Linux Audio

Check our new training course

Embedded Linux training

Mar 31-Apr 8, 2025
Register
Loading...
v5.4
   1/*
   2 * Performance events x86 architecture code
   3 *
   4 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   5 *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
   6 *  Copyright (C) 2009 Jaswinder Singh Rajput
   7 *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
   8 *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
   9 *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
  10 *  Copyright (C) 2009 Google, Inc., Stephane Eranian
  11 *
  12 *  For licencing details see kernel-base/COPYING
  13 */
  14
  15#include <linux/perf_event.h>
  16#include <linux/capability.h>
  17#include <linux/notifier.h>
  18#include <linux/hardirq.h>
  19#include <linux/kprobes.h>
  20#include <linux/export.h>
  21#include <linux/init.h>
  22#include <linux/kdebug.h>
  23#include <linux/sched/mm.h>
  24#include <linux/sched/clock.h>
  25#include <linux/uaccess.h>
  26#include <linux/slab.h>
  27#include <linux/cpu.h>
  28#include <linux/bitops.h>
  29#include <linux/device.h>
  30#include <linux/nospec.h>
 
  31
  32#include <asm/apic.h>
  33#include <asm/stacktrace.h>
  34#include <asm/nmi.h>
  35#include <asm/smp.h>
  36#include <asm/alternative.h>
  37#include <asm/mmu_context.h>
  38#include <asm/tlbflush.h>
  39#include <asm/timer.h>
  40#include <asm/desc.h>
  41#include <asm/ldt.h>
  42#include <asm/unwind.h>
 
 
  43
  44#include "perf_event.h"
  45
  46struct x86_pmu x86_pmu __read_mostly;
 
  47
  48DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
  49	.enabled = 1,
 
  50};
  51
 
  52DEFINE_STATIC_KEY_FALSE(rdpmc_always_available_key);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  53
  54u64 __read_mostly hw_cache_event_ids
  55				[PERF_COUNT_HW_CACHE_MAX]
  56				[PERF_COUNT_HW_CACHE_OP_MAX]
  57				[PERF_COUNT_HW_CACHE_RESULT_MAX];
  58u64 __read_mostly hw_cache_extra_regs
  59				[PERF_COUNT_HW_CACHE_MAX]
  60				[PERF_COUNT_HW_CACHE_OP_MAX]
  61				[PERF_COUNT_HW_CACHE_RESULT_MAX];
  62
  63/*
  64 * Propagate event elapsed time into the generic event.
  65 * Can only be executed on the CPU where the event is active.
  66 * Returns the delta events processed.
  67 */
  68u64 x86_perf_event_update(struct perf_event *event)
  69{
  70	struct hw_perf_event *hwc = &event->hw;
  71	int shift = 64 - x86_pmu.cntval_bits;
  72	u64 prev_raw_count, new_raw_count;
  73	int idx = hwc->idx;
  74	u64 delta;
  75
  76	if (idx == INTEL_PMC_IDX_FIXED_BTS)
  77		return 0;
  78
  79	/*
  80	 * Careful: an NMI might modify the previous event value.
  81	 *
  82	 * Our tactic to handle this is to first atomically read and
  83	 * exchange a new raw count - then add that new-prev delta
  84	 * count to the generic event atomically:
  85	 */
  86again:
  87	prev_raw_count = local64_read(&hwc->prev_count);
  88	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
  89
  90	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
  91					new_raw_count) != prev_raw_count)
  92		goto again;
  93
  94	/*
  95	 * Now we have the new raw value and have updated the prev
  96	 * timestamp already. We can now calculate the elapsed delta
  97	 * (event-)time and add that to the generic event.
  98	 *
  99	 * Careful, not all hw sign-extends above the physical width
 100	 * of the count.
 101	 */
 102	delta = (new_raw_count << shift) - (prev_raw_count << shift);
 103	delta >>= shift;
 104
 105	local64_add(delta, &event->count);
 106	local64_sub(delta, &hwc->period_left);
 107
 108	return new_raw_count;
 109}
 110
 111/*
 112 * Find and validate any extra registers to set up.
 113 */
 114static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
 115{
 
 116	struct hw_perf_event_extra *reg;
 117	struct extra_reg *er;
 118
 119	reg = &event->hw.extra_reg;
 120
 121	if (!x86_pmu.extra_regs)
 122		return 0;
 123
 124	for (er = x86_pmu.extra_regs; er->msr; er++) {
 125		if (er->event != (config & er->config_mask))
 126			continue;
 127		if (event->attr.config1 & ~er->valid_mask)
 128			return -EINVAL;
 129		/* Check if the extra msrs can be safely accessed*/
 130		if (!er->extra_msr_access)
 131			return -ENXIO;
 132
 133		reg->idx = er->idx;
 134		reg->config = event->attr.config1;
 135		reg->reg = er->msr;
 136		break;
 137	}
 138	return 0;
 139}
 140
 141static atomic_t active_events;
 142static atomic_t pmc_refcount;
 143static DEFINE_MUTEX(pmc_reserve_mutex);
 144
 145#ifdef CONFIG_X86_LOCAL_APIC
 146
 147static bool reserve_pmc_hardware(void)
 148{
 
 149	int i;
 150
 151	for (i = 0; i < x86_pmu.num_counters; i++) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 152		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
 153			goto perfctr_fail;
 154	}
 155
 156	for (i = 0; i < x86_pmu.num_counters; i++) {
 157		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
 158			goto eventsel_fail;
 159	}
 160
 161	return true;
 162
 163eventsel_fail:
 164	for (i--; i >= 0; i--)
 
 165		release_evntsel_nmi(x86_pmu_config_addr(i));
 166
 167	i = x86_pmu.num_counters;
 168
 169perfctr_fail:
 170	for (i--; i >= 0; i--)
 
 171		release_perfctr_nmi(x86_pmu_event_addr(i));
 172
 173	return false;
 174}
 175
 176static void release_pmc_hardware(void)
 177{
 
 178	int i;
 179
 180	for (i = 0; i < x86_pmu.num_counters; i++) {
 181		release_perfctr_nmi(x86_pmu_event_addr(i));
 182		release_evntsel_nmi(x86_pmu_config_addr(i));
 183	}
 184}
 185
 186#else
 187
 188static bool reserve_pmc_hardware(void) { return true; }
 189static void release_pmc_hardware(void) {}
 190
 191#endif
 192
 193static bool check_hw_exists(void)
 
 194{
 195	u64 val, val_fail = -1, val_new= ~0;
 196	int i, reg, reg_fail = -1, ret = 0;
 197	int bios_fail = 0;
 198	int reg_safe = -1;
 199
 200	/*
 201	 * Check to see if the BIOS enabled any of the counters, if so
 202	 * complain and bail.
 203	 */
 204	for (i = 0; i < x86_pmu.num_counters; i++) {
 205		reg = x86_pmu_config_addr(i);
 206		ret = rdmsrl_safe(reg, &val);
 207		if (ret)
 208			goto msr_fail;
 209		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
 210			bios_fail = 1;
 211			val_fail = val;
 212			reg_fail = reg;
 213		} else {
 214			reg_safe = i;
 215		}
 216	}
 217
 218	if (x86_pmu.num_counters_fixed) {
 219		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 220		ret = rdmsrl_safe(reg, &val);
 221		if (ret)
 222			goto msr_fail;
 223		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
 224			if (val & (0x03 << i*4)) {
 
 
 225				bios_fail = 1;
 226				val_fail = val;
 227				reg_fail = reg;
 228			}
 229		}
 230	}
 231
 232	/*
 233	 * If all the counters are enabled, the below test will always
 234	 * fail.  The tools will also become useless in this scenario.
 235	 * Just fail and disable the hardware counters.
 236	 */
 237
 238	if (reg_safe == -1) {
 239		reg = reg_safe;
 240		goto msr_fail;
 241	}
 242
 243	/*
 244	 * Read the current value, change it and read it back to see if it
 245	 * matches, this is needed to detect certain hardware emulators
 246	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
 247	 */
 248	reg = x86_pmu_event_addr(reg_safe);
 249	if (rdmsrl_safe(reg, &val))
 250		goto msr_fail;
 251	val ^= 0xffffUL;
 252	ret = wrmsrl_safe(reg, val);
 253	ret |= rdmsrl_safe(reg, &val_new);
 254	if (ret || val != val_new)
 255		goto msr_fail;
 256
 257	/*
 258	 * We still allow the PMU driver to operate:
 259	 */
 260	if (bios_fail) {
 261		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
 262		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
 263			      reg_fail, val_fail);
 264	}
 265
 266	return true;
 267
 268msr_fail:
 269	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
 270		pr_cont("PMU not available due to virtualization, using software events only.\n");
 271	} else {
 272		pr_cont("Broken PMU hardware detected, using software events only.\n");
 273		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
 274		       reg, val_new);
 275	}
 276
 277	return false;
 278}
 279
 280static void hw_perf_event_destroy(struct perf_event *event)
 281{
 282	x86_release_hardware();
 283	atomic_dec(&active_events);
 284}
 285
 286void hw_perf_lbr_event_destroy(struct perf_event *event)
 287{
 288	hw_perf_event_destroy(event);
 289
 290	/* undo the lbr/bts event accounting */
 291	x86_del_exclusive(x86_lbr_exclusive_lbr);
 292}
 293
 294static inline int x86_pmu_initialized(void)
 295{
 296	return x86_pmu.handle_irq != NULL;
 297}
 298
 299static inline int
 300set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
 301{
 302	struct perf_event_attr *attr = &event->attr;
 303	unsigned int cache_type, cache_op, cache_result;
 304	u64 config, val;
 305
 306	config = attr->config;
 307
 308	cache_type = (config >> 0) & 0xff;
 309	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 310		return -EINVAL;
 311	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
 312
 313	cache_op = (config >>  8) & 0xff;
 314	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 315		return -EINVAL;
 316	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
 317
 318	cache_result = (config >> 16) & 0xff;
 319	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 320		return -EINVAL;
 321	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
 322
 323	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
 324
 325	if (val == 0)
 326		return -ENOENT;
 327
 328	if (val == -1)
 329		return -EINVAL;
 330
 331	hwc->config |= val;
 332	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
 333	return x86_pmu_extra_regs(val, event);
 334}
 335
 336int x86_reserve_hardware(void)
 337{
 338	int err = 0;
 339
 340	if (!atomic_inc_not_zero(&pmc_refcount)) {
 341		mutex_lock(&pmc_reserve_mutex);
 342		if (atomic_read(&pmc_refcount) == 0) {
 343			if (!reserve_pmc_hardware())
 344				err = -EBUSY;
 345			else
 346				reserve_ds_buffers();
 
 
 347		}
 348		if (!err)
 349			atomic_inc(&pmc_refcount);
 350		mutex_unlock(&pmc_reserve_mutex);
 351	}
 352
 353	return err;
 354}
 355
 356void x86_release_hardware(void)
 357{
 358	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
 359		release_pmc_hardware();
 360		release_ds_buffers();
 
 361		mutex_unlock(&pmc_reserve_mutex);
 362	}
 363}
 364
 365/*
 366 * Check if we can create event of a certain type (that no conflicting events
 367 * are present).
 368 */
 369int x86_add_exclusive(unsigned int what)
 370{
 371	int i;
 372
 373	/*
 374	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
 375	 * LBR and BTS are still mutually exclusive.
 376	 */
 377	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
 378		return 0;
 379
 380	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
 381		mutex_lock(&pmc_reserve_mutex);
 382		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
 383			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
 384				goto fail_unlock;
 385		}
 386		atomic_inc(&x86_pmu.lbr_exclusive[what]);
 387		mutex_unlock(&pmc_reserve_mutex);
 388	}
 389
 
 390	atomic_inc(&active_events);
 391	return 0;
 392
 393fail_unlock:
 394	mutex_unlock(&pmc_reserve_mutex);
 395	return -EBUSY;
 396}
 397
 398void x86_del_exclusive(unsigned int what)
 399{
 
 
 
 
 
 400	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
 401		return;
 402
 403	atomic_dec(&x86_pmu.lbr_exclusive[what]);
 404	atomic_dec(&active_events);
 405}
 406
 407int x86_setup_perfctr(struct perf_event *event)
 408{
 409	struct perf_event_attr *attr = &event->attr;
 410	struct hw_perf_event *hwc = &event->hw;
 411	u64 config;
 412
 413	if (!is_sampling_event(event)) {
 414		hwc->sample_period = x86_pmu.max_period;
 415		hwc->last_period = hwc->sample_period;
 416		local64_set(&hwc->period_left, hwc->sample_period);
 417	}
 418
 419	if (attr->type == PERF_TYPE_RAW)
 420		return x86_pmu_extra_regs(event->attr.config, event);
 421
 422	if (attr->type == PERF_TYPE_HW_CACHE)
 423		return set_ext_hw_attr(hwc, event);
 424
 425	if (attr->config >= x86_pmu.max_events)
 426		return -EINVAL;
 427
 428	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
 429
 430	/*
 431	 * The generic map:
 432	 */
 433	config = x86_pmu.event_map(attr->config);
 434
 435	if (config == 0)
 436		return -ENOENT;
 437
 438	if (config == -1LL)
 439		return -EINVAL;
 440
 441	hwc->config |= config;
 442
 443	return 0;
 444}
 445
 446/*
 447 * check that branch_sample_type is compatible with
 448 * settings needed for precise_ip > 1 which implies
 449 * using the LBR to capture ALL taken branches at the
 450 * priv levels of the measurement
 451 */
 452static inline int precise_br_compat(struct perf_event *event)
 453{
 454	u64 m = event->attr.branch_sample_type;
 455	u64 b = 0;
 456
 457	/* must capture all branches */
 458	if (!(m & PERF_SAMPLE_BRANCH_ANY))
 459		return 0;
 460
 461	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
 462
 463	if (!event->attr.exclude_user)
 464		b |= PERF_SAMPLE_BRANCH_USER;
 465
 466	if (!event->attr.exclude_kernel)
 467		b |= PERF_SAMPLE_BRANCH_KERNEL;
 468
 469	/*
 470	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
 471	 */
 472
 473	return m == b;
 474}
 475
 476int x86_pmu_max_precise(void)
 477{
 478	int precise = 0;
 479
 480	/* Support for constant skid */
 481	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
 482		precise++;
 483
 484		/* Support for IP fixup */
 485		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
 486			precise++;
 487
 488		if (x86_pmu.pebs_prec_dist)
 489			precise++;
 490	}
 491	return precise;
 492}
 493
 494int x86_pmu_hw_config(struct perf_event *event)
 495{
 496	if (event->attr.precise_ip) {
 497		int precise = x86_pmu_max_precise();
 498
 499		if (event->attr.precise_ip > precise)
 500			return -EOPNOTSUPP;
 501
 502		/* There's no sense in having PEBS for non sampling events: */
 503		if (!is_sampling_event(event))
 504			return -EINVAL;
 505	}
 506	/*
 507	 * check that PEBS LBR correction does not conflict with
 508	 * whatever the user is asking with attr->branch_sample_type
 509	 */
 510	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
 511		u64 *br_type = &event->attr.branch_sample_type;
 512
 513		if (has_branch_stack(event)) {
 514			if (!precise_br_compat(event))
 515				return -EOPNOTSUPP;
 516
 517			/* branch_sample_type is compatible */
 518
 519		} else {
 520			/*
 521			 * user did not specify  branch_sample_type
 522			 *
 523			 * For PEBS fixups, we capture all
 524			 * the branches at the priv level of the
 525			 * event.
 526			 */
 527			*br_type = PERF_SAMPLE_BRANCH_ANY;
 528
 529			if (!event->attr.exclude_user)
 530				*br_type |= PERF_SAMPLE_BRANCH_USER;
 531
 532			if (!event->attr.exclude_kernel)
 533				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
 534		}
 535	}
 536
 537	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
 538		event->attach_state |= PERF_ATTACH_TASK_DATA;
 539
 540	/*
 541	 * Generate PMC IRQs:
 542	 * (keep 'enabled' bit clear for now)
 543	 */
 544	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
 545
 546	/*
 547	 * Count user and OS events unless requested not to
 548	 */
 549	if (!event->attr.exclude_user)
 550		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
 551	if (!event->attr.exclude_kernel)
 552		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
 553
 554	if (event->attr.type == PERF_TYPE_RAW)
 555		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
 556
 557	if (event->attr.sample_period && x86_pmu.limit_period) {
 558		if (x86_pmu.limit_period(event, event->attr.sample_period) >
 559				event->attr.sample_period)
 
 560			return -EINVAL;
 561	}
 562
 563	/* sample_regs_user never support XMM registers */
 564	if (unlikely(event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK))
 565		return -EINVAL;
 566	/*
 567	 * Besides the general purpose registers, XMM registers may
 568	 * be collected in PEBS on some platforms, e.g. Icelake
 569	 */
 570	if (unlikely(event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK)) {
 571		if (!(event->pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS))
 572			return -EINVAL;
 573
 574		if (!event->attr.precise_ip)
 575			return -EINVAL;
 576	}
 577
 578	return x86_setup_perfctr(event);
 579}
 580
 581/*
 582 * Setup the hardware configuration for a given attr_type
 583 */
 584static int __x86_pmu_event_init(struct perf_event *event)
 585{
 586	int err;
 587
 588	if (!x86_pmu_initialized())
 589		return -ENODEV;
 590
 591	err = x86_reserve_hardware();
 592	if (err)
 593		return err;
 594
 595	atomic_inc(&active_events);
 596	event->destroy = hw_perf_event_destroy;
 597
 598	event->hw.idx = -1;
 599	event->hw.last_cpu = -1;
 600	event->hw.last_tag = ~0ULL;
 601
 602	/* mark unused */
 603	event->hw.extra_reg.idx = EXTRA_REG_NONE;
 604	event->hw.branch_reg.idx = EXTRA_REG_NONE;
 605
 606	return x86_pmu.hw_config(event);
 607}
 608
 609void x86_pmu_disable_all(void)
 610{
 611	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 612	int idx;
 613
 614	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 
 615		u64 val;
 616
 617		if (!test_bit(idx, cpuc->active_mask))
 618			continue;
 619		rdmsrl(x86_pmu_config_addr(idx), val);
 620		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
 621			continue;
 622		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 623		wrmsrl(x86_pmu_config_addr(idx), val);
 
 
 624	}
 625}
 626
 
 
 
 
 
 
 627/*
 628 * There may be PMI landing after enabled=0. The PMI hitting could be before or
 629 * after disable_all.
 630 *
 631 * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
 632 * It will not be re-enabled in the NMI handler again, because enabled=0. After
 633 * handling the NMI, disable_all will be called, which will not change the
 634 * state either. If PMI hits after disable_all, the PMU is already disabled
 635 * before entering NMI handler. The NMI handler will not change the state
 636 * either.
 637 *
 638 * So either situation is harmless.
 639 */
 640static void x86_pmu_disable(struct pmu *pmu)
 641{
 642	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 643
 644	if (!x86_pmu_initialized())
 645		return;
 646
 647	if (!cpuc->enabled)
 648		return;
 649
 650	cpuc->n_added = 0;
 651	cpuc->enabled = 0;
 652	barrier();
 653
 654	x86_pmu.disable_all();
 655}
 656
 657void x86_pmu_enable_all(int added)
 658{
 659	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 660	int idx;
 661
 662	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 663		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
 664
 665		if (!test_bit(idx, cpuc->active_mask))
 666			continue;
 667
 668		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
 669	}
 670}
 671
 672static struct pmu pmu;
 673
 674static inline int is_x86_event(struct perf_event *event)
 675{
 676	return event->pmu == &pmu;
 
 
 
 
 
 
 
 
 
 
 677}
 678
 679struct pmu *x86_get_pmu(void)
 680{
 681	return &pmu;
 
 
 
 
 
 
 
 
 
 682}
 683/*
 684 * Event scheduler state:
 685 *
 686 * Assign events iterating over all events and counters, beginning
 687 * with events with least weights first. Keep the current iterator
 688 * state in struct sched_state.
 689 */
 690struct sched_state {
 691	int	weight;
 692	int	event;		/* event index */
 693	int	counter;	/* counter index */
 694	int	unassigned;	/* number of events to be assigned left */
 695	int	nr_gp;		/* number of GP counters used */
 696	unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 697};
 698
 699/* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
 700#define	SCHED_STATES_MAX	2
 701
 702struct perf_sched {
 703	int			max_weight;
 704	int			max_events;
 705	int			max_gp;
 706	int			saved_states;
 707	struct event_constraint	**constraints;
 708	struct sched_state	state;
 709	struct sched_state	saved[SCHED_STATES_MAX];
 710};
 711
 712/*
 713 * Initialize interator that runs through all events and counters.
 714 */
 715static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
 716			    int num, int wmin, int wmax, int gpmax)
 717{
 718	int idx;
 719
 720	memset(sched, 0, sizeof(*sched));
 721	sched->max_events	= num;
 722	sched->max_weight	= wmax;
 723	sched->max_gp		= gpmax;
 724	sched->constraints	= constraints;
 725
 726	for (idx = 0; idx < num; idx++) {
 727		if (constraints[idx]->weight == wmin)
 728			break;
 729	}
 730
 731	sched->state.event	= idx;		/* start with min weight */
 732	sched->state.weight	= wmin;
 733	sched->state.unassigned	= num;
 734}
 735
 736static void perf_sched_save_state(struct perf_sched *sched)
 737{
 738	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
 739		return;
 740
 741	sched->saved[sched->saved_states] = sched->state;
 742	sched->saved_states++;
 743}
 744
 745static bool perf_sched_restore_state(struct perf_sched *sched)
 746{
 747	if (!sched->saved_states)
 748		return false;
 749
 750	sched->saved_states--;
 751	sched->state = sched->saved[sched->saved_states];
 752
 753	/* continue with next counter: */
 754	clear_bit(sched->state.counter++, sched->state.used);
 
 
 
 
 755
 756	return true;
 757}
 758
 759/*
 760 * Select a counter for the current event to schedule. Return true on
 761 * success.
 762 */
 763static bool __perf_sched_find_counter(struct perf_sched *sched)
 764{
 765	struct event_constraint *c;
 766	int idx;
 767
 768	if (!sched->state.unassigned)
 769		return false;
 770
 771	if (sched->state.event >= sched->max_events)
 772		return false;
 773
 774	c = sched->constraints[sched->state.event];
 775	/* Prefer fixed purpose counters */
 776	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
 777		idx = INTEL_PMC_IDX_FIXED;
 778		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
 779			if (!__test_and_set_bit(idx, sched->state.used))
 780				goto done;
 
 
 
 
 
 781		}
 782	}
 783
 784	/* Grab the first unused counter starting with idx */
 785	idx = sched->state.counter;
 786	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
 787		if (!__test_and_set_bit(idx, sched->state.used)) {
 788			if (sched->state.nr_gp++ >= sched->max_gp)
 789				return false;
 790
 791			goto done;
 792		}
 
 
 
 
 
 
 
 
 
 793	}
 794
 795	return false;
 796
 797done:
 798	sched->state.counter = idx;
 799
 800	if (c->overlap)
 801		perf_sched_save_state(sched);
 802
 803	return true;
 804}
 805
 806static bool perf_sched_find_counter(struct perf_sched *sched)
 807{
 808	while (!__perf_sched_find_counter(sched)) {
 809		if (!perf_sched_restore_state(sched))
 810			return false;
 811	}
 812
 813	return true;
 814}
 815
 816/*
 817 * Go through all unassigned events and find the next one to schedule.
 818 * Take events with the least weight first. Return true on success.
 819 */
 820static bool perf_sched_next_event(struct perf_sched *sched)
 821{
 822	struct event_constraint *c;
 823
 824	if (!sched->state.unassigned || !--sched->state.unassigned)
 825		return false;
 826
 827	do {
 828		/* next event */
 829		sched->state.event++;
 830		if (sched->state.event >= sched->max_events) {
 831			/* next weight */
 832			sched->state.event = 0;
 833			sched->state.weight++;
 834			if (sched->state.weight > sched->max_weight)
 835				return false;
 836		}
 837		c = sched->constraints[sched->state.event];
 838	} while (c->weight != sched->state.weight);
 839
 840	sched->state.counter = 0;	/* start with first counter */
 841
 842	return true;
 843}
 844
 845/*
 846 * Assign a counter for each event.
 847 */
 848int perf_assign_events(struct event_constraint **constraints, int n,
 849			int wmin, int wmax, int gpmax, int *assign)
 850{
 851	struct perf_sched sched;
 852
 853	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
 854
 855	do {
 856		if (!perf_sched_find_counter(&sched))
 857			break;	/* failed */
 858		if (assign)
 859			assign[sched.state.event] = sched.state.counter;
 860	} while (perf_sched_next_event(&sched));
 861
 862	return sched.state.unassigned;
 863}
 864EXPORT_SYMBOL_GPL(perf_assign_events);
 865
 866int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 867{
 868	struct event_constraint *c;
 869	unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 870	struct perf_event *e;
 871	int n0, i, wmin, wmax, unsched = 0;
 872	struct hw_perf_event *hwc;
 873
 874	bitmap_zero(used_mask, X86_PMC_IDX_MAX);
 875
 876	/*
 877	 * Compute the number of events already present; see x86_pmu_add(),
 878	 * validate_group() and x86_pmu_commit_txn(). For the former two
 879	 * cpuc->n_events hasn't been updated yet, while for the latter
 880	 * cpuc->n_txn contains the number of events added in the current
 881	 * transaction.
 882	 */
 883	n0 = cpuc->n_events;
 884	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
 885		n0 -= cpuc->n_txn;
 886
 887	if (x86_pmu.start_scheduling)
 888		x86_pmu.start_scheduling(cpuc);
 889
 890	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
 891		c = cpuc->event_constraint[i];
 892
 893		/*
 894		 * Previously scheduled events should have a cached constraint,
 895		 * while new events should not have one.
 896		 */
 897		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
 898
 899		/*
 900		 * Request constraints for new events; or for those events that
 901		 * have a dynamic constraint -- for those the constraint can
 902		 * change due to external factors (sibling state, allow_tfa).
 903		 */
 904		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
 905			c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
 906			cpuc->event_constraint[i] = c;
 907		}
 908
 909		wmin = min(wmin, c->weight);
 910		wmax = max(wmax, c->weight);
 911	}
 912
 913	/*
 914	 * fastpath, try to reuse previous register
 915	 */
 916	for (i = 0; i < n; i++) {
 
 
 917		hwc = &cpuc->event_list[i]->hw;
 918		c = cpuc->event_constraint[i];
 919
 920		/* never assigned */
 921		if (hwc->idx == -1)
 922			break;
 923
 924		/* constraint still honored */
 925		if (!test_bit(hwc->idx, c->idxmsk))
 926			break;
 927
 
 
 
 
 928		/* not already used */
 929		if (test_bit(hwc->idx, used_mask))
 930			break;
 931
 932		__set_bit(hwc->idx, used_mask);
 
 933		if (assign)
 934			assign[i] = hwc->idx;
 935	}
 936
 937	/* slow path */
 938	if (i != n) {
 939		int gpmax = x86_pmu.num_counters;
 940
 941		/*
 942		 * Do not allow scheduling of more than half the available
 943		 * generic counters.
 944		 *
 945		 * This helps avoid counter starvation of sibling thread by
 946		 * ensuring at most half the counters cannot be in exclusive
 947		 * mode. There is no designated counters for the limits. Any
 948		 * N/2 counters can be used. This helps with events with
 949		 * specific counter constraints.
 950		 */
 951		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
 952		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
 953			gpmax /= 2;
 954
 
 
 
 
 
 
 
 
 
 955		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
 956					     wmax, gpmax, assign);
 957	}
 958
 959	/*
 960	 * In case of success (unsched = 0), mark events as committed,
 961	 * so we do not put_constraint() in case new events are added
 962	 * and fail to be scheduled
 963	 *
 964	 * We invoke the lower level commit callback to lock the resource
 965	 *
 966	 * We do not need to do all of this in case we are called to
 967	 * validate an event group (assign == NULL)
 968	 */
 969	if (!unsched && assign) {
 970		for (i = 0; i < n; i++) {
 971			e = cpuc->event_list[i];
 972			if (x86_pmu.commit_scheduling)
 973				x86_pmu.commit_scheduling(cpuc, i, assign[i]);
 974		}
 975	} else {
 976		for (i = n0; i < n; i++) {
 977			e = cpuc->event_list[i];
 978
 979			/*
 980			 * release events that failed scheduling
 981			 */
 982			if (x86_pmu.put_event_constraints)
 983				x86_pmu.put_event_constraints(cpuc, e);
 984
 985			cpuc->event_constraint[i] = NULL;
 986		}
 987	}
 988
 989	if (x86_pmu.stop_scheduling)
 990		x86_pmu.stop_scheduling(cpuc);
 991
 992	return unsched ? -EINVAL : 0;
 993}
 994
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 995/*
 996 * dogrp: true if must collect siblings events (group)
 997 * returns total number of events and error code
 998 */
 999static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1000{
1001	struct perf_event *event;
1002	int n, max_count;
1003
1004	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
1005
1006	/* current number of events already accepted */
1007	n = cpuc->n_events;
1008	if (!cpuc->n_events)
1009		cpuc->pebs_output = 0;
1010
1011	if (!cpuc->is_fake && leader->attr.precise_ip) {
1012		/*
1013		 * For PEBS->PT, if !aux_event, the group leader (PT) went
1014		 * away, the group was broken down and this singleton event
1015		 * can't schedule any more.
1016		 */
1017		if (is_pebs_pt(leader) && !leader->aux_event)
1018			return -EINVAL;
1019
1020		/*
1021		 * pebs_output: 0: no PEBS so far, 1: PT, 2: DS
1022		 */
1023		if (cpuc->pebs_output &&
1024		    cpuc->pebs_output != is_pebs_pt(leader) + 1)
1025			return -EINVAL;
1026
1027		cpuc->pebs_output = is_pebs_pt(leader) + 1;
1028	}
1029
1030	if (is_x86_event(leader)) {
1031		if (n >= max_count)
1032			return -EINVAL;
1033		cpuc->event_list[n] = leader;
1034		n++;
1035	}
 
1036	if (!dogrp)
1037		return n;
1038
1039	for_each_sibling_event(event, leader) {
1040		if (!is_x86_event(event) ||
1041		    event->state <= PERF_EVENT_STATE_OFF)
1042			continue;
1043
1044		if (n >= max_count)
1045			return -EINVAL;
1046
1047		cpuc->event_list[n] = event;
1048		n++;
1049	}
1050	return n;
1051}
1052
1053static inline void x86_assign_hw_event(struct perf_event *event,
1054				struct cpu_hw_events *cpuc, int i)
1055{
1056	struct hw_perf_event *hwc = &event->hw;
 
1057
1058	hwc->idx = cpuc->assign[i];
1059	hwc->last_cpu = smp_processor_id();
1060	hwc->last_tag = ++cpuc->tags[i];
1061
1062	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
 
 
 
 
1063		hwc->config_base = 0;
1064		hwc->event_base	= 0;
1065	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
 
 
 
 
 
 
1066		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1067		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1068		hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1069	} else {
 
 
 
1070		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1071		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1072		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
 
1073	}
1074}
1075
1076/**
1077 * x86_perf_rdpmc_index - Return PMC counter used for event
1078 * @event: the perf_event to which the PMC counter was assigned
1079 *
1080 * The counter assigned to this performance event may change if interrupts
1081 * are enabled. This counter should thus never be used while interrupts are
1082 * enabled. Before this function is used to obtain the assigned counter the
1083 * event should be checked for validity using, for example,
1084 * perf_event_read_local(), within the same interrupt disabled section in
1085 * which this counter is planned to be used.
1086 *
1087 * Return: The index of the performance monitoring counter assigned to
1088 * @perf_event.
1089 */
1090int x86_perf_rdpmc_index(struct perf_event *event)
1091{
1092	lockdep_assert_irqs_disabled();
1093
1094	return event->hw.event_base_rdpmc;
1095}
1096
1097static inline int match_prev_assignment(struct hw_perf_event *hwc,
1098					struct cpu_hw_events *cpuc,
1099					int i)
1100{
1101	return hwc->idx == cpuc->assign[i] &&
1102		hwc->last_cpu == smp_processor_id() &&
1103		hwc->last_tag == cpuc->tags[i];
1104}
1105
1106static void x86_pmu_start(struct perf_event *event, int flags);
1107
1108static void x86_pmu_enable(struct pmu *pmu)
1109{
1110	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1111	struct perf_event *event;
1112	struct hw_perf_event *hwc;
1113	int i, added = cpuc->n_added;
1114
1115	if (!x86_pmu_initialized())
1116		return;
1117
1118	if (cpuc->enabled)
1119		return;
1120
1121	if (cpuc->n_added) {
1122		int n_running = cpuc->n_events - cpuc->n_added;
1123		/*
1124		 * apply assignment obtained either from
1125		 * hw_perf_group_sched_in() or x86_pmu_enable()
1126		 *
1127		 * step1: save events moving to new counters
1128		 */
1129		for (i = 0; i < n_running; i++) {
1130			event = cpuc->event_list[i];
1131			hwc = &event->hw;
1132
1133			/*
1134			 * we can avoid reprogramming counter if:
1135			 * - assigned same counter as last time
1136			 * - running on same CPU as last time
1137			 * - no other event has used the counter since
1138			 */
1139			if (hwc->idx == -1 ||
1140			    match_prev_assignment(hwc, cpuc, i))
1141				continue;
1142
1143			/*
1144			 * Ensure we don't accidentally enable a stopped
1145			 * counter simply because we rescheduled.
1146			 */
1147			if (hwc->state & PERF_HES_STOPPED)
1148				hwc->state |= PERF_HES_ARCH;
1149
1150			x86_pmu_stop(event, PERF_EF_UPDATE);
1151		}
1152
1153		/*
1154		 * step2: reprogram moved events into new counters
1155		 */
1156		for (i = 0; i < cpuc->n_events; i++) {
1157			event = cpuc->event_list[i];
1158			hwc = &event->hw;
1159
1160			if (!match_prev_assignment(hwc, cpuc, i))
1161				x86_assign_hw_event(event, cpuc, i);
1162			else if (i < n_running)
1163				continue;
1164
1165			if (hwc->state & PERF_HES_ARCH)
1166				continue;
1167
 
 
 
 
1168			x86_pmu_start(event, PERF_EF_RELOAD);
1169		}
1170		cpuc->n_added = 0;
1171		perf_events_lapic_init();
1172	}
1173
1174	cpuc->enabled = 1;
1175	barrier();
1176
1177	x86_pmu.enable_all(added);
1178}
1179
1180static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1181
1182/*
1183 * Set the next IRQ period, based on the hwc->period_left value.
1184 * To be called with the event disabled in hw:
1185 */
1186int x86_perf_event_set_period(struct perf_event *event)
1187{
1188	struct hw_perf_event *hwc = &event->hw;
1189	s64 left = local64_read(&hwc->period_left);
1190	s64 period = hwc->sample_period;
1191	int ret = 0, idx = hwc->idx;
1192
1193	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1194		return 0;
1195
1196	/*
1197	 * If we are way outside a reasonable range then just skip forward:
1198	 */
1199	if (unlikely(left <= -period)) {
1200		left = period;
1201		local64_set(&hwc->period_left, left);
1202		hwc->last_period = period;
1203		ret = 1;
1204	}
1205
1206	if (unlikely(left <= 0)) {
1207		left += period;
1208		local64_set(&hwc->period_left, left);
1209		hwc->last_period = period;
1210		ret = 1;
1211	}
1212	/*
1213	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1214	 */
1215	if (unlikely(left < 2))
1216		left = 2;
1217
1218	if (left > x86_pmu.max_period)
1219		left = x86_pmu.max_period;
1220
1221	if (x86_pmu.limit_period)
1222		left = x86_pmu.limit_period(event, left);
1223
1224	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1225
1226	/*
1227	 * The hw event starts counting from this event offset,
1228	 * mark it to be able to extra future deltas:
1229	 */
1230	local64_set(&hwc->prev_count, (u64)-left);
1231
1232	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1233
1234	/*
1235	 * Due to erratum on certan cpu we need
1236	 * a second write to be sure the register
1237	 * is updated properly
1238	 */
1239	if (x86_pmu.perfctr_second_write) {
1240		wrmsrl(hwc->event_base,
1241			(u64)(-left) & x86_pmu.cntval_mask);
1242	}
1243
1244	perf_event_update_userpage(event);
1245
1246	return ret;
1247}
1248
1249void x86_pmu_enable_event(struct perf_event *event)
1250{
1251	if (__this_cpu_read(cpu_hw_events.enabled))
1252		__x86_pmu_enable_event(&event->hw,
1253				       ARCH_PERFMON_EVENTSEL_ENABLE);
1254}
1255
1256/*
1257 * Add a single event to the PMU.
1258 *
1259 * The event is added to the group of enabled events
1260 * but only if it can be scheduled with existing events.
1261 */
1262static int x86_pmu_add(struct perf_event *event, int flags)
1263{
1264	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1265	struct hw_perf_event *hwc;
1266	int assign[X86_PMC_IDX_MAX];
1267	int n, n0, ret;
1268
1269	hwc = &event->hw;
1270
1271	n0 = cpuc->n_events;
1272	ret = n = collect_events(cpuc, event, false);
1273	if (ret < 0)
1274		goto out;
1275
1276	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1277	if (!(flags & PERF_EF_START))
1278		hwc->state |= PERF_HES_ARCH;
1279
1280	/*
1281	 * If group events scheduling transaction was started,
1282	 * skip the schedulability test here, it will be performed
1283	 * at commit time (->commit_txn) as a whole.
1284	 *
1285	 * If commit fails, we'll call ->del() on all events
1286	 * for which ->add() was called.
1287	 */
1288	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1289		goto done_collect;
1290
1291	ret = x86_pmu.schedule_events(cpuc, n, assign);
1292	if (ret)
1293		goto out;
1294	/*
1295	 * copy new assignment, now we know it is possible
1296	 * will be used by hw_perf_enable()
1297	 */
1298	memcpy(cpuc->assign, assign, n*sizeof(int));
1299
1300done_collect:
1301	/*
1302	 * Commit the collect_events() state. See x86_pmu_del() and
1303	 * x86_pmu_*_txn().
1304	 */
1305	cpuc->n_events = n;
1306	cpuc->n_added += n - n0;
1307	cpuc->n_txn += n - n0;
1308
1309	if (x86_pmu.add) {
1310		/*
1311		 * This is before x86_pmu_enable() will call x86_pmu_start(),
1312		 * so we enable LBRs before an event needs them etc..
1313		 */
1314		x86_pmu.add(event);
1315	}
1316
1317	ret = 0;
1318out:
1319	return ret;
1320}
1321
1322static void x86_pmu_start(struct perf_event *event, int flags)
1323{
1324	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1325	int idx = event->hw.idx;
1326
1327	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1328		return;
1329
1330	if (WARN_ON_ONCE(idx == -1))
1331		return;
1332
1333	if (flags & PERF_EF_RELOAD) {
1334		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1335		x86_perf_event_set_period(event);
1336	}
1337
1338	event->hw.state = 0;
1339
1340	cpuc->events[idx] = event;
1341	__set_bit(idx, cpuc->active_mask);
1342	__set_bit(idx, cpuc->running);
1343	x86_pmu.enable(event);
1344	perf_event_update_userpage(event);
1345}
1346
1347void perf_event_print_debug(void)
1348{
1349	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1350	u64 pebs, debugctl;
 
1351	struct cpu_hw_events *cpuc;
1352	unsigned long flags;
1353	int cpu, idx;
1354
1355	if (!x86_pmu.num_counters)
1356		return;
1357
1358	local_irq_save(flags);
1359
1360	cpu = smp_processor_id();
1361	cpuc = &per_cpu(cpu_hw_events, cpu);
 
 
 
 
 
 
1362
1363	if (x86_pmu.version >= 2) {
1364		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1365		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1366		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1367		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1368
1369		pr_info("\n");
1370		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1371		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1372		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1373		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1374		if (x86_pmu.pebs_constraints) {
1375			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1376			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1377		}
1378		if (x86_pmu.lbr_nr) {
1379			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1380			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1381		}
1382	}
1383	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1384
1385	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1386		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1387		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1388
1389		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1390
1391		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1392			cpu, idx, pmc_ctrl);
1393		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1394			cpu, idx, pmc_count);
1395		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1396			cpu, idx, prev_left);
1397	}
1398	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1399		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
 
 
1400
1401		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1402			cpu, idx, pmc_count);
1403	}
1404	local_irq_restore(flags);
1405}
1406
1407void x86_pmu_stop(struct perf_event *event, int flags)
1408{
1409	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1410	struct hw_perf_event *hwc = &event->hw;
1411
1412	if (test_bit(hwc->idx, cpuc->active_mask)) {
1413		x86_pmu.disable(event);
1414		__clear_bit(hwc->idx, cpuc->active_mask);
1415		cpuc->events[hwc->idx] = NULL;
1416		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1417		hwc->state |= PERF_HES_STOPPED;
1418	}
1419
1420	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1421		/*
1422		 * Drain the remaining delta count out of a event
1423		 * that we are disabling:
1424		 */
1425		x86_perf_event_update(event);
1426		hwc->state |= PERF_HES_UPTODATE;
1427	}
1428}
1429
1430static void x86_pmu_del(struct perf_event *event, int flags)
1431{
1432	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 
1433	int i;
1434
1435	/*
1436	 * If we're called during a txn, we only need to undo x86_pmu.add.
1437	 * The events never got scheduled and ->cancel_txn will truncate
1438	 * the event_list.
1439	 *
1440	 * XXX assumes any ->del() called during a TXN will only be on
1441	 * an event added during that same TXN.
1442	 */
1443	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1444		goto do_del;
1445
 
 
1446	/*
1447	 * Not a TXN, therefore cleanup properly.
1448	 */
1449	x86_pmu_stop(event, PERF_EF_UPDATE);
1450
1451	for (i = 0; i < cpuc->n_events; i++) {
1452		if (event == cpuc->event_list[i])
1453			break;
1454	}
1455
1456	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1457		return;
1458
1459	/* If we have a newly added event; make sure to decrease n_added. */
1460	if (i >= cpuc->n_events - cpuc->n_added)
1461		--cpuc->n_added;
1462
1463	if (x86_pmu.put_event_constraints)
1464		x86_pmu.put_event_constraints(cpuc, event);
1465
1466	/* Delete the array entry. */
1467	while (++i < cpuc->n_events) {
1468		cpuc->event_list[i-1] = cpuc->event_list[i];
1469		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
 
1470	}
1471	cpuc->event_constraint[i-1] = NULL;
1472	--cpuc->n_events;
 
 
1473
1474	perf_event_update_userpage(event);
1475
1476do_del:
1477	if (x86_pmu.del) {
1478		/*
1479		 * This is after x86_pmu_stop(); so we disable LBRs after any
1480		 * event can need them etc..
1481		 */
1482		x86_pmu.del(event);
1483	}
1484}
1485
1486int x86_pmu_handle_irq(struct pt_regs *regs)
1487{
1488	struct perf_sample_data data;
1489	struct cpu_hw_events *cpuc;
1490	struct perf_event *event;
1491	int idx, handled = 0;
1492	u64 val;
1493
1494	cpuc = this_cpu_ptr(&cpu_hw_events);
1495
1496	/*
1497	 * Some chipsets need to unmask the LVTPC in a particular spot
1498	 * inside the nmi handler.  As a result, the unmasking was pushed
1499	 * into all the nmi handlers.
1500	 *
1501	 * This generic handler doesn't seem to have any issues where the
1502	 * unmasking occurs so it was left at the top.
1503	 */
1504	apic_write(APIC_LVTPC, APIC_DM_NMI);
1505
1506	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1507		if (!test_bit(idx, cpuc->active_mask))
1508			continue;
1509
1510		event = cpuc->events[idx];
1511
1512		val = x86_perf_event_update(event);
1513		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1514			continue;
1515
1516		/*
1517		 * event overflow
1518		 */
1519		handled++;
1520		perf_sample_data_init(&data, 0, event->hw.last_period);
1521
1522		if (!x86_perf_event_set_period(event))
1523			continue;
1524
 
 
 
 
 
1525		if (perf_event_overflow(event, &data, regs))
1526			x86_pmu_stop(event, 0);
1527	}
1528
1529	if (handled)
1530		inc_irq_stat(apic_perf_irqs);
1531
1532	return handled;
1533}
1534
1535void perf_events_lapic_init(void)
1536{
1537	if (!x86_pmu.apic || !x86_pmu_initialized())
1538		return;
1539
1540	/*
1541	 * Always use NMI for PMU
1542	 */
1543	apic_write(APIC_LVTPC, APIC_DM_NMI);
1544}
1545
1546static int
1547perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1548{
1549	u64 start_clock;
1550	u64 finish_clock;
1551	int ret;
1552
1553	/*
1554	 * All PMUs/events that share this PMI handler should make sure to
1555	 * increment active_events for their events.
1556	 */
1557	if (!atomic_read(&active_events))
1558		return NMI_DONE;
1559
1560	start_clock = sched_clock();
1561	ret = x86_pmu.handle_irq(regs);
1562	finish_clock = sched_clock();
1563
1564	perf_sample_event_took(finish_clock - start_clock);
1565
1566	return ret;
1567}
1568NOKPROBE_SYMBOL(perf_event_nmi_handler);
1569
1570struct event_constraint emptyconstraint;
1571struct event_constraint unconstrained;
1572
1573static int x86_pmu_prepare_cpu(unsigned int cpu)
1574{
1575	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1576	int i;
1577
1578	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1579		cpuc->kfree_on_online[i] = NULL;
1580	if (x86_pmu.cpu_prepare)
1581		return x86_pmu.cpu_prepare(cpu);
1582	return 0;
1583}
1584
1585static int x86_pmu_dead_cpu(unsigned int cpu)
1586{
1587	if (x86_pmu.cpu_dead)
1588		x86_pmu.cpu_dead(cpu);
1589	return 0;
1590}
1591
1592static int x86_pmu_online_cpu(unsigned int cpu)
1593{
1594	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1595	int i;
1596
1597	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1598		kfree(cpuc->kfree_on_online[i]);
1599		cpuc->kfree_on_online[i] = NULL;
1600	}
1601	return 0;
1602}
1603
1604static int x86_pmu_starting_cpu(unsigned int cpu)
1605{
1606	if (x86_pmu.cpu_starting)
1607		x86_pmu.cpu_starting(cpu);
1608	return 0;
1609}
1610
1611static int x86_pmu_dying_cpu(unsigned int cpu)
1612{
1613	if (x86_pmu.cpu_dying)
1614		x86_pmu.cpu_dying(cpu);
1615	return 0;
1616}
1617
1618static void __init pmu_check_apic(void)
1619{
1620	if (boot_cpu_has(X86_FEATURE_APIC))
1621		return;
1622
1623	x86_pmu.apic = 0;
1624	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1625	pr_info("no hardware sampling interrupt available.\n");
1626
1627	/*
1628	 * If we have a PMU initialized but no APIC
1629	 * interrupts, we cannot sample hardware
1630	 * events (user-space has to fall back and
1631	 * sample via a hrtimer based software event):
1632	 */
1633	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1634
1635}
1636
1637static struct attribute_group x86_pmu_format_group __ro_after_init = {
1638	.name = "format",
1639	.attrs = NULL,
1640};
1641
1642ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1643{
1644	struct perf_pmu_events_attr *pmu_attr = \
1645		container_of(attr, struct perf_pmu_events_attr, attr);
1646	u64 config = x86_pmu.event_map(pmu_attr->id);
 
 
 
1647
1648	/* string trumps id */
1649	if (pmu_attr->event_str)
1650		return sprintf(page, "%s", pmu_attr->event_str);
1651
1652	return x86_pmu.events_sysfs_show(page, config);
1653}
1654EXPORT_SYMBOL_GPL(events_sysfs_show);
1655
1656ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1657			  char *page)
1658{
1659	struct perf_pmu_events_ht_attr *pmu_attr =
1660		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1661
1662	/*
1663	 * Report conditional events depending on Hyper-Threading.
1664	 *
1665	 * This is overly conservative as usually the HT special
1666	 * handling is not needed if the other CPU thread is idle.
1667	 *
1668	 * Note this does not (and cannot) handle the case when thread
1669	 * siblings are invisible, for example with virtualization
1670	 * if they are owned by some other guest.  The user tool
1671	 * has to re-read when a thread sibling gets onlined later.
1672	 */
1673	return sprintf(page, "%s",
1674			topology_max_smt_threads() > 1 ?
1675			pmu_attr->event_str_ht :
1676			pmu_attr->event_str_noht);
1677}
1678
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1679EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1680EVENT_ATTR(instructions,		INSTRUCTIONS		);
1681EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1682EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1683EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1684EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1685EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1686EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1687EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1688EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1689
1690static struct attribute *empty_attrs;
1691
1692static struct attribute *events_attr[] = {
1693	EVENT_PTR(CPU_CYCLES),
1694	EVENT_PTR(INSTRUCTIONS),
1695	EVENT_PTR(CACHE_REFERENCES),
1696	EVENT_PTR(CACHE_MISSES),
1697	EVENT_PTR(BRANCH_INSTRUCTIONS),
1698	EVENT_PTR(BRANCH_MISSES),
1699	EVENT_PTR(BUS_CYCLES),
1700	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1701	EVENT_PTR(STALLED_CYCLES_BACKEND),
1702	EVENT_PTR(REF_CPU_CYCLES),
1703	NULL,
1704};
1705
1706/*
1707 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1708 * out of events_attr attributes.
1709 */
1710static umode_t
1711is_visible(struct kobject *kobj, struct attribute *attr, int idx)
1712{
1713	struct perf_pmu_events_attr *pmu_attr;
1714
 
 
 
1715	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
1716	/* str trumps id */
1717	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
1718}
1719
1720static struct attribute_group x86_pmu_events_group __ro_after_init = {
1721	.name = "events",
1722	.attrs = events_attr,
1723	.is_visible = is_visible,
1724};
1725
1726ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1727{
1728	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1729	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1730	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1731	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1732	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1733	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1734	ssize_t ret;
1735
1736	/*
1737	* We have whole page size to spend and just little data
1738	* to write, so we can safely use sprintf.
1739	*/
1740	ret = sprintf(page, "event=0x%02llx", event);
1741
1742	if (umask)
1743		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1744
1745	if (edge)
1746		ret += sprintf(page + ret, ",edge");
1747
1748	if (pc)
1749		ret += sprintf(page + ret, ",pc");
1750
1751	if (any)
1752		ret += sprintf(page + ret, ",any");
1753
1754	if (inv)
1755		ret += sprintf(page + ret, ",inv");
1756
1757	if (cmask)
1758		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1759
1760	ret += sprintf(page + ret, "\n");
1761
1762	return ret;
1763}
1764
1765static struct attribute_group x86_pmu_attr_group;
1766static struct attribute_group x86_pmu_caps_group;
1767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1768static int __init init_hw_perf_events(void)
1769{
1770	struct x86_pmu_quirk *quirk;
1771	int err;
1772
1773	pr_info("Performance Events: ");
1774
1775	switch (boot_cpu_data.x86_vendor) {
1776	case X86_VENDOR_INTEL:
1777		err = intel_pmu_init();
1778		break;
1779	case X86_VENDOR_AMD:
1780		err = amd_pmu_init();
1781		break;
1782	case X86_VENDOR_HYGON:
1783		err = amd_pmu_init();
1784		x86_pmu.name = "HYGON";
1785		break;
 
 
 
 
1786	default:
1787		err = -ENOTSUPP;
1788	}
1789	if (err != 0) {
1790		pr_cont("no PMU driver, software events only.\n");
1791		return 0;
 
1792	}
1793
1794	pmu_check_apic();
1795
1796	/* sanity check that the hardware exists or is emulated */
1797	if (!check_hw_exists())
1798		return 0;
1799
1800	pr_cont("%s PMU driver.\n", x86_pmu.name);
1801
1802	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1803
1804	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1805		quirk->func();
1806
1807	if (!x86_pmu.intel_ctrl)
1808		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
 
 
 
1809
1810	perf_events_lapic_init();
1811	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1812
1813	unconstrained = (struct event_constraint)
1814		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1815				   0, x86_pmu.num_counters, 0, 0);
1816
1817	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1818
1819	if (!x86_pmu.events_sysfs_show)
1820		x86_pmu_events_group.attrs = &empty_attrs;
1821
1822	pmu.attr_update = x86_pmu.attr_update;
1823
1824	pr_info("... version:                %d\n",     x86_pmu.version);
1825	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1826	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1827	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1828	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1829	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1830	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
 
 
 
 
 
 
 
 
 
1831
1832	/*
1833	 * Install callbacks. Core will call them for each online
1834	 * cpu.
1835	 */
1836	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
1837				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1838	if (err)
1839		return err;
1840
1841	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1842				"perf/x86:starting", x86_pmu_starting_cpu,
1843				x86_pmu_dying_cpu);
1844	if (err)
1845		goto out;
1846
1847	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
1848				x86_pmu_online_cpu, NULL);
1849	if (err)
1850		goto out1;
1851
1852	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1853	if (err)
1854		goto out2;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1855
1856	return 0;
1857
1858out2:
1859	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1860out1:
1861	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1862out:
1863	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
 
 
1864	return err;
1865}
1866early_initcall(init_hw_perf_events);
1867
1868static inline void x86_pmu_read(struct perf_event *event)
1869{
1870	if (x86_pmu.read)
1871		return x86_pmu.read(event);
1872	x86_perf_event_update(event);
1873}
1874
1875/*
1876 * Start group events scheduling transaction
1877 * Set the flag to make pmu::enable() not perform the
1878 * schedulability test, it will be performed at commit time
1879 *
1880 * We only support PERF_PMU_TXN_ADD transactions. Save the
1881 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1882 * transactions.
1883 */
1884static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1885{
1886	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1887
1888	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1889
1890	cpuc->txn_flags = txn_flags;
1891	if (txn_flags & ~PERF_PMU_TXN_ADD)
1892		return;
1893
1894	perf_pmu_disable(pmu);
1895	__this_cpu_write(cpu_hw_events.n_txn, 0);
 
 
1896}
1897
1898/*
1899 * Stop group events scheduling transaction
1900 * Clear the flag and pmu::enable() will perform the
1901 * schedulability test.
1902 */
1903static void x86_pmu_cancel_txn(struct pmu *pmu)
1904{
1905	unsigned int txn_flags;
1906	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1907
1908	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1909
1910	txn_flags = cpuc->txn_flags;
1911	cpuc->txn_flags = 0;
1912	if (txn_flags & ~PERF_PMU_TXN_ADD)
1913		return;
1914
1915	/*
1916	 * Truncate collected array by the number of events added in this
1917	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1918	 */
1919	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1920	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
 
 
1921	perf_pmu_enable(pmu);
1922}
1923
1924/*
1925 * Commit group events scheduling transaction
1926 * Perform the group schedulability test as a whole
1927 * Return 0 if success
1928 *
1929 * Does not cancel the transaction on failure; expects the caller to do this.
1930 */
1931static int x86_pmu_commit_txn(struct pmu *pmu)
1932{
1933	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1934	int assign[X86_PMC_IDX_MAX];
1935	int n, ret;
1936
1937	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1938
1939	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1940		cpuc->txn_flags = 0;
1941		return 0;
1942	}
1943
1944	n = cpuc->n_events;
1945
1946	if (!x86_pmu_initialized())
1947		return -EAGAIN;
1948
1949	ret = x86_pmu.schedule_events(cpuc, n, assign);
1950	if (ret)
1951		return ret;
1952
1953	/*
1954	 * copy new assignment, now we know it is possible
1955	 * will be used by hw_perf_enable()
1956	 */
1957	memcpy(cpuc->assign, assign, n*sizeof(int));
1958
1959	cpuc->txn_flags = 0;
1960	perf_pmu_enable(pmu);
1961	return 0;
1962}
1963/*
1964 * a fake_cpuc is used to validate event groups. Due to
1965 * the extra reg logic, we need to also allocate a fake
1966 * per_core and per_cpu structure. Otherwise, group events
1967 * using extra reg may conflict without the kernel being
1968 * able to catch this when the last event gets added to
1969 * the group.
1970 */
1971static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1972{
1973	intel_cpuc_finish(cpuc);
1974	kfree(cpuc);
1975}
1976
1977static struct cpu_hw_events *allocate_fake_cpuc(void)
1978{
1979	struct cpu_hw_events *cpuc;
1980	int cpu = raw_smp_processor_id();
1981
1982	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1983	if (!cpuc)
1984		return ERR_PTR(-ENOMEM);
1985	cpuc->is_fake = 1;
1986
 
 
 
 
 
 
 
 
 
 
 
1987	if (intel_cpuc_prepare(cpuc, cpu))
1988		goto error;
1989
1990	return cpuc;
1991error:
1992	free_fake_cpuc(cpuc);
1993	return ERR_PTR(-ENOMEM);
1994}
1995
1996/*
1997 * validate that we can schedule this event
1998 */
1999static int validate_event(struct perf_event *event)
2000{
2001	struct cpu_hw_events *fake_cpuc;
2002	struct event_constraint *c;
2003	int ret = 0;
2004
2005	fake_cpuc = allocate_fake_cpuc();
2006	if (IS_ERR(fake_cpuc))
2007		return PTR_ERR(fake_cpuc);
2008
2009	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2010
2011	if (!c || !c->weight)
2012		ret = -EINVAL;
2013
2014	if (x86_pmu.put_event_constraints)
2015		x86_pmu.put_event_constraints(fake_cpuc, event);
2016
2017	free_fake_cpuc(fake_cpuc);
2018
2019	return ret;
2020}
2021
2022/*
2023 * validate a single event group
2024 *
2025 * validation include:
2026 *	- check events are compatible which each other
2027 *	- events do not compete for the same counter
2028 *	- number of events <= number of counters
2029 *
2030 * validation ensures the group can be loaded onto the
2031 * PMU if it was the only group available.
2032 */
2033static int validate_group(struct perf_event *event)
2034{
2035	struct perf_event *leader = event->group_leader;
2036	struct cpu_hw_events *fake_cpuc;
2037	int ret = -EINVAL, n;
2038
2039	fake_cpuc = allocate_fake_cpuc();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2040	if (IS_ERR(fake_cpuc))
2041		return PTR_ERR(fake_cpuc);
2042	/*
2043	 * the event is not yet connected with its
2044	 * siblings therefore we must first collect
2045	 * existing siblings, then add the new event
2046	 * before we can simulate the scheduling
2047	 */
2048	n = collect_events(fake_cpuc, leader, true);
2049	if (n < 0)
2050		goto out;
2051
2052	fake_cpuc->n_events = n;
2053	n = collect_events(fake_cpuc, event, false);
2054	if (n < 0)
2055		goto out;
2056
2057	fake_cpuc->n_events = 0;
2058	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2059
2060out:
2061	free_fake_cpuc(fake_cpuc);
2062	return ret;
2063}
2064
2065static int x86_pmu_event_init(struct perf_event *event)
2066{
2067	struct pmu *tmp;
2068	int err;
2069
2070	switch (event->attr.type) {
2071	case PERF_TYPE_RAW:
2072	case PERF_TYPE_HARDWARE:
2073	case PERF_TYPE_HW_CACHE:
2074		break;
2075
2076	default:
2077		return -ENOENT;
 
 
 
 
 
2078	}
2079
2080	err = __x86_pmu_event_init(event);
2081	if (!err) {
2082		/*
2083		 * we temporarily connect event to its pmu
2084		 * such that validate_group() can classify
2085		 * it as an x86 event using is_x86_event()
2086		 */
2087		tmp = event->pmu;
2088		event->pmu = &pmu;
2089
2090		if (event->group_leader != event)
2091			err = validate_group(event);
2092		else
2093			err = validate_event(event);
2094
2095		event->pmu = tmp;
2096	}
2097	if (err) {
2098		if (event->destroy)
2099			event->destroy(event);
 
2100	}
2101
2102	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2103	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2104		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2105
2106	return err;
2107}
2108
2109static void refresh_pce(void *ignored)
2110{
2111	load_mm_cr4_irqsoff(this_cpu_read(cpu_tlbstate.loaded_mm));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2112}
2113
2114static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2115{
2116	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2117		return;
2118
2119	/*
2120	 * This function relies on not being called concurrently in two
2121	 * tasks in the same mm.  Otherwise one task could observe
2122	 * perf_rdpmc_allowed > 1 and return all the way back to
2123	 * userspace with CR4.PCE clear while another task is still
2124	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2125	 *
2126	 * For now, this can't happen because all callers hold mmap_sem
2127	 * for write.  If this changes, we'll need a different solution.
2128	 */
2129	lockdep_assert_held_write(&mm->mmap_sem);
2130
2131	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2132		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2133}
2134
2135static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2136{
2137
2138	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2139		return;
2140
2141	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2142		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2143}
2144
2145static int x86_pmu_event_idx(struct perf_event *event)
2146{
2147	int idx = event->hw.idx;
2148
2149	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2150		return 0;
2151
2152	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2153		idx -= INTEL_PMC_IDX_FIXED;
2154		idx |= 1 << 30;
2155	}
2156
2157	return idx + 1;
2158}
2159
2160static ssize_t get_attr_rdpmc(struct device *cdev,
2161			      struct device_attribute *attr,
2162			      char *buf)
2163{
2164	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2165}
2166
2167static ssize_t set_attr_rdpmc(struct device *cdev,
2168			      struct device_attribute *attr,
2169			      const char *buf, size_t count)
2170{
 
2171	unsigned long val;
2172	ssize_t ret;
2173
2174	ret = kstrtoul(buf, 0, &val);
2175	if (ret)
2176		return ret;
2177
2178	if (val > 2)
2179		return -EINVAL;
2180
2181	if (x86_pmu.attr_rdpmc_broken)
2182		return -ENOTSUPP;
2183
2184	if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
 
 
2185		/*
2186		 * Changing into or out of always available, aka
2187		 * perf-event-bypassing mode.  This path is extremely slow,
2188		 * but only root can trigger it, so it's okay.
2189		 */
 
 
 
 
 
2190		if (val == 2)
2191			static_branch_inc(&rdpmc_always_available_key);
2192		else
2193			static_branch_dec(&rdpmc_always_available_key);
2194		on_each_cpu(refresh_pce, NULL, 1);
2195	}
2196
2197	x86_pmu.attr_rdpmc = val;
 
 
2198
2199	return count;
2200}
2201
2202static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2203
2204static struct attribute *x86_pmu_attrs[] = {
2205	&dev_attr_rdpmc.attr,
2206	NULL,
2207};
2208
2209static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2210	.attrs = x86_pmu_attrs,
2211};
2212
2213static ssize_t max_precise_show(struct device *cdev,
2214				  struct device_attribute *attr,
2215				  char *buf)
2216{
2217	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2218}
2219
2220static DEVICE_ATTR_RO(max_precise);
2221
2222static struct attribute *x86_pmu_caps_attrs[] = {
2223	&dev_attr_max_precise.attr,
2224	NULL
2225};
2226
2227static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2228	.name = "caps",
2229	.attrs = x86_pmu_caps_attrs,
2230};
2231
2232static const struct attribute_group *x86_pmu_attr_groups[] = {
2233	&x86_pmu_attr_group,
2234	&x86_pmu_format_group,
2235	&x86_pmu_events_group,
2236	&x86_pmu_caps_group,
2237	NULL,
2238};
2239
2240static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
 
 
 
 
 
 
2241{
2242	if (x86_pmu.sched_task)
2243		x86_pmu.sched_task(ctx, sched_in);
2244}
2245
2246void perf_check_microcode(void)
2247{
2248	if (x86_pmu.check_microcode)
2249		x86_pmu.check_microcode();
2250}
2251
2252static int x86_pmu_check_period(struct perf_event *event, u64 value)
2253{
2254	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2255		return -EINVAL;
2256
2257	if (value && x86_pmu.limit_period) {
2258		if (x86_pmu.limit_period(event, value) > value)
 
 
2259			return -EINVAL;
2260	}
2261
2262	return 0;
2263}
2264
2265static int x86_pmu_aux_output_match(struct perf_event *event)
2266{
2267	if (!(pmu.capabilities & PERF_PMU_CAP_AUX_OUTPUT))
2268		return 0;
2269
2270	if (x86_pmu.aux_output_match)
2271		return x86_pmu.aux_output_match(event);
2272
2273	return 0;
2274}
2275
 
 
 
 
 
 
 
 
 
2276static struct pmu pmu = {
2277	.pmu_enable		= x86_pmu_enable,
2278	.pmu_disable		= x86_pmu_disable,
2279
2280	.attr_groups		= x86_pmu_attr_groups,
2281
2282	.event_init		= x86_pmu_event_init,
2283
2284	.event_mapped		= x86_pmu_event_mapped,
2285	.event_unmapped		= x86_pmu_event_unmapped,
2286
2287	.add			= x86_pmu_add,
2288	.del			= x86_pmu_del,
2289	.start			= x86_pmu_start,
2290	.stop			= x86_pmu_stop,
2291	.read			= x86_pmu_read,
2292
2293	.start_txn		= x86_pmu_start_txn,
2294	.cancel_txn		= x86_pmu_cancel_txn,
2295	.commit_txn		= x86_pmu_commit_txn,
2296
2297	.event_idx		= x86_pmu_event_idx,
2298	.sched_task		= x86_pmu_sched_task,
2299	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2300	.check_period		= x86_pmu_check_period,
2301
2302	.aux_output_match	= x86_pmu_aux_output_match,
 
 
2303};
2304
2305void arch_perf_update_userpage(struct perf_event *event,
2306			       struct perf_event_mmap_page *userpg, u64 now)
2307{
2308	struct cyc2ns_data data;
2309	u64 offset;
2310
2311	userpg->cap_user_time = 0;
2312	userpg->cap_user_time_zero = 0;
2313	userpg->cap_user_rdpmc =
2314		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2315	userpg->pmc_width = x86_pmu.cntval_bits;
2316
2317	if (!using_native_sched_clock() || !sched_clock_stable())
2318		return;
2319
2320	cyc2ns_read_begin(&data);
2321
2322	offset = data.cyc2ns_offset + __sched_clock_offset;
2323
2324	/*
2325	 * Internal timekeeping for enabled/running/stopped times
2326	 * is always in the local_clock domain.
2327	 */
2328	userpg->cap_user_time = 1;
2329	userpg->time_mult = data.cyc2ns_mul;
2330	userpg->time_shift = data.cyc2ns_shift;
2331	userpg->time_offset = offset - now;
2332
2333	/*
2334	 * cap_user_time_zero doesn't make sense when we're using a different
2335	 * time base for the records.
2336	 */
2337	if (!event->attr.use_clockid) {
2338		userpg->cap_user_time_zero = 1;
2339		userpg->time_zero = offset;
2340	}
2341
2342	cyc2ns_read_end();
2343}
2344
2345/*
2346 * Determine whether the regs were taken from an irq/exception handler rather
2347 * than from perf_arch_fetch_caller_regs().
2348 */
2349static bool perf_hw_regs(struct pt_regs *regs)
2350{
2351	return regs->flags & X86_EFLAGS_FIXED;
2352}
2353
2354void
2355perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2356{
2357	struct unwind_state state;
2358	unsigned long addr;
2359
2360	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2361		/* TODO: We don't support guest os callchain now */
2362		return;
2363	}
2364
2365	if (perf_callchain_store(entry, regs->ip))
2366		return;
2367
2368	if (perf_hw_regs(regs))
2369		unwind_start(&state, current, regs, NULL);
2370	else
2371		unwind_start(&state, current, NULL, (void *)regs->sp);
2372
2373	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2374		addr = unwind_get_return_address(&state);
2375		if (!addr || perf_callchain_store(entry, addr))
2376			return;
2377	}
2378}
2379
2380static inline int
2381valid_user_frame(const void __user *fp, unsigned long size)
2382{
2383	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2384}
2385
2386static unsigned long get_segment_base(unsigned int segment)
2387{
2388	struct desc_struct *desc;
2389	unsigned int idx = segment >> 3;
2390
2391	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2392#ifdef CONFIG_MODIFY_LDT_SYSCALL
2393		struct ldt_struct *ldt;
2394
2395		/* IRQs are off, so this synchronizes with smp_store_release */
2396		ldt = READ_ONCE(current->active_mm->context.ldt);
2397		if (!ldt || idx >= ldt->nr_entries)
2398			return 0;
2399
2400		desc = &ldt->entries[idx];
2401#else
2402		return 0;
2403#endif
2404	} else {
2405		if (idx >= GDT_ENTRIES)
2406			return 0;
2407
2408		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2409	}
2410
2411	return get_desc_base(desc);
2412}
2413
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2414#ifdef CONFIG_IA32_EMULATION
2415
2416#include <linux/compat.h>
2417
2418static inline int
2419perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2420{
2421	/* 32-bit process in 64-bit kernel. */
2422	unsigned long ss_base, cs_base;
2423	struct stack_frame_ia32 frame;
2424	const void __user *fp;
 
2425
2426	if (!test_thread_flag(TIF_IA32))
2427		return 0;
2428
2429	cs_base = get_segment_base(regs->cs);
2430	ss_base = get_segment_base(regs->ss);
2431
2432	fp = compat_ptr(ss_base + regs->bp);
2433	pagefault_disable();
2434	while (entry->nr < entry->max_stack) {
2435		unsigned long bytes;
2436		frame.next_frame     = 0;
2437		frame.return_address = 0;
2438
 
 
 
 
 
 
2439		if (!valid_user_frame(fp, sizeof(frame)))
2440			break;
2441
2442		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2443		if (bytes != 0)
2444			break;
2445		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2446		if (bytes != 0)
2447			break;
2448
2449		perf_callchain_store(entry, cs_base + frame.return_address);
2450		fp = compat_ptr(ss_base + frame.next_frame);
2451	}
2452	pagefault_enable();
2453	return 1;
2454}
2455#else
2456static inline int
2457perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2458{
2459    return 0;
2460}
2461#endif
2462
2463void
2464perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2465{
2466	struct stack_frame frame;
2467	const unsigned long __user *fp;
 
2468
2469	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2470		/* TODO: We don't support guest os callchain now */
2471		return;
2472	}
2473
2474	/*
2475	 * We don't know what to do with VM86 stacks.. ignore them for now.
2476	 */
2477	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2478		return;
2479
2480	fp = (unsigned long __user *)regs->bp;
2481
2482	perf_callchain_store(entry, regs->ip);
2483
2484	if (!nmi_uaccess_okay())
2485		return;
2486
2487	if (perf_callchain_user32(regs, entry))
2488		return;
2489
2490	pagefault_disable();
2491	while (entry->nr < entry->max_stack) {
2492		unsigned long bytes;
2493
2494		frame.next_frame	     = NULL;
2495		frame.return_address = 0;
 
 
 
 
 
 
 
 
 
2496
 
2497		if (!valid_user_frame(fp, sizeof(frame)))
2498			break;
2499
2500		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2501		if (bytes != 0)
2502			break;
2503		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2504		if (bytes != 0)
2505			break;
2506
2507		perf_callchain_store(entry, frame.return_address);
2508		fp = (void __user *)frame.next_frame;
2509	}
2510	pagefault_enable();
2511}
2512
2513/*
2514 * Deal with code segment offsets for the various execution modes:
2515 *
2516 *   VM86 - the good olde 16 bit days, where the linear address is
2517 *          20 bits and we use regs->ip + 0x10 * regs->cs.
2518 *
2519 *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2520 *          to figure out what the 32bit base address is.
2521 *
2522 *    X32 - has TIF_X32 set, but is running in x86_64
2523 *
2524 * X86_64 - CS,DS,SS,ES are all zero based.
2525 */
2526static unsigned long code_segment_base(struct pt_regs *regs)
2527{
2528	/*
2529	 * For IA32 we look at the GDT/LDT segment base to convert the
2530	 * effective IP to a linear address.
2531	 */
2532
2533#ifdef CONFIG_X86_32
2534	/*
2535	 * If we are in VM86 mode, add the segment offset to convert to a
2536	 * linear address.
2537	 */
2538	if (regs->flags & X86_VM_MASK)
2539		return 0x10 * regs->cs;
2540
2541	if (user_mode(regs) && regs->cs != __USER_CS)
2542		return get_segment_base(regs->cs);
2543#else
2544	if (user_mode(regs) && !user_64bit_mode(regs) &&
2545	    regs->cs != __USER32_CS)
2546		return get_segment_base(regs->cs);
2547#endif
2548	return 0;
2549}
2550
2551unsigned long perf_instruction_pointer(struct pt_regs *regs)
2552{
2553	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2554		return perf_guest_cbs->get_guest_ip();
2555
2556	return regs->ip + code_segment_base(regs);
2557}
2558
2559unsigned long perf_misc_flags(struct pt_regs *regs)
2560{
2561	int misc = 0;
 
2562
2563	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2564		if (perf_guest_cbs->is_user_mode())
2565			misc |= PERF_RECORD_MISC_GUEST_USER;
2566		else
2567			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2568	} else {
2569		if (user_mode(regs))
2570			misc |= PERF_RECORD_MISC_USER;
2571		else
2572			misc |= PERF_RECORD_MISC_KERNEL;
2573	}
2574
2575	if (regs->flags & PERF_EFLAGS_EXACT)
2576		misc |= PERF_RECORD_MISC_EXACT_IP;
 
2577
2578	return misc;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2579}
2580
2581void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2582{
 
 
 
 
 
 
 
 
 
 
 
 
2583	cap->version		= x86_pmu.version;
2584	cap->num_counters_gp	= x86_pmu.num_counters;
2585	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2586	cap->bit_width_gp	= x86_pmu.cntval_bits;
2587	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2588	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2589	cap->events_mask_len	= x86_pmu.events_mask_len;
 
2590}
2591EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
v6.13.7
   1/*
   2 * Performance events x86 architecture code
   3 *
   4 *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   5 *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
   6 *  Copyright (C) 2009 Jaswinder Singh Rajput
   7 *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
   8 *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
   9 *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
  10 *  Copyright (C) 2009 Google, Inc., Stephane Eranian
  11 *
  12 *  For licencing details see kernel-base/COPYING
  13 */
  14
  15#include <linux/perf_event.h>
  16#include <linux/capability.h>
  17#include <linux/notifier.h>
  18#include <linux/hardirq.h>
  19#include <linux/kprobes.h>
  20#include <linux/export.h>
  21#include <linux/init.h>
  22#include <linux/kdebug.h>
  23#include <linux/sched/mm.h>
  24#include <linux/sched/clock.h>
  25#include <linux/uaccess.h>
  26#include <linux/slab.h>
  27#include <linux/cpu.h>
  28#include <linux/bitops.h>
  29#include <linux/device.h>
  30#include <linux/nospec.h>
  31#include <linux/static_call.h>
  32
  33#include <asm/apic.h>
  34#include <asm/stacktrace.h>
  35#include <asm/nmi.h>
  36#include <asm/smp.h>
  37#include <asm/alternative.h>
  38#include <asm/mmu_context.h>
  39#include <asm/tlbflush.h>
  40#include <asm/timer.h>
  41#include <asm/desc.h>
  42#include <asm/ldt.h>
  43#include <asm/unwind.h>
  44#include <asm/uprobes.h>
  45#include <asm/ibt.h>
  46
  47#include "perf_event.h"
  48
  49struct x86_pmu x86_pmu __read_mostly;
  50static struct pmu pmu;
  51
  52DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
  53	.enabled = 1,
  54	.pmu = &pmu,
  55};
  56
  57DEFINE_STATIC_KEY_FALSE(rdpmc_never_available_key);
  58DEFINE_STATIC_KEY_FALSE(rdpmc_always_available_key);
  59DEFINE_STATIC_KEY_FALSE(perf_is_hybrid);
  60
  61/*
  62 * This here uses DEFINE_STATIC_CALL_NULL() to get a static_call defined
  63 * from just a typename, as opposed to an actual function.
  64 */
  65DEFINE_STATIC_CALL_NULL(x86_pmu_handle_irq,  *x86_pmu.handle_irq);
  66DEFINE_STATIC_CALL_NULL(x86_pmu_disable_all, *x86_pmu.disable_all);
  67DEFINE_STATIC_CALL_NULL(x86_pmu_enable_all,  *x86_pmu.enable_all);
  68DEFINE_STATIC_CALL_NULL(x86_pmu_enable,	     *x86_pmu.enable);
  69DEFINE_STATIC_CALL_NULL(x86_pmu_disable,     *x86_pmu.disable);
  70
  71DEFINE_STATIC_CALL_NULL(x86_pmu_assign, *x86_pmu.assign);
  72
  73DEFINE_STATIC_CALL_NULL(x86_pmu_add,  *x86_pmu.add);
  74DEFINE_STATIC_CALL_NULL(x86_pmu_del,  *x86_pmu.del);
  75DEFINE_STATIC_CALL_NULL(x86_pmu_read, *x86_pmu.read);
  76
  77DEFINE_STATIC_CALL_NULL(x86_pmu_set_period,   *x86_pmu.set_period);
  78DEFINE_STATIC_CALL_NULL(x86_pmu_update,       *x86_pmu.update);
  79DEFINE_STATIC_CALL_NULL(x86_pmu_limit_period, *x86_pmu.limit_period);
  80
  81DEFINE_STATIC_CALL_NULL(x86_pmu_schedule_events,       *x86_pmu.schedule_events);
  82DEFINE_STATIC_CALL_NULL(x86_pmu_get_event_constraints, *x86_pmu.get_event_constraints);
  83DEFINE_STATIC_CALL_NULL(x86_pmu_put_event_constraints, *x86_pmu.put_event_constraints);
  84
  85DEFINE_STATIC_CALL_NULL(x86_pmu_start_scheduling,  *x86_pmu.start_scheduling);
  86DEFINE_STATIC_CALL_NULL(x86_pmu_commit_scheduling, *x86_pmu.commit_scheduling);
  87DEFINE_STATIC_CALL_NULL(x86_pmu_stop_scheduling,   *x86_pmu.stop_scheduling);
  88
  89DEFINE_STATIC_CALL_NULL(x86_pmu_sched_task,    *x86_pmu.sched_task);
  90DEFINE_STATIC_CALL_NULL(x86_pmu_swap_task_ctx, *x86_pmu.swap_task_ctx);
  91
  92DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs,   *x86_pmu.drain_pebs);
  93DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
  94
  95DEFINE_STATIC_CALL_NULL(x86_pmu_filter, *x86_pmu.filter);
  96
  97/*
  98 * This one is magic, it will get called even when PMU init fails (because
  99 * there is no PMU), in which case it should simply return NULL.
 100 */
 101DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs);
 102
 103u64 __read_mostly hw_cache_event_ids
 104				[PERF_COUNT_HW_CACHE_MAX]
 105				[PERF_COUNT_HW_CACHE_OP_MAX]
 106				[PERF_COUNT_HW_CACHE_RESULT_MAX];
 107u64 __read_mostly hw_cache_extra_regs
 108				[PERF_COUNT_HW_CACHE_MAX]
 109				[PERF_COUNT_HW_CACHE_OP_MAX]
 110				[PERF_COUNT_HW_CACHE_RESULT_MAX];
 111
 112/*
 113 * Propagate event elapsed time into the generic event.
 114 * Can only be executed on the CPU where the event is active.
 115 * Returns the delta events processed.
 116 */
 117u64 x86_perf_event_update(struct perf_event *event)
 118{
 119	struct hw_perf_event *hwc = &event->hw;
 120	int shift = 64 - x86_pmu.cntval_bits;
 121	u64 prev_raw_count, new_raw_count;
 
 122	u64 delta;
 123
 124	if (unlikely(!hwc->event_base))
 125		return 0;
 126
 127	/*
 128	 * Careful: an NMI might modify the previous event value.
 129	 *
 130	 * Our tactic to handle this is to first atomically read and
 131	 * exchange a new raw count - then add that new-prev delta
 132	 * count to the generic event atomically:
 133	 */
 
 134	prev_raw_count = local64_read(&hwc->prev_count);
 135	do {
 136		rdpmcl(hwc->event_base_rdpmc, new_raw_count);
 137	} while (!local64_try_cmpxchg(&hwc->prev_count,
 138				      &prev_raw_count, new_raw_count));
 
 139
 140	/*
 141	 * Now we have the new raw value and have updated the prev
 142	 * timestamp already. We can now calculate the elapsed delta
 143	 * (event-)time and add that to the generic event.
 144	 *
 145	 * Careful, not all hw sign-extends above the physical width
 146	 * of the count.
 147	 */
 148	delta = (new_raw_count << shift) - (prev_raw_count << shift);
 149	delta >>= shift;
 150
 151	local64_add(delta, &event->count);
 152	local64_sub(delta, &hwc->period_left);
 153
 154	return new_raw_count;
 155}
 156
 157/*
 158 * Find and validate any extra registers to set up.
 159 */
 160static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
 161{
 162	struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs);
 163	struct hw_perf_event_extra *reg;
 164	struct extra_reg *er;
 165
 166	reg = &event->hw.extra_reg;
 167
 168	if (!extra_regs)
 169		return 0;
 170
 171	for (er = extra_regs; er->msr; er++) {
 172		if (er->event != (config & er->config_mask))
 173			continue;
 174		if (event->attr.config1 & ~er->valid_mask)
 175			return -EINVAL;
 176		/* Check if the extra msrs can be safely accessed*/
 177		if (!er->extra_msr_access)
 178			return -ENXIO;
 179
 180		reg->idx = er->idx;
 181		reg->config = event->attr.config1;
 182		reg->reg = er->msr;
 183		break;
 184	}
 185	return 0;
 186}
 187
 188static atomic_t active_events;
 189static atomic_t pmc_refcount;
 190static DEFINE_MUTEX(pmc_reserve_mutex);
 191
 192#ifdef CONFIG_X86_LOCAL_APIC
 193
 194static inline u64 get_possible_counter_mask(void)
 195{
 196	u64 cntr_mask = x86_pmu.cntr_mask64;
 197	int i;
 198
 199	if (!is_hybrid())
 200		return cntr_mask;
 201
 202	for (i = 0; i < x86_pmu.num_hybrid_pmus; i++)
 203		cntr_mask |= x86_pmu.hybrid_pmu[i].cntr_mask64;
 204
 205	return cntr_mask;
 206}
 207
 208static bool reserve_pmc_hardware(void)
 209{
 210	u64 cntr_mask = get_possible_counter_mask();
 211	int i, end;
 212
 213	for_each_set_bit(i, (unsigned long *)&cntr_mask, X86_PMC_IDX_MAX) {
 214		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
 215			goto perfctr_fail;
 216	}
 217
 218	for_each_set_bit(i, (unsigned long *)&cntr_mask, X86_PMC_IDX_MAX) {
 219		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
 220			goto eventsel_fail;
 221	}
 222
 223	return true;
 224
 225eventsel_fail:
 226	end = i;
 227	for_each_set_bit(i, (unsigned long *)&cntr_mask, end)
 228		release_evntsel_nmi(x86_pmu_config_addr(i));
 229	i = X86_PMC_IDX_MAX;
 
 230
 231perfctr_fail:
 232	end = i;
 233	for_each_set_bit(i, (unsigned long *)&cntr_mask, end)
 234		release_perfctr_nmi(x86_pmu_event_addr(i));
 235
 236	return false;
 237}
 238
 239static void release_pmc_hardware(void)
 240{
 241	u64 cntr_mask = get_possible_counter_mask();
 242	int i;
 243
 244	for_each_set_bit(i, (unsigned long *)&cntr_mask, X86_PMC_IDX_MAX) {
 245		release_perfctr_nmi(x86_pmu_event_addr(i));
 246		release_evntsel_nmi(x86_pmu_config_addr(i));
 247	}
 248}
 249
 250#else
 251
 252static bool reserve_pmc_hardware(void) { return true; }
 253static void release_pmc_hardware(void) {}
 254
 255#endif
 256
 257bool check_hw_exists(struct pmu *pmu, unsigned long *cntr_mask,
 258		     unsigned long *fixed_cntr_mask)
 259{
 260	u64 val, val_fail = -1, val_new= ~0;
 261	int i, reg, reg_fail = -1, ret = 0;
 262	int bios_fail = 0;
 263	int reg_safe = -1;
 264
 265	/*
 266	 * Check to see if the BIOS enabled any of the counters, if so
 267	 * complain and bail.
 268	 */
 269	for_each_set_bit(i, cntr_mask, X86_PMC_IDX_MAX) {
 270		reg = x86_pmu_config_addr(i);
 271		ret = rdmsrl_safe(reg, &val);
 272		if (ret)
 273			goto msr_fail;
 274		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
 275			bios_fail = 1;
 276			val_fail = val;
 277			reg_fail = reg;
 278		} else {
 279			reg_safe = i;
 280		}
 281	}
 282
 283	if (*(u64 *)fixed_cntr_mask) {
 284		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
 285		ret = rdmsrl_safe(reg, &val);
 286		if (ret)
 287			goto msr_fail;
 288		for_each_set_bit(i, fixed_cntr_mask, X86_PMC_IDX_MAX) {
 289			if (fixed_counter_disabled(i, pmu))
 290				continue;
 291			if (val & (0x03ULL << i*4)) {
 292				bios_fail = 1;
 293				val_fail = val;
 294				reg_fail = reg;
 295			}
 296		}
 297	}
 298
 299	/*
 300	 * If all the counters are enabled, the below test will always
 301	 * fail.  The tools will also become useless in this scenario.
 302	 * Just fail and disable the hardware counters.
 303	 */
 304
 305	if (reg_safe == -1) {
 306		reg = reg_safe;
 307		goto msr_fail;
 308	}
 309
 310	/*
 311	 * Read the current value, change it and read it back to see if it
 312	 * matches, this is needed to detect certain hardware emulators
 313	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
 314	 */
 315	reg = x86_pmu_event_addr(reg_safe);
 316	if (rdmsrl_safe(reg, &val))
 317		goto msr_fail;
 318	val ^= 0xffffUL;
 319	ret = wrmsrl_safe(reg, val);
 320	ret |= rdmsrl_safe(reg, &val_new);
 321	if (ret || val != val_new)
 322		goto msr_fail;
 323
 324	/*
 325	 * We still allow the PMU driver to operate:
 326	 */
 327	if (bios_fail) {
 328		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
 329		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
 330			      reg_fail, val_fail);
 331	}
 332
 333	return true;
 334
 335msr_fail:
 336	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
 337		pr_cont("PMU not available due to virtualization, using software events only.\n");
 338	} else {
 339		pr_cont("Broken PMU hardware detected, using software events only.\n");
 340		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
 341		       reg, val_new);
 342	}
 343
 344	return false;
 345}
 346
 347static void hw_perf_event_destroy(struct perf_event *event)
 348{
 349	x86_release_hardware();
 350	atomic_dec(&active_events);
 351}
 352
 353void hw_perf_lbr_event_destroy(struct perf_event *event)
 354{
 355	hw_perf_event_destroy(event);
 356
 357	/* undo the lbr/bts event accounting */
 358	x86_del_exclusive(x86_lbr_exclusive_lbr);
 359}
 360
 361static inline int x86_pmu_initialized(void)
 362{
 363	return x86_pmu.handle_irq != NULL;
 364}
 365
 366static inline int
 367set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
 368{
 369	struct perf_event_attr *attr = &event->attr;
 370	unsigned int cache_type, cache_op, cache_result;
 371	u64 config, val;
 372
 373	config = attr->config;
 374
 375	cache_type = (config >> 0) & 0xff;
 376	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 377		return -EINVAL;
 378	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
 379
 380	cache_op = (config >>  8) & 0xff;
 381	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 382		return -EINVAL;
 383	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
 384
 385	cache_result = (config >> 16) & 0xff;
 386	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 387		return -EINVAL;
 388	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
 389
 390	val = hybrid_var(event->pmu, hw_cache_event_ids)[cache_type][cache_op][cache_result];
 
 391	if (val == 0)
 392		return -ENOENT;
 393
 394	if (val == -1)
 395		return -EINVAL;
 396
 397	hwc->config |= val;
 398	attr->config1 = hybrid_var(event->pmu, hw_cache_extra_regs)[cache_type][cache_op][cache_result];
 399	return x86_pmu_extra_regs(val, event);
 400}
 401
 402int x86_reserve_hardware(void)
 403{
 404	int err = 0;
 405
 406	if (!atomic_inc_not_zero(&pmc_refcount)) {
 407		mutex_lock(&pmc_reserve_mutex);
 408		if (atomic_read(&pmc_refcount) == 0) {
 409			if (!reserve_pmc_hardware()) {
 410				err = -EBUSY;
 411			} else {
 412				reserve_ds_buffers();
 413				reserve_lbr_buffers();
 414			}
 415		}
 416		if (!err)
 417			atomic_inc(&pmc_refcount);
 418		mutex_unlock(&pmc_reserve_mutex);
 419	}
 420
 421	return err;
 422}
 423
 424void x86_release_hardware(void)
 425{
 426	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
 427		release_pmc_hardware();
 428		release_ds_buffers();
 429		release_lbr_buffers();
 430		mutex_unlock(&pmc_reserve_mutex);
 431	}
 432}
 433
 434/*
 435 * Check if we can create event of a certain type (that no conflicting events
 436 * are present).
 437 */
 438int x86_add_exclusive(unsigned int what)
 439{
 440	int i;
 441
 442	/*
 443	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
 444	 * LBR and BTS are still mutually exclusive.
 445	 */
 446	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
 447		goto out;
 448
 449	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
 450		mutex_lock(&pmc_reserve_mutex);
 451		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
 452			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
 453				goto fail_unlock;
 454		}
 455		atomic_inc(&x86_pmu.lbr_exclusive[what]);
 456		mutex_unlock(&pmc_reserve_mutex);
 457	}
 458
 459out:
 460	atomic_inc(&active_events);
 461	return 0;
 462
 463fail_unlock:
 464	mutex_unlock(&pmc_reserve_mutex);
 465	return -EBUSY;
 466}
 467
 468void x86_del_exclusive(unsigned int what)
 469{
 470	atomic_dec(&active_events);
 471
 472	/*
 473	 * See the comment in x86_add_exclusive().
 474	 */
 475	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
 476		return;
 477
 478	atomic_dec(&x86_pmu.lbr_exclusive[what]);
 
 479}
 480
 481int x86_setup_perfctr(struct perf_event *event)
 482{
 483	struct perf_event_attr *attr = &event->attr;
 484	struct hw_perf_event *hwc = &event->hw;
 485	u64 config;
 486
 487	if (!is_sampling_event(event)) {
 488		hwc->sample_period = x86_pmu.max_period;
 489		hwc->last_period = hwc->sample_period;
 490		local64_set(&hwc->period_left, hwc->sample_period);
 491	}
 492
 493	if (attr->type == event->pmu->type)
 494		return x86_pmu_extra_regs(event->attr.config, event);
 495
 496	if (attr->type == PERF_TYPE_HW_CACHE)
 497		return set_ext_hw_attr(hwc, event);
 498
 499	if (attr->config >= x86_pmu.max_events)
 500		return -EINVAL;
 501
 502	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
 503
 504	/*
 505	 * The generic map:
 506	 */
 507	config = x86_pmu.event_map(attr->config);
 508
 509	if (config == 0)
 510		return -ENOENT;
 511
 512	if (config == -1LL)
 513		return -EINVAL;
 514
 515	hwc->config |= config;
 516
 517	return 0;
 518}
 519
 520/*
 521 * check that branch_sample_type is compatible with
 522 * settings needed for precise_ip > 1 which implies
 523 * using the LBR to capture ALL taken branches at the
 524 * priv levels of the measurement
 525 */
 526static inline int precise_br_compat(struct perf_event *event)
 527{
 528	u64 m = event->attr.branch_sample_type;
 529	u64 b = 0;
 530
 531	/* must capture all branches */
 532	if (!(m & PERF_SAMPLE_BRANCH_ANY))
 533		return 0;
 534
 535	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
 536
 537	if (!event->attr.exclude_user)
 538		b |= PERF_SAMPLE_BRANCH_USER;
 539
 540	if (!event->attr.exclude_kernel)
 541		b |= PERF_SAMPLE_BRANCH_KERNEL;
 542
 543	/*
 544	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
 545	 */
 546
 547	return m == b;
 548}
 549
 550int x86_pmu_max_precise(void)
 551{
 552	int precise = 0;
 553
 554	/* Support for constant skid */
 555	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
 556		precise++;
 557
 558		/* Support for IP fixup */
 559		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
 560			precise++;
 561
 562		if (x86_pmu.pebs_prec_dist)
 563			precise++;
 564	}
 565	return precise;
 566}
 567
 568int x86_pmu_hw_config(struct perf_event *event)
 569{
 570	if (event->attr.precise_ip) {
 571		int precise = x86_pmu_max_precise();
 572
 573		if (event->attr.precise_ip > precise)
 574			return -EOPNOTSUPP;
 575
 576		/* There's no sense in having PEBS for non sampling events: */
 577		if (!is_sampling_event(event))
 578			return -EINVAL;
 579	}
 580	/*
 581	 * check that PEBS LBR correction does not conflict with
 582	 * whatever the user is asking with attr->branch_sample_type
 583	 */
 584	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
 585		u64 *br_type = &event->attr.branch_sample_type;
 586
 587		if (has_branch_stack(event)) {
 588			if (!precise_br_compat(event))
 589				return -EOPNOTSUPP;
 590
 591			/* branch_sample_type is compatible */
 592
 593		} else {
 594			/*
 595			 * user did not specify  branch_sample_type
 596			 *
 597			 * For PEBS fixups, we capture all
 598			 * the branches at the priv level of the
 599			 * event.
 600			 */
 601			*br_type = PERF_SAMPLE_BRANCH_ANY;
 602
 603			if (!event->attr.exclude_user)
 604				*br_type |= PERF_SAMPLE_BRANCH_USER;
 605
 606			if (!event->attr.exclude_kernel)
 607				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
 608		}
 609	}
 610
 611	if (branch_sample_call_stack(event))
 612		event->attach_state |= PERF_ATTACH_TASK_DATA;
 613
 614	/*
 615	 * Generate PMC IRQs:
 616	 * (keep 'enabled' bit clear for now)
 617	 */
 618	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
 619
 620	/*
 621	 * Count user and OS events unless requested not to
 622	 */
 623	if (!event->attr.exclude_user)
 624		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
 625	if (!event->attr.exclude_kernel)
 626		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
 627
 628	if (event->attr.type == event->pmu->type)
 629		event->hw.config |= x86_pmu_get_event_config(event);
 630
 631	if (!event->attr.freq && x86_pmu.limit_period) {
 632		s64 left = event->attr.sample_period;
 633		x86_pmu.limit_period(event, &left);
 634		if (left > event->attr.sample_period)
 635			return -EINVAL;
 636	}
 637
 638	/* sample_regs_user never support XMM registers */
 639	if (unlikely(event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK))
 640		return -EINVAL;
 641	/*
 642	 * Besides the general purpose registers, XMM registers may
 643	 * be collected in PEBS on some platforms, e.g. Icelake
 644	 */
 645	if (unlikely(event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK)) {
 646		if (!(event->pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS))
 647			return -EINVAL;
 648
 649		if (!event->attr.precise_ip)
 650			return -EINVAL;
 651	}
 652
 653	return x86_setup_perfctr(event);
 654}
 655
 656/*
 657 * Setup the hardware configuration for a given attr_type
 658 */
 659static int __x86_pmu_event_init(struct perf_event *event)
 660{
 661	int err;
 662
 663	if (!x86_pmu_initialized())
 664		return -ENODEV;
 665
 666	err = x86_reserve_hardware();
 667	if (err)
 668		return err;
 669
 670	atomic_inc(&active_events);
 671	event->destroy = hw_perf_event_destroy;
 672
 673	event->hw.idx = -1;
 674	event->hw.last_cpu = -1;
 675	event->hw.last_tag = ~0ULL;
 676
 677	/* mark unused */
 678	event->hw.extra_reg.idx = EXTRA_REG_NONE;
 679	event->hw.branch_reg.idx = EXTRA_REG_NONE;
 680
 681	return x86_pmu.hw_config(event);
 682}
 683
 684void x86_pmu_disable_all(void)
 685{
 686	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 687	int idx;
 688
 689	for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
 690		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
 691		u64 val;
 692
 693		if (!test_bit(idx, cpuc->active_mask))
 694			continue;
 695		rdmsrl(x86_pmu_config_addr(idx), val);
 696		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
 697			continue;
 698		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
 699		wrmsrl(x86_pmu_config_addr(idx), val);
 700		if (is_counter_pair(hwc))
 701			wrmsrl(x86_pmu_config_addr(idx + 1), 0);
 702	}
 703}
 704
 705struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr, void *data)
 706{
 707	return static_call(x86_pmu_guest_get_msrs)(nr, data);
 708}
 709EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
 710
 711/*
 712 * There may be PMI landing after enabled=0. The PMI hitting could be before or
 713 * after disable_all.
 714 *
 715 * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
 716 * It will not be re-enabled in the NMI handler again, because enabled=0. After
 717 * handling the NMI, disable_all will be called, which will not change the
 718 * state either. If PMI hits after disable_all, the PMU is already disabled
 719 * before entering NMI handler. The NMI handler will not change the state
 720 * either.
 721 *
 722 * So either situation is harmless.
 723 */
 724static void x86_pmu_disable(struct pmu *pmu)
 725{
 726	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 727
 728	if (!x86_pmu_initialized())
 729		return;
 730
 731	if (!cpuc->enabled)
 732		return;
 733
 734	cpuc->n_added = 0;
 735	cpuc->enabled = 0;
 736	barrier();
 737
 738	static_call(x86_pmu_disable_all)();
 739}
 740
 741void x86_pmu_enable_all(int added)
 742{
 743	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 744	int idx;
 745
 746	for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
 747		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
 748
 749		if (!test_bit(idx, cpuc->active_mask))
 750			continue;
 751
 752		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
 753	}
 754}
 755
 
 
 756static inline int is_x86_event(struct perf_event *event)
 757{
 758	int i;
 759
 760	if (!is_hybrid())
 761		return event->pmu == &pmu;
 762
 763	for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
 764		if (event->pmu == &x86_pmu.hybrid_pmu[i].pmu)
 765			return true;
 766	}
 767
 768	return false;
 769}
 770
 771struct pmu *x86_get_pmu(unsigned int cpu)
 772{
 773	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
 774
 775	/*
 776	 * All CPUs of the hybrid type have been offline.
 777	 * The x86_get_pmu() should not be invoked.
 778	 */
 779	if (WARN_ON_ONCE(!cpuc->pmu))
 780		return &pmu;
 781
 782	return cpuc->pmu;
 783}
 784/*
 785 * Event scheduler state:
 786 *
 787 * Assign events iterating over all events and counters, beginning
 788 * with events with least weights first. Keep the current iterator
 789 * state in struct sched_state.
 790 */
 791struct sched_state {
 792	int	weight;
 793	int	event;		/* event index */
 794	int	counter;	/* counter index */
 795	int	unassigned;	/* number of events to be assigned left */
 796	int	nr_gp;		/* number of GP counters used */
 797	u64	used;
 798};
 799
 800/* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
 801#define	SCHED_STATES_MAX	2
 802
 803struct perf_sched {
 804	int			max_weight;
 805	int			max_events;
 806	int			max_gp;
 807	int			saved_states;
 808	struct event_constraint	**constraints;
 809	struct sched_state	state;
 810	struct sched_state	saved[SCHED_STATES_MAX];
 811};
 812
 813/*
 814 * Initialize iterator that runs through all events and counters.
 815 */
 816static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
 817			    int num, int wmin, int wmax, int gpmax)
 818{
 819	int idx;
 820
 821	memset(sched, 0, sizeof(*sched));
 822	sched->max_events	= num;
 823	sched->max_weight	= wmax;
 824	sched->max_gp		= gpmax;
 825	sched->constraints	= constraints;
 826
 827	for (idx = 0; idx < num; idx++) {
 828		if (constraints[idx]->weight == wmin)
 829			break;
 830	}
 831
 832	sched->state.event	= idx;		/* start with min weight */
 833	sched->state.weight	= wmin;
 834	sched->state.unassigned	= num;
 835}
 836
 837static void perf_sched_save_state(struct perf_sched *sched)
 838{
 839	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
 840		return;
 841
 842	sched->saved[sched->saved_states] = sched->state;
 843	sched->saved_states++;
 844}
 845
 846static bool perf_sched_restore_state(struct perf_sched *sched)
 847{
 848	if (!sched->saved_states)
 849		return false;
 850
 851	sched->saved_states--;
 852	sched->state = sched->saved[sched->saved_states];
 853
 854	/* this assignment didn't work out */
 855	/* XXX broken vs EVENT_PAIR */
 856	sched->state.used &= ~BIT_ULL(sched->state.counter);
 857
 858	/* try the next one */
 859	sched->state.counter++;
 860
 861	return true;
 862}
 863
 864/*
 865 * Select a counter for the current event to schedule. Return true on
 866 * success.
 867 */
 868static bool __perf_sched_find_counter(struct perf_sched *sched)
 869{
 870	struct event_constraint *c;
 871	int idx;
 872
 873	if (!sched->state.unassigned)
 874		return false;
 875
 876	if (sched->state.event >= sched->max_events)
 877		return false;
 878
 879	c = sched->constraints[sched->state.event];
 880	/* Prefer fixed purpose counters */
 881	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
 882		idx = INTEL_PMC_IDX_FIXED;
 883		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
 884			u64 mask = BIT_ULL(idx);
 885
 886			if (sched->state.used & mask)
 887				continue;
 888
 889			sched->state.used |= mask;
 890			goto done;
 891		}
 892	}
 893
 894	/* Grab the first unused counter starting with idx */
 895	idx = sched->state.counter;
 896	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
 897		u64 mask = BIT_ULL(idx);
 
 
 898
 899		if (c->flags & PERF_X86_EVENT_PAIR)
 900			mask |= mask << 1;
 901
 902		if (sched->state.used & mask)
 903			continue;
 904
 905		if (sched->state.nr_gp++ >= sched->max_gp)
 906			return false;
 907
 908		sched->state.used |= mask;
 909		goto done;
 910	}
 911
 912	return false;
 913
 914done:
 915	sched->state.counter = idx;
 916
 917	if (c->overlap)
 918		perf_sched_save_state(sched);
 919
 920	return true;
 921}
 922
 923static bool perf_sched_find_counter(struct perf_sched *sched)
 924{
 925	while (!__perf_sched_find_counter(sched)) {
 926		if (!perf_sched_restore_state(sched))
 927			return false;
 928	}
 929
 930	return true;
 931}
 932
 933/*
 934 * Go through all unassigned events and find the next one to schedule.
 935 * Take events with the least weight first. Return true on success.
 936 */
 937static bool perf_sched_next_event(struct perf_sched *sched)
 938{
 939	struct event_constraint *c;
 940
 941	if (!sched->state.unassigned || !--sched->state.unassigned)
 942		return false;
 943
 944	do {
 945		/* next event */
 946		sched->state.event++;
 947		if (sched->state.event >= sched->max_events) {
 948			/* next weight */
 949			sched->state.event = 0;
 950			sched->state.weight++;
 951			if (sched->state.weight > sched->max_weight)
 952				return false;
 953		}
 954		c = sched->constraints[sched->state.event];
 955	} while (c->weight != sched->state.weight);
 956
 957	sched->state.counter = 0;	/* start with first counter */
 958
 959	return true;
 960}
 961
 962/*
 963 * Assign a counter for each event.
 964 */
 965int perf_assign_events(struct event_constraint **constraints, int n,
 966			int wmin, int wmax, int gpmax, int *assign)
 967{
 968	struct perf_sched sched;
 969
 970	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
 971
 972	do {
 973		if (!perf_sched_find_counter(&sched))
 974			break;	/* failed */
 975		if (assign)
 976			assign[sched.state.event] = sched.state.counter;
 977	} while (perf_sched_next_event(&sched));
 978
 979	return sched.state.unassigned;
 980}
 981EXPORT_SYMBOL_GPL(perf_assign_events);
 982
 983int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 984{
 985	struct event_constraint *c;
 
 986	struct perf_event *e;
 987	int n0, i, wmin, wmax, unsched = 0;
 988	struct hw_perf_event *hwc;
 989	u64 used_mask = 0;
 
 990
 991	/*
 992	 * Compute the number of events already present; see x86_pmu_add(),
 993	 * validate_group() and x86_pmu_commit_txn(). For the former two
 994	 * cpuc->n_events hasn't been updated yet, while for the latter
 995	 * cpuc->n_txn contains the number of events added in the current
 996	 * transaction.
 997	 */
 998	n0 = cpuc->n_events;
 999	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1000		n0 -= cpuc->n_txn;
1001
1002	static_call_cond(x86_pmu_start_scheduling)(cpuc);
 
1003
1004	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
1005		c = cpuc->event_constraint[i];
1006
1007		/*
1008		 * Previously scheduled events should have a cached constraint,
1009		 * while new events should not have one.
1010		 */
1011		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
1012
1013		/*
1014		 * Request constraints for new events; or for those events that
1015		 * have a dynamic constraint -- for those the constraint can
1016		 * change due to external factors (sibling state, allow_tfa).
1017		 */
1018		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
1019			c = static_call(x86_pmu_get_event_constraints)(cpuc, i, cpuc->event_list[i]);
1020			cpuc->event_constraint[i] = c;
1021		}
1022
1023		wmin = min(wmin, c->weight);
1024		wmax = max(wmax, c->weight);
1025	}
1026
1027	/*
1028	 * fastpath, try to reuse previous register
1029	 */
1030	for (i = 0; i < n; i++) {
1031		u64 mask;
1032
1033		hwc = &cpuc->event_list[i]->hw;
1034		c = cpuc->event_constraint[i];
1035
1036		/* never assigned */
1037		if (hwc->idx == -1)
1038			break;
1039
1040		/* constraint still honored */
1041		if (!test_bit(hwc->idx, c->idxmsk))
1042			break;
1043
1044		mask = BIT_ULL(hwc->idx);
1045		if (is_counter_pair(hwc))
1046			mask |= mask << 1;
1047
1048		/* not already used */
1049		if (used_mask & mask)
1050			break;
1051
1052		used_mask |= mask;
1053
1054		if (assign)
1055			assign[i] = hwc->idx;
1056	}
1057
1058	/* slow path */
1059	if (i != n) {
1060		int gpmax = x86_pmu_max_num_counters(cpuc->pmu);
1061
1062		/*
1063		 * Do not allow scheduling of more than half the available
1064		 * generic counters.
1065		 *
1066		 * This helps avoid counter starvation of sibling thread by
1067		 * ensuring at most half the counters cannot be in exclusive
1068		 * mode. There is no designated counters for the limits. Any
1069		 * N/2 counters can be used. This helps with events with
1070		 * specific counter constraints.
1071		 */
1072		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
1073		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
1074			gpmax /= 2;
1075
1076		/*
1077		 * Reduce the amount of available counters to allow fitting
1078		 * the extra Merge events needed by large increment events.
1079		 */
1080		if (x86_pmu.flags & PMU_FL_PAIR) {
1081			gpmax -= cpuc->n_pair;
1082			WARN_ON(gpmax <= 0);
1083		}
1084
1085		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
1086					     wmax, gpmax, assign);
1087	}
1088
1089	/*
1090	 * In case of success (unsched = 0), mark events as committed,
1091	 * so we do not put_constraint() in case new events are added
1092	 * and fail to be scheduled
1093	 *
1094	 * We invoke the lower level commit callback to lock the resource
1095	 *
1096	 * We do not need to do all of this in case we are called to
1097	 * validate an event group (assign == NULL)
1098	 */
1099	if (!unsched && assign) {
1100		for (i = 0; i < n; i++)
1101			static_call_cond(x86_pmu_commit_scheduling)(cpuc, i, assign[i]);
 
 
 
1102	} else {
1103		for (i = n0; i < n; i++) {
1104			e = cpuc->event_list[i];
1105
1106			/*
1107			 * release events that failed scheduling
1108			 */
1109			static_call_cond(x86_pmu_put_event_constraints)(cpuc, e);
 
1110
1111			cpuc->event_constraint[i] = NULL;
1112		}
1113	}
1114
1115	static_call_cond(x86_pmu_stop_scheduling)(cpuc);
 
1116
1117	return unsched ? -EINVAL : 0;
1118}
1119
1120static int add_nr_metric_event(struct cpu_hw_events *cpuc,
1121			       struct perf_event *event)
1122{
1123	if (is_metric_event(event)) {
1124		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
1125			return -EINVAL;
1126		cpuc->n_metric++;
1127		cpuc->n_txn_metric++;
1128	}
1129
1130	return 0;
1131}
1132
1133static void del_nr_metric_event(struct cpu_hw_events *cpuc,
1134				struct perf_event *event)
1135{
1136	if (is_metric_event(event))
1137		cpuc->n_metric--;
1138}
1139
1140static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
1141			 int max_count, int n)
1142{
1143	union perf_capabilities intel_cap = hybrid(cpuc->pmu, intel_cap);
1144
1145	if (intel_cap.perf_metrics && add_nr_metric_event(cpuc, event))
1146		return -EINVAL;
1147
1148	if (n >= max_count + cpuc->n_metric)
1149		return -EINVAL;
1150
1151	cpuc->event_list[n] = event;
1152	if (is_counter_pair(&event->hw)) {
1153		cpuc->n_pair++;
1154		cpuc->n_txn_pair++;
1155	}
1156
1157	return 0;
1158}
1159
1160/*
1161 * dogrp: true if must collect siblings events (group)
1162 * returns total number of events and error code
1163 */
1164static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1165{
1166	struct perf_event *event;
1167	int n, max_count;
1168
1169	max_count = x86_pmu_num_counters(cpuc->pmu) + x86_pmu_num_counters_fixed(cpuc->pmu);
1170
1171	/* current number of events already accepted */
1172	n = cpuc->n_events;
1173	if (!cpuc->n_events)
1174		cpuc->pebs_output = 0;
1175
1176	if (!cpuc->is_fake && leader->attr.precise_ip) {
1177		/*
1178		 * For PEBS->PT, if !aux_event, the group leader (PT) went
1179		 * away, the group was broken down and this singleton event
1180		 * can't schedule any more.
1181		 */
1182		if (is_pebs_pt(leader) && !leader->aux_event)
1183			return -EINVAL;
1184
1185		/*
1186		 * pebs_output: 0: no PEBS so far, 1: PT, 2: DS
1187		 */
1188		if (cpuc->pebs_output &&
1189		    cpuc->pebs_output != is_pebs_pt(leader) + 1)
1190			return -EINVAL;
1191
1192		cpuc->pebs_output = is_pebs_pt(leader) + 1;
1193	}
1194
1195	if (is_x86_event(leader)) {
1196		if (collect_event(cpuc, leader, max_count, n))
1197			return -EINVAL;
 
1198		n++;
1199	}
1200
1201	if (!dogrp)
1202		return n;
1203
1204	for_each_sibling_event(event, leader) {
1205		if (!is_x86_event(event) || event->state <= PERF_EVENT_STATE_OFF)
 
1206			continue;
1207
1208		if (collect_event(cpuc, event, max_count, n))
1209			return -EINVAL;
1210
 
1211		n++;
1212	}
1213	return n;
1214}
1215
1216static inline void x86_assign_hw_event(struct perf_event *event,
1217				struct cpu_hw_events *cpuc, int i)
1218{
1219	struct hw_perf_event *hwc = &event->hw;
1220	int idx;
1221
1222	idx = hwc->idx = cpuc->assign[i];
1223	hwc->last_cpu = smp_processor_id();
1224	hwc->last_tag = ++cpuc->tags[i];
1225
1226	static_call_cond(x86_pmu_assign)(event, idx);
1227
1228	switch (hwc->idx) {
1229	case INTEL_PMC_IDX_FIXED_BTS:
1230	case INTEL_PMC_IDX_FIXED_VLBR:
1231		hwc->config_base = 0;
1232		hwc->event_base	= 0;
1233		break;
1234
1235	case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
1236		/* All the metric events are mapped onto the fixed counter 3. */
1237		idx = INTEL_PMC_IDX_FIXED_SLOTS;
1238		fallthrough;
1239	case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS-1:
1240		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1241		hwc->event_base = x86_pmu_fixed_ctr_addr(idx - INTEL_PMC_IDX_FIXED);
1242		hwc->event_base_rdpmc = (idx - INTEL_PMC_IDX_FIXED) |
1243					INTEL_PMC_FIXED_RDPMC_BASE;
1244		break;
1245
1246	default:
1247		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1248		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1249		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1250		break;
1251	}
1252}
1253
1254/**
1255 * x86_perf_rdpmc_index - Return PMC counter used for event
1256 * @event: the perf_event to which the PMC counter was assigned
1257 *
1258 * The counter assigned to this performance event may change if interrupts
1259 * are enabled. This counter should thus never be used while interrupts are
1260 * enabled. Before this function is used to obtain the assigned counter the
1261 * event should be checked for validity using, for example,
1262 * perf_event_read_local(), within the same interrupt disabled section in
1263 * which this counter is planned to be used.
1264 *
1265 * Return: The index of the performance monitoring counter assigned to
1266 * @perf_event.
1267 */
1268int x86_perf_rdpmc_index(struct perf_event *event)
1269{
1270	lockdep_assert_irqs_disabled();
1271
1272	return event->hw.event_base_rdpmc;
1273}
1274
1275static inline int match_prev_assignment(struct hw_perf_event *hwc,
1276					struct cpu_hw_events *cpuc,
1277					int i)
1278{
1279	return hwc->idx == cpuc->assign[i] &&
1280		hwc->last_cpu == smp_processor_id() &&
1281		hwc->last_tag == cpuc->tags[i];
1282}
1283
1284static void x86_pmu_start(struct perf_event *event, int flags);
1285
1286static void x86_pmu_enable(struct pmu *pmu)
1287{
1288	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1289	struct perf_event *event;
1290	struct hw_perf_event *hwc;
1291	int i, added = cpuc->n_added;
1292
1293	if (!x86_pmu_initialized())
1294		return;
1295
1296	if (cpuc->enabled)
1297		return;
1298
1299	if (cpuc->n_added) {
1300		int n_running = cpuc->n_events - cpuc->n_added;
1301		/*
1302		 * apply assignment obtained either from
1303		 * hw_perf_group_sched_in() or x86_pmu_enable()
1304		 *
1305		 * step1: save events moving to new counters
1306		 */
1307		for (i = 0; i < n_running; i++) {
1308			event = cpuc->event_list[i];
1309			hwc = &event->hw;
1310
1311			/*
1312			 * we can avoid reprogramming counter if:
1313			 * - assigned same counter as last time
1314			 * - running on same CPU as last time
1315			 * - no other event has used the counter since
1316			 */
1317			if (hwc->idx == -1 ||
1318			    match_prev_assignment(hwc, cpuc, i))
1319				continue;
1320
1321			/*
1322			 * Ensure we don't accidentally enable a stopped
1323			 * counter simply because we rescheduled.
1324			 */
1325			if (hwc->state & PERF_HES_STOPPED)
1326				hwc->state |= PERF_HES_ARCH;
1327
1328			x86_pmu_stop(event, PERF_EF_UPDATE);
1329		}
1330
1331		/*
1332		 * step2: reprogram moved events into new counters
1333		 */
1334		for (i = 0; i < cpuc->n_events; i++) {
1335			event = cpuc->event_list[i];
1336			hwc = &event->hw;
1337
1338			if (!match_prev_assignment(hwc, cpuc, i))
1339				x86_assign_hw_event(event, cpuc, i);
1340			else if (i < n_running)
1341				continue;
1342
1343			if (hwc->state & PERF_HES_ARCH)
1344				continue;
1345
1346			/*
1347			 * if cpuc->enabled = 0, then no wrmsr as
1348			 * per x86_pmu_enable_event()
1349			 */
1350			x86_pmu_start(event, PERF_EF_RELOAD);
1351		}
1352		cpuc->n_added = 0;
1353		perf_events_lapic_init();
1354	}
1355
1356	cpuc->enabled = 1;
1357	barrier();
1358
1359	static_call(x86_pmu_enable_all)(added);
1360}
1361
1362DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1363
1364/*
1365 * Set the next IRQ period, based on the hwc->period_left value.
1366 * To be called with the event disabled in hw:
1367 */
1368int x86_perf_event_set_period(struct perf_event *event)
1369{
1370	struct hw_perf_event *hwc = &event->hw;
1371	s64 left = local64_read(&hwc->period_left);
1372	s64 period = hwc->sample_period;
1373	int ret = 0, idx = hwc->idx;
1374
1375	if (unlikely(!hwc->event_base))
1376		return 0;
1377
1378	/*
1379	 * If we are way outside a reasonable range then just skip forward:
1380	 */
1381	if (unlikely(left <= -period)) {
1382		left = period;
1383		local64_set(&hwc->period_left, left);
1384		hwc->last_period = period;
1385		ret = 1;
1386	}
1387
1388	if (unlikely(left <= 0)) {
1389		left += period;
1390		local64_set(&hwc->period_left, left);
1391		hwc->last_period = period;
1392		ret = 1;
1393	}
1394	/*
1395	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1396	 */
1397	if (unlikely(left < 2))
1398		left = 2;
1399
1400	if (left > x86_pmu.max_period)
1401		left = x86_pmu.max_period;
1402
1403	static_call_cond(x86_pmu_limit_period)(event, &left);
 
1404
1405	this_cpu_write(pmc_prev_left[idx], left);
1406
1407	/*
1408	 * The hw event starts counting from this event offset,
1409	 * mark it to be able to extra future deltas:
1410	 */
1411	local64_set(&hwc->prev_count, (u64)-left);
1412
1413	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1414
1415	/*
1416	 * Sign extend the Merge event counter's upper 16 bits since
1417	 * we currently declare a 48-bit counter width
1418	 */
1419	if (is_counter_pair(hwc))
1420		wrmsrl(x86_pmu_event_addr(idx + 1), 0xffff);
 
 
 
1421
1422	perf_event_update_userpage(event);
1423
1424	return ret;
1425}
1426
1427void x86_pmu_enable_event(struct perf_event *event)
1428{
1429	if (__this_cpu_read(cpu_hw_events.enabled))
1430		__x86_pmu_enable_event(&event->hw,
1431				       ARCH_PERFMON_EVENTSEL_ENABLE);
1432}
1433
1434/*
1435 * Add a single event to the PMU.
1436 *
1437 * The event is added to the group of enabled events
1438 * but only if it can be scheduled with existing events.
1439 */
1440static int x86_pmu_add(struct perf_event *event, int flags)
1441{
1442	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1443	struct hw_perf_event *hwc;
1444	int assign[X86_PMC_IDX_MAX];
1445	int n, n0, ret;
1446
1447	hwc = &event->hw;
1448
1449	n0 = cpuc->n_events;
1450	ret = n = collect_events(cpuc, event, false);
1451	if (ret < 0)
1452		goto out;
1453
1454	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1455	if (!(flags & PERF_EF_START))
1456		hwc->state |= PERF_HES_ARCH;
1457
1458	/*
1459	 * If group events scheduling transaction was started,
1460	 * skip the schedulability test here, it will be performed
1461	 * at commit time (->commit_txn) as a whole.
1462	 *
1463	 * If commit fails, we'll call ->del() on all events
1464	 * for which ->add() was called.
1465	 */
1466	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1467		goto done_collect;
1468
1469	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
1470	if (ret)
1471		goto out;
1472	/*
1473	 * copy new assignment, now we know it is possible
1474	 * will be used by hw_perf_enable()
1475	 */
1476	memcpy(cpuc->assign, assign, n*sizeof(int));
1477
1478done_collect:
1479	/*
1480	 * Commit the collect_events() state. See x86_pmu_del() and
1481	 * x86_pmu_*_txn().
1482	 */
1483	cpuc->n_events = n;
1484	cpuc->n_added += n - n0;
1485	cpuc->n_txn += n - n0;
1486
1487	/*
1488	 * This is before x86_pmu_enable() will call x86_pmu_start(),
1489	 * so we enable LBRs before an event needs them etc..
1490	 */
1491	static_call_cond(x86_pmu_add)(event);
 
 
1492
1493	ret = 0;
1494out:
1495	return ret;
1496}
1497
1498static void x86_pmu_start(struct perf_event *event, int flags)
1499{
1500	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1501	int idx = event->hw.idx;
1502
1503	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1504		return;
1505
1506	if (WARN_ON_ONCE(idx == -1))
1507		return;
1508
1509	if (flags & PERF_EF_RELOAD) {
1510		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1511		static_call(x86_pmu_set_period)(event);
1512	}
1513
1514	event->hw.state = 0;
1515
1516	cpuc->events[idx] = event;
1517	__set_bit(idx, cpuc->active_mask);
1518	static_call(x86_pmu_enable)(event);
 
1519	perf_event_update_userpage(event);
1520}
1521
1522void perf_event_print_debug(void)
1523{
1524	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1525	unsigned long *cntr_mask, *fixed_cntr_mask;
1526	struct event_constraint *pebs_constraints;
1527	struct cpu_hw_events *cpuc;
1528	u64 pebs, debugctl;
1529	int cpu, idx;
1530
1531	guard(irqsave)();
 
 
 
1532
1533	cpu = smp_processor_id();
1534	cpuc = &per_cpu(cpu_hw_events, cpu);
1535	cntr_mask = hybrid(cpuc->pmu, cntr_mask);
1536	fixed_cntr_mask = hybrid(cpuc->pmu, fixed_cntr_mask);
1537	pebs_constraints = hybrid(cpuc->pmu, pebs_constraints);
1538
1539	if (!*(u64 *)cntr_mask)
1540		return;
1541
1542	if (x86_pmu.version >= 2) {
1543		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1544		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1545		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1546		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1547
1548		pr_info("\n");
1549		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1550		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1551		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1552		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1553		if (pebs_constraints) {
1554			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1555			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1556		}
1557		if (x86_pmu.lbr_nr) {
1558			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1559			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1560		}
1561	}
1562	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1563
1564	for_each_set_bit(idx, cntr_mask, X86_PMC_IDX_MAX) {
1565		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1566		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1567
1568		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1569
1570		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1571			cpu, idx, pmc_ctrl);
1572		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1573			cpu, idx, pmc_count);
1574		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1575			cpu, idx, prev_left);
1576	}
1577	for_each_set_bit(idx, fixed_cntr_mask, X86_PMC_IDX_MAX) {
1578		if (fixed_counter_disabled(idx, cpuc->pmu))
1579			continue;
1580		rdmsrl(x86_pmu_fixed_ctr_addr(idx), pmc_count);
1581
1582		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1583			cpu, idx, pmc_count);
1584	}
 
1585}
1586
1587void x86_pmu_stop(struct perf_event *event, int flags)
1588{
1589	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1590	struct hw_perf_event *hwc = &event->hw;
1591
1592	if (test_bit(hwc->idx, cpuc->active_mask)) {
1593		static_call(x86_pmu_disable)(event);
1594		__clear_bit(hwc->idx, cpuc->active_mask);
1595		cpuc->events[hwc->idx] = NULL;
1596		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1597		hwc->state |= PERF_HES_STOPPED;
1598	}
1599
1600	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1601		/*
1602		 * Drain the remaining delta count out of a event
1603		 * that we are disabling:
1604		 */
1605		static_call(x86_pmu_update)(event);
1606		hwc->state |= PERF_HES_UPTODATE;
1607	}
1608}
1609
1610static void x86_pmu_del(struct perf_event *event, int flags)
1611{
1612	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1613	union perf_capabilities intel_cap = hybrid(cpuc->pmu, intel_cap);
1614	int i;
1615
1616	/*
1617	 * If we're called during a txn, we only need to undo x86_pmu.add.
1618	 * The events never got scheduled and ->cancel_txn will truncate
1619	 * the event_list.
1620	 *
1621	 * XXX assumes any ->del() called during a TXN will only be on
1622	 * an event added during that same TXN.
1623	 */
1624	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1625		goto do_del;
1626
1627	__set_bit(event->hw.idx, cpuc->dirty);
1628
1629	/*
1630	 * Not a TXN, therefore cleanup properly.
1631	 */
1632	x86_pmu_stop(event, PERF_EF_UPDATE);
1633
1634	for (i = 0; i < cpuc->n_events; i++) {
1635		if (event == cpuc->event_list[i])
1636			break;
1637	}
1638
1639	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1640		return;
1641
1642	/* If we have a newly added event; make sure to decrease n_added. */
1643	if (i >= cpuc->n_events - cpuc->n_added)
1644		--cpuc->n_added;
1645
1646	static_call_cond(x86_pmu_put_event_constraints)(cpuc, event);
 
1647
1648	/* Delete the array entry. */
1649	while (++i < cpuc->n_events) {
1650		cpuc->event_list[i-1] = cpuc->event_list[i];
1651		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1652		cpuc->assign[i-1] = cpuc->assign[i];
1653	}
1654	cpuc->event_constraint[i-1] = NULL;
1655	--cpuc->n_events;
1656	if (intel_cap.perf_metrics)
1657		del_nr_metric_event(cpuc, event);
1658
1659	perf_event_update_userpage(event);
1660
1661do_del:
1662
1663	/*
1664	 * This is after x86_pmu_stop(); so we disable LBRs after any
1665	 * event can need them etc..
1666	 */
1667	static_call_cond(x86_pmu_del)(event);
 
1668}
1669
1670int x86_pmu_handle_irq(struct pt_regs *regs)
1671{
1672	struct perf_sample_data data;
1673	struct cpu_hw_events *cpuc;
1674	struct perf_event *event;
1675	int idx, handled = 0;
1676	u64 val;
1677
1678	cpuc = this_cpu_ptr(&cpu_hw_events);
1679
1680	/*
1681	 * Some chipsets need to unmask the LVTPC in a particular spot
1682	 * inside the nmi handler.  As a result, the unmasking was pushed
1683	 * into all the nmi handlers.
1684	 *
1685	 * This generic handler doesn't seem to have any issues where the
1686	 * unmasking occurs so it was left at the top.
1687	 */
1688	apic_write(APIC_LVTPC, APIC_DM_NMI);
1689
1690	for_each_set_bit(idx, x86_pmu.cntr_mask, X86_PMC_IDX_MAX) {
1691		if (!test_bit(idx, cpuc->active_mask))
1692			continue;
1693
1694		event = cpuc->events[idx];
1695
1696		val = static_call(x86_pmu_update)(event);
1697		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1698			continue;
1699
1700		/*
1701		 * event overflow
1702		 */
1703		handled++;
 
1704
1705		if (!static_call(x86_pmu_set_period)(event))
1706			continue;
1707
1708		perf_sample_data_init(&data, 0, event->hw.last_period);
1709
1710		if (has_branch_stack(event))
1711			perf_sample_save_brstack(&data, event, &cpuc->lbr_stack, NULL);
1712
1713		if (perf_event_overflow(event, &data, regs))
1714			x86_pmu_stop(event, 0);
1715	}
1716
1717	if (handled)
1718		inc_irq_stat(apic_perf_irqs);
1719
1720	return handled;
1721}
1722
1723void perf_events_lapic_init(void)
1724{
1725	if (!x86_pmu.apic || !x86_pmu_initialized())
1726		return;
1727
1728	/*
1729	 * Always use NMI for PMU
1730	 */
1731	apic_write(APIC_LVTPC, APIC_DM_NMI);
1732}
1733
1734static int
1735perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1736{
1737	u64 start_clock;
1738	u64 finish_clock;
1739	int ret;
1740
1741	/*
1742	 * All PMUs/events that share this PMI handler should make sure to
1743	 * increment active_events for their events.
1744	 */
1745	if (!atomic_read(&active_events))
1746		return NMI_DONE;
1747
1748	start_clock = sched_clock();
1749	ret = static_call(x86_pmu_handle_irq)(regs);
1750	finish_clock = sched_clock();
1751
1752	perf_sample_event_took(finish_clock - start_clock);
1753
1754	return ret;
1755}
1756NOKPROBE_SYMBOL(perf_event_nmi_handler);
1757
1758struct event_constraint emptyconstraint;
1759struct event_constraint unconstrained;
1760
1761static int x86_pmu_prepare_cpu(unsigned int cpu)
1762{
1763	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1764	int i;
1765
1766	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1767		cpuc->kfree_on_online[i] = NULL;
1768	if (x86_pmu.cpu_prepare)
1769		return x86_pmu.cpu_prepare(cpu);
1770	return 0;
1771}
1772
1773static int x86_pmu_dead_cpu(unsigned int cpu)
1774{
1775	if (x86_pmu.cpu_dead)
1776		x86_pmu.cpu_dead(cpu);
1777	return 0;
1778}
1779
1780static int x86_pmu_online_cpu(unsigned int cpu)
1781{
1782	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1783	int i;
1784
1785	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1786		kfree(cpuc->kfree_on_online[i]);
1787		cpuc->kfree_on_online[i] = NULL;
1788	}
1789	return 0;
1790}
1791
1792static int x86_pmu_starting_cpu(unsigned int cpu)
1793{
1794	if (x86_pmu.cpu_starting)
1795		x86_pmu.cpu_starting(cpu);
1796	return 0;
1797}
1798
1799static int x86_pmu_dying_cpu(unsigned int cpu)
1800{
1801	if (x86_pmu.cpu_dying)
1802		x86_pmu.cpu_dying(cpu);
1803	return 0;
1804}
1805
1806static void __init pmu_check_apic(void)
1807{
1808	if (boot_cpu_has(X86_FEATURE_APIC))
1809		return;
1810
1811	x86_pmu.apic = 0;
1812	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1813	pr_info("no hardware sampling interrupt available.\n");
1814
1815	/*
1816	 * If we have a PMU initialized but no APIC
1817	 * interrupts, we cannot sample hardware
1818	 * events (user-space has to fall back and
1819	 * sample via a hrtimer based software event):
1820	 */
1821	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1822
1823}
1824
1825static struct attribute_group x86_pmu_format_group __ro_after_init = {
1826	.name = "format",
1827	.attrs = NULL,
1828};
1829
1830ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1831{
1832	struct perf_pmu_events_attr *pmu_attr =
1833		container_of(attr, struct perf_pmu_events_attr, attr);
1834	u64 config = 0;
1835
1836	if (pmu_attr->id < x86_pmu.max_events)
1837		config = x86_pmu.event_map(pmu_attr->id);
1838
1839	/* string trumps id */
1840	if (pmu_attr->event_str)
1841		return sprintf(page, "%s\n", pmu_attr->event_str);
1842
1843	return x86_pmu.events_sysfs_show(page, config);
1844}
1845EXPORT_SYMBOL_GPL(events_sysfs_show);
1846
1847ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1848			  char *page)
1849{
1850	struct perf_pmu_events_ht_attr *pmu_attr =
1851		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1852
1853	/*
1854	 * Report conditional events depending on Hyper-Threading.
1855	 *
1856	 * This is overly conservative as usually the HT special
1857	 * handling is not needed if the other CPU thread is idle.
1858	 *
1859	 * Note this does not (and cannot) handle the case when thread
1860	 * siblings are invisible, for example with virtualization
1861	 * if they are owned by some other guest.  The user tool
1862	 * has to re-read when a thread sibling gets onlined later.
1863	 */
1864	return sprintf(page, "%s",
1865			topology_max_smt_threads() > 1 ?
1866			pmu_attr->event_str_ht :
1867			pmu_attr->event_str_noht);
1868}
1869
1870ssize_t events_hybrid_sysfs_show(struct device *dev,
1871				 struct device_attribute *attr,
1872				 char *page)
1873{
1874	struct perf_pmu_events_hybrid_attr *pmu_attr =
1875		container_of(attr, struct perf_pmu_events_hybrid_attr, attr);
1876	struct x86_hybrid_pmu *pmu;
1877	const char *str, *next_str;
1878	int i;
1879
1880	if (hweight64(pmu_attr->pmu_type) == 1)
1881		return sprintf(page, "%s", pmu_attr->event_str);
1882
1883	/*
1884	 * Hybrid PMUs may support the same event name, but with different
1885	 * event encoding, e.g., the mem-loads event on an Atom PMU has
1886	 * different event encoding from a Core PMU.
1887	 *
1888	 * The event_str includes all event encodings. Each event encoding
1889	 * is divided by ";". The order of the event encodings must follow
1890	 * the order of the hybrid PMU index.
1891	 */
1892	pmu = container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu);
1893
1894	str = pmu_attr->event_str;
1895	for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
1896		if (!(x86_pmu.hybrid_pmu[i].pmu_type & pmu_attr->pmu_type))
1897			continue;
1898		if (x86_pmu.hybrid_pmu[i].pmu_type & pmu->pmu_type) {
1899			next_str = strchr(str, ';');
1900			if (next_str)
1901				return snprintf(page, next_str - str + 1, "%s", str);
1902			else
1903				return sprintf(page, "%s", str);
1904		}
1905		str = strchr(str, ';');
1906		str++;
1907	}
1908
1909	return 0;
1910}
1911EXPORT_SYMBOL_GPL(events_hybrid_sysfs_show);
1912
1913EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1914EVENT_ATTR(instructions,		INSTRUCTIONS		);
1915EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1916EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1917EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1918EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1919EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1920EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1921EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1922EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1923
1924static struct attribute *empty_attrs;
1925
1926static struct attribute *events_attr[] = {
1927	EVENT_PTR(CPU_CYCLES),
1928	EVENT_PTR(INSTRUCTIONS),
1929	EVENT_PTR(CACHE_REFERENCES),
1930	EVENT_PTR(CACHE_MISSES),
1931	EVENT_PTR(BRANCH_INSTRUCTIONS),
1932	EVENT_PTR(BRANCH_MISSES),
1933	EVENT_PTR(BUS_CYCLES),
1934	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1935	EVENT_PTR(STALLED_CYCLES_BACKEND),
1936	EVENT_PTR(REF_CPU_CYCLES),
1937	NULL,
1938};
1939
1940/*
1941 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1942 * out of events_attr attributes.
1943 */
1944static umode_t
1945is_visible(struct kobject *kobj, struct attribute *attr, int idx)
1946{
1947	struct perf_pmu_events_attr *pmu_attr;
1948
1949	if (idx >= x86_pmu.max_events)
1950		return 0;
1951
1952	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
1953	/* str trumps id */
1954	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
1955}
1956
1957static struct attribute_group x86_pmu_events_group __ro_after_init = {
1958	.name = "events",
1959	.attrs = events_attr,
1960	.is_visible = is_visible,
1961};
1962
1963ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1964{
1965	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1966	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1967	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1968	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1969	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1970	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1971	ssize_t ret;
1972
1973	/*
1974	* We have whole page size to spend and just little data
1975	* to write, so we can safely use sprintf.
1976	*/
1977	ret = sprintf(page, "event=0x%02llx", event);
1978
1979	if (umask)
1980		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1981
1982	if (edge)
1983		ret += sprintf(page + ret, ",edge");
1984
1985	if (pc)
1986		ret += sprintf(page + ret, ",pc");
1987
1988	if (any)
1989		ret += sprintf(page + ret, ",any");
1990
1991	if (inv)
1992		ret += sprintf(page + ret, ",inv");
1993
1994	if (cmask)
1995		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1996
1997	ret += sprintf(page + ret, "\n");
1998
1999	return ret;
2000}
2001
2002static struct attribute_group x86_pmu_attr_group;
2003static struct attribute_group x86_pmu_caps_group;
2004
2005static void x86_pmu_static_call_update(void)
2006{
2007	static_call_update(x86_pmu_handle_irq, x86_pmu.handle_irq);
2008	static_call_update(x86_pmu_disable_all, x86_pmu.disable_all);
2009	static_call_update(x86_pmu_enable_all, x86_pmu.enable_all);
2010	static_call_update(x86_pmu_enable, x86_pmu.enable);
2011	static_call_update(x86_pmu_disable, x86_pmu.disable);
2012
2013	static_call_update(x86_pmu_assign, x86_pmu.assign);
2014
2015	static_call_update(x86_pmu_add, x86_pmu.add);
2016	static_call_update(x86_pmu_del, x86_pmu.del);
2017	static_call_update(x86_pmu_read, x86_pmu.read);
2018
2019	static_call_update(x86_pmu_set_period, x86_pmu.set_period);
2020	static_call_update(x86_pmu_update, x86_pmu.update);
2021	static_call_update(x86_pmu_limit_period, x86_pmu.limit_period);
2022
2023	static_call_update(x86_pmu_schedule_events, x86_pmu.schedule_events);
2024	static_call_update(x86_pmu_get_event_constraints, x86_pmu.get_event_constraints);
2025	static_call_update(x86_pmu_put_event_constraints, x86_pmu.put_event_constraints);
2026
2027	static_call_update(x86_pmu_start_scheduling, x86_pmu.start_scheduling);
2028	static_call_update(x86_pmu_commit_scheduling, x86_pmu.commit_scheduling);
2029	static_call_update(x86_pmu_stop_scheduling, x86_pmu.stop_scheduling);
2030
2031	static_call_update(x86_pmu_sched_task, x86_pmu.sched_task);
2032	static_call_update(x86_pmu_swap_task_ctx, x86_pmu.swap_task_ctx);
2033
2034	static_call_update(x86_pmu_drain_pebs, x86_pmu.drain_pebs);
2035	static_call_update(x86_pmu_pebs_aliases, x86_pmu.pebs_aliases);
2036
2037	static_call_update(x86_pmu_guest_get_msrs, x86_pmu.guest_get_msrs);
2038	static_call_update(x86_pmu_filter, x86_pmu.filter);
2039}
2040
2041static void _x86_pmu_read(struct perf_event *event)
2042{
2043	static_call(x86_pmu_update)(event);
2044}
2045
2046void x86_pmu_show_pmu_cap(struct pmu *pmu)
2047{
2048	pr_info("... version:                %d\n",     x86_pmu.version);
2049	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
2050	pr_info("... generic registers:      %d\n",     x86_pmu_num_counters(pmu));
2051	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
2052	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
2053	pr_info("... fixed-purpose events:   %d\n",     x86_pmu_num_counters_fixed(pmu));
2054	pr_info("... event mask:             %016Lx\n", hybrid(pmu, intel_ctrl));
2055}
2056
2057static int __init init_hw_perf_events(void)
2058{
2059	struct x86_pmu_quirk *quirk;
2060	int err;
2061
2062	pr_info("Performance Events: ");
2063
2064	switch (boot_cpu_data.x86_vendor) {
2065	case X86_VENDOR_INTEL:
2066		err = intel_pmu_init();
2067		break;
2068	case X86_VENDOR_AMD:
2069		err = amd_pmu_init();
2070		break;
2071	case X86_VENDOR_HYGON:
2072		err = amd_pmu_init();
2073		x86_pmu.name = "HYGON";
2074		break;
2075	case X86_VENDOR_ZHAOXIN:
2076	case X86_VENDOR_CENTAUR:
2077		err = zhaoxin_pmu_init();
2078		break;
2079	default:
2080		err = -ENOTSUPP;
2081	}
2082	if (err != 0) {
2083		pr_cont("no PMU driver, software events only.\n");
2084		err = 0;
2085		goto out_bad_pmu;
2086	}
2087
2088	pmu_check_apic();
2089
2090	/* sanity check that the hardware exists or is emulated */
2091	if (!check_hw_exists(&pmu, x86_pmu.cntr_mask, x86_pmu.fixed_cntr_mask))
2092		goto out_bad_pmu;
2093
2094	pr_cont("%s PMU driver.\n", x86_pmu.name);
2095
2096	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
2097
2098	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
2099		quirk->func();
2100
2101	if (!x86_pmu.intel_ctrl)
2102		x86_pmu.intel_ctrl = x86_pmu.cntr_mask64;
2103
2104	if (!x86_pmu.config_mask)
2105		x86_pmu.config_mask = X86_RAW_EVENT_MASK;
2106
2107	perf_events_lapic_init();
2108	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
2109
2110	unconstrained = (struct event_constraint)
2111		__EVENT_CONSTRAINT(0, x86_pmu.cntr_mask64,
2112				   0, x86_pmu_num_counters(NULL), 0, 0);
2113
2114	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
2115
2116	if (!x86_pmu.events_sysfs_show)
2117		x86_pmu_events_group.attrs = &empty_attrs;
2118
2119	pmu.attr_update = x86_pmu.attr_update;
2120
2121	if (!is_hybrid())
2122		x86_pmu_show_pmu_cap(NULL);
2123
2124	if (!x86_pmu.read)
2125		x86_pmu.read = _x86_pmu_read;
2126
2127	if (!x86_pmu.guest_get_msrs)
2128		x86_pmu.guest_get_msrs = (void *)&__static_call_return0;
2129
2130	if (!x86_pmu.set_period)
2131		x86_pmu.set_period = x86_perf_event_set_period;
2132
2133	if (!x86_pmu.update)
2134		x86_pmu.update = x86_perf_event_update;
2135
2136	x86_pmu_static_call_update();
2137
2138	/*
2139	 * Install callbacks. Core will call them for each online
2140	 * cpu.
2141	 */
2142	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
2143				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
2144	if (err)
2145		return err;
2146
2147	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
2148				"perf/x86:starting", x86_pmu_starting_cpu,
2149				x86_pmu_dying_cpu);
2150	if (err)
2151		goto out;
2152
2153	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
2154				x86_pmu_online_cpu, NULL);
2155	if (err)
2156		goto out1;
2157
2158	if (!is_hybrid()) {
2159		err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
2160		if (err)
2161			goto out2;
2162	} else {
2163		struct x86_hybrid_pmu *hybrid_pmu;
2164		int i, j;
2165
2166		for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) {
2167			hybrid_pmu = &x86_pmu.hybrid_pmu[i];
2168
2169			hybrid_pmu->pmu = pmu;
2170			hybrid_pmu->pmu.type = -1;
2171			hybrid_pmu->pmu.attr_update = x86_pmu.attr_update;
2172			hybrid_pmu->pmu.capabilities |= PERF_PMU_CAP_EXTENDED_HW_TYPE;
2173
2174			err = perf_pmu_register(&hybrid_pmu->pmu, hybrid_pmu->name,
2175						(hybrid_pmu->pmu_type == hybrid_big) ? PERF_TYPE_RAW : -1);
2176			if (err)
2177				break;
2178		}
2179
2180		if (i < x86_pmu.num_hybrid_pmus) {
2181			for (j = 0; j < i; j++)
2182				perf_pmu_unregister(&x86_pmu.hybrid_pmu[j].pmu);
2183			pr_warn("Failed to register hybrid PMUs\n");
2184			kfree(x86_pmu.hybrid_pmu);
2185			x86_pmu.hybrid_pmu = NULL;
2186			x86_pmu.num_hybrid_pmus = 0;
2187			goto out2;
2188		}
2189	}
2190
2191	return 0;
2192
2193out2:
2194	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
2195out1:
2196	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
2197out:
2198	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
2199out_bad_pmu:
2200	memset(&x86_pmu, 0, sizeof(x86_pmu));
2201	return err;
2202}
2203early_initcall(init_hw_perf_events);
2204
2205static void x86_pmu_read(struct perf_event *event)
2206{
2207	static_call(x86_pmu_read)(event);
 
 
2208}
2209
2210/*
2211 * Start group events scheduling transaction
2212 * Set the flag to make pmu::enable() not perform the
2213 * schedulability test, it will be performed at commit time
2214 *
2215 * We only support PERF_PMU_TXN_ADD transactions. Save the
2216 * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
2217 * transactions.
2218 */
2219static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
2220{
2221	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2222
2223	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
2224
2225	cpuc->txn_flags = txn_flags;
2226	if (txn_flags & ~PERF_PMU_TXN_ADD)
2227		return;
2228
2229	perf_pmu_disable(pmu);
2230	__this_cpu_write(cpu_hw_events.n_txn, 0);
2231	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
2232	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
2233}
2234
2235/*
2236 * Stop group events scheduling transaction
2237 * Clear the flag and pmu::enable() will perform the
2238 * schedulability test.
2239 */
2240static void x86_pmu_cancel_txn(struct pmu *pmu)
2241{
2242	unsigned int txn_flags;
2243	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2244
2245	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2246
2247	txn_flags = cpuc->txn_flags;
2248	cpuc->txn_flags = 0;
2249	if (txn_flags & ~PERF_PMU_TXN_ADD)
2250		return;
2251
2252	/*
2253	 * Truncate collected array by the number of events added in this
2254	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
2255	 */
2256	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
2257	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
2258	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
2259	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
2260	perf_pmu_enable(pmu);
2261}
2262
2263/*
2264 * Commit group events scheduling transaction
2265 * Perform the group schedulability test as a whole
2266 * Return 0 if success
2267 *
2268 * Does not cancel the transaction on failure; expects the caller to do this.
2269 */
2270static int x86_pmu_commit_txn(struct pmu *pmu)
2271{
2272	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2273	int assign[X86_PMC_IDX_MAX];
2274	int n, ret;
2275
2276	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2277
2278	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
2279		cpuc->txn_flags = 0;
2280		return 0;
2281	}
2282
2283	n = cpuc->n_events;
2284
2285	if (!x86_pmu_initialized())
2286		return -EAGAIN;
2287
2288	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
2289	if (ret)
2290		return ret;
2291
2292	/*
2293	 * copy new assignment, now we know it is possible
2294	 * will be used by hw_perf_enable()
2295	 */
2296	memcpy(cpuc->assign, assign, n*sizeof(int));
2297
2298	cpuc->txn_flags = 0;
2299	perf_pmu_enable(pmu);
2300	return 0;
2301}
2302/*
2303 * a fake_cpuc is used to validate event groups. Due to
2304 * the extra reg logic, we need to also allocate a fake
2305 * per_core and per_cpu structure. Otherwise, group events
2306 * using extra reg may conflict without the kernel being
2307 * able to catch this when the last event gets added to
2308 * the group.
2309 */
2310static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2311{
2312	intel_cpuc_finish(cpuc);
2313	kfree(cpuc);
2314}
2315
2316static struct cpu_hw_events *allocate_fake_cpuc(struct pmu *event_pmu)
2317{
2318	struct cpu_hw_events *cpuc;
2319	int cpu;
2320
2321	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2322	if (!cpuc)
2323		return ERR_PTR(-ENOMEM);
2324	cpuc->is_fake = 1;
2325
2326	if (is_hybrid()) {
2327		struct x86_hybrid_pmu *h_pmu;
2328
2329		h_pmu = hybrid_pmu(event_pmu);
2330		if (cpumask_empty(&h_pmu->supported_cpus))
2331			goto error;
2332		cpu = cpumask_first(&h_pmu->supported_cpus);
2333	} else
2334		cpu = raw_smp_processor_id();
2335	cpuc->pmu = event_pmu;
2336
2337	if (intel_cpuc_prepare(cpuc, cpu))
2338		goto error;
2339
2340	return cpuc;
2341error:
2342	free_fake_cpuc(cpuc);
2343	return ERR_PTR(-ENOMEM);
2344}
2345
2346/*
2347 * validate that we can schedule this event
2348 */
2349static int validate_event(struct perf_event *event)
2350{
2351	struct cpu_hw_events *fake_cpuc;
2352	struct event_constraint *c;
2353	int ret = 0;
2354
2355	fake_cpuc = allocate_fake_cpuc(event->pmu);
2356	if (IS_ERR(fake_cpuc))
2357		return PTR_ERR(fake_cpuc);
2358
2359	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2360
2361	if (!c || !c->weight)
2362		ret = -EINVAL;
2363
2364	if (x86_pmu.put_event_constraints)
2365		x86_pmu.put_event_constraints(fake_cpuc, event);
2366
2367	free_fake_cpuc(fake_cpuc);
2368
2369	return ret;
2370}
2371
2372/*
2373 * validate a single event group
2374 *
2375 * validation include:
2376 *	- check events are compatible which each other
2377 *	- events do not compete for the same counter
2378 *	- number of events <= number of counters
2379 *
2380 * validation ensures the group can be loaded onto the
2381 * PMU if it was the only group available.
2382 */
2383static int validate_group(struct perf_event *event)
2384{
2385	struct perf_event *leader = event->group_leader;
2386	struct cpu_hw_events *fake_cpuc;
2387	int ret = -EINVAL, n;
2388
2389	/*
2390	 * Reject events from different hybrid PMUs.
2391	 */
2392	if (is_hybrid()) {
2393		struct perf_event *sibling;
2394		struct pmu *pmu = NULL;
2395
2396		if (is_x86_event(leader))
2397			pmu = leader->pmu;
2398
2399		for_each_sibling_event(sibling, leader) {
2400			if (!is_x86_event(sibling))
2401				continue;
2402			if (!pmu)
2403				pmu = sibling->pmu;
2404			else if (pmu != sibling->pmu)
2405				return ret;
2406		}
2407	}
2408
2409	fake_cpuc = allocate_fake_cpuc(event->pmu);
2410	if (IS_ERR(fake_cpuc))
2411		return PTR_ERR(fake_cpuc);
2412	/*
2413	 * the event is not yet connected with its
2414	 * siblings therefore we must first collect
2415	 * existing siblings, then add the new event
2416	 * before we can simulate the scheduling
2417	 */
2418	n = collect_events(fake_cpuc, leader, true);
2419	if (n < 0)
2420		goto out;
2421
2422	fake_cpuc->n_events = n;
2423	n = collect_events(fake_cpuc, event, false);
2424	if (n < 0)
2425		goto out;
2426
2427	fake_cpuc->n_events = 0;
2428	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2429
2430out:
2431	free_fake_cpuc(fake_cpuc);
2432	return ret;
2433}
2434
2435static int x86_pmu_event_init(struct perf_event *event)
2436{
2437	struct x86_hybrid_pmu *pmu = NULL;
2438	int err;
2439
2440	if ((event->attr.type != event->pmu->type) &&
2441	    (event->attr.type != PERF_TYPE_HARDWARE) &&
2442	    (event->attr.type != PERF_TYPE_HW_CACHE))
 
 
 
 
2443		return -ENOENT;
2444
2445	if (is_hybrid() && (event->cpu != -1)) {
2446		pmu = hybrid_pmu(event->pmu);
2447		if (!cpumask_test_cpu(event->cpu, &pmu->supported_cpus))
2448			return -ENOENT;
2449	}
2450
2451	err = __x86_pmu_event_init(event);
2452	if (!err) {
 
 
 
 
 
 
 
 
2453		if (event->group_leader != event)
2454			err = validate_group(event);
2455		else
2456			err = validate_event(event);
 
 
2457	}
2458	if (err) {
2459		if (event->destroy)
2460			event->destroy(event);
2461		event->destroy = NULL;
2462	}
2463
2464	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2465	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2466		event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
2467
2468	return err;
2469}
2470
2471void perf_clear_dirty_counters(void)
2472{
2473	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2474	int i;
2475
2476	 /* Don't need to clear the assigned counter. */
2477	for (i = 0; i < cpuc->n_events; i++)
2478		__clear_bit(cpuc->assign[i], cpuc->dirty);
2479
2480	if (bitmap_empty(cpuc->dirty, X86_PMC_IDX_MAX))
2481		return;
2482
2483	for_each_set_bit(i, cpuc->dirty, X86_PMC_IDX_MAX) {
2484		if (i >= INTEL_PMC_IDX_FIXED) {
2485			/* Metrics and fake events don't have corresponding HW counters. */
2486			if (!test_bit(i - INTEL_PMC_IDX_FIXED, hybrid(cpuc->pmu, fixed_cntr_mask)))
2487				continue;
2488
2489			wrmsrl(x86_pmu_fixed_ctr_addr(i - INTEL_PMC_IDX_FIXED), 0);
2490		} else {
2491			wrmsrl(x86_pmu_event_addr(i), 0);
2492		}
2493	}
2494
2495	bitmap_zero(cpuc->dirty, X86_PMC_IDX_MAX);
2496}
2497
2498static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2499{
2500	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
2501		return;
2502
2503	/*
2504	 * This function relies on not being called concurrently in two
2505	 * tasks in the same mm.  Otherwise one task could observe
2506	 * perf_rdpmc_allowed > 1 and return all the way back to
2507	 * userspace with CR4.PCE clear while another task is still
2508	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2509	 *
2510	 * For now, this can't happen because all callers hold mmap_lock
2511	 * for write.  If this changes, we'll need a different solution.
2512	 */
2513	mmap_assert_write_locked(mm);
2514
2515	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2516		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2517}
2518
2519static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2520{
2521	if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
 
2522		return;
2523
2524	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2525		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2526}
2527
2528static int x86_pmu_event_idx(struct perf_event *event)
2529{
2530	struct hw_perf_event *hwc = &event->hw;
2531
2532	if (!(hwc->flags & PERF_EVENT_FLAG_USER_READ_CNT))
2533		return 0;
2534
2535	if (is_metric_idx(hwc->idx))
2536		return INTEL_PMC_FIXED_RDPMC_METRICS + 1;
2537	else
2538		return hwc->event_base_rdpmc + 1;
 
 
2539}
2540
2541static ssize_t get_attr_rdpmc(struct device *cdev,
2542			      struct device_attribute *attr,
2543			      char *buf)
2544{
2545	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2546}
2547
2548static ssize_t set_attr_rdpmc(struct device *cdev,
2549			      struct device_attribute *attr,
2550			      const char *buf, size_t count)
2551{
2552	static DEFINE_MUTEX(rdpmc_mutex);
2553	unsigned long val;
2554	ssize_t ret;
2555
2556	ret = kstrtoul(buf, 0, &val);
2557	if (ret)
2558		return ret;
2559
2560	if (val > 2)
2561		return -EINVAL;
2562
2563	if (x86_pmu.attr_rdpmc_broken)
2564		return -ENOTSUPP;
2565
2566	guard(mutex)(&rdpmc_mutex);
2567
2568	if (val != x86_pmu.attr_rdpmc) {
2569		/*
2570		 * Changing into or out of never available or always available,
2571		 * aka perf-event-bypassing mode. This path is extremely slow,
2572		 * but only root can trigger it, so it's okay.
2573		 */
2574		if (val == 0)
2575			static_branch_inc(&rdpmc_never_available_key);
2576		else if (x86_pmu.attr_rdpmc == 0)
2577			static_branch_dec(&rdpmc_never_available_key);
2578
2579		if (val == 2)
2580			static_branch_inc(&rdpmc_always_available_key);
2581		else if (x86_pmu.attr_rdpmc == 2)
2582			static_branch_dec(&rdpmc_always_available_key);
 
 
2583
2584		on_each_cpu(cr4_update_pce, NULL, 1);
2585		x86_pmu.attr_rdpmc = val;
2586	}
2587
2588	return count;
2589}
2590
2591static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2592
2593static struct attribute *x86_pmu_attrs[] = {
2594	&dev_attr_rdpmc.attr,
2595	NULL,
2596};
2597
2598static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2599	.attrs = x86_pmu_attrs,
2600};
2601
2602static ssize_t max_precise_show(struct device *cdev,
2603				  struct device_attribute *attr,
2604				  char *buf)
2605{
2606	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2607}
2608
2609static DEVICE_ATTR_RO(max_precise);
2610
2611static struct attribute *x86_pmu_caps_attrs[] = {
2612	&dev_attr_max_precise.attr,
2613	NULL
2614};
2615
2616static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2617	.name = "caps",
2618	.attrs = x86_pmu_caps_attrs,
2619};
2620
2621static const struct attribute_group *x86_pmu_attr_groups[] = {
2622	&x86_pmu_attr_group,
2623	&x86_pmu_format_group,
2624	&x86_pmu_events_group,
2625	&x86_pmu_caps_group,
2626	NULL,
2627};
2628
2629static void x86_pmu_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
2630{
2631	static_call_cond(x86_pmu_sched_task)(pmu_ctx, sched_in);
2632}
2633
2634static void x86_pmu_swap_task_ctx(struct perf_event_pmu_context *prev_epc,
2635				  struct perf_event_pmu_context *next_epc)
2636{
2637	static_call_cond(x86_pmu_swap_task_ctx)(prev_epc, next_epc);
 
2638}
2639
2640void perf_check_microcode(void)
2641{
2642	if (x86_pmu.check_microcode)
2643		x86_pmu.check_microcode();
2644}
2645
2646static int x86_pmu_check_period(struct perf_event *event, u64 value)
2647{
2648	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2649		return -EINVAL;
2650
2651	if (value && x86_pmu.limit_period) {
2652		s64 left = value;
2653		x86_pmu.limit_period(event, &left);
2654		if (left > value)
2655			return -EINVAL;
2656	}
2657
2658	return 0;
2659}
2660
2661static int x86_pmu_aux_output_match(struct perf_event *event)
2662{
2663	if (!(pmu.capabilities & PERF_PMU_CAP_AUX_OUTPUT))
2664		return 0;
2665
2666	if (x86_pmu.aux_output_match)
2667		return x86_pmu.aux_output_match(event);
2668
2669	return 0;
2670}
2671
2672static bool x86_pmu_filter(struct pmu *pmu, int cpu)
2673{
2674	bool ret = false;
2675
2676	static_call_cond(x86_pmu_filter)(pmu, cpu, &ret);
2677
2678	return ret;
2679}
2680
2681static struct pmu pmu = {
2682	.pmu_enable		= x86_pmu_enable,
2683	.pmu_disable		= x86_pmu_disable,
2684
2685	.attr_groups		= x86_pmu_attr_groups,
2686
2687	.event_init		= x86_pmu_event_init,
2688
2689	.event_mapped		= x86_pmu_event_mapped,
2690	.event_unmapped		= x86_pmu_event_unmapped,
2691
2692	.add			= x86_pmu_add,
2693	.del			= x86_pmu_del,
2694	.start			= x86_pmu_start,
2695	.stop			= x86_pmu_stop,
2696	.read			= x86_pmu_read,
2697
2698	.start_txn		= x86_pmu_start_txn,
2699	.cancel_txn		= x86_pmu_cancel_txn,
2700	.commit_txn		= x86_pmu_commit_txn,
2701
2702	.event_idx		= x86_pmu_event_idx,
2703	.sched_task		= x86_pmu_sched_task,
2704	.swap_task_ctx		= x86_pmu_swap_task_ctx,
2705	.check_period		= x86_pmu_check_period,
2706
2707	.aux_output_match	= x86_pmu_aux_output_match,
2708
2709	.filter			= x86_pmu_filter,
2710};
2711
2712void arch_perf_update_userpage(struct perf_event *event,
2713			       struct perf_event_mmap_page *userpg, u64 now)
2714{
2715	struct cyc2ns_data data;
2716	u64 offset;
2717
2718	userpg->cap_user_time = 0;
2719	userpg->cap_user_time_zero = 0;
2720	userpg->cap_user_rdpmc =
2721		!!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT);
2722	userpg->pmc_width = x86_pmu.cntval_bits;
2723
2724	if (!using_native_sched_clock() || !sched_clock_stable())
2725		return;
2726
2727	cyc2ns_read_begin(&data);
2728
2729	offset = data.cyc2ns_offset + __sched_clock_offset;
2730
2731	/*
2732	 * Internal timekeeping for enabled/running/stopped times
2733	 * is always in the local_clock domain.
2734	 */
2735	userpg->cap_user_time = 1;
2736	userpg->time_mult = data.cyc2ns_mul;
2737	userpg->time_shift = data.cyc2ns_shift;
2738	userpg->time_offset = offset - now;
2739
2740	/*
2741	 * cap_user_time_zero doesn't make sense when we're using a different
2742	 * time base for the records.
2743	 */
2744	if (!event->attr.use_clockid) {
2745		userpg->cap_user_time_zero = 1;
2746		userpg->time_zero = offset;
2747	}
2748
2749	cyc2ns_read_end();
2750}
2751
2752/*
2753 * Determine whether the regs were taken from an irq/exception handler rather
2754 * than from perf_arch_fetch_caller_regs().
2755 */
2756static bool perf_hw_regs(struct pt_regs *regs)
2757{
2758	return regs->flags & X86_EFLAGS_FIXED;
2759}
2760
2761void
2762perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2763{
2764	struct unwind_state state;
2765	unsigned long addr;
2766
2767	if (perf_guest_state()) {
2768		/* TODO: We don't support guest os callchain now */
2769		return;
2770	}
2771
2772	if (perf_callchain_store(entry, regs->ip))
2773		return;
2774
2775	if (perf_hw_regs(regs))
2776		unwind_start(&state, current, regs, NULL);
2777	else
2778		unwind_start(&state, current, NULL, (void *)regs->sp);
2779
2780	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2781		addr = unwind_get_return_address(&state);
2782		if (!addr || perf_callchain_store(entry, addr))
2783			return;
2784	}
2785}
2786
2787static inline int
2788valid_user_frame(const void __user *fp, unsigned long size)
2789{
2790	return __access_ok(fp, size);
2791}
2792
2793static unsigned long get_segment_base(unsigned int segment)
2794{
2795	struct desc_struct *desc;
2796	unsigned int idx = segment >> 3;
2797
2798	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2799#ifdef CONFIG_MODIFY_LDT_SYSCALL
2800		struct ldt_struct *ldt;
2801
2802		/* IRQs are off, so this synchronizes with smp_store_release */
2803		ldt = READ_ONCE(current->active_mm->context.ldt);
2804		if (!ldt || idx >= ldt->nr_entries)
2805			return 0;
2806
2807		desc = &ldt->entries[idx];
2808#else
2809		return 0;
2810#endif
2811	} else {
2812		if (idx >= GDT_ENTRIES)
2813			return 0;
2814
2815		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2816	}
2817
2818	return get_desc_base(desc);
2819}
2820
2821#ifdef CONFIG_UPROBES
2822/*
2823 * Heuristic-based check if uprobe is installed at the function entry.
2824 *
2825 * Under assumption of user code being compiled with frame pointers,
2826 * `push %rbp/%ebp` is a good indicator that we indeed are.
2827 *
2828 * Similarly, `endbr64` (assuming 64-bit mode) is also a common pattern.
2829 * If we get this wrong, captured stack trace might have one extra bogus
2830 * entry, but the rest of stack trace will still be meaningful.
2831 */
2832static bool is_uprobe_at_func_entry(struct pt_regs *regs)
2833{
2834	struct arch_uprobe *auprobe;
2835
2836	if (!current->utask)
2837		return false;
2838
2839	auprobe = current->utask->auprobe;
2840	if (!auprobe)
2841		return false;
2842
2843	/* push %rbp/%ebp */
2844	if (auprobe->insn[0] == 0x55)
2845		return true;
2846
2847	/* endbr64 (64-bit only) */
2848	if (user_64bit_mode(regs) && is_endbr(*(u32 *)auprobe->insn))
2849		return true;
2850
2851	return false;
2852}
2853
2854#else
2855static bool is_uprobe_at_func_entry(struct pt_regs *regs)
2856{
2857	return false;
2858}
2859#endif /* CONFIG_UPROBES */
2860
2861#ifdef CONFIG_IA32_EMULATION
2862
2863#include <linux/compat.h>
2864
2865static inline int
2866perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2867{
2868	/* 32-bit process in 64-bit kernel. */
2869	unsigned long ss_base, cs_base;
2870	struct stack_frame_ia32 frame;
2871	const struct stack_frame_ia32 __user *fp;
2872	u32 ret_addr;
2873
2874	if (user_64bit_mode(regs))
2875		return 0;
2876
2877	cs_base = get_segment_base(regs->cs);
2878	ss_base = get_segment_base(regs->ss);
2879
2880	fp = compat_ptr(ss_base + regs->bp);
2881	pagefault_disable();
 
 
 
 
2882
2883	/* see perf_callchain_user() below for why we do this */
2884	if (is_uprobe_at_func_entry(regs) &&
2885	    !get_user(ret_addr, (const u32 __user *)regs->sp))
2886		perf_callchain_store(entry, ret_addr);
2887
2888	while (entry->nr < entry->max_stack) {
2889		if (!valid_user_frame(fp, sizeof(frame)))
2890			break;
2891
2892		if (__get_user(frame.next_frame, &fp->next_frame))
 
2893			break;
2894		if (__get_user(frame.return_address, &fp->return_address))
 
2895			break;
2896
2897		perf_callchain_store(entry, cs_base + frame.return_address);
2898		fp = compat_ptr(ss_base + frame.next_frame);
2899	}
2900	pagefault_enable();
2901	return 1;
2902}
2903#else
2904static inline int
2905perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2906{
2907    return 0;
2908}
2909#endif
2910
2911void
2912perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2913{
2914	struct stack_frame frame;
2915	const struct stack_frame __user *fp;
2916	unsigned long ret_addr;
2917
2918	if (perf_guest_state()) {
2919		/* TODO: We don't support guest os callchain now */
2920		return;
2921	}
2922
2923	/*
2924	 * We don't know what to do with VM86 stacks.. ignore them for now.
2925	 */
2926	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2927		return;
2928
2929	fp = (void __user *)regs->bp;
2930
2931	perf_callchain_store(entry, regs->ip);
2932
2933	if (!nmi_uaccess_okay())
2934		return;
2935
2936	if (perf_callchain_user32(regs, entry))
2937		return;
2938
2939	pagefault_disable();
 
 
2940
2941	/*
2942	 * If we are called from uprobe handler, and we are indeed at the very
2943	 * entry to user function (which is normally a `push %rbp` instruction,
2944	 * under assumption of application being compiled with frame pointers),
2945	 * we should read return address from *regs->sp before proceeding
2946	 * to follow frame pointers, otherwise we'll skip immediate caller
2947	 * as %rbp is not yet setup.
2948	 */
2949	if (is_uprobe_at_func_entry(regs) &&
2950	    !get_user(ret_addr, (const unsigned long __user *)regs->sp))
2951		perf_callchain_store(entry, ret_addr);
2952
2953	while (entry->nr < entry->max_stack) {
2954		if (!valid_user_frame(fp, sizeof(frame)))
2955			break;
2956
2957		if (__get_user(frame.next_frame, &fp->next_frame))
 
2958			break;
2959		if (__get_user(frame.return_address, &fp->return_address))
 
2960			break;
2961
2962		perf_callchain_store(entry, frame.return_address);
2963		fp = (void __user *)frame.next_frame;
2964	}
2965	pagefault_enable();
2966}
2967
2968/*
2969 * Deal with code segment offsets for the various execution modes:
2970 *
2971 *   VM86 - the good olde 16 bit days, where the linear address is
2972 *          20 bits and we use regs->ip + 0x10 * regs->cs.
2973 *
2974 *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2975 *          to figure out what the 32bit base address is.
2976 *
2977 *    X32 - has TIF_X32 set, but is running in x86_64
2978 *
2979 * X86_64 - CS,DS,SS,ES are all zero based.
2980 */
2981static unsigned long code_segment_base(struct pt_regs *regs)
2982{
2983	/*
2984	 * For IA32 we look at the GDT/LDT segment base to convert the
2985	 * effective IP to a linear address.
2986	 */
2987
2988#ifdef CONFIG_X86_32
2989	/*
2990	 * If we are in VM86 mode, add the segment offset to convert to a
2991	 * linear address.
2992	 */
2993	if (regs->flags & X86_VM_MASK)
2994		return 0x10 * regs->cs;
2995
2996	if (user_mode(regs) && regs->cs != __USER_CS)
2997		return get_segment_base(regs->cs);
2998#else
2999	if (user_mode(regs) && !user_64bit_mode(regs) &&
3000	    regs->cs != __USER32_CS)
3001		return get_segment_base(regs->cs);
3002#endif
3003	return 0;
3004}
3005
3006unsigned long perf_arch_instruction_pointer(struct pt_regs *regs)
3007{
 
 
 
3008	return regs->ip + code_segment_base(regs);
3009}
3010
3011static unsigned long common_misc_flags(struct pt_regs *regs)
3012{
3013	if (regs->flags & PERF_EFLAGS_EXACT)
3014		return PERF_RECORD_MISC_EXACT_IP;
3015
3016	return 0;
3017}
 
 
 
 
 
 
 
 
 
3018
3019static unsigned long guest_misc_flags(struct pt_regs *regs)
3020{
3021	unsigned long guest_state = perf_guest_state();
3022
3023	if (!(guest_state & PERF_GUEST_ACTIVE))
3024		return 0;
3025
3026	if (guest_state & PERF_GUEST_USER)
3027		return PERF_RECORD_MISC_GUEST_USER;
3028	else
3029		return PERF_RECORD_MISC_GUEST_KERNEL;
3030
3031}
3032
3033static unsigned long host_misc_flags(struct pt_regs *regs)
3034{
3035	if (user_mode(regs))
3036		return PERF_RECORD_MISC_USER;
3037	else
3038		return PERF_RECORD_MISC_KERNEL;
3039}
3040
3041unsigned long perf_arch_guest_misc_flags(struct pt_regs *regs)
3042{
3043	unsigned long flags = common_misc_flags(regs);
3044
3045	flags |= guest_misc_flags(regs);
3046
3047	return flags;
3048}
3049
3050unsigned long perf_arch_misc_flags(struct pt_regs *regs)
3051{
3052	unsigned long flags = common_misc_flags(regs);
3053
3054	flags |= host_misc_flags(regs);
3055
3056	return flags;
3057}
3058
3059void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
3060{
3061	/* This API doesn't currently support enumerating hybrid PMUs. */
3062	if (WARN_ON_ONCE(cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) ||
3063	    !x86_pmu_initialized()) {
3064		memset(cap, 0, sizeof(*cap));
3065		return;
3066	}
3067
3068	/*
3069	 * Note, hybrid CPU models get tracked as having hybrid PMUs even when
3070	 * all E-cores are disabled via BIOS.  When E-cores are disabled, the
3071	 * base PMU holds the correct number of counters for P-cores.
3072	 */
3073	cap->version		= x86_pmu.version;
3074	cap->num_counters_gp	= x86_pmu_num_counters(NULL);
3075	cap->num_counters_fixed	= x86_pmu_num_counters_fixed(NULL);
3076	cap->bit_width_gp	= x86_pmu.cntval_bits;
3077	cap->bit_width_fixed	= x86_pmu.cntval_bits;
3078	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
3079	cap->events_mask_len	= x86_pmu.events_mask_len;
3080	cap->pebs_ept		= x86_pmu.pebs_ept;
3081}
3082EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
3083
3084u64 perf_get_hw_event_config(int hw_event)
3085{
3086	int max = x86_pmu.max_events;
3087
3088	if (hw_event < max)
3089		return x86_pmu.event_map(array_index_nospec(hw_event, max));
3090
3091	return 0;
3092}
3093EXPORT_SYMBOL_GPL(perf_get_hw_event_config);