git.oblomov.eu Git - linux-2.6/blob - arch/powerpc/kernel/perf_counter.c

   1 /*
   2  * Performance counter support - powerpc architecture code
   3  *
   4  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/perf_counter.h>
  14 #include <linux/percpu.h>
  15 #include <linux/hardirq.h>
  16 #include <asm/reg.h>
  17 #include <asm/pmc.h>
  18 #include <asm/machdep.h>
  19 #include <asm/firmware.h>
  20
  21 struct cpu_hw_counters {
  22         int n_counters;
  23         int n_percpu;
  24         int disabled;
  25         int n_added;
  26         struct perf_counter *counter[MAX_HWCOUNTERS];
  27         unsigned int events[MAX_HWCOUNTERS];
  28         u64 mmcr[3];
  29         u8 pmcs_enabled;
  30 };
  31 DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters);
  32
  33 struct power_pmu *ppmu;
  34
  35 /*
  36  * Normally, to ignore kernel events we set the FCS (freeze counters
  37  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
  38  * hypervisor bit set in the MSR, or if we are running on a processor
  39  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
  40  * then we need to use the FCHV bit to ignore kernel events.
  41  */
  42 static unsigned int freeze_counters_kernel = MMCR0_FCS;
  43
  44 void perf_counter_print_debug(void)
  45 {
  46 }
  47
  48 /*
  49  * Read one performance monitor counter (PMC).
  50  */
  51 static unsigned long read_pmc(int idx)
  52 {
  53         unsigned long val;
  54
  55         switch (idx) {
  56         case 1:
  57                 val = mfspr(SPRN_PMC1);
  58                 break;
  59         case 2:
  60                 val = mfspr(SPRN_PMC2);
  61                 break;
  62         case 3:
  63                 val = mfspr(SPRN_PMC3);
  64                 break;
  65         case 4:
  66                 val = mfspr(SPRN_PMC4);
  67                 break;
  68         case 5:
  69                 val = mfspr(SPRN_PMC5);
  70                 break;
  71         case 6:
  72                 val = mfspr(SPRN_PMC6);
  73                 break;
  74         case 7:
  75                 val = mfspr(SPRN_PMC7);
  76                 break;
  77         case 8:
  78                 val = mfspr(SPRN_PMC8);
  79                 break;
  80         default:
  81                 printk(KERN_ERR "oops trying to read PMC%d\n", idx);
  82                 val = 0;
  83         }
  84         return val;
  85 }
  86
  87 /*
  88  * Write one PMC.
  89  */
  90 static void write_pmc(int idx, unsigned long val)
  91 {
  92         switch (idx) {
  93         case 1:
  94                 mtspr(SPRN_PMC1, val);
  95                 break;
  96         case 2:
  97                 mtspr(SPRN_PMC2, val);
  98                 break;
  99         case 3:
 100                 mtspr(SPRN_PMC3, val);
 101                 break;
 102         case 4:
 103                 mtspr(SPRN_PMC4, val);
 104                 break;
 105         case 5:
 106                 mtspr(SPRN_PMC5, val);
 107                 break;
 108         case 6:
 109                 mtspr(SPRN_PMC6, val);
 110                 break;
 111         case 7:
 112                 mtspr(SPRN_PMC7, val);
 113                 break;
 114         case 8:
 115                 mtspr(SPRN_PMC8, val);
 116                 break;
 117         default:
 118                 printk(KERN_ERR "oops trying to write PMC%d\n", idx);
 119         }
 120 }
 121
 122 /*
 123  * Check if a set of events can all go on the PMU at once.
 124  * If they can't, this will look at alternative codes for the events
 125  * and see if any combination of alternative codes is feasible.
 126  * The feasible set is returned in event[].
 127  */
 128 static int power_check_constraints(unsigned int event[], int n_ev)
 129 {
 130         u64 mask, value, nv;
 131         unsigned int alternatives[MAX_HWCOUNTERS][MAX_EVENT_ALTERNATIVES];
 132         u64 amasks[MAX_HWCOUNTERS][MAX_EVENT_ALTERNATIVES];
 133         u64 avalues[MAX_HWCOUNTERS][MAX_EVENT_ALTERNATIVES];
 134         u64 smasks[MAX_HWCOUNTERS], svalues[MAX_HWCOUNTERS];
 135         int n_alt[MAX_HWCOUNTERS], choice[MAX_HWCOUNTERS];
 136         int i, j;
 137         u64 addf = ppmu->add_fields;
 138         u64 tadd = ppmu->test_adder;
 139
 140         if (n_ev > ppmu->n_counter)
 141                 return -1;
 142
 143         /* First see if the events will go on as-is */
 144         for (i = 0; i < n_ev; ++i) {
 145                 alternatives[i][0] = event[i];
 146                 if (ppmu->get_constraint(event[i], &amasks[i][0],
 147                                          &avalues[i][0]))
 148                         return -1;
 149                 choice[i] = 0;
 150         }
 151         value = mask = 0;
 152         for (i = 0; i < n_ev; ++i) {
 153                 nv = (value | avalues[i][0]) + (value & avalues[i][0] & addf);
 154                 if ((((nv + tadd) ^ value) & mask) != 0 ||
 155                     (((nv + tadd) ^ avalues[i][0]) & amasks[i][0]) != 0)
 156                         break;
 157                 value = nv;
 158                 mask |= amasks[i][0];
 159         }
 160         if (i == n_ev)
 161                 return 0;       /* all OK */
 162
 163         /* doesn't work, gather alternatives... */
 164         if (!ppmu->get_alternatives)
 165                 return -1;
 166         for (i = 0; i < n_ev; ++i) {
 167                 n_alt[i] = ppmu->get_alternatives(event[i], alternatives[i]);
 168                 for (j = 1; j < n_alt[i]; ++j)
 169                         ppmu->get_constraint(alternatives[i][j],
 170                                              &amasks[i][j], &avalues[i][j]);
 171         }
 172
 173         /* enumerate all possibilities and see if any will work */
 174         i = 0;
 175         j = -1;
 176         value = mask = nv = 0;
 177         while (i < n_ev) {
 178                 if (j >= 0) {
 179                         /* we're backtracking, restore context */
 180                         value = svalues[i];
 181                         mask = smasks[i];
 182                         j = choice[i];
 183                 }
 184                 /*
 185                  * See if any alternative k for event i,
 186                  * where k > j, will satisfy the constraints.
 187                  */
 188                 while (++j < n_alt[i]) {
 189                         nv = (value | avalues[i][j]) +
 190                                 (value & avalues[i][j] & addf);
 191                         if ((((nv + tadd) ^ value) & mask) == 0 &&
 192                             (((nv + tadd) ^ avalues[i][j])
 193                              & amasks[i][j]) == 0)
 194                                 break;
 195                 }
 196                 if (j >= n_alt[i]) {
 197                         /*
 198                          * No feasible alternative, backtrack
 199                          * to event i-1 and continue enumerating its
 200                          * alternatives from where we got up to.
 201                          */
 202                         if (--i < 0)
 203                                 return -1;
 204                 } else {
 205                         /*
 206                          * Found a feasible alternative for event i,
 207                          * remember where we got up to with this event,
 208                          * go on to the next event, and start with
 209                          * the first alternative for it.
 210                          */
 211                         choice[i] = j;
 212                         svalues[i] = value;
 213                         smasks[i] = mask;
 214                         value = nv;
 215                         mask |= amasks[i][j];
 216                         ++i;
 217                         j = -1;
 218                 }
 219         }
 220
 221         /* OK, we have a feasible combination, tell the caller the solution */
 222         for (i = 0; i < n_ev; ++i)
 223                 event[i] = alternatives[i][choice[i]];
 224         return 0;
 225 }
 226
 227 /*
 228  * Check if newly-added counters have consistent settings for
 229  * exclude_{user,kernel,hv} with each other and any previously
 230  * added counters.
 231  */
 232 static int check_excludes(struct perf_counter **ctrs, int n_prev, int n_new)
 233 {
 234         int eu, ek, eh;
 235         int i, n;
 236         struct perf_counter *counter;
 237
 238         n = n_prev + n_new;
 239         if (n <= 1)
 240                 return 0;
 241
 242         eu = ctrs[0]->hw_event.exclude_user;
 243         ek = ctrs[0]->hw_event.exclude_kernel;
 244         eh = ctrs[0]->hw_event.exclude_hv;
 245         if (n_prev == 0)
 246                 n_prev = 1;
 247         for (i = n_prev; i < n; ++i) {
 248                 counter = ctrs[i];
 249                 if (counter->hw_event.exclude_user != eu ||
 250                     counter->hw_event.exclude_kernel != ek ||
 251                     counter->hw_event.exclude_hv != eh)
 252                         return -EAGAIN;
 253         }
 254         return 0;
 255 }
 256
 257 static void power_perf_read(struct perf_counter *counter)
 258 {
 259         long val, delta, prev;
 260
 261         if (!counter->hw.idx)
 262                 return;
 263         /*
 264          * Performance monitor interrupts come even when interrupts
 265          * are soft-disabled, as long as interrupts are hard-enabled.
 266          * Therefore we treat them like NMIs.
 267          */
 268         do {
 269                 prev = atomic64_read(&counter->hw.prev_count);
 270                 barrier();
 271                 val = read_pmc(counter->hw.idx);
 272         } while (atomic64_cmpxchg(&counter->hw.prev_count, prev, val) != prev);
 273
 274         /* The counters are only 32 bits wide */
 275         delta = (val - prev) & 0xfffffffful;
 276         atomic64_add(delta, &counter->count);
 277         atomic64_sub(delta, &counter->hw.period_left);
 278 }
 279
 280 /*
 281  * Disable all counters to prevent PMU interrupts and to allow
 282  * counters to be added or removed.
 283  */
 284 u64 hw_perf_save_disable(void)
 285 {
 286         struct cpu_hw_counters *cpuhw;
 287         unsigned long ret;
 288         unsigned long flags;
 289
 290         local_irq_save(flags);
 291         cpuhw = &__get_cpu_var(cpu_hw_counters);
 292
 293         ret = cpuhw->disabled;
 294         if (!ret) {
 295                 cpuhw->disabled = 1;
 296                 cpuhw->n_added = 0;
 297
 298                 /*
 299                  * Check if we ever enabled the PMU on this cpu.
 300                  */
 301                 if (!cpuhw->pmcs_enabled) {
 302                         if (ppc_md.enable_pmcs)
 303                                 ppc_md.enable_pmcs();
 304                         cpuhw->pmcs_enabled = 1;
 305                 }
 306
 307                 /*
 308                  * Set the 'freeze counters' bit.
 309                  * The barrier is to make sure the mtspr has been
 310                  * executed and the PMU has frozen the counters
 311                  * before we return.
 312                  */
 313                 mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
 314                 mb();
 315         }
 316         local_irq_restore(flags);
 317         return ret;
 318 }
 319
 320 /*
 321  * Re-enable all counters if disable == 0.
 322  * If we were previously disabled and counters were added, then
 323  * put the new config on the PMU.
 324  */
 325 void hw_perf_restore(u64 disable)
 326 {
 327         struct perf_counter *counter;
 328         struct cpu_hw_counters *cpuhw;
 329         unsigned long flags;
 330         long i;
 331         unsigned long val;
 332         s64 left;
 333         unsigned int hwc_index[MAX_HWCOUNTERS];
 334
 335         if (disable)
 336                 return;
 337         local_irq_save(flags);
 338         cpuhw = &__get_cpu_var(cpu_hw_counters);
 339         cpuhw->disabled = 0;
 340
 341         /*
 342          * If we didn't change anything, or only removed counters,
 343          * no need to recalculate MMCR* settings and reset the PMCs.
 344          * Just reenable the PMU with the current MMCR* settings
 345          * (possibly updated for removal of counters).
 346          */
 347         if (!cpuhw->n_added) {
 348                 mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
 349                 mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 350                 mtspr(SPRN_MMCR0, cpuhw->mmcr[0]);
 351                 if (cpuhw->n_counters == 0)
 352                         get_lppaca()->pmcregs_in_use = 0;
 353                 goto out;
 354         }
 355
 356         /*
 357          * Compute MMCR* values for the new set of counters
 358          */
 359         if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_counters, hwc_index,
 360                                cpuhw->mmcr)) {
 361                 /* shouldn't ever get here */
 362                 printk(KERN_ERR "oops compute_mmcr failed\n");
 363                 goto out;
 364         }
 365
 366         /*
 367          * Add in MMCR0 freeze bits corresponding to the
 368          * hw_event.exclude_* bits for the first counter.
 369          * We have already checked that all counters have the
 370          * same values for these bits as the first counter.
 371          */
 372         counter = cpuhw->counter[0];
 373         if (counter->hw_event.exclude_user)
 374                 cpuhw->mmcr[0] |= MMCR0_FCP;
 375         if (counter->hw_event.exclude_kernel)
 376                 cpuhw->mmcr[0] |= freeze_counters_kernel;
 377         if (counter->hw_event.exclude_hv)
 378                 cpuhw->mmcr[0] |= MMCR0_FCHV;
 379
 380         /*
 381          * Write the new configuration to MMCR* with the freeze
 382          * bit set and set the hardware counters to their initial values.
 383          * Then unfreeze the counters.
 384          */
 385         get_lppaca()->pmcregs_in_use = 1;
 386         mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
 387         mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
 388         mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
 389                                 | MMCR0_FC);
 390
 391         /*
 392          * Read off any pre-existing counters that need to move
 393          * to another PMC.
 394          */
 395         for (i = 0; i < cpuhw->n_counters; ++i) {
 396                 counter = cpuhw->counter[i];
 397                 if (counter->hw.idx && counter->hw.idx != hwc_index[i] + 1) {
 398                         power_perf_read(counter);
 399                         write_pmc(counter->hw.idx, 0);
 400                         counter->hw.idx = 0;
 401                 }
 402         }
 403
 404         /*
 405          * Initialize the PMCs for all the new and moved counters.
 406          */
 407         for (i = 0; i < cpuhw->n_counters; ++i) {
 408                 counter = cpuhw->counter[i];
 409                 if (counter->hw.idx)
 410                         continue;
 411                 val = 0;
 412                 if (counter->hw_event.irq_period) {
 413                         left = atomic64_read(&counter->hw.period_left);
 414                         if (left < 0x80000000L)
 415                                 val = 0x80000000L - left;
 416                 }
 417                 atomic64_set(&counter->hw.prev_count, val);
 418                 counter->hw.idx = hwc_index[i] + 1;
 419                 write_pmc(counter->hw.idx, val);
 420         }
 421         mb();
 422         cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
 423         mtspr(SPRN_MMCR0, cpuhw->mmcr[0]);
 424
 425  out:
 426         local_irq_restore(flags);
 427 }
 428
 429 static int collect_events(struct perf_counter *group, int max_count,
 430                           struct perf_counter *ctrs[], unsigned int *events)
 431 {
 432         int n = 0;
 433         struct perf_counter *counter;
 434
 435         if (!is_software_counter(group)) {
 436                 if (n >= max_count)
 437                         return -1;
 438                 ctrs[n] = group;
 439                 events[n++] = group->hw.config;
 440         }
 441         list_for_each_entry(counter, &group->sibling_list, list_entry) {
 442                 if (!is_software_counter(counter) &&
 443                     counter->state != PERF_COUNTER_STATE_OFF) {
 444                         if (n >= max_count)
 445                                 return -1;
 446                         ctrs[n] = counter;
 447                         events[n++] = counter->hw.config;
 448                 }
 449         }
 450         return n;
 451 }
 452
 453 static void counter_sched_in(struct perf_counter *counter, int cpu)
 454 {
 455         counter->state = PERF_COUNTER_STATE_ACTIVE;
 456         counter->oncpu = cpu;
 457         if (is_software_counter(counter))
 458                 counter->hw_ops->enable(counter);
 459 }
 460
 461 /*
 462  * Called to enable a whole group of counters.
 463  * Returns 1 if the group was enabled, or -EAGAIN if it could not be.
 464  * Assumes the caller has disabled interrupts and has
 465  * frozen the PMU with hw_perf_save_disable.
 466  */
 467 int hw_perf_group_sched_in(struct perf_counter *group_leader,
 468                struct perf_cpu_context *cpuctx,
 469                struct perf_counter_context *ctx, int cpu)
 470 {
 471         struct cpu_hw_counters *cpuhw;
 472         long i, n, n0;
 473         struct perf_counter *sub;
 474
 475         cpuhw = &__get_cpu_var(cpu_hw_counters);
 476         n0 = cpuhw->n_counters;
 477         n = collect_events(group_leader, ppmu->n_counter - n0,
 478                            &cpuhw->counter[n0], &cpuhw->events[n0]);
 479         if (n < 0)
 480                 return -EAGAIN;
 481         if (check_excludes(cpuhw->counter, n0, n))
 482                 return -EAGAIN;
 483         if (power_check_constraints(cpuhw->events, n + n0))
 484                 return -EAGAIN;
 485         cpuhw->n_counters = n0 + n;
 486         cpuhw->n_added += n;
 487
 488         /*
 489          * OK, this group can go on; update counter states etc.,
 490          * and enable any software counters
 491          */
 492         for (i = n0; i < n0 + n; ++i)
 493                 cpuhw->counter[i]->hw.config = cpuhw->events[i];
 494         cpuctx->active_oncpu += n;
 495         n = 1;
 496         counter_sched_in(group_leader, cpu);
 497         list_for_each_entry(sub, &group_leader->sibling_list, list_entry) {
 498                 if (sub->state != PERF_COUNTER_STATE_OFF) {
 499                         counter_sched_in(sub, cpu);
 500                         ++n;
 501                 }
 502         }
 503         ctx->nr_active += n;
 504
 505         return 1;
 506 }
 507
 508 /*
 509  * Add a counter to the PMU.
 510  * If all counters are not already frozen, then we disable and
 511  * re-enable the PMU in order to get hw_perf_restore to do the
 512  * actual work of reconfiguring the PMU.
 513  */
 514 static int power_perf_enable(struct perf_counter *counter)
 515 {
 516         struct cpu_hw_counters *cpuhw;
 517         unsigned long flags;
 518         u64 pmudis;
 519         int n0;
 520         int ret = -EAGAIN;
 521
 522         local_irq_save(flags);
 523         pmudis = hw_perf_save_disable();
 524
 525         /*
 526          * Add the counter to the list (if there is room)
 527          * and check whether the total set is still feasible.
 528          */
 529         cpuhw = &__get_cpu_var(cpu_hw_counters);
 530         n0 = cpuhw->n_counters;
 531         if (n0 >= ppmu->n_counter)
 532                 goto out;
 533         cpuhw->counter[n0] = counter;
 534         cpuhw->events[n0] = counter->hw.config;
 535         if (check_excludes(cpuhw->counter, n0, 1))
 536                 goto out;
 537         if (power_check_constraints(cpuhw->events, n0 + 1))
 538                 goto out;
 539
 540         counter->hw.config = cpuhw->events[n0];
 541         ++cpuhw->n_counters;
 542         ++cpuhw->n_added;
 543
 544         ret = 0;
 545  out:
 546         hw_perf_restore(pmudis);
 547         local_irq_restore(flags);
 548         return ret;
 549 }
 550
 551 /*
 552  * Remove a counter from the PMU.
 553  */
 554 static void power_perf_disable(struct perf_counter *counter)
 555 {
 556         struct cpu_hw_counters *cpuhw;
 557         long i;
 558         u64 pmudis;
 559         unsigned long flags;
 560
 561         local_irq_save(flags);
 562         pmudis = hw_perf_save_disable();
 563
 564         power_perf_read(counter);
 565
 566         cpuhw = &__get_cpu_var(cpu_hw_counters);
 567         for (i = 0; i < cpuhw->n_counters; ++i) {
 568                 if (counter == cpuhw->counter[i]) {
 569                         while (++i < cpuhw->n_counters)
 570                                 cpuhw->counter[i-1] = cpuhw->counter[i];
 571                         --cpuhw->n_counters;
 572                         ppmu->disable_pmc(counter->hw.idx - 1, cpuhw->mmcr);
 573                         write_pmc(counter->hw.idx, 0);
 574                         counter->hw.idx = 0;
 575                         break;
 576                 }
 577         }
 578         if (cpuhw->n_counters == 0) {
 579                 /* disable exceptions if no counters are running */
 580                 cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
 581         }
 582
 583         hw_perf_restore(pmudis);
 584         local_irq_restore(flags);
 585 }
 586
 587 struct hw_perf_counter_ops power_perf_ops = {
 588         .enable = power_perf_enable,
 589         .disable = power_perf_disable,
 590         .read = power_perf_read
 591 };
 592
 593 const struct hw_perf_counter_ops *
 594 hw_perf_counter_init(struct perf_counter *counter)
 595 {
 596         unsigned long ev;
 597         struct perf_counter *ctrs[MAX_HWCOUNTERS];
 598         unsigned int events[MAX_HWCOUNTERS];
 599         int n;
 600
 601         if (!ppmu)
 602                 return NULL;
 603         if ((s64)counter->hw_event.irq_period < 0)
 604                 return NULL;
 605         ev = counter->hw_event.type;
 606         if (!counter->hw_event.raw) {
 607                 if (ev >= ppmu->n_generic ||
 608                     ppmu->generic_events[ev] == 0)
 609                         return NULL;
 610                 ev = ppmu->generic_events[ev];
 611         }
 612         counter->hw.config_base = ev;
 613         counter->hw.idx = 0;
 614
 615         /*
 616          * If we are not running on a hypervisor, force the
 617          * exclude_hv bit to 0 so that we don't care what
 618          * the user set it to.
 619          */
 620         if (!firmware_has_feature(FW_FEATURE_LPAR))
 621                 counter->hw_event.exclude_hv = 0;
 622
 623         /*
 624          * If this is in a group, check if it can go on with all the
 625          * other hardware counters in the group.  We assume the counter
 626          * hasn't been linked into its leader's sibling list at this point.
 627          */
 628         n = 0;
 629         if (counter->group_leader != counter) {
 630                 n = collect_events(counter->group_leader, ppmu->n_counter - 1,
 631                                    ctrs, events);
 632                 if (n < 0)
 633                         return NULL;
 634         }
 635         events[n] = ev;
 636         ctrs[n] = counter;
 637         if (check_excludes(ctrs, n, 1))
 638                 return NULL;
 639         if (power_check_constraints(events, n + 1))
 640                 return NULL;
 641
 642         counter->hw.config = events[n];
 643         atomic64_set(&counter->hw.period_left, counter->hw_event.irq_period);
 644         return &power_perf_ops;
 645 }
 646
 647 /*
 648  * Handle wakeups.
 649  */
 650 void perf_counter_do_pending(void)
 651 {
 652         int i;
 653         struct cpu_hw_counters *cpuhw = &__get_cpu_var(cpu_hw_counters);
 654         struct perf_counter *counter;
 655
 656         set_perf_counter_pending(0);
 657         for (i = 0; i < cpuhw->n_counters; ++i) {
 658                 counter = cpuhw->counter[i];
 659                 if (counter && counter->wakeup_pending) {
 660                         counter->wakeup_pending = 0;
 661                         wake_up(&counter->waitq);
 662                 }
 663         }
 664 }
 665
 666 /*
 667  * Record data for an irq counter.
 668  * This function was lifted from the x86 code; maybe it should
 669  * go in the core?
 670  */
 671 static void perf_store_irq_data(struct perf_counter *counter, u64 data)
 672 {
 673         struct perf_data *irqdata = counter->irqdata;
 674
 675         if (irqdata->len > PERF_DATA_BUFLEN - sizeof(u64)) {
 676                 irqdata->overrun++;
 677         } else {
 678                 u64 *p = (u64 *) &irqdata->data[irqdata->len];
 679
 680                 *p = data;
 681                 irqdata->len += sizeof(u64);
 682         }
 683 }
 684
 685 /*
 686  * Record all the values of the counters in a group
 687  */
 688 static void perf_handle_group(struct perf_counter *counter)
 689 {
 690         struct perf_counter *leader, *sub;
 691
 692         leader = counter->group_leader;
 693         list_for_each_entry(sub, &leader->sibling_list, list_entry) {
 694                 if (sub != counter)
 695                         sub->hw_ops->read(sub);
 696                 perf_store_irq_data(counter, sub->hw_event.type);
 697                 perf_store_irq_data(counter, atomic64_read(&sub->count));
 698         }
 699 }
 700
 701 /*
 702  * A counter has overflowed; update its count and record
 703  * things if requested.  Note that interrupts are hard-disabled
 704  * here so there is no possibility of being interrupted.
 705  */
 706 static void record_and_restart(struct perf_counter *counter, long val,
 707                                struct pt_regs *regs)
 708 {
 709         s64 prev, delta, left;
 710         int record = 0;
 711
 712         /* we don't have to worry about interrupts here */
 713         prev = atomic64_read(&counter->hw.prev_count);
 714         delta = (val - prev) & 0xfffffffful;
 715         atomic64_add(delta, &counter->count);
 716
 717         /*
 718          * See if the total period for this counter has expired,
 719          * and update for the next period.
 720          */
 721         val = 0;
 722         left = atomic64_read(&counter->hw.period_left) - delta;
 723         if (counter->hw_event.irq_period) {
 724                 if (left <= 0) {
 725                         left += counter->hw_event.irq_period;
 726                         if (left <= 0)
 727                                 left = counter->hw_event.irq_period;
 728                         record = 1;
 729                 }
 730                 if (left < 0x80000000L)
 731                         val = 0x80000000L - left;
 732         }
 733         write_pmc(counter->hw.idx, val);
 734         atomic64_set(&counter->hw.prev_count, val);
 735         atomic64_set(&counter->hw.period_left, left);
 736
 737         /*
 738          * Finally record data if requested.
 739          */
 740         if (record) {
 741                 switch (counter->hw_event.record_type) {
 742                 case PERF_RECORD_SIMPLE:
 743                         break;
 744                 case PERF_RECORD_IRQ:
 745                         perf_store_irq_data(counter, instruction_pointer(regs));
 746                         counter->wakeup_pending = 1;
 747                         break;
 748                 case PERF_RECORD_GROUP:
 749                         perf_handle_group(counter);
 750                         counter->wakeup_pending = 1;
 751                         break;
 752                 }
 753         }
 754 }
 755
 756 /*
 757  * Performance monitor interrupt stuff
 758  */
 759 static void perf_counter_interrupt(struct pt_regs *regs)
 760 {
 761         int i;
 762         struct cpu_hw_counters *cpuhw = &__get_cpu_var(cpu_hw_counters);
 763         struct perf_counter *counter;
 764         long val;
 765         int need_wakeup = 0, found = 0;
 766
 767         for (i = 0; i < cpuhw->n_counters; ++i) {
 768                 counter = cpuhw->counter[i];
 769                 val = read_pmc(counter->hw.idx);
 770                 if ((int)val < 0) {
 771                         /* counter has overflowed */
 772                         found = 1;
 773                         record_and_restart(counter, val, regs);
 774                         if (counter->wakeup_pending)
 775                                 need_wakeup = 1;
 776                 }
 777         }
 778
 779         /*
 780          * In case we didn't find and reset the counter that caused
 781          * the interrupt, scan all counters and reset any that are
 782          * negative, to avoid getting continual interrupts.
 783          * Any that we processed in the previous loop will not be negative.
 784          */
 785         if (!found) {
 786                 for (i = 0; i < ppmu->n_counter; ++i) {
 787                         val = read_pmc(i + 1);
 788                         if ((int)val < 0)
 789                                 write_pmc(i + 1, 0);
 790                 }
 791         }
 792
 793         /*
 794          * Reset MMCR0 to its normal value.  This will set PMXE and
 795          * clear FC (freeze counters) and PMAO (perf mon alert occurred)
 796          * and thus allow interrupts to occur again.
 797          * XXX might want to use MSR.PM to keep the counters frozen until
 798          * we get back out of this interrupt.
 799          */
 800         mtspr(SPRN_MMCR0, cpuhw->mmcr[0]);
 801
 802         /*
 803          * If we need a wakeup, check whether interrupts were soft-enabled
 804          * when we took the interrupt.  If they were, we can wake stuff up
 805          * immediately; otherwise we'll have to set a flag and do the
 806          * wakeup when interrupts get soft-enabled.
 807          */
 808         if (need_wakeup) {
 809                 if (regs->softe) {
 810                         irq_enter();
 811                         perf_counter_do_pending();
 812                         irq_exit();
 813                 } else {
 814                         set_perf_counter_pending(1);
 815                 }
 816         }
 817 }
 818
 819 void hw_perf_counter_setup(int cpu)
 820 {
 821         struct cpu_hw_counters *cpuhw = &per_cpu(cpu_hw_counters, cpu);
 822
 823         memset(cpuhw, 0, sizeof(*cpuhw));
 824         cpuhw->mmcr[0] = MMCR0_FC;
 825 }
 826
 827 extern struct power_pmu ppc970_pmu;
 828 extern struct power_pmu power5_pmu;
 829 extern struct power_pmu power6_pmu;
 830
 831 static int init_perf_counters(void)
 832 {
 833         unsigned long pvr;
 834
 835         if (reserve_pmc_hardware(perf_counter_interrupt)) {
 836                 printk(KERN_ERR "Couldn't init performance monitor subsystem\n");
 837                 return -EBUSY;
 838         }
 839
 840         /* XXX should get this from cputable */
 841         pvr = mfspr(SPRN_PVR);
 842         switch (PVR_VER(pvr)) {
 843         case PV_970:
 844         case PV_970FX:
 845         case PV_970MP:
 846                 ppmu = &ppc970_pmu;
 847                 break;
 848         case PV_POWER5:
 849                 ppmu = &power5_pmu;
 850                 break;
 851         case 0x3e:
 852                 ppmu = &power6_pmu;
 853                 break;
 854         }
 855
 856         /*
 857          * Use FCHV to ignore kernel events if MSR.HV is set.
 858          */
 859         if (mfmsr() & MSR_HV)
 860                 freeze_counters_kernel = MMCR0_FCHV;
 861
 862         return 0;
 863 }
 864
 865 arch_initcall(init_perf_counters);