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/highmem.h>
  26 #include <linux/cpu.h>
  27 #include <linux/bitops.h>
  28
  29 #include <asm/apic.h>
  30 #include <asm/stacktrace.h>
  31 #include <asm/nmi.h>
  32 #include <asm/compat.h>
  33
  34 #if 0
  35 #undef wrmsrl
  36 #define wrmsrl(msr, val)                                        \
  37 do {                                                            \
  38         trace_printk("wrmsrl(%lx, %lx)\n", (unsigned long)(msr),\
  39                         (unsigned long)(val));                  \
  40         native_write_msr((msr), (u32)((u64)(val)),              \
  41                         (u32)((u64)(val) >> 32));               \
  42 } while (0)
  43 #endif
  44
  45 /*
  46  * best effort, GUP based copy_from_user() that assumes IRQ or NMI context
  47  */
  48 static unsigned long
  49 copy_from_user_nmi(void *to, const void __user *from, unsigned long n)
  50 {
  51         unsigned long offset, addr = (unsigned long)from;
  52         unsigned long size, len = 0;
  53         struct page *page;
  54         void *map;
  55         int ret;
  56
  57         do {
  58                 ret = __get_user_pages_fast(addr, 1, 0, &page);
  59                 if (!ret)
  60                         break;
  61
  62                 offset = addr & (PAGE_SIZE - 1);
  63                 size = min(PAGE_SIZE - offset, n - len);
  64
  65                 map = kmap_atomic(page);
  66                 memcpy(to, map+offset, size);
  67                 kunmap_atomic(map);
  68                 put_page(page);
  69
  70                 len  += size;
  71                 to   += size;
  72                 addr += size;
  73
  74         } while (len < n);
  75
  76         return len;
  77 }
  78
  79 struct event_constraint {
  80         union {
  81                 unsigned long   idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
  82                 u64             idxmsk64;
  83         };
  84         u64     code;
  85         u64     cmask;
  86         int     weight;
  87 };
  88
  89 struct amd_nb {
  90         int nb_id;  /* NorthBridge id */
  91         int refcnt; /* reference count */
  92         struct perf_event *owners[X86_PMC_IDX_MAX];
  93         struct event_constraint event_constraints[X86_PMC_IDX_MAX];
  94 };
  95
  96 #define MAX_LBR_ENTRIES         16
  97
  98 struct cpu_hw_events {
  99         /*
 100          * Generic x86 PMC bits
 101          */
 102         struct perf_event       *events[X86_PMC_IDX_MAX]; /* in counter order */
 103         unsigned long           active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 104         unsigned long           running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 105         int                     enabled;
 106
 107         int                     n_events;
 108         int                     n_added;
 109         int                     n_txn;
 110         int                     assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 111         u64                     tags[X86_PMC_IDX_MAX];
 112         struct perf_event       *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
 113
 114         unsigned int            group_flag;
 115
 116         /*
 117          * Intel DebugStore bits
 118          */
 119         struct debug_store      *ds;
 120         u64                     pebs_enabled;
 121
 122         /*
 123          * Intel LBR bits
 124          */
 125         int                             lbr_users;
 126         void                            *lbr_context;
 127         struct perf_branch_stack        lbr_stack;
 128         struct perf_branch_entry        lbr_entries[MAX_LBR_ENTRIES];
 129
 130         /*
 131          * AMD specific bits
 132          */
 133         struct amd_nb           *amd_nb;
 134 };
 135
 136 #define __EVENT_CONSTRAINT(c, n, m, w) {\
 137         { .idxmsk64 = (n) },            \
 138         .code = (c),                    \
 139         .cmask = (m),                   \
 140         .weight = (w),                  \
 141 }
 142
 143 #define EVENT_CONSTRAINT(c, n, m)       \
 144         __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n))
 145
 146 /*
 147  * Constraint on the Event code.
 148  */
 149 #define INTEL_EVENT_CONSTRAINT(c, n)    \
 150         EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
 151
 152 /*
 153  * Constraint on the Event code + UMask + fixed-mask
 154  *
 155  * filter mask to validate fixed counter events.
 156  * the following filters disqualify for fixed counters:
 157  *  - inv
 158  *  - edge
 159  *  - cnt-mask
 160  *  The other filters are supported by fixed counters.
 161  *  The any-thread option is supported starting with v3.
 162  */
 163 #define FIXED_EVENT_CONSTRAINT(c, n)    \
 164         EVENT_CONSTRAINT(c, (1ULL << (32+n)), X86_RAW_EVENT_MASK)
 165
 166 /*
 167  * Constraint on the Event code + UMask
 168  */
 169 #define PEBS_EVENT_CONSTRAINT(c, n)     \
 170         EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
 171
 172 #define EVENT_CONSTRAINT_END            \
 173         EVENT_CONSTRAINT(0, 0, 0)
 174
 175 #define for_each_event_constraint(e, c) \
 176         for ((e) = (c); (e)->weight; (e)++)
 177
 178 union perf_capabilities {
 179         struct {
 180                 u64     lbr_format    : 6;
 181                 u64     pebs_trap     : 1;
 182                 u64     pebs_arch_reg : 1;
 183                 u64     pebs_format   : 4;
 184                 u64     smm_freeze    : 1;
 185         };
 186         u64     capabilities;
 187 };
 188
 189 /*
 190  * struct x86_pmu - generic x86 pmu
 191  */
 192 struct x86_pmu {
 193         /*
 194          * Generic x86 PMC bits
 195          */
 196         const char      *name;
 197         int             version;
 198         int             (*handle_irq)(struct pt_regs *);
 199         void            (*disable_all)(void);
 200         void            (*enable_all)(int added);
 201         void            (*enable)(struct perf_event *);
 202         void            (*disable)(struct perf_event *);
 203         int             (*hw_config)(struct perf_event *event);
 204         int             (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
 205         unsigned        eventsel;
 206         unsigned        perfctr;
 207         u64             (*event_map)(int);
 208         int             max_events;
 209         int             num_counters;
 210         int             num_counters_fixed;
 211         int             cntval_bits;
 212         u64             cntval_mask;
 213         int             apic;
 214         u64             max_period;
 215         struct event_constraint *
 216                         (*get_event_constraints)(struct cpu_hw_events *cpuc,
 217                                                  struct perf_event *event);
 218
 219         void            (*put_event_constraints)(struct cpu_hw_events *cpuc,
 220                                                  struct perf_event *event);
 221         struct event_constraint *event_constraints;
 222         void            (*quirks)(void);
 223         int             perfctr_second_write;
 224
 225         int             (*cpu_prepare)(int cpu);
 226         void            (*cpu_starting)(int cpu);
 227         void            (*cpu_dying)(int cpu);
 228         void            (*cpu_dead)(int cpu);
 229
 230         /*
 231          * Intel Arch Perfmon v2+
 232          */
 233         u64                     intel_ctrl;
 234         union perf_capabilities intel_cap;
 235
 236         /*
 237          * Intel DebugStore bits
 238          */
 239         int             bts, pebs;
 240         int             pebs_record_size;
 241         void            (*drain_pebs)(struct pt_regs *regs);
 242         struct event_constraint *pebs_constraints;
 243
 244         /*
 245          * Intel LBR
 246          */
 247         unsigned long   lbr_tos, lbr_from, lbr_to; /* MSR base regs       */
 248         int             lbr_nr;                    /* hardware stack size */
 249 };
 250
 251 static struct x86_pmu x86_pmu __read_mostly;
 252
 253 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
 254         .enabled = 1,
 255 };
 256
 257 static int x86_perf_event_set_period(struct perf_event *event);
 258
 259 /*
 260  * Generalized hw caching related hw_event table, filled
 261  * in on a per model basis. A value of 0 means
 262  * 'not supported', -1 means 'hw_event makes no sense on
 263  * this CPU', any other value means the raw hw_event
 264  * ID.
 265  */
 266
 267 #define C(x) PERF_COUNT_HW_CACHE_##x
 268
 269 static u64 __read_mostly hw_cache_event_ids
 270                                 [PERF_COUNT_HW_CACHE_MAX]
 271                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 272                                 [PERF_COUNT_HW_CACHE_RESULT_MAX];
 273
 274 /*
 275  * Propagate event elapsed time into the generic event.
 276  * Can only be executed on the CPU where the event is active.
 277  * Returns the delta events processed.
 278  */
 279 static u64
 280 x86_perf_event_update(struct perf_event *event)
 281 {
 282         struct hw_perf_event *hwc = &event->hw;
 283         int shift = 64 - x86_pmu.cntval_bits;
 284         u64 prev_raw_count, new_raw_count;
 285         int idx = hwc->idx;
 286         s64 delta;
 287
 288         if (idx == X86_PMC_IDX_FIXED_BTS)
 289                 return 0;
 290
 291         /*
 292          * Careful: an NMI might modify the previous event value.
 293          *
 294          * Our tactic to handle this is to first atomically read and
 295          * exchange a new raw count - then add that new-prev delta
 296          * count to the generic event atomically:
 297          */
 298 again:
 299         prev_raw_count = local64_read(&hwc->prev_count);
 300         rdmsrl(hwc->event_base + idx, new_raw_count);
 301
 302         if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 303                                         new_raw_count) != prev_raw_count)
 304                 goto again;
 305
 306         /*
 307          * Now we have the new raw value and have updated the prev
 308          * timestamp already. We can now calculate the elapsed delta
 309          * (event-)time and add that to the generic event.
 310          *
 311          * Careful, not all hw sign-extends above the physical width
 312          * of the count.
 313          */
 314         delta = (new_raw_count << shift) - (prev_raw_count << shift);
 315         delta >>= shift;
 316
 317         local64_add(delta, &event->count);
 318         local64_sub(delta, &hwc->period_left);
 319
 320         return new_raw_count;
 321 }
 322
 323 static atomic_t active_events;
 324 static DEFINE_MUTEX(pmc_reserve_mutex);
 325
 326 #ifdef CONFIG_X86_LOCAL_APIC
 327
 328 static bool reserve_pmc_hardware(void)
 329 {
 330         int i;
 331
 332         if (nmi_watchdog == NMI_LOCAL_APIC)
 333                 disable_lapic_nmi_watchdog();
 334
 335         for (i = 0; i < x86_pmu.num_counters; i++) {
 336                 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
 337                         goto perfctr_fail;
 338         }
 339
 340         for (i = 0; i < x86_pmu.num_counters; i++) {
 341                 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
 342                         goto eventsel_fail;
 343         }
 344
 345         return true;
 346
 347 eventsel_fail:
 348         for (i--; i >= 0; i--)
 349                 release_evntsel_nmi(x86_pmu.eventsel + i);
 350
 351         i = x86_pmu.num_counters;
 352
 353 perfctr_fail:
 354         for (i--; i >= 0; i--)
 355                 release_perfctr_nmi(x86_pmu.perfctr + i);
 356
 357         if (nmi_watchdog == NMI_LOCAL_APIC)
 358                 enable_lapic_nmi_watchdog();
 359
 360         return false;
 361 }
 362
 363 static void release_pmc_hardware(void)
 364 {
 365         int i;
 366
 367         for (i = 0; i < x86_pmu.num_counters; i++) {
 368                 release_perfctr_nmi(x86_pmu.perfctr + i);
 369                 release_evntsel_nmi(x86_pmu.eventsel + i);
 370         }
 371
 372         if (nmi_watchdog == NMI_LOCAL_APIC)
 373                 enable_lapic_nmi_watchdog();
 374 }
 375
 376 #else
 377
 378 static bool reserve_pmc_hardware(void) { return true; }
 379 static void release_pmc_hardware(void) {}
 380
 381 #endif
 382
 383 static int reserve_ds_buffers(void);
 384 static void release_ds_buffers(void);
 385
 386 static void hw_perf_event_destroy(struct perf_event *event)
 387 {
 388         if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
 389                 release_pmc_hardware();
 390                 release_ds_buffers();
 391                 mutex_unlock(&pmc_reserve_mutex);
 392         }
 393 }
 394
 395 static inline int x86_pmu_initialized(void)
 396 {
 397         return x86_pmu.handle_irq != NULL;
 398 }
 399
 400 static inline int
 401 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event_attr *attr)
 402 {
 403         unsigned int cache_type, cache_op, cache_result;
 404         u64 config, val;
 405
 406         config = attr->config;
 407
 408         cache_type = (config >>  0) & 0xff;
 409         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 410                 return -EINVAL;
 411
 412         cache_op = (config >>  8) & 0xff;
 413         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 414                 return -EINVAL;
 415
 416         cache_result = (config >> 16) & 0xff;
 417         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 418                 return -EINVAL;
 419
 420         val = hw_cache_event_ids[cache_type][cache_op][cache_result];
 421
 422         if (val == 0)
 423                 return -ENOENT;
 424
 425         if (val == -1)
 426                 return -EINVAL;
 427
 428         hwc->config |= val;
 429
 430         return 0;
 431 }
 432
 433 static int x86_setup_perfctr(struct perf_event *event)
 434 {
 435         struct perf_event_attr *attr = &event->attr;
 436         struct hw_perf_event *hwc = &event->hw;
 437         u64 config;
 438
 439         if (!hwc->sample_period) {
 440                 hwc->sample_period = x86_pmu.max_period;
 441                 hwc->last_period = hwc->sample_period;
 442                 local64_set(&hwc->period_left, hwc->sample_period);
 443         } else {
 444                 /*
 445                  * If we have a PMU initialized but no APIC
 446                  * interrupts, we cannot sample hardware
 447                  * events (user-space has to fall back and
 448                  * sample via a hrtimer based software event):
 449                  */
 450                 if (!x86_pmu.apic)
 451                         return -EOPNOTSUPP;
 452         }
 453
 454         if (attr->type == PERF_TYPE_RAW)
 455                 return 0;
 456
 457         if (attr->type == PERF_TYPE_HW_CACHE)
 458                 return set_ext_hw_attr(hwc, attr);
 459
 460         if (attr->config >= x86_pmu.max_events)
 461                 return -EINVAL;
 462
 463         /*
 464          * The generic map:
 465          */
 466         config = x86_pmu.event_map(attr->config);
 467
 468         if (config == 0)
 469                 return -ENOENT;
 470
 471         if (config == -1LL)
 472                 return -EINVAL;
 473
 474         /*
 475          * Branch tracing:
 476          */
 477         if ((attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
 478             (hwc->sample_period == 1)) {
 479                 /* BTS is not supported by this architecture. */
 480                 if (!x86_pmu.bts)
 481                         return -EOPNOTSUPP;
 482
 483                 /* BTS is currently only allowed for user-mode. */
 484                 if (!attr->exclude_kernel)
 485                         return -EOPNOTSUPP;
 486         }
 487
 488         hwc->config |= config;
 489
 490         return 0;
 491 }
 492
 493 static int x86_pmu_hw_config(struct perf_event *event)
 494 {
 495         if (event->attr.precise_ip) {
 496                 int precise = 0;
 497
 498                 /* Support for constant skid */
 499                 if (x86_pmu.pebs)
 500                         precise++;
 501
 502                 /* Support for IP fixup */
 503                 if (x86_pmu.lbr_nr)
 504                         precise++;
 505
 506                 if (event->attr.precise_ip > precise)
 507                         return -EOPNOTSUPP;
 508         }
 509
 510         /*
 511          * Generate PMC IRQs:
 512          * (keep 'enabled' bit clear for now)
 513          */
 514         event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
 515
 516         /*
 517          * Count user and OS events unless requested not to
 518          */
 519         if (!event->attr.exclude_user)
 520                 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
 521         if (!event->attr.exclude_kernel)
 522                 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
 523
 524         if (event->attr.type == PERF_TYPE_RAW)
 525                 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
 526
 527         return x86_setup_perfctr(event);
 528 }
 529
 530 /*
 531  * Setup the hardware configuration for a given attr_type
 532  */
 533 static int __x86_pmu_event_init(struct perf_event *event)
 534 {
 535         int err;
 536
 537         if (!x86_pmu_initialized())
 538                 return -ENODEV;
 539
 540         err = 0;
 541         if (!atomic_inc_not_zero(&active_events)) {
 542                 mutex_lock(&pmc_reserve_mutex);
 543                 if (atomic_read(&active_events) == 0) {
 544                         if (!reserve_pmc_hardware())
 545                                 err = -EBUSY;
 546                         else {
 547                                 err = reserve_ds_buffers();
 548                                 if (err)
 549                                         release_pmc_hardware();
 550                         }
 551                 }
 552                 if (!err)
 553                         atomic_inc(&active_events);
 554                 mutex_unlock(&pmc_reserve_mutex);
 555         }
 556         if (err)
 557                 return err;
 558
 559         event->destroy = hw_perf_event_destroy;
 560
 561         event->hw.idx = -1;
 562         event->hw.last_cpu = -1;
 563         event->hw.last_tag = ~0ULL;
 564
 565         return x86_pmu.hw_config(event);
 566 }
 567
 568 static void x86_pmu_disable_all(void)
 569 {
 570         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 571         int idx;
 572
 573         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 574                 u64 val;
 575
 576                 if (!test_bit(idx, cpuc->active_mask))
 577                         continue;
 578                 rdmsrl(x86_pmu.eventsel + idx, val);
 579                 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
 580                         continue;
 581                 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 582                 wrmsrl(x86_pmu.eventsel + idx, val);
 583         }
 584 }
 585
 586 static void x86_pmu_disable(struct pmu *pmu)
 587 {
 588         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 589
 590         if (!x86_pmu_initialized())
 591                 return;
 592
 593         if (!cpuc->enabled)
 594                 return;
 595
 596         cpuc->n_added = 0;
 597         cpuc->enabled = 0;
 598         barrier();
 599
 600         x86_pmu.disable_all();
 601 }
 602
 603 static void x86_pmu_enable_all(int added)
 604 {
 605         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 606         int idx;
 607
 608         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 609                 struct perf_event *event = cpuc->events[idx];
 610                 u64 val;
 611
 612                 if (!test_bit(idx, cpuc->active_mask))
 613                         continue;
 614
 615                 val = event->hw.config;
 616                 val |= ARCH_PERFMON_EVENTSEL_ENABLE;
 617                 wrmsrl(x86_pmu.eventsel + idx, val);
 618         }
 619 }
 620
 621 static struct pmu pmu;
 622
 623 static inline int is_x86_event(struct perf_event *event)
 624 {
 625         return event->pmu == &pmu;
 626 }
 627
 628 static int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 629 {
 630         struct event_constraint *c, *constraints[X86_PMC_IDX_MAX];
 631         unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 632         int i, j, w, wmax, num = 0;
 633         struct hw_perf_event *hwc;
 634
 635         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
 636
 637         for (i = 0; i < n; i++) {
 638                 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
 639                 constraints[i] = c;
 640         }
 641
 642         /*
 643          * fastpath, try to reuse previous register
 644          */
 645         for (i = 0; i < n; i++) {
 646                 hwc = &cpuc->event_list[i]->hw;
 647                 c = constraints[i];
 648
 649                 /* never assigned */
 650                 if (hwc->idx == -1)
 651                         break;
 652
 653                 /* constraint still honored */
 654                 if (!test_bit(hwc->idx, c->idxmsk))
 655                         break;
 656
 657                 /* not already used */
 658                 if (test_bit(hwc->idx, used_mask))
 659                         break;
 660
 661                 __set_bit(hwc->idx, used_mask);
 662                 if (assign)
 663                         assign[i] = hwc->idx;
 664         }
 665         if (i == n)
 666                 goto done;
 667
 668         /*
 669          * begin slow path
 670          */
 671
 672         bitmap_zero(used_mask, X86_PMC_IDX_MAX);
 673
 674         /*
 675          * weight = number of possible counters
 676          *
 677          * 1    = most constrained, only works on one counter
 678          * wmax = least constrained, works on any counter
 679          *
 680          * assign events to counters starting with most
 681          * constrained events.
 682          */
 683         wmax = x86_pmu.num_counters;
 684
 685         /*
 686          * when fixed event counters are present,
 687          * wmax is incremented by 1 to account
 688          * for one more choice
 689          */
 690         if (x86_pmu.num_counters_fixed)
 691                 wmax++;
 692
 693         for (w = 1, num = n; num && w <= wmax; w++) {
 694                 /* for each event */
 695                 for (i = 0; num && i < n; i++) {
 696                         c = constraints[i];
 697                         hwc = &cpuc->event_list[i]->hw;
 698
 699                         if (c->weight != w)
 700                                 continue;
 701
 702                         for_each_set_bit(j, c->idxmsk, X86_PMC_IDX_MAX) {
 703                                 if (!test_bit(j, used_mask))
 704                                         break;
 705                         }
 706
 707                         if (j == X86_PMC_IDX_MAX)
 708                                 break;
 709
 710                         __set_bit(j, used_mask);
 711
 712                         if (assign)
 713                                 assign[i] = j;
 714                         num--;
 715                 }
 716         }
 717 done:
 718         /*
 719          * scheduling failed or is just a simulation,
 720          * free resources if necessary
 721          */
 722         if (!assign || num) {
 723                 for (i = 0; i < n; i++) {
 724                         if (x86_pmu.put_event_constraints)
 725                                 x86_pmu.put_event_constraints(cpuc, cpuc->event_list[i]);
 726                 }
 727         }
 728         return num ? -ENOSPC : 0;
 729 }
 730
 731 /*
 732  * dogrp: true if must collect siblings events (group)
 733  * returns total number of events and error code
 734  */
 735 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
 736 {
 737         struct perf_event *event;
 738         int n, max_count;
 739
 740         max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
 741
 742         /* current number of events already accepted */
 743         n = cpuc->n_events;
 744
 745         if (is_x86_event(leader)) {
 746                 if (n >= max_count)
 747                         return -ENOSPC;
 748                 cpuc->event_list[n] = leader;
 749                 n++;
 750         }
 751         if (!dogrp)
 752                 return n;
 753
 754         list_for_each_entry(event, &leader->sibling_list, group_entry) {
 755                 if (!is_x86_event(event) ||
 756                     event->state <= PERF_EVENT_STATE_OFF)
 757                         continue;
 758
 759                 if (n >= max_count)
 760                         return -ENOSPC;
 761
 762                 cpuc->event_list[n] = event;
 763                 n++;
 764         }
 765         return n;
 766 }
 767
 768 static inline void x86_assign_hw_event(struct perf_event *event,
 769                                 struct cpu_hw_events *cpuc, int i)
 770 {
 771         struct hw_perf_event *hwc = &event->hw;
 772
 773         hwc->idx = cpuc->assign[i];
 774         hwc->last_cpu = smp_processor_id();
 775         hwc->last_tag = ++cpuc->tags[i];
 776
 777         if (hwc->idx == X86_PMC_IDX_FIXED_BTS) {
 778                 hwc->config_base = 0;
 779                 hwc->event_base = 0;
 780         } else if (hwc->idx >= X86_PMC_IDX_FIXED) {
 781                 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 782                 /*
 783                  * We set it so that event_base + idx in wrmsr/rdmsr maps to
 784                  * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
 785                  */
 786                 hwc->event_base =
 787                         MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
 788         } else {
 789                 hwc->config_base = x86_pmu.eventsel;
 790                 hwc->event_base  = x86_pmu.perfctr;
 791         }
 792 }
 793
 794 static inline int match_prev_assignment(struct hw_perf_event *hwc,
 795                                         struct cpu_hw_events *cpuc,
 796                                         int i)
 797 {
 798         return hwc->idx == cpuc->assign[i] &&
 799                 hwc->last_cpu == smp_processor_id() &&
 800                 hwc->last_tag == cpuc->tags[i];
 801 }
 802
 803 static void x86_pmu_start(struct perf_event *event, int flags);
 804 static void x86_pmu_stop(struct perf_event *event, int flags);
 805
 806 static void x86_pmu_enable(struct pmu *pmu)
 807 {
 808         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 809         struct perf_event *event;
 810         struct hw_perf_event *hwc;
 811         int i, added = cpuc->n_added;
 812
 813         if (!x86_pmu_initialized())
 814                 return;
 815
 816         if (cpuc->enabled)
 817                 return;
 818
 819         if (cpuc->n_added) {
 820                 int n_running = cpuc->n_events - cpuc->n_added;
 821                 /*
 822                  * apply assignment obtained either from
 823                  * hw_perf_group_sched_in() or x86_pmu_enable()
 824                  *
 825                  * step1: save events moving to new counters
 826                  * step2: reprogram moved events into new counters
 827                  */
 828                 for (i = 0; i < n_running; i++) {
 829                         event = cpuc->event_list[i];
 830                         hwc = &event->hw;
 831
 832                         /*
 833                          * we can avoid reprogramming counter if:
 834                          * - assigned same counter as last time
 835                          * - running on same CPU as last time
 836                          * - no other event has used the counter since
 837                          */
 838                         if (hwc->idx == -1 ||
 839                             match_prev_assignment(hwc, cpuc, i))
 840                                 continue;
 841
 842                         /*
 843                          * Ensure we don't accidentally enable a stopped
 844                          * counter simply because we rescheduled.
 845                          */
 846                         if (hwc->state & PERF_HES_STOPPED)
 847                                 hwc->state |= PERF_HES_ARCH;
 848
 849                         x86_pmu_stop(event, PERF_EF_UPDATE);
 850                 }
 851
 852                 for (i = 0; i < cpuc->n_events; i++) {
 853                         event = cpuc->event_list[i];
 854                         hwc = &event->hw;
 855
 856                         if (!match_prev_assignment(hwc, cpuc, i))
 857                                 x86_assign_hw_event(event, cpuc, i);
 858                         else if (i < n_running)
 859                                 continue;
 860
 861                         if (hwc->state & PERF_HES_ARCH)
 862                                 continue;
 863
 864                         x86_pmu_start(event, PERF_EF_RELOAD);
 865                 }
 866                 cpuc->n_added = 0;
 867                 perf_events_lapic_init();
 868         }
 869
 870         cpuc->enabled = 1;
 871         barrier();
 872
 873         x86_pmu.enable_all(added);
 874 }
 875
 876 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
 877                                           u64 enable_mask)
 878 {
 879         wrmsrl(hwc->config_base + hwc->idx, hwc->config | enable_mask);
 880 }
 881
 882 static inline void x86_pmu_disable_event(struct perf_event *event)
 883 {
 884         struct hw_perf_event *hwc = &event->hw;
 885
 886         wrmsrl(hwc->config_base + hwc->idx, hwc->config);
 887 }
 888
 889 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
 890
 891 /*
 892  * Set the next IRQ period, based on the hwc->period_left value.
 893  * To be called with the event disabled in hw:
 894  */
 895 static int
 896 x86_perf_event_set_period(struct perf_event *event)
 897 {
 898         struct hw_perf_event *hwc = &event->hw;
 899         s64 left = local64_read(&hwc->period_left);
 900         s64 period = hwc->sample_period;
 901         int ret = 0, idx = hwc->idx;
 902
 903         if (idx == X86_PMC_IDX_FIXED_BTS)
 904                 return 0;
 905
 906         /*
 907          * If we are way outside a reasonable range then just skip forward:
 908          */
 909         if (unlikely(left <= -period)) {
 910                 left = period;
 911                 local64_set(&hwc->period_left, left);
 912                 hwc->last_period = period;
 913                 ret = 1;
 914         }
 915
 916         if (unlikely(left <= 0)) {
 917                 left += period;
 918                 local64_set(&hwc->period_left, left);
 919                 hwc->last_period = period;
 920                 ret = 1;
 921         }
 922         /*
 923          * Quirk: certain CPUs dont like it if just 1 hw_event is left:
 924          */
 925         if (unlikely(left < 2))
 926                 left = 2;
 927
 928         if (left > x86_pmu.max_period)
 929                 left = x86_pmu.max_period;
 930
 931         per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
 932
 933         /*
 934          * The hw event starts counting from this event offset,
 935          * mark it to be able to extra future deltas:
 936          */
 937         local64_set(&hwc->prev_count, (u64)-left);
 938
 939         wrmsrl(hwc->event_base + idx, (u64)(-left) & x86_pmu.cntval_mask);
 940
 941         /*
 942          * Due to erratum on certan cpu we need
 943          * a second write to be sure the register
 944          * is updated properly
 945          */
 946         if (x86_pmu.perfctr_second_write) {
 947                 wrmsrl(hwc->event_base + idx,
 948                         (u64)(-left) & x86_pmu.cntval_mask);
 949         }
 950
 951         perf_event_update_userpage(event);
 952
 953         return ret;
 954 }
 955
 956 static void x86_pmu_enable_event(struct perf_event *event)
 957 {
 958         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 959         if (cpuc->enabled)
 960                 __x86_pmu_enable_event(&event->hw,
 961                                        ARCH_PERFMON_EVENTSEL_ENABLE);
 962 }
 963
 964 /*
 965  * Add a single event to the PMU.
 966  *
 967  * The event is added to the group of enabled events
 968  * but only if it can be scehduled with existing events.
 969  */
 970 static int x86_pmu_add(struct perf_event *event, int flags)
 971 {
 972         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 973         struct hw_perf_event *hwc;
 974         int assign[X86_PMC_IDX_MAX];
 975         int n, n0, ret;
 976
 977         hwc = &event->hw;
 978
 979         perf_pmu_disable(event->pmu);
 980         n0 = cpuc->n_events;
 981         ret = n = collect_events(cpuc, event, false);
 982         if (ret < 0)
 983                 goto out;
 984
 985         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 986         if (!(flags & PERF_EF_START))
 987                 hwc->state |= PERF_HES_ARCH;
 988
 989         /*
 990          * If group events scheduling transaction was started,
 991          * skip the schedulability test here, it will be peformed
 992          * at commit time (->commit_txn) as a whole
 993          */
 994         if (cpuc->group_flag & PERF_EVENT_TXN)
 995                 goto done_collect;
 996
 997         ret = x86_pmu.schedule_events(cpuc, n, assign);
 998         if (ret)
 999                 goto out;
1000         /*
1001          * copy new assignment, now we know it is possible
1002          * will be used by hw_perf_enable()
1003          */
1004         memcpy(cpuc->assign, assign, n*sizeof(int));
1005
1006 done_collect:
1007         cpuc->n_events = n;
1008         cpuc->n_added += n - n0;
1009         cpuc->n_txn += n - n0;
1010
1011         ret = 0;
1012 out:
1013         perf_pmu_enable(event->pmu);
1014         return ret;
1015 }
1016
1017 static void x86_pmu_start(struct perf_event *event, int flags)
1018 {
1019         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1020         int idx = event->hw.idx;
1021
1022         if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1023                 return;
1024
1025         if (WARN_ON_ONCE(idx == -1))
1026                 return;
1027
1028         if (flags & PERF_EF_RELOAD) {
1029                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1030                 x86_perf_event_set_period(event);
1031         }
1032
1033         event->hw.state = 0;
1034
1035         cpuc->events[idx] = event;
1036         __set_bit(idx, cpuc->active_mask);
1037         __set_bit(idx, cpuc->running);
1038         x86_pmu.enable(event);
1039         perf_event_update_userpage(event);
1040 }
1041
1042 void perf_event_print_debug(void)
1043 {
1044         u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1045         u64 pebs;
1046         struct cpu_hw_events *cpuc;
1047         unsigned long flags;
1048         int cpu, idx;
1049
1050         if (!x86_pmu.num_counters)
1051                 return;
1052
1053         local_irq_save(flags);
1054
1055         cpu = smp_processor_id();
1056         cpuc = &per_cpu(cpu_hw_events, cpu);
1057
1058         if (x86_pmu.version >= 2) {
1059                 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1060                 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1061                 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1062                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1063                 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1064
1065                 pr_info("\n");
1066                 pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1067                 pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1068                 pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1069                 pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1070                 pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1071         }
1072         pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1073
1074         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1075                 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
1076                 rdmsrl(x86_pmu.perfctr  + idx, pmc_count);
1077
1078                 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1079
1080                 pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1081                         cpu, idx, pmc_ctrl);
1082                 pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1083                         cpu, idx, pmc_count);
1084                 pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1085                         cpu, idx, prev_left);
1086         }
1087         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1088                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1089
1090                 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1091                         cpu, idx, pmc_count);
1092         }
1093         local_irq_restore(flags);
1094 }
1095
1096 static void x86_pmu_stop(struct perf_event *event, int flags)
1097 {
1098         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1099         struct hw_perf_event *hwc = &event->hw;
1100
1101         if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1102                 x86_pmu.disable(event);
1103                 cpuc->events[hwc->idx] = NULL;
1104                 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1105                 hwc->state |= PERF_HES_STOPPED;
1106         }
1107
1108         if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1109                 /*
1110                  * Drain the remaining delta count out of a event
1111                  * that we are disabling:
1112                  */
1113                 x86_perf_event_update(event);
1114                 hwc->state |= PERF_HES_UPTODATE;
1115         }
1116 }
1117
1118 static void x86_pmu_del(struct perf_event *event, int flags)
1119 {
1120         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1121         int i;
1122
1123         /*
1124          * If we're called during a txn, we don't need to do anything.
1125          * The events never got scheduled and ->cancel_txn will truncate
1126          * the event_list.
1127          */
1128         if (cpuc->group_flag & PERF_EVENT_TXN)
1129                 return;
1130
1131         x86_pmu_stop(event, PERF_EF_UPDATE);
1132
1133         for (i = 0; i < cpuc->n_events; i++) {
1134                 if (event == cpuc->event_list[i]) {
1135
1136                         if (x86_pmu.put_event_constraints)
1137                                 x86_pmu.put_event_constraints(cpuc, event);
1138
1139                         while (++i < cpuc->n_events)
1140                                 cpuc->event_list[i-1] = cpuc->event_list[i];
1141
1142                         --cpuc->n_events;
1143                         break;
1144                 }
1145         }
1146         perf_event_update_userpage(event);
1147 }
1148
1149 static int x86_pmu_handle_irq(struct pt_regs *regs)
1150 {
1151         struct perf_sample_data data;
1152         struct cpu_hw_events *cpuc;
1153         struct perf_event *event;
1154         int idx, handled = 0;
1155         u64 val;
1156
1157         perf_sample_data_init(&data, 0);
1158
1159         cpuc = &__get_cpu_var(cpu_hw_events);
1160
1161         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1162                 if (!test_bit(idx, cpuc->active_mask)) {
1163                         /*
1164                          * Though we deactivated the counter some cpus
1165                          * might still deliver spurious interrupts still
1166                          * in flight. Catch them:
1167                          */
1168                         if (__test_and_clear_bit(idx, cpuc->running))
1169                                 handled++;
1170                         continue;
1171                 }
1172
1173                 event = cpuc->events[idx];
1174
1175                 val = x86_perf_event_update(event);
1176                 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1177                         continue;
1178
1179                 /*
1180                  * event overflow
1181                  */
1182                 handled++;
1183                 data.period     = event->hw.last_period;
1184
1185                 if (!x86_perf_event_set_period(event))
1186                         continue;
1187
1188                 if (perf_event_overflow(event, 1, &data, regs))
1189                         x86_pmu_stop(event, 0);
1190         }
1191
1192         if (handled)
1193                 inc_irq_stat(apic_perf_irqs);
1194
1195         return handled;
1196 }
1197
1198 void perf_events_lapic_init(void)
1199 {
1200         if (!x86_pmu.apic || !x86_pmu_initialized())
1201                 return;
1202
1203         /*
1204          * Always use NMI for PMU
1205          */
1206         apic_write(APIC_LVTPC, APIC_DM_NMI);
1207 }
1208
1209 struct pmu_nmi_state {
1210         unsigned int    marked;
1211         int             handled;
1212 };
1213
1214 static DEFINE_PER_CPU(struct pmu_nmi_state, pmu_nmi);
1215
1216 static int __kprobes
1217 perf_event_nmi_handler(struct notifier_block *self,
1218                          unsigned long cmd, void *__args)
1219 {
1220         struct die_args *args = __args;
1221         unsigned int this_nmi;
1222         int handled;
1223
1224         if (!atomic_read(&active_events))
1225                 return NOTIFY_DONE;
1226
1227         switch (cmd) {
1228         case DIE_NMI:
1229         case DIE_NMI_IPI:
1230                 break;
1231         case DIE_NMIUNKNOWN:
1232                 this_nmi = percpu_read(irq_stat.__nmi_count);
1233                 if (this_nmi != __get_cpu_var(pmu_nmi).marked)
1234                         /* let the kernel handle the unknown nmi */
1235                         return NOTIFY_DONE;
1236                 /*
1237                  * This one is a PMU back-to-back nmi. Two events
1238                  * trigger 'simultaneously' raising two back-to-back
1239                  * NMIs. If the first NMI handles both, the latter
1240                  * will be empty and daze the CPU. So, we drop it to
1241                  * avoid false-positive 'unknown nmi' messages.
1242                  */
1243                 return NOTIFY_STOP;
1244         default:
1245                 return NOTIFY_DONE;
1246         }
1247
1248         apic_write(APIC_LVTPC, APIC_DM_NMI);
1249
1250         handled = x86_pmu.handle_irq(args->regs);
1251         if (!handled)
1252                 return NOTIFY_DONE;
1253
1254         this_nmi = percpu_read(irq_stat.__nmi_count);
1255         if ((handled > 1) ||
1256                 /* the next nmi could be a back-to-back nmi */
1257             ((__get_cpu_var(pmu_nmi).marked == this_nmi) &&
1258              (__get_cpu_var(pmu_nmi).handled > 1))) {
1259                 /*
1260                  * We could have two subsequent back-to-back nmis: The
1261                  * first handles more than one counter, the 2nd
1262                  * handles only one counter and the 3rd handles no
1263                  * counter.
1264                  *
1265                  * This is the 2nd nmi because the previous was
1266                  * handling more than one counter. We will mark the
1267                  * next (3rd) and then drop it if unhandled.
1268                  */
1269                 __get_cpu_var(pmu_nmi).marked   = this_nmi + 1;
1270                 __get_cpu_var(pmu_nmi).handled  = handled;
1271         }
1272
1273         return NOTIFY_STOP;
1274 }
1275
1276 static __read_mostly struct notifier_block perf_event_nmi_notifier = {
1277         .notifier_call          = perf_event_nmi_handler,
1278         .next                   = NULL,
1279         .priority               = 1
1280 };
1281
1282 static struct event_constraint unconstrained;
1283 static struct event_constraint emptyconstraint;
1284
1285 static struct event_constraint *
1286 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1287 {
1288         struct event_constraint *c;
1289
1290         if (x86_pmu.event_constraints) {
1291                 for_each_event_constraint(c, x86_pmu.event_constraints) {
1292                         if ((event->hw.config & c->cmask) == c->code)
1293                                 return c;
1294                 }
1295         }
1296
1297         return &unconstrained;
1298 }
1299
1300 #include "perf_event_amd.c"
1301 #include "perf_event_p6.c"
1302 #include "perf_event_p4.c"
1303 #include "perf_event_intel_lbr.c"
1304 #include "perf_event_intel_ds.c"
1305 #include "perf_event_intel.c"
1306
1307 static int __cpuinit
1308 x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1309 {
1310         unsigned int cpu = (long)hcpu;
1311         int ret = NOTIFY_OK;
1312
1313         switch (action & ~CPU_TASKS_FROZEN) {
1314         case CPU_UP_PREPARE:
1315                 if (x86_pmu.cpu_prepare)
1316                         ret = x86_pmu.cpu_prepare(cpu);
1317                 break;
1318
1319         case CPU_STARTING:
1320                 if (x86_pmu.cpu_starting)
1321                         x86_pmu.cpu_starting(cpu);
1322                 break;
1323
1324         case CPU_DYING:
1325                 if (x86_pmu.cpu_dying)
1326                         x86_pmu.cpu_dying(cpu);
1327                 break;
1328
1329         case CPU_UP_CANCELED:
1330         case CPU_DEAD:
1331                 if (x86_pmu.cpu_dead)
1332                         x86_pmu.cpu_dead(cpu);
1333                 break;
1334
1335         default:
1336                 break;
1337         }
1338
1339         return ret;
1340 }
1341
1342 static void __init pmu_check_apic(void)
1343 {
1344         if (cpu_has_apic)
1345                 return;
1346
1347         x86_pmu.apic = 0;
1348         pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1349         pr_info("no hardware sampling interrupt available.\n");
1350 }
1351
1352 void __init init_hw_perf_events(void)
1353 {
1354         struct event_constraint *c;
1355         int err;
1356
1357         pr_info("Performance Events: ");
1358
1359         switch (boot_cpu_data.x86_vendor) {
1360         case X86_VENDOR_INTEL:
1361                 err = intel_pmu_init();
1362                 break;
1363         case X86_VENDOR_AMD:
1364                 err = amd_pmu_init();
1365                 break;
1366         default:
1367                 return;
1368         }
1369         if (err != 0) {
1370                 pr_cont("no PMU driver, software events only.\n");
1371                 return;
1372         }
1373
1374         pmu_check_apic();
1375
1376         pr_cont("%s PMU driver.\n", x86_pmu.name);
1377
1378         if (x86_pmu.quirks)
1379                 x86_pmu.quirks();
1380
1381         if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1382                 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!",
1383                      x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1384                 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1385         }
1386         x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1387
1388         if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1389                 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!",
1390                      x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1391                 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1392         }
1393
1394         x86_pmu.intel_ctrl |=
1395                 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1396
1397         perf_events_lapic_init();
1398         register_die_notifier(&perf_event_nmi_notifier);
1399
1400         unconstrained = (struct event_constraint)
1401                 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1402                                    0, x86_pmu.num_counters);
1403
1404         if (x86_pmu.event_constraints) {
1405                 for_each_event_constraint(c, x86_pmu.event_constraints) {
1406                         if (c->cmask != X86_RAW_EVENT_MASK)
1407                                 continue;
1408
1409                         c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1;
1410                         c->weight += x86_pmu.num_counters;
1411                 }
1412         }
1413
1414         pr_info("... version:                %d\n",     x86_pmu.version);
1415         pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1416         pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1417         pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1418         pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1419         pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1420         pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1421
1422         perf_pmu_register(&pmu);
1423         perf_cpu_notifier(x86_pmu_notifier);
1424 }
1425
1426 static inline void x86_pmu_read(struct perf_event *event)
1427 {
1428         x86_perf_event_update(event);
1429 }
1430
1431 /*
1432  * Start group events scheduling transaction
1433  * Set the flag to make pmu::enable() not perform the
1434  * schedulability test, it will be performed at commit time
1435  */
1436 static void x86_pmu_start_txn(struct pmu *pmu)
1437 {
1438         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1439
1440         perf_pmu_disable(pmu);
1441         cpuc->group_flag |= PERF_EVENT_TXN;
1442         cpuc->n_txn = 0;
1443 }
1444
1445 /*
1446  * Stop group events scheduling transaction
1447  * Clear the flag and pmu::enable() will perform the
1448  * schedulability test.
1449  */
1450 static void x86_pmu_cancel_txn(struct pmu *pmu)
1451 {
1452         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1453
1454         cpuc->group_flag &= ~PERF_EVENT_TXN;
1455         /*
1456          * Truncate the collected events.
1457          */
1458         cpuc->n_added -= cpuc->n_txn;
1459         cpuc->n_events -= cpuc->n_txn;
1460         perf_pmu_enable(pmu);
1461 }
1462
1463 /*
1464  * Commit group events scheduling transaction
1465  * Perform the group schedulability test as a whole
1466  * Return 0 if success
1467  */
1468 static int x86_pmu_commit_txn(struct pmu *pmu)
1469 {
1470         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1471         int assign[X86_PMC_IDX_MAX];
1472         int n, ret;
1473
1474         n = cpuc->n_events;
1475
1476         if (!x86_pmu_initialized())
1477                 return -EAGAIN;
1478
1479         ret = x86_pmu.schedule_events(cpuc, n, assign);
1480         if (ret)
1481                 return ret;
1482
1483         /*
1484          * copy new assignment, now we know it is possible
1485          * will be used by hw_perf_enable()
1486          */
1487         memcpy(cpuc->assign, assign, n*sizeof(int));
1488
1489         cpuc->group_flag &= ~PERF_EVENT_TXN;
1490         perf_pmu_enable(pmu);
1491         return 0;
1492 }
1493
1494 /*
1495  * validate that we can schedule this event
1496  */
1497 static int validate_event(struct perf_event *event)
1498 {
1499         struct cpu_hw_events *fake_cpuc;
1500         struct event_constraint *c;
1501         int ret = 0;
1502
1503         fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1504         if (!fake_cpuc)
1505                 return -ENOMEM;
1506
1507         c = x86_pmu.get_event_constraints(fake_cpuc, event);
1508
1509         if (!c || !c->weight)
1510                 ret = -ENOSPC;
1511
1512         if (x86_pmu.put_event_constraints)
1513                 x86_pmu.put_event_constraints(fake_cpuc, event);
1514
1515         kfree(fake_cpuc);
1516
1517         return ret;
1518 }
1519
1520 /*
1521  * validate a single event group
1522  *
1523  * validation include:
1524  *      - check events are compatible which each other
1525  *      - events do not compete for the same counter
1526  *      - number of events <= number of counters
1527  *
1528  * validation ensures the group can be loaded onto the
1529  * PMU if it was the only group available.
1530  */
1531 static int validate_group(struct perf_event *event)
1532 {
1533         struct perf_event *leader = event->group_leader;
1534         struct cpu_hw_events *fake_cpuc;
1535         int ret, n;
1536
1537         ret = -ENOMEM;
1538         fake_cpuc = kmalloc(sizeof(*fake_cpuc), GFP_KERNEL | __GFP_ZERO);
1539         if (!fake_cpuc)
1540                 goto out;
1541
1542         /*
1543          * the event is not yet connected with its
1544          * siblings therefore we must first collect
1545          * existing siblings, then add the new event
1546          * before we can simulate the scheduling
1547          */
1548         ret = -ENOSPC;
1549         n = collect_events(fake_cpuc, leader, true);
1550         if (n < 0)
1551                 goto out_free;
1552
1553         fake_cpuc->n_events = n;
1554         n = collect_events(fake_cpuc, event, false);
1555         if (n < 0)
1556                 goto out_free;
1557
1558         fake_cpuc->n_events = n;
1559
1560         ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1561
1562 out_free:
1563         kfree(fake_cpuc);
1564 out:
1565         return ret;
1566 }
1567
1568 int x86_pmu_event_init(struct perf_event *event)
1569 {
1570         struct pmu *tmp;
1571         int err;
1572
1573         switch (event->attr.type) {
1574         case PERF_TYPE_RAW:
1575         case PERF_TYPE_HARDWARE:
1576         case PERF_TYPE_HW_CACHE:
1577                 break;
1578
1579         default:
1580                 return -ENOENT;
1581         }
1582
1583         err = __x86_pmu_event_init(event);
1584         if (!err) {
1585                 /*
1586                  * we temporarily connect event to its pmu
1587                  * such that validate_group() can classify
1588                  * it as an x86 event using is_x86_event()
1589                  */
1590                 tmp = event->pmu;
1591                 event->pmu = &pmu;
1592
1593                 if (event->group_leader != event)
1594                         err = validate_group(event);
1595                 else
1596                         err = validate_event(event);
1597
1598                 event->pmu = tmp;
1599         }
1600         if (err) {
1601                 if (event->destroy)
1602                         event->destroy(event);
1603         }
1604
1605         return err;
1606 }
1607
1608 static struct pmu pmu = {
1609         .pmu_enable     = x86_pmu_enable,
1610         .pmu_disable    = x86_pmu_disable,
1611
1612         .event_init     = x86_pmu_event_init,
1613
1614         .add            = x86_pmu_add,
1615         .del            = x86_pmu_del,
1616         .start          = x86_pmu_start,
1617         .stop           = x86_pmu_stop,
1618         .read           = x86_pmu_read,
1619
1620         .start_txn      = x86_pmu_start_txn,
1621         .cancel_txn     = x86_pmu_cancel_txn,
1622         .commit_txn     = x86_pmu_commit_txn,
1623 };
1624
1625 /*
1626  * callchain support
1627  */
1628
1629 static void
1630 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1631 {
1632         /* Ignore warnings */
1633 }
1634
1635 static void backtrace_warning(void *data, char *msg)
1636 {
1637         /* Ignore warnings */
1638 }
1639
1640 static int backtrace_stack(void *data, char *name)
1641 {
1642         return 0;
1643 }
1644
1645 static void backtrace_address(void *data, unsigned long addr, int reliable)
1646 {
1647         struct perf_callchain_entry *entry = data;
1648
1649         perf_callchain_store(entry, addr);
1650 }
1651
1652 static const struct stacktrace_ops backtrace_ops = {
1653         .warning                = backtrace_warning,
1654         .warning_symbol         = backtrace_warning_symbol,
1655         .stack                  = backtrace_stack,
1656         .address                = backtrace_address,
1657         .walk_stack             = print_context_stack_bp,
1658 };
1659
1660 void
1661 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1662 {
1663         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1664                 /* TODO: We don't support guest os callchain now */
1665                 return;
1666         }
1667
1668         perf_callchain_store(entry, regs->ip);
1669
1670         dump_trace(NULL, regs, NULL, regs->bp, &backtrace_ops, entry);
1671 }
1672
1673 #ifdef CONFIG_COMPAT
1674 static inline int
1675 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1676 {
1677         /* 32-bit process in 64-bit kernel. */
1678         struct stack_frame_ia32 frame;
1679         const void __user *fp;
1680
1681         if (!test_thread_flag(TIF_IA32))
1682                 return 0;
1683
1684         fp = compat_ptr(regs->bp);
1685         while (entry->nr < PERF_MAX_STACK_DEPTH) {
1686                 unsigned long bytes;
1687                 frame.next_frame     = 0;
1688                 frame.return_address = 0;
1689
1690                 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1691                 if (bytes != sizeof(frame))
1692                         break;
1693
1694                 if (fp < compat_ptr(regs->sp))
1695                         break;
1696
1697                 perf_callchain_store(entry, frame.return_address);
1698                 fp = compat_ptr(frame.next_frame);
1699         }
1700         return 1;
1701 }
1702 #else
1703 static inline int
1704 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
1705 {
1706     return 0;
1707 }
1708 #endif
1709
1710 void
1711 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1712 {
1713         struct stack_frame frame;
1714         const void __user *fp;
1715
1716         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1717                 /* TODO: We don't support guest os callchain now */
1718                 return;
1719         }
1720
1721         fp = (void __user *)regs->bp;
1722
1723         perf_callchain_store(entry, regs->ip);
1724
1725         if (perf_callchain_user32(regs, entry))
1726                 return;
1727
1728         while (entry->nr < PERF_MAX_STACK_DEPTH) {
1729                 unsigned long bytes;
1730                 frame.next_frame             = NULL;
1731                 frame.return_address = 0;
1732
1733                 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
1734                 if (bytes != sizeof(frame))
1735                         break;
1736
1737                 if ((unsigned long)fp < regs->sp)
1738                         break;
1739
1740                 perf_callchain_store(entry, frame.return_address);
1741                 fp = frame.next_frame;
1742         }
1743 }
1744
1745 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1746 {
1747         unsigned long ip;
1748
1749         if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
1750                 ip = perf_guest_cbs->get_guest_ip();
1751         else
1752                 ip = instruction_pointer(regs);
1753
1754         return ip;
1755 }
1756
1757 unsigned long perf_misc_flags(struct pt_regs *regs)
1758 {
1759         int misc = 0;
1760
1761         if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1762                 if (perf_guest_cbs->is_user_mode())
1763                         misc |= PERF_RECORD_MISC_GUEST_USER;
1764                 else
1765                         misc |= PERF_RECORD_MISC_GUEST_KERNEL;
1766         } else {
1767                 if (user_mode(regs))
1768                         misc |= PERF_RECORD_MISC_USER;
1769                 else
1770                         misc |= PERF_RECORD_MISC_KERNEL;
1771         }
1772
1773         if (regs->flags & PERF_EFLAGS_EXACT)
1774                 misc |= PERF_RECORD_MISC_EXACT_IP;
1775
1776         return misc;
1777 }