1/* 2 * Linux performance counter support for MIPS. 3 * 4 * Copyright (C) 2010 MIPS Technologies, Inc. 5 * Copyright (C) 2011 Cavium Networks, Inc. 6 * Author: Deng-Cheng Zhu 7 * 8 * This code is based on the implementation for ARM, which is in turn 9 * based on the sparc64 perf event code and the x86 code. Performance 10 * counter access is based on the MIPS Oprofile code. And the callchain 11 * support references the code of MIPS stacktrace.c. 12 * 13 * This program is free software; you can redistribute it and/or modify 14 * it under the terms of the GNU General Public License version 2 as 15 * published by the Free Software Foundation. 16 */ 17 18#include <linux/cpumask.h> 19#include <linux/interrupt.h> 20#include <linux/smp.h> 21#include <linux/kernel.h> 22#include <linux/perf_event.h> 23#include <linux/uaccess.h> 24 25#include <asm/irq.h> 26#include <asm/irq_regs.h> 27#include <asm/stacktrace.h> 28#include <asm/time.h> /* For perf_irq */ 29 30#define MIPS_MAX_HWEVENTS 4 31#define MIPS_TCS_PER_COUNTER 2 32#define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1) 33 34struct cpu_hw_events { 35 /* Array of events on this cpu. */ 36 struct perf_event *events[MIPS_MAX_HWEVENTS]; 37 38 /* 39 * Set the bit (indexed by the counter number) when the counter 40 * is used for an event. 41 */ 42 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)]; 43 44 /* 45 * Software copy of the control register for each performance counter. 46 * MIPS CPUs vary in performance counters. They use this differently, 47 * and even may not use it. 48 */ 49 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS]; 50}; 51DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { 52 .saved_ctrl = {0}, 53}; 54 55/* The description of MIPS performance events. */ 56struct mips_perf_event { 57 unsigned int event_id; 58 /* 59 * MIPS performance counters are indexed starting from 0. 60 * CNTR_EVEN indicates the indexes of the counters to be used are 61 * even numbers. 62 */ 63 unsigned int cntr_mask; 64 #define CNTR_EVEN 0x55555555 65 #define CNTR_ODD 0xaaaaaaaa 66 #define CNTR_ALL 0xffffffff 67#ifdef CONFIG_MIPS_MT_SMP 68 enum { 69 T = 0, 70 V = 1, 71 P = 2, 72 } range; 73#else 74 #define T 75 #define V 76 #define P 77#endif 78}; 79 80static struct mips_perf_event raw_event; 81static DEFINE_MUTEX(raw_event_mutex); 82 83#define C(x) PERF_COUNT_HW_CACHE_##x 84 85struct mips_pmu { 86 u64 max_period; 87 u64 valid_count; 88 u64 overflow; 89 const char *name; 90 int irq; 91 u64 (*read_counter)(unsigned int idx); 92 void (*write_counter)(unsigned int idx, u64 val); 93 const struct mips_perf_event *(*map_raw_event)(u64 config); 94 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX]; 95 const struct mips_perf_event (*cache_event_map) 96 [PERF_COUNT_HW_CACHE_MAX] 97 [PERF_COUNT_HW_CACHE_OP_MAX] 98 [PERF_COUNT_HW_CACHE_RESULT_MAX]; 99 unsigned int num_counters; 100}; 101 102static struct mips_pmu mipspmu; 103 104#define M_CONFIG1_PC (1 << 4) 105 106#define M_PERFCTL_EXL (1 << 0) 107#define M_PERFCTL_KERNEL (1 << 1) 108#define M_PERFCTL_SUPERVISOR (1 << 2) 109#define M_PERFCTL_USER (1 << 3) 110#define M_PERFCTL_INTERRUPT_ENABLE (1 << 4) 111#define M_PERFCTL_EVENT(event) (((event) & 0x3ff) << 5) 112#define M_PERFCTL_VPEID(vpe) ((vpe) << 16) 113 114#ifdef CONFIG_CPU_BMIPS5000 115#define M_PERFCTL_MT_EN(filter) 0 116#else /* !CONFIG_CPU_BMIPS5000 */ 117#define M_PERFCTL_MT_EN(filter) ((filter) << 20) 118#endif /* CONFIG_CPU_BMIPS5000 */ 119 120#define M_TC_EN_ALL M_PERFCTL_MT_EN(0) 121#define M_TC_EN_VPE M_PERFCTL_MT_EN(1) 122#define M_TC_EN_TC M_PERFCTL_MT_EN(2) 123#define M_PERFCTL_TCID(tcid) ((tcid) << 22) 124#define M_PERFCTL_WIDE (1 << 30) 125#define M_PERFCTL_MORE (1 << 31) 126#define M_PERFCTL_TC (1 << 30) 127 128#define M_PERFCTL_COUNT_EVENT_WHENEVER (M_PERFCTL_EXL | \ 129 M_PERFCTL_KERNEL | \ 130 M_PERFCTL_USER | \ 131 M_PERFCTL_SUPERVISOR | \ 132 M_PERFCTL_INTERRUPT_ENABLE) 133 134#ifdef CONFIG_MIPS_MT_SMP 135#define M_PERFCTL_CONFIG_MASK 0x3fff801f 136#else 137#define M_PERFCTL_CONFIG_MASK 0x1f 138#endif 139#define M_PERFCTL_EVENT_MASK 0xfe0 140 141 142#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 143static int cpu_has_mipsmt_pertccounters; 144 145static DEFINE_RWLOCK(pmuint_rwlock); 146 147#if defined(CONFIG_CPU_BMIPS5000) 148#define vpe_id() (cpu_has_mipsmt_pertccounters ? \ 149 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK)) 150#else 151/* 152 * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because 153 * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs. 154 */ 155#define vpe_id() (cpu_has_mipsmt_pertccounters ? \ 156 0 : smp_processor_id()) 157#endif 158 159/* Copied from op_model_mipsxx.c */ 160static unsigned int vpe_shift(void) 161{ 162 if (num_possible_cpus() > 1) 163 return 1; 164 165 return 0; 166} 167 168static unsigned int counters_total_to_per_cpu(unsigned int counters) 169{ 170 return counters >> vpe_shift(); 171} 172 173#else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */ 174#define vpe_id() 0 175 176#endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */ 177 178static void resume_local_counters(void); 179static void pause_local_counters(void); 180static irqreturn_t mipsxx_pmu_handle_irq(int, void *); 181static int mipsxx_pmu_handle_shared_irq(void); 182 183static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx) 184{ 185 if (vpe_id() == 1) 186 idx = (idx + 2) & 3; 187 return idx; 188} 189 190static u64 mipsxx_pmu_read_counter(unsigned int idx) 191{ 192 idx = mipsxx_pmu_swizzle_perf_idx(idx); 193 194 switch (idx) { 195 case 0: 196 /* 197 * The counters are unsigned, we must cast to truncate 198 * off the high bits. 199 */ 200 return (u32)read_c0_perfcntr0(); 201 case 1: 202 return (u32)read_c0_perfcntr1(); 203 case 2: 204 return (u32)read_c0_perfcntr2(); 205 case 3: 206 return (u32)read_c0_perfcntr3(); 207 default: 208 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 209 return 0; 210 } 211} 212 213static u64 mipsxx_pmu_read_counter_64(unsigned int idx) 214{ 215 idx = mipsxx_pmu_swizzle_perf_idx(idx); 216 217 switch (idx) { 218 case 0: 219 return read_c0_perfcntr0_64(); 220 case 1: 221 return read_c0_perfcntr1_64(); 222 case 2: 223 return read_c0_perfcntr2_64(); 224 case 3: 225 return read_c0_perfcntr3_64(); 226 default: 227 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 228 return 0; 229 } 230} 231 232static void mipsxx_pmu_write_counter(unsigned int idx, u64 val) 233{ 234 idx = mipsxx_pmu_swizzle_perf_idx(idx); 235 236 switch (idx) { 237 case 0: 238 write_c0_perfcntr0(val); 239 return; 240 case 1: 241 write_c0_perfcntr1(val); 242 return; 243 case 2: 244 write_c0_perfcntr2(val); 245 return; 246 case 3: 247 write_c0_perfcntr3(val); 248 return; 249 } 250} 251 252static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val) 253{ 254 idx = mipsxx_pmu_swizzle_perf_idx(idx); 255 256 switch (idx) { 257 case 0: 258 write_c0_perfcntr0_64(val); 259 return; 260 case 1: 261 write_c0_perfcntr1_64(val); 262 return; 263 case 2: 264 write_c0_perfcntr2_64(val); 265 return; 266 case 3: 267 write_c0_perfcntr3_64(val); 268 return; 269 } 270} 271 272static unsigned int mipsxx_pmu_read_control(unsigned int idx) 273{ 274 idx = mipsxx_pmu_swizzle_perf_idx(idx); 275 276 switch (idx) { 277 case 0: 278 return read_c0_perfctrl0(); 279 case 1: 280 return read_c0_perfctrl1(); 281 case 2: 282 return read_c0_perfctrl2(); 283 case 3: 284 return read_c0_perfctrl3(); 285 default: 286 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 287 return 0; 288 } 289} 290 291static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val) 292{ 293 idx = mipsxx_pmu_swizzle_perf_idx(idx); 294 295 switch (idx) { 296 case 0: 297 write_c0_perfctrl0(val); 298 return; 299 case 1: 300 write_c0_perfctrl1(val); 301 return; 302 case 2: 303 write_c0_perfctrl2(val); 304 return; 305 case 3: 306 write_c0_perfctrl3(val); 307 return; 308 } 309} 310 311static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc, 312 struct hw_perf_event *hwc) 313{ 314 int i; 315 316 /* 317 * We only need to care the counter mask. The range has been 318 * checked definitely. 319 */ 320 unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff; 321 322 for (i = mipspmu.num_counters - 1; i >= 0; i--) { 323 /* 324 * Note that some MIPS perf events can be counted by both 325 * even and odd counters, wheresas many other are only by 326 * even _or_ odd counters. This introduces an issue that 327 * when the former kind of event takes the counter the 328 * latter kind of event wants to use, then the "counter 329 * allocation" for the latter event will fail. In fact if 330 * they can be dynamically swapped, they both feel happy. 331 * But here we leave this issue alone for now. 332 */ 333 if (test_bit(i, &cntr_mask) && 334 !test_and_set_bit(i, cpuc->used_mask)) 335 return i; 336 } 337 338 return -EAGAIN; 339} 340 341static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx) 342{ 343 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 344 345 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 346 347 cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) | 348 (evt->config_base & M_PERFCTL_CONFIG_MASK) | 349 /* Make sure interrupt enabled. */ 350 M_PERFCTL_INTERRUPT_ENABLE; 351 if (IS_ENABLED(CONFIG_CPU_BMIPS5000)) 352 /* enable the counter for the calling thread */ 353 cpuc->saved_ctrl[idx] |= 354 (1 << (12 + vpe_id())) | M_PERFCTL_TC; 355 356 /* 357 * We do not actually let the counter run. Leave it until start(). 358 */ 359} 360 361static void mipsxx_pmu_disable_event(int idx) 362{ 363 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 364 unsigned long flags; 365 366 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 367 368 local_irq_save(flags); 369 cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) & 370 ~M_PERFCTL_COUNT_EVENT_WHENEVER; 371 mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]); 372 local_irq_restore(flags); 373} 374 375static int mipspmu_event_set_period(struct perf_event *event, 376 struct hw_perf_event *hwc, 377 int idx) 378{ 379 u64 left = local64_read(&hwc->period_left); 380 u64 period = hwc->sample_period; 381 int ret = 0; 382 383 if (unlikely((left + period) & (1ULL << 63))) { 384 /* left underflowed by more than period. */ 385 left = period; 386 local64_set(&hwc->period_left, left); 387 hwc->last_period = period; 388 ret = 1; 389 } else if (unlikely((left + period) <= period)) { 390 /* left underflowed by less than period. */ 391 left += period; 392 local64_set(&hwc->period_left, left); 393 hwc->last_period = period; 394 ret = 1; 395 } 396 397 if (left > mipspmu.max_period) { 398 left = mipspmu.max_period; 399 local64_set(&hwc->period_left, left); 400 } 401 402 local64_set(&hwc->prev_count, mipspmu.overflow - left); 403 404 mipspmu.write_counter(idx, mipspmu.overflow - left); 405 406 perf_event_update_userpage(event); 407 408 return ret; 409} 410 411static void mipspmu_event_update(struct perf_event *event, 412 struct hw_perf_event *hwc, 413 int idx) 414{ 415 u64 prev_raw_count, new_raw_count; 416 u64 delta; 417 418again: 419 prev_raw_count = local64_read(&hwc->prev_count); 420 new_raw_count = mipspmu.read_counter(idx); 421 422 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, 423 new_raw_count) != prev_raw_count) 424 goto again; 425 426 delta = new_raw_count - prev_raw_count; 427 428 local64_add(delta, &event->count); 429 local64_sub(delta, &hwc->period_left); 430} 431 432static void mipspmu_start(struct perf_event *event, int flags) 433{ 434 struct hw_perf_event *hwc = &event->hw; 435 436 if (flags & PERF_EF_RELOAD) 437 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); 438 439 hwc->state = 0; 440 441 /* Set the period for the event. */ 442 mipspmu_event_set_period(event, hwc, hwc->idx); 443 444 /* Enable the event. */ 445 mipsxx_pmu_enable_event(hwc, hwc->idx); 446} 447 448static void mipspmu_stop(struct perf_event *event, int flags) 449{ 450 struct hw_perf_event *hwc = &event->hw; 451 452 if (!(hwc->state & PERF_HES_STOPPED)) { 453 /* We are working on a local event. */ 454 mipsxx_pmu_disable_event(hwc->idx); 455 barrier(); 456 mipspmu_event_update(event, hwc, hwc->idx); 457 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; 458 } 459} 460 461static int mipspmu_add(struct perf_event *event, int flags) 462{ 463 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 464 struct hw_perf_event *hwc = &event->hw; 465 int idx; 466 int err = 0; 467 468 perf_pmu_disable(event->pmu); 469 470 /* To look for a free counter for this event. */ 471 idx = mipsxx_pmu_alloc_counter(cpuc, hwc); 472 if (idx < 0) { 473 err = idx; 474 goto out; 475 } 476 477 /* 478 * If there is an event in the counter we are going to use then 479 * make sure it is disabled. 480 */ 481 event->hw.idx = idx; 482 mipsxx_pmu_disable_event(idx); 483 cpuc->events[idx] = event; 484 485 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; 486 if (flags & PERF_EF_START) 487 mipspmu_start(event, PERF_EF_RELOAD); 488 489 /* Propagate our changes to the userspace mapping. */ 490 perf_event_update_userpage(event); 491 492out: 493 perf_pmu_enable(event->pmu); 494 return err; 495} 496 497static void mipspmu_del(struct perf_event *event, int flags) 498{ 499 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 500 struct hw_perf_event *hwc = &event->hw; 501 int idx = hwc->idx; 502 503 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 504 505 mipspmu_stop(event, PERF_EF_UPDATE); 506 cpuc->events[idx] = NULL; 507 clear_bit(idx, cpuc->used_mask); 508 509 perf_event_update_userpage(event); 510} 511 512static void mipspmu_read(struct perf_event *event) 513{ 514 struct hw_perf_event *hwc = &event->hw; 515 516 /* Don't read disabled counters! */ 517 if (hwc->idx < 0) 518 return; 519 520 mipspmu_event_update(event, hwc, hwc->idx); 521} 522 523static void mipspmu_enable(struct pmu *pmu) 524{ 525#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 526 write_unlock(&pmuint_rwlock); 527#endif 528 resume_local_counters(); 529} 530 531/* 532 * MIPS performance counters can be per-TC. The control registers can 533 * not be directly accessed accross CPUs. Hence if we want to do global 534 * control, we need cross CPU calls. on_each_cpu() can help us, but we 535 * can not make sure this function is called with interrupts enabled. So 536 * here we pause local counters and then grab a rwlock and leave the 537 * counters on other CPUs alone. If any counter interrupt raises while 538 * we own the write lock, simply pause local counters on that CPU and 539 * spin in the handler. Also we know we won't be switched to another 540 * CPU after pausing local counters and before grabbing the lock. 541 */ 542static void mipspmu_disable(struct pmu *pmu) 543{ 544 pause_local_counters(); 545#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 546 write_lock(&pmuint_rwlock); 547#endif 548} 549 550static atomic_t active_events = ATOMIC_INIT(0); 551static DEFINE_MUTEX(pmu_reserve_mutex); 552static int (*save_perf_irq)(void); 553 554static int mipspmu_get_irq(void) 555{ 556 int err; 557 558 if (mipspmu.irq >= 0) { 559 /* Request my own irq handler. */ 560 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq, 561 IRQF_PERCPU | IRQF_NOBALANCING | 562 IRQF_NO_THREAD | IRQF_NO_SUSPEND | 563 IRQF_SHARED, 564 "mips_perf_pmu", &mipspmu); 565 if (err) { 566 pr_warn("Unable to request IRQ%d for MIPS performance counters!\n", 567 mipspmu.irq); 568 } 569 } else if (cp0_perfcount_irq < 0) { 570 /* 571 * We are sharing the irq number with the timer interrupt. 572 */ 573 save_perf_irq = perf_irq; 574 perf_irq = mipsxx_pmu_handle_shared_irq; 575 err = 0; 576 } else { 577 pr_warn("The platform hasn't properly defined its interrupt controller\n"); 578 err = -ENOENT; 579 } 580 581 return err; 582} 583 584static void mipspmu_free_irq(void) 585{ 586 if (mipspmu.irq >= 0) 587 free_irq(mipspmu.irq, &mipspmu); 588 else if (cp0_perfcount_irq < 0) 589 perf_irq = save_perf_irq; 590} 591 592/* 593 * mipsxx/rm9000/loongson2 have different performance counters, they have 594 * specific low-level init routines. 595 */ 596static void reset_counters(void *arg); 597static int __hw_perf_event_init(struct perf_event *event); 598 599static void hw_perf_event_destroy(struct perf_event *event) 600{ 601 if (atomic_dec_and_mutex_lock(&active_events, 602 &pmu_reserve_mutex)) { 603 /* 604 * We must not call the destroy function with interrupts 605 * disabled. 606 */ 607 on_each_cpu(reset_counters, 608 (void *)(long)mipspmu.num_counters, 1); 609 mipspmu_free_irq(); 610 mutex_unlock(&pmu_reserve_mutex); 611 } 612} 613 614static int mipspmu_event_init(struct perf_event *event) 615{ 616 int err = 0; 617 618 /* does not support taken branch sampling */ 619 if (has_branch_stack(event)) 620 return -EOPNOTSUPP; 621 622 switch (event->attr.type) { 623 case PERF_TYPE_RAW: 624 case PERF_TYPE_HARDWARE: 625 case PERF_TYPE_HW_CACHE: 626 break; 627 628 default: 629 return -ENOENT; 630 } 631 632 if (event->cpu >= nr_cpumask_bits || 633 (event->cpu >= 0 && !cpu_online(event->cpu))) 634 return -ENODEV; 635 636 if (!atomic_inc_not_zero(&active_events)) { 637 mutex_lock(&pmu_reserve_mutex); 638 if (atomic_read(&active_events) == 0) 639 err = mipspmu_get_irq(); 640 641 if (!err) 642 atomic_inc(&active_events); 643 mutex_unlock(&pmu_reserve_mutex); 644 } 645 646 if (err) 647 return err; 648 649 return __hw_perf_event_init(event); 650} 651 652static struct pmu pmu = { 653 .pmu_enable = mipspmu_enable, 654 .pmu_disable = mipspmu_disable, 655 .event_init = mipspmu_event_init, 656 .add = mipspmu_add, 657 .del = mipspmu_del, 658 .start = mipspmu_start, 659 .stop = mipspmu_stop, 660 .read = mipspmu_read, 661}; 662 663static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev) 664{ 665/* 666 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for 667 * event_id. 668 */ 669#ifdef CONFIG_MIPS_MT_SMP 670 return ((unsigned int)pev->range << 24) | 671 (pev->cntr_mask & 0xffff00) | 672 (pev->event_id & 0xff); 673#else 674 return (pev->cntr_mask & 0xffff00) | 675 (pev->event_id & 0xff); 676#endif 677} 678 679static const struct mips_perf_event *mipspmu_map_general_event(int idx) 680{ 681 682 if ((*mipspmu.general_event_map)[idx].cntr_mask == 0) 683 return ERR_PTR(-EOPNOTSUPP); 684 return &(*mipspmu.general_event_map)[idx]; 685} 686 687static const struct mips_perf_event *mipspmu_map_cache_event(u64 config) 688{ 689 unsigned int cache_type, cache_op, cache_result; 690 const struct mips_perf_event *pev; 691 692 cache_type = (config >> 0) & 0xff; 693 if (cache_type >= PERF_COUNT_HW_CACHE_MAX) 694 return ERR_PTR(-EINVAL); 695 696 cache_op = (config >> 8) & 0xff; 697 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) 698 return ERR_PTR(-EINVAL); 699 700 cache_result = (config >> 16) & 0xff; 701 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 702 return ERR_PTR(-EINVAL); 703 704 pev = &((*mipspmu.cache_event_map) 705 [cache_type] 706 [cache_op] 707 [cache_result]); 708 709 if (pev->cntr_mask == 0) 710 return ERR_PTR(-EOPNOTSUPP); 711 712 return pev; 713 714} 715 716static int validate_group(struct perf_event *event) 717{ 718 struct perf_event *sibling, *leader = event->group_leader; 719 struct cpu_hw_events fake_cpuc; 720 721 memset(&fake_cpuc, 0, sizeof(fake_cpuc)); 722 723 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0) 724 return -EINVAL; 725 726 list_for_each_entry(sibling, &leader->sibling_list, group_entry) { 727 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0) 728 return -EINVAL; 729 } 730 731 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0) 732 return -EINVAL; 733 734 return 0; 735} 736 737/* This is needed by specific irq handlers in perf_event_*.c */ 738static void handle_associated_event(struct cpu_hw_events *cpuc, 739 int idx, struct perf_sample_data *data, 740 struct pt_regs *regs) 741{ 742 struct perf_event *event = cpuc->events[idx]; 743 struct hw_perf_event *hwc = &event->hw; 744 745 mipspmu_event_update(event, hwc, idx); 746 data->period = event->hw.last_period; 747 if (!mipspmu_event_set_period(event, hwc, idx)) 748 return; 749 750 if (perf_event_overflow(event, data, regs)) 751 mipsxx_pmu_disable_event(idx); 752} 753 754 755static int __n_counters(void) 756{ 757 if (!(read_c0_config1() & M_CONFIG1_PC)) 758 return 0; 759 if (!(read_c0_perfctrl0() & M_PERFCTL_MORE)) 760 return 1; 761 if (!(read_c0_perfctrl1() & M_PERFCTL_MORE)) 762 return 2; 763 if (!(read_c0_perfctrl2() & M_PERFCTL_MORE)) 764 return 3; 765 766 return 4; 767} 768 769static int n_counters(void) 770{ 771 int counters; 772 773 switch (current_cpu_type()) { 774 case CPU_R10000: 775 counters = 2; 776 break; 777 778 case CPU_R12000: 779 case CPU_R14000: 780 case CPU_R16000: 781 counters = 4; 782 break; 783 784 default: 785 counters = __n_counters(); 786 } 787 788 return counters; 789} 790 791static void reset_counters(void *arg) 792{ 793 int counters = (int)(long)arg; 794 switch (counters) { 795 case 4: 796 mipsxx_pmu_write_control(3, 0); 797 mipspmu.write_counter(3, 0); 798 case 3: 799 mipsxx_pmu_write_control(2, 0); 800 mipspmu.write_counter(2, 0); 801 case 2: 802 mipsxx_pmu_write_control(1, 0); 803 mipspmu.write_counter(1, 0); 804 case 1: 805 mipsxx_pmu_write_control(0, 0); 806 mipspmu.write_counter(0, 0); 807 } 808} 809 810/* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */ 811static const struct mips_perf_event mipsxxcore_event_map 812 [PERF_COUNT_HW_MAX] = { 813 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, 814 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, 815 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T }, 816 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T }, 817}; 818 819/* 74K/proAptiv core has different branch event code. */ 820static const struct mips_perf_event mipsxxcore_event_map2 821 [PERF_COUNT_HW_MAX] = { 822 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, 823 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, 824 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T }, 825 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T }, 826}; 827 828static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = { 829 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN }, 830 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD }, 831 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN }, 832 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD }, 833}; 834 835static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = { 836 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL }, 837 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL }, 838 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL }, 839 [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL }, 840 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL }, 841 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL }, 842 [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL }, 843}; 844 845static const struct mips_perf_event bmips5000_event_map 846 [PERF_COUNT_HW_MAX] = { 847 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T }, 848 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, 849 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T }, 850}; 851 852static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = { 853 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL }, 854 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */ 855 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */ 856 [PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */ 857 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */ 858 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */ 859}; 860 861/* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */ 862static const struct mips_perf_event mipsxxcore_cache_map 863 [PERF_COUNT_HW_CACHE_MAX] 864 [PERF_COUNT_HW_CACHE_OP_MAX] 865 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 866[C(L1D)] = { 867 /* 868 * Like some other architectures (e.g. ARM), the performance 869 * counters don't differentiate between read and write 870 * accesses/misses, so this isn't strictly correct, but it's the 871 * best we can do. Writes and reads get combined. 872 */ 873 [C(OP_READ)] = { 874 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, 875 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, 876 }, 877 [C(OP_WRITE)] = { 878 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, 879 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, 880 }, 881}, 882[C(L1I)] = { 883 [C(OP_READ)] = { 884 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, 885 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, 886 }, 887 [C(OP_WRITE)] = { 888 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, 889 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, 890 }, 891 [C(OP_PREFETCH)] = { 892 [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T }, 893 /* 894 * Note that MIPS has only "hit" events countable for 895 * the prefetch operation. 896 */ 897 }, 898}, 899[C(LL)] = { 900 [C(OP_READ)] = { 901 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, 902 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, 903 }, 904 [C(OP_WRITE)] = { 905 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, 906 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, 907 }, 908}, 909[C(DTLB)] = { 910 [C(OP_READ)] = { 911 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 912 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 913 }, 914 [C(OP_WRITE)] = { 915 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 916 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 917 }, 918}, 919[C(ITLB)] = { 920 [C(OP_READ)] = { 921 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, 922 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, 923 }, 924 [C(OP_WRITE)] = { 925 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, 926 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, 927 }, 928}, 929[C(BPU)] = { 930 /* Using the same code for *HW_BRANCH* */ 931 [C(OP_READ)] = { 932 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, 933 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 934 }, 935 [C(OP_WRITE)] = { 936 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, 937 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 938 }, 939}, 940}; 941 942/* 74K/proAptiv core has completely different cache event map. */ 943static const struct mips_perf_event mipsxxcore_cache_map2 944 [PERF_COUNT_HW_CACHE_MAX] 945 [PERF_COUNT_HW_CACHE_OP_MAX] 946 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 947[C(L1D)] = { 948 /* 949 * Like some other architectures (e.g. ARM), the performance 950 * counters don't differentiate between read and write 951 * accesses/misses, so this isn't strictly correct, but it's the 952 * best we can do. Writes and reads get combined. 953 */ 954 [C(OP_READ)] = { 955 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, 956 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, 957 }, 958 [C(OP_WRITE)] = { 959 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, 960 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, 961 }, 962}, 963[C(L1I)] = { 964 [C(OP_READ)] = { 965 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 966 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 967 }, 968 [C(OP_WRITE)] = { 969 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 970 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 971 }, 972 [C(OP_PREFETCH)] = { 973 [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T }, 974 /* 975 * Note that MIPS has only "hit" events countable for 976 * the prefetch operation. 977 */ 978 }, 979}, 980[C(LL)] = { 981 [C(OP_READ)] = { 982 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, 983 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P }, 984 }, 985 [C(OP_WRITE)] = { 986 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, 987 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P }, 988 }, 989}, 990/* 991 * 74K core does not have specific DTLB events. proAptiv core has 992 * "speculative" DTLB events which are numbered 0x63 (even/odd) and 993 * not included here. One can use raw events if really needed. 994 */ 995[C(ITLB)] = { 996 [C(OP_READ)] = { 997 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, 998 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, 999 }, 1000 [C(OP_WRITE)] = { 1001 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, 1002 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, 1003 }, 1004}, 1005[C(BPU)] = { 1006 /* Using the same code for *HW_BRANCH* */ 1007 [C(OP_READ)] = { 1008 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, 1009 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, 1010 }, 1011 [C(OP_WRITE)] = { 1012 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, 1013 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, 1014 }, 1015}, 1016}; 1017 1018static const struct mips_perf_event loongson3_cache_map 1019 [PERF_COUNT_HW_CACHE_MAX] 1020 [PERF_COUNT_HW_CACHE_OP_MAX] 1021 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1022[C(L1D)] = { 1023 /* 1024 * Like some other architectures (e.g. ARM), the performance 1025 * counters don't differentiate between read and write 1026 * accesses/misses, so this isn't strictly correct, but it's the 1027 * best we can do. Writes and reads get combined. 1028 */ 1029 [C(OP_READ)] = { 1030 [C(RESULT_MISS)] = { 0x04, CNTR_ODD }, 1031 }, 1032 [C(OP_WRITE)] = { 1033 [C(RESULT_MISS)] = { 0x04, CNTR_ODD }, 1034 }, 1035}, 1036[C(L1I)] = { 1037 [C(OP_READ)] = { 1038 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN }, 1039 }, 1040 [C(OP_WRITE)] = { 1041 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN }, 1042 }, 1043}, 1044[C(DTLB)] = { 1045 [C(OP_READ)] = { 1046 [C(RESULT_MISS)] = { 0x09, CNTR_ODD }, 1047 }, 1048 [C(OP_WRITE)] = { 1049 [C(RESULT_MISS)] = { 0x09, CNTR_ODD }, 1050 }, 1051}, 1052[C(ITLB)] = { 1053 [C(OP_READ)] = { 1054 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD }, 1055 }, 1056 [C(OP_WRITE)] = { 1057 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD }, 1058 }, 1059}, 1060[C(BPU)] = { 1061 /* Using the same code for *HW_BRANCH* */ 1062 [C(OP_READ)] = { 1063 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN }, 1064 [C(RESULT_MISS)] = { 0x02, CNTR_ODD }, 1065 }, 1066 [C(OP_WRITE)] = { 1067 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN }, 1068 [C(RESULT_MISS)] = { 0x02, CNTR_ODD }, 1069 }, 1070}, 1071}; 1072 1073/* BMIPS5000 */ 1074static const struct mips_perf_event bmips5000_cache_map 1075 [PERF_COUNT_HW_CACHE_MAX] 1076 [PERF_COUNT_HW_CACHE_OP_MAX] 1077 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1078[C(L1D)] = { 1079 /* 1080 * Like some other architectures (e.g. ARM), the performance 1081 * counters don't differentiate between read and write 1082 * accesses/misses, so this isn't strictly correct, but it's the 1083 * best we can do. Writes and reads get combined. 1084 */ 1085 [C(OP_READ)] = { 1086 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T }, 1087 [C(RESULT_MISS)] = { 12, CNTR_ODD, T }, 1088 }, 1089 [C(OP_WRITE)] = { 1090 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T }, 1091 [C(RESULT_MISS)] = { 12, CNTR_ODD, T }, 1092 }, 1093}, 1094[C(L1I)] = { 1095 [C(OP_READ)] = { 1096 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T }, 1097 [C(RESULT_MISS)] = { 10, CNTR_ODD, T }, 1098 }, 1099 [C(OP_WRITE)] = { 1100 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T }, 1101 [C(RESULT_MISS)] = { 10, CNTR_ODD, T }, 1102 }, 1103 [C(OP_PREFETCH)] = { 1104 [C(RESULT_ACCESS)] = { 23, CNTR_EVEN, T }, 1105 /* 1106 * Note that MIPS has only "hit" events countable for 1107 * the prefetch operation. 1108 */ 1109 }, 1110}, 1111[C(LL)] = { 1112 [C(OP_READ)] = { 1113 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P }, 1114 [C(RESULT_MISS)] = { 28, CNTR_ODD, P }, 1115 }, 1116 [C(OP_WRITE)] = { 1117 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P }, 1118 [C(RESULT_MISS)] = { 28, CNTR_ODD, P }, 1119 }, 1120}, 1121[C(BPU)] = { 1122 /* Using the same code for *HW_BRANCH* */ 1123 [C(OP_READ)] = { 1124 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 1125 }, 1126 [C(OP_WRITE)] = { 1127 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 1128 }, 1129}, 1130}; 1131 1132 1133static const struct mips_perf_event octeon_cache_map 1134 [PERF_COUNT_HW_CACHE_MAX] 1135 [PERF_COUNT_HW_CACHE_OP_MAX] 1136 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1137[C(L1D)] = { 1138 [C(OP_READ)] = { 1139 [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL }, 1140 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, 1141 }, 1142 [C(OP_WRITE)] = { 1143 [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL }, 1144 }, 1145}, 1146[C(L1I)] = { 1147 [C(OP_READ)] = { 1148 [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL }, 1149 }, 1150 [C(OP_PREFETCH)] = { 1151 [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL }, 1152 }, 1153}, 1154[C(DTLB)] = { 1155 /* 1156 * Only general DTLB misses are counted use the same event for 1157 * read and write. 1158 */ 1159 [C(OP_READ)] = { 1160 [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, 1161 }, 1162 [C(OP_WRITE)] = { 1163 [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, 1164 }, 1165}, 1166[C(ITLB)] = { 1167 [C(OP_READ)] = { 1168 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, 1169 }, 1170}, 1171}; 1172 1173static const struct mips_perf_event xlp_cache_map 1174 [PERF_COUNT_HW_CACHE_MAX] 1175 [PERF_COUNT_HW_CACHE_OP_MAX] 1176 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1177[C(L1D)] = { 1178 [C(OP_READ)] = { 1179 [C(RESULT_ACCESS)] = { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */ 1180 [C(RESULT_MISS)] = { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */ 1181 }, 1182 [C(OP_WRITE)] = { 1183 [C(RESULT_ACCESS)] = { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */ 1184 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */ 1185 }, 1186}, 1187[C(L1I)] = { 1188 [C(OP_READ)] = { 1189 [C(RESULT_ACCESS)] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */ 1190 [C(RESULT_MISS)] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */ 1191 }, 1192}, 1193[C(LL)] = { 1194 [C(OP_READ)] = { 1195 [C(RESULT_ACCESS)] = { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */ 1196 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */ 1197 }, 1198 [C(OP_WRITE)] = { 1199 [C(RESULT_ACCESS)] = { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */ 1200 [C(RESULT_MISS)] = { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */ 1201 }, 1202}, 1203[C(DTLB)] = { 1204 /* 1205 * Only general DTLB misses are counted use the same event for 1206 * read and write. 1207 */ 1208 [C(OP_READ)] = { 1209 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */ 1210 }, 1211 [C(OP_WRITE)] = { 1212 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */ 1213 }, 1214}, 1215[C(ITLB)] = { 1216 [C(OP_READ)] = { 1217 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */ 1218 }, 1219 [C(OP_WRITE)] = { 1220 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */ 1221 }, 1222}, 1223[C(BPU)] = { 1224 [C(OP_READ)] = { 1225 [C(RESULT_MISS)] = { 0x25, CNTR_ALL }, 1226 }, 1227}, 1228}; 1229 1230#ifdef CONFIG_MIPS_MT_SMP 1231static void check_and_calc_range(struct perf_event *event, 1232 const struct mips_perf_event *pev) 1233{ 1234 struct hw_perf_event *hwc = &event->hw; 1235 1236 if (event->cpu >= 0) { 1237 if (pev->range > V) { 1238 /* 1239 * The user selected an event that is processor 1240 * wide, while expecting it to be VPE wide. 1241 */ 1242 hwc->config_base |= M_TC_EN_ALL; 1243 } else { 1244 /* 1245 * FIXME: cpu_data[event->cpu].vpe_id reports 0 1246 * for both CPUs. 1247 */ 1248 hwc->config_base |= M_PERFCTL_VPEID(event->cpu); 1249 hwc->config_base |= M_TC_EN_VPE; 1250 } 1251 } else 1252 hwc->config_base |= M_TC_EN_ALL; 1253} 1254#else 1255static void check_and_calc_range(struct perf_event *event, 1256 const struct mips_perf_event *pev) 1257{ 1258} 1259#endif 1260 1261static int __hw_perf_event_init(struct perf_event *event) 1262{ 1263 struct perf_event_attr *attr = &event->attr; 1264 struct hw_perf_event *hwc = &event->hw; 1265 const struct mips_perf_event *pev; 1266 int err; 1267 1268 /* Returning MIPS event descriptor for generic perf event. */ 1269 if (PERF_TYPE_HARDWARE == event->attr.type) { 1270 if (event->attr.config >= PERF_COUNT_HW_MAX) 1271 return -EINVAL; 1272 pev = mipspmu_map_general_event(event->attr.config); 1273 } else if (PERF_TYPE_HW_CACHE == event->attr.type) { 1274 pev = mipspmu_map_cache_event(event->attr.config); 1275 } else if (PERF_TYPE_RAW == event->attr.type) { 1276 /* We are working on the global raw event. */ 1277 mutex_lock(&raw_event_mutex); 1278 pev = mipspmu.map_raw_event(event->attr.config); 1279 } else { 1280 /* The event type is not (yet) supported. */ 1281 return -EOPNOTSUPP; 1282 } 1283 1284 if (IS_ERR(pev)) { 1285 if (PERF_TYPE_RAW == event->attr.type) 1286 mutex_unlock(&raw_event_mutex); 1287 return PTR_ERR(pev); 1288 } 1289 1290 /* 1291 * We allow max flexibility on how each individual counter shared 1292 * by the single CPU operates (the mode exclusion and the range). 1293 */ 1294 hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE; 1295 1296 /* Calculate range bits and validate it. */ 1297 if (num_possible_cpus() > 1) 1298 check_and_calc_range(event, pev); 1299 1300 hwc->event_base = mipspmu_perf_event_encode(pev); 1301 if (PERF_TYPE_RAW == event->attr.type) 1302 mutex_unlock(&raw_event_mutex); 1303 1304 if (!attr->exclude_user) 1305 hwc->config_base |= M_PERFCTL_USER; 1306 if (!attr->exclude_kernel) { 1307 hwc->config_base |= M_PERFCTL_KERNEL; 1308 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */ 1309 hwc->config_base |= M_PERFCTL_EXL; 1310 } 1311 if (!attr->exclude_hv) 1312 hwc->config_base |= M_PERFCTL_SUPERVISOR; 1313 1314 hwc->config_base &= M_PERFCTL_CONFIG_MASK; 1315 /* 1316 * The event can belong to another cpu. We do not assign a local 1317 * counter for it for now. 1318 */ 1319 hwc->idx = -1; 1320 hwc->config = 0; 1321 1322 if (!hwc->sample_period) { 1323 hwc->sample_period = mipspmu.max_period; 1324 hwc->last_period = hwc->sample_period; 1325 local64_set(&hwc->period_left, hwc->sample_period); 1326 } 1327 1328 err = 0; 1329 if (event->group_leader != event) 1330 err = validate_group(event); 1331 1332 event->destroy = hw_perf_event_destroy; 1333 1334 if (err) 1335 event->destroy(event); 1336 1337 return err; 1338} 1339 1340static void pause_local_counters(void) 1341{ 1342 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 1343 int ctr = mipspmu.num_counters; 1344 unsigned long flags; 1345 1346 local_irq_save(flags); 1347 do { 1348 ctr--; 1349 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr); 1350 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] & 1351 ~M_PERFCTL_COUNT_EVENT_WHENEVER); 1352 } while (ctr > 0); 1353 local_irq_restore(flags); 1354} 1355 1356static void resume_local_counters(void) 1357{ 1358 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 1359 int ctr = mipspmu.num_counters; 1360 1361 do { 1362 ctr--; 1363 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]); 1364 } while (ctr > 0); 1365} 1366 1367static int mipsxx_pmu_handle_shared_irq(void) 1368{ 1369 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 1370 struct perf_sample_data data; 1371 unsigned int counters = mipspmu.num_counters; 1372 u64 counter; 1373 int handled = IRQ_NONE; 1374 struct pt_regs *regs; 1375 1376 if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI)) 1377 return handled; 1378 /* 1379 * First we pause the local counters, so that when we are locked 1380 * here, the counters are all paused. When it gets locked due to 1381 * perf_disable(), the timer interrupt handler will be delayed. 1382 * 1383 * See also mipsxx_pmu_start(). 1384 */ 1385 pause_local_counters(); 1386#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 1387 read_lock(&pmuint_rwlock); 1388#endif 1389 1390 regs = get_irq_regs(); 1391 1392 perf_sample_data_init(&data, 0, 0); 1393 1394 switch (counters) { 1395#define HANDLE_COUNTER(n) \ 1396 case n + 1: \ 1397 if (test_bit(n, cpuc->used_mask)) { \ 1398 counter = mipspmu.read_counter(n); \ 1399 if (counter & mipspmu.overflow) { \ 1400 handle_associated_event(cpuc, n, &data, regs); \ 1401 handled = IRQ_HANDLED; \ 1402 } \ 1403 } 1404 HANDLE_COUNTER(3) 1405 HANDLE_COUNTER(2) 1406 HANDLE_COUNTER(1) 1407 HANDLE_COUNTER(0) 1408 } 1409 1410 /* 1411 * Do all the work for the pending perf events. We can do this 1412 * in here because the performance counter interrupt is a regular 1413 * interrupt, not NMI. 1414 */ 1415 if (handled == IRQ_HANDLED) 1416 irq_work_run(); 1417 1418#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 1419 read_unlock(&pmuint_rwlock); 1420#endif 1421 resume_local_counters(); 1422 return handled; 1423} 1424 1425static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev) 1426{ 1427 return mipsxx_pmu_handle_shared_irq(); 1428} 1429 1430/* 24K */ 1431#define IS_BOTH_COUNTERS_24K_EVENT(b) \ 1432 ((b) == 0 || (b) == 1 || (b) == 11) 1433 1434/* 34K */ 1435#define IS_BOTH_COUNTERS_34K_EVENT(b) \ 1436 ((b) == 0 || (b) == 1 || (b) == 11) 1437#ifdef CONFIG_MIPS_MT_SMP 1438#define IS_RANGE_P_34K_EVENT(r, b) \ 1439 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ 1440 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \ 1441 (r) == 176 || ((b) >= 50 && (b) <= 55) || \ 1442 ((b) >= 64 && (b) <= 67)) 1443#define IS_RANGE_V_34K_EVENT(r) ((r) == 47) 1444#endif 1445 1446/* 74K */ 1447#define IS_BOTH_COUNTERS_74K_EVENT(b) \ 1448 ((b) == 0 || (b) == 1) 1449 1450/* proAptiv */ 1451#define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b) \ 1452 ((b) == 0 || (b) == 1) 1453/* P5600 */ 1454#define IS_BOTH_COUNTERS_P5600_EVENT(b) \ 1455 ((b) == 0 || (b) == 1) 1456 1457/* 1004K */ 1458#define IS_BOTH_COUNTERS_1004K_EVENT(b) \ 1459 ((b) == 0 || (b) == 1 || (b) == 11) 1460#ifdef CONFIG_MIPS_MT_SMP 1461#define IS_RANGE_P_1004K_EVENT(r, b) \ 1462 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ 1463 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \ 1464 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \ 1465 (r) == 188 || (b) == 61 || (b) == 62 || \ 1466 ((b) >= 64 && (b) <= 67)) 1467#define IS_RANGE_V_1004K_EVENT(r) ((r) == 47) 1468#endif 1469 1470/* interAptiv */ 1471#define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b) \ 1472 ((b) == 0 || (b) == 1 || (b) == 11) 1473#ifdef CONFIG_MIPS_MT_SMP 1474/* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */ 1475#define IS_RANGE_P_INTERAPTIV_EVENT(r, b) \ 1476 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ 1477 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 || \ 1478 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 && \ 1479 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 || \ 1480 ((b) >= 64 && (b) <= 67)) 1481#define IS_RANGE_V_INTERAPTIV_EVENT(r) ((r) == 47 || (r) == 175) 1482#endif 1483 1484/* BMIPS5000 */ 1485#define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b) \ 1486 ((b) == 0 || (b) == 1) 1487 1488 1489/* 1490 * For most cores the user can use 0-255 raw events, where 0-127 for the events 1491 * of even counters, and 128-255 for odd counters. Note that bit 7 is used to 1492 * indicate the even/odd bank selector. So, for example, when user wants to take 1493 * the Event Num of 15 for odd counters (by referring to the user manual), then 1494 * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F) 1495 * to be used. 1496 * 1497 * Some newer cores have even more events, in which case the user can use raw 1498 * events 0-511, where 0-255 are for the events of even counters, and 256-511 1499 * are for odd counters, so bit 8 is used to indicate the even/odd bank selector. 1500 */ 1501static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config) 1502{ 1503 /* currently most cores have 7-bit event numbers */ 1504 unsigned int raw_id = config & 0xff; 1505 unsigned int base_id = raw_id & 0x7f; 1506 1507 switch (current_cpu_type()) { 1508 case CPU_24K: 1509 if (IS_BOTH_COUNTERS_24K_EVENT(base_id)) 1510 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1511 else 1512 raw_event.cntr_mask = 1513 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1514#ifdef CONFIG_MIPS_MT_SMP 1515 /* 1516 * This is actually doing nothing. Non-multithreading 1517 * CPUs will not check and calculate the range. 1518 */ 1519 raw_event.range = P; 1520#endif 1521 break; 1522 case CPU_34K: 1523 if (IS_BOTH_COUNTERS_34K_EVENT(base_id)) 1524 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1525 else 1526 raw_event.cntr_mask = 1527 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1528#ifdef CONFIG_MIPS_MT_SMP 1529 if (IS_RANGE_P_34K_EVENT(raw_id, base_id)) 1530 raw_event.range = P; 1531 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id))) 1532 raw_event.range = V; 1533 else 1534 raw_event.range = T; 1535#endif 1536 break; 1537 case CPU_74K: 1538 case CPU_1074K: 1539 if (IS_BOTH_COUNTERS_74K_EVENT(base_id)) 1540 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1541 else 1542 raw_event.cntr_mask = 1543 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1544#ifdef CONFIG_MIPS_MT_SMP 1545 raw_event.range = P; 1546#endif 1547 break; 1548 case CPU_PROAPTIV: 1549 if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id)) 1550 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1551 else 1552 raw_event.cntr_mask = 1553 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1554#ifdef CONFIG_MIPS_MT_SMP 1555 raw_event.range = P; 1556#endif 1557 break; 1558 case CPU_P5600: 1559 /* 8-bit event numbers */ 1560 raw_id = config & 0x1ff; 1561 base_id = raw_id & 0xff; 1562 if (IS_BOTH_COUNTERS_P5600_EVENT(base_id)) 1563 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1564 else 1565 raw_event.cntr_mask = 1566 raw_id > 255 ? CNTR_ODD : CNTR_EVEN; 1567#ifdef CONFIG_MIPS_MT_SMP 1568 raw_event.range = P; 1569#endif 1570 break; 1571 case CPU_1004K: 1572 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id)) 1573 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1574 else 1575 raw_event.cntr_mask = 1576 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1577#ifdef CONFIG_MIPS_MT_SMP 1578 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id)) 1579 raw_event.range = P; 1580 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id))) 1581 raw_event.range = V; 1582 else 1583 raw_event.range = T; 1584#endif 1585 break; 1586 case CPU_INTERAPTIV: 1587 if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id)) 1588 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1589 else 1590 raw_event.cntr_mask = 1591 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1592#ifdef CONFIG_MIPS_MT_SMP 1593 if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id)) 1594 raw_event.range = P; 1595 else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id))) 1596 raw_event.range = V; 1597 else 1598 raw_event.range = T; 1599#endif 1600 break; 1601 case CPU_BMIPS5000: 1602 if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id)) 1603 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1604 else 1605 raw_event.cntr_mask = 1606 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1607 break; 1608 case CPU_LOONGSON3: 1609 raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1610 break; 1611 } 1612 1613 raw_event.event_id = base_id; 1614 1615 return &raw_event; 1616} 1617 1618static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config) 1619{ 1620 unsigned int raw_id = config & 0xff; 1621 unsigned int base_id = raw_id & 0x7f; 1622 1623 1624 raw_event.cntr_mask = CNTR_ALL; 1625 raw_event.event_id = base_id; 1626 1627 if (current_cpu_type() == CPU_CAVIUM_OCTEON2) { 1628 if (base_id > 0x42) 1629 return ERR_PTR(-EOPNOTSUPP); 1630 } else { 1631 if (base_id > 0x3a) 1632 return ERR_PTR(-EOPNOTSUPP); 1633 } 1634 1635 switch (base_id) { 1636 case 0x00: 1637 case 0x0f: 1638 case 0x1e: 1639 case 0x1f: 1640 case 0x2f: 1641 case 0x34: 1642 case 0x3b ... 0x3f: 1643 return ERR_PTR(-EOPNOTSUPP); 1644 default: 1645 break; 1646 } 1647 1648 return &raw_event; 1649} 1650 1651static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config) 1652{ 1653 unsigned int raw_id = config & 0xff; 1654 1655 /* Only 1-63 are defined */ 1656 if ((raw_id < 0x01) || (raw_id > 0x3f)) 1657 return ERR_PTR(-EOPNOTSUPP); 1658 1659 raw_event.cntr_mask = CNTR_ALL; 1660 raw_event.event_id = raw_id; 1661 1662 return &raw_event; 1663} 1664 1665static int __init 1666init_hw_perf_events(void) 1667{ 1668 int counters, irq; 1669 int counter_bits; 1670 1671 pr_info("Performance counters: "); 1672 1673 counters = n_counters(); 1674 if (counters == 0) { 1675 pr_cont("No available PMU.\n"); 1676 return -ENODEV; 1677 } 1678 1679#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS 1680 cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19); 1681 if (!cpu_has_mipsmt_pertccounters) 1682 counters = counters_total_to_per_cpu(counters); 1683#endif 1684 1685 if (get_c0_perfcount_int) 1686 irq = get_c0_perfcount_int(); 1687 else if (cp0_perfcount_irq >= 0) 1688 irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq; 1689 else 1690 irq = -1; 1691 1692 mipspmu.map_raw_event = mipsxx_pmu_map_raw_event; 1693 1694 switch (current_cpu_type()) { 1695 case CPU_24K: 1696 mipspmu.name = "mips/24K"; 1697 mipspmu.general_event_map = &mipsxxcore_event_map; 1698 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1699 break; 1700 case CPU_34K: 1701 mipspmu.name = "mips/34K"; 1702 mipspmu.general_event_map = &mipsxxcore_event_map; 1703 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1704 break; 1705 case CPU_74K: 1706 mipspmu.name = "mips/74K"; 1707 mipspmu.general_event_map = &mipsxxcore_event_map2; 1708 mipspmu.cache_event_map = &mipsxxcore_cache_map2; 1709 break; 1710 case CPU_PROAPTIV: 1711 mipspmu.name = "mips/proAptiv"; 1712 mipspmu.general_event_map = &mipsxxcore_event_map2; 1713 mipspmu.cache_event_map = &mipsxxcore_cache_map2; 1714 break; 1715 case CPU_P5600: 1716 mipspmu.name = "mips/P5600"; 1717 mipspmu.general_event_map = &mipsxxcore_event_map2; 1718 mipspmu.cache_event_map = &mipsxxcore_cache_map2; 1719 break; 1720 case CPU_1004K: 1721 mipspmu.name = "mips/1004K"; 1722 mipspmu.general_event_map = &mipsxxcore_event_map; 1723 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1724 break; 1725 case CPU_1074K: 1726 mipspmu.name = "mips/1074K"; 1727 mipspmu.general_event_map = &mipsxxcore_event_map; 1728 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1729 break; 1730 case CPU_INTERAPTIV: 1731 mipspmu.name = "mips/interAptiv"; 1732 mipspmu.general_event_map = &mipsxxcore_event_map; 1733 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1734 break; 1735 case CPU_LOONGSON1: 1736 mipspmu.name = "mips/loongson1"; 1737 mipspmu.general_event_map = &mipsxxcore_event_map; 1738 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1739 break; 1740 case CPU_LOONGSON3: 1741 mipspmu.name = "mips/loongson3"; 1742 mipspmu.general_event_map = &loongson3_event_map; 1743 mipspmu.cache_event_map = &loongson3_cache_map; 1744 break; 1745 case CPU_CAVIUM_OCTEON: 1746 case CPU_CAVIUM_OCTEON_PLUS: 1747 case CPU_CAVIUM_OCTEON2: 1748 mipspmu.name = "octeon"; 1749 mipspmu.general_event_map = &octeon_event_map; 1750 mipspmu.cache_event_map = &octeon_cache_map; 1751 mipspmu.map_raw_event = octeon_pmu_map_raw_event; 1752 break; 1753 case CPU_BMIPS5000: 1754 mipspmu.name = "BMIPS5000"; 1755 mipspmu.general_event_map = &bmips5000_event_map; 1756 mipspmu.cache_event_map = &bmips5000_cache_map; 1757 break; 1758 case CPU_XLP: 1759 mipspmu.name = "xlp"; 1760 mipspmu.general_event_map = &xlp_event_map; 1761 mipspmu.cache_event_map = &xlp_cache_map; 1762 mipspmu.map_raw_event = xlp_pmu_map_raw_event; 1763 break; 1764 default: 1765 pr_cont("Either hardware does not support performance " 1766 "counters, or not yet implemented.\n"); 1767 return -ENODEV; 1768 } 1769 1770 mipspmu.num_counters = counters; 1771 mipspmu.irq = irq; 1772 1773 if (read_c0_perfctrl0() & M_PERFCTL_WIDE) { 1774 mipspmu.max_period = (1ULL << 63) - 1; 1775 mipspmu.valid_count = (1ULL << 63) - 1; 1776 mipspmu.overflow = 1ULL << 63; 1777 mipspmu.read_counter = mipsxx_pmu_read_counter_64; 1778 mipspmu.write_counter = mipsxx_pmu_write_counter_64; 1779 counter_bits = 64; 1780 } else { 1781 mipspmu.max_period = (1ULL << 31) - 1; 1782 mipspmu.valid_count = (1ULL << 31) - 1; 1783 mipspmu.overflow = 1ULL << 31; 1784 mipspmu.read_counter = mipsxx_pmu_read_counter; 1785 mipspmu.write_counter = mipsxx_pmu_write_counter; 1786 counter_bits = 32; 1787 } 1788 1789 on_each_cpu(reset_counters, (void *)(long)counters, 1); 1790 1791 pr_cont("%s PMU enabled, %d %d-bit counters available to each " 1792 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq, 1793 irq < 0 ? " (share with timer interrupt)" : ""); 1794 1795 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW); 1796 1797 return 0; 1798} 1799early_initcall(init_hw_perf_events); 1800