
 572                 return err;
 573
 574         atomic_inc(&active_events);
 575         event->destroy = hw_perf_event_destroy;
 576
 577         event->hw.idx = -1;
 578         event->hw.last_cpu = -1;
 579         event->hw.last_tag = ~0ULL;
 580
 581         /* mark unused */
 582         event->hw.extra_reg.idx = EXTRA_REG_NONE;
 583         event->hw.branch_reg.idx = EXTRA_REG_NONE;
 584
 585         return x86_pmu.hw_config(event);
 586 }
 587
 588 void x86_pmu_disable_all(void)
 589 {
 590         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 591         int idx;
 592
 593         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 594                 u64 val;
 595
 596                 if (!test_bit(idx, cpuc->active_mask))
 597                         continue;
 598                 rdmsrl(x86_pmu_config_addr(idx), val);
 599                 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
 600                         continue;
 601                 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 602                 wrmsrl(x86_pmu_config_addr(idx), val);
 603         }
 604 }
 605
 606 /*
 607  * There may be PMI landing after enabled=0. The PMI hitting could be before or
 608  * after disable_all.
 609  *
 610  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
 611  * It will not be re-enabled in the NMI handler again, because enabled=0. After
 612  * handling the NMI, disable_all will be called, which will not change the
 613  * state either. If PMI hits after disable_all, the PMU is already disabled
 614  * before entering NMI handler. The NMI handler will not change the state
 615  * either.
 616  *
 617  * So either situation is harmless.
 618  */
 619 static void x86_pmu_disable(struct pmu *pmu)
 620 {
 621         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 622
 623         if (!x86_pmu_initialized())
 624                 return;
 625
 626         if (!cpuc->enabled)
 627                 return;
 628
 629         cpuc->n_added = 0;
 630         cpuc->enabled = 0;
 631         barrier();
 632
 633         x86_pmu.disable_all();
 634 }
 635
 636 void x86_pmu_enable_all(int added)
 637 {
 638         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 639         int idx;
 640
 641         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 642                 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
 643
 644                 if (!test_bit(idx, cpuc->active_mask))
 645                         continue;
 646
 647                 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
 648         }
 649 }
 650
 651 static struct pmu pmu;
 652
 653 static inline int is_x86_event(struct perf_event *event)
 654 {
 655         return event->pmu == &pmu;
 656 }
 657
 658 /*
 659  * Event scheduler state:
 660  *
 661  * Assign events iterating over all events and counters, beginning
 662  * with events with least weights first. Keep the current iterator
 663  * state in struct sched_state.
 664  */
 665 struct sched_state {
 666         int     weight;
 667         int     event;          /* event index */
 668         int     counter;        /* counter index */
 669         int     unassigned;     /* number of events to be assigned left */
 670         int     nr_gp;          /* number of GP counters used */
 671         unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 672 };
 673
 674 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
 675 #define SCHED_STATES_MAX        2
 676
 677 struct perf_sched {
 678         int                     max_weight;
 679         int                     max_events;
 680         int                     max_gp;
 681         int                     saved_states;
 682         struct event_constraint **constraints;
 683         struct sched_state      state;
 684         struct sched_state      saved[SCHED_STATES_MAX];
 685 };
 686
 687 /*
 688  * Initialize interator that runs through all events and counters.
 689  */
 690 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
 691                             int num, int wmin, int wmax, int gpmax)
 692 {
 693         int idx;
 694
 695         memset(sched, 0, sizeof(*sched));
 696         sched->max_events       = num;
 697         sched->max_weight       = wmax;
 698         sched->max_gp           = gpmax;
 699         sched->constraints      = constraints;
 700
 701         for (idx = 0; idx < num; idx++) {
 702                 if (constraints[idx]->weight == wmin)
 703                         break;
 704         }
 705
 706         sched->state.event      = idx;          /* start with min weight */
 707         sched->state.weight     = wmin;
 708         sched->state.unassigned = num;
 709 }
 710
 711 static void perf_sched_save_state(struct perf_sched *sched)
 712 {
 713         if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
 714                 return;
 715
 716         sched->saved[sched->saved_states] = sched->state;
 717         sched->saved_states++;
 718 }
 719
 720 static bool perf_sched_restore_state(struct perf_sched *sched)
 721 {
 722         if (!sched->saved_states)
 723                 return false;
 724
 725         sched->saved_states--;
 726         sched->state = sched->saved[sched->saved_states];
 727
 728         /* continue with next counter: */
 729         clear_bit(sched->state.counter++, sched->state.used);
 730
 731         return true;
 732 }
 733
 734 /*
 735  * Select a counter for the current event to schedule. Return true on
 736  * success.
 737  */
 738 static bool __perf_sched_find_counter(struct perf_sched *sched)
 739 {
 740         struct event_constraint *c;
 741         int idx;
 742
 743         if (!sched->state.unassigned)
 744                 return false;
 745
 746         if (sched->state.event >= sched->max_events)
 747                 return false;
 748
 749         c = sched->constraints[sched->state.event];
 750         /* Prefer fixed purpose counters */
 751         if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
 752                 idx = INTEL_PMC_IDX_FIXED;
 753                 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
 754                         if (!__test_and_set_bit(idx, sched->state.used))
 755                                 goto done;
 756                 }
 757         }
 758
 759         /* Grab the first unused counter starting with idx */
 760         idx = sched->state.counter;
 761         for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
 762                 if (!__test_and_set_bit(idx, sched->state.used)) {
 763                         if (sched->state.nr_gp++ >= sched->max_gp)
 764                                 return false;
 765
 766                         goto done;
 767                 }
 768         }
 769
 770         return false;
 771
 772 done:
 773         sched->state.counter = idx;
 774
 775         if (c->overlap)
 776                 perf_sched_save_state(sched);
 777
 778         return true;
 779 }
 780
 781 static bool perf_sched_find_counter(struct perf_sched *sched)
 782 {
 783         while (!__perf_sched_find_counter(sched)) {
 784                 if (!perf_sched_restore_state(sched))
 785                         return false;
 786         }
 787
 788         return true;
 789 }
 790
 791 /*
 792  * Go through all unassigned events and find the next one to schedule.
 793  * Take events with the least weight first. Return true on success.
 794  */
 795 static bool perf_sched_next_event(struct perf_sched *sched)
 796 {
 797         struct event_constraint *c;
 798
 799         if (!sched->state.unassigned || !--sched->state.unassigned)
 800                 return false;
 801
 802         do {
 803                 /* next event */
 804                 sched->state.event++;
 805                 if (sched->state.event >= sched->max_events) {
 806                         /* next weight */
 807                         sched->state.event = 0;
 808                         sched->state.weight++;
 809                         if (sched->state.weight > sched->max_weight)
 810                                 return false;
 811                 }
 812                 c = sched->constraints[sched->state.event];
 813         } while (c->weight != sched->state.weight);
 814
 815         sched->state.counter = 0;       /* start with first counter */
 816
 817         return true;
 818 }
 819
 820 /*
 821  * Assign a counter for each event.
 822  */
 823 int perf_assign_events(struct event_constraint **constraints, int n,
 824                         int wmin, int wmax, int gpmax, int *assign)
 825 {
 826         struct perf_sched sched;
 827
 828         perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
 829
 830         do {
 831                 if (!perf_sched_find_counter(&sched))
 832                         break;  /* failed */
 833                 if (assign)
 834                         assign[sched.state.event] = sched.state.counter;
 835         } while (perf_sched_next_event(&sched));
 836
 837         return sched.state.unassigned;
 838 }
 839 EXPORT_SYMBOL_GPL(perf_assign_events);
 840
 841 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 842 {
 843         struct event_constraint *c;
 844         unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 845         struct perf_event *e;
 846         int i, wmin, wmax, unsched = 0;
 847         struct hw_perf_event *hwc;
 848
 849         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
 850
 851         if (x86_pmu.start_scheduling)
 852                 x86_pmu.start_scheduling(cpuc);
 853
 854         for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
 855                 cpuc->event_constraint[i] = NULL;
 856                 c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
 857                 cpuc->event_constraint[i] = c;
 858
 859                 wmin = min(wmin, c->weight);
 860                 wmax = max(wmax, c->weight);
 861         }
 862
 863         /*
 864          * fastpath, try to reuse previous register
 865          */
 866         for (i = 0; i < n; i++) {
 867                 hwc = &cpuc->event_list[i]->hw;
 868                 c = cpuc->event_constraint[i];
 869
 870                 /* never assigned */
 871                 if (hwc->idx == -1)
 872                         break;
 873
 874                 /* constraint still honored */
 875                 if (!test_bit(hwc->idx, c->idxmsk))
 876                         break;
 877
 878                 /* not already used */
 879                 if (test_bit(hwc->idx, used_mask))
 880                         break;
 881
 882                 __set_bit(hwc->idx, used_mask);
 883                 if (assign)
 884                         assign[i] = hwc->idx;
 885         }
 886
 887         /* slow path */
 888         if (i != n) {
 889                 int gpmax = x86_pmu.num_counters;
 890
 891                 /*
 892                  * Do not allow scheduling of more than half the available
 893                  * generic counters.
 894                  *
 895                  * This helps avoid counter starvation of sibling thread by
 896                  * ensuring at most half the counters cannot be in exclusive
 897                  * mode. There is no designated counters for the limits. Any
 898                  * N/2 counters can be used. This helps with events with
 899                  * specific counter constraints.
 900                  */
 901                 if (is_ht_workaround_enabled() && !cpuc->is_fake &&
 902                     READ_ONCE(cpuc->excl_cntrs->exclusive_present))
 903                         gpmax /= 2;
 904
 905                 unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
 906                                              wmax, gpmax, assign);
 907         }
 908
 909         /*
 910          * In case of success (unsched = 0), mark events as committed,
 911          * so we do not put_constraint() in case new events are added
 912          * and fail to be scheduled
 913          *
 914          * We invoke the lower level commit callback to lock the resource
 915          *
 916          * We do not need to do all of this in case we are called to
 917          * validate an event group (assign == NULL)
 918          */
 919         if (!unsched && assign) {
 920                 for (i = 0; i < n; i++) {
 921                         e = cpuc->event_list[i];
 922                         e->hw.flags |= PERF_X86_EVENT_COMMITTED;
 923                         if (x86_pmu.commit_scheduling)
 924                                 x86_pmu.commit_scheduling(cpuc, i, assign[i]);
 925                 }
 926         } else {
 927                 for (i = 0; i < n; i++) {
 928                         e = cpuc->event_list[i];
 929                         /*
 930                          * do not put_constraint() on comitted events,
 931                          * because they are good to go
 932                          */
 933                         if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
 934                                 continue;
 935
 936                         /*
 937                          * release events that failed scheduling
 938                          */
 939                         if (x86_pmu.put_event_constraints)
 940                                 x86_pmu.put_event_constraints(cpuc, e);
 941                 }
 942         }
 943
 944         if (x86_pmu.stop_scheduling)
 945                 x86_pmu.stop_scheduling(cpuc);
 946
 947         return unsched ? -EINVAL : 0;
 948 }
 949
 950 /*
 951  * dogrp: true if must collect siblings events (group)
 952  * returns total number of events and error code
 953  */
 954 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
 955 {
 956         struct perf_event *event;
 957         int n, max_count;
 958
 959         max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
 960
 961         /* current number of events already accepted */
 962         n = cpuc->n_events;
 963
 964         if (is_x86_event(leader)) {
 965                 if (n >= max_count)
 966                         return -EINVAL;
 967                 cpuc->event_list[n] = leader;
 968                 n++;
 969         }
 970         if (!dogrp)
 971                 return n;
 972
 973         list_for_each_entry(event, &leader->sibling_list, group_entry) {
 974                 if (!is_x86_event(event) ||
 975                     event->state <= PERF_EVENT_STATE_OFF)
 976                         continue;
 977
 978                 if (n >= max_count)
 979                         return -EINVAL;
 980
 981                 cpuc->event_list[n] = event;
 982                 n++;
 983         }
 984         return n;
 985 }
 986
 987 static inline void x86_assign_hw_event(struct perf_event *event,
 988                                 struct cpu_hw_events *cpuc, int i)
 989 {
 990         struct hw_perf_event *hwc = &event->hw;
 991
 992         hwc->idx = cpuc->assign[i];
 993         hwc->last_cpu = smp_processor_id();
 994         hwc->last_tag = ++cpuc->tags[i];
 995
 996         if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
 997                 hwc->config_base = 0;
 998                 hwc->event_base = 0;
 999         } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1000                 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1001                 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1002                 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1003         } else {
1004                 hwc->config_base = x86_pmu_config_addr(hwc->idx);
1005                 hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1006                 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1007         }
1008 }
1009
1010 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1011                                         struct cpu_hw_events *cpuc,
1012                                         int i)
1013 {
1014         return hwc->idx == cpuc->assign[i] &&
1015                 hwc->last_cpu == smp_processor_id() &&
1016                 hwc->last_tag == cpuc->tags[i];
1017 }
1018
1019 static void x86_pmu_start(struct perf_event *event, int flags);
1020
1021 static void x86_pmu_enable(struct pmu *pmu)
1022 {
1023         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1024         struct perf_event *event;
1025         struct hw_perf_event *hwc;
1026         int i, added = cpuc->n_added;
1027
1028         if (!x86_pmu_initialized())
1029                 return;
1030
1031         if (cpuc->enabled)
1032                 return;
1033
1034         if (cpuc->n_added) {
1035                 int n_running = cpuc->n_events - cpuc->n_added;
1036                 /*
1037                  * apply assignment obtained either from
1038                  * hw_perf_group_sched_in() or x86_pmu_enable()
1039                  *
1040                  * step1: save events moving to new counters
1041                  */
1042                 for (i = 0; i < n_running; i++) {
1043                         event = cpuc->event_list[i];
1044                         hwc = &event->hw;
1045
1046                         /*
1047                          * we can avoid reprogramming counter if:
1048                          * - assigned same counter as last time
1049                          * - running on same CPU as last time
1050                          * - no other event has used the counter since
1051                          */
1052                         if (hwc->idx == -1 ||
1053                             match_prev_assignment(hwc, cpuc, i))
1054                                 continue;
1055
1056                         /*
1057                          * Ensure we don't accidentally enable a stopped
1058                          * counter simply because we rescheduled.
1059                          */
1060                         if (hwc->state & PERF_HES_STOPPED)
1061                                 hwc->state |= PERF_HES_ARCH;
1062
1063                         x86_pmu_stop(event, PERF_EF_UPDATE);
1064                 }
1065
1066                 /*
1067                  * step2: reprogram moved events into new counters
1068                  */
1069                 for (i = 0; i < cpuc->n_events; i++) {
1070                         event = cpuc->event_list[i];
1071                         hwc = &event->hw;
1072
1073                         if (!match_prev_assignment(hwc, cpuc, i))
1074                                 x86_assign_hw_event(event, cpuc, i);
1075                         else if (i < n_running)
1076                                 continue;
1077
1078                         if (hwc->state & PERF_HES_ARCH)
1079                                 continue;
1080
1081                         x86_pmu_start(event, PERF_EF_RELOAD);
1082                 }
1083                 cpuc->n_added = 0;
1084                 perf_events_lapic_init();
1085         }
1086
1087         cpuc->enabled = 1;
1088         barrier();
1089
1090         x86_pmu.enable_all(added);
1091 }
1092
1093 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1094
1095 /*
1096  * Set the next IRQ period, based on the hwc->period_left value.
1097  * To be called with the event disabled in hw:
1098  */
1099 int x86_perf_event_set_period(struct perf_event *event)
1100 {
1101         struct hw_perf_event *hwc = &event->hw;
1102         s64 left = local64_read(&hwc->period_left);
1103         s64 period = hwc->sample_period;
1104         int ret = 0, idx = hwc->idx;
1105
1106         if (idx == INTEL_PMC_IDX_FIXED_BTS)
1107                 return 0;
1108
1109         /*
1110          * If we are way outside a reasonable range then just skip forward:
1111          */
1112         if (unlikely(left <= -period)) {
1113                 left = period;
1114                 local64_set(&hwc->period_left, left);
1115                 hwc->last_period = period;
1116                 ret = 1;
1117         }
1118
1119         if (unlikely(left <= 0)) {
1120                 left += period;
1121                 local64_set(&hwc->period_left, left);
1122                 hwc->last_period = period;
1123                 ret = 1;
1124         }
1125         /*
1126          * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1127          */
1128         if (unlikely(left < 2))
1129                 left = 2;
1130
1131         if (left > x86_pmu.max_period)
1132                 left = x86_pmu.max_period;
1133
1134         if (x86_pmu.limit_period)
1135                 left = x86_pmu.limit_period(event, left);
1136
1137         per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1138
1139         if (!(hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) ||
1140             local64_read(&hwc->prev_count) != (u64)-left) {
1141                 /*
1142                  * The hw event starts counting from this event offset,
1143                  * mark it to be able to extra future deltas:
1144                  */
1145                 local64_set(&hwc->prev_count, (u64)-left);
1146
1147                 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1148         }
1149
1150         /*
1151          * Due to erratum on certan cpu we need
1152          * a second write to be sure the register
1153          * is updated properly
1154          */
1155         if (x86_pmu.perfctr_second_write) {
1156                 wrmsrl(hwc->event_base,
1157                         (u64)(-left) & x86_pmu.cntval_mask);
1158         }
1159
1160         perf_event_update_userpage(event);
1161
1162         return ret;
1163 }
1164
1165 void x86_pmu_enable_event(struct perf_event *event)
1166 {
1167         if (__this_cpu_read(cpu_hw_events.enabled))
1168                 __x86_pmu_enable_event(&event->hw,
1169                                        ARCH_PERFMON_EVENTSEL_ENABLE);
1170 }
1171
1172 /*
1173  * Add a single event to the PMU.
1174  *
1175  * The event is added to the group of enabled events
1176  * but only if it can be scehduled with existing events.
1177  */
1178 static int x86_pmu_add(struct perf_event *event, int flags)
1179 {
1180         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1181         struct hw_perf_event *hwc;
1182         int assign[X86_PMC_IDX_MAX];
1183         int n, n0, ret;
1184
1185         hwc = &event->hw;
1186
1187         n0 = cpuc->n_events;
1188         ret = n = collect_events(cpuc, event, false);
1189         if (ret < 0)
1190                 goto out;
1191
1192         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1193         if (!(flags & PERF_EF_START))
1194                 hwc->state |= PERF_HES_ARCH;
1195
1196         /*
1197          * If group events scheduling transaction was started,
1198          * skip the schedulability test here, it will be performed
1199          * at commit time (->commit_txn) as a whole.
1200          */
1201         if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1202                 goto done_collect;
1203
1204         ret = x86_pmu.schedule_events(cpuc, n, assign);
1205         if (ret)
1206                 goto out;
1207         /*
1208          * copy new assignment, now we know it is possible
1209          * will be used by hw_perf_enable()
1210          */
1211         memcpy(cpuc->assign, assign, n*sizeof(int));
1212
1213 done_collect:
1214         /*
1215          * Commit the collect_events() state. See x86_pmu_del() and
1216          * x86_pmu_*_txn().
1217          */
1218         cpuc->n_events = n;
1219         cpuc->n_added += n - n0;
1220         cpuc->n_txn += n - n0;
1221
1222         ret = 0;
1223 out:
1224         return ret;
1225 }
1226
1227 static void x86_pmu_start(struct perf_event *event, int flags)
1228 {
1229         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1230         int idx = event->hw.idx;
1231
1232         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1233                 return;
1234
1235         if (WARN_ON_ONCE(idx == -1))
1236                 return;
1237
1238         if (flags & PERF_EF_RELOAD) {
1239                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1240                 x86_perf_event_set_period(event);
1241         }
1242
1243         event->hw.state = 0;
1244
1245         cpuc->events[idx] = event;
1246         __set_bit(idx, cpuc->active_mask);
1247         __set_bit(idx, cpuc->running);
1248         x86_pmu.enable(event);
1249         perf_event_update_userpage(event);
1250 }
1251
1252 void perf_event_print_debug(void)
1253 {
1254         u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1255         u64 pebs, debugctl;
1256         struct cpu_hw_events *cpuc;
1257         unsigned long flags;
1258         int cpu, idx;
1259
1260         if (!x86_pmu.num_counters)
1261                 return;
1262
1263         local_irq_save(flags);
1264
1265         cpu = smp_processor_id();
1266         cpuc = &per_cpu(cpu_hw_events, cpu);
1267
1268         if (x86_pmu.version >= 2) {
1269                 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1270                 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1271                 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1272                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1273
1274                 pr_info("\n");
1275                 pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1276                 pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1277                 pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1278                 pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1279                 if (x86_pmu.pebs_constraints) {
1280                         rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1281                         pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1282                 }
1283                 if (x86_pmu.lbr_nr) {
1284                         rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1285                         pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1286                 }
1287         }
1288         pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1289
1290         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1291                 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1292                 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1293
1294                 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1295
1296                 pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1297                         cpu, idx, pmc_ctrl);
1298                 pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1299                         cpu, idx, pmc_count);
1300                 pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1301                         cpu, idx, prev_left);
1302         }
1303         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1304                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1305
1306                 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1307                         cpu, idx, pmc_count);
1308         }
1309         local_irq_restore(flags);
1310 }
1311
1312 void x86_pmu_stop(struct perf_event *event, int flags)
1313 {
1314         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1315         struct hw_perf_event *hwc = &event->hw;
1316
1317         if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1318                 x86_pmu.disable(event);
1319                 cpuc->events[hwc->idx] = NULL;
1320                 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1321                 hwc->state |= PERF_HES_STOPPED;
1322         }
1323
1324         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1325                 /*
1326                  * Drain the remaining delta count out of a event
1327                  * that we are disabling:
1328                  */
1329                 x86_perf_event_update(event);
1330                 hwc->state |= PERF_HES_UPTODATE;
1331         }
1332 }
1333
1334 static void x86_pmu_del(struct perf_event *event, int flags)
1335 {
1336         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1337         int i;
1338
1339         /*
1340          * event is descheduled
1341          */
1342         event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1343
1344         /*
1345          * If we're called during a txn, we don't need to do anything.
1346          * The events never got scheduled and ->cancel_txn will truncate
1347          * the event_list.
1348          *
1349          * XXX assumes any ->del() called during a TXN will only be on
1350          * an event added during that same TXN.
1351          */
1352         if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1353                 return;
1354
1355         /*
1356          * Not a TXN, therefore cleanup properly.
1357          */
1358         x86_pmu_stop(event, PERF_EF_UPDATE);
1359
1360         for (i = 0; i < cpuc->n_events; i++) {
1361                 if (event == cpuc->event_list[i])
1362                         break;
1363         }
1364
1365         if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1366                 return;
1367
1368         /* If we have a newly added event; make sure to decrease n_added. */
1369         if (i >= cpuc->n_events - cpuc->n_added)
1370                 --cpuc->n_added;
1371
1372         if (x86_pmu.put_event_constraints)
1373                 x86_pmu.put_event_constraints(cpuc, event);
1374
1375         /* Delete the array entry. */
1376         while (++i < cpuc->n_events) {
1377                 cpuc->event_list[i-1] = cpuc->event_list[i];
1378                 cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1379         }
1380         --cpuc->n_events;
1381
1382         perf_event_update_userpage(event);
1383 }
1384
1385 int x86_pmu_handle_irq(struct pt_regs *regs)
1386 {
1387         struct perf_sample_data data;
1388         struct cpu_hw_events *cpuc;
1389         struct perf_event *event;
1390         int idx, handled = 0;
1391         u64 val;
1392
1393         cpuc = this_cpu_ptr(&cpu_hw_events);
1394
1395         /*
1396          * Some chipsets need to unmask the LVTPC in a particular spot
1397          * inside the nmi handler.  As a result, the unmasking was pushed
1398          * into all the nmi handlers.
1399          *
1400          * This generic handler doesn't seem to have any issues where the
1401          * unmasking occurs so it was left at the top.
1402          */
1403         apic_write(APIC_LVTPC, APIC_DM_NMI);
1404
1405         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1406                 if (!test_bit(idx, cpuc->active_mask)) {
1407                         /*
1408                          * Though we deactivated the counter some cpus
1409                          * might still deliver spurious interrupts still
1410                          * in flight. Catch them:
1411                          */
1412                         if (__test_and_clear_bit(idx, cpuc->running))
1413                                 handled++;
1414                         continue;
1415                 }
1416
1417                 event = cpuc->events[idx];
1418
1419                 val = x86_perf_event_update(event);
1420                 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1421                         continue;
1422
1423                 /*
1424                  * event overflow
1425                  */
1426                 handled++;
1427                 perf_sample_data_init(&data, 0, event->hw.last_period);
1428
1429                 if (!x86_perf_event_set_period(event))
1430                         continue;
1431
1432                 if (perf_event_overflow(event, &data, regs))
1433                         x86_pmu_stop(event, 0);
1434         }
1435
1436         if (handled)
1437                 inc_irq_stat(apic_perf_irqs);
1438
1439         return handled;
1440 }
1441
1442 void perf_events_lapic_init(void)
1443 {
1444         if (!x86_pmu.apic || !x86_pmu_initialized())
1445                 return;
1446
1447         /*
1448          * Always use NMI for PMU
1449          */
1450         apic_write(APIC_LVTPC, APIC_DM_NMI);
1451 }
1452
1453 static int
1454 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1455 {
1456         u64 start_clock;
1457         u64 finish_clock;
1458         int ret;
1459
1460         /*
1461          * All PMUs/events that share this PMI handler should make sure to
1462          * increment active_events for their events.
1463          */
1464         if (!atomic_read(&active_events))
1465                 return NMI_DONE;
1466
1467         start_clock = sched_clock();
1468         ret = x86_pmu.handle_irq(regs);
1469         finish_clock = sched_clock();
1470
1471         perf_sample_event_took(finish_clock - start_clock);
1472
1473         return ret;
1474 }
1475 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1476
1477 struct event_constraint emptyconstraint;
1478 struct event_constraint unconstrained;
1479
1480 static int
1481 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1482 {
1483         unsigned int cpu = (long)hcpu;
1484         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1485         int i, ret = NOTIFY_OK;
1486
1487         switch (action & ~CPU_TASKS_FROZEN) {
1488         case CPU_UP_PREPARE:
1489                 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1490                         cpuc->kfree_on_online[i] = NULL;
1491                 if (x86_pmu.cpu_prepare)
1492                         ret = x86_pmu.cpu_prepare(cpu);
1493                 break;
1494
1495         case CPU_STARTING:
1496                 if (x86_pmu.cpu_starting)
1497                         x86_pmu.cpu_starting(cpu);
1498                 break;
1499
1500         case CPU_ONLINE:
1501                 for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1502                         kfree(cpuc->kfree_on_online[i]);
1503                         cpuc->kfree_on_online[i] = NULL;
1504                 }
1505                 break;
1506
1507         case CPU_DYING:
1508                 if (x86_pmu.cpu_dying)
1509                         x86_pmu.cpu_dying(cpu);
1510                 break;
1511
1512         case CPU_UP_CANCELED:
1513         case CPU_DEAD:
1514                 if (x86_pmu.cpu_dead)
1515                         x86_pmu.cpu_dead(cpu);
1516                 break;
1517
1518         default:
1519                 break;
1520         }
1521
1522         return ret;
1523 }
1524
1525 static void __init pmu_check_apic(void)
1526 {
1527         if (boot_cpu_has(X86_FEATURE_APIC))
1528                 return;
1529
1530         x86_pmu.apic = 0;
1531         pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1532         pr_info("no hardware sampling interrupt available.\n");
1533
1534         /*
1535          * If we have a PMU initialized but no APIC
1536          * interrupts, we cannot sample hardware
1537          * events (user-space has to fall back and
1538          * sample via a hrtimer based software event):
1539          */
1540         pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1541
1542 }
1543
1544 static struct attribute_group x86_pmu_format_group = {
1545         .name = "format",
1546         .attrs = NULL,
1547 };
1548
1549 /*
1550  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1551  * out of events_attr attributes.
1552  */
1553 static void __init filter_events(struct attribute **attrs)
1554 {
1555         struct device_attribute *d;
1556         struct perf_pmu_events_attr *pmu_attr;
1557         int offset = 0;
1558         int i, j;
1559
1560         for (i = 0; attrs[i]; i++) {
1561                 d = (struct device_attribute *)attrs[i];
1562                 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1563                 /* str trumps id */
1564                 if (pmu_attr->event_str)
1565                         continue;
1566                 if (x86_pmu.event_map(i + offset))
1567                         continue;
1568
1569                 for (j = i; attrs[j]; j++)
1570                         attrs[j] = attrs[j + 1];
1571
1572                 /* Check the shifted attr. */
1573                 i--;
1574
1575                 /*
1576                  * event_map() is index based, the attrs array is organized
1577                  * by increasing event index. If we shift the events, then
1578                  * we need to compensate for the event_map(), otherwise
1579                  * we are looking up the wrong event in the map
1580                  */
1581                 offset++;
1582         }
1583 }
1584
1585 /* Merge two pointer arrays */
1586 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1587 {
1588         struct attribute **new;
1589         int j, i;
1590
1591         for (j = 0; a[j]; j++)
1592                 ;
1593         for (i = 0; b[i]; i++)
1594                 j++;
1595         j++;
1596
1597         new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1598         if (!new)
1599                 return NULL;
1600
1601         j = 0;
1602         for (i = 0; a[i]; i++)
1603                 new[j++] = a[i];
1604         for (i = 0; b[i]; i++)
1605                 new[j++] = b[i];
1606         new[j] = NULL;
1607
1608         return new;
1609 }
1610
1611 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1612 {
1613         struct perf_pmu_events_attr *pmu_attr = \
1614                 container_of(attr, struct perf_pmu_events_attr, attr);
1615         u64 config = x86_pmu.event_map(pmu_attr->id);
1616
1617         /* string trumps id */
1618         if (pmu_attr->event_str)
1619                 return sprintf(page, "%s", pmu_attr->event_str);
1620
1621         return x86_pmu.events_sysfs_show(page, config);
1622 }
1623 EXPORT_SYMBOL_GPL(events_sysfs_show);
1624
1625 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1626                           char *page)
1627 {
1628         struct perf_pmu_events_ht_attr *pmu_attr =
1629                 container_of(attr, struct perf_pmu_events_ht_attr, attr);
1630
1631         /*
1632          * Report conditional events depending on Hyper-Threading.
1633          *
1634          * This is overly conservative as usually the HT special
1635          * handling is not needed if the other CPU thread is idle.
1636          *
1637          * Note this does not (and cannot) handle the case when thread
1638          * siblings are invisible, for example with virtualization
1639          * if they are owned by some other guest.  The user tool
1640          * has to re-read when a thread sibling gets onlined later.
1641          */
1642         return sprintf(page, "%s",
1643                         topology_max_smt_threads() > 1 ?
1644                         pmu_attr->event_str_ht :
1645                         pmu_attr->event_str_noht);
1646 }
1647
1648 EVENT_ATTR(cpu-cycles,                  CPU_CYCLES              );
1649 EVENT_ATTR(instructions,                INSTRUCTIONS            );
1650 EVENT_ATTR(cache-references,            CACHE_REFERENCES        );
1651 EVENT_ATTR(cache-misses,                CACHE_MISSES            );
1652 EVENT_ATTR(branch-instructions,         BRANCH_INSTRUCTIONS     );
1653 EVENT_ATTR(branch-misses,               BRANCH_MISSES           );
1654 EVENT_ATTR(bus-cycles,                  BUS_CYCLES              );
1655 EVENT_ATTR(stalled-cycles-frontend,     STALLED_CYCLES_FRONTEND );
1656 EVENT_ATTR(stalled-cycles-backend,      STALLED_CYCLES_BACKEND  );
1657 EVENT_ATTR(ref-cycles,                  REF_CPU_CYCLES          );
1658
1659 static struct attribute *empty_attrs;
1660
1661 static struct attribute *events_attr[] = {
1662         EVENT_PTR(CPU_CYCLES),
1663         EVENT_PTR(INSTRUCTIONS),
1664         EVENT_PTR(CACHE_REFERENCES),
1665         EVENT_PTR(CACHE_MISSES),
1666         EVENT_PTR(BRANCH_INSTRUCTIONS),
1667         EVENT_PTR(BRANCH_MISSES),
1668         EVENT_PTR(BUS_CYCLES),
1669         EVENT_PTR(STALLED_CYCLES_FRONTEND),
1670         EVENT_PTR(STALLED_CYCLES_BACKEND),
1671         EVENT_PTR(REF_CPU_CYCLES),
1672         NULL,
1673 };
1674
1675 static struct attribute_group x86_pmu_events_group = {
1676         .name = "events",
1677         .attrs = events_attr,
1678 };
1679
1680 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1681 {
1682         u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1683         u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1684         bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1685         bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1686         bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1687         bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1688         ssize_t ret;
1689
1690         /*
1691         * We have whole page size to spend and just little data
1692         * to write, so we can safely use sprintf.
1693         */
1694         ret = sprintf(page, "event=0x%02llx", event);
1695
1696         if (umask)
1697                 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1698
1699         if (edge)
1700                 ret += sprintf(page + ret, ",edge");
1701
1702         if (pc)
1703                 ret += sprintf(page + ret, ",pc");
1704
1705         if (any)
1706                 ret += sprintf(page + ret, ",any");
1707
1708         if (inv)
1709                 ret += sprintf(page + ret, ",inv");
1710
1711         if (cmask)
1712                 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1713
1714         ret += sprintf(page + ret, "\n");
1715
1716         return ret;
1717 }
1718
1719 static int __init init_hw_perf_events(void)
1720 {
1721         struct x86_pmu_quirk *quirk;
1722         int err;
1723
1724         pr_info("Performance Events: ");
1725
1726         switch (boot_cpu_data.x86_vendor) {
1727         case X86_VENDOR_INTEL:
1728                 err = intel_pmu_init();
1729                 break;
1730         case X86_VENDOR_AMD:
1731                 err = amd_pmu_init();
1732                 break;
1733         default:
1734                 err = -ENOTSUPP;
1735         }
1736         if (err != 0) {
1737                 pr_cont("no PMU driver, software events only.\n");
1738                 return 0;
1739         }
1740
1741         pmu_check_apic();
1742
1743         /* sanity check that the hardware exists or is emulated */
1744         if (!check_hw_exists())
1745                 return 0;
1746
1747         pr_cont("%s PMU driver.\n", x86_pmu.name);
1748
1749         x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1750
1751         for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1752                 quirk->func();
1753
1754         if (!x86_pmu.intel_ctrl)
1755                 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1756
1757         perf_events_lapic_init();
1758         register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1759
1760         unconstrained = (struct event_constraint)
1761                 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1762                                    0, x86_pmu.num_counters, 0, 0);
1763
1764         x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1765
1766         if (x86_pmu.event_attrs)
1767                 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1768
1769         if (!x86_pmu.events_sysfs_show)
1770                 x86_pmu_events_group.attrs = &empty_attrs;
1771         else
1772                 filter_events(x86_pmu_events_group.attrs);
1773
1774         if (x86_pmu.cpu_events) {
1775                 struct attribute **tmp;
1776
1777                 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1778                 if (!WARN_ON(!tmp))
1779                         x86_pmu_events_group.attrs = tmp;
1780         }
1781
1782         pr_info("... version:                %d\n",     x86_pmu.version);
1783         pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1784         pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1785         pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1786         pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1787         pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1788         pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1789
1790         perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1791         perf_cpu_notifier(x86_pmu_notifier);
1792
1793         return 0;
1794 }
1795 early_initcall(init_hw_perf_events);
1796
1797 static inline void x86_pmu_read(struct perf_event *event)
1798 {
1799         x86_perf_event_update(event);
1800 }
1801
1802 /*
1803  * Start group events scheduling transaction
1804  * Set the flag to make pmu::enable() not perform the
1805  * schedulability test, it will be performed at commit time
1806  *
1807  * We only support PERF_PMU_TXN_ADD transactions. Save the
1808  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1809  * transactions.
1810  */
1811 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1812 {
1813         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1814
1815         WARN_ON_ONCE(cpuc->txn_flags);          /* txn already in flight */
1816
1817         cpuc->txn_flags = txn_flags;
1818         if (txn_flags & ~PERF_PMU_TXN_ADD)
1819                 return;
1820
1821         perf_pmu_disable(pmu);
1822         __this_cpu_write(cpu_hw_events.n_txn, 0);
1823 }
1824
1825 /*
1826  * Stop group events scheduling transaction
1827  * Clear the flag and pmu::enable() will perform the
1828  * schedulability test.
1829  */
1830 static void x86_pmu_cancel_txn(struct pmu *pmu)
1831 {
1832         unsigned int txn_flags;
1833         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1834
1835         WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1836
1837         txn_flags = cpuc->txn_flags;
1838         cpuc->txn_flags = 0;
1839         if (txn_flags & ~PERF_PMU_TXN_ADD)
1840                 return;
1841
1842         /*
1843          * Truncate collected array by the number of events added in this
1844          * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1845          */
1846         __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1847         __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1848         perf_pmu_enable(pmu);
1849 }
1850
1851 /*
1852  * Commit group events scheduling transaction
1853  * Perform the group schedulability test as a whole
1854  * Return 0 if success
1855  *
1856  * Does not cancel the transaction on failure; expects the caller to do this.
1857  */
1858 static int x86_pmu_commit_txn(struct pmu *pmu)
1859 {
1860         struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1861         int assign[X86_PMC_IDX_MAX];
1862         int n, ret;
1863
1864         WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
1865
1866         if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1867                 cpuc->txn_flags = 0;
1868                 return 0;
1869         }
1870
1871         n = cpuc->n_events;
1872
1873         if (!x86_pmu_initialized())
1874                 return -EAGAIN;
1875
1876         ret = x86_pmu.schedule_events(cpuc, n, assign);
1877         if (ret)
1878                 return ret;
1879
1880         /*
1881          * copy new assignment, now we know it is possible
1882          * will be used by hw_perf_enable()
1883          */
1884         memcpy(cpuc->assign, assign, n*sizeof(int));
1885
1886         cpuc->txn_flags = 0;
1887         perf_pmu_enable(pmu);
1888         return 0;
1889 }
1890 /*
1891  * a fake_cpuc is used to validate event groups. Due to
1892  * the extra reg logic, we need to also allocate a fake
1893  * per_core and per_cpu structure. Otherwise, group events
1894  * using extra reg may conflict without the kernel being
1895  * able to catch this when the last event gets added to
1896  * the group.
1897  */
1898 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1899 {
1900         kfree(cpuc->shared_regs);
1901         kfree(cpuc);
1902 }
1903
1904 static struct cpu_hw_events *allocate_fake_cpuc(void)
1905 {
1906         struct cpu_hw_events *cpuc;
1907         int cpu = raw_smp_processor_id();
1908
1909         cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1910         if (!cpuc)
1911                 return ERR_PTR(-ENOMEM);
1912
1913         /* only needed, if we have extra_regs */
1914         if (x86_pmu.extra_regs) {
1915                 cpuc->shared_regs = allocate_shared_regs(cpu);
1916                 if (!cpuc->shared_regs)
1917                         goto error;
1918         }
1919         cpuc->is_fake = 1;
1920         return cpuc;
1921 error:
1922         free_fake_cpuc(cpuc);
1923         return ERR_PTR(-ENOMEM);
1924 }
1925
1926 /*
1927  * validate that we can schedule this event
1928  */
1929 static int validate_event(struct perf_event *event)
1930 {
1931         struct cpu_hw_events *fake_cpuc;
1932         struct event_constraint *c;
1933         int ret = 0;
1934
1935         fake_cpuc = allocate_fake_cpuc();
1936         if (IS_ERR(fake_cpuc))
1937                 return PTR_ERR(fake_cpuc);
1938
1939         c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
1940
1941         if (!c || !c->weight)
1942                 ret = -EINVAL;
1943
1944         if (x86_pmu.put_event_constraints)
1945                 x86_pmu.put_event_constraints(fake_cpuc, event);
1946
1947         free_fake_cpuc(fake_cpuc);
1948
1949         return ret;
1950 }
1951
1952 /*
1953  * validate a single event group
1954  *
1955  * validation include:
1956  *      - check events are compatible which each other
1957  *      - events do not compete for the same counter
1958  *      - number of events <= number of counters
1959  *
1960  * validation ensures the group can be loaded onto the
1961  * PMU if it was the only group available.
1962  */
1963 static int validate_group(struct perf_event *event)
1964 {
1965         struct perf_event *leader = event->group_leader;
1966         struct cpu_hw_events *fake_cpuc;
1967         int ret = -EINVAL, n;
1968
1969         fake_cpuc = allocate_fake_cpuc();
1970         if (IS_ERR(fake_cpuc))
1971                 return PTR_ERR(fake_cpuc);
1972         /*
1973          * the event is not yet connected with its
1974          * siblings therefore we must first collect
1975          * existing siblings, then add the new event
1976          * before we can simulate the scheduling
1977          */
1978         n = collect_events(fake_cpuc, leader, true);
1979         if (n < 0)
1980                 goto out;
1981
1982         fake_cpuc->n_events = n;
1983         n = collect_events(fake_cpuc, event, false);
1984         if (n < 0)
1985                 goto out;
1986
1987         fake_cpuc->n_events = n;
1988
1989         ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1990
1991 out:
1992         free_fake_cpuc(fake_cpuc);
1993         return ret;
1994 }
1995
1996 static int x86_pmu_event_init(struct perf_event *event)
1997 {
1998         struct pmu *tmp;
1999         int err;
2000
2001         switch (event->attr.type) {
2002         case PERF_TYPE_RAW:
2003         case PERF_TYPE_HARDWARE:
2004         case PERF_TYPE_HW_CACHE:
2005                 break;
2006
2007         default:
2008                 return -ENOENT;
2009         }
2010
2011         err = __x86_pmu_event_init(event);
2012         if (!err) {
2013                 /*
2014                  * we temporarily connect event to its pmu
2015                  * such that validate_group() can classify
2016                  * it as an x86 event using is_x86_event()
2017                  */
2018                 tmp = event->pmu;
2019                 event->pmu = &pmu;
2020
2021                 if (event->group_leader != event)
2022                         err = validate_group(event);
2023                 else
2024                         err = validate_event(event);
2025
2026                 event->pmu = tmp;
2027         }
2028         if (err) {
2029                 if (event->destroy)
2030                         event->destroy(event);
2031         }
2032
2033         if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
2034                 event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2035
2036         return err;
2037 }
2038
2039 static void refresh_pce(void *ignored)
2040 {
2041         if (current->mm)
2042                 load_mm_cr4(current->mm);
2043 }
2044
2045 static void x86_pmu_event_mapped(struct perf_event *event)
2046 {
2047         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2048                 return;
2049
2050         if (atomic_inc_return(&current->mm->context.perf_rdpmc_allowed) == 1)
2051                 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2052 }
2053
2054 static void x86_pmu_event_unmapped(struct perf_event *event)
2055 {
2056         if (!current->mm)
2057                 return;
2058
2059         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2060                 return;
2061
2062         if (atomic_dec_and_test(&current->mm->context.perf_rdpmc_allowed))
2063                 on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2064 }
2065
2066 static int x86_pmu_event_idx(struct perf_event *event)
2067 {
2068         int idx = event->hw.idx;
2069
2070         if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2071                 return 0;
2072
2073         if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2074                 idx -= INTEL_PMC_IDX_FIXED;
2075                 idx |= 1 << 30;
2076         }
2077
2078         return idx + 1;
2079 }
2080
2081 static ssize_t get_attr_rdpmc(struct device *cdev,
2082                               struct device_attribute *attr,
2083                               char *buf)
2084 {
2085         return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2086 }
2087
2088 static ssize_t set_attr_rdpmc(struct device *cdev,
2089                               struct device_attribute *attr,
2090                               const char *buf, size_t count)
2091 {
2092         unsigned long val;
2093         ssize_t ret;
2094
2095         ret = kstrtoul(buf, 0, &val);
2096         if (ret)
2097                 return ret;
2098
2099         if (val > 2)
2100                 return -EINVAL;
2101
2102         if (x86_pmu.attr_rdpmc_broken)
2103                 return -ENOTSUPP;
2104
2105         if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2106                 /*
2107                  * Changing into or out of always available, aka
2108                  * perf-event-bypassing mode.  This path is extremely slow,
2109                  * but only root can trigger it, so it's okay.
2110                  */
2111                 if (val == 2)
2112                         static_key_slow_inc(&rdpmc_always_available);
2113                 else
2114                         static_key_slow_dec(&rdpmc_always_available);
2115                 on_each_cpu(refresh_pce, NULL, 1);
2116         }
2117
2118         x86_pmu.attr_rdpmc = val;
2119
2120         return count;
2121 }
2122
2123 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2124
2125 static struct attribute *x86_pmu_attrs[] = {
2126         &dev_attr_rdpmc.attr,
2127         NULL,
2128 };
2129
2130 static struct attribute_group x86_pmu_attr_group = {
2131         .attrs = x86_pmu_attrs,
2132 };
2133
2134 static const struct attribute_group *x86_pmu_attr_groups[] = {
2135         &x86_pmu_attr_group,
2136         &x86_pmu_format_group,
2137         &x86_pmu_events_group,
2138         NULL,
2139 };
2140
2141 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2142 {
2143         if (x86_pmu.sched_task)
2144                 x86_pmu.sched_task(ctx, sched_in);
2145 }
2146
2147 void perf_check_microcode(void)
2148 {
2149         if (x86_pmu.check_microcode)
2150                 x86_pmu.check_microcode();
2151 }
2152 EXPORT_SYMBOL_GPL(perf_check_microcode);
2153
2154 static struct pmu pmu = {
2155         .pmu_enable             = x86_pmu_enable,
2156         .pmu_disable            = x86_pmu_disable,
2157
2158         .attr_groups            = x86_pmu_attr_groups,
2159
2160         .event_init             = x86_pmu_event_init,
2161
2162         .event_mapped           = x86_pmu_event_mapped,
2163         .event_unmapped         = x86_pmu_event_unmapped,
2164
2165         .add                    = x86_pmu_add,
2166         .del                    = x86_pmu_del,
2167         .start                  = x86_pmu_start,
2168         .stop                   = x86_pmu_stop,
2169         .read                   = x86_pmu_read,
2170
2171         .start_txn              = x86_pmu_start_txn,
2172         .cancel_txn             = x86_pmu_cancel_txn,
2173         .commit_txn             = x86_pmu_commit_txn,
2174
2175         .event_idx              = x86_pmu_event_idx,
2176         .sched_task             = x86_pmu_sched_task,
2177         .task_ctx_size          = sizeof(struct x86_perf_task_context),
2178 };
2179
2180 void arch_perf_update_userpage(struct perf_event *event,
2181                                struct perf_event_mmap_page *userpg, u64 now)
2182 {
2183         struct cyc2ns_data *data;
2184
2185         userpg->cap_user_time = 0;
2186         userpg->cap_user_time_zero = 0;
2187         userpg->cap_user_rdpmc =
2188                 !!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2189         userpg->pmc_width = x86_pmu.cntval_bits;
2190
2191         if (!sched_clock_stable())
2192                 return;
2193
2194         data = cyc2ns_read_begin();
2195
2196         /*
2197          * Internal timekeeping for enabled/running/stopped times
2198          * is always in the local_clock domain.
2199          */
2200         userpg->cap_user_time = 1;
2201         userpg->time_mult = data->cyc2ns_mul;
2202         userpg->time_shift = data->cyc2ns_shift;
2203         userpg->time_offset = data->cyc2ns_offset - now;
2204
2205         /*
2206          * cap_user_time_zero doesn't make sense when we're using a different
2207          * time base for the records.
2208          */
2209         if (!event->attr.use_clockid) {
2210                 userpg->cap_user_time_zero = 1;
2211                 userpg->time_zero = data->cyc2ns_offset;
2212         }
2213
2214         cyc2ns_read_end(data);
2215 }
2216
2217 /*
2218  * callchain support
2219  */
2220
2221 static int backtrace_stack(void *data, char *name)
2222 {
2223         return 0;
2224 }
2225
2226 static int backtrace_address(void *data, unsigned long addr, int reliable)
2227 {
2228         struct perf_callchain_entry_ctx *entry = data;
2229
2230         return perf_callchain_store(entry, addr);
2231 }
2232
2233 static const struct stacktrace_ops backtrace_ops = {
2234         .stack                  = backtrace_stack,
2235         .address                = backtrace_address,
2236         .walk_stack             = print_context_stack_bp,
2237 };
2238
2239 void
2240 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2241 {
2242         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2243                 /* TODO: We don't support guest os callchain now */
2244                 return;
2245         }
2246
2247         perf_callchain_store(entry, regs->ip);
2248
2249         dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2250 }
2251
2252 static inline int
2253 valid_user_frame(const void __user *fp, unsigned long size)
2254 {
2255         return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2256 }
2257
2258 static unsigned long get_segment_base(unsigned int segment)
2259 {
2260         struct desc_struct *desc;
2261         int idx = segment >> 3;
2262
2263         if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2264 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2265                 struct ldt_struct *ldt;
2266
2267                 if (idx > LDT_ENTRIES)
2268                         return 0;
2269
2270                 /* IRQs are off, so this synchronizes with smp_store_release */
2271                 ldt = lockless_dereference(current->active_mm->context.ldt);
2272                 if (!ldt || idx > ldt->size)
2273                         return 0;
2274
2275                 desc = &ldt->entries[idx];
2276 #else
2277                 return 0;
2278 #endif
2279         } else {
2280                 if (idx > GDT_ENTRIES)
2281                         return 0;
2282
2283                 desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2284         }
2285
2286         return get_desc_base(desc);
2287 }
2288
2289 #ifdef CONFIG_IA32_EMULATION
2290
2291 #include <asm/compat.h>
2292
2293 static inline int
2294 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2295 {
2296         /* 32-bit process in 64-bit kernel. */
2297         unsigned long ss_base, cs_base;
2298         struct stack_frame_ia32 frame;
2299         const void __user *fp;
2300
2301         if (!test_thread_flag(TIF_IA32))
2302                 return 0;
2303
2304         cs_base = get_segment_base(regs->cs);
2305         ss_base = get_segment_base(regs->ss);
2306
2307         fp = compat_ptr(ss_base + regs->bp);
2308         pagefault_disable();
2309         while (entry->nr < entry->max_stack) {
2310                 unsigned long bytes;
2311                 frame.next_frame     = 0;
2312                 frame.return_address = 0;
2313
2314                 if (!access_ok(VERIFY_READ, fp, 8))
2315                         break;
2316
2317                 bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2318                 if (bytes != 0)
2319                         break;
2320                 bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2321                 if (bytes != 0)
2322                         break;
2323
2324                 if (!valid_user_frame(fp, sizeof(frame)))
2325                         break;
2326
2327                 perf_callchain_store(entry, cs_base + frame.return_address);
2328                 fp = compat_ptr(ss_base + frame.next_frame);
2329         }
2330         pagefault_enable();
2331         return 1;
2332 }
2333 #else
2334 static inline int
2335 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2336 {
2337     return 0;
2338 }
2339 #endif
2340
2341 void
2342 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2343 {
2344         struct stack_frame frame;
2345         const unsigned long __user *fp;
2346
2347         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2348                 /* TODO: We don't support guest os callchain now */
2349                 return;
2350         }
2351
2352         /*
2353          * We don't know what to do with VM86 stacks.. ignore them for now.
2354          */
2355         if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2356                 return;
2357
2358         fp = (unsigned long __user *)regs->bp;
2359
2360         perf_callchain_store(entry, regs->ip);
2361
2362         if (!current->mm)
2363                 return;
2364
2365         if (perf_callchain_user32(regs, entry))
2366                 return;
2367
2368         pagefault_disable();
2369         while (entry->nr < entry->max_stack) {
2370                 unsigned long bytes;
2371
2372                 frame.next_frame             = NULL;
2373                 frame.return_address = 0;
2374
2375                 if (!access_ok(VERIFY_READ, fp, sizeof(*fp) * 2))
2376                         break;
2377
2378                 bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2379                 if (bytes != 0)
2380                         break;
2381                 bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2382                 if (bytes != 0)
2383                         break;
2384
2385                 if (!valid_user_frame(fp, sizeof(frame)))
2386                         break;
2387
2388                 perf_callchain_store(entry, frame.return_address);
2389                 fp = (void __user *)frame.next_frame;
2390         }
2391         pagefault_enable();
2392 }
2393
2394 /*
2395  * Deal with code segment offsets for the various execution modes:
2396  *
2397  *   VM86 - the good olde 16 bit days, where the linear address is
2398  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2399  *
2400  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2401  *          to figure out what the 32bit base address is.
2402  *
2403  *    X32 - has TIF_X32 set, but is running in x86_64
2404  *
2405  * X86_64 - CS,DS,SS,ES are all zero based.
2406  */
2407 static unsigned long code_segment_base(struct pt_regs *regs)
2408 {
2409         /*
2410          * For IA32 we look at the GDT/LDT segment base to convert the
2411          * effective IP to a linear address.
2412          */
2413
2414 #ifdef CONFIG_X86_32
2415         /*
2416          * If we are in VM86 mode, add the segment offset to convert to a
2417          * linear address.
2418          */
2419         if (regs->flags & X86_VM_MASK)
2420                 return 0x10 * regs->cs;
2421
2422         if (user_mode(regs) && regs->cs != __USER_CS)
2423                 return get_segment_base(regs->cs);
2424 #else
2425         if (user_mode(regs) && !user_64bit_mode(regs) &&
2426             regs->cs != __USER32_CS)
2427                 return get_segment_base(regs->cs);
2428 #endif
2429         return 0;
2430 }
2431
2432 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2433 {
2434         if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2435                 return perf_guest_cbs->get_guest_ip();
2436
2437         return regs->ip + code_segment_base(regs);
2438 }
2439
2440 unsigned long perf_misc_flags(struct pt_regs *regs)
2441 {
2442         int misc = 0;
2443
2444         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2445                 if (perf_guest_cbs->is_user_mode())
2446                         misc |= PERF_RECORD_MISC_GUEST_USER;
2447                 else
2448                         misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2449         } else {
2450                 if (user_mode(regs))
2451                         misc |= PERF_RECORD_MISC_USER;
2452                 else
2453                         misc |= PERF_RECORD_MISC_KERNEL;
2454         }
2455
2456         if (regs->flags & PERF_EFLAGS_EXACT)
2457                 misc |= PERF_RECORD_MISC_EXACT_IP;
2458
2459         return misc;
2460 }
2461
2462 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2463 {
2464         cap->version            = x86_pmu.version;
2465         cap->num_counters_gp    = x86_pmu.num_counters;
2466         cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2467         cap->bit_width_gp       = x86_pmu.cntval_bits;
2468         cap->bit_width_fixed    = x86_pmu.cntval_bits;
2469         cap->events_mask        = (unsigned int)x86_pmu.events_maskl;
2470         cap->events_mask_len    = x86_pmu.events_mask_len;
2471 }
2472 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);