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 32 struct cpu_hw_events { 33 /* Array of events on this cpu. */ 34 struct perf_event *events[MIPS_MAX_HWEVENTS]; 35 36 /* 37 * Set the bit (indexed by the counter number) when the counter 38 * is used for an event. 39 */ 40 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)]; 41 42 /* 43 * Software copy of the control register for each performance counter. 44 * MIPS CPUs vary in performance counters. They use this differently, 45 * and even may not use it. 46 */ 47 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS]; 48 }; 49 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = { 50 .saved_ctrl = {0}, 51 }; 52 53 /* The description of MIPS performance events. */ 54 struct mips_perf_event { 55 unsigned int event_id; 56 /* 57 * MIPS performance counters are indexed starting from 0. 58 * CNTR_EVEN indicates the indexes of the counters to be used are 59 * even numbers. 60 */ 61 unsigned int cntr_mask; 62 #define CNTR_EVEN 0x55555555 63 #define CNTR_ODD 0xaaaaaaaa 64 #define CNTR_ALL 0xffffffff 65 #ifdef CONFIG_MIPS_MT_SMP 66 enum { 67 T = 0, 68 V = 1, 69 P = 2, 70 } range; 71 #else 72 #define T 73 #define V 74 #define P 75 #endif 76 }; 77 78 static struct mips_perf_event raw_event; 79 static DEFINE_MUTEX(raw_event_mutex); 80 81 #define UNSUPPORTED_PERF_EVENT_ID 0xffffffff 82 #define C(x) PERF_COUNT_HW_CACHE_##x 83 84 struct mips_pmu { 85 u64 max_period; 86 u64 valid_count; 87 u64 overflow; 88 const char *name; 89 int irq; 90 u64 (*read_counter)(unsigned int idx); 91 void (*write_counter)(unsigned int idx, u64 val); 92 const struct mips_perf_event *(*map_raw_event)(u64 config); 93 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX]; 94 const struct mips_perf_event (*cache_event_map) 95 [PERF_COUNT_HW_CACHE_MAX] 96 [PERF_COUNT_HW_CACHE_OP_MAX] 97 [PERF_COUNT_HW_CACHE_RESULT_MAX]; 98 unsigned int num_counters; 99 }; 100 101 static struct mips_pmu mipspmu; 102 103 #define M_CONFIG1_PC (1 << 4) 104 105 #define M_PERFCTL_EXL (1 << 0) 106 #define M_PERFCTL_KERNEL (1 << 1) 107 #define M_PERFCTL_SUPERVISOR (1 << 2) 108 #define M_PERFCTL_USER (1 << 3) 109 #define M_PERFCTL_INTERRUPT_ENABLE (1 << 4) 110 #define M_PERFCTL_EVENT(event) (((event) & 0x3ff) << 5) 111 #define M_PERFCTL_VPEID(vpe) ((vpe) << 16) 112 #define M_PERFCTL_MT_EN(filter) ((filter) << 20) 113 #define M_TC_EN_ALL M_PERFCTL_MT_EN(0) 114 #define M_TC_EN_VPE M_PERFCTL_MT_EN(1) 115 #define M_TC_EN_TC M_PERFCTL_MT_EN(2) 116 #define M_PERFCTL_TCID(tcid) ((tcid) << 22) 117 #define M_PERFCTL_WIDE (1 << 30) 118 #define M_PERFCTL_MORE (1 << 31) 119 120 #define M_PERFCTL_COUNT_EVENT_WHENEVER (M_PERFCTL_EXL | \ 121 M_PERFCTL_KERNEL | \ 122 M_PERFCTL_USER | \ 123 M_PERFCTL_SUPERVISOR | \ 124 M_PERFCTL_INTERRUPT_ENABLE) 125 126 #ifdef CONFIG_MIPS_MT_SMP 127 #define M_PERFCTL_CONFIG_MASK 0x3fff801f 128 #else 129 #define M_PERFCTL_CONFIG_MASK 0x1f 130 #endif 131 #define M_PERFCTL_EVENT_MASK 0xfe0 132 133 134 #ifdef CONFIG_MIPS_MT_SMP 135 static int cpu_has_mipsmt_pertccounters; 136 137 static DEFINE_RWLOCK(pmuint_rwlock); 138 139 /* 140 * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because 141 * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs. 142 */ 143 #if defined(CONFIG_HW_PERF_EVENTS) 144 #define vpe_id() (cpu_has_mipsmt_pertccounters ? \ 145 0 : smp_processor_id()) 146 #else 147 #define vpe_id() (cpu_has_mipsmt_pertccounters ? \ 148 0 : cpu_data[smp_processor_id()].vpe_id) 149 #endif 150 151 /* Copied from op_model_mipsxx.c */ 152 static unsigned int vpe_shift(void) 153 { 154 if (num_possible_cpus() > 1) 155 return 1; 156 157 return 0; 158 } 159 160 static unsigned int counters_total_to_per_cpu(unsigned int counters) 161 { 162 return counters >> vpe_shift(); 163 } 164 165 static unsigned int counters_per_cpu_to_total(unsigned int counters) 166 { 167 return counters << vpe_shift(); 168 } 169 170 #else /* !CONFIG_MIPS_MT_SMP */ 171 #define vpe_id() 0 172 173 #endif /* CONFIG_MIPS_MT_SMP */ 174 175 static void resume_local_counters(void); 176 static void pause_local_counters(void); 177 static irqreturn_t mipsxx_pmu_handle_irq(int, void *); 178 static int mipsxx_pmu_handle_shared_irq(void); 179 180 static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx) 181 { 182 if (vpe_id() == 1) 183 idx = (idx + 2) & 3; 184 return idx; 185 } 186 187 static u64 mipsxx_pmu_read_counter(unsigned int idx) 188 { 189 idx = mipsxx_pmu_swizzle_perf_idx(idx); 190 191 switch (idx) { 192 case 0: 193 /* 194 * The counters are unsigned, we must cast to truncate 195 * off the high bits. 196 */ 197 return (u32)read_c0_perfcntr0(); 198 case 1: 199 return (u32)read_c0_perfcntr1(); 200 case 2: 201 return (u32)read_c0_perfcntr2(); 202 case 3: 203 return (u32)read_c0_perfcntr3(); 204 default: 205 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 206 return 0; 207 } 208 } 209 210 static u64 mipsxx_pmu_read_counter_64(unsigned int idx) 211 { 212 idx = mipsxx_pmu_swizzle_perf_idx(idx); 213 214 switch (idx) { 215 case 0: 216 return read_c0_perfcntr0_64(); 217 case 1: 218 return read_c0_perfcntr1_64(); 219 case 2: 220 return read_c0_perfcntr2_64(); 221 case 3: 222 return read_c0_perfcntr3_64(); 223 default: 224 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 225 return 0; 226 } 227 } 228 229 static void mipsxx_pmu_write_counter(unsigned int idx, u64 val) 230 { 231 idx = mipsxx_pmu_swizzle_perf_idx(idx); 232 233 switch (idx) { 234 case 0: 235 write_c0_perfcntr0(val); 236 return; 237 case 1: 238 write_c0_perfcntr1(val); 239 return; 240 case 2: 241 write_c0_perfcntr2(val); 242 return; 243 case 3: 244 write_c0_perfcntr3(val); 245 return; 246 } 247 } 248 249 static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val) 250 { 251 idx = mipsxx_pmu_swizzle_perf_idx(idx); 252 253 switch (idx) { 254 case 0: 255 write_c0_perfcntr0_64(val); 256 return; 257 case 1: 258 write_c0_perfcntr1_64(val); 259 return; 260 case 2: 261 write_c0_perfcntr2_64(val); 262 return; 263 case 3: 264 write_c0_perfcntr3_64(val); 265 return; 266 } 267 } 268 269 static unsigned int mipsxx_pmu_read_control(unsigned int idx) 270 { 271 idx = mipsxx_pmu_swizzle_perf_idx(idx); 272 273 switch (idx) { 274 case 0: 275 return read_c0_perfctrl0(); 276 case 1: 277 return read_c0_perfctrl1(); 278 case 2: 279 return read_c0_perfctrl2(); 280 case 3: 281 return read_c0_perfctrl3(); 282 default: 283 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx); 284 return 0; 285 } 286 } 287 288 static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val) 289 { 290 idx = mipsxx_pmu_swizzle_perf_idx(idx); 291 292 switch (idx) { 293 case 0: 294 write_c0_perfctrl0(val); 295 return; 296 case 1: 297 write_c0_perfctrl1(val); 298 return; 299 case 2: 300 write_c0_perfctrl2(val); 301 return; 302 case 3: 303 write_c0_perfctrl3(val); 304 return; 305 } 306 } 307 308 static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc, 309 struct hw_perf_event *hwc) 310 { 311 int i; 312 313 /* 314 * We only need to care the counter mask. The range has been 315 * checked definitely. 316 */ 317 unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff; 318 319 for (i = mipspmu.num_counters - 1; i >= 0; i--) { 320 /* 321 * Note that some MIPS perf events can be counted by both 322 * even and odd counters, wheresas many other are only by 323 * even _or_ odd counters. This introduces an issue that 324 * when the former kind of event takes the counter the 325 * latter kind of event wants to use, then the "counter 326 * allocation" for the latter event will fail. In fact if 327 * they can be dynamically swapped, they both feel happy. 328 * But here we leave this issue alone for now. 329 */ 330 if (test_bit(i, &cntr_mask) && 331 !test_and_set_bit(i, cpuc->used_mask)) 332 return i; 333 } 334 335 return -EAGAIN; 336 } 337 338 static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx) 339 { 340 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 341 342 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 343 344 cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) | 345 (evt->config_base & M_PERFCTL_CONFIG_MASK) | 346 /* Make sure interrupt enabled. */ 347 M_PERFCTL_INTERRUPT_ENABLE; 348 /* 349 * We do not actually let the counter run. Leave it until start(). 350 */ 351 } 352 353 static void mipsxx_pmu_disable_event(int idx) 354 { 355 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 356 unsigned long flags; 357 358 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 359 360 local_irq_save(flags); 361 cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) & 362 ~M_PERFCTL_COUNT_EVENT_WHENEVER; 363 mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]); 364 local_irq_restore(flags); 365 } 366 367 static int mipspmu_event_set_period(struct perf_event *event, 368 struct hw_perf_event *hwc, 369 int idx) 370 { 371 u64 left = local64_read(&hwc->period_left); 372 u64 period = hwc->sample_period; 373 int ret = 0; 374 375 if (unlikely((left + period) & (1ULL << 63))) { 376 /* left underflowed by more than period. */ 377 left = period; 378 local64_set(&hwc->period_left, left); 379 hwc->last_period = period; 380 ret = 1; 381 } else if (unlikely((left + period) <= period)) { 382 /* left underflowed by less than period. */ 383 left += period; 384 local64_set(&hwc->period_left, left); 385 hwc->last_period = period; 386 ret = 1; 387 } 388 389 if (left > mipspmu.max_period) { 390 left = mipspmu.max_period; 391 local64_set(&hwc->period_left, left); 392 } 393 394 local64_set(&hwc->prev_count, mipspmu.overflow - left); 395 396 mipspmu.write_counter(idx, mipspmu.overflow - left); 397 398 perf_event_update_userpage(event); 399 400 return ret; 401 } 402 403 static void mipspmu_event_update(struct perf_event *event, 404 struct hw_perf_event *hwc, 405 int idx) 406 { 407 u64 prev_raw_count, new_raw_count; 408 u64 delta; 409 410 again: 411 prev_raw_count = local64_read(&hwc->prev_count); 412 new_raw_count = mipspmu.read_counter(idx); 413 414 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, 415 new_raw_count) != prev_raw_count) 416 goto again; 417 418 delta = new_raw_count - prev_raw_count; 419 420 local64_add(delta, &event->count); 421 local64_sub(delta, &hwc->period_left); 422 } 423 424 static void mipspmu_start(struct perf_event *event, int flags) 425 { 426 struct hw_perf_event *hwc = &event->hw; 427 428 if (flags & PERF_EF_RELOAD) 429 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); 430 431 hwc->state = 0; 432 433 /* Set the period for the event. */ 434 mipspmu_event_set_period(event, hwc, hwc->idx); 435 436 /* Enable the event. */ 437 mipsxx_pmu_enable_event(hwc, hwc->idx); 438 } 439 440 static void mipspmu_stop(struct perf_event *event, int flags) 441 { 442 struct hw_perf_event *hwc = &event->hw; 443 444 if (!(hwc->state & PERF_HES_STOPPED)) { 445 /* We are working on a local event. */ 446 mipsxx_pmu_disable_event(hwc->idx); 447 barrier(); 448 mipspmu_event_update(event, hwc, hwc->idx); 449 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; 450 } 451 } 452 453 static int mipspmu_add(struct perf_event *event, int flags) 454 { 455 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 456 struct hw_perf_event *hwc = &event->hw; 457 int idx; 458 int err = 0; 459 460 perf_pmu_disable(event->pmu); 461 462 /* To look for a free counter for this event. */ 463 idx = mipsxx_pmu_alloc_counter(cpuc, hwc); 464 if (idx < 0) { 465 err = idx; 466 goto out; 467 } 468 469 /* 470 * If there is an event in the counter we are going to use then 471 * make sure it is disabled. 472 */ 473 event->hw.idx = idx; 474 mipsxx_pmu_disable_event(idx); 475 cpuc->events[idx] = event; 476 477 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; 478 if (flags & PERF_EF_START) 479 mipspmu_start(event, PERF_EF_RELOAD); 480 481 /* Propagate our changes to the userspace mapping. */ 482 perf_event_update_userpage(event); 483 484 out: 485 perf_pmu_enable(event->pmu); 486 return err; 487 } 488 489 static void mipspmu_del(struct perf_event *event, int flags) 490 { 491 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 492 struct hw_perf_event *hwc = &event->hw; 493 int idx = hwc->idx; 494 495 WARN_ON(idx < 0 || idx >= mipspmu.num_counters); 496 497 mipspmu_stop(event, PERF_EF_UPDATE); 498 cpuc->events[idx] = NULL; 499 clear_bit(idx, cpuc->used_mask); 500 501 perf_event_update_userpage(event); 502 } 503 504 static void mipspmu_read(struct perf_event *event) 505 { 506 struct hw_perf_event *hwc = &event->hw; 507 508 /* Don't read disabled counters! */ 509 if (hwc->idx < 0) 510 return; 511 512 mipspmu_event_update(event, hwc, hwc->idx); 513 } 514 515 static void mipspmu_enable(struct pmu *pmu) 516 { 517 #ifdef CONFIG_MIPS_MT_SMP 518 write_unlock(&pmuint_rwlock); 519 #endif 520 resume_local_counters(); 521 } 522 523 /* 524 * MIPS performance counters can be per-TC. The control registers can 525 * not be directly accessed accross CPUs. Hence if we want to do global 526 * control, we need cross CPU calls. on_each_cpu() can help us, but we 527 * can not make sure this function is called with interrupts enabled. So 528 * here we pause local counters and then grab a rwlock and leave the 529 * counters on other CPUs alone. If any counter interrupt raises while 530 * we own the write lock, simply pause local counters on that CPU and 531 * spin in the handler. Also we know we won't be switched to another 532 * CPU after pausing local counters and before grabbing the lock. 533 */ 534 static void mipspmu_disable(struct pmu *pmu) 535 { 536 pause_local_counters(); 537 #ifdef CONFIG_MIPS_MT_SMP 538 write_lock(&pmuint_rwlock); 539 #endif 540 } 541 542 static atomic_t active_events = ATOMIC_INIT(0); 543 static DEFINE_MUTEX(pmu_reserve_mutex); 544 static int (*save_perf_irq)(void); 545 546 static int mipspmu_get_irq(void) 547 { 548 int err; 549 550 if (mipspmu.irq >= 0) { 551 /* Request my own irq handler. */ 552 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq, 553 IRQF_PERCPU | IRQF_NOBALANCING, 554 "mips_perf_pmu", NULL); 555 if (err) { 556 pr_warning("Unable to request IRQ%d for MIPS " 557 "performance counters!\n", mipspmu.irq); 558 } 559 } else if (cp0_perfcount_irq < 0) { 560 /* 561 * We are sharing the irq number with the timer interrupt. 562 */ 563 save_perf_irq = perf_irq; 564 perf_irq = mipsxx_pmu_handle_shared_irq; 565 err = 0; 566 } else { 567 pr_warning("The platform hasn't properly defined its " 568 "interrupt controller.\n"); 569 err = -ENOENT; 570 } 571 572 return err; 573 } 574 575 static void mipspmu_free_irq(void) 576 { 577 if (mipspmu.irq >= 0) 578 free_irq(mipspmu.irq, NULL); 579 else if (cp0_perfcount_irq < 0) 580 perf_irq = save_perf_irq; 581 } 582 583 /* 584 * mipsxx/rm9000/loongson2 have different performance counters, they have 585 * specific low-level init routines. 586 */ 587 static void reset_counters(void *arg); 588 static int __hw_perf_event_init(struct perf_event *event); 589 590 static void hw_perf_event_destroy(struct perf_event *event) 591 { 592 if (atomic_dec_and_mutex_lock(&active_events, 593 &pmu_reserve_mutex)) { 594 /* 595 * We must not call the destroy function with interrupts 596 * disabled. 597 */ 598 on_each_cpu(reset_counters, 599 (void *)(long)mipspmu.num_counters, 1); 600 mipspmu_free_irq(); 601 mutex_unlock(&pmu_reserve_mutex); 602 } 603 } 604 605 static int mipspmu_event_init(struct perf_event *event) 606 { 607 int err = 0; 608 609 /* does not support taken branch sampling */ 610 if (has_branch_stack(event)) 611 return -EOPNOTSUPP; 612 613 switch (event->attr.type) { 614 case PERF_TYPE_RAW: 615 case PERF_TYPE_HARDWARE: 616 case PERF_TYPE_HW_CACHE: 617 break; 618 619 default: 620 return -ENOENT; 621 } 622 623 if (event->cpu >= nr_cpumask_bits || 624 (event->cpu >= 0 && !cpu_online(event->cpu))) 625 return -ENODEV; 626 627 if (!atomic_inc_not_zero(&active_events)) { 628 mutex_lock(&pmu_reserve_mutex); 629 if (atomic_read(&active_events) == 0) 630 err = mipspmu_get_irq(); 631 632 if (!err) 633 atomic_inc(&active_events); 634 mutex_unlock(&pmu_reserve_mutex); 635 } 636 637 if (err) 638 return err; 639 640 return __hw_perf_event_init(event); 641 } 642 643 static struct pmu pmu = { 644 .pmu_enable = mipspmu_enable, 645 .pmu_disable = mipspmu_disable, 646 .event_init = mipspmu_event_init, 647 .add = mipspmu_add, 648 .del = mipspmu_del, 649 .start = mipspmu_start, 650 .stop = mipspmu_stop, 651 .read = mipspmu_read, 652 }; 653 654 static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev) 655 { 656 /* 657 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for 658 * event_id. 659 */ 660 #ifdef CONFIG_MIPS_MT_SMP 661 return ((unsigned int)pev->range << 24) | 662 (pev->cntr_mask & 0xffff00) | 663 (pev->event_id & 0xff); 664 #else 665 return (pev->cntr_mask & 0xffff00) | 666 (pev->event_id & 0xff); 667 #endif 668 } 669 670 static const struct mips_perf_event *mipspmu_map_general_event(int idx) 671 { 672 const struct mips_perf_event *pev; 673 674 pev = ((*mipspmu.general_event_map)[idx].event_id == 675 UNSUPPORTED_PERF_EVENT_ID ? ERR_PTR(-EOPNOTSUPP) : 676 &(*mipspmu.general_event_map)[idx]); 677 678 return pev; 679 } 680 681 static const struct mips_perf_event *mipspmu_map_cache_event(u64 config) 682 { 683 unsigned int cache_type, cache_op, cache_result; 684 const struct mips_perf_event *pev; 685 686 cache_type = (config >> 0) & 0xff; 687 if (cache_type >= PERF_COUNT_HW_CACHE_MAX) 688 return ERR_PTR(-EINVAL); 689 690 cache_op = (config >> 8) & 0xff; 691 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) 692 return ERR_PTR(-EINVAL); 693 694 cache_result = (config >> 16) & 0xff; 695 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 696 return ERR_PTR(-EINVAL); 697 698 pev = &((*mipspmu.cache_event_map) 699 [cache_type] 700 [cache_op] 701 [cache_result]); 702 703 if (pev->event_id == UNSUPPORTED_PERF_EVENT_ID) 704 return ERR_PTR(-EOPNOTSUPP); 705 706 return pev; 707 708 } 709 710 static int validate_group(struct perf_event *event) 711 { 712 struct perf_event *sibling, *leader = event->group_leader; 713 struct cpu_hw_events fake_cpuc; 714 715 memset(&fake_cpuc, 0, sizeof(fake_cpuc)); 716 717 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0) 718 return -EINVAL; 719 720 list_for_each_entry(sibling, &leader->sibling_list, group_entry) { 721 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0) 722 return -EINVAL; 723 } 724 725 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0) 726 return -EINVAL; 727 728 return 0; 729 } 730 731 /* This is needed by specific irq handlers in perf_event_*.c */ 732 static void handle_associated_event(struct cpu_hw_events *cpuc, 733 int idx, struct perf_sample_data *data, 734 struct pt_regs *regs) 735 { 736 struct perf_event *event = cpuc->events[idx]; 737 struct hw_perf_event *hwc = &event->hw; 738 739 mipspmu_event_update(event, hwc, idx); 740 data->period = event->hw.last_period; 741 if (!mipspmu_event_set_period(event, hwc, idx)) 742 return; 743 744 if (perf_event_overflow(event, data, regs)) 745 mipsxx_pmu_disable_event(idx); 746 } 747 748 749 static int __n_counters(void) 750 { 751 if (!(read_c0_config1() & M_CONFIG1_PC)) 752 return 0; 753 if (!(read_c0_perfctrl0() & M_PERFCTL_MORE)) 754 return 1; 755 if (!(read_c0_perfctrl1() & M_PERFCTL_MORE)) 756 return 2; 757 if (!(read_c0_perfctrl2() & M_PERFCTL_MORE)) 758 return 3; 759 760 return 4; 761 } 762 763 static int n_counters(void) 764 { 765 int counters; 766 767 switch (current_cpu_type()) { 768 case CPU_R10000: 769 counters = 2; 770 break; 771 772 case CPU_R12000: 773 case CPU_R14000: 774 counters = 4; 775 break; 776 777 default: 778 counters = __n_counters(); 779 } 780 781 return counters; 782 } 783 784 static void reset_counters(void *arg) 785 { 786 int counters = (int)(long)arg; 787 switch (counters) { 788 case 4: 789 mipsxx_pmu_write_control(3, 0); 790 mipspmu.write_counter(3, 0); 791 case 3: 792 mipsxx_pmu_write_control(2, 0); 793 mipspmu.write_counter(2, 0); 794 case 2: 795 mipsxx_pmu_write_control(1, 0); 796 mipspmu.write_counter(1, 0); 797 case 1: 798 mipsxx_pmu_write_control(0, 0); 799 mipspmu.write_counter(0, 0); 800 } 801 } 802 803 /* 24K/34K/1004K cores can share the same event map. */ 804 static const struct mips_perf_event mipsxxcore_event_map 805 [PERF_COUNT_HW_MAX] = { 806 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, 807 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, 808 [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID }, 809 [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID }, 810 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T }, 811 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T }, 812 [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID }, 813 }; 814 815 /* 74K core has different branch event code. */ 816 static const struct mips_perf_event mipsxx74Kcore_event_map 817 [PERF_COUNT_HW_MAX] = { 818 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P }, 819 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T }, 820 [PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID }, 821 [PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID }, 822 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T }, 823 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T }, 824 [PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID }, 825 }; 826 827 static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = { 828 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL }, 829 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL }, 830 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL }, 831 [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL }, 832 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL }, 833 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL }, 834 [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL }, 835 }; 836 837 /* 24K/34K/1004K cores can share the same cache event map. */ 838 static const struct mips_perf_event mipsxxcore_cache_map 839 [PERF_COUNT_HW_CACHE_MAX] 840 [PERF_COUNT_HW_CACHE_OP_MAX] 841 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 842 [C(L1D)] = { 843 /* 844 * Like some other architectures (e.g. ARM), the performance 845 * counters don't differentiate between read and write 846 * accesses/misses, so this isn't strictly correct, but it's the 847 * best we can do. Writes and reads get combined. 848 */ 849 [C(OP_READ)] = { 850 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, 851 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, 852 }, 853 [C(OP_WRITE)] = { 854 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T }, 855 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T }, 856 }, 857 [C(OP_PREFETCH)] = { 858 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 859 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 860 }, 861 }, 862 [C(L1I)] = { 863 [C(OP_READ)] = { 864 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, 865 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, 866 }, 867 [C(OP_WRITE)] = { 868 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T }, 869 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T }, 870 }, 871 [C(OP_PREFETCH)] = { 872 [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T }, 873 /* 874 * Note that MIPS has only "hit" events countable for 875 * the prefetch operation. 876 */ 877 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 878 }, 879 }, 880 [C(LL)] = { 881 [C(OP_READ)] = { 882 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, 883 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, 884 }, 885 [C(OP_WRITE)] = { 886 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P }, 887 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P }, 888 }, 889 [C(OP_PREFETCH)] = { 890 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 891 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 892 }, 893 }, 894 [C(DTLB)] = { 895 [C(OP_READ)] = { 896 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 897 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 898 }, 899 [C(OP_WRITE)] = { 900 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 901 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 902 }, 903 [C(OP_PREFETCH)] = { 904 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 905 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 906 }, 907 }, 908 [C(ITLB)] = { 909 [C(OP_READ)] = { 910 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, 911 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, 912 }, 913 [C(OP_WRITE)] = { 914 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T }, 915 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T }, 916 }, 917 [C(OP_PREFETCH)] = { 918 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 919 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 920 }, 921 }, 922 [C(BPU)] = { 923 /* Using the same code for *HW_BRANCH* */ 924 [C(OP_READ)] = { 925 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, 926 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 927 }, 928 [C(OP_WRITE)] = { 929 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T }, 930 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T }, 931 }, 932 [C(OP_PREFETCH)] = { 933 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 934 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 935 }, 936 }, 937 [C(NODE)] = { 938 [C(OP_READ)] = { 939 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 940 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 941 }, 942 [C(OP_WRITE)] = { 943 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 944 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 945 }, 946 [C(OP_PREFETCH)] = { 947 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 948 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 949 }, 950 }, 951 }; 952 953 /* 74K core has completely different cache event map. */ 954 static const struct mips_perf_event mipsxx74Kcore_cache_map 955 [PERF_COUNT_HW_CACHE_MAX] 956 [PERF_COUNT_HW_CACHE_OP_MAX] 957 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 958 [C(L1D)] = { 959 /* 960 * Like some other architectures (e.g. ARM), the performance 961 * counters don't differentiate between read and write 962 * accesses/misses, so this isn't strictly correct, but it's the 963 * best we can do. Writes and reads get combined. 964 */ 965 [C(OP_READ)] = { 966 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, 967 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, 968 }, 969 [C(OP_WRITE)] = { 970 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T }, 971 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T }, 972 }, 973 [C(OP_PREFETCH)] = { 974 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 975 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 976 }, 977 }, 978 [C(L1I)] = { 979 [C(OP_READ)] = { 980 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 981 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 982 }, 983 [C(OP_WRITE)] = { 984 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T }, 985 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T }, 986 }, 987 [C(OP_PREFETCH)] = { 988 [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T }, 989 /* 990 * Note that MIPS has only "hit" events countable for 991 * the prefetch operation. 992 */ 993 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 994 }, 995 }, 996 [C(LL)] = { 997 [C(OP_READ)] = { 998 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, 999 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN | CNTR_ODD, P }, 1000 }, 1001 [C(OP_WRITE)] = { 1002 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P }, 1003 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN | CNTR_ODD, P }, 1004 }, 1005 [C(OP_PREFETCH)] = { 1006 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1007 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1008 }, 1009 }, 1010 [C(DTLB)] = { 1011 /* 74K core does not have specific DTLB events. */ 1012 [C(OP_READ)] = { 1013 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1014 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1015 }, 1016 [C(OP_WRITE)] = { 1017 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1018 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1019 }, 1020 [C(OP_PREFETCH)] = { 1021 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1022 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1023 }, 1024 }, 1025 [C(ITLB)] = { 1026 [C(OP_READ)] = { 1027 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, 1028 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, 1029 }, 1030 [C(OP_WRITE)] = { 1031 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T }, 1032 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T }, 1033 }, 1034 [C(OP_PREFETCH)] = { 1035 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1036 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1037 }, 1038 }, 1039 [C(BPU)] = { 1040 /* Using the same code for *HW_BRANCH* */ 1041 [C(OP_READ)] = { 1042 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, 1043 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, 1044 }, 1045 [C(OP_WRITE)] = { 1046 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T }, 1047 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T }, 1048 }, 1049 [C(OP_PREFETCH)] = { 1050 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1051 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1052 }, 1053 }, 1054 [C(NODE)] = { 1055 [C(OP_READ)] = { 1056 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1057 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1058 }, 1059 [C(OP_WRITE)] = { 1060 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1061 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1062 }, 1063 [C(OP_PREFETCH)] = { 1064 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1065 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1066 }, 1067 }, 1068 }; 1069 1070 1071 static const struct mips_perf_event octeon_cache_map 1072 [PERF_COUNT_HW_CACHE_MAX] 1073 [PERF_COUNT_HW_CACHE_OP_MAX] 1074 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1075 [C(L1D)] = { 1076 [C(OP_READ)] = { 1077 [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL }, 1078 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, 1079 }, 1080 [C(OP_WRITE)] = { 1081 [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL }, 1082 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1083 }, 1084 [C(OP_PREFETCH)] = { 1085 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1086 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1087 }, 1088 }, 1089 [C(L1I)] = { 1090 [C(OP_READ)] = { 1091 [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL }, 1092 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1093 }, 1094 [C(OP_WRITE)] = { 1095 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1096 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1097 }, 1098 [C(OP_PREFETCH)] = { 1099 [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL }, 1100 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1101 }, 1102 }, 1103 [C(LL)] = { 1104 [C(OP_READ)] = { 1105 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1106 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1107 }, 1108 [C(OP_WRITE)] = { 1109 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1110 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1111 }, 1112 [C(OP_PREFETCH)] = { 1113 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1114 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1115 }, 1116 }, 1117 [C(DTLB)] = { 1118 /* 1119 * Only general DTLB misses are counted use the same event for 1120 * read and write. 1121 */ 1122 [C(OP_READ)] = { 1123 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1124 [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, 1125 }, 1126 [C(OP_WRITE)] = { 1127 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1128 [C(RESULT_MISS)] = { 0x35, CNTR_ALL }, 1129 }, 1130 [C(OP_PREFETCH)] = { 1131 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1132 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1133 }, 1134 }, 1135 [C(ITLB)] = { 1136 [C(OP_READ)] = { 1137 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1138 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, 1139 }, 1140 [C(OP_WRITE)] = { 1141 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1142 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1143 }, 1144 [C(OP_PREFETCH)] = { 1145 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1146 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1147 }, 1148 }, 1149 [C(BPU)] = { 1150 /* Using the same code for *HW_BRANCH* */ 1151 [C(OP_READ)] = { 1152 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1153 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1154 }, 1155 [C(OP_WRITE)] = { 1156 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1157 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1158 }, 1159 [C(OP_PREFETCH)] = { 1160 [C(RESULT_ACCESS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1161 [C(RESULT_MISS)] = { UNSUPPORTED_PERF_EVENT_ID }, 1162 }, 1163 }, 1164 }; 1165 1166 #ifdef CONFIG_MIPS_MT_SMP 1167 static void check_and_calc_range(struct perf_event *event, 1168 const struct mips_perf_event *pev) 1169 { 1170 struct hw_perf_event *hwc = &event->hw; 1171 1172 if (event->cpu >= 0) { 1173 if (pev->range > V) { 1174 /* 1175 * The user selected an event that is processor 1176 * wide, while expecting it to be VPE wide. 1177 */ 1178 hwc->config_base |= M_TC_EN_ALL; 1179 } else { 1180 /* 1181 * FIXME: cpu_data[event->cpu].vpe_id reports 0 1182 * for both CPUs. 1183 */ 1184 hwc->config_base |= M_PERFCTL_VPEID(event->cpu); 1185 hwc->config_base |= M_TC_EN_VPE; 1186 } 1187 } else 1188 hwc->config_base |= M_TC_EN_ALL; 1189 } 1190 #else 1191 static void check_and_calc_range(struct perf_event *event, 1192 const struct mips_perf_event *pev) 1193 { 1194 } 1195 #endif 1196 1197 static int __hw_perf_event_init(struct perf_event *event) 1198 { 1199 struct perf_event_attr *attr = &event->attr; 1200 struct hw_perf_event *hwc = &event->hw; 1201 const struct mips_perf_event *pev; 1202 int err; 1203 1204 /* Returning MIPS event descriptor for generic perf event. */ 1205 if (PERF_TYPE_HARDWARE == event->attr.type) { 1206 if (event->attr.config >= PERF_COUNT_HW_MAX) 1207 return -EINVAL; 1208 pev = mipspmu_map_general_event(event->attr.config); 1209 } else if (PERF_TYPE_HW_CACHE == event->attr.type) { 1210 pev = mipspmu_map_cache_event(event->attr.config); 1211 } else if (PERF_TYPE_RAW == event->attr.type) { 1212 /* We are working on the global raw event. */ 1213 mutex_lock(&raw_event_mutex); 1214 pev = mipspmu.map_raw_event(event->attr.config); 1215 } else { 1216 /* The event type is not (yet) supported. */ 1217 return -EOPNOTSUPP; 1218 } 1219 1220 if (IS_ERR(pev)) { 1221 if (PERF_TYPE_RAW == event->attr.type) 1222 mutex_unlock(&raw_event_mutex); 1223 return PTR_ERR(pev); 1224 } 1225 1226 /* 1227 * We allow max flexibility on how each individual counter shared 1228 * by the single CPU operates (the mode exclusion and the range). 1229 */ 1230 hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE; 1231 1232 /* Calculate range bits and validate it. */ 1233 if (num_possible_cpus() > 1) 1234 check_and_calc_range(event, pev); 1235 1236 hwc->event_base = mipspmu_perf_event_encode(pev); 1237 if (PERF_TYPE_RAW == event->attr.type) 1238 mutex_unlock(&raw_event_mutex); 1239 1240 if (!attr->exclude_user) 1241 hwc->config_base |= M_PERFCTL_USER; 1242 if (!attr->exclude_kernel) { 1243 hwc->config_base |= M_PERFCTL_KERNEL; 1244 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */ 1245 hwc->config_base |= M_PERFCTL_EXL; 1246 } 1247 if (!attr->exclude_hv) 1248 hwc->config_base |= M_PERFCTL_SUPERVISOR; 1249 1250 hwc->config_base &= M_PERFCTL_CONFIG_MASK; 1251 /* 1252 * The event can belong to another cpu. We do not assign a local 1253 * counter for it for now. 1254 */ 1255 hwc->idx = -1; 1256 hwc->config = 0; 1257 1258 if (!hwc->sample_period) { 1259 hwc->sample_period = mipspmu.max_period; 1260 hwc->last_period = hwc->sample_period; 1261 local64_set(&hwc->period_left, hwc->sample_period); 1262 } 1263 1264 err = 0; 1265 if (event->group_leader != event) 1266 err = validate_group(event); 1267 1268 event->destroy = hw_perf_event_destroy; 1269 1270 if (err) 1271 event->destroy(event); 1272 1273 return err; 1274 } 1275 1276 static void pause_local_counters(void) 1277 { 1278 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 1279 int ctr = mipspmu.num_counters; 1280 unsigned long flags; 1281 1282 local_irq_save(flags); 1283 do { 1284 ctr--; 1285 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr); 1286 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] & 1287 ~M_PERFCTL_COUNT_EVENT_WHENEVER); 1288 } while (ctr > 0); 1289 local_irq_restore(flags); 1290 } 1291 1292 static void resume_local_counters(void) 1293 { 1294 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 1295 int ctr = mipspmu.num_counters; 1296 1297 do { 1298 ctr--; 1299 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]); 1300 } while (ctr > 0); 1301 } 1302 1303 static int mipsxx_pmu_handle_shared_irq(void) 1304 { 1305 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); 1306 struct perf_sample_data data; 1307 unsigned int counters = mipspmu.num_counters; 1308 u64 counter; 1309 int handled = IRQ_NONE; 1310 struct pt_regs *regs; 1311 1312 if (cpu_has_mips_r2 && !(read_c0_cause() & (1 << 26))) 1313 return handled; 1314 /* 1315 * First we pause the local counters, so that when we are locked 1316 * here, the counters are all paused. When it gets locked due to 1317 * perf_disable(), the timer interrupt handler will be delayed. 1318 * 1319 * See also mipsxx_pmu_start(). 1320 */ 1321 pause_local_counters(); 1322 #ifdef CONFIG_MIPS_MT_SMP 1323 read_lock(&pmuint_rwlock); 1324 #endif 1325 1326 regs = get_irq_regs(); 1327 1328 perf_sample_data_init(&data, 0); 1329 1330 switch (counters) { 1331 #define HANDLE_COUNTER(n) \ 1332 case n + 1: \ 1333 if (test_bit(n, cpuc->used_mask)) { \ 1334 counter = mipspmu.read_counter(n); \ 1335 if (counter & mipspmu.overflow) { \ 1336 handle_associated_event(cpuc, n, &data, regs); \ 1337 handled = IRQ_HANDLED; \ 1338 } \ 1339 } 1340 HANDLE_COUNTER(3) 1341 HANDLE_COUNTER(2) 1342 HANDLE_COUNTER(1) 1343 HANDLE_COUNTER(0) 1344 } 1345 1346 /* 1347 * Do all the work for the pending perf events. We can do this 1348 * in here because the performance counter interrupt is a regular 1349 * interrupt, not NMI. 1350 */ 1351 if (handled == IRQ_HANDLED) 1352 irq_work_run(); 1353 1354 #ifdef CONFIG_MIPS_MT_SMP 1355 read_unlock(&pmuint_rwlock); 1356 #endif 1357 resume_local_counters(); 1358 return handled; 1359 } 1360 1361 static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev) 1362 { 1363 return mipsxx_pmu_handle_shared_irq(); 1364 } 1365 1366 /* 24K */ 1367 #define IS_BOTH_COUNTERS_24K_EVENT(b) \ 1368 ((b) == 0 || (b) == 1 || (b) == 11) 1369 1370 /* 34K */ 1371 #define IS_BOTH_COUNTERS_34K_EVENT(b) \ 1372 ((b) == 0 || (b) == 1 || (b) == 11) 1373 #ifdef CONFIG_MIPS_MT_SMP 1374 #define IS_RANGE_P_34K_EVENT(r, b) \ 1375 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ 1376 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \ 1377 (r) == 176 || ((b) >= 50 && (b) <= 55) || \ 1378 ((b) >= 64 && (b) <= 67)) 1379 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47) 1380 #endif 1381 1382 /* 74K */ 1383 #define IS_BOTH_COUNTERS_74K_EVENT(b) \ 1384 ((b) == 0 || (b) == 1) 1385 1386 /* 1004K */ 1387 #define IS_BOTH_COUNTERS_1004K_EVENT(b) \ 1388 ((b) == 0 || (b) == 1 || (b) == 11) 1389 #ifdef CONFIG_MIPS_MT_SMP 1390 #define IS_RANGE_P_1004K_EVENT(r, b) \ 1391 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \ 1392 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \ 1393 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \ 1394 (r) == 188 || (b) == 61 || (b) == 62 || \ 1395 ((b) >= 64 && (b) <= 67)) 1396 #define IS_RANGE_V_1004K_EVENT(r) ((r) == 47) 1397 #endif 1398 1399 /* 1400 * User can use 0-255 raw events, where 0-127 for the events of even 1401 * counters, and 128-255 for odd counters. Note that bit 7 is used to 1402 * indicate the parity. So, for example, when user wants to take the 1403 * Event Num of 15 for odd counters (by referring to the user manual), 1404 * then 128 needs to be added to 15 as the input for the event config, 1405 * i.e., 143 (0x8F) to be used. 1406 */ 1407 static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config) 1408 { 1409 unsigned int raw_id = config & 0xff; 1410 unsigned int base_id = raw_id & 0x7f; 1411 1412 raw_event.event_id = base_id; 1413 1414 switch (current_cpu_type()) { 1415 case CPU_24K: 1416 if (IS_BOTH_COUNTERS_24K_EVENT(base_id)) 1417 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1418 else 1419 raw_event.cntr_mask = 1420 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1421 #ifdef CONFIG_MIPS_MT_SMP 1422 /* 1423 * This is actually doing nothing. Non-multithreading 1424 * CPUs will not check and calculate the range. 1425 */ 1426 raw_event.range = P; 1427 #endif 1428 break; 1429 case CPU_34K: 1430 if (IS_BOTH_COUNTERS_34K_EVENT(base_id)) 1431 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1432 else 1433 raw_event.cntr_mask = 1434 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1435 #ifdef CONFIG_MIPS_MT_SMP 1436 if (IS_RANGE_P_34K_EVENT(raw_id, base_id)) 1437 raw_event.range = P; 1438 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id))) 1439 raw_event.range = V; 1440 else 1441 raw_event.range = T; 1442 #endif 1443 break; 1444 case CPU_74K: 1445 if (IS_BOTH_COUNTERS_74K_EVENT(base_id)) 1446 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1447 else 1448 raw_event.cntr_mask = 1449 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1450 #ifdef CONFIG_MIPS_MT_SMP 1451 raw_event.range = P; 1452 #endif 1453 break; 1454 case CPU_1004K: 1455 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id)) 1456 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD; 1457 else 1458 raw_event.cntr_mask = 1459 raw_id > 127 ? CNTR_ODD : CNTR_EVEN; 1460 #ifdef CONFIG_MIPS_MT_SMP 1461 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id)) 1462 raw_event.range = P; 1463 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id))) 1464 raw_event.range = V; 1465 else 1466 raw_event.range = T; 1467 #endif 1468 break; 1469 } 1470 1471 return &raw_event; 1472 } 1473 1474 static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config) 1475 { 1476 unsigned int raw_id = config & 0xff; 1477 unsigned int base_id = raw_id & 0x7f; 1478 1479 1480 raw_event.cntr_mask = CNTR_ALL; 1481 raw_event.event_id = base_id; 1482 1483 if (current_cpu_type() == CPU_CAVIUM_OCTEON2) { 1484 if (base_id > 0x42) 1485 return ERR_PTR(-EOPNOTSUPP); 1486 } else { 1487 if (base_id > 0x3a) 1488 return ERR_PTR(-EOPNOTSUPP); 1489 } 1490 1491 switch (base_id) { 1492 case 0x00: 1493 case 0x0f: 1494 case 0x1e: 1495 case 0x1f: 1496 case 0x2f: 1497 case 0x34: 1498 case 0x3b ... 0x3f: 1499 return ERR_PTR(-EOPNOTSUPP); 1500 default: 1501 break; 1502 } 1503 1504 return &raw_event; 1505 } 1506 1507 static int __init 1508 init_hw_perf_events(void) 1509 { 1510 int counters, irq; 1511 int counter_bits; 1512 1513 pr_info("Performance counters: "); 1514 1515 counters = n_counters(); 1516 if (counters == 0) { 1517 pr_cont("No available PMU.\n"); 1518 return -ENODEV; 1519 } 1520 1521 #ifdef CONFIG_MIPS_MT_SMP 1522 cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19); 1523 if (!cpu_has_mipsmt_pertccounters) 1524 counters = counters_total_to_per_cpu(counters); 1525 #endif 1526 1527 #ifdef MSC01E_INT_BASE 1528 if (cpu_has_veic) { 1529 /* 1530 * Using platform specific interrupt controller defines. 1531 */ 1532 irq = MSC01E_INT_BASE + MSC01E_INT_PERFCTR; 1533 } else { 1534 #endif 1535 if (cp0_perfcount_irq >= 0) 1536 irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq; 1537 else 1538 irq = -1; 1539 #ifdef MSC01E_INT_BASE 1540 } 1541 #endif 1542 1543 mipspmu.map_raw_event = mipsxx_pmu_map_raw_event; 1544 1545 switch (current_cpu_type()) { 1546 case CPU_24K: 1547 mipspmu.name = "mips/24K"; 1548 mipspmu.general_event_map = &mipsxxcore_event_map; 1549 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1550 break; 1551 case CPU_34K: 1552 mipspmu.name = "mips/34K"; 1553 mipspmu.general_event_map = &mipsxxcore_event_map; 1554 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1555 break; 1556 case CPU_74K: 1557 mipspmu.name = "mips/74K"; 1558 mipspmu.general_event_map = &mipsxx74Kcore_event_map; 1559 mipspmu.cache_event_map = &mipsxx74Kcore_cache_map; 1560 break; 1561 case CPU_1004K: 1562 mipspmu.name = "mips/1004K"; 1563 mipspmu.general_event_map = &mipsxxcore_event_map; 1564 mipspmu.cache_event_map = &mipsxxcore_cache_map; 1565 break; 1566 case CPU_CAVIUM_OCTEON: 1567 case CPU_CAVIUM_OCTEON_PLUS: 1568 case CPU_CAVIUM_OCTEON2: 1569 mipspmu.name = "octeon"; 1570 mipspmu.general_event_map = &octeon_event_map; 1571 mipspmu.cache_event_map = &octeon_cache_map; 1572 mipspmu.map_raw_event = octeon_pmu_map_raw_event; 1573 break; 1574 default: 1575 pr_cont("Either hardware does not support performance " 1576 "counters, or not yet implemented.\n"); 1577 return -ENODEV; 1578 } 1579 1580 mipspmu.num_counters = counters; 1581 mipspmu.irq = irq; 1582 1583 if (read_c0_perfctrl0() & M_PERFCTL_WIDE) { 1584 mipspmu.max_period = (1ULL << 63) - 1; 1585 mipspmu.valid_count = (1ULL << 63) - 1; 1586 mipspmu.overflow = 1ULL << 63; 1587 mipspmu.read_counter = mipsxx_pmu_read_counter_64; 1588 mipspmu.write_counter = mipsxx_pmu_write_counter_64; 1589 counter_bits = 64; 1590 } else { 1591 mipspmu.max_period = (1ULL << 31) - 1; 1592 mipspmu.valid_count = (1ULL << 31) - 1; 1593 mipspmu.overflow = 1ULL << 31; 1594 mipspmu.read_counter = mipsxx_pmu_read_counter; 1595 mipspmu.write_counter = mipsxx_pmu_write_counter; 1596 counter_bits = 32; 1597 } 1598 1599 on_each_cpu(reset_counters, (void *)(long)counters, 1); 1600 1601 pr_cont("%s PMU enabled, %d %d-bit counters available to each " 1602 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq, 1603 irq < 0 ? " (share with timer interrupt)" : ""); 1604 1605 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW); 1606 1607 return 0; 1608 } 1609 early_initcall(init_hw_perf_events); 1610