Linux Audio

Check our new training course

Loading...
v3.5.6
 
   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
  32struct 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};
  49DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
  50	.saved_ctrl = {0},
  51};
  52
  53/* The description of MIPS performance events. */
  54struct 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
  78static struct mips_perf_event raw_event;
  79static DEFINE_MUTEX(raw_event_mutex);
  80
  81#define UNSUPPORTED_PERF_EVENT_ID 0xffffffff
  82#define C(x) PERF_COUNT_HW_CACHE_##x
  83
  84struct 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
 101static 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
 135static int cpu_has_mipsmt_pertccounters;
 136
 137static 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 */
 152static unsigned int vpe_shift(void)
 153{
 154	if (num_possible_cpus() > 1)
 155		return 1;
 156
 157	return 0;
 158}
 159
 160static unsigned int counters_total_to_per_cpu(unsigned int counters)
 161{
 162	return counters >> vpe_shift();
 163}
 164
 165#else /* !CONFIG_MIPS_MT_SMP */
 166#define vpe_id()	0
 167
 168#endif /* CONFIG_MIPS_MT_SMP */
 169
 170static void resume_local_counters(void);
 171static void pause_local_counters(void);
 172static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
 173static int mipsxx_pmu_handle_shared_irq(void);
 174
 175static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
 176{
 177	if (vpe_id() == 1)
 178		idx = (idx + 2) & 3;
 179	return idx;
 180}
 181
 182static u64 mipsxx_pmu_read_counter(unsigned int idx)
 183{
 184	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 185
 186	switch (idx) {
 187	case 0:
 188		/*
 189		 * The counters are unsigned, we must cast to truncate
 190		 * off the high bits.
 191		 */
 192		return (u32)read_c0_perfcntr0();
 193	case 1:
 194		return (u32)read_c0_perfcntr1();
 195	case 2:
 196		return (u32)read_c0_perfcntr2();
 197	case 3:
 198		return (u32)read_c0_perfcntr3();
 199	default:
 200		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 201		return 0;
 202	}
 203}
 204
 205static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
 206{
 207	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 208
 209	switch (idx) {
 210	case 0:
 211		return read_c0_perfcntr0_64();
 212	case 1:
 213		return read_c0_perfcntr1_64();
 214	case 2:
 215		return read_c0_perfcntr2_64();
 216	case 3:
 217		return read_c0_perfcntr3_64();
 218	default:
 219		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 220		return 0;
 221	}
 222}
 223
 224static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
 225{
 226	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 227
 228	switch (idx) {
 229	case 0:
 230		write_c0_perfcntr0(val);
 231		return;
 232	case 1:
 233		write_c0_perfcntr1(val);
 234		return;
 235	case 2:
 236		write_c0_perfcntr2(val);
 237		return;
 238	case 3:
 239		write_c0_perfcntr3(val);
 240		return;
 241	}
 242}
 243
 244static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
 245{
 246	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 247
 248	switch (idx) {
 249	case 0:
 250		write_c0_perfcntr0_64(val);
 251		return;
 252	case 1:
 253		write_c0_perfcntr1_64(val);
 254		return;
 255	case 2:
 256		write_c0_perfcntr2_64(val);
 257		return;
 258	case 3:
 259		write_c0_perfcntr3_64(val);
 260		return;
 261	}
 262}
 263
 264static unsigned int mipsxx_pmu_read_control(unsigned int idx)
 265{
 266	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 267
 268	switch (idx) {
 269	case 0:
 270		return read_c0_perfctrl0();
 271	case 1:
 272		return read_c0_perfctrl1();
 273	case 2:
 274		return read_c0_perfctrl2();
 275	case 3:
 276		return read_c0_perfctrl3();
 277	default:
 278		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 279		return 0;
 280	}
 281}
 282
 283static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
 284{
 285	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 286
 287	switch (idx) {
 288	case 0:
 289		write_c0_perfctrl0(val);
 290		return;
 291	case 1:
 292		write_c0_perfctrl1(val);
 293		return;
 294	case 2:
 295		write_c0_perfctrl2(val);
 296		return;
 297	case 3:
 298		write_c0_perfctrl3(val);
 299		return;
 300	}
 301}
 302
 303static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
 304				    struct hw_perf_event *hwc)
 305{
 306	int i;
 307
 308	/*
 309	 * We only need to care the counter mask. The range has been
 310	 * checked definitely.
 311	 */
 312	unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
 313
 314	for (i = mipspmu.num_counters - 1; i >= 0; i--) {
 315		/*
 316		 * Note that some MIPS perf events can be counted by both
 317		 * even and odd counters, wheresas many other are only by
 318		 * even _or_ odd counters. This introduces an issue that
 319		 * when the former kind of event takes the counter the
 320		 * latter kind of event wants to use, then the "counter
 321		 * allocation" for the latter event will fail. In fact if
 322		 * they can be dynamically swapped, they both feel happy.
 323		 * But here we leave this issue alone for now.
 324		 */
 325		if (test_bit(i, &cntr_mask) &&
 326			!test_and_set_bit(i, cpuc->used_mask))
 327			return i;
 328	}
 329
 330	return -EAGAIN;
 331}
 332
 333static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
 334{
 335	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 
 
 336
 337	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 338
 339	cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
 340		(evt->config_base & M_PERFCTL_CONFIG_MASK) |
 341		/* Make sure interrupt enabled. */
 342		M_PERFCTL_INTERRUPT_ENABLE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 343	/*
 344	 * We do not actually let the counter run. Leave it until start().
 345	 */
 346}
 347
 348static void mipsxx_pmu_disable_event(int idx)
 349{
 350	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 351	unsigned long flags;
 352
 353	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 354
 355	local_irq_save(flags);
 356	cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
 357		~M_PERFCTL_COUNT_EVENT_WHENEVER;
 358	mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
 359	local_irq_restore(flags);
 360}
 361
 362static int mipspmu_event_set_period(struct perf_event *event,
 363				    struct hw_perf_event *hwc,
 364				    int idx)
 365{
 366	u64 left = local64_read(&hwc->period_left);
 367	u64 period = hwc->sample_period;
 368	int ret = 0;
 369
 370	if (unlikely((left + period) & (1ULL << 63))) {
 371		/* left underflowed by more than period. */
 372		left = period;
 373		local64_set(&hwc->period_left, left);
 374		hwc->last_period = period;
 375		ret = 1;
 376	} else	if (unlikely((left + period) <= period)) {
 377		/* left underflowed by less than period. */
 378		left += period;
 379		local64_set(&hwc->period_left, left);
 380		hwc->last_period = period;
 381		ret = 1;
 382	}
 383
 384	if (left > mipspmu.max_period) {
 385		left = mipspmu.max_period;
 386		local64_set(&hwc->period_left, left);
 387	}
 388
 389	local64_set(&hwc->prev_count, mipspmu.overflow - left);
 390
 391	mipspmu.write_counter(idx, mipspmu.overflow - left);
 392
 393	perf_event_update_userpage(event);
 394
 395	return ret;
 396}
 397
 398static void mipspmu_event_update(struct perf_event *event,
 399				 struct hw_perf_event *hwc,
 400				 int idx)
 401{
 402	u64 prev_raw_count, new_raw_count;
 403	u64 delta;
 404
 405again:
 406	prev_raw_count = local64_read(&hwc->prev_count);
 407	new_raw_count = mipspmu.read_counter(idx);
 408
 409	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 410				new_raw_count) != prev_raw_count)
 411		goto again;
 412
 413	delta = new_raw_count - prev_raw_count;
 414
 415	local64_add(delta, &event->count);
 416	local64_sub(delta, &hwc->period_left);
 417}
 418
 419static void mipspmu_start(struct perf_event *event, int flags)
 420{
 421	struct hw_perf_event *hwc = &event->hw;
 422
 423	if (flags & PERF_EF_RELOAD)
 424		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 425
 426	hwc->state = 0;
 427
 428	/* Set the period for the event. */
 429	mipspmu_event_set_period(event, hwc, hwc->idx);
 430
 431	/* Enable the event. */
 432	mipsxx_pmu_enable_event(hwc, hwc->idx);
 433}
 434
 435static void mipspmu_stop(struct perf_event *event, int flags)
 436{
 437	struct hw_perf_event *hwc = &event->hw;
 438
 439	if (!(hwc->state & PERF_HES_STOPPED)) {
 440		/* We are working on a local event. */
 441		mipsxx_pmu_disable_event(hwc->idx);
 442		barrier();
 443		mipspmu_event_update(event, hwc, hwc->idx);
 444		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
 445	}
 446}
 447
 448static int mipspmu_add(struct perf_event *event, int flags)
 449{
 450	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 451	struct hw_perf_event *hwc = &event->hw;
 452	int idx;
 453	int err = 0;
 454
 455	perf_pmu_disable(event->pmu);
 456
 457	/* To look for a free counter for this event. */
 458	idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
 459	if (idx < 0) {
 460		err = idx;
 461		goto out;
 462	}
 463
 464	/*
 465	 * If there is an event in the counter we are going to use then
 466	 * make sure it is disabled.
 467	 */
 468	event->hw.idx = idx;
 469	mipsxx_pmu_disable_event(idx);
 470	cpuc->events[idx] = event;
 471
 472	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
 473	if (flags & PERF_EF_START)
 474		mipspmu_start(event, PERF_EF_RELOAD);
 475
 476	/* Propagate our changes to the userspace mapping. */
 477	perf_event_update_userpage(event);
 478
 479out:
 480	perf_pmu_enable(event->pmu);
 481	return err;
 482}
 483
 484static void mipspmu_del(struct perf_event *event, int flags)
 485{
 486	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 487	struct hw_perf_event *hwc = &event->hw;
 488	int idx = hwc->idx;
 489
 490	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 491
 492	mipspmu_stop(event, PERF_EF_UPDATE);
 493	cpuc->events[idx] = NULL;
 494	clear_bit(idx, cpuc->used_mask);
 495
 496	perf_event_update_userpage(event);
 497}
 498
 499static void mipspmu_read(struct perf_event *event)
 500{
 501	struct hw_perf_event *hwc = &event->hw;
 502
 503	/* Don't read disabled counters! */
 504	if (hwc->idx < 0)
 505		return;
 506
 507	mipspmu_event_update(event, hwc, hwc->idx);
 508}
 509
 510static void mipspmu_enable(struct pmu *pmu)
 511{
 512#ifdef CONFIG_MIPS_MT_SMP
 513	write_unlock(&pmuint_rwlock);
 514#endif
 515	resume_local_counters();
 516}
 517
 518/*
 519 * MIPS performance counters can be per-TC. The control registers can
 520 * not be directly accessed accross CPUs. Hence if we want to do global
 521 * control, we need cross CPU calls. on_each_cpu() can help us, but we
 522 * can not make sure this function is called with interrupts enabled. So
 523 * here we pause local counters and then grab a rwlock and leave the
 524 * counters on other CPUs alone. If any counter interrupt raises while
 525 * we own the write lock, simply pause local counters on that CPU and
 526 * spin in the handler. Also we know we won't be switched to another
 527 * CPU after pausing local counters and before grabbing the lock.
 528 */
 529static void mipspmu_disable(struct pmu *pmu)
 530{
 531	pause_local_counters();
 532#ifdef CONFIG_MIPS_MT_SMP
 533	write_lock(&pmuint_rwlock);
 534#endif
 535}
 536
 537static atomic_t active_events = ATOMIC_INIT(0);
 538static DEFINE_MUTEX(pmu_reserve_mutex);
 539static int (*save_perf_irq)(void);
 540
 541static int mipspmu_get_irq(void)
 542{
 543	int err;
 544
 545	if (mipspmu.irq >= 0) {
 546		/* Request my own irq handler. */
 547		err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
 548			IRQF_PERCPU | IRQF_NOBALANCING,
 549			"mips_perf_pmu", NULL);
 
 
 550		if (err) {
 551			pr_warning("Unable to request IRQ%d for MIPS "
 552			   "performance counters!\n", mipspmu.irq);
 553		}
 554	} else if (cp0_perfcount_irq < 0) {
 555		/*
 556		 * We are sharing the irq number with the timer interrupt.
 557		 */
 558		save_perf_irq = perf_irq;
 559		perf_irq = mipsxx_pmu_handle_shared_irq;
 560		err = 0;
 561	} else {
 562		pr_warning("The platform hasn't properly defined its "
 563			"interrupt controller.\n");
 564		err = -ENOENT;
 565	}
 566
 567	return err;
 568}
 569
 570static void mipspmu_free_irq(void)
 571{
 572	if (mipspmu.irq >= 0)
 573		free_irq(mipspmu.irq, NULL);
 574	else if (cp0_perfcount_irq < 0)
 575		perf_irq = save_perf_irq;
 576}
 577
 578/*
 579 * mipsxx/rm9000/loongson2 have different performance counters, they have
 580 * specific low-level init routines.
 581 */
 582static void reset_counters(void *arg);
 583static int __hw_perf_event_init(struct perf_event *event);
 584
 585static void hw_perf_event_destroy(struct perf_event *event)
 586{
 587	if (atomic_dec_and_mutex_lock(&active_events,
 588				&pmu_reserve_mutex)) {
 589		/*
 590		 * We must not call the destroy function with interrupts
 591		 * disabled.
 592		 */
 593		on_each_cpu(reset_counters,
 594			(void *)(long)mipspmu.num_counters, 1);
 595		mipspmu_free_irq();
 596		mutex_unlock(&pmu_reserve_mutex);
 597	}
 598}
 599
 600static int mipspmu_event_init(struct perf_event *event)
 601{
 602	int err = 0;
 603
 604	/* does not support taken branch sampling */
 605	if (has_branch_stack(event))
 606		return -EOPNOTSUPP;
 607
 608	switch (event->attr.type) {
 609	case PERF_TYPE_RAW:
 610	case PERF_TYPE_HARDWARE:
 611	case PERF_TYPE_HW_CACHE:
 612		break;
 613
 614	default:
 615		return -ENOENT;
 616	}
 617
 618	if (event->cpu >= nr_cpumask_bits ||
 619	    (event->cpu >= 0 && !cpu_online(event->cpu)))
 620		return -ENODEV;
 621
 622	if (!atomic_inc_not_zero(&active_events)) {
 623		mutex_lock(&pmu_reserve_mutex);
 624		if (atomic_read(&active_events) == 0)
 625			err = mipspmu_get_irq();
 626
 627		if (!err)
 628			atomic_inc(&active_events);
 629		mutex_unlock(&pmu_reserve_mutex);
 630	}
 631
 632	if (err)
 633		return err;
 634
 635	return __hw_perf_event_init(event);
 636}
 637
 638static struct pmu pmu = {
 639	.pmu_enable	= mipspmu_enable,
 640	.pmu_disable	= mipspmu_disable,
 641	.event_init	= mipspmu_event_init,
 642	.add		= mipspmu_add,
 643	.del		= mipspmu_del,
 644	.start		= mipspmu_start,
 645	.stop		= mipspmu_stop,
 646	.read		= mipspmu_read,
 647};
 648
 649static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
 650{
 651/*
 652 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
 653 * event_id.
 654 */
 655#ifdef CONFIG_MIPS_MT_SMP
 656	return ((unsigned int)pev->range << 24) |
 657		(pev->cntr_mask & 0xffff00) |
 658		(pev->event_id & 0xff);
 659#else
 660	return (pev->cntr_mask & 0xffff00) |
 661		(pev->event_id & 0xff);
 662#endif
 
 663}
 664
 665static const struct mips_perf_event *mipspmu_map_general_event(int idx)
 666{
 667	const struct mips_perf_event *pev;
 668
 669	pev = ((*mipspmu.general_event_map)[idx].event_id ==
 670		UNSUPPORTED_PERF_EVENT_ID ? ERR_PTR(-EOPNOTSUPP) :
 671		&(*mipspmu.general_event_map)[idx]);
 672
 673	return pev;
 674}
 675
 676static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
 677{
 678	unsigned int cache_type, cache_op, cache_result;
 679	const struct mips_perf_event *pev;
 680
 681	cache_type = (config >> 0) & 0xff;
 682	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 683		return ERR_PTR(-EINVAL);
 684
 685	cache_op = (config >> 8) & 0xff;
 686	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 687		return ERR_PTR(-EINVAL);
 688
 689	cache_result = (config >> 16) & 0xff;
 690	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 691		return ERR_PTR(-EINVAL);
 692
 693	pev = &((*mipspmu.cache_event_map)
 694					[cache_type]
 695					[cache_op]
 696					[cache_result]);
 697
 698	if (pev->event_id == UNSUPPORTED_PERF_EVENT_ID)
 699		return ERR_PTR(-EOPNOTSUPP);
 700
 701	return pev;
 702
 703}
 704
 705static int validate_group(struct perf_event *event)
 706{
 707	struct perf_event *sibling, *leader = event->group_leader;
 708	struct cpu_hw_events fake_cpuc;
 709
 710	memset(&fake_cpuc, 0, sizeof(fake_cpuc));
 711
 712	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
 713		return -EINVAL;
 714
 715	list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
 716		if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
 717			return -EINVAL;
 718	}
 719
 720	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
 721		return -EINVAL;
 722
 723	return 0;
 724}
 725
 726/* This is needed by specific irq handlers in perf_event_*.c */
 727static void handle_associated_event(struct cpu_hw_events *cpuc,
 728				    int idx, struct perf_sample_data *data,
 729				    struct pt_regs *regs)
 730{
 731	struct perf_event *event = cpuc->events[idx];
 732	struct hw_perf_event *hwc = &event->hw;
 733
 734	mipspmu_event_update(event, hwc, idx);
 735	data->period = event->hw.last_period;
 736	if (!mipspmu_event_set_period(event, hwc, idx))
 737		return;
 738
 739	if (perf_event_overflow(event, data, regs))
 740		mipsxx_pmu_disable_event(idx);
 741}
 742
 743
 744static int __n_counters(void)
 745{
 746	if (!(read_c0_config1() & M_CONFIG1_PC))
 747		return 0;
 748	if (!(read_c0_perfctrl0() & M_PERFCTL_MORE))
 749		return 1;
 750	if (!(read_c0_perfctrl1() & M_PERFCTL_MORE))
 751		return 2;
 752	if (!(read_c0_perfctrl2() & M_PERFCTL_MORE))
 753		return 3;
 754
 755	return 4;
 756}
 757
 758static int n_counters(void)
 759{
 760	int counters;
 761
 762	switch (current_cpu_type()) {
 763	case CPU_R10000:
 764		counters = 2;
 765		break;
 766
 767	case CPU_R12000:
 768	case CPU_R14000:
 
 769		counters = 4;
 770		break;
 771
 772	default:
 773		counters = __n_counters();
 774	}
 775
 776	return counters;
 777}
 778
 779static void reset_counters(void *arg)
 780{
 781	int counters = (int)(long)arg;
 782	switch (counters) {
 783	case 4:
 784		mipsxx_pmu_write_control(3, 0);
 785		mipspmu.write_counter(3, 0);
 
 786	case 3:
 787		mipsxx_pmu_write_control(2, 0);
 788		mipspmu.write_counter(2, 0);
 
 789	case 2:
 790		mipsxx_pmu_write_control(1, 0);
 791		mipspmu.write_counter(1, 0);
 
 792	case 1:
 793		mipsxx_pmu_write_control(0, 0);
 794		mipspmu.write_counter(0, 0);
 
 795	}
 796}
 797
 798/* 24K/34K/1004K cores can share the same event map. */
 799static const struct mips_perf_event mipsxxcore_event_map
 800				[PERF_COUNT_HW_MAX] = {
 801	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
 802	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
 803	[PERF_COUNT_HW_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
 804	[PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
 805	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
 806	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
 807	[PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
 808};
 809
 810/* 74K core has different branch event code. */
 811static const struct mips_perf_event mipsxx74Kcore_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_CACHE_REFERENCES] = { UNSUPPORTED_PERF_EVENT_ID },
 816	[PERF_COUNT_HW_CACHE_MISSES] = { UNSUPPORTED_PERF_EVENT_ID },
 817	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
 818	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
 819	[PERF_COUNT_HW_BUS_CYCLES] = { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 820};
 821
 822static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
 823	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
 824	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
 825	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
 826	[PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL  },
 827	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
 828	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
 829	[PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
 830};
 831
 832/* 24K/34K/1004K cores can share the same cache event map. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 833static const struct mips_perf_event mipsxxcore_cache_map
 834				[PERF_COUNT_HW_CACHE_MAX]
 835				[PERF_COUNT_HW_CACHE_OP_MAX]
 836				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
 837[C(L1D)] = {
 838	/*
 839	 * Like some other architectures (e.g. ARM), the performance
 840	 * counters don't differentiate between read and write
 841	 * accesses/misses, so this isn't strictly correct, but it's the
 842	 * best we can do. Writes and reads get combined.
 843	 */
 844	[C(OP_READ)] = {
 845		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
 846		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
 847	},
 848	[C(OP_WRITE)] = {
 849		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
 850		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
 851	},
 852	[C(OP_PREFETCH)] = {
 853		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 854		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 855	},
 856},
 857[C(L1I)] = {
 858	[C(OP_READ)] = {
 859		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
 860		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
 861	},
 862	[C(OP_WRITE)] = {
 863		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
 864		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
 865	},
 866	[C(OP_PREFETCH)] = {
 867		[C(RESULT_ACCESS)]	= { 0x14, CNTR_EVEN, T },
 868		/*
 869		 * Note that MIPS has only "hit" events countable for
 870		 * the prefetch operation.
 871		 */
 872		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 873	},
 874},
 875[C(LL)] = {
 876	[C(OP_READ)] = {
 877		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
 878		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
 879	},
 880	[C(OP_WRITE)] = {
 881		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
 882		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
 883	},
 884	[C(OP_PREFETCH)] = {
 885		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 886		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 887	},
 888},
 889[C(DTLB)] = {
 890	[C(OP_READ)] = {
 891		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 892		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 893	},
 894	[C(OP_WRITE)] = {
 895		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 896		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 897	},
 898	[C(OP_PREFETCH)] = {
 899		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 900		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 901	},
 902},
 903[C(ITLB)] = {
 904	[C(OP_READ)] = {
 905		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
 906		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
 907	},
 908	[C(OP_WRITE)] = {
 909		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
 910		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
 911	},
 912	[C(OP_PREFETCH)] = {
 913		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 914		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 915	},
 916},
 917[C(BPU)] = {
 918	/* Using the same code for *HW_BRANCH* */
 919	[C(OP_READ)] = {
 920		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
 921		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
 922	},
 923	[C(OP_WRITE)] = {
 924		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
 925		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
 926	},
 927	[C(OP_PREFETCH)] = {
 928		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 929		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 930	},
 931},
 932[C(NODE)] = {
 933	[C(OP_READ)] = {
 934		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 935		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 936	},
 937	[C(OP_WRITE)] = {
 938		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 939		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 940	},
 941	[C(OP_PREFETCH)] = {
 942		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 943		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 944	},
 945},
 946};
 947
 948/* 74K core has completely different cache event map. */
 949static const struct mips_perf_event mipsxx74Kcore_cache_map
 950				[PERF_COUNT_HW_CACHE_MAX]
 951				[PERF_COUNT_HW_CACHE_OP_MAX]
 952				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
 953[C(L1D)] = {
 954	/*
 955	 * Like some other architectures (e.g. ARM), the performance
 956	 * counters don't differentiate between read and write
 957	 * accesses/misses, so this isn't strictly correct, but it's the
 958	 * best we can do. Writes and reads get combined.
 959	 */
 960	[C(OP_READ)] = {
 961		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
 962		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
 963	},
 964	[C(OP_WRITE)] = {
 965		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
 966		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
 967	},
 968	[C(OP_PREFETCH)] = {
 969		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 970		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 971	},
 972},
 973[C(L1I)] = {
 974	[C(OP_READ)] = {
 975		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 976		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 977	},
 978	[C(OP_WRITE)] = {
 979		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 980		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 981	},
 982	[C(OP_PREFETCH)] = {
 983		[C(RESULT_ACCESS)]	= { 0x34, CNTR_EVEN, T },
 984		/*
 985		 * Note that MIPS has only "hit" events countable for
 986		 * the prefetch operation.
 987		 */
 988		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 989	},
 990},
 991[C(LL)] = {
 992	[C(OP_READ)] = {
 993		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
 994		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN | CNTR_ODD, P },
 995	},
 996	[C(OP_WRITE)] = {
 997		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
 998		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN | CNTR_ODD, P },
 999	},
1000	[C(OP_PREFETCH)] = {
1001		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1002		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1003	},
1004},
1005[C(DTLB)] = {
1006	/* 74K core does not have specific DTLB events. */
1007	[C(OP_READ)] = {
1008		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1009		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1010	},
1011	[C(OP_WRITE)] = {
1012		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1013		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1014	},
1015	[C(OP_PREFETCH)] = {
1016		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1017		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1018	},
1019},
 
 
 
 
 
1020[C(ITLB)] = {
1021	[C(OP_READ)] = {
1022		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1023		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1024	},
1025	[C(OP_WRITE)] = {
1026		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1027		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1028	},
1029	[C(OP_PREFETCH)] = {
1030		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1031		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1032	},
1033},
1034[C(BPU)] = {
1035	/* Using the same code for *HW_BRANCH* */
1036	[C(OP_READ)] = {
1037		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1038		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1039	},
1040	[C(OP_WRITE)] = {
1041		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1042		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1043	},
1044	[C(OP_PREFETCH)] = {
1045		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1046		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
 
 
 
 
 
 
 
 
1047	},
1048},
1049[C(NODE)] = {
 
 
 
 
 
 
 
1050	[C(OP_READ)] = {
1051		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1052		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1053	},
1054	[C(OP_WRITE)] = {
1055		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1056		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1057	},
1058	[C(OP_PREFETCH)] = {
1059		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1060		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
1061	},
1062},
1063};
1064
1065
1066static const struct mips_perf_event octeon_cache_map
1067				[PERF_COUNT_HW_CACHE_MAX]
1068				[PERF_COUNT_HW_CACHE_OP_MAX]
1069				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1070[C(L1D)] = {
 
 
 
 
 
 
1071	[C(OP_READ)] = {
1072		[C(RESULT_ACCESS)]	= { 0x2b, CNTR_ALL },
1073		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL },
1074	},
1075	[C(OP_WRITE)] = {
1076		[C(RESULT_ACCESS)]	= { 0x30, CNTR_ALL },
1077		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1078	},
1079	[C(OP_PREFETCH)] = {
1080		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1081		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1082	},
1083},
1084[C(L1I)] = {
1085	[C(OP_READ)] = {
1086		[C(RESULT_ACCESS)]	= { 0x18, CNTR_ALL },
1087		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1088	},
1089	[C(OP_WRITE)] = {
1090		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1091		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1092	},
1093	[C(OP_PREFETCH)] = {
1094		[C(RESULT_ACCESS)]	= { 0x19, CNTR_ALL },
1095		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
1096	},
1097},
1098[C(LL)] = {
1099	[C(OP_READ)] = {
1100		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1101		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1102	},
1103	[C(OP_WRITE)] = {
1104		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1105		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
1106	},
1107	[C(OP_PREFETCH)] = {
1108		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1109		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1110	},
1111},
1112[C(DTLB)] = {
1113	/*
1114	 * Only general DTLB misses are counted use the same event for
1115	 * read and write.
1116	 */
1117	[C(OP_READ)] = {
1118		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1119		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1120	},
1121	[C(OP_WRITE)] = {
1122		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1123		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1124	},
1125	[C(OP_PREFETCH)] = {
1126		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1127		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1128	},
1129},
1130[C(ITLB)] = {
1131	[C(OP_READ)] = {
1132		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1133		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL },
1134	},
 
 
 
 
 
 
 
 
 
 
 
 
1135	[C(OP_WRITE)] = {
1136		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1137		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1138	},
1139	[C(OP_PREFETCH)] = {
1140		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1141		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
1142	},
1143},
1144[C(BPU)] = {
1145	/* Using the same code for *HW_BRANCH* */
1146	[C(OP_READ)] = {
1147		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1148		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1149	},
1150	[C(OP_WRITE)] = {
1151		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1152		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1153	},
1154	[C(OP_PREFETCH)] = {
1155		[C(RESULT_ACCESS)]	= { UNSUPPORTED_PERF_EVENT_ID },
1156		[C(RESULT_MISS)]	= { UNSUPPORTED_PERF_EVENT_ID },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1157	},
1158},
1159};
1160
1161#ifdef CONFIG_MIPS_MT_SMP
1162static void check_and_calc_range(struct perf_event *event,
1163				 const struct mips_perf_event *pev)
1164{
1165	struct hw_perf_event *hwc = &event->hw;
1166
1167	if (event->cpu >= 0) {
1168		if (pev->range > V) {
1169			/*
1170			 * The user selected an event that is processor
1171			 * wide, while expecting it to be VPE wide.
1172			 */
1173			hwc->config_base |= M_TC_EN_ALL;
1174		} else {
1175			/*
1176			 * FIXME: cpu_data[event->cpu].vpe_id reports 0
1177			 * for both CPUs.
1178			 */
1179			hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
1180			hwc->config_base |= M_TC_EN_VPE;
1181		}
1182	} else
1183		hwc->config_base |= M_TC_EN_ALL;
1184}
1185#else
1186static void check_and_calc_range(struct perf_event *event,
1187				 const struct mips_perf_event *pev)
1188{
1189}
1190#endif
1191
1192static int __hw_perf_event_init(struct perf_event *event)
1193{
1194	struct perf_event_attr *attr = &event->attr;
1195	struct hw_perf_event *hwc = &event->hw;
1196	const struct mips_perf_event *pev;
1197	int err;
1198
1199	/* Returning MIPS event descriptor for generic perf event. */
1200	if (PERF_TYPE_HARDWARE == event->attr.type) {
1201		if (event->attr.config >= PERF_COUNT_HW_MAX)
1202			return -EINVAL;
1203		pev = mipspmu_map_general_event(event->attr.config);
1204	} else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1205		pev = mipspmu_map_cache_event(event->attr.config);
1206	} else if (PERF_TYPE_RAW == event->attr.type) {
1207		/* We are working on the global raw event. */
1208		mutex_lock(&raw_event_mutex);
1209		pev = mipspmu.map_raw_event(event->attr.config);
1210	} else {
1211		/* The event type is not (yet) supported. */
1212		return -EOPNOTSUPP;
1213	}
1214
1215	if (IS_ERR(pev)) {
1216		if (PERF_TYPE_RAW == event->attr.type)
1217			mutex_unlock(&raw_event_mutex);
1218		return PTR_ERR(pev);
1219	}
1220
1221	/*
1222	 * We allow max flexibility on how each individual counter shared
1223	 * by the single CPU operates (the mode exclusion and the range).
1224	 */
1225	hwc->config_base = M_PERFCTL_INTERRUPT_ENABLE;
1226
1227	/* Calculate range bits and validate it. */
1228	if (num_possible_cpus() > 1)
1229		check_and_calc_range(event, pev);
1230
1231	hwc->event_base = mipspmu_perf_event_encode(pev);
1232	if (PERF_TYPE_RAW == event->attr.type)
1233		mutex_unlock(&raw_event_mutex);
1234
1235	if (!attr->exclude_user)
1236		hwc->config_base |= M_PERFCTL_USER;
1237	if (!attr->exclude_kernel) {
1238		hwc->config_base |= M_PERFCTL_KERNEL;
1239		/* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1240		hwc->config_base |= M_PERFCTL_EXL;
1241	}
1242	if (!attr->exclude_hv)
1243		hwc->config_base |= M_PERFCTL_SUPERVISOR;
1244
1245	hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1246	/*
1247	 * The event can belong to another cpu. We do not assign a local
1248	 * counter for it for now.
1249	 */
1250	hwc->idx = -1;
1251	hwc->config = 0;
1252
1253	if (!hwc->sample_period) {
1254		hwc->sample_period  = mipspmu.max_period;
1255		hwc->last_period    = hwc->sample_period;
1256		local64_set(&hwc->period_left, hwc->sample_period);
1257	}
1258
1259	err = 0;
1260	if (event->group_leader != event)
1261		err = validate_group(event);
1262
1263	event->destroy = hw_perf_event_destroy;
1264
1265	if (err)
1266		event->destroy(event);
1267
1268	return err;
1269}
1270
1271static void pause_local_counters(void)
1272{
1273	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1274	int ctr = mipspmu.num_counters;
1275	unsigned long flags;
1276
1277	local_irq_save(flags);
1278	do {
1279		ctr--;
1280		cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1281		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1282					 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1283	} while (ctr > 0);
1284	local_irq_restore(flags);
1285}
1286
1287static void resume_local_counters(void)
1288{
1289	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1290	int ctr = mipspmu.num_counters;
1291
1292	do {
1293		ctr--;
1294		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1295	} while (ctr > 0);
1296}
1297
1298static int mipsxx_pmu_handle_shared_irq(void)
1299{
1300	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1301	struct perf_sample_data data;
1302	unsigned int counters = mipspmu.num_counters;
1303	u64 counter;
1304	int handled = IRQ_NONE;
1305	struct pt_regs *regs;
1306
1307	if (cpu_has_mips_r2 && !(read_c0_cause() & (1 << 26)))
1308		return handled;
1309	/*
1310	 * First we pause the local counters, so that when we are locked
1311	 * here, the counters are all paused. When it gets locked due to
1312	 * perf_disable(), the timer interrupt handler will be delayed.
1313	 *
1314	 * See also mipsxx_pmu_start().
1315	 */
1316	pause_local_counters();
1317#ifdef CONFIG_MIPS_MT_SMP
1318	read_lock(&pmuint_rwlock);
1319#endif
1320
1321	regs = get_irq_regs();
1322
1323	perf_sample_data_init(&data, 0, 0);
1324
1325	switch (counters) {
1326#define HANDLE_COUNTER(n)						\
1327	case n + 1:							\
1328		if (test_bit(n, cpuc->used_mask)) {			\
1329			counter = mipspmu.read_counter(n);		\
1330			if (counter & mipspmu.overflow) {		\
1331				handle_associated_event(cpuc, n, &data, regs); \
1332				handled = IRQ_HANDLED;			\
1333			}						\
1334		}
1335	HANDLE_COUNTER(3)
1336	HANDLE_COUNTER(2)
1337	HANDLE_COUNTER(1)
1338	HANDLE_COUNTER(0)
1339	}
1340
 
 
 
 
 
1341	/*
1342	 * Do all the work for the pending perf events. We can do this
1343	 * in here because the performance counter interrupt is a regular
1344	 * interrupt, not NMI.
1345	 */
1346	if (handled == IRQ_HANDLED)
1347		irq_work_run();
1348
1349#ifdef CONFIG_MIPS_MT_SMP
1350	read_unlock(&pmuint_rwlock);
1351#endif
1352	resume_local_counters();
1353	return handled;
1354}
1355
1356static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1357{
1358	return mipsxx_pmu_handle_shared_irq();
1359}
1360
1361/* 24K */
1362#define IS_BOTH_COUNTERS_24K_EVENT(b)					\
1363	((b) == 0 || (b) == 1 || (b) == 11)
1364
1365/* 34K */
1366#define IS_BOTH_COUNTERS_34K_EVENT(b)					\
1367	((b) == 0 || (b) == 1 || (b) == 11)
1368#ifdef CONFIG_MIPS_MT_SMP
1369#define IS_RANGE_P_34K_EVENT(r, b)					\
1370	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1371	 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||		\
1372	 (r) == 176 || ((b) >= 50 && (b) <= 55) ||			\
1373	 ((b) >= 64 && (b) <= 67))
1374#define IS_RANGE_V_34K_EVENT(r)	((r) == 47)
1375#endif
1376
1377/* 74K */
1378#define IS_BOTH_COUNTERS_74K_EVENT(b)					\
1379	((b) == 0 || (b) == 1)
1380
 
 
 
 
 
 
 
1381/* 1004K */
1382#define IS_BOTH_COUNTERS_1004K_EVENT(b)					\
1383	((b) == 0 || (b) == 1 || (b) == 11)
1384#ifdef CONFIG_MIPS_MT_SMP
1385#define IS_RANGE_P_1004K_EVENT(r, b)					\
1386	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1387	 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||		\
1388	 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||	\
1389	 (r) == 188 || (b) == 61 || (b) == 62 ||			\
1390	 ((b) >= 64 && (b) <= 67))
1391#define IS_RANGE_V_1004K_EVENT(r)	((r) == 47)
1392#endif
1393
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1394/*
1395 * User can use 0-255 raw events, where 0-127 for the events of even
1396 * counters, and 128-255 for odd counters. Note that bit 7 is used to
1397 * indicate the parity. So, for example, when user wants to take the
1398 * Event Num of 15 for odd counters (by referring to the user manual),
1399 * then 128 needs to be added to 15 as the input for the event config,
1400 * i.e., 143 (0x8F) to be used.
 
 
 
 
1401 */
1402static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1403{
 
1404	unsigned int raw_id = config & 0xff;
1405	unsigned int base_id = raw_id & 0x7f;
1406
1407	raw_event.event_id = base_id;
1408
1409	switch (current_cpu_type()) {
1410	case CPU_24K:
1411		if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1412			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1413		else
1414			raw_event.cntr_mask =
1415				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1416#ifdef CONFIG_MIPS_MT_SMP
1417		/*
1418		 * This is actually doing nothing. Non-multithreading
1419		 * CPUs will not check and calculate the range.
1420		 */
1421		raw_event.range = P;
1422#endif
1423		break;
1424	case CPU_34K:
1425		if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1426			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1427		else
1428			raw_event.cntr_mask =
1429				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1430#ifdef CONFIG_MIPS_MT_SMP
1431		if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1432			raw_event.range = P;
1433		else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1434			raw_event.range = V;
1435		else
1436			raw_event.range = T;
1437#endif
1438		break;
1439	case CPU_74K:
 
1440		if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1441			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1442		else
1443			raw_event.cntr_mask =
1444				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1445#ifdef CONFIG_MIPS_MT_SMP
1446		raw_event.range = P;
1447#endif
1448		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1449	case CPU_1004K:
1450		if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1451			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1452		else
1453			raw_event.cntr_mask =
1454				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1455#ifdef CONFIG_MIPS_MT_SMP
1456		if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1457			raw_event.range = P;
1458		else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1459			raw_event.range = V;
1460		else
1461			raw_event.range = T;
1462#endif
1463		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1464	}
1465
 
 
1466	return &raw_event;
1467}
1468
1469static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1470{
1471	unsigned int raw_id = config & 0xff;
1472	unsigned int base_id = raw_id & 0x7f;
1473
1474
1475	raw_event.cntr_mask = CNTR_ALL;
1476	raw_event.event_id = base_id;
1477
1478	if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1479		if (base_id > 0x42)
1480			return ERR_PTR(-EOPNOTSUPP);
1481	} else {
1482		if (base_id > 0x3a)
1483			return ERR_PTR(-EOPNOTSUPP);
1484	}
1485
1486	switch (base_id) {
1487	case 0x00:
1488	case 0x0f:
1489	case 0x1e:
1490	case 0x1f:
1491	case 0x2f:
1492	case 0x34:
1493	case 0x3b ... 0x3f:
1494		return ERR_PTR(-EOPNOTSUPP);
1495	default:
1496		break;
1497	}
1498
1499	return &raw_event;
1500}
1501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1502static int __init
1503init_hw_perf_events(void)
1504{
1505	int counters, irq;
1506	int counter_bits;
1507
1508	pr_info("Performance counters: ");
1509
1510	counters = n_counters();
1511	if (counters == 0) {
1512		pr_cont("No available PMU.\n");
1513		return -ENODEV;
1514	}
1515
1516#ifdef CONFIG_MIPS_MT_SMP
1517	cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1518	if (!cpu_has_mipsmt_pertccounters)
1519		counters = counters_total_to_per_cpu(counters);
1520#endif
1521
1522#ifdef MSC01E_INT_BASE
1523	if (cpu_has_veic) {
1524		/*
1525		 * Using platform specific interrupt controller defines.
1526		 */
1527		irq = MSC01E_INT_BASE + MSC01E_INT_PERFCTR;
1528	} else {
1529#endif
1530		if ((cp0_perfcount_irq >= 0) &&
1531				(cp0_compare_irq != cp0_perfcount_irq))
1532			irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1533		else
1534			irq = -1;
1535#ifdef MSC01E_INT_BASE
1536	}
1537#endif
1538
1539	mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1540
1541	switch (current_cpu_type()) {
1542	case CPU_24K:
1543		mipspmu.name = "mips/24K";
1544		mipspmu.general_event_map = &mipsxxcore_event_map;
1545		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1546		break;
1547	case CPU_34K:
1548		mipspmu.name = "mips/34K";
1549		mipspmu.general_event_map = &mipsxxcore_event_map;
1550		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1551		break;
1552	case CPU_74K:
1553		mipspmu.name = "mips/74K";
1554		mipspmu.general_event_map = &mipsxx74Kcore_event_map;
1555		mipspmu.cache_event_map = &mipsxx74Kcore_cache_map;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1556		break;
1557	case CPU_1004K:
1558		mipspmu.name = "mips/1004K";
1559		mipspmu.general_event_map = &mipsxxcore_event_map;
1560		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1561		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1562	case CPU_CAVIUM_OCTEON:
1563	case CPU_CAVIUM_OCTEON_PLUS:
1564	case CPU_CAVIUM_OCTEON2:
1565		mipspmu.name = "octeon";
1566		mipspmu.general_event_map = &octeon_event_map;
1567		mipspmu.cache_event_map = &octeon_cache_map;
1568		mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1569		break;
 
 
 
 
 
 
 
 
 
 
 
1570	default:
1571		pr_cont("Either hardware does not support performance "
1572			"counters, or not yet implemented.\n");
1573		return -ENODEV;
1574	}
1575
1576	mipspmu.num_counters = counters;
1577	mipspmu.irq = irq;
1578
1579	if (read_c0_perfctrl0() & M_PERFCTL_WIDE) {
1580		mipspmu.max_period = (1ULL << 63) - 1;
1581		mipspmu.valid_count = (1ULL << 63) - 1;
1582		mipspmu.overflow = 1ULL << 63;
1583		mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1584		mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1585		counter_bits = 64;
1586	} else {
1587		mipspmu.max_period = (1ULL << 31) - 1;
1588		mipspmu.valid_count = (1ULL << 31) - 1;
1589		mipspmu.overflow = 1ULL << 31;
1590		mipspmu.read_counter = mipsxx_pmu_read_counter;
1591		mipspmu.write_counter = mipsxx_pmu_write_counter;
1592		counter_bits = 32;
1593	}
1594
1595	on_each_cpu(reset_counters, (void *)(long)counters, 1);
1596
1597	pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1598		"CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1599		irq < 0 ? " (share with timer interrupt)" : "");
1600
1601	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1602
1603	return 0;
1604}
1605early_initcall(init_hw_perf_events);
v5.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Linux performance counter support for MIPS.
   4 *
   5 * Copyright (C) 2010 MIPS Technologies, Inc.
   6 * Copyright (C) 2011 Cavium Networks, Inc.
   7 * Author: Deng-Cheng Zhu
   8 *
   9 * This code is based on the implementation for ARM, which is in turn
  10 * based on the sparc64 perf event code and the x86 code. Performance
  11 * counter access is based on the MIPS Oprofile code. And the callchain
  12 * support references the code of MIPS stacktrace.c.
 
 
 
 
  13 */
  14
  15#include <linux/cpumask.h>
  16#include <linux/interrupt.h>
  17#include <linux/smp.h>
  18#include <linux/kernel.h>
  19#include <linux/perf_event.h>
  20#include <linux/uaccess.h>
  21
  22#include <asm/irq.h>
  23#include <asm/irq_regs.h>
  24#include <asm/stacktrace.h>
  25#include <asm/time.h> /* For perf_irq */
  26
  27#define MIPS_MAX_HWEVENTS 4
  28#define MIPS_TCS_PER_COUNTER 2
  29#define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
  30
  31struct cpu_hw_events {
  32	/* Array of events on this cpu. */
  33	struct perf_event	*events[MIPS_MAX_HWEVENTS];
  34
  35	/*
  36	 * Set the bit (indexed by the counter number) when the counter
  37	 * is used for an event.
  38	 */
  39	unsigned long		used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
  40
  41	/*
  42	 * Software copy of the control register for each performance counter.
  43	 * MIPS CPUs vary in performance counters. They use this differently,
  44	 * and even may not use it.
  45	 */
  46	unsigned int		saved_ctrl[MIPS_MAX_HWEVENTS];
  47};
  48DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
  49	.saved_ctrl = {0},
  50};
  51
  52/* The description of MIPS performance events. */
  53struct mips_perf_event {
  54	unsigned int event_id;
  55	/*
  56	 * MIPS performance counters are indexed starting from 0.
  57	 * CNTR_EVEN indicates the indexes of the counters to be used are
  58	 * even numbers.
  59	 */
  60	unsigned int cntr_mask;
  61	#define CNTR_EVEN	0x55555555
  62	#define CNTR_ODD	0xaaaaaaaa
  63	#define CNTR_ALL	0xffffffff
 
  64	enum {
  65		T  = 0,
  66		V  = 1,
  67		P  = 2,
  68	} range;
 
 
 
 
 
  69};
  70
  71static struct mips_perf_event raw_event;
  72static DEFINE_MUTEX(raw_event_mutex);
  73
 
  74#define C(x) PERF_COUNT_HW_CACHE_##x
  75
  76struct mips_pmu {
  77	u64		max_period;
  78	u64		valid_count;
  79	u64		overflow;
  80	const char	*name;
  81	int		irq;
  82	u64		(*read_counter)(unsigned int idx);
  83	void		(*write_counter)(unsigned int idx, u64 val);
  84	const struct mips_perf_event *(*map_raw_event)(u64 config);
  85	const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
  86	const struct mips_perf_event (*cache_event_map)
  87				[PERF_COUNT_HW_CACHE_MAX]
  88				[PERF_COUNT_HW_CACHE_OP_MAX]
  89				[PERF_COUNT_HW_CACHE_RESULT_MAX];
  90	unsigned int	num_counters;
  91};
  92
  93static struct mips_pmu mipspmu;
  94
  95#define M_PERFCTL_EVENT(event)		(((event) << MIPS_PERFCTRL_EVENT_S) & \
  96					 MIPS_PERFCTRL_EVENT)
  97#define M_PERFCTL_VPEID(vpe)		((vpe)	  << MIPS_PERFCTRL_VPEID_S)
  98
  99#ifdef CONFIG_CPU_BMIPS5000
 100#define M_PERFCTL_MT_EN(filter)		0
 101#else /* !CONFIG_CPU_BMIPS5000 */
 102#define M_PERFCTL_MT_EN(filter)		(filter)
 103#endif /* CONFIG_CPU_BMIPS5000 */
 104
 105#define	   M_TC_EN_ALL			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_ALL)
 106#define	   M_TC_EN_VPE			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_VPE)
 107#define	   M_TC_EN_TC			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_TC)
 108
 109#define M_PERFCTL_COUNT_EVENT_WHENEVER	(MIPS_PERFCTRL_EXL |		\
 110					 MIPS_PERFCTRL_K |		\
 111					 MIPS_PERFCTRL_U |		\
 112					 MIPS_PERFCTRL_S |		\
 113					 MIPS_PERFCTRL_IE)
 
 
 
 114
 115#ifdef CONFIG_MIPS_MT_SMP
 116#define M_PERFCTL_CONFIG_MASK		0x3fff801f
 117#else
 118#define M_PERFCTL_CONFIG_MASK		0x1f
 119#endif
 
 120
 121
 122#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
 
 
 123static DEFINE_RWLOCK(pmuint_rwlock);
 124
 125#if defined(CONFIG_CPU_BMIPS5000)
 
 
 
 
 126#define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
 127			 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
 128#else
 129#define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
 130			 0 : cpu_vpe_id(&current_cpu_data))
 131#endif
 132
 133/* Copied from op_model_mipsxx.c */
 134static unsigned int vpe_shift(void)
 135{
 136	if (num_possible_cpus() > 1)
 137		return 1;
 138
 139	return 0;
 140}
 141
 142static unsigned int counters_total_to_per_cpu(unsigned int counters)
 143{
 144	return counters >> vpe_shift();
 145}
 146
 147#else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
 148#define vpe_id()	0
 149
 150#endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
 151
 152static void resume_local_counters(void);
 153static void pause_local_counters(void);
 154static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
 155static int mipsxx_pmu_handle_shared_irq(void);
 156
 157static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
 158{
 159	if (vpe_id() == 1)
 160		idx = (idx + 2) & 3;
 161	return idx;
 162}
 163
 164static u64 mipsxx_pmu_read_counter(unsigned int idx)
 165{
 166	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 167
 168	switch (idx) {
 169	case 0:
 170		/*
 171		 * The counters are unsigned, we must cast to truncate
 172		 * off the high bits.
 173		 */
 174		return (u32)read_c0_perfcntr0();
 175	case 1:
 176		return (u32)read_c0_perfcntr1();
 177	case 2:
 178		return (u32)read_c0_perfcntr2();
 179	case 3:
 180		return (u32)read_c0_perfcntr3();
 181	default:
 182		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 183		return 0;
 184	}
 185}
 186
 187static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
 188{
 189	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 190
 191	switch (idx) {
 192	case 0:
 193		return read_c0_perfcntr0_64();
 194	case 1:
 195		return read_c0_perfcntr1_64();
 196	case 2:
 197		return read_c0_perfcntr2_64();
 198	case 3:
 199		return read_c0_perfcntr3_64();
 200	default:
 201		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 202		return 0;
 203	}
 204}
 205
 206static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
 207{
 208	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 209
 210	switch (idx) {
 211	case 0:
 212		write_c0_perfcntr0(val);
 213		return;
 214	case 1:
 215		write_c0_perfcntr1(val);
 216		return;
 217	case 2:
 218		write_c0_perfcntr2(val);
 219		return;
 220	case 3:
 221		write_c0_perfcntr3(val);
 222		return;
 223	}
 224}
 225
 226static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
 227{
 228	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 229
 230	switch (idx) {
 231	case 0:
 232		write_c0_perfcntr0_64(val);
 233		return;
 234	case 1:
 235		write_c0_perfcntr1_64(val);
 236		return;
 237	case 2:
 238		write_c0_perfcntr2_64(val);
 239		return;
 240	case 3:
 241		write_c0_perfcntr3_64(val);
 242		return;
 243	}
 244}
 245
 246static unsigned int mipsxx_pmu_read_control(unsigned int idx)
 247{
 248	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 249
 250	switch (idx) {
 251	case 0:
 252		return read_c0_perfctrl0();
 253	case 1:
 254		return read_c0_perfctrl1();
 255	case 2:
 256		return read_c0_perfctrl2();
 257	case 3:
 258		return read_c0_perfctrl3();
 259	default:
 260		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
 261		return 0;
 262	}
 263}
 264
 265static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
 266{
 267	idx = mipsxx_pmu_swizzle_perf_idx(idx);
 268
 269	switch (idx) {
 270	case 0:
 271		write_c0_perfctrl0(val);
 272		return;
 273	case 1:
 274		write_c0_perfctrl1(val);
 275		return;
 276	case 2:
 277		write_c0_perfctrl2(val);
 278		return;
 279	case 3:
 280		write_c0_perfctrl3(val);
 281		return;
 282	}
 283}
 284
 285static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
 286				    struct hw_perf_event *hwc)
 287{
 288	int i;
 289
 290	/*
 291	 * We only need to care the counter mask. The range has been
 292	 * checked definitely.
 293	 */
 294	unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
 295
 296	for (i = mipspmu.num_counters - 1; i >= 0; i--) {
 297		/*
 298		 * Note that some MIPS perf events can be counted by both
 299		 * even and odd counters, wheresas many other are only by
 300		 * even _or_ odd counters. This introduces an issue that
 301		 * when the former kind of event takes the counter the
 302		 * latter kind of event wants to use, then the "counter
 303		 * allocation" for the latter event will fail. In fact if
 304		 * they can be dynamically swapped, they both feel happy.
 305		 * But here we leave this issue alone for now.
 306		 */
 307		if (test_bit(i, &cntr_mask) &&
 308			!test_and_set_bit(i, cpuc->used_mask))
 309			return i;
 310	}
 311
 312	return -EAGAIN;
 313}
 314
 315static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
 316{
 317	struct perf_event *event = container_of(evt, struct perf_event, hw);
 318	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 319	unsigned int range = evt->event_base >> 24;
 320
 321	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 322
 323	cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
 324		(evt->config_base & M_PERFCTL_CONFIG_MASK) |
 325		/* Make sure interrupt enabled. */
 326		MIPS_PERFCTRL_IE;
 327
 328	if (IS_ENABLED(CONFIG_CPU_BMIPS5000)) {
 329		/* enable the counter for the calling thread */
 330		cpuc->saved_ctrl[idx] |=
 331			(1 << (12 + vpe_id())) | BRCM_PERFCTRL_TC;
 332	} else if (IS_ENABLED(CONFIG_MIPS_MT_SMP) && range > V) {
 333		/* The counter is processor wide. Set it up to count all TCs. */
 334		pr_debug("Enabling perf counter for all TCs\n");
 335		cpuc->saved_ctrl[idx] |= M_TC_EN_ALL;
 336	} else {
 337		unsigned int cpu, ctrl;
 338
 339		/*
 340		 * Set up the counter for a particular CPU when event->cpu is
 341		 * a valid CPU number. Otherwise set up the counter for the CPU
 342		 * scheduling this thread.
 343		 */
 344		cpu = (event->cpu >= 0) ? event->cpu : smp_processor_id();
 345
 346		ctrl = M_PERFCTL_VPEID(cpu_vpe_id(&cpu_data[cpu]));
 347		ctrl |= M_TC_EN_VPE;
 348		cpuc->saved_ctrl[idx] |= ctrl;
 349		pr_debug("Enabling perf counter for CPU%d\n", cpu);
 350	}
 351	/*
 352	 * We do not actually let the counter run. Leave it until start().
 353	 */
 354}
 355
 356static void mipsxx_pmu_disable_event(int idx)
 357{
 358	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 359	unsigned long flags;
 360
 361	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 362
 363	local_irq_save(flags);
 364	cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
 365		~M_PERFCTL_COUNT_EVENT_WHENEVER;
 366	mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
 367	local_irq_restore(flags);
 368}
 369
 370static int mipspmu_event_set_period(struct perf_event *event,
 371				    struct hw_perf_event *hwc,
 372				    int idx)
 373{
 374	u64 left = local64_read(&hwc->period_left);
 375	u64 period = hwc->sample_period;
 376	int ret = 0;
 377
 378	if (unlikely((left + period) & (1ULL << 63))) {
 379		/* left underflowed by more than period. */
 380		left = period;
 381		local64_set(&hwc->period_left, left);
 382		hwc->last_period = period;
 383		ret = 1;
 384	} else	if (unlikely((left + period) <= period)) {
 385		/* left underflowed by less than period. */
 386		left += period;
 387		local64_set(&hwc->period_left, left);
 388		hwc->last_period = period;
 389		ret = 1;
 390	}
 391
 392	if (left > mipspmu.max_period) {
 393		left = mipspmu.max_period;
 394		local64_set(&hwc->period_left, left);
 395	}
 396
 397	local64_set(&hwc->prev_count, mipspmu.overflow - left);
 398
 399	mipspmu.write_counter(idx, mipspmu.overflow - left);
 400
 401	perf_event_update_userpage(event);
 402
 403	return ret;
 404}
 405
 406static void mipspmu_event_update(struct perf_event *event,
 407				 struct hw_perf_event *hwc,
 408				 int idx)
 409{
 410	u64 prev_raw_count, new_raw_count;
 411	u64 delta;
 412
 413again:
 414	prev_raw_count = local64_read(&hwc->prev_count);
 415	new_raw_count = mipspmu.read_counter(idx);
 416
 417	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 418				new_raw_count) != prev_raw_count)
 419		goto again;
 420
 421	delta = new_raw_count - prev_raw_count;
 422
 423	local64_add(delta, &event->count);
 424	local64_sub(delta, &hwc->period_left);
 425}
 426
 427static void mipspmu_start(struct perf_event *event, int flags)
 428{
 429	struct hw_perf_event *hwc = &event->hw;
 430
 431	if (flags & PERF_EF_RELOAD)
 432		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 433
 434	hwc->state = 0;
 435
 436	/* Set the period for the event. */
 437	mipspmu_event_set_period(event, hwc, hwc->idx);
 438
 439	/* Enable the event. */
 440	mipsxx_pmu_enable_event(hwc, hwc->idx);
 441}
 442
 443static void mipspmu_stop(struct perf_event *event, int flags)
 444{
 445	struct hw_perf_event *hwc = &event->hw;
 446
 447	if (!(hwc->state & PERF_HES_STOPPED)) {
 448		/* We are working on a local event. */
 449		mipsxx_pmu_disable_event(hwc->idx);
 450		barrier();
 451		mipspmu_event_update(event, hwc, hwc->idx);
 452		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
 453	}
 454}
 455
 456static int mipspmu_add(struct perf_event *event, int flags)
 457{
 458	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 459	struct hw_perf_event *hwc = &event->hw;
 460	int idx;
 461	int err = 0;
 462
 463	perf_pmu_disable(event->pmu);
 464
 465	/* To look for a free counter for this event. */
 466	idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
 467	if (idx < 0) {
 468		err = idx;
 469		goto out;
 470	}
 471
 472	/*
 473	 * If there is an event in the counter we are going to use then
 474	 * make sure it is disabled.
 475	 */
 476	event->hw.idx = idx;
 477	mipsxx_pmu_disable_event(idx);
 478	cpuc->events[idx] = event;
 479
 480	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
 481	if (flags & PERF_EF_START)
 482		mipspmu_start(event, PERF_EF_RELOAD);
 483
 484	/* Propagate our changes to the userspace mapping. */
 485	perf_event_update_userpage(event);
 486
 487out:
 488	perf_pmu_enable(event->pmu);
 489	return err;
 490}
 491
 492static void mipspmu_del(struct perf_event *event, int flags)
 493{
 494	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 495	struct hw_perf_event *hwc = &event->hw;
 496	int idx = hwc->idx;
 497
 498	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
 499
 500	mipspmu_stop(event, PERF_EF_UPDATE);
 501	cpuc->events[idx] = NULL;
 502	clear_bit(idx, cpuc->used_mask);
 503
 504	perf_event_update_userpage(event);
 505}
 506
 507static void mipspmu_read(struct perf_event *event)
 508{
 509	struct hw_perf_event *hwc = &event->hw;
 510
 511	/* Don't read disabled counters! */
 512	if (hwc->idx < 0)
 513		return;
 514
 515	mipspmu_event_update(event, hwc, hwc->idx);
 516}
 517
 518static void mipspmu_enable(struct pmu *pmu)
 519{
 520#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
 521	write_unlock(&pmuint_rwlock);
 522#endif
 523	resume_local_counters();
 524}
 525
 526/*
 527 * MIPS performance counters can be per-TC. The control registers can
 528 * not be directly accessed across CPUs. Hence if we want to do global
 529 * control, we need cross CPU calls. on_each_cpu() can help us, but we
 530 * can not make sure this function is called with interrupts enabled. So
 531 * here we pause local counters and then grab a rwlock and leave the
 532 * counters on other CPUs alone. If any counter interrupt raises while
 533 * we own the write lock, simply pause local counters on that CPU and
 534 * spin in the handler. Also we know we won't be switched to another
 535 * CPU after pausing local counters and before grabbing the lock.
 536 */
 537static void mipspmu_disable(struct pmu *pmu)
 538{
 539	pause_local_counters();
 540#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
 541	write_lock(&pmuint_rwlock);
 542#endif
 543}
 544
 545static atomic_t active_events = ATOMIC_INIT(0);
 546static DEFINE_MUTEX(pmu_reserve_mutex);
 547static int (*save_perf_irq)(void);
 548
 549static int mipspmu_get_irq(void)
 550{
 551	int err;
 552
 553	if (mipspmu.irq >= 0) {
 554		/* Request my own irq handler. */
 555		err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
 556				  IRQF_PERCPU | IRQF_NOBALANCING |
 557				  IRQF_NO_THREAD | IRQF_NO_SUSPEND |
 558				  IRQF_SHARED,
 559				  "mips_perf_pmu", &mipspmu);
 560		if (err) {
 561			pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
 562				mipspmu.irq);
 563		}
 564	} else if (cp0_perfcount_irq < 0) {
 565		/*
 566		 * We are sharing the irq number with the timer interrupt.
 567		 */
 568		save_perf_irq = perf_irq;
 569		perf_irq = mipsxx_pmu_handle_shared_irq;
 570		err = 0;
 571	} else {
 572		pr_warn("The platform hasn't properly defined its interrupt controller\n");
 
 573		err = -ENOENT;
 574	}
 575
 576	return err;
 577}
 578
 579static void mipspmu_free_irq(void)
 580{
 581	if (mipspmu.irq >= 0)
 582		free_irq(mipspmu.irq, &mipspmu);
 583	else if (cp0_perfcount_irq < 0)
 584		perf_irq = save_perf_irq;
 585}
 586
 587/*
 588 * mipsxx/rm9000/loongson2 have different performance counters, they have
 589 * specific low-level init routines.
 590 */
 591static void reset_counters(void *arg);
 592static int __hw_perf_event_init(struct perf_event *event);
 593
 594static void hw_perf_event_destroy(struct perf_event *event)
 595{
 596	if (atomic_dec_and_mutex_lock(&active_events,
 597				&pmu_reserve_mutex)) {
 598		/*
 599		 * We must not call the destroy function with interrupts
 600		 * disabled.
 601		 */
 602		on_each_cpu(reset_counters,
 603			(void *)(long)mipspmu.num_counters, 1);
 604		mipspmu_free_irq();
 605		mutex_unlock(&pmu_reserve_mutex);
 606	}
 607}
 608
 609static int mipspmu_event_init(struct perf_event *event)
 610{
 611	int err = 0;
 612
 613	/* does not support taken branch sampling */
 614	if (has_branch_stack(event))
 615		return -EOPNOTSUPP;
 616
 617	switch (event->attr.type) {
 618	case PERF_TYPE_RAW:
 619	case PERF_TYPE_HARDWARE:
 620	case PERF_TYPE_HW_CACHE:
 621		break;
 622
 623	default:
 624		return -ENOENT;
 625	}
 626
 627	if (event->cpu >= 0 && !cpu_online(event->cpu))
 
 628		return -ENODEV;
 629
 630	if (!atomic_inc_not_zero(&active_events)) {
 631		mutex_lock(&pmu_reserve_mutex);
 632		if (atomic_read(&active_events) == 0)
 633			err = mipspmu_get_irq();
 634
 635		if (!err)
 636			atomic_inc(&active_events);
 637		mutex_unlock(&pmu_reserve_mutex);
 638	}
 639
 640	if (err)
 641		return err;
 642
 643	return __hw_perf_event_init(event);
 644}
 645
 646static struct pmu pmu = {
 647	.pmu_enable	= mipspmu_enable,
 648	.pmu_disable	= mipspmu_disable,
 649	.event_init	= mipspmu_event_init,
 650	.add		= mipspmu_add,
 651	.del		= mipspmu_del,
 652	.start		= mipspmu_start,
 653	.stop		= mipspmu_stop,
 654	.read		= mipspmu_read,
 655};
 656
 657static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
 658{
 659/*
 660 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
 661 * event_id.
 662 */
 663#ifdef CONFIG_MIPS_MT_SMP
 664	if (num_possible_cpus() > 1)
 665		return ((unsigned int)pev->range << 24) |
 666			(pev->cntr_mask & 0xffff00) |
 667			(pev->event_id & 0xff);
 668	else
 669#endif /* CONFIG_MIPS_MT_SMP */
 670		return ((pev->cntr_mask & 0xffff00) |
 671			(pev->event_id & 0xff));
 672}
 673
 674static const struct mips_perf_event *mipspmu_map_general_event(int idx)
 675{
 
 676
 677	if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
 678		return ERR_PTR(-EOPNOTSUPP);
 679	return &(*mipspmu.general_event_map)[idx];
 
 
 680}
 681
 682static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
 683{
 684	unsigned int cache_type, cache_op, cache_result;
 685	const struct mips_perf_event *pev;
 686
 687	cache_type = (config >> 0) & 0xff;
 688	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 689		return ERR_PTR(-EINVAL);
 690
 691	cache_op = (config >> 8) & 0xff;
 692	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 693		return ERR_PTR(-EINVAL);
 694
 695	cache_result = (config >> 16) & 0xff;
 696	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 697		return ERR_PTR(-EINVAL);
 698
 699	pev = &((*mipspmu.cache_event_map)
 700					[cache_type]
 701					[cache_op]
 702					[cache_result]);
 703
 704	if (pev->cntr_mask == 0)
 705		return ERR_PTR(-EOPNOTSUPP);
 706
 707	return pev;
 708
 709}
 710
 711static int validate_group(struct perf_event *event)
 712{
 713	struct perf_event *sibling, *leader = event->group_leader;
 714	struct cpu_hw_events fake_cpuc;
 715
 716	memset(&fake_cpuc, 0, sizeof(fake_cpuc));
 717
 718	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
 719		return -EINVAL;
 720
 721	for_each_sibling_event(sibling, leader) {
 722		if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
 723			return -EINVAL;
 724	}
 725
 726	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
 727		return -EINVAL;
 728
 729	return 0;
 730}
 731
 732/* This is needed by specific irq handlers in perf_event_*.c */
 733static void handle_associated_event(struct cpu_hw_events *cpuc,
 734				    int idx, struct perf_sample_data *data,
 735				    struct pt_regs *regs)
 736{
 737	struct perf_event *event = cpuc->events[idx];
 738	struct hw_perf_event *hwc = &event->hw;
 739
 740	mipspmu_event_update(event, hwc, idx);
 741	data->period = event->hw.last_period;
 742	if (!mipspmu_event_set_period(event, hwc, idx))
 743		return;
 744
 745	if (perf_event_overflow(event, data, regs))
 746		mipsxx_pmu_disable_event(idx);
 747}
 748
 749
 750static int __n_counters(void)
 751{
 752	if (!cpu_has_perf)
 753		return 0;
 754	if (!(read_c0_perfctrl0() & MIPS_PERFCTRL_M))
 755		return 1;
 756	if (!(read_c0_perfctrl1() & MIPS_PERFCTRL_M))
 757		return 2;
 758	if (!(read_c0_perfctrl2() & MIPS_PERFCTRL_M))
 759		return 3;
 760
 761	return 4;
 762}
 763
 764static int n_counters(void)
 765{
 766	int counters;
 767
 768	switch (current_cpu_type()) {
 769	case CPU_R10000:
 770		counters = 2;
 771		break;
 772
 773	case CPU_R12000:
 774	case CPU_R14000:
 775	case CPU_R16000:
 776		counters = 4;
 777		break;
 778
 779	default:
 780		counters = __n_counters();
 781	}
 782
 783	return counters;
 784}
 785
 786static void reset_counters(void *arg)
 787{
 788	int counters = (int)(long)arg;
 789	switch (counters) {
 790	case 4:
 791		mipsxx_pmu_write_control(3, 0);
 792		mipspmu.write_counter(3, 0);
 793		/* fall through */
 794	case 3:
 795		mipsxx_pmu_write_control(2, 0);
 796		mipspmu.write_counter(2, 0);
 797		/* fall through */
 798	case 2:
 799		mipsxx_pmu_write_control(1, 0);
 800		mipspmu.write_counter(1, 0);
 801		/* fall through */
 802	case 1:
 803		mipsxx_pmu_write_control(0, 0);
 804		mipspmu.write_counter(0, 0);
 805		/* fall through */
 806	}
 807}
 808
 809/* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
 810static const struct mips_perf_event mipsxxcore_event_map
 811				[PERF_COUNT_HW_MAX] = {
 812	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
 813	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
 
 
 814	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
 815	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
 
 816};
 817
 818/* 74K/proAptiv core has different branch event code. */
 819static const struct mips_perf_event mipsxxcore_event_map2
 820				[PERF_COUNT_HW_MAX] = {
 821	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
 822	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
 
 
 823	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
 824	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
 825};
 826
 827static const struct mips_perf_event i6x00_event_map[PERF_COUNT_HW_MAX] = {
 828	[PERF_COUNT_HW_CPU_CYCLES]          = { 0x00, CNTR_EVEN | CNTR_ODD },
 829	[PERF_COUNT_HW_INSTRUCTIONS]        = { 0x01, CNTR_EVEN | CNTR_ODD },
 830	/* These only count dcache, not icache */
 831	[PERF_COUNT_HW_CACHE_REFERENCES]    = { 0x45, CNTR_EVEN | CNTR_ODD },
 832	[PERF_COUNT_HW_CACHE_MISSES]        = { 0x48, CNTR_EVEN | CNTR_ODD },
 833	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD },
 834	[PERF_COUNT_HW_BRANCH_MISSES]       = { 0x16, CNTR_EVEN | CNTR_ODD },
 835};
 836
 837static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = {
 838	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
 839	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
 840	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
 841	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
 842};
 843
 844static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
 845	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
 846	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
 847	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
 848	[PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL	 },
 849	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
 850	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
 851	[PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
 852};
 853
 854static const struct mips_perf_event bmips5000_event_map
 855				[PERF_COUNT_HW_MAX] = {
 856	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
 857	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
 858	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
 859};
 860
 861static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = {
 862	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
 863	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */
 864	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
 865	[PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
 866	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */
 867	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */
 868};
 869
 870/* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
 871static const struct mips_perf_event mipsxxcore_cache_map
 872				[PERF_COUNT_HW_CACHE_MAX]
 873				[PERF_COUNT_HW_CACHE_OP_MAX]
 874				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
 875[C(L1D)] = {
 876	/*
 877	 * Like some other architectures (e.g. ARM), the performance
 878	 * counters don't differentiate between read and write
 879	 * accesses/misses, so this isn't strictly correct, but it's the
 880	 * best we can do. Writes and reads get combined.
 881	 */
 882	[C(OP_READ)] = {
 883		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
 884		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
 885	},
 886	[C(OP_WRITE)] = {
 887		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
 888		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
 889	},
 
 
 
 
 890},
 891[C(L1I)] = {
 892	[C(OP_READ)] = {
 893		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
 894		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
 895	},
 896	[C(OP_WRITE)] = {
 897		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
 898		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
 899	},
 900	[C(OP_PREFETCH)] = {
 901		[C(RESULT_ACCESS)]	= { 0x14, CNTR_EVEN, T },
 902		/*
 903		 * Note that MIPS has only "hit" events countable for
 904		 * the prefetch operation.
 905		 */
 
 906	},
 907},
 908[C(LL)] = {
 909	[C(OP_READ)] = {
 910		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
 911		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
 912	},
 913	[C(OP_WRITE)] = {
 914		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
 915		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
 916	},
 
 
 
 
 917},
 918[C(DTLB)] = {
 919	[C(OP_READ)] = {
 920		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 921		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 922	},
 923	[C(OP_WRITE)] = {
 924		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 925		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 926	},
 
 
 
 
 927},
 928[C(ITLB)] = {
 929	[C(OP_READ)] = {
 930		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
 931		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
 932	},
 933	[C(OP_WRITE)] = {
 934		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
 935		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
 936	},
 
 
 
 
 937},
 938[C(BPU)] = {
 939	/* Using the same code for *HW_BRANCH* */
 940	[C(OP_READ)] = {
 941		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
 942		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
 943	},
 944	[C(OP_WRITE)] = {
 945		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
 946		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
 947	},
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 948},
 949};
 950
 951/* 74K/proAptiv core has completely different cache event map. */
 952static const struct mips_perf_event mipsxxcore_cache_map2
 953				[PERF_COUNT_HW_CACHE_MAX]
 954				[PERF_COUNT_HW_CACHE_OP_MAX]
 955				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
 956[C(L1D)] = {
 957	/*
 958	 * Like some other architectures (e.g. ARM), the performance
 959	 * counters don't differentiate between read and write
 960	 * accesses/misses, so this isn't strictly correct, but it's the
 961	 * best we can do. Writes and reads get combined.
 962	 */
 963	[C(OP_READ)] = {
 964		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
 965		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
 966	},
 967	[C(OP_WRITE)] = {
 968		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
 969		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
 970	},
 
 
 
 
 971},
 972[C(L1I)] = {
 973	[C(OP_READ)] = {
 974		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 975		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 976	},
 977	[C(OP_WRITE)] = {
 978		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
 979		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
 980	},
 981	[C(OP_PREFETCH)] = {
 982		[C(RESULT_ACCESS)]	= { 0x34, CNTR_EVEN, T },
 983		/*
 984		 * Note that MIPS has only "hit" events countable for
 985		 * the prefetch operation.
 986		 */
 
 987	},
 988},
 989[C(LL)] = {
 990	[C(OP_READ)] = {
 991		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
 992		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
 993	},
 994	[C(OP_WRITE)] = {
 995		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
 996		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 997	},
 998},
 999/*
1000 * 74K core does not have specific DTLB events. proAptiv core has
1001 * "speculative" DTLB events which are numbered 0x63 (even/odd) and
1002 * not included here. One can use raw events if really needed.
1003 */
1004[C(ITLB)] = {
1005	[C(OP_READ)] = {
1006		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1007		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1008	},
1009	[C(OP_WRITE)] = {
1010		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1011		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1012	},
 
 
 
 
1013},
1014[C(BPU)] = {
1015	/* Using the same code for *HW_BRANCH* */
1016	[C(OP_READ)] = {
1017		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1018		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1019	},
1020	[C(OP_WRITE)] = {
1021		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1022		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1023	},
1024},
1025};
1026
1027static const struct mips_perf_event i6x00_cache_map
1028				[PERF_COUNT_HW_CACHE_MAX]
1029				[PERF_COUNT_HW_CACHE_OP_MAX]
1030				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1031[C(L1D)] = {
1032	[C(OP_READ)] = {
1033		[C(RESULT_ACCESS)]	= { 0x46, CNTR_EVEN | CNTR_ODD },
1034		[C(RESULT_MISS)]	= { 0x49, CNTR_EVEN | CNTR_ODD },
1035	},
1036	[C(OP_WRITE)] = {
1037		[C(RESULT_ACCESS)]	= { 0x47, CNTR_EVEN | CNTR_ODD },
1038		[C(RESULT_MISS)]	= { 0x4a, CNTR_EVEN | CNTR_ODD },
1039	},
1040},
1041[C(L1I)] = {
1042	[C(OP_READ)] = {
1043		[C(RESULT_ACCESS)]	= { 0x84, CNTR_EVEN | CNTR_ODD },
1044		[C(RESULT_MISS)]	= { 0x85, CNTR_EVEN | CNTR_ODD },
1045	},
1046},
1047[C(DTLB)] = {
1048	/* Can't distinguish read & write */
1049	[C(OP_READ)] = {
1050		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1051		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1052	},
1053	[C(OP_WRITE)] = {
1054		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1055		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1056	},
1057},
1058[C(BPU)] = {
1059	/* Conditional branches / mispredicted */
1060	[C(OP_READ)] = {
1061		[C(RESULT_ACCESS)]	= { 0x15, CNTR_EVEN | CNTR_ODD },
1062		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN | CNTR_ODD },
1063	},
1064},
1065};
1066
1067static const struct mips_perf_event loongson3_cache_map
 
1068				[PERF_COUNT_HW_CACHE_MAX]
1069				[PERF_COUNT_HW_CACHE_OP_MAX]
1070				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1071[C(L1D)] = {
1072	/*
1073	 * Like some other architectures (e.g. ARM), the performance
1074	 * counters don't differentiate between read and write
1075	 * accesses/misses, so this isn't strictly correct, but it's the
1076	 * best we can do. Writes and reads get combined.
1077	 */
1078	[C(OP_READ)] = {
1079		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
 
1080	},
1081	[C(OP_WRITE)] = {
1082		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
 
1083	},
1084},
1085[C(L1I)] = {
1086	[C(OP_READ)] = {
1087		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1088	},
1089	[C(OP_WRITE)] = {
1090		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1091	},
1092},
1093[C(DTLB)] = {
1094	[C(OP_READ)] = {
1095		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1096	},
1097	[C(OP_WRITE)] = {
1098		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1099	},
1100},
1101[C(ITLB)] = {
1102	[C(OP_READ)] = {
1103		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1104	},
1105	[C(OP_WRITE)] = {
1106		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1107	},
1108},
1109[C(BPU)] = {
1110	/* Using the same code for *HW_BRANCH* */
1111	[C(OP_READ)] = {
1112		[C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN },
1113		[C(RESULT_MISS)]        = { 0x02, CNTR_ODD },
1114	},
1115	[C(OP_WRITE)] = {
1116		[C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN },
1117		[C(RESULT_MISS)]        = { 0x02, CNTR_ODD },
1118	},
1119},
1120};
1121
1122/* BMIPS5000 */
1123static const struct mips_perf_event bmips5000_cache_map
1124				[PERF_COUNT_HW_CACHE_MAX]
1125				[PERF_COUNT_HW_CACHE_OP_MAX]
1126				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1127[C(L1D)] = {
1128	/*
1129	 * Like some other architectures (e.g. ARM), the performance
1130	 * counters don't differentiate between read and write
1131	 * accesses/misses, so this isn't strictly correct, but it's the
1132	 * best we can do. Writes and reads get combined.
1133	 */
1134	[C(OP_READ)] = {
1135		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1136		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1137	},
1138	[C(OP_WRITE)] = {
1139		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1140		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1141	},
1142},
1143[C(L1I)] = {
1144	[C(OP_READ)] = {
1145		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1146		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1147	},
1148	[C(OP_WRITE)] = {
1149		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1150		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1151	},
1152	[C(OP_PREFETCH)] = {
1153		[C(RESULT_ACCESS)]	= { 23, CNTR_EVEN, T },
1154		/*
1155		 * Note that MIPS has only "hit" events countable for
1156		 * the prefetch operation.
1157		 */
1158	},
1159},
1160[C(LL)] = {
1161	[C(OP_READ)] = {
1162		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1163		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1164	},
1165	[C(OP_WRITE)] = {
1166		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1167		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1168	},
1169},
1170[C(BPU)] = {
1171	/* Using the same code for *HW_BRANCH* */
1172	[C(OP_READ)] = {
1173		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1174	},
1175	[C(OP_WRITE)] = {
1176		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1177	},
1178},
1179};
1180
1181
1182static const struct mips_perf_event octeon_cache_map
1183				[PERF_COUNT_HW_CACHE_MAX]
1184				[PERF_COUNT_HW_CACHE_OP_MAX]
1185				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1186[C(L1D)] = {
1187	[C(OP_READ)] = {
1188		[C(RESULT_ACCESS)]	= { 0x2b, CNTR_ALL },
1189		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL },
1190	},
1191	[C(OP_WRITE)] = {
1192		[C(RESULT_ACCESS)]	= { 0x30, CNTR_ALL },
1193	},
1194},
1195[C(L1I)] = {
1196	[C(OP_READ)] = {
1197		[C(RESULT_ACCESS)]	= { 0x18, CNTR_ALL },
1198	},
1199	[C(OP_PREFETCH)] = {
1200		[C(RESULT_ACCESS)]	= { 0x19, CNTR_ALL },
 
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)]	= { 0x35, CNTR_ALL },
1210	},
1211	[C(OP_WRITE)] = {
 
1212		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1213	},
 
 
 
 
1214},
1215[C(ITLB)] = {
1216	[C(OP_READ)] = {
 
1217		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL },
1218	},
1219},
1220};
1221
1222static const struct mips_perf_event xlp_cache_map
1223				[PERF_COUNT_HW_CACHE_MAX]
1224				[PERF_COUNT_HW_CACHE_OP_MAX]
1225				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1226[C(L1D)] = {
1227	[C(OP_READ)] = {
1228		[C(RESULT_ACCESS)]	= { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */
1229		[C(RESULT_MISS)]	= { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */
1230	},
1231	[C(OP_WRITE)] = {
1232		[C(RESULT_ACCESS)]	= { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */
1233		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */
1234	},
1235},
1236[C(L1I)] = {
1237	[C(OP_READ)] = {
1238		[C(RESULT_ACCESS)]	= { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
1239		[C(RESULT_MISS)]	= { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
1240	},
1241},
1242[C(LL)] = {
 
1243	[C(OP_READ)] = {
1244		[C(RESULT_ACCESS)]	= { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */
1245		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */
1246	},
1247	[C(OP_WRITE)] = {
1248		[C(RESULT_ACCESS)]	= { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */
1249		[C(RESULT_MISS)]	= { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */
1250	},
1251},
1252[C(DTLB)] = {
1253	/*
1254	 * Only general DTLB misses are counted use the same event for
1255	 * read and write.
1256	 */
1257	[C(OP_READ)] = {
1258		[C(RESULT_MISS)]	= { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1259	},
1260	[C(OP_WRITE)] = {
1261		[C(RESULT_MISS)]	= { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1262	},
1263},
1264[C(ITLB)] = {
1265	[C(OP_READ)] = {
1266		[C(RESULT_MISS)]	= { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1267	},
1268	[C(OP_WRITE)] = {
1269		[C(RESULT_MISS)]	= { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1270	},
1271},
1272[C(BPU)] = {
1273	[C(OP_READ)] = {
1274		[C(RESULT_MISS)]	= { 0x25, CNTR_ALL },
1275	},
1276},
1277};
1278
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1279static int __hw_perf_event_init(struct perf_event *event)
1280{
1281	struct perf_event_attr *attr = &event->attr;
1282	struct hw_perf_event *hwc = &event->hw;
1283	const struct mips_perf_event *pev;
1284	int err;
1285
1286	/* Returning MIPS event descriptor for generic perf event. */
1287	if (PERF_TYPE_HARDWARE == event->attr.type) {
1288		if (event->attr.config >= PERF_COUNT_HW_MAX)
1289			return -EINVAL;
1290		pev = mipspmu_map_general_event(event->attr.config);
1291	} else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1292		pev = mipspmu_map_cache_event(event->attr.config);
1293	} else if (PERF_TYPE_RAW == event->attr.type) {
1294		/* We are working on the global raw event. */
1295		mutex_lock(&raw_event_mutex);
1296		pev = mipspmu.map_raw_event(event->attr.config);
1297	} else {
1298		/* The event type is not (yet) supported. */
1299		return -EOPNOTSUPP;
1300	}
1301
1302	if (IS_ERR(pev)) {
1303		if (PERF_TYPE_RAW == event->attr.type)
1304			mutex_unlock(&raw_event_mutex);
1305		return PTR_ERR(pev);
1306	}
1307
1308	/*
1309	 * We allow max flexibility on how each individual counter shared
1310	 * by the single CPU operates (the mode exclusion and the range).
1311	 */
1312	hwc->config_base = MIPS_PERFCTRL_IE;
 
 
 
 
1313
1314	hwc->event_base = mipspmu_perf_event_encode(pev);
1315	if (PERF_TYPE_RAW == event->attr.type)
1316		mutex_unlock(&raw_event_mutex);
1317
1318	if (!attr->exclude_user)
1319		hwc->config_base |= MIPS_PERFCTRL_U;
1320	if (!attr->exclude_kernel) {
1321		hwc->config_base |= MIPS_PERFCTRL_K;
1322		/* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1323		hwc->config_base |= MIPS_PERFCTRL_EXL;
1324	}
1325	if (!attr->exclude_hv)
1326		hwc->config_base |= MIPS_PERFCTRL_S;
1327
1328	hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1329	/*
1330	 * The event can belong to another cpu. We do not assign a local
1331	 * counter for it for now.
1332	 */
1333	hwc->idx = -1;
1334	hwc->config = 0;
1335
1336	if (!hwc->sample_period) {
1337		hwc->sample_period  = mipspmu.max_period;
1338		hwc->last_period    = hwc->sample_period;
1339		local64_set(&hwc->period_left, hwc->sample_period);
1340	}
1341
1342	err = 0;
1343	if (event->group_leader != event)
1344		err = validate_group(event);
1345
1346	event->destroy = hw_perf_event_destroy;
1347
1348	if (err)
1349		event->destroy(event);
1350
1351	return err;
1352}
1353
1354static void pause_local_counters(void)
1355{
1356	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1357	int ctr = mipspmu.num_counters;
1358	unsigned long flags;
1359
1360	local_irq_save(flags);
1361	do {
1362		ctr--;
1363		cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1364		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1365					 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1366	} while (ctr > 0);
1367	local_irq_restore(flags);
1368}
1369
1370static void resume_local_counters(void)
1371{
1372	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1373	int ctr = mipspmu.num_counters;
1374
1375	do {
1376		ctr--;
1377		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1378	} while (ctr > 0);
1379}
1380
1381static int mipsxx_pmu_handle_shared_irq(void)
1382{
1383	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1384	struct perf_sample_data data;
1385	unsigned int counters = mipspmu.num_counters;
1386	u64 counter;
1387	int n, handled = IRQ_NONE;
1388	struct pt_regs *regs;
1389
1390	if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1391		return handled;
1392	/*
1393	 * First we pause the local counters, so that when we are locked
1394	 * here, the counters are all paused. When it gets locked due to
1395	 * perf_disable(), the timer interrupt handler will be delayed.
1396	 *
1397	 * See also mipsxx_pmu_start().
1398	 */
1399	pause_local_counters();
1400#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1401	read_lock(&pmuint_rwlock);
1402#endif
1403
1404	regs = get_irq_regs();
1405
1406	perf_sample_data_init(&data, 0, 0);
1407
1408	for (n = counters - 1; n >= 0; n--) {
1409		if (!test_bit(n, cpuc->used_mask))
1410			continue;
1411
1412		counter = mipspmu.read_counter(n);
1413		if (!(counter & mipspmu.overflow))
1414			continue;
1415
1416		handle_associated_event(cpuc, n, &data, regs);
1417		handled = IRQ_HANDLED;
 
 
 
 
1418	}
1419
1420#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1421	read_unlock(&pmuint_rwlock);
1422#endif
1423	resume_local_counters();
1424
1425	/*
1426	 * Do all the work for the pending perf events. We can do this
1427	 * in here because the performance counter interrupt is a regular
1428	 * interrupt, not NMI.
1429	 */
1430	if (handled == IRQ_HANDLED)
1431		irq_work_run();
1432
 
 
 
 
1433	return handled;
1434}
1435
1436static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1437{
1438	return mipsxx_pmu_handle_shared_irq();
1439}
1440
1441/* 24K */
1442#define IS_BOTH_COUNTERS_24K_EVENT(b)					\
1443	((b) == 0 || (b) == 1 || (b) == 11)
1444
1445/* 34K */
1446#define IS_BOTH_COUNTERS_34K_EVENT(b)					\
1447	((b) == 0 || (b) == 1 || (b) == 11)
1448#ifdef CONFIG_MIPS_MT_SMP
1449#define IS_RANGE_P_34K_EVENT(r, b)					\
1450	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1451	 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||		\
1452	 (r) == 176 || ((b) >= 50 && (b) <= 55) ||			\
1453	 ((b) >= 64 && (b) <= 67))
1454#define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1455#endif
1456
1457/* 74K */
1458#define IS_BOTH_COUNTERS_74K_EVENT(b)					\
1459	((b) == 0 || (b) == 1)
1460
1461/* proAptiv */
1462#define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b)				\
1463	((b) == 0 || (b) == 1)
1464/* P5600 */
1465#define IS_BOTH_COUNTERS_P5600_EVENT(b)					\
1466	((b) == 0 || (b) == 1)
1467
1468/* 1004K */
1469#define IS_BOTH_COUNTERS_1004K_EVENT(b)					\
1470	((b) == 0 || (b) == 1 || (b) == 11)
1471#ifdef CONFIG_MIPS_MT_SMP
1472#define IS_RANGE_P_1004K_EVENT(r, b)					\
1473	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1474	 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||		\
1475	 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||	\
1476	 (r) == 188 || (b) == 61 || (b) == 62 ||			\
1477	 ((b) >= 64 && (b) <= 67))
1478#define IS_RANGE_V_1004K_EVENT(r)	((r) == 47)
1479#endif
1480
1481/* interAptiv */
1482#define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b)				\
1483	((b) == 0 || (b) == 1 || (b) == 11)
1484#ifdef CONFIG_MIPS_MT_SMP
1485/* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1486#define IS_RANGE_P_INTERAPTIV_EVENT(r, b)				\
1487	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1488	 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 ||		\
1489	 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 &&		\
1490	 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 ||		\
1491	 ((b) >= 64 && (b) <= 67))
1492#define IS_RANGE_V_INTERAPTIV_EVENT(r)	((r) == 47 || (r) == 175)
1493#endif
1494
1495/* BMIPS5000 */
1496#define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b)				\
1497	((b) == 0 || (b) == 1)
1498
1499
1500/*
1501 * For most cores the user can use 0-255 raw events, where 0-127 for the events
1502 * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1503 * indicate the even/odd bank selector. So, for example, when user wants to take
1504 * the Event Num of 15 for odd counters (by referring to the user manual), then
1505 * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1506 * to be used.
1507 *
1508 * Some newer cores have even more events, in which case the user can use raw
1509 * events 0-511, where 0-255 are for the events of even counters, and 256-511
1510 * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1511 */
1512static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1513{
1514	/* currently most cores have 7-bit event numbers */
1515	unsigned int raw_id = config & 0xff;
1516	unsigned int base_id = raw_id & 0x7f;
1517
 
 
1518	switch (current_cpu_type()) {
1519	case CPU_24K:
1520		if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1521			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1522		else
1523			raw_event.cntr_mask =
1524				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1525#ifdef CONFIG_MIPS_MT_SMP
1526		/*
1527		 * This is actually doing nothing. Non-multithreading
1528		 * CPUs will not check and calculate the range.
1529		 */
1530		raw_event.range = P;
1531#endif
1532		break;
1533	case CPU_34K:
1534		if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1535			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1536		else
1537			raw_event.cntr_mask =
1538				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1539#ifdef CONFIG_MIPS_MT_SMP
1540		if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1541			raw_event.range = P;
1542		else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1543			raw_event.range = V;
1544		else
1545			raw_event.range = T;
1546#endif
1547		break;
1548	case CPU_74K:
1549	case CPU_1074K:
1550		if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1551			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1552		else
1553			raw_event.cntr_mask =
1554				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1555#ifdef CONFIG_MIPS_MT_SMP
1556		raw_event.range = P;
1557#endif
1558		break;
1559	case CPU_PROAPTIV:
1560		if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1561			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1562		else
1563			raw_event.cntr_mask =
1564				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1565#ifdef CONFIG_MIPS_MT_SMP
1566		raw_event.range = P;
1567#endif
1568		break;
1569	case CPU_P5600:
1570	case CPU_P6600:
1571		/* 8-bit event numbers */
1572		raw_id = config & 0x1ff;
1573		base_id = raw_id & 0xff;
1574		if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1575			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1576		else
1577			raw_event.cntr_mask =
1578				raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1579#ifdef CONFIG_MIPS_MT_SMP
1580		raw_event.range = P;
1581#endif
1582		break;
1583	case CPU_I6400:
1584	case CPU_I6500:
1585		/* 8-bit event numbers */
1586		base_id = config & 0xff;
1587		raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1588		break;
1589	case CPU_1004K:
1590		if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1591			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1592		else
1593			raw_event.cntr_mask =
1594				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1595#ifdef CONFIG_MIPS_MT_SMP
1596		if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1597			raw_event.range = P;
1598		else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1599			raw_event.range = V;
1600		else
1601			raw_event.range = T;
1602#endif
1603		break;
1604	case CPU_INTERAPTIV:
1605		if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1606			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1607		else
1608			raw_event.cntr_mask =
1609				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1610#ifdef CONFIG_MIPS_MT_SMP
1611		if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1612			raw_event.range = P;
1613		else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1614			raw_event.range = V;
1615		else
1616			raw_event.range = T;
1617#endif
1618		break;
1619	case CPU_BMIPS5000:
1620		if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1621			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1622		else
1623			raw_event.cntr_mask =
1624				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1625		break;
1626	case CPU_LOONGSON3:
1627		raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1628	break;
1629	}
1630
1631	raw_event.event_id = base_id;
1632
1633	return &raw_event;
1634}
1635
1636static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1637{
1638	unsigned int raw_id = config & 0xff;
1639	unsigned int base_id = raw_id & 0x7f;
1640
1641
1642	raw_event.cntr_mask = CNTR_ALL;
1643	raw_event.event_id = base_id;
1644
1645	if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1646		if (base_id > 0x42)
1647			return ERR_PTR(-EOPNOTSUPP);
1648	} else {
1649		if (base_id > 0x3a)
1650			return ERR_PTR(-EOPNOTSUPP);
1651	}
1652
1653	switch (base_id) {
1654	case 0x00:
1655	case 0x0f:
1656	case 0x1e:
1657	case 0x1f:
1658	case 0x2f:
1659	case 0x34:
1660	case 0x3b ... 0x3f:
1661		return ERR_PTR(-EOPNOTSUPP);
1662	default:
1663		break;
1664	}
1665
1666	return &raw_event;
1667}
1668
1669static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config)
1670{
1671	unsigned int raw_id = config & 0xff;
1672
1673	/* Only 1-63 are defined */
1674	if ((raw_id < 0x01) || (raw_id > 0x3f))
1675		return ERR_PTR(-EOPNOTSUPP);
1676
1677	raw_event.cntr_mask = CNTR_ALL;
1678	raw_event.event_id = raw_id;
1679
1680	return &raw_event;
1681}
1682
1683static int __init
1684init_hw_perf_events(void)
1685{
1686	int counters, irq;
1687	int counter_bits;
1688
1689	pr_info("Performance counters: ");
1690
1691	counters = n_counters();
1692	if (counters == 0) {
1693		pr_cont("No available PMU.\n");
1694		return -ENODEV;
1695	}
1696
1697#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
 
1698	if (!cpu_has_mipsmt_pertccounters)
1699		counters = counters_total_to_per_cpu(counters);
1700#endif
1701
1702	if (get_c0_perfcount_int)
1703		irq = get_c0_perfcount_int();
1704	else if (cp0_perfcount_irq >= 0)
1705		irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1706	else
1707		irq = -1;
 
 
 
 
 
 
 
 
 
 
1708
1709	mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1710
1711	switch (current_cpu_type()) {
1712	case CPU_24K:
1713		mipspmu.name = "mips/24K";
1714		mipspmu.general_event_map = &mipsxxcore_event_map;
1715		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1716		break;
1717	case CPU_34K:
1718		mipspmu.name = "mips/34K";
1719		mipspmu.general_event_map = &mipsxxcore_event_map;
1720		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1721		break;
1722	case CPU_74K:
1723		mipspmu.name = "mips/74K";
1724		mipspmu.general_event_map = &mipsxxcore_event_map2;
1725		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1726		break;
1727	case CPU_PROAPTIV:
1728		mipspmu.name = "mips/proAptiv";
1729		mipspmu.general_event_map = &mipsxxcore_event_map2;
1730		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1731		break;
1732	case CPU_P5600:
1733		mipspmu.name = "mips/P5600";
1734		mipspmu.general_event_map = &mipsxxcore_event_map2;
1735		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1736		break;
1737	case CPU_P6600:
1738		mipspmu.name = "mips/P6600";
1739		mipspmu.general_event_map = &mipsxxcore_event_map2;
1740		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1741		break;
1742	case CPU_I6400:
1743		mipspmu.name = "mips/I6400";
1744		mipspmu.general_event_map = &i6x00_event_map;
1745		mipspmu.cache_event_map = &i6x00_cache_map;
1746		break;
1747	case CPU_I6500:
1748		mipspmu.name = "mips/I6500";
1749		mipspmu.general_event_map = &i6x00_event_map;
1750		mipspmu.cache_event_map = &i6x00_cache_map;
1751		break;
1752	case CPU_1004K:
1753		mipspmu.name = "mips/1004K";
1754		mipspmu.general_event_map = &mipsxxcore_event_map;
1755		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1756		break;
1757	case CPU_1074K:
1758		mipspmu.name = "mips/1074K";
1759		mipspmu.general_event_map = &mipsxxcore_event_map;
1760		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1761		break;
1762	case CPU_INTERAPTIV:
1763		mipspmu.name = "mips/interAptiv";
1764		mipspmu.general_event_map = &mipsxxcore_event_map;
1765		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1766		break;
1767	case CPU_LOONGSON1:
1768		mipspmu.name = "mips/loongson1";
1769		mipspmu.general_event_map = &mipsxxcore_event_map;
1770		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1771		break;
1772	case CPU_LOONGSON3:
1773		mipspmu.name = "mips/loongson3";
1774		mipspmu.general_event_map = &loongson3_event_map;
1775		mipspmu.cache_event_map = &loongson3_cache_map;
1776		break;
1777	case CPU_CAVIUM_OCTEON:
1778	case CPU_CAVIUM_OCTEON_PLUS:
1779	case CPU_CAVIUM_OCTEON2:
1780		mipspmu.name = "octeon";
1781		mipspmu.general_event_map = &octeon_event_map;
1782		mipspmu.cache_event_map = &octeon_cache_map;
1783		mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1784		break;
1785	case CPU_BMIPS5000:
1786		mipspmu.name = "BMIPS5000";
1787		mipspmu.general_event_map = &bmips5000_event_map;
1788		mipspmu.cache_event_map = &bmips5000_cache_map;
1789		break;
1790	case CPU_XLP:
1791		mipspmu.name = "xlp";
1792		mipspmu.general_event_map = &xlp_event_map;
1793		mipspmu.cache_event_map = &xlp_cache_map;
1794		mipspmu.map_raw_event = xlp_pmu_map_raw_event;
1795		break;
1796	default:
1797		pr_cont("Either hardware does not support performance "
1798			"counters, or not yet implemented.\n");
1799		return -ENODEV;
1800	}
1801
1802	mipspmu.num_counters = counters;
1803	mipspmu.irq = irq;
1804
1805	if (read_c0_perfctrl0() & MIPS_PERFCTRL_W) {
1806		mipspmu.max_period = (1ULL << 63) - 1;
1807		mipspmu.valid_count = (1ULL << 63) - 1;
1808		mipspmu.overflow = 1ULL << 63;
1809		mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1810		mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1811		counter_bits = 64;
1812	} else {
1813		mipspmu.max_period = (1ULL << 31) - 1;
1814		mipspmu.valid_count = (1ULL << 31) - 1;
1815		mipspmu.overflow = 1ULL << 31;
1816		mipspmu.read_counter = mipsxx_pmu_read_counter;
1817		mipspmu.write_counter = mipsxx_pmu_write_counter;
1818		counter_bits = 32;
1819	}
1820
1821	on_each_cpu(reset_counters, (void *)(long)counters, 1);
1822
1823	pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1824		"CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1825		irq < 0 ? " (share with timer interrupt)" : "");
1826
1827	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1828
1829	return 0;
1830}
1831early_initcall(init_hw_perf_events);