arch/sparc/kernel/perf_counter.c

   1 /* Performance counter support for sparc64.
   2  *
   3  * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
   4  *
   5  * This code is based almost entirely upon the x86 perf counter
   6  * code, which is:
   7  *
   8  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   9  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
  10  *  Copyright (C) 2009 Jaswinder Singh Rajput
  11  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
  12  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  13  */
  14
  15 #include <linux/perf_counter.h>
  16 #include <linux/kprobes.h>
  17 #include <linux/kernel.h>
  18 #include <linux/kdebug.h>
  19 #include <linux/mutex.h>
  20
  21 #include <asm/cpudata.h>
  22 #include <asm/atomic.h>
  23 #include <asm/nmi.h>
  24 #include <asm/pcr.h>
  25
  26 /* Sparc64 chips have two performance counters, 32-bits each, with
  27  * overflow interrupts generated on transition from 0xffffffff to 0.
  28  * The counters are accessed in one go using a 64-bit register.
  29  *
  30  * Both counters are controlled using a single control register.  The
  31  * only way to stop all sampling is to clear all of the context (user,
  32  * supervisor, hypervisor) sampling enable bits.  But these bits apply
  33  * to both counters, thus the two counters can't be enabled/disabled
  34  * individually.
  35  *
  36  * The control register has two event fields, one for each of the two
  37  * counters.  It's thus nearly impossible to have one counter going
  38  * while keeping the other one stopped.  Therefore it is possible to
  39  * get overflow interrupts for counters not currently "in use" and
  40  * that condition must be checked in the overflow interrupt handler.
  41  *
  42  * So we use a hack, in that we program inactive counters with the
  43  * "sw_count0" and "sw_count1" events.  These count how many times
  44  * the instruction "sethi %hi(0xfc000), %g0" is executed.  It's an
  45  * unusual way to encode a NOP and therefore will not trigger in
  46  * normal code.
  47  */
  48
  49 #define MAX_HWCOUNTERS                  2
  50 #define MAX_PERIOD                      ((1UL << 32) - 1)
  51
  52 #define PIC_UPPER_INDEX                 0
  53 #define PIC_LOWER_INDEX                 1
  54
  55 struct cpu_hw_counters {
  56         struct perf_counter     *counters[MAX_HWCOUNTERS];
  57         unsigned long           used_mask[BITS_TO_LONGS(MAX_HWCOUNTERS)];
  58         unsigned long           active_mask[BITS_TO_LONGS(MAX_HWCOUNTERS)];
  59         int enabled;
  60 };
  61 DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = { .enabled = 1, };
  62
  63 struct perf_event_map {
  64         u16     encoding;
  65         u8      pic_mask;
  66 #define PIC_NONE        0x00
  67 #define PIC_UPPER       0x01
  68 #define PIC_LOWER       0x02
  69 };
  70
  71 struct sparc_pmu {
  72         const struct perf_event_map     *(*event_map)(int);
  73         int                             max_events;
  74         int                             upper_shift;
  75         int                             lower_shift;
  76         int                             event_mask;
  77         int                             hv_bit;
  78         int                             irq_bit;
  79         int                             upper_nop;
  80         int                             lower_nop;
  81 };
  82
  83 static const struct perf_event_map ultra3i_perfmon_event_map[] = {
  84         [PERF_COUNT_HW_CPU_CYCLES] = { 0x0000, PIC_UPPER | PIC_LOWER },
  85         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x0001, PIC_UPPER | PIC_LOWER },
  86         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0009, PIC_LOWER },
  87         [PERF_COUNT_HW_CACHE_MISSES] = { 0x0009, PIC_UPPER },
  88 };
  89
  90 static const struct perf_event_map *ultra3i_event_map(int event)
  91 {
  92         return &ultra3i_perfmon_event_map[event];
  93 }
  94
  95 static const struct sparc_pmu ultra3i_pmu = {
  96         .event_map      = ultra3i_event_map,
  97         .max_events     = ARRAY_SIZE(ultra3i_perfmon_event_map),
  98         .upper_shift    = 11,
  99         .lower_shift    = 4,
 100         .event_mask     = 0x3f,
 101         .upper_nop      = 0x1c,
 102         .lower_nop      = 0x14,
 103 };
 104
 105 static const struct perf_event_map niagara2_perfmon_event_map[] = {
 106         [PERF_COUNT_HW_CPU_CYCLES] = { 0x02ff, PIC_UPPER | PIC_LOWER },
 107         [PERF_COUNT_HW_INSTRUCTIONS] = { 0x02ff, PIC_UPPER | PIC_LOWER },
 108         [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x0208, PIC_UPPER | PIC_LOWER },
 109         [PERF_COUNT_HW_CACHE_MISSES] = { 0x0302, PIC_UPPER | PIC_LOWER },
 110         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x0201, PIC_UPPER | PIC_LOWER },
 111         [PERF_COUNT_HW_BRANCH_MISSES] = { 0x0202, PIC_UPPER | PIC_LOWER },
 112 };
 113
 114 static const struct perf_event_map *niagara2_event_map(int event)
 115 {
 116         return &niagara2_perfmon_event_map[event];
 117 }
 118
 119 static const struct sparc_pmu niagara2_pmu = {
 120         .event_map      = niagara2_event_map,
 121         .max_events     = ARRAY_SIZE(niagara2_perfmon_event_map),
 122         .upper_shift    = 19,
 123         .lower_shift    = 6,
 124         .event_mask     = 0xfff,
 125         .hv_bit         = 0x8,
 126         .irq_bit        = 0x03,
 127         .upper_nop      = 0x220,
 128         .lower_nop      = 0x220,
 129 };
 130
 131 static const struct sparc_pmu *sparc_pmu __read_mostly;
 132
 133 static u64 event_encoding(u64 event, int idx)
 134 {
 135         if (idx == PIC_UPPER_INDEX)
 136                 event <<= sparc_pmu->upper_shift;
 137         else
 138                 event <<= sparc_pmu->lower_shift;
 139         return event;
 140 }
 141
 142 static u64 mask_for_index(int idx)
 143 {
 144         return event_encoding(sparc_pmu->event_mask, idx);
 145 }
 146
 147 static u64 nop_for_index(int idx)
 148 {
 149         return event_encoding(idx == PIC_UPPER_INDEX ?
 150                               sparc_pmu->upper_nop :
 151                               sparc_pmu->lower_nop, idx);
 152 }
 153
 154 static inline void sparc_pmu_enable_counter(struct hw_perf_counter *hwc,
 155                                             int idx)
 156 {
 157         u64 val, mask = mask_for_index(idx);
 158
 159         val = pcr_ops->read();
 160         pcr_ops->write((val & ~mask) | hwc->config);
 161 }
 162
 163 static inline void sparc_pmu_disable_counter(struct hw_perf_counter *hwc,
 164                                              int idx)
 165 {
 166         u64 mask = mask_for_index(idx);
 167         u64 nop = nop_for_index(idx);
 168         u64 val = pcr_ops->read();
 169
 170         pcr_ops->write((val & ~mask) | nop);
 171 }
 172
 173 void hw_perf_enable(void)
 174 {
 175         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 176         u64 val;
 177         int i;
 178
 179         if (cpuc->enabled)
 180                 return;
 181
 182         cpuc->enabled = 1;
 183         barrier();
 184
 185         val = pcr_ops->read();
 186
 187         for (i = 0; i < MAX_HWCOUNTERS; i++) {
 188                 struct perf_counter *cp = cpuc->counters[i];
 189                 struct hw_perf_counter *hwc;
 190
 191                 if (!cp)
 192                         continue;
 193                 hwc = &cp->hw;
 194                 val |= hwc->config_base;
 195         }
 196
 197         pcr_ops->write(val);
 198 }
 199
 200 void hw_perf_disable(void)
 201 {
 202         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 203         u64 val;
 204
 205         if (!cpuc->enabled)
 206                 return;
 207
 208         cpuc->enabled = 0;
 209
 210         val = pcr_ops->read();
 211         val &= ~(PCR_UTRACE | PCR_STRACE |
 212                  sparc_pmu->hv_bit | sparc_pmu->irq_bit);
 213         pcr_ops->write(val);
 214 }
 215
 216 static u32 read_pmc(int idx)
 217 {
 218         u64 val;
 219
 220         read_pic(val);
 221         if (idx == PIC_UPPER_INDEX)
 222                 val >>= 32;
 223
 224         return val & 0xffffffff;
 225 }
 226
 227 static void write_pmc(int idx, u64 val)
 228 {
 229         u64 shift, mask, pic;
 230
 231         shift = 0;
 232         if (idx == PIC_UPPER_INDEX)
 233                 shift = 32;
 234
 235         mask = ((u64) 0xffffffff) << shift;
 236         val <<= shift;
 237
 238         read_pic(pic);
 239         pic &= ~mask;
 240         pic |= val;
 241         write_pic(pic);
 242 }
 243
 244 static int sparc_perf_counter_set_period(struct perf_counter *counter,
 245                                          struct hw_perf_counter *hwc, int idx)
 246 {
 247         s64 left = atomic64_read(&hwc->period_left);
 248         s64 period = hwc->sample_period;
 249         int ret = 0;
 250
 251         if (unlikely(left <= -period)) {
 252                 left = period;
 253                 atomic64_set(&hwc->period_left, left);
 254                 hwc->last_period = period;
 255                 ret = 1;
 256         }
 257
 258         if (unlikely(left <= 0)) {
 259                 left += period;
 260                 atomic64_set(&hwc->period_left, left);
 261                 hwc->last_period = period;
 262                 ret = 1;
 263         }
 264         if (left > MAX_PERIOD)
 265                 left = MAX_PERIOD;
 266
 267         atomic64_set(&hwc->prev_count, (u64)-left);
 268
 269         write_pmc(idx, (u64)(-left) & 0xffffffff);
 270
 271         perf_counter_update_userpage(counter);
 272
 273         return ret;
 274 }
 275
 276 static int sparc_pmu_enable(struct perf_counter *counter)
 277 {
 278         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 279         struct hw_perf_counter *hwc = &counter->hw;
 280         int idx = hwc->idx;
 281
 282         if (test_and_set_bit(idx, cpuc->used_mask))
 283                 return -EAGAIN;
 284
 285         sparc_pmu_disable_counter(hwc, idx);
 286
 287         cpuc->counters[idx] = counter;
 288         set_bit(idx, cpuc->active_mask);
 289
 290         sparc_perf_counter_set_period(counter, hwc, idx);
 291         sparc_pmu_enable_counter(hwc, idx);
 292         perf_counter_update_userpage(counter);
 293         return 0;
 294 }
 295
 296 static u64 sparc_perf_counter_update(struct perf_counter *counter,
 297                                      struct hw_perf_counter *hwc, int idx)
 298 {
 299         int shift = 64 - 32;
 300         u64 prev_raw_count, new_raw_count;
 301         s64 delta;
 302
 303 again:
 304         prev_raw_count = atomic64_read(&hwc->prev_count);
 305         new_raw_count = read_pmc(idx);
 306
 307         if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
 308                              new_raw_count) != prev_raw_count)
 309                 goto again;
 310
 311         delta = (new_raw_count << shift) - (prev_raw_count << shift);
 312         delta >>= shift;
 313
 314         atomic64_add(delta, &counter->count);
 315         atomic64_sub(delta, &hwc->period_left);
 316
 317         return new_raw_count;
 318 }
 319
 320 static void sparc_pmu_disable(struct perf_counter *counter)
 321 {
 322         struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 323         struct hw_perf_counter *hwc = &counter->hw;
 324         int idx = hwc->idx;
 325
 326         clear_bit(idx, cpuc->active_mask);
 327         sparc_pmu_disable_counter(hwc, idx);
 328
 329         barrier();
 330
 331         sparc_perf_counter_update(counter, hwc, idx);
 332         cpuc->counters[idx] = NULL;
 333         clear_bit(idx, cpuc->used_mask);
 334
 335         perf_counter_update_userpage(counter);
 336 }
 337
 338 static void sparc_pmu_read(struct perf_counter *counter)
 339 {
 340         struct hw_perf_counter *hwc = &counter->hw;
 341         sparc_perf_counter_update(counter, hwc, hwc->idx);
 342 }
 343
 344 static void sparc_pmu_unthrottle(struct perf_counter *counter)
 345 {
 346         struct hw_perf_counter *hwc = &counter->hw;
 347         sparc_pmu_enable_counter(hwc, hwc->idx);
 348 }
 349
 350 static atomic_t active_counters = ATOMIC_INIT(0);
 351 static DEFINE_MUTEX(pmc_grab_mutex);
 352
 353 void perf_counter_grab_pmc(void)
 354 {
 355         if (atomic_inc_not_zero(&active_counters))
 356                 return;
 357
 358         mutex_lock(&pmc_grab_mutex);
 359         if (atomic_read(&active_counters) == 0) {
 360                 if (atomic_read(&nmi_active) > 0) {
 361                         on_each_cpu(stop_nmi_watchdog, NULL, 1);
 362                         BUG_ON(atomic_read(&nmi_active) != 0);
 363                 }
 364                 atomic_inc(&active_counters);
 365         }
 366         mutex_unlock(&pmc_grab_mutex);
 367 }
 368
 369 void perf_counter_release_pmc(void)
 370 {
 371         if (atomic_dec_and_mutex_lock(&active_counters, &pmc_grab_mutex)) {
 372                 if (atomic_read(&nmi_active) == 0)
 373                         on_each_cpu(start_nmi_watchdog, NULL, 1);
 374                 mutex_unlock(&pmc_grab_mutex);
 375         }
 376 }
 377
 378 static void hw_perf_counter_destroy(struct perf_counter *counter)
 379 {
 380         perf_counter_release_pmc();
 381 }
 382
 383 static int __hw_perf_counter_init(struct perf_counter *counter)
 384 {
 385         struct perf_counter_attr *attr = &counter->attr;
 386         struct hw_perf_counter *hwc = &counter->hw;
 387         const struct perf_event_map *pmap;
 388         u64 enc;
 389
 390         if (atomic_read(&nmi_active) < 0)
 391                 return -ENODEV;
 392
 393         if (attr->type != PERF_TYPE_HARDWARE)
 394                 return -EOPNOTSUPP;
 395
 396         if (attr->config >= sparc_pmu->max_events)
 397                 return -EINVAL;
 398
 399         perf_counter_grab_pmc();
 400         counter->destroy = hw_perf_counter_destroy;
 401
 402         /* We save the enable bits in the config_base.  So to
 403          * turn off sampling just write 'config', and to enable
 404          * things write 'config | config_base'.
 405          */
 406         hwc->config_base = sparc_pmu->irq_bit;
 407         if (!attr->exclude_user)
 408                 hwc->config_base |= PCR_UTRACE;
 409         if (!attr->exclude_kernel)
 410                 hwc->config_base |= PCR_STRACE;
 411         if (!attr->exclude_hv)
 412                 hwc->config_base |= sparc_pmu->hv_bit;
 413
 414         if (!hwc->sample_period) {
 415                 hwc->sample_period = MAX_PERIOD;
 416                 hwc->last_period = hwc->sample_period;
 417                 atomic64_set(&hwc->period_left, hwc->sample_period);
 418         }
 419
 420         pmap = sparc_pmu->event_map(attr->config);
 421
 422         enc = pmap->encoding;
 423         if (pmap->pic_mask & PIC_UPPER) {
 424                 hwc->idx = PIC_UPPER_INDEX;
 425                 enc <<= sparc_pmu->upper_shift;
 426         } else {
 427                 hwc->idx = PIC_LOWER_INDEX;
 428                 enc <<= sparc_pmu->lower_shift;
 429         }
 430
 431         hwc->config |= enc;
 432         return 0;
 433 }
 434
 435 static const struct pmu pmu = {
 436         .enable         = sparc_pmu_enable,
 437         .disable        = sparc_pmu_disable,
 438         .read           = sparc_pmu_read,
 439         .unthrottle     = sparc_pmu_unthrottle,
 440 };
 441
 442 const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
 443 {
 444         int err = __hw_perf_counter_init(counter);
 445
 446         if (err)
 447                 return ERR_PTR(err);
 448         return &pmu;
 449 }
 450
 451 void perf_counter_print_debug(void)
 452 {
 453         unsigned long flags;
 454         u64 pcr, pic;
 455         int cpu;
 456
 457         if (!sparc_pmu)
 458                 return;
 459
 460         local_irq_save(flags);
 461
 462         cpu = smp_processor_id();
 463
 464         pcr = pcr_ops->read();
 465         read_pic(pic);
 466
 467         pr_info("\n");
 468         pr_info("CPU#%d: PCR[%016llx] PIC[%016llx]\n",
 469                 cpu, pcr, pic);
 470
 471         local_irq_restore(flags);
 472 }
 473
 474 static int __kprobes perf_counter_nmi_handler(struct notifier_block *self,
 475                                               unsigned long cmd, void *__args)
 476 {
 477         struct die_args *args = __args;
 478         struct perf_sample_data data;
 479         struct cpu_hw_counters *cpuc;
 480         struct pt_regs *regs;
 481         int idx;
 482
 483         if (!atomic_read(&active_counters))
 484                 return NOTIFY_DONE;
 485
 486         switch (cmd) {
 487         case DIE_NMI:
 488                 break;
 489
 490         default:
 491                 return NOTIFY_DONE;
 492         }
 493
 494         regs = args->regs;
 495
 496         data.addr = 0;
 497
 498         cpuc = &__get_cpu_var(cpu_hw_counters);
 499         for (idx = 0; idx < MAX_HWCOUNTERS; idx++) {
 500                 struct perf_counter *counter = cpuc->counters[idx];
 501                 struct hw_perf_counter *hwc;
 502                 u64 val;
 503
 504                 if (!test_bit(idx, cpuc->active_mask))
 505                         continue;
 506                 hwc = &counter->hw;
 507                 val = sparc_perf_counter_update(counter, hwc, idx);
 508                 if (val & (1ULL << 31))
 509                         continue;
 510
 511                 data.period = counter->hw.last_period;
 512                 if (!sparc_perf_counter_set_period(counter, hwc, idx))
 513                         continue;
 514
 515                 if (perf_counter_overflow(counter, 1, &data, regs))
 516                         sparc_pmu_disable_counter(hwc, idx);
 517         }
 518
 519         return NOTIFY_STOP;
 520 }
 521
 522 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
 523         .notifier_call          = perf_counter_nmi_handler,
 524 };
 525
 526 static bool __init supported_pmu(void)
 527 {
 528         if (!strcmp(sparc_pmu_type, "ultra3i")) {
 529                 sparc_pmu = &ultra3i_pmu;
 530                 return true;
 531         }
 532         if (!strcmp(sparc_pmu_type, "niagara2")) {
 533                 sparc_pmu = &niagara2_pmu;
 534                 return true;
 535         }
 536         return false;
 537 }
 538
 539 void __init init_hw_perf_counters(void)
 540 {
 541         pr_info("Performance counters: ");
 542
 543         if (!supported_pmu()) {
 544                 pr_cont("No support for PMU type '%s'\n", sparc_pmu_type);
 545                 return;
 546         }
 547
 548         pr_cont("Supported PMU type is '%s'\n", sparc_pmu_type);
 549
 550         /* All sparc64 PMUs currently have 2 counters.  But this simple
 551          * driver only supports one active counter at a time.
 552          */
 553         perf_max_counters = 1;
 554
 555         register_die_notifier(&perf_counter_nmi_notifier);
 556 }