arch/x86/kernel/cpu/perf_counter.c

   1 /*
   2  * Performance counter 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  *
  10  *  For licencing details see kernel-base/COPYING
  11  */
  12
  13 #include <linux/perf_counter.h>
  14 #include <linux/capability.h>
  15 #include <linux/notifier.h>
  16 #include <linux/hardirq.h>
  17 #include <linux/kprobes.h>
  18 #include <linux/module.h>
  19 #include <linux/kdebug.h>
  20 #include <linux/sched.h>
  21 #include <linux/uaccess.h>
  22
  23 #include <asm/apic.h>
  24 #include <asm/stacktrace.h>
  25 #include <asm/nmi.h>
  26
  27 static u64 perf_counter_mask __read_mostly;
  28
  29 struct cpu_hw_counters {
  30         struct perf_counter     *counters[X86_PMC_IDX_MAX];
  31         unsigned long           used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
  32         unsigned long           active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
  33         unsigned long           interrupts;
  34         u64                     throttle_ctrl;
  35         int                     enabled;
  36 };
  37
  38 /*
  39  * struct x86_pmu - generic x86 pmu
  40  */
  41 struct x86_pmu {
  42         const char      *name;
  43         int             version;
  44         int             (*handle_irq)(struct pt_regs *, int);
  45         u64             (*save_disable_all)(void);
  46         void            (*restore_all)(u64);
  47         void            (*enable)(struct hw_perf_counter *, int);
  48         void            (*disable)(struct hw_perf_counter *, int);
  49         unsigned        eventsel;
  50         unsigned        perfctr;
  51         u64             (*event_map)(int);
  52         u64             (*raw_event)(u64);
  53         int             max_events;
  54         int             num_counters;
  55         int             num_counters_fixed;
  56         int             counter_bits;
  57         u64             counter_mask;
  58         u64             max_period;
  59 };
  60
  61 static struct x86_pmu x86_pmu __read_mostly;
  62
  63 static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = {
  64         .enabled = 1,
  65 };
  66
  67 /*
  68  * Intel PerfMon v3. Used on Core2 and later.
  69  */
  70 static const u64 intel_perfmon_event_map[] =
  71 {
  72   [PERF_COUNT_CPU_CYCLES]               = 0x003c,
  73   [PERF_COUNT_INSTRUCTIONS]             = 0x00c0,
  74   [PERF_COUNT_CACHE_REFERENCES]         = 0x4f2e,
  75   [PERF_COUNT_CACHE_MISSES]             = 0x412e,
  76   [PERF_COUNT_BRANCH_INSTRUCTIONS]      = 0x00c4,
  77   [PERF_COUNT_BRANCH_MISSES]            = 0x00c5,
  78   [PERF_COUNT_BUS_CYCLES]               = 0x013c,
  79 };
  80
  81 static u64 intel_pmu_event_map(int event)
  82 {
  83         return intel_perfmon_event_map[event];
  84 }
  85
  86 static u64 intel_pmu_raw_event(u64 event)
  87 {
  88 #define CORE_EVNTSEL_EVENT_MASK         0x000000FFULL
  89 #define CORE_EVNTSEL_UNIT_MASK          0x0000FF00ULL
  90 #define CORE_EVNTSEL_COUNTER_MASK       0xFF000000ULL
  91
  92 #define CORE_EVNTSEL_MASK               \
  93         (CORE_EVNTSEL_EVENT_MASK |      \
  94          CORE_EVNTSEL_UNIT_MASK  |      \
  95          CORE_EVNTSEL_COUNTER_MASK)
  96
  97         return event & CORE_EVNTSEL_MASK;
  98 }
  99
 100 /*
 101  * AMD Performance Monitor K7 and later.
 102  */
 103 static const u64 amd_perfmon_event_map[] =
 104 {
 105   [PERF_COUNT_CPU_CYCLES]               = 0x0076,
 106   [PERF_COUNT_INSTRUCTIONS]             = 0x00c0,
 107   [PERF_COUNT_CACHE_REFERENCES]         = 0x0080,
 108   [PERF_COUNT_CACHE_MISSES]             = 0x0081,
 109   [PERF_COUNT_BRANCH_INSTRUCTIONS]      = 0x00c4,
 110   [PERF_COUNT_BRANCH_MISSES]            = 0x00c5,
 111 };
 112
 113 static u64 amd_pmu_event_map(int event)
 114 {
 115         return amd_perfmon_event_map[event];
 116 }
 117
 118 static u64 amd_pmu_raw_event(u64 event)
 119 {
 120 #define K7_EVNTSEL_EVENT_MASK   0x7000000FFULL
 121 #define K7_EVNTSEL_UNIT_MASK    0x00000FF00ULL
 122 #define K7_EVNTSEL_COUNTER_MASK 0x0FF000000ULL
 123
 124 #define K7_EVNTSEL_MASK                 \
 125         (K7_EVNTSEL_EVENT_MASK |        \
 126          K7_EVNTSEL_UNIT_MASK  |        \
 127          K7_EVNTSEL_COUNTER_MASK)
 128
 129         return event & K7_EVNTSEL_MASK;
 130 }
 131
 132 /*
 133  * Propagate counter elapsed time into the generic counter.
 134  * Can only be executed on the CPU where the counter is active.
 135  * Returns the delta events processed.
 136  */
 137 static u64
 138 x86_perf_counter_update(struct perf_counter *counter,
 139                         struct hw_perf_counter *hwc, int idx)
 140 {
 141         u64 prev_raw_count, new_raw_count, delta;
 142
 143         /*
 144          * Careful: an NMI might modify the previous counter value.
 145          *
 146          * Our tactic to handle this is to first atomically read and
 147          * exchange a new raw count - then add that new-prev delta
 148          * count to the generic counter atomically:
 149          */
 150 again:
 151         prev_raw_count = atomic64_read(&hwc->prev_count);
 152         rdmsrl(hwc->counter_base + idx, new_raw_count);
 153
 154         if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
 155                                         new_raw_count) != prev_raw_count)
 156                 goto again;
 157
 158         /*
 159          * Now we have the new raw value and have updated the prev
 160          * timestamp already. We can now calculate the elapsed delta
 161          * (counter-)time and add that to the generic counter.
 162          *
 163          * Careful, not all hw sign-extends above the physical width
 164          * of the count, so we do that by clipping the delta to 32 bits:
 165          */
 166         delta = (u64)(u32)((s32)new_raw_count - (s32)prev_raw_count);
 167
 168         atomic64_add(delta, &counter->count);
 169         atomic64_sub(delta, &hwc->period_left);
 170
 171         return new_raw_count;
 172 }
 173
 174 static atomic_t active_counters;
 175 static DEFINE_MUTEX(pmc_reserve_mutex);
 176
 177 static bool reserve_pmc_hardware(void)
 178 {
 179         int i;
 180
 181         if (nmi_watchdog == NMI_LOCAL_APIC)
 182                 disable_lapic_nmi_watchdog();
 183
 184         for (i = 0; i < x86_pmu.num_counters; i++) {
 185                 if (!reserve_perfctr_nmi(x86_pmu.perfctr + i))
 186                         goto perfctr_fail;
 187         }
 188
 189         for (i = 0; i < x86_pmu.num_counters; i++) {
 190                 if (!reserve_evntsel_nmi(x86_pmu.eventsel + i))
 191                         goto eventsel_fail;
 192         }
 193
 194         return true;
 195
 196 eventsel_fail:
 197         for (i--; i >= 0; i--)
 198                 release_evntsel_nmi(x86_pmu.eventsel + i);
 199
 200         i = x86_pmu.num_counters;
 201
 202 perfctr_fail:
 203         for (i--; i >= 0; i--)
 204                 release_perfctr_nmi(x86_pmu.perfctr + i);
 205
 206         if (nmi_watchdog == NMI_LOCAL_APIC)
 207                 enable_lapic_nmi_watchdog();
 208
 209         return false;
 210 }
 211
 212 static void release_pmc_hardware(void)
 213 {
 214         int i;
 215
 216         for (i = 0; i < x86_pmu.num_counters; i++) {
 217                 release_perfctr_nmi(x86_pmu.perfctr + i);
 218                 release_evntsel_nmi(x86_pmu.eventsel + i);
 219         }
 220
 221         if (nmi_watchdog == NMI_LOCAL_APIC)
 222                 enable_lapic_nmi_watchdog();
 223 }
 224
 225 static void hw_perf_counter_destroy(struct perf_counter *counter)
 226 {
 227         if (atomic_dec_and_mutex_lock(&active_counters, &pmc_reserve_mutex)) {
 228                 release_pmc_hardware();
 229                 mutex_unlock(&pmc_reserve_mutex);
 230         }
 231 }
 232
 233 static inline int x86_pmu_initialized(void)
 234 {
 235         return x86_pmu.handle_irq != NULL;
 236 }
 237
 238 /*
 239  * Setup the hardware configuration for a given hw_event_type
 240  */
 241 static int __hw_perf_counter_init(struct perf_counter *counter)
 242 {
 243         struct perf_counter_hw_event *hw_event = &counter->hw_event;
 244         struct hw_perf_counter *hwc = &counter->hw;
 245         int err;
 246
 247         if (!x86_pmu_initialized())
 248                 return -ENODEV;
 249
 250         err = 0;
 251         if (!atomic_inc_not_zero(&active_counters)) {
 252                 mutex_lock(&pmc_reserve_mutex);
 253                 if (atomic_read(&active_counters) == 0 && !reserve_pmc_hardware())
 254                         err = -EBUSY;
 255                 else
 256                         atomic_inc(&active_counters);
 257                 mutex_unlock(&pmc_reserve_mutex);
 258         }
 259         if (err)
 260                 return err;
 261
 262         /*
 263          * Generate PMC IRQs:
 264          * (keep 'enabled' bit clear for now)
 265          */
 266         hwc->config = ARCH_PERFMON_EVENTSEL_INT;
 267
 268         /*
 269          * Count user and OS events unless requested not to.
 270          */
 271         if (!hw_event->exclude_user)
 272                 hwc->config |= ARCH_PERFMON_EVENTSEL_USR;
 273         if (!hw_event->exclude_kernel)
 274                 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
 275
 276         /*
 277          * If privileged enough, allow NMI events:
 278          */
 279         hwc->nmi = 0;
 280         if (capable(CAP_SYS_ADMIN) && hw_event->nmi)
 281                 hwc->nmi = 1;
 282
 283         hwc->irq_period = hw_event->irq_period;
 284         if ((s64)hwc->irq_period <= 0 || hwc->irq_period > x86_pmu.max_period)
 285                 hwc->irq_period = x86_pmu.max_period;
 286
 287         atomic64_set(&hwc->period_left, hwc->irq_period);
 288
 289         /*
 290          * Raw event type provide the config in the event structure
 291          */
 292         if (perf_event_raw(hw_event)) {
 293                 hwc->config |= x86_pmu.raw_event(perf_event_config(hw_event));
 294         } else {
 295                 if (perf_event_id(hw_event) >= x86_pmu.max_events)
 296                         return -EINVAL;
 297                 /*
 298                  * The generic map:
 299                  */
 300                 hwc->config |= x86_pmu.event_map(perf_event_id(hw_event));
 301         }
 302
 303         counter->destroy = hw_perf_counter_destroy;
 304
 305         return 0;
 306 }
 307
 308 static u64 intel_pmu_save_disable_all(void)
 309 {
 310         u64 ctrl;
 311
 312         rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
 313         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
 314
 315         return ctrl;
 316 }
 317
 318 static u64 amd_pmu_save_disable_all(void)
 319 {
 320         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 321         int enabled, idx;
 322
 323         enabled = cpuc->enabled;
 324         cpuc->enabled = 0;
 325         /*
 326          * ensure we write the disable before we start disabling the
 327          * counters proper, so that amd_pmu_enable_counter() does the
 328          * right thing.
 329          */
 330         barrier();
 331
 332         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 333                 u64 val;
 334
 335                 if (!test_bit(idx, cpuc->active_mask))
 336                         continue;
 337                 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
 338                 if (!(val & ARCH_PERFMON_EVENTSEL0_ENABLE))
 339                         continue;
 340                 val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
 341                 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
 342         }
 343
 344         return enabled;
 345 }
 346
 347 u64 hw_perf_save_disable(void)
 348 {
 349         if (!x86_pmu_initialized())
 350                 return 0;
 351         return x86_pmu.save_disable_all();
 352 }
 353 /*
 354  * Exported because of ACPI idle
 355  */
 356 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
 357
 358 static void intel_pmu_restore_all(u64 ctrl)
 359 {
 360         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
 361 }
 362
 363 static void amd_pmu_restore_all(u64 ctrl)
 364 {
 365         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 366         int idx;
 367
 368         cpuc->enabled = ctrl;
 369         barrier();
 370         if (!ctrl)
 371                 return;
 372
 373         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 374                 u64 val;
 375
 376                 if (!test_bit(idx, cpuc->active_mask))
 377                         continue;
 378                 rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
 379                 if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
 380                         continue;
 381                 val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 382                 wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
 383         }
 384 }
 385
 386 void hw_perf_restore(u64 ctrl)
 387 {
 388         if (!x86_pmu_initialized())
 389                 return;
 390         x86_pmu.restore_all(ctrl);
 391 }
 392 /*
 393  * Exported because of ACPI idle
 394  */
 395 EXPORT_SYMBOL_GPL(hw_perf_restore);
 396
 397 static inline u64 intel_pmu_get_status(void)
 398 {
 399         u64 status;
 400
 401         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
 402
 403         return status;
 404 }
 405
 406 static inline void intel_pmu_ack_status(u64 ack)
 407 {
 408         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
 409 }
 410
 411 static inline void x86_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
 412 {
 413         int err;
 414         err = checking_wrmsrl(hwc->config_base + idx,
 415                               hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE);
 416 }
 417
 418 static inline void x86_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
 419 {
 420         int err;
 421         err = checking_wrmsrl(hwc->config_base + idx,
 422                               hwc->config);
 423 }
 424
 425 static inline void
 426 intel_pmu_disable_fixed(struct hw_perf_counter *hwc, int __idx)
 427 {
 428         int idx = __idx - X86_PMC_IDX_FIXED;
 429         u64 ctrl_val, mask;
 430         int err;
 431
 432         mask = 0xfULL << (idx * 4);
 433
 434         rdmsrl(hwc->config_base, ctrl_val);
 435         ctrl_val &= ~mask;
 436         err = checking_wrmsrl(hwc->config_base, ctrl_val);
 437 }
 438
 439 static inline void
 440 intel_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
 441 {
 442         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 443                 intel_pmu_disable_fixed(hwc, idx);
 444                 return;
 445         }
 446
 447         x86_pmu_disable_counter(hwc, idx);
 448 }
 449
 450 static inline void
 451 amd_pmu_disable_counter(struct hw_perf_counter *hwc, int idx)
 452 {
 453         x86_pmu_disable_counter(hwc, idx);
 454 }
 455
 456 static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
 457
 458 /*
 459  * Set the next IRQ period, based on the hwc->period_left value.
 460  * To be called with the counter disabled in hw:
 461  */
 462 static void
 463 x86_perf_counter_set_period(struct perf_counter *counter,
 464                              struct hw_perf_counter *hwc, int idx)
 465 {
 466         s64 left = atomic64_read(&hwc->period_left);
 467         s64 period = hwc->irq_period;
 468         int err;
 469
 470         /*
 471          * If we are way outside a reasoable range then just skip forward:
 472          */
 473         if (unlikely(left <= -period)) {
 474                 left = period;
 475                 atomic64_set(&hwc->period_left, left);
 476         }
 477
 478         if (unlikely(left <= 0)) {
 479                 left += period;
 480                 atomic64_set(&hwc->period_left, left);
 481         }
 482
 483         per_cpu(prev_left[idx], smp_processor_id()) = left;
 484
 485         /*
 486          * The hw counter starts counting from this counter offset,
 487          * mark it to be able to extra future deltas:
 488          */
 489         atomic64_set(&hwc->prev_count, (u64)-left);
 490
 491         err = checking_wrmsrl(hwc->counter_base + idx,
 492                              (u64)(-left) & x86_pmu.counter_mask);
 493 }
 494
 495 static inline void
 496 intel_pmu_enable_fixed(struct hw_perf_counter *hwc, int __idx)
 497 {
 498         int idx = __idx - X86_PMC_IDX_FIXED;
 499         u64 ctrl_val, bits, mask;
 500         int err;
 501
 502         /*
 503          * Enable IRQ generation (0x8),
 504          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
 505          * if requested:
 506          */
 507         bits = 0x8ULL;
 508         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
 509                 bits |= 0x2;
 510         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
 511                 bits |= 0x1;
 512         bits <<= (idx * 4);
 513         mask = 0xfULL << (idx * 4);
 514
 515         rdmsrl(hwc->config_base, ctrl_val);
 516         ctrl_val &= ~mask;
 517         ctrl_val |= bits;
 518         err = checking_wrmsrl(hwc->config_base, ctrl_val);
 519 }
 520
 521 static void intel_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
 522 {
 523         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 524                 intel_pmu_enable_fixed(hwc, idx);
 525                 return;
 526         }
 527
 528         x86_pmu_enable_counter(hwc, idx);
 529 }
 530
 531 static void amd_pmu_enable_counter(struct hw_perf_counter *hwc, int idx)
 532 {
 533         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 534
 535         if (cpuc->enabled)
 536                 x86_pmu_enable_counter(hwc, idx);
 537         else
 538                 x86_pmu_disable_counter(hwc, idx);
 539 }
 540
 541 static int
 542 fixed_mode_idx(struct perf_counter *counter, struct hw_perf_counter *hwc)
 543 {
 544         unsigned int event;
 545
 546         if (!x86_pmu.num_counters_fixed)
 547                 return -1;
 548
 549         if (unlikely(hwc->nmi))
 550                 return -1;
 551
 552         event = hwc->config & ARCH_PERFMON_EVENT_MASK;
 553
 554         if (unlikely(event == x86_pmu.event_map(PERF_COUNT_INSTRUCTIONS)))
 555                 return X86_PMC_IDX_FIXED_INSTRUCTIONS;
 556         if (unlikely(event == x86_pmu.event_map(PERF_COUNT_CPU_CYCLES)))
 557                 return X86_PMC_IDX_FIXED_CPU_CYCLES;
 558         if (unlikely(event == x86_pmu.event_map(PERF_COUNT_BUS_CYCLES)))
 559                 return X86_PMC_IDX_FIXED_BUS_CYCLES;
 560
 561         return -1;
 562 }
 563
 564 /*
 565  * Find a PMC slot for the freshly enabled / scheduled in counter:
 566  */
 567 static int x86_pmu_enable(struct perf_counter *counter)
 568 {
 569         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 570         struct hw_perf_counter *hwc = &counter->hw;
 571         int idx;
 572
 573         idx = fixed_mode_idx(counter, hwc);
 574         if (idx >= 0) {
 575                 /*
 576                  * Try to get the fixed counter, if that is already taken
 577                  * then try to get a generic counter:
 578                  */
 579                 if (test_and_set_bit(idx, cpuc->used_mask))
 580                         goto try_generic;
 581
 582                 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 583                 /*
 584                  * We set it so that counter_base + idx in wrmsr/rdmsr maps to
 585                  * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
 586                  */
 587                 hwc->counter_base =
 588                         MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
 589                 hwc->idx = idx;
 590         } else {
 591                 idx = hwc->idx;
 592                 /* Try to get the previous generic counter again */
 593                 if (test_and_set_bit(idx, cpuc->used_mask)) {
 594 try_generic:
 595                         idx = find_first_zero_bit(cpuc->used_mask,
 596                                                   x86_pmu.num_counters);
 597                         if (idx == x86_pmu.num_counters)
 598                                 return -EAGAIN;
 599
 600                         set_bit(idx, cpuc->used_mask);
 601                         hwc->idx = idx;
 602                 }
 603                 hwc->config_base  = x86_pmu.eventsel;
 604                 hwc->counter_base = x86_pmu.perfctr;
 605         }
 606
 607         perf_counters_lapic_init(hwc->nmi);
 608
 609         x86_pmu.disable(hwc, idx);
 610
 611         cpuc->counters[idx] = counter;
 612         set_bit(idx, cpuc->active_mask);
 613
 614         x86_perf_counter_set_period(counter, hwc, idx);
 615         x86_pmu.enable(hwc, idx);
 616
 617         return 0;
 618 }
 619
 620 void perf_counter_print_debug(void)
 621 {
 622         u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
 623         struct cpu_hw_counters *cpuc;
 624         int cpu, idx;
 625
 626         if (!x86_pmu.num_counters)
 627                 return;
 628
 629         local_irq_disable();
 630
 631         cpu = smp_processor_id();
 632         cpuc = &per_cpu(cpu_hw_counters, cpu);
 633
 634         if (x86_pmu.version >= 2) {
 635                 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
 636                 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
 637                 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
 638                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
 639
 640                 pr_info("\n");
 641                 pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
 642                 pr_info("CPU#%d: status:     %016llx\n", cpu, status);
 643                 pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
 644                 pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
 645         }
 646         pr_info("CPU#%d: used:       %016llx\n", cpu, *(u64 *)cpuc->used_mask);
 647
 648         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 649                 rdmsrl(x86_pmu.eventsel + idx, pmc_ctrl);
 650                 rdmsrl(x86_pmu.perfctr  + idx, pmc_count);
 651
 652                 prev_left = per_cpu(prev_left[idx], cpu);
 653
 654                 pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
 655                         cpu, idx, pmc_ctrl);
 656                 pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
 657                         cpu, idx, pmc_count);
 658                 pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
 659                         cpu, idx, prev_left);
 660         }
 661         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
 662                 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
 663
 664                 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
 665                         cpu, idx, pmc_count);
 666         }
 667         local_irq_enable();
 668 }
 669
 670 static void x86_pmu_disable(struct perf_counter *counter)
 671 {
 672         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 673         struct hw_perf_counter *hwc = &counter->hw;
 674         int idx = hwc->idx;
 675
 676         /*
 677          * Must be done before we disable, otherwise the nmi handler
 678          * could reenable again:
 679          */
 680         clear_bit(idx, cpuc->active_mask);
 681         x86_pmu.disable(hwc, idx);
 682
 683         /*
 684          * Make sure the cleared pointer becomes visible before we
 685          * (potentially) free the counter:
 686          */
 687         barrier();
 688
 689         /*
 690          * Drain the remaining delta count out of a counter
 691          * that we are disabling:
 692          */
 693         x86_perf_counter_update(counter, hwc, idx);
 694         cpuc->counters[idx] = NULL;
 695         clear_bit(idx, cpuc->used_mask);
 696 }
 697
 698 /*
 699  * Save and restart an expired counter. Called by NMI contexts,
 700  * so it has to be careful about preempting normal counter ops:
 701  */
 702 static void intel_pmu_save_and_restart(struct perf_counter *counter)
 703 {
 704         struct hw_perf_counter *hwc = &counter->hw;
 705         int idx = hwc->idx;
 706
 707         x86_perf_counter_update(counter, hwc, idx);
 708         x86_perf_counter_set_period(counter, hwc, idx);
 709
 710         if (counter->state == PERF_COUNTER_STATE_ACTIVE)
 711                 intel_pmu_enable_counter(hwc, idx);
 712 }
 713
 714 /*
 715  * Maximum interrupt frequency of 100KHz per CPU
 716  */
 717 #define PERFMON_MAX_INTERRUPTS (100000/HZ)
 718
 719 /*
 720  * This handler is triggered by the local APIC, so the APIC IRQ handling
 721  * rules apply:
 722  */
 723 static int intel_pmu_handle_irq(struct pt_regs *regs, int nmi)
 724 {
 725         int bit, cpu = smp_processor_id();
 726         u64 ack, status;
 727         struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
 728         int ret = 0;
 729
 730         cpuc->throttle_ctrl = intel_pmu_save_disable_all();
 731
 732         status = intel_pmu_get_status();
 733         if (!status)
 734                 goto out;
 735
 736         ret = 1;
 737 again:
 738         inc_irq_stat(apic_perf_irqs);
 739         ack = status;
 740         for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
 741                 struct perf_counter *counter = cpuc->counters[bit];
 742
 743                 clear_bit(bit, (unsigned long *) &status);
 744                 if (!test_bit(bit, cpuc->active_mask))
 745                         continue;
 746
 747                 intel_pmu_save_and_restart(counter);
 748                 if (perf_counter_overflow(counter, nmi, regs, 0))
 749                         intel_pmu_disable_counter(&counter->hw, bit);
 750         }
 751
 752         intel_pmu_ack_status(ack);
 753
 754         /*
 755          * Repeat if there is more work to be done:
 756          */
 757         status = intel_pmu_get_status();
 758         if (status)
 759                 goto again;
 760 out:
 761         /*
 762          * Restore - do not reenable when global enable is off or throttled:
 763          */
 764         if (++cpuc->interrupts < PERFMON_MAX_INTERRUPTS)
 765                 intel_pmu_restore_all(cpuc->throttle_ctrl);
 766
 767         return ret;
 768 }
 769
 770 static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi)
 771 {
 772         int cpu = smp_processor_id();
 773         struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
 774         u64 val;
 775         int handled = 0;
 776         struct perf_counter *counter;
 777         struct hw_perf_counter *hwc;
 778         int idx;
 779
 780         ++cpuc->interrupts;
 781         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 782                 if (!test_bit(idx, cpuc->active_mask))
 783                         continue;
 784                 counter = cpuc->counters[idx];
 785                 hwc = &counter->hw;
 786                 val = x86_perf_counter_update(counter, hwc, idx);
 787                 if (val & (1ULL << (x86_pmu.counter_bits - 1)))
 788                         continue;
 789                 /* counter overflow */
 790                 x86_perf_counter_set_period(counter, hwc, idx);
 791                 handled = 1;
 792                 inc_irq_stat(apic_perf_irqs);
 793                 if (perf_counter_overflow(counter, nmi, regs, 0))
 794                         amd_pmu_disable_counter(hwc, idx);
 795                 else if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS)
 796                         /*
 797                          * do not reenable when throttled, but reload
 798                          * the register
 799                          */
 800                         amd_pmu_disable_counter(hwc, idx);
 801                 else if (counter->state == PERF_COUNTER_STATE_ACTIVE)
 802                         amd_pmu_enable_counter(hwc, idx);
 803         }
 804         return handled;
 805 }
 806
 807 void perf_counter_unthrottle(void)
 808 {
 809         struct cpu_hw_counters *cpuc;
 810
 811         if (!x86_pmu_initialized())
 812                 return;
 813
 814         cpuc = &__get_cpu_var(cpu_hw_counters);
 815         if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
 816                 if (printk_ratelimit())
 817                         printk(KERN_WARNING "PERFMON: max interrupts exceeded!\n");
 818                 hw_perf_restore(cpuc->throttle_ctrl);
 819         }
 820         cpuc->interrupts = 0;
 821 }
 822
 823 void smp_perf_counter_interrupt(struct pt_regs *regs)
 824 {
 825         irq_enter();
 826         apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
 827         ack_APIC_irq();
 828         x86_pmu.handle_irq(regs, 0);
 829         irq_exit();
 830 }
 831
 832 void smp_perf_pending_interrupt(struct pt_regs *regs)
 833 {
 834         irq_enter();
 835         ack_APIC_irq();
 836         inc_irq_stat(apic_pending_irqs);
 837         perf_counter_do_pending();
 838         irq_exit();
 839 }
 840
 841 void set_perf_counter_pending(void)
 842 {
 843         apic->send_IPI_self(LOCAL_PENDING_VECTOR);
 844 }
 845
 846 void perf_counters_lapic_init(int nmi)
 847 {
 848         u32 apic_val;
 849
 850         if (!x86_pmu_initialized())
 851                 return;
 852
 853         /*
 854          * Enable the performance counter vector in the APIC LVT:
 855          */
 856         apic_val = apic_read(APIC_LVTERR);
 857
 858         apic_write(APIC_LVTERR, apic_val | APIC_LVT_MASKED);
 859         if (nmi)
 860                 apic_write(APIC_LVTPC, APIC_DM_NMI);
 861         else
 862                 apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
 863         apic_write(APIC_LVTERR, apic_val);
 864 }
 865
 866 static int __kprobes
 867 perf_counter_nmi_handler(struct notifier_block *self,
 868                          unsigned long cmd, void *__args)
 869 {
 870         struct die_args *args = __args;
 871         struct pt_regs *regs;
 872         int ret;
 873
 874         if (!atomic_read(&active_counters))
 875                 return NOTIFY_DONE;
 876
 877         switch (cmd) {
 878         case DIE_NMI:
 879         case DIE_NMI_IPI:
 880                 break;
 881
 882         default:
 883                 return NOTIFY_DONE;
 884         }
 885
 886         regs = args->regs;
 887
 888         apic_write(APIC_LVTPC, APIC_DM_NMI);
 889         ret = x86_pmu.handle_irq(regs, 1);
 890
 891         return ret ? NOTIFY_STOP : NOTIFY_OK;
 892 }
 893
 894 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
 895         .notifier_call          = perf_counter_nmi_handler,
 896         .next                   = NULL,
 897         .priority               = 1
 898 };
 899
 900 static struct x86_pmu intel_pmu = {
 901         .name                   = "Intel",
 902         .handle_irq             = intel_pmu_handle_irq,
 903         .save_disable_all       = intel_pmu_save_disable_all,
 904         .restore_all            = intel_pmu_restore_all,
 905         .enable                 = intel_pmu_enable_counter,
 906         .disable                = intel_pmu_disable_counter,
 907         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
 908         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
 909         .event_map              = intel_pmu_event_map,
 910         .raw_event              = intel_pmu_raw_event,
 911         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
 912         /*
 913          * Intel PMCs cannot be accessed sanely above 32 bit width,
 914          * so we install an artificial 1<<31 period regardless of
 915          * the generic counter period:
 916          */
 917         .max_period             = (1ULL << 31) - 1,
 918 };
 919
 920 static struct x86_pmu amd_pmu = {
 921         .name                   = "AMD",
 922         .handle_irq             = amd_pmu_handle_irq,
 923         .save_disable_all       = amd_pmu_save_disable_all,
 924         .restore_all            = amd_pmu_restore_all,
 925         .enable                 = amd_pmu_enable_counter,
 926         .disable                = amd_pmu_disable_counter,
 927         .eventsel               = MSR_K7_EVNTSEL0,
 928         .perfctr                = MSR_K7_PERFCTR0,
 929         .event_map              = amd_pmu_event_map,
 930         .raw_event              = amd_pmu_raw_event,
 931         .max_events             = ARRAY_SIZE(amd_perfmon_event_map),
 932         .num_counters           = 4,
 933         .counter_bits           = 48,
 934         .counter_mask           = (1ULL << 48) - 1,
 935         /* use highest bit to detect overflow */
 936         .max_period             = (1ULL << 47) - 1,
 937 };
 938
 939 static int intel_pmu_init(void)
 940 {
 941         union cpuid10_edx edx;
 942         union cpuid10_eax eax;
 943         unsigned int unused;
 944         unsigned int ebx;
 945         int version;
 946
 947         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 948                 return -ENODEV;
 949
 950         /*
 951          * Check whether the Architectural PerfMon supports
 952          * Branch Misses Retired Event or not.
 953          */
 954         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
 955         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
 956                 return -ENODEV;
 957
 958         version = eax.split.version_id;
 959         if (version < 2)
 960                 return -ENODEV;
 961
 962         x86_pmu = intel_pmu;
 963         x86_pmu.version = version;
 964         x86_pmu.num_counters = eax.split.num_counters;
 965
 966         /*
 967          * Quirk: v2 perfmon does not report fixed-purpose counters, so
 968          * assume at least 3 counters:
 969          */
 970         x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
 971
 972         x86_pmu.counter_bits = eax.split.bit_width;
 973         x86_pmu.counter_mask = (1ULL << eax.split.bit_width) - 1;
 974
 975         return 0;
 976 }
 977
 978 static int amd_pmu_init(void)
 979 {
 980         x86_pmu = amd_pmu;
 981         return 0;
 982 }
 983
 984 void __init init_hw_perf_counters(void)
 985 {
 986         int err;
 987
 988         switch (boot_cpu_data.x86_vendor) {
 989         case X86_VENDOR_INTEL:
 990                 err = intel_pmu_init();
 991                 break;
 992         case X86_VENDOR_AMD:
 993                 err = amd_pmu_init();
 994                 break;
 995         default:
 996                 return;
 997         }
 998         if (err != 0)
 999                 return;
1000
1001         pr_info("%s Performance Monitoring support detected.\n", x86_pmu.name);
1002         pr_info("... version:         %d\n", x86_pmu.version);
1003         pr_info("... bit width:       %d\n", x86_pmu.counter_bits);
1004
1005         pr_info("... num counters:    %d\n", x86_pmu.num_counters);
1006         if (x86_pmu.num_counters > X86_PMC_MAX_GENERIC) {
1007                 x86_pmu.num_counters = X86_PMC_MAX_GENERIC;
1008                 WARN(1, KERN_ERR "hw perf counters %d > max(%d), clipping!",
1009                      x86_pmu.num_counters, X86_PMC_MAX_GENERIC);
1010         }
1011         perf_counter_mask = (1 << x86_pmu.num_counters) - 1;
1012         perf_max_counters = x86_pmu.num_counters;
1013
1014         pr_info("... value mask:      %016Lx\n", x86_pmu.counter_mask);
1015         pr_info("... max period:      %016Lx\n", x86_pmu.max_period);
1016
1017         if (x86_pmu.num_counters_fixed > X86_PMC_MAX_FIXED) {
1018                 x86_pmu.num_counters_fixed = X86_PMC_MAX_FIXED;
1019                 WARN(1, KERN_ERR "hw perf counters fixed %d > max(%d), clipping!",
1020                      x86_pmu.num_counters_fixed, X86_PMC_MAX_FIXED);
1021         }
1022         pr_info("... fixed counters:  %d\n", x86_pmu.num_counters_fixed);
1023
1024         perf_counter_mask |=
1025                 ((1LL << x86_pmu.num_counters_fixed)-1) << X86_PMC_IDX_FIXED;
1026
1027         pr_info("... counter mask:    %016Lx\n", perf_counter_mask);
1028
1029         perf_counters_lapic_init(0);
1030         register_die_notifier(&perf_counter_nmi_notifier);
1031 }
1032
1033 static inline void x86_pmu_read(struct perf_counter *counter)
1034 {
1035         x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
1036 }
1037
1038 static const struct pmu pmu = {
1039         .enable         = x86_pmu_enable,
1040         .disable        = x86_pmu_disable,
1041         .read           = x86_pmu_read,
1042 };
1043
1044 const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
1045 {
1046         int err;
1047
1048         err = __hw_perf_counter_init(counter);
1049         if (err)
1050                 return ERR_PTR(err);
1051
1052         return &pmu;
1053 }
1054
1055 /*
1056  * callchain support
1057  */
1058
1059 static inline
1060 void callchain_store(struct perf_callchain_entry *entry, unsigned long ip)
1061 {
1062         if (entry->nr < MAX_STACK_DEPTH)
1063                 entry->ip[entry->nr++] = ip;
1064 }
1065
1066 static DEFINE_PER_CPU(struct perf_callchain_entry, irq_entry);
1067 static DEFINE_PER_CPU(struct perf_callchain_entry, nmi_entry);
1068
1069
1070 static void
1071 backtrace_warning_symbol(void *data, char *msg, unsigned long symbol)
1072 {
1073         /* Ignore warnings */
1074 }
1075
1076 static void backtrace_warning(void *data, char *msg)
1077 {
1078         /* Ignore warnings */
1079 }
1080
1081 static int backtrace_stack(void *data, char *name)
1082 {
1083         /* Don't bother with IRQ stacks for now */
1084         return -1;
1085 }
1086
1087 static void backtrace_address(void *data, unsigned long addr, int reliable)
1088 {
1089         struct perf_callchain_entry *entry = data;
1090
1091         if (reliable)
1092                 callchain_store(entry, addr);
1093 }
1094
1095 static const struct stacktrace_ops backtrace_ops = {
1096         .warning                = backtrace_warning,
1097         .warning_symbol         = backtrace_warning_symbol,
1098         .stack                  = backtrace_stack,
1099         .address                = backtrace_address,
1100 };
1101
1102 static void
1103 perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
1104 {
1105         unsigned long bp;
1106         char *stack;
1107         int nr = entry->nr;
1108
1109         callchain_store(entry, instruction_pointer(regs));
1110
1111         stack = ((char *)regs + sizeof(struct pt_regs));
1112 #ifdef CONFIG_FRAME_POINTER
1113         bp = frame_pointer(regs);
1114 #else
1115         bp = 0;
1116 #endif
1117
1118         dump_trace(NULL, regs, (void *)stack, bp, &backtrace_ops, entry);
1119
1120         entry->kernel = entry->nr - nr;
1121 }
1122
1123
1124 struct stack_frame {
1125         const void __user       *next_fp;
1126         unsigned long           return_address;
1127 };
1128
1129 static int copy_stack_frame(const void __user *fp, struct stack_frame *frame)
1130 {
1131         int ret;
1132
1133         if (!access_ok(VERIFY_READ, fp, sizeof(*frame)))
1134                 return 0;
1135
1136         ret = 1;
1137         pagefault_disable();
1138         if (__copy_from_user_inatomic(frame, fp, sizeof(*frame)))
1139                 ret = 0;
1140         pagefault_enable();
1141
1142         return ret;
1143 }
1144
1145 static void
1146 perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry)
1147 {
1148         struct stack_frame frame;
1149         const void __user *fp;
1150         int nr = entry->nr;
1151
1152         regs = (struct pt_regs *)current->thread.sp0 - 1;
1153         fp   = (void __user *)regs->bp;
1154
1155         callchain_store(entry, regs->ip);
1156
1157         while (entry->nr < MAX_STACK_DEPTH) {
1158                 frame.next_fp        = NULL;
1159                 frame.return_address = 0;
1160
1161                 if (!copy_stack_frame(fp, &frame))
1162                         break;
1163
1164                 if ((unsigned long)fp < user_stack_pointer(regs))
1165                         break;
1166
1167                 callchain_store(entry, frame.return_address);
1168                 fp = frame.next_fp;
1169         }
1170
1171         entry->user = entry->nr - nr;
1172 }
1173
1174 static void
1175 perf_do_callchain(struct pt_regs *regs, struct perf_callchain_entry *entry)
1176 {
1177         int is_user;
1178
1179         if (!regs)
1180                 return;
1181
1182         is_user = user_mode(regs);
1183
1184         if (!current || current->pid == 0)
1185                 return;
1186
1187         if (is_user && current->state != TASK_RUNNING)
1188                 return;
1189
1190         if (!is_user)
1191                 perf_callchain_kernel(regs, entry);
1192
1193         if (current->mm)
1194                 perf_callchain_user(regs, entry);
1195 }
1196
1197 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1198 {
1199         struct perf_callchain_entry *entry;
1200
1201         if (in_nmi())
1202                 entry = &__get_cpu_var(nmi_entry);
1203         else
1204                 entry = &__get_cpu_var(irq_entry);
1205
1206         entry->nr = 0;
1207         entry->hv = 0;
1208         entry->kernel = 0;
1209         entry->user = 0;
1210
1211         perf_do_callchain(regs, entry);
1212
1213         return entry;
1214 }