Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v4.6
 
   1/*
   2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3 *
   4 * Parts came from builtin-{top,stat,record}.c, see those files for further
   5 * copyright notes.
   6 *
   7 * Released under the GPL v2. (and only v2, not any later version)
   8 */
   9
  10#include <byteswap.h>
 
 
  11#include <linux/bitops.h>
 
  12#include <api/fs/tracing_path.h>
  13#include <traceevent/event-parse.h>
  14#include <linux/hw_breakpoint.h>
  15#include <linux/perf_event.h>
 
  16#include <linux/err.h>
 
 
  17#include <sys/resource.h>
 
 
 
 
  18#include "asm/bug.h"
 
  19#include "callchain.h"
  20#include "cgroup.h"
 
 
  21#include "evsel.h"
 
 
 
  22#include "evlist.h"
  23#include "util.h"
  24#include "cpumap.h"
  25#include "thread_map.h"
  26#include "target.h"
  27#include "perf_regs.h"
 
  28#include "debug.h"
  29#include "trace-event.h"
  30#include "stat.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  31
  32static struct {
  33	bool sample_id_all;
  34	bool exclude_guest;
  35	bool mmap2;
  36	bool cloexec;
  37	bool clockid;
  38	bool clockid_wrong;
  39	bool lbr_flags;
  40} perf_missing_features;
  41
  42static clockid_t clockid;
  43
  44static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  45{
  46	return 0;
  47}
  48
  49static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
 
 
  50{
  51}
  52
  53static struct {
  54	size_t	size;
  55	int	(*init)(struct perf_evsel *evsel);
  56	void	(*fini)(struct perf_evsel *evsel);
  57} perf_evsel__object = {
  58	.size = sizeof(struct perf_evsel),
  59	.init = perf_evsel__no_extra_init,
  60	.fini = perf_evsel__no_extra_fini,
  61};
  62
  63int perf_evsel__object_config(size_t object_size,
  64			      int (*init)(struct perf_evsel *evsel),
  65			      void (*fini)(struct perf_evsel *evsel))
  66{
  67
  68	if (object_size == 0)
  69		goto set_methods;
  70
  71	if (perf_evsel__object.size > object_size)
  72		return -EINVAL;
  73
  74	perf_evsel__object.size = object_size;
  75
  76set_methods:
  77	if (init != NULL)
  78		perf_evsel__object.init = init;
  79
  80	if (fini != NULL)
  81		perf_evsel__object.fini = fini;
  82
  83	return 0;
  84}
  85
  86#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  87
  88int __perf_evsel__sample_size(u64 sample_type)
  89{
  90	u64 mask = sample_type & PERF_SAMPLE_MASK;
  91	int size = 0;
  92	int i;
  93
  94	for (i = 0; i < 64; i++) {
  95		if (mask & (1ULL << i))
  96			size++;
  97	}
  98
  99	size *= sizeof(u64);
 100
 101	return size;
 102}
 103
 104/**
 105 * __perf_evsel__calc_id_pos - calculate id_pos.
 106 * @sample_type: sample type
 107 *
 108 * This function returns the position of the event id (PERF_SAMPLE_ID or
 109 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 110 * sample_event.
 111 */
 112static int __perf_evsel__calc_id_pos(u64 sample_type)
 113{
 114	int idx = 0;
 115
 116	if (sample_type & PERF_SAMPLE_IDENTIFIER)
 117		return 0;
 118
 119	if (!(sample_type & PERF_SAMPLE_ID))
 120		return -1;
 121
 122	if (sample_type & PERF_SAMPLE_IP)
 123		idx += 1;
 124
 125	if (sample_type & PERF_SAMPLE_TID)
 126		idx += 1;
 127
 128	if (sample_type & PERF_SAMPLE_TIME)
 129		idx += 1;
 130
 131	if (sample_type & PERF_SAMPLE_ADDR)
 132		idx += 1;
 133
 134	return idx;
 135}
 136
 137/**
 138 * __perf_evsel__calc_is_pos - calculate is_pos.
 139 * @sample_type: sample type
 140 *
 141 * This function returns the position (counting backwards) of the event id
 142 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 143 * sample_id_all is used there is an id sample appended to non-sample events.
 144 */
 145static int __perf_evsel__calc_is_pos(u64 sample_type)
 146{
 147	int idx = 1;
 148
 149	if (sample_type & PERF_SAMPLE_IDENTIFIER)
 150		return 1;
 151
 152	if (!(sample_type & PERF_SAMPLE_ID))
 153		return -1;
 154
 155	if (sample_type & PERF_SAMPLE_CPU)
 156		idx += 1;
 157
 158	if (sample_type & PERF_SAMPLE_STREAM_ID)
 159		idx += 1;
 160
 161	return idx;
 162}
 163
 164void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 165{
 166	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 167	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 168}
 169
 170void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 171				  enum perf_event_sample_format bit)
 172{
 173	if (!(evsel->attr.sample_type & bit)) {
 174		evsel->attr.sample_type |= bit;
 175		evsel->sample_size += sizeof(u64);
 176		perf_evsel__calc_id_pos(evsel);
 177	}
 178}
 179
 180void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 181				    enum perf_event_sample_format bit)
 182{
 183	if (evsel->attr.sample_type & bit) {
 184		evsel->attr.sample_type &= ~bit;
 185		evsel->sample_size -= sizeof(u64);
 186		perf_evsel__calc_id_pos(evsel);
 187	}
 188}
 189
 190void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 191			       bool can_sample_identifier)
 192{
 193	if (can_sample_identifier) {
 194		perf_evsel__reset_sample_bit(evsel, ID);
 195		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 196	} else {
 197		perf_evsel__set_sample_bit(evsel, ID);
 198	}
 199	evsel->attr.read_format |= PERF_FORMAT_ID;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 200}
 201
 202void perf_evsel__init(struct perf_evsel *evsel,
 203		      struct perf_event_attr *attr, int idx)
 204{
 205	evsel->idx	   = idx;
 206	evsel->tracking	   = !idx;
 207	evsel->attr	   = *attr;
 208	evsel->leader	   = evsel;
 209	evsel->unit	   = "";
 210	evsel->scale	   = 1.0;
 
 211	evsel->evlist	   = NULL;
 
 212	evsel->bpf_fd	   = -1;
 213	INIT_LIST_HEAD(&evsel->node);
 214	INIT_LIST_HEAD(&evsel->config_terms);
 
 
 215	perf_evsel__object.init(evsel);
 216	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 217	perf_evsel__calc_id_pos(evsel);
 218	evsel->cmdline_group_boundary = false;
 
 
 
 
 
 
 219}
 220
 221struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 222{
 223	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 224
 225	if (evsel != NULL)
 226		perf_evsel__init(evsel, attr, idx);
 
 
 
 
 
 
 
 227
 228	if (perf_evsel__is_bpf_output(evsel)) {
 229		evsel->attr.sample_type |= PERF_SAMPLE_RAW;
 230		evsel->attr.sample_period = 1;
 
 231	}
 232
 233	return evsel;
 234}
 235
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 236/*
 237 * Returns pointer with encoded error via <linux/err.h> interface.
 238 */
 239struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 
 240{
 241	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 242	int err = -ENOMEM;
 243
 244	if (evsel == NULL) {
 245		goto out_err;
 246	} else {
 247		struct perf_event_attr attr = {
 248			.type	       = PERF_TYPE_TRACEPOINT,
 249			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 250					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 251		};
 252
 253		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 254			goto out_free;
 255
 256		evsel->tp_format = trace_event__tp_format(sys, name);
 257		if (IS_ERR(evsel->tp_format)) {
 258			err = PTR_ERR(evsel->tp_format);
 259			goto out_free;
 260		}
 261
 262		event_attr_init(&attr);
 263		attr.config = evsel->tp_format->id;
 264		attr.sample_period = 1;
 265		perf_evsel__init(evsel, &attr, idx);
 266	}
 267
 268	return evsel;
 269
 270out_free:
 271	zfree(&evsel->name);
 272	free(evsel);
 273out_err:
 274	return ERR_PTR(err);
 275}
 
 276
 277const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 278	"cycles",
 279	"instructions",
 280	"cache-references",
 281	"cache-misses",
 282	"branches",
 283	"branch-misses",
 284	"bus-cycles",
 285	"stalled-cycles-frontend",
 286	"stalled-cycles-backend",
 287	"ref-cycles",
 288};
 289
 290static const char *__perf_evsel__hw_name(u64 config)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 291{
 292	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 293		return perf_evsel__hw_names[config];
 294
 295	return "unknown-hardware";
 296}
 297
 298static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 299{
 300	int colon = 0, r = 0;
 301	struct perf_event_attr *attr = &evsel->attr;
 302	bool exclude_guest_default = false;
 303
 304#define MOD_PRINT(context, mod)	do {					\
 305		if (!attr->exclude_##context) {				\
 306			if (!colon) colon = ++r;			\
 307			r += scnprintf(bf + r, size - r, "%c", mod);	\
 308		} } while(0)
 309
 310	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 311		MOD_PRINT(kernel, 'k');
 312		MOD_PRINT(user, 'u');
 313		MOD_PRINT(hv, 'h');
 314		exclude_guest_default = true;
 315	}
 316
 317	if (attr->precise_ip) {
 318		if (!colon)
 319			colon = ++r;
 320		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 321		exclude_guest_default = true;
 322	}
 323
 324	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 325		MOD_PRINT(host, 'H');
 326		MOD_PRINT(guest, 'G');
 327	}
 328#undef MOD_PRINT
 329	if (colon)
 330		bf[colon - 1] = ':';
 331	return r;
 332}
 333
 334static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 335{
 336	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 337	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 338}
 339
 340const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 
 
 
 
 
 
 341	"cpu-clock",
 342	"task-clock",
 343	"page-faults",
 344	"context-switches",
 345	"cpu-migrations",
 346	"minor-faults",
 347	"major-faults",
 348	"alignment-faults",
 349	"emulation-faults",
 350	"dummy",
 351};
 352
 353static const char *__perf_evsel__sw_name(u64 config)
 354{
 355	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 356		return perf_evsel__sw_names[config];
 357	return "unknown-software";
 358}
 359
 360static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 361{
 362	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 363	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 364}
 365
 366static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 
 
 
 
 
 367{
 368	int r;
 369
 370	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 371
 372	if (type & HW_BREAKPOINT_R)
 373		r += scnprintf(bf + r, size - r, "r");
 374
 375	if (type & HW_BREAKPOINT_W)
 376		r += scnprintf(bf + r, size - r, "w");
 377
 378	if (type & HW_BREAKPOINT_X)
 379		r += scnprintf(bf + r, size - r, "x");
 380
 381	return r;
 382}
 383
 384static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 385{
 386	struct perf_event_attr *attr = &evsel->attr;
 387	int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 388	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 389}
 390
 391const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 392				[PERF_EVSEL__MAX_ALIASES] = {
 393 { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
 394 { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
 395 { "LLC",	"L2",							},
 396 { "dTLB",	"d-tlb",	"Data-TLB",				},
 397 { "iTLB",	"i-tlb",	"Instruction-TLB",			},
 398 { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
 399 { "node",								},
 400};
 401
 402const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 403				   [PERF_EVSEL__MAX_ALIASES] = {
 404 { "load",	"loads",	"read",					},
 405 { "store",	"stores",	"write",				},
 406 { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
 407};
 408
 409const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 410				       [PERF_EVSEL__MAX_ALIASES] = {
 411 { "refs",	"Reference",	"ops",		"access",		},
 412 { "misses",	"miss",							},
 413};
 414
 415#define C(x)		PERF_COUNT_HW_CACHE_##x
 416#define CACHE_READ	(1 << C(OP_READ))
 417#define CACHE_WRITE	(1 << C(OP_WRITE))
 418#define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
 419#define COP(x)		(1 << x)
 420
 421/*
 422 * cache operartion stat
 423 * L1I : Read and prefetch only
 424 * ITLB and BPU : Read-only
 425 */
 426static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 427 [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 428 [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
 429 [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 430 [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 431 [C(ITLB)]	= (CACHE_READ),
 432 [C(BPU)]	= (CACHE_READ),
 433 [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 434};
 435
 436bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 437{
 438	if (perf_evsel__hw_cache_stat[type] & COP(op))
 439		return true;	/* valid */
 440	else
 441		return false;	/* invalid */
 442}
 443
 444int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 445					    char *bf, size_t size)
 446{
 447	if (result) {
 448		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 449				 perf_evsel__hw_cache_op[op][0],
 450				 perf_evsel__hw_cache_result[result][0]);
 451	}
 452
 453	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 454			 perf_evsel__hw_cache_op[op][1]);
 455}
 456
 457static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 458{
 459	u8 op, result, type = (config >>  0) & 0xff;
 460	const char *err = "unknown-ext-hardware-cache-type";
 461
 462	if (type > PERF_COUNT_HW_CACHE_MAX)
 463		goto out_err;
 464
 465	op = (config >>  8) & 0xff;
 466	err = "unknown-ext-hardware-cache-op";
 467	if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 468		goto out_err;
 469
 470	result = (config >> 16) & 0xff;
 471	err = "unknown-ext-hardware-cache-result";
 472	if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 473		goto out_err;
 474
 475	err = "invalid-cache";
 476	if (!perf_evsel__is_cache_op_valid(type, op))
 477		goto out_err;
 478
 479	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 480out_err:
 481	return scnprintf(bf, size, "%s", err);
 482}
 483
 484static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 485{
 486	int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 487	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 488}
 489
 490static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 491{
 492	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 493	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 494}
 495
 496const char *perf_evsel__name(struct perf_evsel *evsel)
 497{
 498	char bf[128];
 499
 
 
 
 500	if (evsel->name)
 501		return evsel->name;
 502
 503	switch (evsel->attr.type) {
 504	case PERF_TYPE_RAW:
 505		perf_evsel__raw_name(evsel, bf, sizeof(bf));
 506		break;
 507
 508	case PERF_TYPE_HARDWARE:
 509		perf_evsel__hw_name(evsel, bf, sizeof(bf));
 510		break;
 511
 512	case PERF_TYPE_HW_CACHE:
 513		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 514		break;
 515
 516	case PERF_TYPE_SOFTWARE:
 517		perf_evsel__sw_name(evsel, bf, sizeof(bf));
 
 
 
 518		break;
 519
 520	case PERF_TYPE_TRACEPOINT:
 521		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 522		break;
 523
 524	case PERF_TYPE_BREAKPOINT:
 525		perf_evsel__bp_name(evsel, bf, sizeof(bf));
 526		break;
 527
 528	default:
 529		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 530			  evsel->attr.type);
 531		break;
 532	}
 533
 534	evsel->name = strdup(bf);
 535
 536	return evsel->name ?: "unknown";
 
 
 
 537}
 538
 539const char *perf_evsel__group_name(struct perf_evsel *evsel)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 540{
 541	return evsel->group_name ?: "anon group";
 542}
 543
 544int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 
 
 
 
 
 
 
 
 
 
 545{
 546	int ret;
 547	struct perf_evsel *pos;
 548	const char *group_name = perf_evsel__group_name(evsel);
 549
 550	ret = scnprintf(buf, size, "%s", group_name);
 
 551
 552	ret += scnprintf(buf + ret, size - ret, " { %s",
 553			 perf_evsel__name(evsel));
 554
 555	for_each_group_member(pos, evsel)
 556		ret += scnprintf(buf + ret, size - ret, ", %s",
 557				 perf_evsel__name(pos));
 558
 559	ret += scnprintf(buf + ret, size - ret, " }");
 
 560
 561	return ret;
 562}
 563
 564static void
 565perf_evsel__config_callgraph(struct perf_evsel *evsel,
 566			     struct record_opts *opts,
 567			     struct callchain_param *param)
 568{
 569	bool function = perf_evsel__is_function_event(evsel);
 570	struct perf_event_attr *attr = &evsel->attr;
 
 571
 572	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 573
 
 
 
 
 
 
 574	if (param->record_mode == CALLCHAIN_LBR) {
 575		if (!opts->branch_stack) {
 576			if (attr->exclude_user) {
 577				pr_warning("LBR callstack option is only available "
 578					   "to get user callchain information. "
 579					   "Falling back to framepointers.\n");
 580			} else {
 581				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 582				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 583							PERF_SAMPLE_BRANCH_CALL_STACK |
 584							PERF_SAMPLE_BRANCH_NO_CYCLES |
 585							PERF_SAMPLE_BRANCH_NO_FLAGS;
 
 586			}
 587		} else
 588			 pr_warning("Cannot use LBR callstack with branch stack. "
 589				    "Falling back to framepointers.\n");
 590	}
 591
 592	if (param->record_mode == CALLCHAIN_DWARF) {
 593		if (!function) {
 594			perf_evsel__set_sample_bit(evsel, REGS_USER);
 595			perf_evsel__set_sample_bit(evsel, STACK_USER);
 596			attr->sample_regs_user = PERF_REGS_MASK;
 
 
 
 
 
 
 
 
 597			attr->sample_stack_user = param->dump_size;
 598			attr->exclude_callchain_user = 1;
 599		} else {
 600			pr_info("Cannot use DWARF unwind for function trace event,"
 601				" falling back to framepointers.\n");
 602		}
 603	}
 604
 605	if (function) {
 606		pr_info("Disabling user space callchains for function trace event.\n");
 607		attr->exclude_callchain_user = 1;
 608	}
 609}
 610
 611static void
 612perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 613			    struct callchain_param *param)
 614{
 615	struct perf_event_attr *attr = &evsel->attr;
 
 
 
 
 
 
 616
 617	perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 618	if (param->record_mode == CALLCHAIN_LBR) {
 619		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 620		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 621					      PERF_SAMPLE_BRANCH_CALL_STACK);
 
 622	}
 623	if (param->record_mode == CALLCHAIN_DWARF) {
 624		perf_evsel__reset_sample_bit(evsel, REGS_USER);
 625		perf_evsel__reset_sample_bit(evsel, STACK_USER);
 626	}
 627}
 628
 629static void apply_config_terms(struct perf_evsel *evsel,
 630			       struct record_opts *opts)
 631{
 632	struct perf_evsel_config_term *term;
 633	struct list_head *config_terms = &evsel->config_terms;
 634	struct perf_event_attr *attr = &evsel->attr;
 635	struct callchain_param param;
 636	u32 dump_size = 0;
 637	char *callgraph_buf = NULL;
 638
 639	/* callgraph default */
 640	param.record_mode = callchain_param.record_mode;
 
 
 
 
 
 641
 642	list_for_each_entry(term, config_terms, list) {
 643		switch (term->type) {
 644		case PERF_EVSEL__CONFIG_TERM_PERIOD:
 645			attr->sample_period = term->val.period;
 646			attr->freq = 0;
 
 
 
 647			break;
 648		case PERF_EVSEL__CONFIG_TERM_FREQ:
 649			attr->sample_freq = term->val.freq;
 650			attr->freq = 1;
 
 
 
 651			break;
 652		case PERF_EVSEL__CONFIG_TERM_TIME:
 653			if (term->val.time)
 654				perf_evsel__set_sample_bit(evsel, TIME);
 655			else
 656				perf_evsel__reset_sample_bit(evsel, TIME);
 657			break;
 658		case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 659			callgraph_buf = term->val.callgraph;
 660			break;
 661		case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 
 
 
 
 
 
 
 
 662			dump_size = term->val.stack_user;
 663			break;
 664		case PERF_EVSEL__CONFIG_TERM_INHERIT:
 
 
 
 
 
 
 665			/*
 666			 * attr->inherit should has already been set by
 667			 * perf_evsel__config. If user explicitly set
 668			 * inherit using config terms, override global
 669			 * opt->no_inherit setting.
 670			 */
 671			attr->inherit = term->val.inherit ? 1 : 0;
 672			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 673		default:
 674			break;
 675		}
 676	}
 677
 678	/* User explicitly set per-event callgraph, clear the old setting and reset. */
 679	if ((callgraph_buf != NULL) || (dump_size > 0)) {
 
 
 
 
 
 
 
 680
 681		/* parse callgraph parameters */
 682		if (callgraph_buf != NULL) {
 683			if (!strcmp(callgraph_buf, "no")) {
 684				param.enabled = false;
 685				param.record_mode = CALLCHAIN_NONE;
 686			} else {
 687				param.enabled = true;
 688				if (parse_callchain_record(callgraph_buf, &param)) {
 689					pr_err("per-event callgraph setting for %s failed. "
 690					       "Apply callgraph global setting for it\n",
 691					       evsel->name);
 692					return;
 693				}
 
 
 694			}
 695		}
 696		if (dump_size > 0) {
 697			dump_size = round_up(dump_size, sizeof(u64));
 698			param.dump_size = dump_size;
 699		}
 700
 701		/* If global callgraph set, clear it */
 702		if (callchain_param.enabled)
 703			perf_evsel__reset_callgraph(evsel, &callchain_param);
 704
 705		/* set perf-event callgraph */
 706		if (param.enabled)
 707			perf_evsel__config_callgraph(evsel, opts, &param);
 
 
 
 
 
 
 708	}
 709}
 710
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 711/*
 712 * The enable_on_exec/disabled value strategy:
 713 *
 714 *  1) For any type of traced program:
 715 *    - all independent events and group leaders are disabled
 716 *    - all group members are enabled
 717 *
 718 *     Group members are ruled by group leaders. They need to
 719 *     be enabled, because the group scheduling relies on that.
 720 *
 721 *  2) For traced programs executed by perf:
 722 *     - all independent events and group leaders have
 723 *       enable_on_exec set
 724 *     - we don't specifically enable or disable any event during
 725 *       the record command
 726 *
 727 *     Independent events and group leaders are initially disabled
 728 *     and get enabled by exec. Group members are ruled by group
 729 *     leaders as stated in 1).
 730 *
 731 *  3) For traced programs attached by perf (pid/tid):
 732 *     - we specifically enable or disable all events during
 733 *       the record command
 734 *
 735 *     When attaching events to already running traced we
 736 *     enable/disable events specifically, as there's no
 737 *     initial traced exec call.
 738 */
 739void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 
 740{
 741	struct perf_evsel *leader = evsel->leader;
 742	struct perf_event_attr *attr = &evsel->attr;
 743	int track = evsel->tracking;
 744	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 745
 746	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 747	attr->inherit	    = !opts->no_inherit;
 
 
 748
 749	perf_evsel__set_sample_bit(evsel, IP);
 750	perf_evsel__set_sample_bit(evsel, TID);
 751
 752	if (evsel->sample_read) {
 753		perf_evsel__set_sample_bit(evsel, READ);
 754
 755		/*
 756		 * We need ID even in case of single event, because
 757		 * PERF_SAMPLE_READ process ID specific data.
 758		 */
 759		perf_evsel__set_sample_id(evsel, false);
 760
 761		/*
 762		 * Apply group format only if we belong to group
 763		 * with more than one members.
 764		 */
 765		if (leader->nr_members > 1) {
 766			attr->read_format |= PERF_FORMAT_GROUP;
 767			attr->inherit = 0;
 768		}
 769	}
 770
 771	/*
 772	 * We default some events to have a default interval. But keep
 773	 * it a weak assumption overridable by the user.
 774	 */
 775	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 776				     opts->user_interval != ULLONG_MAX)) {
 777		if (opts->freq) {
 778			perf_evsel__set_sample_bit(evsel, PERIOD);
 779			attr->freq		= 1;
 780			attr->sample_freq	= opts->freq;
 781		} else {
 782			attr->sample_period = opts->default_interval;
 783		}
 784	}
 785
 786	/*
 787	 * Disable sampling for all group members other
 788	 * than leader in case leader 'leads' the sampling.
 789	 */
 790	if ((leader != evsel) && leader->sample_read) {
 791		attr->sample_freq   = 0;
 792		attr->sample_period = 0;
 793	}
 794
 795	if (opts->no_samples)
 796		attr->sample_freq = 0;
 797
 798	if (opts->inherit_stat)
 
 
 
 
 799		attr->inherit_stat = 1;
 
 800
 801	if (opts->sample_address) {
 802		perf_evsel__set_sample_bit(evsel, ADDR);
 803		attr->mmap_data = track;
 804	}
 805
 806	/*
 807	 * We don't allow user space callchains for  function trace
 808	 * event, due to issues with page faults while tracing page
 809	 * fault handler and its overall trickiness nature.
 810	 */
 811	if (perf_evsel__is_function_event(evsel))
 812		evsel->attr.exclude_callchain_user = 1;
 813
 814	if (callchain_param.enabled && !evsel->no_aux_samples)
 815		perf_evsel__config_callgraph(evsel, opts, &callchain_param);
 816
 817	if (opts->sample_intr_regs) {
 
 818		attr->sample_regs_intr = opts->sample_intr_regs;
 819		perf_evsel__set_sample_bit(evsel, REGS_INTR);
 820	}
 821
 822	if (target__has_cpu(&opts->target))
 823		perf_evsel__set_sample_bit(evsel, CPU);
 
 
 
 824
 825	if (opts->period)
 826		perf_evsel__set_sample_bit(evsel, PERIOD);
 827
 828	/*
 829	 * When the user explicitely disabled time don't force it here.
 830	 */
 831	if (opts->sample_time &&
 832	    (!perf_missing_features.sample_id_all &&
 833	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 834	     opts->sample_time_set)))
 835		perf_evsel__set_sample_bit(evsel, TIME);
 836
 837	if (opts->raw_samples && !evsel->no_aux_samples) {
 838		perf_evsel__set_sample_bit(evsel, TIME);
 839		perf_evsel__set_sample_bit(evsel, RAW);
 840		perf_evsel__set_sample_bit(evsel, CPU);
 841	}
 842
 843	if (opts->sample_address)
 844		perf_evsel__set_sample_bit(evsel, DATA_SRC);
 
 
 
 845
 846	if (opts->no_buffering) {
 847		attr->watermark = 0;
 848		attr->wakeup_events = 1;
 849	}
 850	if (opts->branch_stack && !evsel->no_aux_samples) {
 851		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 852		attr->branch_sample_type = opts->branch_stack;
 853	}
 854
 855	if (opts->sample_weight)
 856		perf_evsel__set_sample_bit(evsel, WEIGHT);
 857
 858	attr->task  = track;
 859	attr->mmap  = track;
 860	attr->mmap2 = track && !perf_missing_features.mmap2;
 861	attr->comm  = track;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 862
 863	if (opts->record_switch_events)
 864		attr->context_switch = track;
 865
 866	if (opts->sample_transaction)
 867		perf_evsel__set_sample_bit(evsel, TRANSACTION);
 868
 869	if (opts->running_time) {
 870		evsel->attr.read_format |=
 871			PERF_FORMAT_TOTAL_TIME_ENABLED |
 872			PERF_FORMAT_TOTAL_TIME_RUNNING;
 873	}
 874
 875	/*
 876	 * XXX see the function comment above
 877	 *
 878	 * Disabling only independent events or group leaders,
 879	 * keeping group members enabled.
 880	 */
 881	if (perf_evsel__is_group_leader(evsel))
 882		attr->disabled = 1;
 883
 884	/*
 885	 * Setting enable_on_exec for independent events and
 886	 * group leaders for traced executed by perf.
 887	 */
 888	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 889		!opts->initial_delay)
 890		attr->enable_on_exec = 1;
 891
 892	if (evsel->immediate) {
 893		attr->disabled = 0;
 894		attr->enable_on_exec = 0;
 895	}
 896
 897	clockid = opts->clockid;
 898	if (opts->use_clockid) {
 899		attr->use_clockid = 1;
 900		attr->clockid = opts->clockid;
 901	}
 902
 903	if (evsel->precise_max)
 904		perf_event_attr__set_max_precise_ip(attr);
 905
 906	if (opts->all_user) {
 907		attr->exclude_kernel = 1;
 908		attr->exclude_user   = 0;
 909	}
 910
 911	if (opts->all_kernel) {
 912		attr->exclude_kernel = 0;
 913		attr->exclude_user   = 1;
 914	}
 915
 
 
 
 916	/*
 917	 * Apply event specific term settings,
 918	 * it overloads any global configuration.
 919	 */
 920	apply_config_terms(evsel, opts);
 921}
 922
 923static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 924{
 925	int cpu, thread;
 926
 927	if (evsel->system_wide)
 928		nthreads = 1;
 929
 930	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 931
 932	if (evsel->fd) {
 933		for (cpu = 0; cpu < ncpus; cpu++) {
 934			for (thread = 0; thread < nthreads; thread++) {
 935				FD(evsel, cpu, thread) = -1;
 936			}
 937		}
 938	}
 939
 940	return evsel->fd != NULL ? 0 : -ENOMEM;
 941}
 942
 943static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 944			  int ioc,  void *arg)
 945{
 946	int cpu, thread;
 947
 948	if (evsel->system_wide)
 949		nthreads = 1;
 950
 951	for (cpu = 0; cpu < ncpus; cpu++) {
 952		for (thread = 0; thread < nthreads; thread++) {
 953			int fd = FD(evsel, cpu, thread),
 954			    err = ioctl(fd, ioc, arg);
 955
 956			if (err)
 957				return err;
 958		}
 959	}
 960
 961	return 0;
 962}
 963
 964int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 965			     const char *filter)
 966{
 967	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 968				     PERF_EVENT_IOC_SET_FILTER,
 969				     (void *)filter);
 970}
 971
 972int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 973{
 974	char *new_filter = strdup(filter);
 975
 976	if (new_filter != NULL) {
 977		free(evsel->filter);
 978		evsel->filter = new_filter;
 979		return 0;
 980	}
 981
 982	return -1;
 983}
 984
 985int perf_evsel__append_filter(struct perf_evsel *evsel,
 986			      const char *op, const char *filter)
 987{
 988	char *new_filter;
 989
 990	if (evsel->filter == NULL)
 991		return perf_evsel__set_filter(evsel, filter);
 992
 993	if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 994		free(evsel->filter);
 995		evsel->filter = new_filter;
 996		return 0;
 997	}
 998
 999	return -1;
1000}
1001
1002int perf_evsel__enable(struct perf_evsel *evsel)
1003{
1004	int nthreads = thread_map__nr(evsel->threads);
1005	int ncpus = cpu_map__nr(evsel->cpus);
1006
1007	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1008				     PERF_EVENT_IOC_ENABLE,
1009				     0);
1010}
1011
1012int perf_evsel__disable(struct perf_evsel *evsel)
1013{
1014	int nthreads = thread_map__nr(evsel->threads);
1015	int ncpus = cpu_map__nr(evsel->cpus);
1016
1017	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1018				     PERF_EVENT_IOC_DISABLE,
1019				     0);
1020}
1021
1022int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 
1023{
1024	if (ncpus == 0 || nthreads == 0)
1025		return 0;
1026
1027	if (evsel->system_wide)
1028		nthreads = 1;
1029
1030	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1031	if (evsel->sample_id == NULL)
1032		return -ENOMEM;
1033
1034	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1035	if (evsel->id == NULL) {
1036		xyarray__delete(evsel->sample_id);
1037		evsel->sample_id = NULL;
1038		return -ENOMEM;
1039	}
1040
1041	return 0;
 
 
1042}
1043
1044static void perf_evsel__free_fd(struct perf_evsel *evsel)
 
1045{
1046	xyarray__delete(evsel->fd);
1047	evsel->fd = NULL;
1048}
1049
1050static void perf_evsel__free_id(struct perf_evsel *evsel)
1051{
1052	xyarray__delete(evsel->sample_id);
1053	evsel->sample_id = NULL;
1054	zfree(&evsel->id);
 
 
 
 
 
 
 
 
1055}
1056
1057static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1058{
1059	struct perf_evsel_config_term *term, *h;
1060
1061	list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1062		list_del(&term->list);
 
 
1063		free(term);
1064	}
1065}
1066
1067void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1068{
1069	int cpu, thread;
1070
1071	if (evsel->system_wide)
1072		nthreads = 1;
1073
1074	for (cpu = 0; cpu < ncpus; cpu++)
1075		for (thread = 0; thread < nthreads; ++thread) {
1076			close(FD(evsel, cpu, thread));
1077			FD(evsel, cpu, thread) = -1;
1078		}
1079}
1080
1081void perf_evsel__exit(struct perf_evsel *evsel)
1082{
1083	assert(list_empty(&evsel->node));
1084	assert(evsel->evlist == NULL);
1085	perf_evsel__free_fd(evsel);
1086	perf_evsel__free_id(evsel);
1087	perf_evsel__free_config_terms(evsel);
1088	close_cgroup(evsel->cgrp);
1089	cpu_map__put(evsel->cpus);
1090	cpu_map__put(evsel->own_cpus);
1091	thread_map__put(evsel->threads);
 
 
 
1092	zfree(&evsel->group_name);
1093	zfree(&evsel->name);
 
 
 
 
 
 
 
 
 
1094	perf_evsel__object.fini(evsel);
1095}
1096
1097void perf_evsel__delete(struct perf_evsel *evsel)
1098{
1099	perf_evsel__exit(evsel);
 
 
 
1100	free(evsel);
1101}
1102
1103void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1104				struct perf_counts_values *count)
1105{
1106	struct perf_counts_values tmp;
1107
1108	if (!evsel->prev_raw_counts)
1109		return;
1110
1111	if (cpu == -1) {
1112		tmp = evsel->prev_raw_counts->aggr;
1113		evsel->prev_raw_counts->aggr = *count;
1114	} else {
1115		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1116		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1117	}
1118
1119	count->val = count->val - tmp.val;
1120	count->ena = count->ena - tmp.ena;
1121	count->run = count->run - tmp.run;
1122}
1123
1124void perf_counts_values__scale(struct perf_counts_values *count,
1125			       bool scale, s8 *pscaled)
1126{
1127	s8 scaled = 0;
1128
1129	if (scale) {
1130		if (count->run == 0) {
1131			scaled = -1;
1132			count->val = 0;
1133		} else if (count->run < count->ena) {
1134			scaled = 1;
1135			count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1136		}
1137	} else
1138		count->ena = count->run = 0;
1139
1140	if (pscaled)
1141		*pscaled = scaled;
 
 
 
 
 
 
 
 
 
 
 
1142}
1143
1144int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1145		     struct perf_counts_values *count)
1146{
1147	memset(count, 0, sizeof(*count));
 
 
 
 
1148
1149	if (FD(evsel, cpu, thread) < 0)
1150		return -EINVAL;
1151
1152	if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1153		return -errno;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1154
1155	return 0;
1156}
1157
1158int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1159			      int cpu, int thread, bool scale)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1160{
1161	struct perf_counts_values count;
1162	size_t nv = scale ? 3 : 1;
1163
1164	if (FD(evsel, cpu, thread) < 0)
1165		return -EINVAL;
1166
1167	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1168		return -ENOMEM;
1169
1170	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1171		return -errno;
1172
1173	perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1174	perf_counts_values__scale(&count, scale, NULL);
1175	*perf_counts(evsel->counts, cpu, thread) = count;
1176	return 0;
1177}
1178
1179static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1180{
1181	struct perf_evsel *leader = evsel->leader;
1182	int fd;
1183
1184	if (perf_evsel__is_group_leader(evsel))
1185		return -1;
1186
1187	/*
1188	 * Leader must be already processed/open,
1189	 * if not it's a bug.
1190	 */
1191	BUG_ON(!leader->fd);
1192
1193	fd = FD(leader, cpu, thread);
1194	BUG_ON(fd == -1);
 
1195
1196	return fd;
1197}
1198
1199struct bit_names {
1200	int bit;
1201	const char *name;
1202};
 
 
1203
1204static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1205{
1206	bool first_bit = true;
1207	int i = 0;
1208
1209	do {
1210		if (value & bits[i].bit) {
1211			buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1212			first_bit = false;
1213		}
1214	} while (bits[++i].name != NULL);
1215}
1216
1217static void __p_sample_type(char *buf, size_t size, u64 value)
1218{
1219#define bit_name(n) { PERF_SAMPLE_##n, #n }
1220	struct bit_names bits[] = {
1221		bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1222		bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1223		bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1224		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1225		bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1226		bit_name(WEIGHT),
1227		{ .name = NULL, }
1228	};
1229#undef bit_name
1230	__p_bits(buf, size, value, bits);
1231}
1232
1233static void __p_read_format(char *buf, size_t size, u64 value)
 
 
1234{
1235#define bit_name(n) { PERF_FORMAT_##n, #n }
1236	struct bit_names bits[] = {
1237		bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1238		bit_name(ID), bit_name(GROUP),
1239		{ .name = NULL, }
1240	};
1241#undef bit_name
1242	__p_bits(buf, size, value, bits);
1243}
1244
1245#define BUF_SIZE		1024
 
1246
1247#define p_hex(val)		snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1248#define p_unsigned(val)		snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1249#define p_signed(val)		snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1250#define p_sample_type(val)	__p_sample_type(buf, BUF_SIZE, val)
1251#define p_read_format(val)	__p_read_format(buf, BUF_SIZE, val)
1252
1253#define PRINT_ATTRn(_n, _f, _p)				\
1254do {							\
1255	if (attr->_f) {					\
1256		_p(attr->_f);				\
1257		ret += attr__fprintf(fp, _n, buf, priv);\
1258	}						\
1259} while (0)
1260
1261#define PRINT_ATTRf(_f, _p)	PRINT_ATTRn(#_f, _f, _p)
 
 
 
 
 
 
 
 
1262
1263int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1264			     attr__fprintf_f attr__fprintf, void *priv)
1265{
1266	char buf[BUF_SIZE];
1267	int ret = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1268
1269	PRINT_ATTRf(type, p_unsigned);
1270	PRINT_ATTRf(size, p_unsigned);
1271	PRINT_ATTRf(config, p_hex);
1272	PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1273	PRINT_ATTRf(sample_type, p_sample_type);
1274	PRINT_ATTRf(read_format, p_read_format);
1275
1276	PRINT_ATTRf(disabled, p_unsigned);
1277	PRINT_ATTRf(inherit, p_unsigned);
1278	PRINT_ATTRf(pinned, p_unsigned);
1279	PRINT_ATTRf(exclusive, p_unsigned);
1280	PRINT_ATTRf(exclude_user, p_unsigned);
1281	PRINT_ATTRf(exclude_kernel, p_unsigned);
1282	PRINT_ATTRf(exclude_hv, p_unsigned);
1283	PRINT_ATTRf(exclude_idle, p_unsigned);
1284	PRINT_ATTRf(mmap, p_unsigned);
1285	PRINT_ATTRf(comm, p_unsigned);
1286	PRINT_ATTRf(freq, p_unsigned);
1287	PRINT_ATTRf(inherit_stat, p_unsigned);
1288	PRINT_ATTRf(enable_on_exec, p_unsigned);
1289	PRINT_ATTRf(task, p_unsigned);
1290	PRINT_ATTRf(watermark, p_unsigned);
1291	PRINT_ATTRf(precise_ip, p_unsigned);
1292	PRINT_ATTRf(mmap_data, p_unsigned);
1293	PRINT_ATTRf(sample_id_all, p_unsigned);
1294	PRINT_ATTRf(exclude_host, p_unsigned);
1295	PRINT_ATTRf(exclude_guest, p_unsigned);
1296	PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1297	PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1298	PRINT_ATTRf(mmap2, p_unsigned);
1299	PRINT_ATTRf(comm_exec, p_unsigned);
1300	PRINT_ATTRf(use_clockid, p_unsigned);
1301	PRINT_ATTRf(context_switch, p_unsigned);
1302
1303	PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1304	PRINT_ATTRf(bp_type, p_unsigned);
1305	PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1306	PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1307	PRINT_ATTRf(branch_sample_type, p_unsigned);
1308	PRINT_ATTRf(sample_regs_user, p_hex);
1309	PRINT_ATTRf(sample_stack_user, p_unsigned);
1310	PRINT_ATTRf(clockid, p_signed);
1311	PRINT_ATTRf(sample_regs_intr, p_hex);
1312	PRINT_ATTRf(aux_watermark, p_unsigned);
1313
1314	return ret;
 
 
 
 
 
1315}
1316
1317static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1318				void *priv __attribute__((unused)))
1319{
1320	return fprintf(fp, "  %-32s %s\n", name, val);
1321}
1322
1323static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1324			      struct thread_map *threads)
1325{
1326	int cpu, thread, nthreads;
1327	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1328	int pid = -1, err;
1329	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
 
 
 
1330
1331	if (evsel->system_wide)
1332		nthreads = 1;
1333	else
1334		nthreads = threads->nr;
 
1335
1336	if (evsel->fd == NULL &&
1337	    perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1338		return -ENOMEM;
 
 
 
 
 
1339
1340	if (evsel->cgrp) {
1341		flags |= PERF_FLAG_PID_CGROUP;
1342		pid = evsel->cgrp->fd;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1343	}
1344
1345fallback_missing_features:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1346	if (perf_missing_features.clockid_wrong)
1347		evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1348	if (perf_missing_features.clockid) {
1349		evsel->attr.use_clockid = 0;
1350		evsel->attr.clockid = 0;
1351	}
1352	if (perf_missing_features.cloexec)
1353		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1354	if (perf_missing_features.mmap2)
1355		evsel->attr.mmap2 = 0;
1356	if (perf_missing_features.exclude_guest)
1357		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1358	if (perf_missing_features.lbr_flags)
1359		evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1360				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1361retry_sample_id:
 
 
 
 
 
 
 
1362	if (perf_missing_features.sample_id_all)
1363		evsel->attr.sample_id_all = 0;
 
1364
1365	if (verbose >= 2) {
1366		fprintf(stderr, "%.60s\n", graph_dotted_line);
1367		fprintf(stderr, "perf_event_attr:\n");
1368		perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1369		fprintf(stderr, "%.60s\n", graph_dotted_line);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1370	}
 
1371
1372	for (cpu = 0; cpu < cpus->nr; cpu++) {
 
 
 
 
 
 
1373
1374		for (thread = 0; thread < nthreads; thread++) {
1375			int group_fd;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1376
1377			if (!evsel->cgrp && !evsel->system_wide)
1378				pid = thread_map__pid(threads, thread);
1379
1380			group_fd = get_group_fd(evsel, cpu, thread);
 
 
 
1381retry_open:
1382			pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1383				  pid, cpus->map[cpu], group_fd, flags);
 
 
 
1384
1385			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1386								     pid,
1387								     cpus->map[cpu],
1388								     group_fd, flags);
1389			if (FD(evsel, cpu, thread) < 0) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1390				err = -errno;
1391				pr_debug2("sys_perf_event_open failed, error %d\n",
 
1392					  err);
1393				goto try_fallback;
1394			}
1395
 
 
 
 
 
 
 
 
 
 
 
1396			if (evsel->bpf_fd >= 0) {
1397				int evt_fd = FD(evsel, cpu, thread);
1398				int bpf_fd = evsel->bpf_fd;
1399
1400				err = ioctl(evt_fd,
1401					    PERF_EVENT_IOC_SET_BPF,
1402					    bpf_fd);
1403				if (err && errno != EEXIST) {
1404					pr_err("failed to attach bpf fd %d: %s\n",
1405					       bpf_fd, strerror(errno));
1406					err = -EINVAL;
1407					goto out_close;
1408				}
1409			}
1410
1411			set_rlimit = NO_CHANGE;
1412
1413			/*
1414			 * If we succeeded but had to kill clockid, fail and
1415			 * have perf_evsel__open_strerror() print us a nice
1416			 * error.
1417			 */
1418			if (perf_missing_features.clockid ||
1419			    perf_missing_features.clockid_wrong) {
1420				err = -EINVAL;
1421				goto out_close;
1422			}
1423		}
1424	}
1425
1426	return 0;
1427
1428try_fallback:
 
 
 
 
 
 
 
 
 
 
 
 
1429	/*
1430	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1431	 * of them try to increase the limits.
1432	 */
1433	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1434		struct rlimit l;
1435		int old_errno = errno;
1436
1437		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1438			if (set_rlimit == NO_CHANGE)
1439				l.rlim_cur = l.rlim_max;
1440			else {
1441				l.rlim_cur = l.rlim_max + 1000;
1442				l.rlim_max = l.rlim_cur;
1443			}
1444			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1445				set_rlimit++;
1446				errno = old_errno;
1447				goto retry_open;
1448			}
1449		}
1450		errno = old_errno;
1451	}
1452
1453	if (err != -EINVAL || cpu > 0 || thread > 0)
1454		goto out_close;
1455
1456	/*
1457	 * Must probe features in the order they were added to the
1458	 * perf_event_attr interface.
1459	 */
1460	if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1461		perf_missing_features.clockid_wrong = true;
1462		goto fallback_missing_features;
1463	} else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1464		perf_missing_features.clockid = true;
1465		goto fallback_missing_features;
1466	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1467		perf_missing_features.cloexec = true;
1468		goto fallback_missing_features;
1469	} else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1470		perf_missing_features.mmap2 = true;
1471		goto fallback_missing_features;
1472	} else if (!perf_missing_features.exclude_guest &&
1473		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1474		perf_missing_features.exclude_guest = true;
1475		goto fallback_missing_features;
1476	} else if (!perf_missing_features.sample_id_all) {
1477		perf_missing_features.sample_id_all = true;
1478		goto retry_sample_id;
1479	} else if (!perf_missing_features.lbr_flags &&
1480			(evsel->attr.branch_sample_type &
1481			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1482			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1483		perf_missing_features.lbr_flags = true;
1484		goto fallback_missing_features;
1485	}
1486
1487out_close:
 
 
 
 
1488	do {
1489		while (--thread >= 0) {
1490			close(FD(evsel, cpu, thread));
1491			FD(evsel, cpu, thread) = -1;
 
1492		}
1493		thread = nthreads;
1494	} while (--cpu >= 0);
 
1495	return err;
1496}
1497
1498void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
 
1499{
1500	if (evsel->fd == NULL)
1501		return;
1502
1503	perf_evsel__close_fd(evsel, ncpus, nthreads);
1504	perf_evsel__free_fd(evsel);
1505}
1506
1507static struct {
1508	struct cpu_map map;
1509	int cpus[1];
1510} empty_cpu_map = {
1511	.map.nr	= 1,
1512	.cpus	= { -1, },
1513};
1514
1515static struct {
1516	struct thread_map map;
1517	int threads[1];
1518} empty_thread_map = {
1519	.map.nr	 = 1,
1520	.threads = { -1, },
1521};
1522
1523int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1524		     struct thread_map *threads)
1525{
1526	if (cpus == NULL) {
1527		/* Work around old compiler warnings about strict aliasing */
1528		cpus = &empty_cpu_map.map;
1529	}
1530
1531	if (threads == NULL)
1532		threads = &empty_thread_map.map;
1533
1534	return __perf_evsel__open(evsel, cpus, threads);
1535}
1536
1537int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1538			     struct cpu_map *cpus)
1539{
1540	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
 
 
 
1541}
1542
1543int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1544				struct thread_map *threads)
1545{
1546	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1547}
1548
1549static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1550				       const union perf_event *event,
1551				       struct perf_sample *sample)
1552{
1553	u64 type = evsel->attr.sample_type;
1554	const u64 *array = event->sample.array;
1555	bool swapped = evsel->needs_swap;
1556	union u64_swap u;
1557
1558	array += ((event->header.size -
1559		   sizeof(event->header)) / sizeof(u64)) - 1;
1560
1561	if (type & PERF_SAMPLE_IDENTIFIER) {
1562		sample->id = *array;
1563		array--;
1564	}
1565
1566	if (type & PERF_SAMPLE_CPU) {
1567		u.val64 = *array;
1568		if (swapped) {
1569			/* undo swap of u64, then swap on individual u32s */
1570			u.val64 = bswap_64(u.val64);
1571			u.val32[0] = bswap_32(u.val32[0]);
1572		}
1573
1574		sample->cpu = u.val32[0];
1575		array--;
1576	}
1577
1578	if (type & PERF_SAMPLE_STREAM_ID) {
1579		sample->stream_id = *array;
1580		array--;
1581	}
1582
1583	if (type & PERF_SAMPLE_ID) {
1584		sample->id = *array;
1585		array--;
1586	}
1587
1588	if (type & PERF_SAMPLE_TIME) {
1589		sample->time = *array;
1590		array--;
1591	}
1592
1593	if (type & PERF_SAMPLE_TID) {
1594		u.val64 = *array;
1595		if (swapped) {
1596			/* undo swap of u64, then swap on individual u32s */
1597			u.val64 = bswap_64(u.val64);
1598			u.val32[0] = bswap_32(u.val32[0]);
1599			u.val32[1] = bswap_32(u.val32[1]);
1600		}
1601
1602		sample->pid = u.val32[0];
1603		sample->tid = u.val32[1];
1604		array--;
1605	}
1606
1607	return 0;
1608}
1609
1610static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1611			    u64 size)
1612{
1613	return size > max_size || offset + size > endp;
1614}
1615
1616#define OVERFLOW_CHECK(offset, size, max_size)				\
1617	do {								\
1618		if (overflow(endp, (max_size), (offset), (size)))	\
1619			return -EFAULT;					\
1620	} while (0)
1621
1622#define OVERFLOW_CHECK_u64(offset) \
1623	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1624
1625int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1626			     struct perf_sample *data)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1627{
1628	u64 type = evsel->attr.sample_type;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1629	bool swapped = evsel->needs_swap;
1630	const u64 *array;
1631	u16 max_size = event->header.size;
1632	const void *endp = (void *)event + max_size;
1633	u64 sz;
1634
1635	/*
1636	 * used for cross-endian analysis. See git commit 65014ab3
1637	 * for why this goofiness is needed.
1638	 */
1639	union u64_swap u;
1640
1641	memset(data, 0, sizeof(*data));
1642	data->cpu = data->pid = data->tid = -1;
1643	data->stream_id = data->id = data->time = -1ULL;
1644	data->period = evsel->attr.sample_period;
1645	data->weight = 0;
1646	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 
 
 
 
1647
1648	if (event->header.type != PERF_RECORD_SAMPLE) {
1649		if (!evsel->attr.sample_id_all)
1650			return 0;
1651		return perf_evsel__parse_id_sample(evsel, event, data);
1652	}
1653
1654	array = event->sample.array;
1655
1656	/*
1657	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1658	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1659	 * check the format does not go past the end of the event.
1660	 */
1661	if (evsel->sample_size + sizeof(event->header) > event->header.size)
1662		return -EFAULT;
1663
1664	data->id = -1ULL;
1665	if (type & PERF_SAMPLE_IDENTIFIER) {
1666		data->id = *array;
1667		array++;
1668	}
1669
1670	if (type & PERF_SAMPLE_IP) {
1671		data->ip = *array;
1672		array++;
1673	}
1674
1675	if (type & PERF_SAMPLE_TID) {
1676		u.val64 = *array;
1677		if (swapped) {
1678			/* undo swap of u64, then swap on individual u32s */
1679			u.val64 = bswap_64(u.val64);
1680			u.val32[0] = bswap_32(u.val32[0]);
1681			u.val32[1] = bswap_32(u.val32[1]);
1682		}
1683
1684		data->pid = u.val32[0];
1685		data->tid = u.val32[1];
1686		array++;
1687	}
1688
1689	if (type & PERF_SAMPLE_TIME) {
1690		data->time = *array;
1691		array++;
1692	}
1693
1694	data->addr = 0;
1695	if (type & PERF_SAMPLE_ADDR) {
1696		data->addr = *array;
1697		array++;
1698	}
1699
1700	if (type & PERF_SAMPLE_ID) {
1701		data->id = *array;
1702		array++;
1703	}
1704
1705	if (type & PERF_SAMPLE_STREAM_ID) {
1706		data->stream_id = *array;
1707		array++;
1708	}
1709
1710	if (type & PERF_SAMPLE_CPU) {
1711
1712		u.val64 = *array;
1713		if (swapped) {
1714			/* undo swap of u64, then swap on individual u32s */
1715			u.val64 = bswap_64(u.val64);
1716			u.val32[0] = bswap_32(u.val32[0]);
1717		}
1718
1719		data->cpu = u.val32[0];
1720		array++;
1721	}
1722
1723	if (type & PERF_SAMPLE_PERIOD) {
1724		data->period = *array;
1725		array++;
1726	}
1727
1728	if (type & PERF_SAMPLE_READ) {
1729		u64 read_format = evsel->attr.read_format;
1730
1731		OVERFLOW_CHECK_u64(array);
1732		if (read_format & PERF_FORMAT_GROUP)
1733			data->read.group.nr = *array;
1734		else
1735			data->read.one.value = *array;
1736
1737		array++;
1738
1739		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1740			OVERFLOW_CHECK_u64(array);
1741			data->read.time_enabled = *array;
1742			array++;
1743		}
1744
1745		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1746			OVERFLOW_CHECK_u64(array);
1747			data->read.time_running = *array;
1748			array++;
1749		}
1750
1751		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1752		if (read_format & PERF_FORMAT_GROUP) {
1753			const u64 max_group_nr = UINT64_MAX /
1754					sizeof(struct sample_read_value);
1755
1756			if (data->read.group.nr > max_group_nr)
1757				return -EFAULT;
1758			sz = data->read.group.nr *
1759			     sizeof(struct sample_read_value);
1760			OVERFLOW_CHECK(array, sz, max_size);
1761			data->read.group.values =
1762					(struct sample_read_value *)array;
1763			array = (void *)array + sz;
1764		} else {
1765			OVERFLOW_CHECK_u64(array);
1766			data->read.one.id = *array;
1767			array++;
 
 
 
 
 
 
1768		}
1769	}
1770
1771	if (type & PERF_SAMPLE_CALLCHAIN) {
1772		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1773
1774		OVERFLOW_CHECK_u64(array);
1775		data->callchain = (struct ip_callchain *)array++;
1776		if (data->callchain->nr > max_callchain_nr)
1777			return -EFAULT;
1778		sz = data->callchain->nr * sizeof(u64);
1779		OVERFLOW_CHECK(array, sz, max_size);
1780		array = (void *)array + sz;
1781	}
1782
1783	if (type & PERF_SAMPLE_RAW) {
1784		OVERFLOW_CHECK_u64(array);
1785		u.val64 = *array;
1786		if (WARN_ONCE(swapped,
1787			      "Endianness of raw data not corrected!\n")) {
1788			/* undo swap of u64, then swap on individual u32s */
 
 
 
 
 
1789			u.val64 = bswap_64(u.val64);
1790			u.val32[0] = bswap_32(u.val32[0]);
1791			u.val32[1] = bswap_32(u.val32[1]);
1792		}
1793		data->raw_size = u.val32[0];
 
 
 
 
 
 
 
 
1794		array = (void *)array + sizeof(u32);
1795
1796		OVERFLOW_CHECK(array, data->raw_size, max_size);
1797		data->raw_data = (void *)array;
1798		array = (void *)array + data->raw_size;
1799	}
1800
1801	if (type & PERF_SAMPLE_BRANCH_STACK) {
1802		const u64 max_branch_nr = UINT64_MAX /
1803					  sizeof(struct branch_entry);
 
 
1804
1805		OVERFLOW_CHECK_u64(array);
1806		data->branch_stack = (struct branch_stack *)array++;
1807
1808		if (data->branch_stack->nr > max_branch_nr)
1809			return -EFAULT;
 
1810		sz = data->branch_stack->nr * sizeof(struct branch_entry);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1811		OVERFLOW_CHECK(array, sz, max_size);
1812		array = (void *)array + sz;
 
 
 
 
 
 
 
 
 
 
1813	}
1814
1815	if (type & PERF_SAMPLE_REGS_USER) {
1816		OVERFLOW_CHECK_u64(array);
1817		data->user_regs.abi = *array;
1818		array++;
1819
1820		if (data->user_regs.abi) {
1821			u64 mask = evsel->attr.sample_regs_user;
1822
1823			sz = hweight_long(mask) * sizeof(u64);
1824			OVERFLOW_CHECK(array, sz, max_size);
1825			data->user_regs.mask = mask;
1826			data->user_regs.regs = (u64 *)array;
1827			array = (void *)array + sz;
1828		}
1829	}
1830
1831	if (type & PERF_SAMPLE_STACK_USER) {
1832		OVERFLOW_CHECK_u64(array);
1833		sz = *array++;
1834
1835		data->user_stack.offset = ((char *)(array - 1)
1836					  - (char *) event);
1837
1838		if (!sz) {
1839			data->user_stack.size = 0;
1840		} else {
1841			OVERFLOW_CHECK(array, sz, max_size);
1842			data->user_stack.data = (char *)array;
1843			array = (void *)array + sz;
1844			OVERFLOW_CHECK_u64(array);
1845			data->user_stack.size = *array++;
1846			if (WARN_ONCE(data->user_stack.size > sz,
1847				      "user stack dump failure\n"))
1848				return -EFAULT;
1849		}
1850	}
1851
1852	data->weight = 0;
1853	if (type & PERF_SAMPLE_WEIGHT) {
1854		OVERFLOW_CHECK_u64(array);
1855		data->weight = *array;
1856		array++;
1857	}
1858
1859	data->data_src = PERF_MEM_DATA_SRC_NONE;
1860	if (type & PERF_SAMPLE_DATA_SRC) {
1861		OVERFLOW_CHECK_u64(array);
1862		data->data_src = *array;
1863		array++;
1864	}
1865
1866	data->transaction = 0;
1867	if (type & PERF_SAMPLE_TRANSACTION) {
1868		OVERFLOW_CHECK_u64(array);
1869		data->transaction = *array;
1870		array++;
1871	}
1872
1873	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1874	if (type & PERF_SAMPLE_REGS_INTR) {
1875		OVERFLOW_CHECK_u64(array);
1876		data->intr_regs.abi = *array;
1877		array++;
1878
1879		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1880			u64 mask = evsel->attr.sample_regs_intr;
1881
1882			sz = hweight_long(mask) * sizeof(u64);
1883			OVERFLOW_CHECK(array, sz, max_size);
1884			data->intr_regs.mask = mask;
1885			data->intr_regs.regs = (u64 *)array;
1886			array = (void *)array + sz;
1887		}
1888	}
1889
1890	return 0;
1891}
1892
1893size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1894				     u64 read_format)
1895{
1896	size_t sz, result = sizeof(struct sample_event);
1897
1898	if (type & PERF_SAMPLE_IDENTIFIER)
1899		result += sizeof(u64);
1900
1901	if (type & PERF_SAMPLE_IP)
1902		result += sizeof(u64);
1903
1904	if (type & PERF_SAMPLE_TID)
1905		result += sizeof(u64);
1906
1907	if (type & PERF_SAMPLE_TIME)
1908		result += sizeof(u64);
1909
1910	if (type & PERF_SAMPLE_ADDR)
1911		result += sizeof(u64);
1912
1913	if (type & PERF_SAMPLE_ID)
1914		result += sizeof(u64);
1915
1916	if (type & PERF_SAMPLE_STREAM_ID)
1917		result += sizeof(u64);
1918
1919	if (type & PERF_SAMPLE_CPU)
1920		result += sizeof(u64);
1921
1922	if (type & PERF_SAMPLE_PERIOD)
1923		result += sizeof(u64);
1924
1925	if (type & PERF_SAMPLE_READ) {
1926		result += sizeof(u64);
1927		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1928			result += sizeof(u64);
1929		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1930			result += sizeof(u64);
1931		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1932		if (read_format & PERF_FORMAT_GROUP) {
1933			sz = sample->read.group.nr *
1934			     sizeof(struct sample_read_value);
1935			result += sz;
1936		} else {
1937			result += sizeof(u64);
1938		}
1939	}
1940
1941	if (type & PERF_SAMPLE_CALLCHAIN) {
1942		sz = (sample->callchain->nr + 1) * sizeof(u64);
1943		result += sz;
1944	}
1945
1946	if (type & PERF_SAMPLE_RAW) {
1947		result += sizeof(u32);
1948		result += sample->raw_size;
1949	}
1950
1951	if (type & PERF_SAMPLE_BRANCH_STACK) {
1952		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1953		sz += sizeof(u64);
1954		result += sz;
1955	}
1956
1957	if (type & PERF_SAMPLE_REGS_USER) {
1958		if (sample->user_regs.abi) {
1959			result += sizeof(u64);
1960			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1961			result += sz;
1962		} else {
1963			result += sizeof(u64);
1964		}
1965	}
1966
1967	if (type & PERF_SAMPLE_STACK_USER) {
1968		sz = sample->user_stack.size;
1969		result += sizeof(u64);
1970		if (sz) {
1971			result += sz;
1972			result += sizeof(u64);
1973		}
1974	}
1975
1976	if (type & PERF_SAMPLE_WEIGHT)
1977		result += sizeof(u64);
1978
1979	if (type & PERF_SAMPLE_DATA_SRC)
1980		result += sizeof(u64);
1981
1982	if (type & PERF_SAMPLE_TRANSACTION)
1983		result += sizeof(u64);
1984
1985	if (type & PERF_SAMPLE_REGS_INTR) {
1986		if (sample->intr_regs.abi) {
1987			result += sizeof(u64);
1988			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1989			result += sz;
1990		} else {
1991			result += sizeof(u64);
1992		}
1993	}
1994
1995	return result;
1996}
1997
1998int perf_event__synthesize_sample(union perf_event *event, u64 type,
1999				  u64 read_format,
2000				  const struct perf_sample *sample,
2001				  bool swapped)
2002{
2003	u64 *array;
2004	size_t sz;
2005	/*
2006	 * used for cross-endian analysis. See git commit 65014ab3
2007	 * for why this goofiness is needed.
2008	 */
2009	union u64_swap u;
2010
2011	array = event->sample.array;
2012
2013	if (type & PERF_SAMPLE_IDENTIFIER) {
2014		*array = sample->id;
2015		array++;
2016	}
2017
2018	if (type & PERF_SAMPLE_IP) {
2019		*array = sample->ip;
2020		array++;
2021	}
2022
2023	if (type & PERF_SAMPLE_TID) {
2024		u.val32[0] = sample->pid;
2025		u.val32[1] = sample->tid;
2026		if (swapped) {
2027			/*
2028			 * Inverse of what is done in perf_evsel__parse_sample
2029			 */
2030			u.val32[0] = bswap_32(u.val32[0]);
2031			u.val32[1] = bswap_32(u.val32[1]);
2032			u.val64 = bswap_64(u.val64);
2033		}
2034
2035		*array = u.val64;
2036		array++;
2037	}
 
2038
2039	if (type & PERF_SAMPLE_TIME) {
2040		*array = sample->time;
2041		array++;
2042	}
2043
2044	if (type & PERF_SAMPLE_ADDR) {
2045		*array = sample->addr;
2046		array++;
2047	}
2048
2049	if (type & PERF_SAMPLE_ID) {
2050		*array = sample->id;
2051		array++;
2052	}
2053
2054	if (type & PERF_SAMPLE_STREAM_ID) {
2055		*array = sample->stream_id;
2056		array++;
2057	}
2058
2059	if (type & PERF_SAMPLE_CPU) {
2060		u.val32[0] = sample->cpu;
2061		if (swapped) {
2062			/*
2063			 * Inverse of what is done in perf_evsel__parse_sample
2064			 */
2065			u.val32[0] = bswap_32(u.val32[0]);
2066			u.val64 = bswap_64(u.val64);
2067		}
2068		*array = u.val64;
2069		array++;
2070	}
2071
2072	if (type & PERF_SAMPLE_PERIOD) {
2073		*array = sample->period;
2074		array++;
2075	}
2076
2077	if (type & PERF_SAMPLE_READ) {
2078		if (read_format & PERF_FORMAT_GROUP)
2079			*array = sample->read.group.nr;
2080		else
2081			*array = sample->read.one.value;
2082		array++;
2083
2084		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2085			*array = sample->read.time_enabled;
2086			array++;
2087		}
2088
2089		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2090			*array = sample->read.time_running;
2091			array++;
2092		}
2093
2094		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2095		if (read_format & PERF_FORMAT_GROUP) {
2096			sz = sample->read.group.nr *
2097			     sizeof(struct sample_read_value);
2098			memcpy(array, sample->read.group.values, sz);
2099			array = (void *)array + sz;
2100		} else {
2101			*array = sample->read.one.id;
2102			array++;
2103		}
2104	}
2105
2106	if (type & PERF_SAMPLE_CALLCHAIN) {
2107		sz = (sample->callchain->nr + 1) * sizeof(u64);
2108		memcpy(array, sample->callchain, sz);
2109		array = (void *)array + sz;
2110	}
2111
2112	if (type & PERF_SAMPLE_RAW) {
2113		u.val32[0] = sample->raw_size;
2114		if (WARN_ONCE(swapped,
2115			      "Endianness of raw data not corrected!\n")) {
2116			/*
2117			 * Inverse of what is done in perf_evsel__parse_sample
2118			 */
2119			u.val32[0] = bswap_32(u.val32[0]);
2120			u.val32[1] = bswap_32(u.val32[1]);
2121			u.val64 = bswap_64(u.val64);
2122		}
2123		*array = u.val64;
2124		array = (void *)array + sizeof(u32);
2125
2126		memcpy(array, sample->raw_data, sample->raw_size);
2127		array = (void *)array + sample->raw_size;
2128	}
 
2129
2130	if (type & PERF_SAMPLE_BRANCH_STACK) {
2131		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2132		sz += sizeof(u64);
2133		memcpy(array, sample->branch_stack, sz);
2134		array = (void *)array + sz;
2135	}
2136
2137	if (type & PERF_SAMPLE_REGS_USER) {
2138		if (sample->user_regs.abi) {
2139			*array++ = sample->user_regs.abi;
2140			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2141			memcpy(array, sample->user_regs.regs, sz);
2142			array = (void *)array + sz;
2143		} else {
2144			*array++ = 0;
2145		}
2146	}
2147
2148	if (type & PERF_SAMPLE_STACK_USER) {
2149		sz = sample->user_stack.size;
2150		*array++ = sz;
2151		if (sz) {
2152			memcpy(array, sample->user_stack.data, sz);
2153			array = (void *)array + sz;
2154			*array++ = sz;
2155		}
2156	}
2157
2158	if (type & PERF_SAMPLE_WEIGHT) {
2159		*array = sample->weight;
2160		array++;
2161	}
2162
2163	if (type & PERF_SAMPLE_DATA_SRC) {
2164		*array = sample->data_src;
2165		array++;
2166	}
2167
2168	if (type & PERF_SAMPLE_TRANSACTION) {
2169		*array = sample->transaction;
2170		array++;
2171	}
2172
2173	if (type & PERF_SAMPLE_REGS_INTR) {
2174		if (sample->intr_regs.abi) {
2175			*array++ = sample->intr_regs.abi;
2176			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2177			memcpy(array, sample->intr_regs.regs, sz);
2178			array = (void *)array + sz;
2179		} else {
2180			*array++ = 0;
2181		}
2182	}
2183
2184	return 0;
 
 
 
2185}
2186
2187struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2188{
2189	return pevent_find_field(evsel->tp_format, name);
2190}
2191
2192void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2193			 const char *name)
2194{
2195	struct format_field *field = perf_evsel__field(evsel, name);
2196	int offset;
2197
2198	if (!field)
2199		return NULL;
2200
2201	offset = field->offset;
2202
2203	if (field->flags & FIELD_IS_DYNAMIC) {
2204		offset = *(int *)(sample->raw_data + field->offset);
2205		offset &= 0xffff;
 
 
2206	}
2207
2208	return sample->raw_data + offset;
2209}
2210
2211u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2212		       const char *name)
2213{
2214	struct format_field *field = perf_evsel__field(evsel, name);
2215	void *ptr;
2216	u64 value;
2217
2218	if (!field)
2219		return 0;
2220
2221	ptr = sample->raw_data + field->offset;
2222
2223	switch (field->size) {
2224	case 1:
2225		return *(u8 *)ptr;
2226	case 2:
2227		value = *(u16 *)ptr;
2228		break;
2229	case 4:
2230		value = *(u32 *)ptr;
2231		break;
2232	case 8:
2233		memcpy(&value, ptr, sizeof(u64));
2234		break;
2235	default:
2236		return 0;
2237	}
2238
2239	if (!evsel->needs_swap)
2240		return value;
2241
2242	switch (field->size) {
2243	case 2:
2244		return bswap_16(value);
2245	case 4:
2246		return bswap_32(value);
2247	case 8:
2248		return bswap_64(value);
2249	default:
2250		return 0;
2251	}
2252
2253	return 0;
2254}
2255
2256static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2257{
2258	va_list args;
2259	int ret = 0;
2260
2261	if (!*first) {
2262		ret += fprintf(fp, ",");
2263	} else {
2264		ret += fprintf(fp, ":");
2265		*first = false;
2266	}
2267
2268	va_start(args, fmt);
2269	ret += vfprintf(fp, fmt, args);
2270	va_end(args);
2271	return ret;
2272}
2273
2274static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2275{
2276	return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2277}
2278
2279int perf_evsel__fprintf(struct perf_evsel *evsel,
2280			struct perf_attr_details *details, FILE *fp)
2281{
2282	bool first = true;
2283	int printed = 0;
2284
2285	if (details->event_group) {
2286		struct perf_evsel *pos;
2287
2288		if (!perf_evsel__is_group_leader(evsel))
2289			return 0;
2290
2291		if (evsel->nr_members > 1)
2292			printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2293
2294		printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2295		for_each_group_member(pos, evsel)
2296			printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2297
2298		if (evsel->nr_members > 1)
2299			printed += fprintf(fp, "}");
2300		goto out;
2301	}
2302
2303	printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2304
2305	if (details->verbose) {
2306		printed += perf_event_attr__fprintf(fp, &evsel->attr,
2307						    __print_attr__fprintf, &first);
2308	} else if (details->freq) {
2309		const char *term = "sample_freq";
2310
2311		if (!evsel->attr.freq)
2312			term = "sample_period";
2313
2314		printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2315					 term, (u64)evsel->attr.sample_freq);
2316	}
2317
2318	if (details->trace_fields) {
2319		struct format_field *field;
2320
2321		if (evsel->attr.type != PERF_TYPE_TRACEPOINT) {
2322			printed += comma_fprintf(fp, &first, " (not a tracepoint)");
2323			goto out;
2324		}
2325
2326		field = evsel->tp_format->format.fields;
2327		if (field == NULL) {
2328			printed += comma_fprintf(fp, &first, " (no trace field)");
2329			goto out;
2330		}
2331
2332		printed += comma_fprintf(fp, &first, " trace_fields: %s", field->name);
2333
2334		field = field->next;
2335		while (field) {
2336			printed += comma_fprintf(fp, &first, "%s", field->name);
2337			field = field->next;
2338		}
2339	}
2340out:
2341	fputc('\n', fp);
2342	return ++printed;
2343}
2344
2345bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2346			  char *msg, size_t msgsize)
 
 
2347{
2348	int paranoid;
2349
2350	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2351	    evsel->attr.type   == PERF_TYPE_HARDWARE &&
2352	    evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2353		/*
2354		 * If it's cycles then fall back to hrtimer based
2355		 * cpu-clock-tick sw counter, which is always available even if
2356		 * no PMU support.
2357		 *
2358		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2359		 * b0a873e).
2360		 */
2361		scnprintf(msg, msgsize, "%s",
2362"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2363
2364		evsel->attr.type   = PERF_TYPE_SOFTWARE;
2365		evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
 
 
2366
2367		zfree(&evsel->name);
2368		return true;
2369	} else if (err == EACCES && !evsel->attr.exclude_kernel &&
2370		   (paranoid = perf_event_paranoid()) > 1) {
2371		const char *name = perf_evsel__name(evsel);
2372		char *new_name;
 
 
 
 
 
 
 
 
 
 
2373
2374		if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2375			return false;
2376
2377		if (evsel->name)
2378			free(evsel->name);
2379		evsel->name = new_name;
2380		scnprintf(msg, msgsize,
2381"kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2382		evsel->attr.exclude_kernel = 1;
 
 
2383
2384		return true;
2385	}
2386
2387	return false;
2388}
2389
2390int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2391			      int err, char *msg, size_t size)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2392{
2393	char sbuf[STRERR_BUFSIZE];
 
 
2394
2395	switch (err) {
2396	case EPERM:
2397	case EACCES:
2398		return scnprintf(msg, size,
2399		 "You may not have permission to collect %sstats.\n\n"
2400		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2401		 "which controls use of the performance events system by\n"
2402		 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2403		 "The current value is %d:\n\n"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2404		 "  -1: Allow use of (almost) all events by all users\n"
2405		 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2406		 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2407		 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2408				 target->system_wide ? "system-wide " : "",
2409				 perf_event_paranoid());
 
 
2410	case ENOENT:
2411		return scnprintf(msg, size, "The %s event is not supported.",
2412				 perf_evsel__name(evsel));
2413	case EMFILE:
2414		return scnprintf(msg, size, "%s",
2415			 "Too many events are opened.\n"
2416			 "Probably the maximum number of open file descriptors has been reached.\n"
2417			 "Hint: Try again after reducing the number of events.\n"
2418			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
 
 
 
 
 
 
 
 
2419	case ENODEV:
2420		if (target->cpu_list)
2421			return scnprintf(msg, size, "%s",
2422	 "No such device - did you specify an out-of-range profile CPU?\n");
2423		break;
2424	case EOPNOTSUPP:
2425		if (evsel->attr.precise_ip)
 
 
 
 
 
 
 
 
 
 
 
 
2426			return scnprintf(msg, size, "%s",
2427	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2428#if defined(__i386__) || defined(__x86_64__)
2429		if (evsel->attr.type == PERF_TYPE_HARDWARE)
2430			return scnprintf(msg, size, "%s",
2431	"No hardware sampling interrupt available.\n"
2432	"No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2433#endif
2434		break;
2435	case EBUSY:
2436		if (find_process("oprofiled"))
2437			return scnprintf(msg, size,
2438	"The PMU counters are busy/taken by another profiler.\n"
2439	"We found oprofile daemon running, please stop it and try again.");
2440		break;
2441	case EINVAL:
 
 
 
 
 
 
2442		if (perf_missing_features.clockid)
2443			return scnprintf(msg, size, "clockid feature not supported.");
2444		if (perf_missing_features.clockid_wrong)
2445			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
 
 
 
 
 
 
 
2446		break;
 
 
 
2447	default:
2448		break;
2449	}
2450
 
 
 
 
2451	return scnprintf(msg, size,
2452	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2453	"/bin/dmesg may provide additional information.\n"
2454	"No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2455			 err, strerror_r(err, sbuf, sizeof(sbuf)),
2456			 perf_evsel__name(evsel));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2457}
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4 *
   5 * Parts came from builtin-{top,stat,record}.c, see those files for further
   6 * copyright notes.
 
 
   7 */
   8
   9#include <byteswap.h>
  10#include <errno.h>
  11#include <inttypes.h>
  12#include <linux/bitops.h>
  13#include <api/fs/fs.h>
  14#include <api/fs/tracing_path.h>
 
  15#include <linux/hw_breakpoint.h>
  16#include <linux/perf_event.h>
  17#include <linux/compiler.h>
  18#include <linux/err.h>
  19#include <linux/zalloc.h>
  20#include <sys/ioctl.h>
  21#include <sys/resource.h>
  22#include <sys/types.h>
  23#include <dirent.h>
  24#include <stdlib.h>
  25#include <perf/evsel.h>
  26#include "asm/bug.h"
  27#include "bpf_counter.h"
  28#include "callchain.h"
  29#include "cgroup.h"
  30#include "counts.h"
  31#include "event.h"
  32#include "evsel.h"
  33#include "util/env.h"
  34#include "util/evsel_config.h"
  35#include "util/evsel_fprintf.h"
  36#include "evlist.h"
  37#include <perf/cpumap.h>
 
  38#include "thread_map.h"
  39#include "target.h"
  40#include "perf_regs.h"
  41#include "record.h"
  42#include "debug.h"
  43#include "trace-event.h"
  44#include "stat.h"
  45#include "string2.h"
  46#include "memswap.h"
  47#include "util.h"
  48#include "util/hashmap.h"
  49#include "off_cpu.h"
  50#include "pmu.h"
  51#include "pmus.h"
  52#include "rlimit.h"
  53#include "../perf-sys.h"
  54#include "util/parse-branch-options.h"
  55#include "util/bpf-filter.h"
  56#include <internal/xyarray.h>
  57#include <internal/lib.h>
  58#include <internal/threadmap.h>
  59
  60#include <linux/ctype.h>
  61
  62#ifdef HAVE_LIBTRACEEVENT
  63#include <traceevent/event-parse.h>
  64#endif
  65
  66struct perf_missing_features perf_missing_features;
 
 
  67
  68static clockid_t clockid;
  69
  70static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  71	NULL,
  72	"duration_time",
  73	"user_time",
  74	"system_time",
  75};
  76
  77const char *perf_tool_event__to_str(enum perf_tool_event ev)
  78{
  79	if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  80		return perf_tool_event__tool_names[ev];
  81
  82	return NULL;
  83}
  84
  85enum perf_tool_event perf_tool_event__from_str(const char *str)
  86{
  87	int i;
  88
  89	perf_tool_event__for_each_event(i) {
  90		if (!strcmp(str, perf_tool_event__tool_names[i]))
  91			return i;
  92	}
  93	return PERF_TOOL_NONE;
  94}
  95
  96
  97static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  98{
  99	return 0;
 100}
 101
 102void __weak test_attr__ready(void) { }
 103
 104static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 105{
 106}
 107
 108static struct {
 109	size_t	size;
 110	int	(*init)(struct evsel *evsel);
 111	void	(*fini)(struct evsel *evsel);
 112} perf_evsel__object = {
 113	.size = sizeof(struct evsel),
 114	.init = evsel__no_extra_init,
 115	.fini = evsel__no_extra_fini,
 116};
 117
 118int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 119			 void (*fini)(struct evsel *evsel))
 
 120{
 121
 122	if (object_size == 0)
 123		goto set_methods;
 124
 125	if (perf_evsel__object.size > object_size)
 126		return -EINVAL;
 127
 128	perf_evsel__object.size = object_size;
 129
 130set_methods:
 131	if (init != NULL)
 132		perf_evsel__object.init = init;
 133
 134	if (fini != NULL)
 135		perf_evsel__object.fini = fini;
 136
 137	return 0;
 138}
 139
 140#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 141
 142int __evsel__sample_size(u64 sample_type)
 143{
 144	u64 mask = sample_type & PERF_SAMPLE_MASK;
 145	int size = 0;
 146	int i;
 147
 148	for (i = 0; i < 64; i++) {
 149		if (mask & (1ULL << i))
 150			size++;
 151	}
 152
 153	size *= sizeof(u64);
 154
 155	return size;
 156}
 157
 158/**
 159 * __perf_evsel__calc_id_pos - calculate id_pos.
 160 * @sample_type: sample type
 161 *
 162 * This function returns the position of the event id (PERF_SAMPLE_ID or
 163 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 164 * perf_record_sample.
 165 */
 166static int __perf_evsel__calc_id_pos(u64 sample_type)
 167{
 168	int idx = 0;
 169
 170	if (sample_type & PERF_SAMPLE_IDENTIFIER)
 171		return 0;
 172
 173	if (!(sample_type & PERF_SAMPLE_ID))
 174		return -1;
 175
 176	if (sample_type & PERF_SAMPLE_IP)
 177		idx += 1;
 178
 179	if (sample_type & PERF_SAMPLE_TID)
 180		idx += 1;
 181
 182	if (sample_type & PERF_SAMPLE_TIME)
 183		idx += 1;
 184
 185	if (sample_type & PERF_SAMPLE_ADDR)
 186		idx += 1;
 187
 188	return idx;
 189}
 190
 191/**
 192 * __perf_evsel__calc_is_pos - calculate is_pos.
 193 * @sample_type: sample type
 194 *
 195 * This function returns the position (counting backwards) of the event id
 196 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 197 * sample_id_all is used there is an id sample appended to non-sample events.
 198 */
 199static int __perf_evsel__calc_is_pos(u64 sample_type)
 200{
 201	int idx = 1;
 202
 203	if (sample_type & PERF_SAMPLE_IDENTIFIER)
 204		return 1;
 205
 206	if (!(sample_type & PERF_SAMPLE_ID))
 207		return -1;
 208
 209	if (sample_type & PERF_SAMPLE_CPU)
 210		idx += 1;
 211
 212	if (sample_type & PERF_SAMPLE_STREAM_ID)
 213		idx += 1;
 214
 215	return idx;
 216}
 217
 218void evsel__calc_id_pos(struct evsel *evsel)
 219{
 220	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 221	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 222}
 223
 224void __evsel__set_sample_bit(struct evsel *evsel,
 225				  enum perf_event_sample_format bit)
 226{
 227	if (!(evsel->core.attr.sample_type & bit)) {
 228		evsel->core.attr.sample_type |= bit;
 229		evsel->sample_size += sizeof(u64);
 230		evsel__calc_id_pos(evsel);
 231	}
 232}
 233
 234void __evsel__reset_sample_bit(struct evsel *evsel,
 235				    enum perf_event_sample_format bit)
 236{
 237	if (evsel->core.attr.sample_type & bit) {
 238		evsel->core.attr.sample_type &= ~bit;
 239		evsel->sample_size -= sizeof(u64);
 240		evsel__calc_id_pos(evsel);
 241	}
 242}
 243
 244void evsel__set_sample_id(struct evsel *evsel,
 245			       bool can_sample_identifier)
 246{
 247	if (can_sample_identifier) {
 248		evsel__reset_sample_bit(evsel, ID);
 249		evsel__set_sample_bit(evsel, IDENTIFIER);
 250	} else {
 251		evsel__set_sample_bit(evsel, ID);
 252	}
 253	evsel->core.attr.read_format |= PERF_FORMAT_ID;
 254}
 255
 256/**
 257 * evsel__is_function_event - Return whether given evsel is a function
 258 * trace event
 259 *
 260 * @evsel - evsel selector to be tested
 261 *
 262 * Return %true if event is function trace event
 263 */
 264bool evsel__is_function_event(struct evsel *evsel)
 265{
 266#define FUNCTION_EVENT "ftrace:function"
 267
 268	return evsel->name &&
 269	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 270
 271#undef FUNCTION_EVENT
 272}
 273
 274void evsel__init(struct evsel *evsel,
 275		 struct perf_event_attr *attr, int idx)
 276{
 277	perf_evsel__init(&evsel->core, attr, idx);
 278	evsel->tracking	   = !idx;
 279	evsel->unit	   = strdup("");
 
 
 280	evsel->scale	   = 1.0;
 281	evsel->max_events  = ULONG_MAX;
 282	evsel->evlist	   = NULL;
 283	evsel->bpf_obj	   = NULL;
 284	evsel->bpf_fd	   = -1;
 
 285	INIT_LIST_HEAD(&evsel->config_terms);
 286	INIT_LIST_HEAD(&evsel->bpf_counter_list);
 287	INIT_LIST_HEAD(&evsel->bpf_filters);
 288	perf_evsel__object.init(evsel);
 289	evsel->sample_size = __evsel__sample_size(attr->sample_type);
 290	evsel__calc_id_pos(evsel);
 291	evsel->cmdline_group_boundary = false;
 292	evsel->metric_events = NULL;
 293	evsel->per_pkg_mask  = NULL;
 294	evsel->collect_stat  = false;
 295	evsel->pmu_name      = NULL;
 296	evsel->group_pmu_name = NULL;
 297	evsel->skippable     = false;
 298}
 299
 300struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 301{
 302	struct evsel *evsel = zalloc(perf_evsel__object.size);
 303
 304	if (!evsel)
 305		return NULL;
 306	evsel__init(evsel, attr, idx);
 307
 308	if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 309		evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 310					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 311		evsel->core.attr.sample_period = 1;
 312	}
 313
 314	if (evsel__is_clock(evsel)) {
 315		free((char *)evsel->unit);
 316		evsel->unit = strdup("msec");
 317		evsel->scale = 1e-6;
 318	}
 319
 320	return evsel;
 321}
 322
 323int copy_config_terms(struct list_head *dst, struct list_head *src)
 324{
 325	struct evsel_config_term *pos, *tmp;
 326
 327	list_for_each_entry(pos, src, list) {
 328		tmp = malloc(sizeof(*tmp));
 329		if (tmp == NULL)
 330			return -ENOMEM;
 331
 332		*tmp = *pos;
 333		if (tmp->free_str) {
 334			tmp->val.str = strdup(pos->val.str);
 335			if (tmp->val.str == NULL) {
 336				free(tmp);
 337				return -ENOMEM;
 338			}
 339		}
 340		list_add_tail(&tmp->list, dst);
 341	}
 342	return 0;
 343}
 344
 345static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 346{
 347	return copy_config_terms(&dst->config_terms, &src->config_terms);
 348}
 349
 350/**
 351 * evsel__clone - create a new evsel copied from @orig
 352 * @orig: original evsel
 353 *
 354 * The assumption is that @orig is not configured nor opened yet.
 355 * So we only care about the attributes that can be set while it's parsed.
 356 */
 357struct evsel *evsel__clone(struct evsel *orig)
 358{
 359	struct evsel *evsel;
 360
 361	BUG_ON(orig->core.fd);
 362	BUG_ON(orig->counts);
 363	BUG_ON(orig->priv);
 364	BUG_ON(orig->per_pkg_mask);
 365
 366	/* cannot handle BPF objects for now */
 367	if (orig->bpf_obj)
 368		return NULL;
 369
 370	evsel = evsel__new(&orig->core.attr);
 371	if (evsel == NULL)
 372		return NULL;
 373
 374	evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 375	evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 376	evsel->core.threads = perf_thread_map__get(orig->core.threads);
 377	evsel->core.nr_members = orig->core.nr_members;
 378	evsel->core.system_wide = orig->core.system_wide;
 379	evsel->core.requires_cpu = orig->core.requires_cpu;
 380	evsel->core.is_pmu_core = orig->core.is_pmu_core;
 381
 382	if (orig->name) {
 383		evsel->name = strdup(orig->name);
 384		if (evsel->name == NULL)
 385			goto out_err;
 386	}
 387	if (orig->group_name) {
 388		evsel->group_name = strdup(orig->group_name);
 389		if (evsel->group_name == NULL)
 390			goto out_err;
 391	}
 392	if (orig->pmu_name) {
 393		evsel->pmu_name = strdup(orig->pmu_name);
 394		if (evsel->pmu_name == NULL)
 395			goto out_err;
 396	}
 397	if (orig->group_pmu_name) {
 398		evsel->group_pmu_name = strdup(orig->group_pmu_name);
 399		if (evsel->group_pmu_name == NULL)
 400			goto out_err;
 401	}
 402	if (orig->filter) {
 403		evsel->filter = strdup(orig->filter);
 404		if (evsel->filter == NULL)
 405			goto out_err;
 406	}
 407	if (orig->metric_id) {
 408		evsel->metric_id = strdup(orig->metric_id);
 409		if (evsel->metric_id == NULL)
 410			goto out_err;
 411	}
 412	evsel->cgrp = cgroup__get(orig->cgrp);
 413#ifdef HAVE_LIBTRACEEVENT
 414	evsel->tp_format = orig->tp_format;
 415#endif
 416	evsel->handler = orig->handler;
 417	evsel->core.leader = orig->core.leader;
 418
 419	evsel->max_events = orig->max_events;
 420	evsel->tool_event = orig->tool_event;
 421	free((char *)evsel->unit);
 422	evsel->unit = strdup(orig->unit);
 423	if (evsel->unit == NULL)
 424		goto out_err;
 425
 426	evsel->scale = orig->scale;
 427	evsel->snapshot = orig->snapshot;
 428	evsel->per_pkg = orig->per_pkg;
 429	evsel->percore = orig->percore;
 430	evsel->precise_max = orig->precise_max;
 431	evsel->is_libpfm_event = orig->is_libpfm_event;
 432
 433	evsel->exclude_GH = orig->exclude_GH;
 434	evsel->sample_read = orig->sample_read;
 435	evsel->auto_merge_stats = orig->auto_merge_stats;
 436	evsel->collect_stat = orig->collect_stat;
 437	evsel->weak_group = orig->weak_group;
 438	evsel->use_config_name = orig->use_config_name;
 439	evsel->pmu = orig->pmu;
 440
 441	if (evsel__copy_config_terms(evsel, orig) < 0)
 442		goto out_err;
 443
 444	return evsel;
 445
 446out_err:
 447	evsel__delete(evsel);
 448	return NULL;
 449}
 450
 451/*
 452 * Returns pointer with encoded error via <linux/err.h> interface.
 453 */
 454#ifdef HAVE_LIBTRACEEVENT
 455struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 456{
 457	struct evsel *evsel = zalloc(perf_evsel__object.size);
 458	int err = -ENOMEM;
 459
 460	if (evsel == NULL) {
 461		goto out_err;
 462	} else {
 463		struct perf_event_attr attr = {
 464			.type	       = PERF_TYPE_TRACEPOINT,
 465			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 466					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 467		};
 468
 469		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 470			goto out_free;
 471
 472		evsel->tp_format = trace_event__tp_format(sys, name);
 473		if (IS_ERR(evsel->tp_format)) {
 474			err = PTR_ERR(evsel->tp_format);
 475			goto out_free;
 476		}
 477
 478		event_attr_init(&attr);
 479		attr.config = evsel->tp_format->id;
 480		attr.sample_period = 1;
 481		evsel__init(evsel, &attr, idx);
 482	}
 483
 484	return evsel;
 485
 486out_free:
 487	zfree(&evsel->name);
 488	free(evsel);
 489out_err:
 490	return ERR_PTR(err);
 491}
 492#endif
 493
 494const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 495	"cycles",
 496	"instructions",
 497	"cache-references",
 498	"cache-misses",
 499	"branches",
 500	"branch-misses",
 501	"bus-cycles",
 502	"stalled-cycles-frontend",
 503	"stalled-cycles-backend",
 504	"ref-cycles",
 505};
 506
 507char *evsel__bpf_counter_events;
 508
 509bool evsel__match_bpf_counter_events(const char *name)
 510{
 511	int name_len;
 512	bool match;
 513	char *ptr;
 514
 515	if (!evsel__bpf_counter_events)
 516		return false;
 517
 518	ptr = strstr(evsel__bpf_counter_events, name);
 519	name_len = strlen(name);
 520
 521	/* check name matches a full token in evsel__bpf_counter_events */
 522	match = (ptr != NULL) &&
 523		((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 524		((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 525
 526	return match;
 527}
 528
 529static const char *__evsel__hw_name(u64 config)
 530{
 531	if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 532		return evsel__hw_names[config];
 533
 534	return "unknown-hardware";
 535}
 536
 537static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 538{
 539	int colon = 0, r = 0;
 540	struct perf_event_attr *attr = &evsel->core.attr;
 541	bool exclude_guest_default = false;
 542
 543#define MOD_PRINT(context, mod)	do {					\
 544		if (!attr->exclude_##context) {				\
 545			if (!colon) colon = ++r;			\
 546			r += scnprintf(bf + r, size - r, "%c", mod);	\
 547		} } while(0)
 548
 549	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 550		MOD_PRINT(kernel, 'k');
 551		MOD_PRINT(user, 'u');
 552		MOD_PRINT(hv, 'h');
 553		exclude_guest_default = true;
 554	}
 555
 556	if (attr->precise_ip) {
 557		if (!colon)
 558			colon = ++r;
 559		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 560		exclude_guest_default = true;
 561	}
 562
 563	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 564		MOD_PRINT(host, 'H');
 565		MOD_PRINT(guest, 'G');
 566	}
 567#undef MOD_PRINT
 568	if (colon)
 569		bf[colon - 1] = ':';
 570	return r;
 571}
 572
 573int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 574{
 575	return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 
 576}
 577
 578static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 579{
 580	int r = arch_evsel__hw_name(evsel, bf, size);
 581	return r + evsel__add_modifiers(evsel, bf + r, size - r);
 582}
 583
 584const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 585	"cpu-clock",
 586	"task-clock",
 587	"page-faults",
 588	"context-switches",
 589	"cpu-migrations",
 590	"minor-faults",
 591	"major-faults",
 592	"alignment-faults",
 593	"emulation-faults",
 594	"dummy",
 595};
 596
 597static const char *__evsel__sw_name(u64 config)
 598{
 599	if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 600		return evsel__sw_names[config];
 601	return "unknown-software";
 602}
 603
 604static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 605{
 606	int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 607	return r + evsel__add_modifiers(evsel, bf + r, size - r);
 608}
 609
 610static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 611{
 612	return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 613}
 614
 615static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 616{
 617	int r;
 618
 619	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 620
 621	if (type & HW_BREAKPOINT_R)
 622		r += scnprintf(bf + r, size - r, "r");
 623
 624	if (type & HW_BREAKPOINT_W)
 625		r += scnprintf(bf + r, size - r, "w");
 626
 627	if (type & HW_BREAKPOINT_X)
 628		r += scnprintf(bf + r, size - r, "x");
 629
 630	return r;
 631}
 632
 633static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 634{
 635	struct perf_event_attr *attr = &evsel->core.attr;
 636	int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 637	return r + evsel__add_modifiers(evsel, bf + r, size - r);
 638}
 639
 640const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 
 641 { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
 642 { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
 643 { "LLC",	"L2",							},
 644 { "dTLB",	"d-tlb",	"Data-TLB",				},
 645 { "iTLB",	"i-tlb",	"Instruction-TLB",			},
 646 { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
 647 { "node",								},
 648};
 649
 650const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 
 651 { "load",	"loads",	"read",					},
 652 { "store",	"stores",	"write",				},
 653 { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
 654};
 655
 656const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 
 657 { "refs",	"Reference",	"ops",		"access",		},
 658 { "misses",	"miss",							},
 659};
 660
 661#define C(x)		PERF_COUNT_HW_CACHE_##x
 662#define CACHE_READ	(1 << C(OP_READ))
 663#define CACHE_WRITE	(1 << C(OP_WRITE))
 664#define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
 665#define COP(x)		(1 << x)
 666
 667/*
 668 * cache operation stat
 669 * L1I : Read and prefetch only
 670 * ITLB and BPU : Read-only
 671 */
 672static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 673 [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 674 [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
 675 [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 676 [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 677 [C(ITLB)]	= (CACHE_READ),
 678 [C(BPU)]	= (CACHE_READ),
 679 [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 680};
 681
 682bool evsel__is_cache_op_valid(u8 type, u8 op)
 683{
 684	if (evsel__hw_cache_stat[type] & COP(op))
 685		return true;	/* valid */
 686	else
 687		return false;	/* invalid */
 688}
 689
 690int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 
 691{
 692	if (result) {
 693		return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 694				 evsel__hw_cache_op[op][0],
 695				 evsel__hw_cache_result[result][0]);
 696	}
 697
 698	return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 699			 evsel__hw_cache_op[op][1]);
 700}
 701
 702static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 703{
 704	u8 op, result, type = (config >>  0) & 0xff;
 705	const char *err = "unknown-ext-hardware-cache-type";
 706
 707	if (type >= PERF_COUNT_HW_CACHE_MAX)
 708		goto out_err;
 709
 710	op = (config >>  8) & 0xff;
 711	err = "unknown-ext-hardware-cache-op";
 712	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 713		goto out_err;
 714
 715	result = (config >> 16) & 0xff;
 716	err = "unknown-ext-hardware-cache-result";
 717	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 718		goto out_err;
 719
 720	err = "invalid-cache";
 721	if (!evsel__is_cache_op_valid(type, op))
 722		goto out_err;
 723
 724	return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 725out_err:
 726	return scnprintf(bf, size, "%s", err);
 727}
 728
 729static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 730{
 731	int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 732	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 733}
 734
 735static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 736{
 737	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 738	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 739}
 740
 741const char *evsel__name(struct evsel *evsel)
 742{
 743	char bf[128];
 744
 745	if (!evsel)
 746		goto out_unknown;
 747
 748	if (evsel->name)
 749		return evsel->name;
 750
 751	switch (evsel->core.attr.type) {
 752	case PERF_TYPE_RAW:
 753		evsel__raw_name(evsel, bf, sizeof(bf));
 754		break;
 755
 756	case PERF_TYPE_HARDWARE:
 757		evsel__hw_name(evsel, bf, sizeof(bf));
 758		break;
 759
 760	case PERF_TYPE_HW_CACHE:
 761		evsel__hw_cache_name(evsel, bf, sizeof(bf));
 762		break;
 763
 764	case PERF_TYPE_SOFTWARE:
 765		if (evsel__is_tool(evsel))
 766			evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 767		else
 768			evsel__sw_name(evsel, bf, sizeof(bf));
 769		break;
 770
 771	case PERF_TYPE_TRACEPOINT:
 772		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 773		break;
 774
 775	case PERF_TYPE_BREAKPOINT:
 776		evsel__bp_name(evsel, bf, sizeof(bf));
 777		break;
 778
 779	default:
 780		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 781			  evsel->core.attr.type);
 782		break;
 783	}
 784
 785	evsel->name = strdup(bf);
 786
 787	if (evsel->name)
 788		return evsel->name;
 789out_unknown:
 790	return "unknown";
 791}
 792
 793bool evsel__name_is(struct evsel *evsel, const char *name)
 794{
 795	return !strcmp(evsel__name(evsel), name);
 796}
 797
 798const char *evsel__metric_id(const struct evsel *evsel)
 799{
 800	if (evsel->metric_id)
 801		return evsel->metric_id;
 802
 803	if (evsel__is_tool(evsel))
 804		return perf_tool_event__to_str(evsel->tool_event);
 805
 806	return "unknown";
 807}
 808
 809const char *evsel__group_name(struct evsel *evsel)
 810{
 811	return evsel->group_name ?: "anon group";
 812}
 813
 814/*
 815 * Returns the group details for the specified leader,
 816 * with following rules.
 817 *
 818 *  For record -e '{cycles,instructions}'
 819 *    'anon group { cycles:u, instructions:u }'
 820 *
 821 *  For record -e 'cycles,instructions' and report --group
 822 *    'cycles:u, instructions:u'
 823 */
 824int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 825{
 826	int ret = 0;
 827	struct evsel *pos;
 828	const char *group_name = evsel__group_name(evsel);
 829
 830	if (!evsel->forced_leader)
 831		ret = scnprintf(buf, size, "%s { ", group_name);
 832
 833	ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 
 834
 835	for_each_group_member(pos, evsel)
 836		ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 
 837
 838	if (!evsel->forced_leader)
 839		ret += scnprintf(buf + ret, size - ret, " }");
 840
 841	return ret;
 842}
 843
 844static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 845				      struct callchain_param *param)
 
 
 846{
 847	bool function = evsel__is_function_event(evsel);
 848	struct perf_event_attr *attr = &evsel->core.attr;
 849	const char *arch = perf_env__arch(evsel__env(evsel));
 850
 851	evsel__set_sample_bit(evsel, CALLCHAIN);
 852
 853	attr->sample_max_stack = param->max_stack;
 854
 855	if (opts->kernel_callchains)
 856		attr->exclude_callchain_user = 1;
 857	if (opts->user_callchains)
 858		attr->exclude_callchain_kernel = 1;
 859	if (param->record_mode == CALLCHAIN_LBR) {
 860		if (!opts->branch_stack) {
 861			if (attr->exclude_user) {
 862				pr_warning("LBR callstack option is only available "
 863					   "to get user callchain information. "
 864					   "Falling back to framepointers.\n");
 865			} else {
 866				evsel__set_sample_bit(evsel, BRANCH_STACK);
 867				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 868							PERF_SAMPLE_BRANCH_CALL_STACK |
 869							PERF_SAMPLE_BRANCH_NO_CYCLES |
 870							PERF_SAMPLE_BRANCH_NO_FLAGS |
 871							PERF_SAMPLE_BRANCH_HW_INDEX;
 872			}
 873		} else
 874			 pr_warning("Cannot use LBR callstack with branch stack. "
 875				    "Falling back to framepointers.\n");
 876	}
 877
 878	if (param->record_mode == CALLCHAIN_DWARF) {
 879		if (!function) {
 880			evsel__set_sample_bit(evsel, REGS_USER);
 881			evsel__set_sample_bit(evsel, STACK_USER);
 882			if (opts->sample_user_regs &&
 883			    DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) {
 884				attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch);
 885				pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 886					   "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 887					   "so the minimal registers set (IP, SP) is explicitly forced.\n");
 888			} else {
 889				attr->sample_regs_user |= arch__user_reg_mask();
 890			}
 891			attr->sample_stack_user = param->dump_size;
 892			attr->exclude_callchain_user = 1;
 893		} else {
 894			pr_info("Cannot use DWARF unwind for function trace event,"
 895				" falling back to framepointers.\n");
 896		}
 897	}
 898
 899	if (function) {
 900		pr_info("Disabling user space callchains for function trace event.\n");
 901		attr->exclude_callchain_user = 1;
 902	}
 903}
 904
 905void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 906			     struct callchain_param *param)
 
 907{
 908	if (param->enabled)
 909		return __evsel__config_callchain(evsel, opts, param);
 910}
 911
 912static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 913{
 914	struct perf_event_attr *attr = &evsel->core.attr;
 915
 916	evsel__reset_sample_bit(evsel, CALLCHAIN);
 917	if (param->record_mode == CALLCHAIN_LBR) {
 918		evsel__reset_sample_bit(evsel, BRANCH_STACK);
 919		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 920					      PERF_SAMPLE_BRANCH_CALL_STACK |
 921					      PERF_SAMPLE_BRANCH_HW_INDEX);
 922	}
 923	if (param->record_mode == CALLCHAIN_DWARF) {
 924		evsel__reset_sample_bit(evsel, REGS_USER);
 925		evsel__reset_sample_bit(evsel, STACK_USER);
 926	}
 927}
 928
 929static void evsel__apply_config_terms(struct evsel *evsel,
 930				      struct record_opts *opts, bool track)
 931{
 932	struct evsel_config_term *term;
 933	struct list_head *config_terms = &evsel->config_terms;
 934	struct perf_event_attr *attr = &evsel->core.attr;
 
 
 
 
 935	/* callgraph default */
 936	struct callchain_param param = {
 937		.record_mode = callchain_param.record_mode,
 938	};
 939	u32 dump_size = 0;
 940	int max_stack = 0;
 941	const char *callgraph_buf = NULL;
 942
 943	list_for_each_entry(term, config_terms, list) {
 944		switch (term->type) {
 945		case EVSEL__CONFIG_TERM_PERIOD:
 946			if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 947				attr->sample_period = term->val.period;
 948				attr->freq = 0;
 949				evsel__reset_sample_bit(evsel, PERIOD);
 950			}
 951			break;
 952		case EVSEL__CONFIG_TERM_FREQ:
 953			if (!(term->weak && opts->user_freq != UINT_MAX)) {
 954				attr->sample_freq = term->val.freq;
 955				attr->freq = 1;
 956				evsel__set_sample_bit(evsel, PERIOD);
 957			}
 958			break;
 959		case EVSEL__CONFIG_TERM_TIME:
 960			if (term->val.time)
 961				evsel__set_sample_bit(evsel, TIME);
 962			else
 963				evsel__reset_sample_bit(evsel, TIME);
 964			break;
 965		case EVSEL__CONFIG_TERM_CALLGRAPH:
 966			callgraph_buf = term->val.str;
 967			break;
 968		case EVSEL__CONFIG_TERM_BRANCH:
 969			if (term->val.str && strcmp(term->val.str, "no")) {
 970				evsel__set_sample_bit(evsel, BRANCH_STACK);
 971				parse_branch_str(term->val.str,
 972						 &attr->branch_sample_type);
 973			} else
 974				evsel__reset_sample_bit(evsel, BRANCH_STACK);
 975			break;
 976		case EVSEL__CONFIG_TERM_STACK_USER:
 977			dump_size = term->val.stack_user;
 978			break;
 979		case EVSEL__CONFIG_TERM_MAX_STACK:
 980			max_stack = term->val.max_stack;
 981			break;
 982		case EVSEL__CONFIG_TERM_MAX_EVENTS:
 983			evsel->max_events = term->val.max_events;
 984			break;
 985		case EVSEL__CONFIG_TERM_INHERIT:
 986			/*
 987			 * attr->inherit should has already been set by
 988			 * evsel__config. If user explicitly set
 989			 * inherit using config terms, override global
 990			 * opt->no_inherit setting.
 991			 */
 992			attr->inherit = term->val.inherit ? 1 : 0;
 993			break;
 994		case EVSEL__CONFIG_TERM_OVERWRITE:
 995			attr->write_backward = term->val.overwrite ? 1 : 0;
 996			break;
 997		case EVSEL__CONFIG_TERM_DRV_CFG:
 998			break;
 999		case EVSEL__CONFIG_TERM_PERCORE:
1000			break;
1001		case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1002			attr->aux_output = term->val.aux_output ? 1 : 0;
1003			break;
1004		case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1005			/* Already applied by auxtrace */
1006			break;
1007		case EVSEL__CONFIG_TERM_CFG_CHG:
1008			break;
1009		default:
1010			break;
1011		}
1012	}
1013
1014	/* User explicitly set per-event callgraph, clear the old setting and reset. */
1015	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1016		bool sample_address = false;
1017
1018		if (max_stack) {
1019			param.max_stack = max_stack;
1020			if (callgraph_buf == NULL)
1021				callgraph_buf = "fp";
1022		}
1023
1024		/* parse callgraph parameters */
1025		if (callgraph_buf != NULL) {
1026			if (!strcmp(callgraph_buf, "no")) {
1027				param.enabled = false;
1028				param.record_mode = CALLCHAIN_NONE;
1029			} else {
1030				param.enabled = true;
1031				if (parse_callchain_record(callgraph_buf, &param)) {
1032					pr_err("per-event callgraph setting for %s failed. "
1033					       "Apply callgraph global setting for it\n",
1034					       evsel->name);
1035					return;
1036				}
1037				if (param.record_mode == CALLCHAIN_DWARF)
1038					sample_address = true;
1039			}
1040		}
1041		if (dump_size > 0) {
1042			dump_size = round_up(dump_size, sizeof(u64));
1043			param.dump_size = dump_size;
1044		}
1045
1046		/* If global callgraph set, clear it */
1047		if (callchain_param.enabled)
1048			evsel__reset_callgraph(evsel, &callchain_param);
1049
1050		/* set perf-event callgraph */
1051		if (param.enabled) {
1052			if (sample_address) {
1053				evsel__set_sample_bit(evsel, ADDR);
1054				evsel__set_sample_bit(evsel, DATA_SRC);
1055				evsel->core.attr.mmap_data = track;
1056			}
1057			evsel__config_callchain(evsel, opts, &param);
1058		}
1059	}
1060}
1061
1062struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1063{
1064	struct evsel_config_term *term, *found_term = NULL;
1065
1066	list_for_each_entry(term, &evsel->config_terms, list) {
1067		if (term->type == type)
1068			found_term = term;
1069	}
1070
1071	return found_term;
1072}
1073
1074void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1075{
1076	evsel__set_sample_bit(evsel, WEIGHT);
1077}
1078
1079void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1080				    struct perf_event_attr *attr __maybe_unused)
1081{
1082}
1083
1084static void evsel__set_default_freq_period(struct record_opts *opts,
1085					   struct perf_event_attr *attr)
1086{
1087	if (opts->freq) {
1088		attr->freq = 1;
1089		attr->sample_freq = opts->freq;
1090	} else {
1091		attr->sample_period = opts->default_interval;
1092	}
1093}
1094
1095static bool evsel__is_offcpu_event(struct evsel *evsel)
1096{
1097	return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1098}
1099
1100/*
1101 * The enable_on_exec/disabled value strategy:
1102 *
1103 *  1) For any type of traced program:
1104 *    - all independent events and group leaders are disabled
1105 *    - all group members are enabled
1106 *
1107 *     Group members are ruled by group leaders. They need to
1108 *     be enabled, because the group scheduling relies on that.
1109 *
1110 *  2) For traced programs executed by perf:
1111 *     - all independent events and group leaders have
1112 *       enable_on_exec set
1113 *     - we don't specifically enable or disable any event during
1114 *       the record command
1115 *
1116 *     Independent events and group leaders are initially disabled
1117 *     and get enabled by exec. Group members are ruled by group
1118 *     leaders as stated in 1).
1119 *
1120 *  3) For traced programs attached by perf (pid/tid):
1121 *     - we specifically enable or disable all events during
1122 *       the record command
1123 *
1124 *     When attaching events to already running traced we
1125 *     enable/disable events specifically, as there's no
1126 *     initial traced exec call.
1127 */
1128void evsel__config(struct evsel *evsel, struct record_opts *opts,
1129		   struct callchain_param *callchain)
1130{
1131	struct evsel *leader = evsel__leader(evsel);
1132	struct perf_event_attr *attr = &evsel->core.attr;
1133	int track = evsel->tracking;
1134	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1135
1136	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1137	attr->inherit	    = !opts->no_inherit;
1138	attr->write_backward = opts->overwrite ? 1 : 0;
1139	attr->read_format   = PERF_FORMAT_LOST;
1140
1141	evsel__set_sample_bit(evsel, IP);
1142	evsel__set_sample_bit(evsel, TID);
1143
1144	if (evsel->sample_read) {
1145		evsel__set_sample_bit(evsel, READ);
1146
1147		/*
1148		 * We need ID even in case of single event, because
1149		 * PERF_SAMPLE_READ process ID specific data.
1150		 */
1151		evsel__set_sample_id(evsel, false);
1152
1153		/*
1154		 * Apply group format only if we belong to group
1155		 * with more than one members.
1156		 */
1157		if (leader->core.nr_members > 1) {
1158			attr->read_format |= PERF_FORMAT_GROUP;
1159			attr->inherit = 0;
1160		}
1161	}
1162
1163	/*
1164	 * We default some events to have a default interval. But keep
1165	 * it a weak assumption overridable by the user.
1166	 */
1167	if ((evsel->is_libpfm_event && !attr->sample_period) ||
1168	    (!evsel->is_libpfm_event && (!attr->sample_period ||
1169					 opts->user_freq != UINT_MAX ||
1170					 opts->user_interval != ULLONG_MAX)))
1171		evsel__set_default_freq_period(opts, attr);
 
 
 
 
 
1172
1173	/*
1174	 * If attr->freq was set (here or earlier), ask for period
1175	 * to be sampled.
1176	 */
1177	if (attr->freq)
1178		evsel__set_sample_bit(evsel, PERIOD);
 
 
1179
1180	if (opts->no_samples)
1181		attr->sample_freq = 0;
1182
1183	if (opts->inherit_stat) {
1184		evsel->core.attr.read_format |=
1185			PERF_FORMAT_TOTAL_TIME_ENABLED |
1186			PERF_FORMAT_TOTAL_TIME_RUNNING |
1187			PERF_FORMAT_ID;
1188		attr->inherit_stat = 1;
1189	}
1190
1191	if (opts->sample_address) {
1192		evsel__set_sample_bit(evsel, ADDR);
1193		attr->mmap_data = track;
1194	}
1195
1196	/*
1197	 * We don't allow user space callchains for  function trace
1198	 * event, due to issues with page faults while tracing page
1199	 * fault handler and its overall trickiness nature.
1200	 */
1201	if (evsel__is_function_event(evsel))
1202		evsel->core.attr.exclude_callchain_user = 1;
1203
1204	if (callchain && callchain->enabled && !evsel->no_aux_samples)
1205		evsel__config_callchain(evsel, opts, callchain);
1206
1207	if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1208	    !evsel__is_dummy_event(evsel)) {
1209		attr->sample_regs_intr = opts->sample_intr_regs;
1210		evsel__set_sample_bit(evsel, REGS_INTR);
1211	}
1212
1213	if (opts->sample_user_regs && !evsel->no_aux_samples &&
1214	    !evsel__is_dummy_event(evsel)) {
1215		attr->sample_regs_user |= opts->sample_user_regs;
1216		evsel__set_sample_bit(evsel, REGS_USER);
1217	}
1218
1219	if (target__has_cpu(&opts->target) || opts->sample_cpu)
1220		evsel__set_sample_bit(evsel, CPU);
1221
1222	/*
1223	 * When the user explicitly disabled time don't force it here.
1224	 */
1225	if (opts->sample_time &&
1226	    (!perf_missing_features.sample_id_all &&
1227	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1228	     opts->sample_time_set)))
1229		evsel__set_sample_bit(evsel, TIME);
1230
1231	if (opts->raw_samples && !evsel->no_aux_samples) {
1232		evsel__set_sample_bit(evsel, TIME);
1233		evsel__set_sample_bit(evsel, RAW);
1234		evsel__set_sample_bit(evsel, CPU);
1235	}
1236
1237	if (opts->sample_address)
1238		evsel__set_sample_bit(evsel, DATA_SRC);
1239
1240	if (opts->sample_phys_addr)
1241		evsel__set_sample_bit(evsel, PHYS_ADDR);
1242
1243	if (opts->no_buffering) {
1244		attr->watermark = 0;
1245		attr->wakeup_events = 1;
1246	}
1247	if (opts->branch_stack && !evsel->no_aux_samples) {
1248		evsel__set_sample_bit(evsel, BRANCH_STACK);
1249		attr->branch_sample_type = opts->branch_stack;
1250	}
1251
1252	if (opts->sample_weight)
1253		arch_evsel__set_sample_weight(evsel);
1254
1255	attr->task     = track;
1256	attr->mmap     = track;
1257	attr->mmap2    = track && !perf_missing_features.mmap2;
1258	attr->comm     = track;
1259	attr->build_id = track && opts->build_id;
1260
1261	/*
1262	 * ksymbol is tracked separately with text poke because it needs to be
1263	 * system wide and enabled immediately.
1264	 */
1265	if (!opts->text_poke)
1266		attr->ksymbol = track && !perf_missing_features.ksymbol;
1267	attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1268
1269	if (opts->record_namespaces)
1270		attr->namespaces  = track;
1271
1272	if (opts->record_cgroup) {
1273		attr->cgroup = track && !perf_missing_features.cgroup;
1274		evsel__set_sample_bit(evsel, CGROUP);
1275	}
1276
1277	if (opts->sample_data_page_size)
1278		evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1279
1280	if (opts->sample_code_page_size)
1281		evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1282
1283	if (opts->record_switch_events)
1284		attr->context_switch = track;
1285
1286	if (opts->sample_transaction)
1287		evsel__set_sample_bit(evsel, TRANSACTION);
1288
1289	if (opts->running_time) {
1290		evsel->core.attr.read_format |=
1291			PERF_FORMAT_TOTAL_TIME_ENABLED |
1292			PERF_FORMAT_TOTAL_TIME_RUNNING;
1293	}
1294
1295	/*
1296	 * XXX see the function comment above
1297	 *
1298	 * Disabling only independent events or group leaders,
1299	 * keeping group members enabled.
1300	 */
1301	if (evsel__is_group_leader(evsel))
1302		attr->disabled = 1;
1303
1304	/*
1305	 * Setting enable_on_exec for independent events and
1306	 * group leaders for traced executed by perf.
1307	 */
1308	if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1309	    !opts->target.initial_delay)
1310		attr->enable_on_exec = 1;
1311
1312	if (evsel->immediate) {
1313		attr->disabled = 0;
1314		attr->enable_on_exec = 0;
1315	}
1316
1317	clockid = opts->clockid;
1318	if (opts->use_clockid) {
1319		attr->use_clockid = 1;
1320		attr->clockid = opts->clockid;
1321	}
1322
1323	if (evsel->precise_max)
1324		attr->precise_ip = 3;
1325
1326	if (opts->all_user) {
1327		attr->exclude_kernel = 1;
1328		attr->exclude_user   = 0;
1329	}
1330
1331	if (opts->all_kernel) {
1332		attr->exclude_kernel = 0;
1333		attr->exclude_user   = 1;
1334	}
1335
1336	if (evsel->core.own_cpus || evsel->unit)
1337		evsel->core.attr.read_format |= PERF_FORMAT_ID;
1338
1339	/*
1340	 * Apply event specific term settings,
1341	 * it overloads any global configuration.
1342	 */
1343	evsel__apply_config_terms(evsel, opts, track);
 
 
 
 
 
1344
1345	evsel->ignore_missing_thread = opts->ignore_missing_thread;
 
1346
1347	/* The --period option takes the precedence. */
1348	if (opts->period_set) {
1349		if (opts->period)
1350			evsel__set_sample_bit(evsel, PERIOD);
1351		else
1352			evsel__reset_sample_bit(evsel, PERIOD);
 
 
1353	}
1354
1355	/*
1356	 * A dummy event never triggers any actual counter and therefore
1357	 * cannot be used with branch_stack.
1358	 *
1359	 * For initial_delay, a dummy event is added implicitly.
1360	 * The software event will trigger -EOPNOTSUPP error out,
1361	 * if BRANCH_STACK bit is set.
1362	 */
1363	if (evsel__is_dummy_event(evsel))
1364		evsel__reset_sample_bit(evsel, BRANCH_STACK);
 
 
 
 
 
 
 
 
 
 
1365
1366	if (evsel__is_offcpu_event(evsel))
1367		evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1368
1369	arch__post_evsel_config(evsel, attr);
 
 
 
 
 
1370}
1371
1372int evsel__set_filter(struct evsel *evsel, const char *filter)
1373{
1374	char *new_filter = strdup(filter);
1375
1376	if (new_filter != NULL) {
1377		free(evsel->filter);
1378		evsel->filter = new_filter;
1379		return 0;
1380	}
1381
1382	return -1;
1383}
1384
1385static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
 
1386{
1387	char *new_filter;
1388
1389	if (evsel->filter == NULL)
1390		return evsel__set_filter(evsel, filter);
1391
1392	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1393		free(evsel->filter);
1394		evsel->filter = new_filter;
1395		return 0;
1396	}
1397
1398	return -1;
1399}
1400
1401int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1402{
1403	return evsel__append_filter(evsel, "(%s) && (%s)", filter);
 
 
 
 
 
1404}
1405
1406int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1407{
1408	return evsel__append_filter(evsel, "%s,%s", filter);
 
 
 
 
 
1409}
1410
1411/* Caller has to clear disabled after going through all CPUs. */
1412int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1413{
1414	return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1415}
 
 
 
 
 
 
 
1416
1417int evsel__enable(struct evsel *evsel)
1418{
1419	int err = perf_evsel__enable(&evsel->core);
 
 
 
1420
1421	if (!err)
1422		evsel->disabled = false;
1423	return err;
1424}
1425
1426/* Caller has to set disabled after going through all CPUs. */
1427int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1428{
1429	return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
 
1430}
1431
1432int evsel__disable(struct evsel *evsel)
1433{
1434	int err = perf_evsel__disable(&evsel->core);
1435	/*
1436	 * We mark it disabled here so that tools that disable a event can
1437	 * ignore events after they disable it. I.e. the ring buffer may have
1438	 * already a few more events queued up before the kernel got the stop
1439	 * request.
1440	 */
1441	if (!err)
1442		evsel->disabled = true;
1443
1444	return err;
1445}
1446
1447void free_config_terms(struct list_head *config_terms)
1448{
1449	struct evsel_config_term *term, *h;
1450
1451	list_for_each_entry_safe(term, h, config_terms, list) {
1452		list_del_init(&term->list);
1453		if (term->free_str)
1454			zfree(&term->val.str);
1455		free(term);
1456	}
1457}
1458
1459static void evsel__free_config_terms(struct evsel *evsel)
1460{
1461	free_config_terms(&evsel->config_terms);
 
 
 
 
 
 
 
 
 
1462}
1463
1464void evsel__exit(struct evsel *evsel)
1465{
1466	assert(list_empty(&evsel->core.node));
1467	assert(evsel->evlist == NULL);
1468	bpf_counter__destroy(evsel);
1469	perf_bpf_filter__destroy(evsel);
1470	evsel__free_counts(evsel);
1471	perf_evsel__free_fd(&evsel->core);
1472	perf_evsel__free_id(&evsel->core);
1473	evsel__free_config_terms(evsel);
1474	cgroup__put(evsel->cgrp);
1475	perf_cpu_map__put(evsel->core.cpus);
1476	perf_cpu_map__put(evsel->core.own_cpus);
1477	perf_thread_map__put(evsel->core.threads);
1478	zfree(&evsel->group_name);
1479	zfree(&evsel->name);
1480	zfree(&evsel->filter);
1481	zfree(&evsel->pmu_name);
1482	zfree(&evsel->group_pmu_name);
1483	zfree(&evsel->unit);
1484	zfree(&evsel->metric_id);
1485	evsel__zero_per_pkg(evsel);
1486	hashmap__free(evsel->per_pkg_mask);
1487	evsel->per_pkg_mask = NULL;
1488	zfree(&evsel->metric_events);
1489	perf_evsel__object.fini(evsel);
1490}
1491
1492void evsel__delete(struct evsel *evsel)
1493{
1494	if (!evsel)
1495		return;
1496
1497	evsel__exit(evsel);
1498	free(evsel);
1499}
1500
1501void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1502			   struct perf_counts_values *count)
1503{
1504	struct perf_counts_values tmp;
1505
1506	if (!evsel->prev_raw_counts)
1507		return;
1508
1509	tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1510	*perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
 
 
 
 
 
1511
1512	count->val = count->val - tmp.val;
1513	count->ena = count->ena - tmp.ena;
1514	count->run = count->run - tmp.run;
1515}
1516
1517static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
 
1518{
1519	struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1520
1521	return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1522}
 
 
 
 
 
 
 
 
1523
1524static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1525			     u64 val, u64 ena, u64 run, u64 lost)
1526{
1527	struct perf_counts_values *count;
1528
1529	count = perf_counts(counter->counts, cpu_map_idx, thread);
1530
1531	count->val    = val;
1532	count->ena    = ena;
1533	count->run    = run;
1534	count->lost   = lost;
1535
1536	perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1537}
1538
1539static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
 
1540{
1541	u64 read_format = leader->core.attr.read_format;
1542	struct sample_read_value *v;
1543	u64 nr, ena = 0, run = 0, lost = 0;
1544
1545	nr = *data++;
1546
1547	if (nr != (u64) leader->core.nr_members)
1548		return -EINVAL;
1549
1550	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1551		ena = *data++;
1552
1553	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1554		run = *data++;
1555
1556	v = (void *)data;
1557	sample_read_group__for_each(v, nr, read_format) {
1558		struct evsel *counter;
1559
1560		counter = evlist__id2evsel(leader->evlist, v->id);
1561		if (!counter)
1562			return -EINVAL;
1563
1564		if (read_format & PERF_FORMAT_LOST)
1565			lost = v->lost;
1566
1567		evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1568	}
1569
1570	return 0;
1571}
1572
1573static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1574{
1575	struct perf_stat_evsel *ps = leader->stats;
1576	u64 read_format = leader->core.attr.read_format;
1577	int size = perf_evsel__read_size(&leader->core);
1578	u64 *data = ps->group_data;
1579
1580	if (!(read_format & PERF_FORMAT_ID))
1581		return -EINVAL;
1582
1583	if (!evsel__is_group_leader(leader))
1584		return -EINVAL;
1585
1586	if (!data) {
1587		data = zalloc(size);
1588		if (!data)
1589			return -ENOMEM;
1590
1591		ps->group_data = data;
1592	}
1593
1594	if (FD(leader, cpu_map_idx, thread) < 0)
1595		return -EINVAL;
1596
1597	if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1598		return -errno;
1599
1600	return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1601}
1602
1603int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1604{
1605	u64 read_format = evsel->core.attr.read_format;
1606
1607	if (read_format & PERF_FORMAT_GROUP)
1608		return evsel__read_group(evsel, cpu_map_idx, thread);
1609
1610	return evsel__read_one(evsel, cpu_map_idx, thread);
1611}
1612
1613int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1614{
1615	struct perf_counts_values count;
1616	size_t nv = scale ? 3 : 1;
1617
1618	if (FD(evsel, cpu_map_idx, thread) < 0)
1619		return -EINVAL;
1620
1621	if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1622		return -ENOMEM;
1623
1624	if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1625		return -errno;
1626
1627	evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1628	perf_counts_values__scale(&count, scale, NULL);
1629	*perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1630	return 0;
1631}
1632
1633static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1634				  int cpu_map_idx)
1635{
1636	struct perf_cpu cpu;
1637
1638	cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1639	return perf_cpu_map__idx(other->core.cpus, cpu);
1640}
1641
1642static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1643{
1644	struct evsel *leader = evsel__leader(evsel);
1645
1646	if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1647	    (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1648		return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1649	}
1650
1651	return cpu_map_idx;
1652}
1653
1654static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1655{
1656	struct evsel *leader = evsel__leader(evsel);
1657	int fd;
1658
1659	if (evsel__is_group_leader(evsel))
1660		return -1;
1661
1662	/*
1663	 * Leader must be already processed/open,
1664	 * if not it's a bug.
1665	 */
1666	BUG_ON(!leader->core.fd);
1667
1668	cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1669	if (cpu_map_idx == -1)
1670		return -1;
1671
1672	fd = FD(leader, cpu_map_idx, thread);
1673	BUG_ON(fd == -1 && !leader->skippable);
1674
1675	/*
1676	 * When the leader has been skipped, return -2 to distinguish from no
1677	 * group leader case.
1678	 */
1679	return fd == -1 ? -2 : fd;
1680}
1681
1682static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1683{
1684	for (int cpu = 0; cpu < nr_cpus; cpu++)
1685		for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1686			FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1687}
1688
1689static int update_fds(struct evsel *evsel,
1690		      int nr_cpus, int cpu_map_idx,
1691		      int nr_threads, int thread_idx)
1692{
1693	struct evsel *pos;
 
 
 
 
 
 
 
 
1694
1695	if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1696		return -EINVAL;
1697
1698	evlist__for_each_entry(evsel->evlist, pos) {
1699		nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
 
 
 
1700
1701		evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
 
 
 
 
 
 
1702
1703		/*
1704		 * Since fds for next evsel has not been created,
1705		 * there is no need to iterate whole event list.
1706		 */
1707		if (pos == evsel)
1708			break;
1709	}
1710	return 0;
1711}
1712
1713static bool evsel__ignore_missing_thread(struct evsel *evsel,
1714					 int nr_cpus, int cpu_map_idx,
1715					 struct perf_thread_map *threads,
1716					 int thread, int err)
1717{
1718	pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1719
1720	if (!evsel->ignore_missing_thread)
1721		return false;
1722
1723	/* The system wide setup does not work with threads. */
1724	if (evsel->core.system_wide)
1725		return false;
1726
1727	/* The -ESRCH is perf event syscall errno for pid's not found. */
1728	if (err != -ESRCH)
1729		return false;
1730
1731	/* If there's only one thread, let it fail. */
1732	if (threads->nr == 1)
1733		return false;
1734
1735	/*
1736	 * We should remove fd for missing_thread first
1737	 * because thread_map__remove() will decrease threads->nr.
1738	 */
1739	if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1740		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1741
1742	if (thread_map__remove(threads, thread))
1743		return false;
1744
1745	pr_warning("WARNING: Ignored open failure for pid %d\n",
1746		   ignore_pid);
1747	return true;
1748}
1749
1750static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1751				void *priv __maybe_unused)
1752{
1753	return fprintf(fp, "  %-32s %s\n", name, val);
1754}
1755
1756static void display_attr(struct perf_event_attr *attr)
 
1757{
1758	if (verbose >= 2 || debug_peo_args) {
1759		fprintf(stderr, "%.60s\n", graph_dotted_line);
1760		fprintf(stderr, "perf_event_attr:\n");
1761		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1762		fprintf(stderr, "%.60s\n", graph_dotted_line);
1763	}
1764}
1765
1766bool evsel__precise_ip_fallback(struct evsel *evsel)
1767{
1768	/* Do not try less precise if not requested. */
1769	if (!evsel->precise_max)
1770		return false;
1771
1772	/*
1773	 * We tried all the precise_ip values, and it's
1774	 * still failing, so leave it to standard fallback.
1775	 */
1776	if (!evsel->core.attr.precise_ip) {
1777		evsel->core.attr.precise_ip = evsel->precise_ip_original;
1778		return false;
1779	}
1780
1781	if (!evsel->precise_ip_original)
1782		evsel->precise_ip_original = evsel->core.attr.precise_ip;
1783
1784	evsel->core.attr.precise_ip--;
1785	pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1786	display_attr(&evsel->core.attr);
1787	return true;
1788}
1789
1790static struct perf_cpu_map *empty_cpu_map;
1791static struct perf_thread_map *empty_thread_map;
1792
1793static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1794		struct perf_thread_map *threads)
1795{
1796	int nthreads = perf_thread_map__nr(threads);
1797
1798	if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1799	    (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1800		return -EINVAL;
1801
1802	if (cpus == NULL) {
1803		if (empty_cpu_map == NULL) {
1804			empty_cpu_map = perf_cpu_map__new_any_cpu();
1805			if (empty_cpu_map == NULL)
1806				return -ENOMEM;
1807		}
1808
1809		cpus = empty_cpu_map;
1810	}
1811
1812	if (threads == NULL) {
1813		if (empty_thread_map == NULL) {
1814			empty_thread_map = thread_map__new_by_tid(-1);
1815			if (empty_thread_map == NULL)
1816				return -ENOMEM;
1817		}
1818
1819		threads = empty_thread_map;
1820	}
1821
1822	if (evsel->core.fd == NULL &&
1823	    perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1824		return -ENOMEM;
1825
1826	evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1827	if (evsel->cgrp)
1828		evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1829
1830	return 0;
1831}
1832
1833static void evsel__disable_missing_features(struct evsel *evsel)
1834{
1835	if (perf_missing_features.branch_counters)
1836		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_COUNTERS;
1837	if (perf_missing_features.read_lost)
1838		evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1839	if (perf_missing_features.weight_struct) {
1840		evsel__set_sample_bit(evsel, WEIGHT);
1841		evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1842	}
1843	if (perf_missing_features.clockid_wrong)
1844		evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1845	if (perf_missing_features.clockid) {
1846		evsel->core.attr.use_clockid = 0;
1847		evsel->core.attr.clockid = 0;
1848	}
1849	if (perf_missing_features.cloexec)
1850		evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1851	if (perf_missing_features.mmap2)
1852		evsel->core.attr.mmap2 = 0;
1853	if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1854		evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1855	if (perf_missing_features.lbr_flags)
1856		evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1857				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1858	if (perf_missing_features.group_read && evsel->core.attr.inherit)
1859		evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1860	if (perf_missing_features.ksymbol)
1861		evsel->core.attr.ksymbol = 0;
1862	if (perf_missing_features.bpf)
1863		evsel->core.attr.bpf_event = 0;
1864	if (perf_missing_features.branch_hw_idx)
1865		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1866	if (perf_missing_features.sample_id_all)
1867		evsel->core.attr.sample_id_all = 0;
1868}
1869
1870int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1871			struct perf_thread_map *threads)
1872{
1873	int err;
1874
1875	err = __evsel__prepare_open(evsel, cpus, threads);
1876	if (err)
1877		return err;
1878
1879	evsel__disable_missing_features(evsel);
1880
1881	return err;
1882}
1883
1884bool evsel__detect_missing_features(struct evsel *evsel)
1885{
1886	/*
1887	 * Must probe features in the order they were added to the
1888	 * perf_event_attr interface.
1889	 */
1890	if (!perf_missing_features.branch_counters &&
1891	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS)) {
1892		perf_missing_features.branch_counters = true;
1893		pr_debug2("switching off branch counters support\n");
1894		return true;
1895	} else if (!perf_missing_features.read_lost &&
1896	    (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1897		perf_missing_features.read_lost = true;
1898		pr_debug2("switching off PERF_FORMAT_LOST support\n");
1899		return true;
1900	} else if (!perf_missing_features.weight_struct &&
1901	    (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1902		perf_missing_features.weight_struct = true;
1903		pr_debug2("switching off weight struct support\n");
1904		return true;
1905	} else if (!perf_missing_features.code_page_size &&
1906	    (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1907		perf_missing_features.code_page_size = true;
1908		pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1909		return false;
1910	} else if (!perf_missing_features.data_page_size &&
1911	    (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1912		perf_missing_features.data_page_size = true;
1913		pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1914		return false;
1915	} else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1916		perf_missing_features.cgroup = true;
1917		pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1918		return false;
1919	} else if (!perf_missing_features.branch_hw_idx &&
1920	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1921		perf_missing_features.branch_hw_idx = true;
1922		pr_debug2("switching off branch HW index support\n");
1923		return true;
1924	} else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1925		perf_missing_features.aux_output = true;
1926		pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1927		return false;
1928	} else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1929		perf_missing_features.bpf = true;
1930		pr_debug2_peo("switching off bpf_event\n");
1931		return true;
1932	} else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1933		perf_missing_features.ksymbol = true;
1934		pr_debug2_peo("switching off ksymbol\n");
1935		return true;
1936	} else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1937		perf_missing_features.write_backward = true;
1938		pr_debug2_peo("switching off write_backward\n");
1939		return false;
1940	} else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1941		perf_missing_features.clockid_wrong = true;
1942		pr_debug2_peo("switching off clockid\n");
1943		return true;
1944	} else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1945		perf_missing_features.clockid = true;
1946		pr_debug2_peo("switching off use_clockid\n");
1947		return true;
1948	} else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1949		perf_missing_features.cloexec = true;
1950		pr_debug2_peo("switching off cloexec flag\n");
1951		return true;
1952	} else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1953		perf_missing_features.mmap2 = true;
1954		pr_debug2_peo("switching off mmap2\n");
1955		return true;
1956	} else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1957		if (evsel->pmu == NULL)
1958			evsel->pmu = evsel__find_pmu(evsel);
1959
1960		if (evsel->pmu)
1961			evsel->pmu->missing_features.exclude_guest = true;
1962		else {
1963			/* we cannot find PMU, disable attrs now */
1964			evsel->core.attr.exclude_host = false;
1965			evsel->core.attr.exclude_guest = false;
1966		}
1967
1968		if (evsel->exclude_GH) {
1969			pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1970			return false;
1971		}
1972		if (!perf_missing_features.exclude_guest) {
1973			perf_missing_features.exclude_guest = true;
1974			pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1975		}
1976		return true;
1977	} else if (!perf_missing_features.sample_id_all) {
1978		perf_missing_features.sample_id_all = true;
1979		pr_debug2_peo("switching off sample_id_all\n");
1980		return true;
1981	} else if (!perf_missing_features.lbr_flags &&
1982			(evsel->core.attr.branch_sample_type &
1983			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1984			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1985		perf_missing_features.lbr_flags = true;
1986		pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1987		return true;
1988	} else if (!perf_missing_features.group_read &&
1989		    evsel->core.attr.inherit &&
1990		   (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1991		   evsel__is_group_leader(evsel)) {
1992		perf_missing_features.group_read = true;
1993		pr_debug2_peo("switching off group read\n");
1994		return true;
1995	} else {
1996		return false;
1997	}
1998}
1999
2000static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2001		struct perf_thread_map *threads,
2002		int start_cpu_map_idx, int end_cpu_map_idx)
2003{
2004	int idx, thread, nthreads;
2005	int pid = -1, err, old_errno;
2006	enum rlimit_action set_rlimit = NO_CHANGE;
2007
2008	err = __evsel__prepare_open(evsel, cpus, threads);
2009	if (err)
2010		return err;
2011
2012	if (cpus == NULL)
2013		cpus = empty_cpu_map;
2014
2015	if (threads == NULL)
2016		threads = empty_thread_map;
2017
2018	nthreads = perf_thread_map__nr(threads);
2019
2020	if (evsel->cgrp)
2021		pid = evsel->cgrp->fd;
2022
2023fallback_missing_features:
2024	evsel__disable_missing_features(evsel);
2025
2026	pr_debug3("Opening: %s\n", evsel__name(evsel));
2027	display_attr(&evsel->core.attr);
2028
2029	for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2030
2031		for (thread = 0; thread < nthreads; thread++) {
2032			int fd, group_fd;
2033retry_open:
2034			if (thread >= nthreads)
2035				break;
2036
2037			if (!evsel->cgrp && !evsel->core.system_wide)
2038				pid = perf_thread_map__pid(threads, thread);
2039
2040			group_fd = get_group_fd(evsel, idx, thread);
2041
2042			if (group_fd == -2) {
2043				pr_debug("broken group leader for %s\n", evsel->name);
2044				err = -EINVAL;
2045				goto out_close;
2046			}
2047
2048			test_attr__ready();
2049
2050			/* Debug message used by test scripts */
2051			pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2052				pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2053
2054			fd = sys_perf_event_open(&evsel->core.attr, pid,
2055						perf_cpu_map__cpu(cpus, idx).cpu,
2056						group_fd, evsel->open_flags);
2057
2058			FD(evsel, idx, thread) = fd;
2059
2060			if (fd < 0) {
2061				err = -errno;
2062
2063				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2064					  err);
2065				goto try_fallback;
2066			}
2067
2068			bpf_counter__install_pe(evsel, idx, fd);
2069
2070			if (unlikely(test_attr__enabled)) {
2071				test_attr__open(&evsel->core.attr, pid,
2072						perf_cpu_map__cpu(cpus, idx),
2073						fd, group_fd, evsel->open_flags);
2074			}
2075
2076			/* Debug message used by test scripts */
2077			pr_debug2_peo(" = %d\n", fd);
2078
2079			if (evsel->bpf_fd >= 0) {
2080				int evt_fd = fd;
2081				int bpf_fd = evsel->bpf_fd;
2082
2083				err = ioctl(evt_fd,
2084					    PERF_EVENT_IOC_SET_BPF,
2085					    bpf_fd);
2086				if (err && errno != EEXIST) {
2087					pr_err("failed to attach bpf fd %d: %s\n",
2088					       bpf_fd, strerror(errno));
2089					err = -EINVAL;
2090					goto out_close;
2091				}
2092			}
2093
2094			set_rlimit = NO_CHANGE;
2095
2096			/*
2097			 * If we succeeded but had to kill clockid, fail and
2098			 * have evsel__open_strerror() print us a nice error.
 
2099			 */
2100			if (perf_missing_features.clockid ||
2101			    perf_missing_features.clockid_wrong) {
2102				err = -EINVAL;
2103				goto out_close;
2104			}
2105		}
2106	}
2107
2108	return 0;
2109
2110try_fallback:
2111	if (evsel__precise_ip_fallback(evsel))
2112		goto retry_open;
2113
2114	if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2115					 idx, threads, thread, err)) {
2116		/* We just removed 1 thread, so lower the upper nthreads limit. */
2117		nthreads--;
2118
2119		/* ... and pretend like nothing have happened. */
2120		err = 0;
2121		goto retry_open;
2122	}
2123	/*
2124	 * perf stat needs between 5 and 22 fds per CPU. When we run out
2125	 * of them try to increase the limits.
2126	 */
2127	if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
2128		goto retry_open;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2129
2130	if (err != -EINVAL || idx > 0 || thread > 0)
2131		goto out_close;
2132
2133	if (evsel__detect_missing_features(evsel))
 
 
 
 
 
2134		goto fallback_missing_features;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2135out_close:
2136	if (err)
2137		threads->err_thread = thread;
2138
2139	old_errno = errno;
2140	do {
2141		while (--thread >= 0) {
2142			if (FD(evsel, idx, thread) >= 0)
2143				close(FD(evsel, idx, thread));
2144			FD(evsel, idx, thread) = -1;
2145		}
2146		thread = nthreads;
2147	} while (--idx >= 0);
2148	errno = old_errno;
2149	return err;
2150}
2151
2152int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2153		struct perf_thread_map *threads)
2154{
2155	return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
 
 
 
 
2156}
2157
2158void evsel__close(struct evsel *evsel)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2159{
2160	perf_evsel__close(&evsel->core);
2161	perf_evsel__free_id(&evsel->core);
 
 
 
 
 
 
 
2162}
2163
2164int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
 
2165{
2166	if (cpu_map_idx == -1)
2167		return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2168
2169	return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2170}
2171
2172int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
 
2173{
2174	return evsel__open(evsel, NULL, threads);
2175}
2176
2177static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2178				       const union perf_event *event,
2179				       struct perf_sample *sample)
2180{
2181	u64 type = evsel->core.attr.sample_type;
2182	const __u64 *array = event->sample.array;
2183	bool swapped = evsel->needs_swap;
2184	union u64_swap u;
2185
2186	array += ((event->header.size -
2187		   sizeof(event->header)) / sizeof(u64)) - 1;
2188
2189	if (type & PERF_SAMPLE_IDENTIFIER) {
2190		sample->id = *array;
2191		array--;
2192	}
2193
2194	if (type & PERF_SAMPLE_CPU) {
2195		u.val64 = *array;
2196		if (swapped) {
2197			/* undo swap of u64, then swap on individual u32s */
2198			u.val64 = bswap_64(u.val64);
2199			u.val32[0] = bswap_32(u.val32[0]);
2200		}
2201
2202		sample->cpu = u.val32[0];
2203		array--;
2204	}
2205
2206	if (type & PERF_SAMPLE_STREAM_ID) {
2207		sample->stream_id = *array;
2208		array--;
2209	}
2210
2211	if (type & PERF_SAMPLE_ID) {
2212		sample->id = *array;
2213		array--;
2214	}
2215
2216	if (type & PERF_SAMPLE_TIME) {
2217		sample->time = *array;
2218		array--;
2219	}
2220
2221	if (type & PERF_SAMPLE_TID) {
2222		u.val64 = *array;
2223		if (swapped) {
2224			/* undo swap of u64, then swap on individual u32s */
2225			u.val64 = bswap_64(u.val64);
2226			u.val32[0] = bswap_32(u.val32[0]);
2227			u.val32[1] = bswap_32(u.val32[1]);
2228		}
2229
2230		sample->pid = u.val32[0];
2231		sample->tid = u.val32[1];
2232		array--;
2233	}
2234
2235	return 0;
2236}
2237
2238static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2239			    u64 size)
2240{
2241	return size > max_size || offset + size > endp;
2242}
2243
2244#define OVERFLOW_CHECK(offset, size, max_size)				\
2245	do {								\
2246		if (overflow(endp, (max_size), (offset), (size)))	\
2247			return -EFAULT;					\
2248	} while (0)
2249
2250#define OVERFLOW_CHECK_u64(offset) \
2251	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2252
2253static int
2254perf_event__check_size(union perf_event *event, unsigned int sample_size)
2255{
2256	/*
2257	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2258	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2259	 * check the format does not go past the end of the event.
2260	 */
2261	if (sample_size + sizeof(event->header) > event->header.size)
2262		return -EFAULT;
2263
2264	return 0;
2265}
2266
2267void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2268					  const __u64 *array,
2269					  u64 type __maybe_unused)
2270{
2271	data->weight = *array;
2272}
2273
2274u64 evsel__bitfield_swap_branch_flags(u64 value)
2275{
2276	u64 new_val = 0;
2277
2278	/*
2279	 * branch_flags
2280	 * union {
2281	 * 	u64 values;
2282	 * 	struct {
2283	 * 		mispred:1	//target mispredicted
2284	 * 		predicted:1	//target predicted
2285	 * 		in_tx:1		//in transaction
2286	 * 		abort:1		//transaction abort
2287	 * 		cycles:16	//cycle count to last branch
2288	 * 		type:4		//branch type
2289	 * 		spec:2		//branch speculation info
2290	 * 		new_type:4	//additional branch type
2291	 * 		priv:3		//privilege level
2292	 * 		reserved:31
2293	 * 	}
2294	 * }
2295	 *
2296	 * Avoid bswap64() the entire branch_flag.value,
2297	 * as it has variable bit-field sizes. Instead the
2298	 * macro takes the bit-field position/size,
2299	 * swaps it based on the host endianness.
2300	 */
2301	if (host_is_bigendian()) {
2302		new_val = bitfield_swap(value, 0, 1);
2303		new_val |= bitfield_swap(value, 1, 1);
2304		new_val |= bitfield_swap(value, 2, 1);
2305		new_val |= bitfield_swap(value, 3, 1);
2306		new_val |= bitfield_swap(value, 4, 16);
2307		new_val |= bitfield_swap(value, 20, 4);
2308		new_val |= bitfield_swap(value, 24, 2);
2309		new_val |= bitfield_swap(value, 26, 4);
2310		new_val |= bitfield_swap(value, 30, 3);
2311		new_val |= bitfield_swap(value, 33, 31);
2312	} else {
2313		new_val = bitfield_swap(value, 63, 1);
2314		new_val |= bitfield_swap(value, 62, 1);
2315		new_val |= bitfield_swap(value, 61, 1);
2316		new_val |= bitfield_swap(value, 60, 1);
2317		new_val |= bitfield_swap(value, 44, 16);
2318		new_val |= bitfield_swap(value, 40, 4);
2319		new_val |= bitfield_swap(value, 38, 2);
2320		new_val |= bitfield_swap(value, 34, 4);
2321		new_val |= bitfield_swap(value, 31, 3);
2322		new_val |= bitfield_swap(value, 0, 31);
2323	}
2324
2325	return new_val;
2326}
2327
2328static inline bool evsel__has_branch_counters(const struct evsel *evsel)
2329{
2330	struct evsel *cur, *leader = evsel__leader(evsel);
2331
2332	/* The branch counters feature only supports group */
2333	if (!leader || !evsel->evlist)
2334		return false;
2335
2336	evlist__for_each_entry(evsel->evlist, cur) {
2337		if ((leader == evsel__leader(cur)) &&
2338		    (cur->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS))
2339			return true;
2340	}
2341	return false;
2342}
2343
2344int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2345			struct perf_sample *data)
2346{
2347	u64 type = evsel->core.attr.sample_type;
2348	bool swapped = evsel->needs_swap;
2349	const __u64 *array;
2350	u16 max_size = event->header.size;
2351	const void *endp = (void *)event + max_size;
2352	u64 sz;
2353
2354	/*
2355	 * used for cross-endian analysis. See git commit 65014ab3
2356	 * for why this goofiness is needed.
2357	 */
2358	union u64_swap u;
2359
2360	memset(data, 0, sizeof(*data));
2361	data->cpu = data->pid = data->tid = -1;
2362	data->stream_id = data->id = data->time = -1ULL;
2363	data->period = evsel->core.attr.sample_period;
 
2364	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2365	data->misc    = event->header.misc;
2366	data->id = -1ULL;
2367	data->data_src = PERF_MEM_DATA_SRC_NONE;
2368	data->vcpu = -1;
2369
2370	if (event->header.type != PERF_RECORD_SAMPLE) {
2371		if (!evsel->core.attr.sample_id_all)
2372			return 0;
2373		return perf_evsel__parse_id_sample(evsel, event, data);
2374	}
2375
2376	array = event->sample.array;
2377
2378	if (perf_event__check_size(event, evsel->sample_size))
 
 
 
 
 
2379		return -EFAULT;
2380
 
2381	if (type & PERF_SAMPLE_IDENTIFIER) {
2382		data->id = *array;
2383		array++;
2384	}
2385
2386	if (type & PERF_SAMPLE_IP) {
2387		data->ip = *array;
2388		array++;
2389	}
2390
2391	if (type & PERF_SAMPLE_TID) {
2392		u.val64 = *array;
2393		if (swapped) {
2394			/* undo swap of u64, then swap on individual u32s */
2395			u.val64 = bswap_64(u.val64);
2396			u.val32[0] = bswap_32(u.val32[0]);
2397			u.val32[1] = bswap_32(u.val32[1]);
2398		}
2399
2400		data->pid = u.val32[0];
2401		data->tid = u.val32[1];
2402		array++;
2403	}
2404
2405	if (type & PERF_SAMPLE_TIME) {
2406		data->time = *array;
2407		array++;
2408	}
2409
 
2410	if (type & PERF_SAMPLE_ADDR) {
2411		data->addr = *array;
2412		array++;
2413	}
2414
2415	if (type & PERF_SAMPLE_ID) {
2416		data->id = *array;
2417		array++;
2418	}
2419
2420	if (type & PERF_SAMPLE_STREAM_ID) {
2421		data->stream_id = *array;
2422		array++;
2423	}
2424
2425	if (type & PERF_SAMPLE_CPU) {
2426
2427		u.val64 = *array;
2428		if (swapped) {
2429			/* undo swap of u64, then swap on individual u32s */
2430			u.val64 = bswap_64(u.val64);
2431			u.val32[0] = bswap_32(u.val32[0]);
2432		}
2433
2434		data->cpu = u.val32[0];
2435		array++;
2436	}
2437
2438	if (type & PERF_SAMPLE_PERIOD) {
2439		data->period = *array;
2440		array++;
2441	}
2442
2443	if (type & PERF_SAMPLE_READ) {
2444		u64 read_format = evsel->core.attr.read_format;
2445
2446		OVERFLOW_CHECK_u64(array);
2447		if (read_format & PERF_FORMAT_GROUP)
2448			data->read.group.nr = *array;
2449		else
2450			data->read.one.value = *array;
2451
2452		array++;
2453
2454		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2455			OVERFLOW_CHECK_u64(array);
2456			data->read.time_enabled = *array;
2457			array++;
2458		}
2459
2460		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2461			OVERFLOW_CHECK_u64(array);
2462			data->read.time_running = *array;
2463			array++;
2464		}
2465
2466		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2467		if (read_format & PERF_FORMAT_GROUP) {
2468			const u64 max_group_nr = UINT64_MAX /
2469					sizeof(struct sample_read_value);
2470
2471			if (data->read.group.nr > max_group_nr)
2472				return -EFAULT;
2473
2474			sz = data->read.group.nr * sample_read_value_size(read_format);
2475			OVERFLOW_CHECK(array, sz, max_size);
2476			data->read.group.values =
2477					(struct sample_read_value *)array;
2478			array = (void *)array + sz;
2479		} else {
2480			OVERFLOW_CHECK_u64(array);
2481			data->read.one.id = *array;
2482			array++;
2483
2484			if (read_format & PERF_FORMAT_LOST) {
2485				OVERFLOW_CHECK_u64(array);
2486				data->read.one.lost = *array;
2487				array++;
2488			}
2489		}
2490	}
2491
2492	if (type & PERF_SAMPLE_CALLCHAIN) {
2493		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2494
2495		OVERFLOW_CHECK_u64(array);
2496		data->callchain = (struct ip_callchain *)array++;
2497		if (data->callchain->nr > max_callchain_nr)
2498			return -EFAULT;
2499		sz = data->callchain->nr * sizeof(u64);
2500		OVERFLOW_CHECK(array, sz, max_size);
2501		array = (void *)array + sz;
2502	}
2503
2504	if (type & PERF_SAMPLE_RAW) {
2505		OVERFLOW_CHECK_u64(array);
2506		u.val64 = *array;
2507
2508		/*
2509		 * Undo swap of u64, then swap on individual u32s,
2510		 * get the size of the raw area and undo all of the
2511		 * swap. The pevent interface handles endianness by
2512		 * itself.
2513		 */
2514		if (swapped) {
2515			u.val64 = bswap_64(u.val64);
2516			u.val32[0] = bswap_32(u.val32[0]);
2517			u.val32[1] = bswap_32(u.val32[1]);
2518		}
2519		data->raw_size = u.val32[0];
2520
2521		/*
2522		 * The raw data is aligned on 64bits including the
2523		 * u32 size, so it's safe to use mem_bswap_64.
2524		 */
2525		if (swapped)
2526			mem_bswap_64((void *) array, data->raw_size);
2527
2528		array = (void *)array + sizeof(u32);
2529
2530		OVERFLOW_CHECK(array, data->raw_size, max_size);
2531		data->raw_data = (void *)array;
2532		array = (void *)array + data->raw_size;
2533	}
2534
2535	if (type & PERF_SAMPLE_BRANCH_STACK) {
2536		const u64 max_branch_nr = UINT64_MAX /
2537					  sizeof(struct branch_entry);
2538		struct branch_entry *e;
2539		unsigned int i;
2540
2541		OVERFLOW_CHECK_u64(array);
2542		data->branch_stack = (struct branch_stack *)array++;
2543
2544		if (data->branch_stack->nr > max_branch_nr)
2545			return -EFAULT;
2546
2547		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2548		if (evsel__has_branch_hw_idx(evsel)) {
2549			sz += sizeof(u64);
2550			e = &data->branch_stack->entries[0];
2551		} else {
2552			data->no_hw_idx = true;
2553			/*
2554			 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2555			 * only nr and entries[] will be output by kernel.
2556			 */
2557			e = (struct branch_entry *)&data->branch_stack->hw_idx;
2558		}
2559
2560		if (swapped) {
2561			/*
2562			 * struct branch_flag does not have endian
2563			 * specific bit field definition. And bswap
2564			 * will not resolve the issue, since these
2565			 * are bit fields.
2566			 *
2567			 * evsel__bitfield_swap_branch_flags() uses a
2568			 * bitfield_swap macro to swap the bit position
2569			 * based on the host endians.
2570			 */
2571			for (i = 0; i < data->branch_stack->nr; i++, e++)
2572				e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2573		}
2574
2575		OVERFLOW_CHECK(array, sz, max_size);
2576		array = (void *)array + sz;
2577
2578		if (evsel__has_branch_counters(evsel)) {
2579			OVERFLOW_CHECK_u64(array);
2580
2581			data->branch_stack_cntr = (u64 *)array;
2582			sz = data->branch_stack->nr * sizeof(u64);
2583
2584			OVERFLOW_CHECK(array, sz, max_size);
2585			array = (void *)array + sz;
2586		}
2587	}
2588
2589	if (type & PERF_SAMPLE_REGS_USER) {
2590		OVERFLOW_CHECK_u64(array);
2591		data->user_regs.abi = *array;
2592		array++;
2593
2594		if (data->user_regs.abi) {
2595			u64 mask = evsel->core.attr.sample_regs_user;
2596
2597			sz = hweight64(mask) * sizeof(u64);
2598			OVERFLOW_CHECK(array, sz, max_size);
2599			data->user_regs.mask = mask;
2600			data->user_regs.regs = (u64 *)array;
2601			array = (void *)array + sz;
2602		}
2603	}
2604
2605	if (type & PERF_SAMPLE_STACK_USER) {
2606		OVERFLOW_CHECK_u64(array);
2607		sz = *array++;
2608
2609		data->user_stack.offset = ((char *)(array - 1)
2610					  - (char *) event);
2611
2612		if (!sz) {
2613			data->user_stack.size = 0;
2614		} else {
2615			OVERFLOW_CHECK(array, sz, max_size);
2616			data->user_stack.data = (char *)array;
2617			array = (void *)array + sz;
2618			OVERFLOW_CHECK_u64(array);
2619			data->user_stack.size = *array++;
2620			if (WARN_ONCE(data->user_stack.size > sz,
2621				      "user stack dump failure\n"))
2622				return -EFAULT;
2623		}
2624	}
2625
2626	if (type & PERF_SAMPLE_WEIGHT_TYPE) {
 
2627		OVERFLOW_CHECK_u64(array);
2628		arch_perf_parse_sample_weight(data, array, type);
2629		array++;
2630	}
2631
 
2632	if (type & PERF_SAMPLE_DATA_SRC) {
2633		OVERFLOW_CHECK_u64(array);
2634		data->data_src = *array;
2635		array++;
2636	}
2637
 
2638	if (type & PERF_SAMPLE_TRANSACTION) {
2639		OVERFLOW_CHECK_u64(array);
2640		data->transaction = *array;
2641		array++;
2642	}
2643
2644	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2645	if (type & PERF_SAMPLE_REGS_INTR) {
2646		OVERFLOW_CHECK_u64(array);
2647		data->intr_regs.abi = *array;
2648		array++;
2649
2650		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2651			u64 mask = evsel->core.attr.sample_regs_intr;
2652
2653			sz = hweight64(mask) * sizeof(u64);
2654			OVERFLOW_CHECK(array, sz, max_size);
2655			data->intr_regs.mask = mask;
2656			data->intr_regs.regs = (u64 *)array;
2657			array = (void *)array + sz;
2658		}
2659	}
2660
2661	data->phys_addr = 0;
2662	if (type & PERF_SAMPLE_PHYS_ADDR) {
2663		data->phys_addr = *array;
2664		array++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2665	}
2666
2667	data->cgroup = 0;
2668	if (type & PERF_SAMPLE_CGROUP) {
2669		data->cgroup = *array;
2670		array++;
2671	}
2672
2673	data->data_page_size = 0;
2674	if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2675		data->data_page_size = *array;
2676		array++;
 
 
 
 
2677	}
2678
2679	data->code_page_size = 0;
2680	if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2681		data->code_page_size = *array;
2682		array++;
 
 
 
2683	}
2684
2685	if (type & PERF_SAMPLE_AUX) {
2686		OVERFLOW_CHECK_u64(array);
2687		sz = *array++;
 
 
 
 
 
2688
2689		OVERFLOW_CHECK(array, sz, max_size);
2690		/* Undo swap of data */
2691		if (swapped)
2692			mem_bswap_64((char *)array, sz);
2693		data->aux_sample.size = sz;
2694		data->aux_sample.data = (char *)array;
2695		array = (void *)array + sz;
 
2696	}
2697
2698	return 0;
2699}
2700
2701int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2702				  u64 *timestamp)
 
 
2703{
2704	u64 type = evsel->core.attr.sample_type;
2705	const __u64 *array;
 
 
 
 
 
 
 
2706
2707	if (!(type & PERF_SAMPLE_TIME))
2708		return -1;
 
 
 
 
 
 
 
2709
2710	if (event->header.type != PERF_RECORD_SAMPLE) {
2711		struct perf_sample data = {
2712			.time = -1ULL,
2713		};
 
 
 
 
 
 
 
2714
2715		if (!evsel->core.attr.sample_id_all)
2716			return -1;
2717		if (perf_evsel__parse_id_sample(evsel, event, &data))
2718			return -1;
2719
2720		*timestamp = data.time;
2721		return 0;
 
2722	}
2723
2724	array = event->sample.array;
 
 
 
2725
2726	if (perf_event__check_size(event, evsel->sample_size))
2727		return -EFAULT;
 
 
2728
2729	if (type & PERF_SAMPLE_IDENTIFIER)
 
2730		array++;
 
2731
2732	if (type & PERF_SAMPLE_IP)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2733		array++;
 
2734
2735	if (type & PERF_SAMPLE_TID)
 
 
 
 
2736		array++;
2737
2738	if (type & PERF_SAMPLE_TIME)
2739		*timestamp = *array;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2740
2741	return 0;
2742}
 
 
 
 
 
 
 
 
 
 
 
2743
2744u16 evsel__id_hdr_size(struct evsel *evsel)
2745{
2746	u64 sample_type = evsel->core.attr.sample_type;
2747	u16 size = 0;
2748
2749	if (sample_type & PERF_SAMPLE_TID)
2750		size += sizeof(u64);
 
 
 
 
2751
2752	if (sample_type & PERF_SAMPLE_TIME)
2753		size += sizeof(u64);
 
 
 
 
 
 
 
 
2754
2755	if (sample_type & PERF_SAMPLE_ID)
2756		size += sizeof(u64);
 
 
 
 
 
 
 
2757
2758	if (sample_type & PERF_SAMPLE_STREAM_ID)
2759		size += sizeof(u64);
 
 
2760
2761	if (sample_type & PERF_SAMPLE_CPU)
2762		size += sizeof(u64);
 
 
2763
2764	if (sample_type & PERF_SAMPLE_IDENTIFIER)
2765		size += sizeof(u64);
 
 
2766
2767	return size;
2768}
 
 
 
 
 
 
 
 
2769
2770#ifdef HAVE_LIBTRACEEVENT
2771struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2772{
2773	return tep_find_field(evsel->tp_format, name);
2774}
2775
2776struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name)
2777{
2778	return tep_find_common_field(evsel->tp_format, name);
2779}
2780
2781void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
 
2782{
2783	struct tep_format_field *field = evsel__field(evsel, name);
2784	int offset;
2785
2786	if (!field)
2787		return NULL;
2788
2789	offset = field->offset;
2790
2791	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2792		offset = *(int *)(sample->raw_data + field->offset);
2793		offset &= 0xffff;
2794		if (tep_field_is_relative(field->flags))
2795			offset += field->offset + field->size;
2796	}
2797
2798	return sample->raw_data + offset;
2799}
2800
2801u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2802			 bool needs_swap)
2803{
 
 
2804	u64 value;
2805	void *ptr = sample->raw_data + field->offset;
 
 
 
 
2806
2807	switch (field->size) {
2808	case 1:
2809		return *(u8 *)ptr;
2810	case 2:
2811		value = *(u16 *)ptr;
2812		break;
2813	case 4:
2814		value = *(u32 *)ptr;
2815		break;
2816	case 8:
2817		memcpy(&value, ptr, sizeof(u64));
2818		break;
2819	default:
2820		return 0;
2821	}
2822
2823	if (!needs_swap)
2824		return value;
2825
2826	switch (field->size) {
2827	case 2:
2828		return bswap_16(value);
2829	case 4:
2830		return bswap_32(value);
2831	case 8:
2832		return bswap_64(value);
2833	default:
2834		return 0;
2835	}
2836
2837	return 0;
2838}
2839
2840u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2841{
2842	struct tep_format_field *field = evsel__field(evsel, name);
 
 
 
 
 
 
 
 
2843
2844	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
 
 
 
 
 
 
 
 
2845}
2846
2847u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name)
 
2848{
2849	struct tep_format_field *field = evsel__common_field(evsel, name);
 
 
 
 
2850
2851	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2852}
2853
2854#endif
2855
2856bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
2857		     char *msg, size_t msgsize)
2858{
2859	int paranoid;
2860
2861	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2862	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2863	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2864		/*
2865		 * If it's cycles then fall back to hrtimer based cpu-clock sw
2866		 * counter, which is always available even if no PMU support.
 
2867		 *
2868		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2869		 * b0a873e).
2870		 */
2871		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2872		evsel->core.attr.config = target__has_cpu(target)
2873			? PERF_COUNT_SW_CPU_CLOCK
2874			: PERF_COUNT_SW_TASK_CLOCK;
2875		scnprintf(msg, msgsize,
2876			"The cycles event is not supported, trying to fall back to %s",
2877			target__has_cpu(target) ? "cpu-clock" : "task-clock");
2878
2879		zfree(&evsel->name);
2880		return true;
2881	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2882		   (paranoid = perf_event_paranoid()) > 1) {
2883		const char *name = evsel__name(evsel);
2884		char *new_name;
2885		const char *sep = ":";
2886
2887		/* If event has exclude user then don't exclude kernel. */
2888		if (evsel->core.attr.exclude_user)
2889			return false;
2890
2891		/* Is there already the separator in the name. */
2892		if (strchr(name, '/') ||
2893		    (strchr(name, ':') && !evsel->is_libpfm_event))
2894			sep = "";
2895
2896		if (asprintf(&new_name, "%s%su", name, sep) < 0)
2897			return false;
2898
2899		free(evsel->name);
 
2900		evsel->name = new_name;
2901		scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2902			  "to fall back to excluding kernel and hypervisor "
2903			  " samples", paranoid);
2904		evsel->core.attr.exclude_kernel = 1;
2905		evsel->core.attr.exclude_hv     = 1;
2906
2907		return true;
2908	}
2909
2910	return false;
2911}
2912
2913static bool find_process(const char *name)
2914{
2915	size_t len = strlen(name);
2916	DIR *dir;
2917	struct dirent *d;
2918	int ret = -1;
2919
2920	dir = opendir(procfs__mountpoint());
2921	if (!dir)
2922		return false;
2923
2924	/* Walk through the directory. */
2925	while (ret && (d = readdir(dir)) != NULL) {
2926		char path[PATH_MAX];
2927		char *data;
2928		size_t size;
2929
2930		if ((d->d_type != DT_DIR) ||
2931		     !strcmp(".", d->d_name) ||
2932		     !strcmp("..", d->d_name))
2933			continue;
2934
2935		scnprintf(path, sizeof(path), "%s/%s/comm",
2936			  procfs__mountpoint(), d->d_name);
2937
2938		if (filename__read_str(path, &data, &size))
2939			continue;
2940
2941		ret = strncmp(name, data, len);
2942		free(data);
2943	}
2944
2945	closedir(dir);
2946	return ret ? false : true;
2947}
2948
2949int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
2950				     char *msg __maybe_unused,
2951				     size_t size __maybe_unused)
2952{
2953	return 0;
2954}
2955
2956int evsel__open_strerror(struct evsel *evsel, struct target *target,
2957			 int err, char *msg, size_t size)
2958{
2959	char sbuf[STRERR_BUFSIZE];
2960	int printed = 0, enforced = 0;
2961	int ret;
2962
2963	switch (err) {
2964	case EPERM:
2965	case EACCES:
2966		printed += scnprintf(msg + printed, size - printed,
2967			"Access to performance monitoring and observability operations is limited.\n");
2968
2969		if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2970			if (enforced) {
2971				printed += scnprintf(msg + printed, size - printed,
2972					"Enforced MAC policy settings (SELinux) can limit access to performance\n"
2973					"monitoring and observability operations. Inspect system audit records for\n"
2974					"more perf_event access control information and adjusting the policy.\n");
2975			}
2976		}
2977
2978		if (err == EPERM)
2979			printed += scnprintf(msg, size,
2980				"No permission to enable %s event.\n\n", evsel__name(evsel));
2981
2982		return scnprintf(msg + printed, size - printed,
2983		 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2984		 "access to performance monitoring and observability operations for processes\n"
2985		 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2986		 "More information can be found at 'Perf events and tool security' document:\n"
2987		 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2988		 "perf_event_paranoid setting is %d:\n"
2989		 "  -1: Allow use of (almost) all events by all users\n"
2990		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2991		 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2992		 ">= 1: Disallow CPU event access\n"
2993		 ">= 2: Disallow kernel profiling\n"
2994		 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2995		 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2996		 perf_event_paranoid());
2997	case ENOENT:
2998		return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
 
2999	case EMFILE:
3000		return scnprintf(msg, size, "%s",
3001			 "Too many events are opened.\n"
3002			 "Probably the maximum number of open file descriptors has been reached.\n"
3003			 "Hint: Try again after reducing the number of events.\n"
3004			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3005	case ENOMEM:
3006		if (evsel__has_callchain(evsel) &&
3007		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3008			return scnprintf(msg, size,
3009					 "Not enough memory to setup event with callchain.\n"
3010					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3011					 "Hint: Current value: %d", sysctl__max_stack());
3012		break;
3013	case ENODEV:
3014		if (target->cpu_list)
3015			return scnprintf(msg, size, "%s",
3016	 "No such device - did you specify an out-of-range profile CPU?");
3017		break;
3018	case EOPNOTSUPP:
3019		if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3020			return scnprintf(msg, size,
3021	"%s: PMU Hardware or event type doesn't support branch stack sampling.",
3022					 evsel__name(evsel));
3023		if (evsel->core.attr.aux_output)
3024			return scnprintf(msg, size,
3025	"%s: PMU Hardware doesn't support 'aux_output' feature",
3026					 evsel__name(evsel));
3027		if (evsel->core.attr.sample_period != 0)
3028			return scnprintf(msg, size,
3029	"%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3030					 evsel__name(evsel));
3031		if (evsel->core.attr.precise_ip)
3032			return scnprintf(msg, size, "%s",
3033	"\'precise\' request may not be supported. Try removing 'p' modifier.");
3034#if defined(__i386__) || defined(__x86_64__)
3035		if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3036			return scnprintf(msg, size, "%s",
3037	"No hardware sampling interrupt available.\n");
 
3038#endif
3039		break;
3040	case EBUSY:
3041		if (find_process("oprofiled"))
3042			return scnprintf(msg, size,
3043	"The PMU counters are busy/taken by another profiler.\n"
3044	"We found oprofile daemon running, please stop it and try again.");
3045		break;
3046	case EINVAL:
3047		if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3048			return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3049		if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3050			return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3051		if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3052			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3053		if (perf_missing_features.clockid)
3054			return scnprintf(msg, size, "clockid feature not supported.");
3055		if (perf_missing_features.clockid_wrong)
3056			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3057		if (perf_missing_features.aux_output)
3058			return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3059		if (!target__has_cpu(target))
3060			return scnprintf(msg, size,
3061	"Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3062					evsel__name(evsel));
3063
3064		break;
3065	case ENODATA:
3066		return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3067				 "Please add an auxiliary event in front of the load latency event.");
3068	default:
3069		break;
3070	}
3071
3072	ret = arch_evsel__open_strerror(evsel, msg, size);
3073	if (ret)
3074		return ret;
3075
3076	return scnprintf(msg, size,
3077	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3078	"/bin/dmesg | grep -i perf may provide additional information.\n",
3079			 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3080}
3081
3082struct perf_env *evsel__env(struct evsel *evsel)
3083{
3084	if (evsel && evsel->evlist && evsel->evlist->env)
3085		return evsel->evlist->env;
3086	return &perf_env;
3087}
3088
3089static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3090{
3091	int cpu_map_idx, thread;
3092
3093	for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3094		for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3095		     thread++) {
3096			int fd = FD(evsel, cpu_map_idx, thread);
3097
3098			if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3099						   cpu_map_idx, thread, fd) < 0)
3100				return -1;
3101		}
3102	}
3103
3104	return 0;
3105}
3106
3107int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3108{
3109	struct perf_cpu_map *cpus = evsel->core.cpus;
3110	struct perf_thread_map *threads = evsel->core.threads;
3111
3112	if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3113		return -ENOMEM;
3114
3115	return store_evsel_ids(evsel, evlist);
3116}
3117
3118void evsel__zero_per_pkg(struct evsel *evsel)
3119{
3120	struct hashmap_entry *cur;
3121	size_t bkt;
3122
3123	if (evsel->per_pkg_mask) {
3124		hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3125			zfree(&cur->pkey);
3126
3127		hashmap__clear(evsel->per_pkg_mask);
3128	}
3129}
3130
3131/**
3132 * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3133 *                    will be false on hybrid systems for hardware and legacy
3134 *                    cache events.
3135 */
3136bool evsel__is_hybrid(const struct evsel *evsel)
3137{
3138	if (perf_pmus__num_core_pmus() == 1)
3139		return false;
3140
3141	return evsel->core.is_pmu_core;
3142}
3143
3144struct evsel *evsel__leader(const struct evsel *evsel)
3145{
3146	return container_of(evsel->core.leader, struct evsel, core);
3147}
3148
3149bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3150{
3151	return evsel->core.leader == &leader->core;
3152}
3153
3154bool evsel__is_leader(struct evsel *evsel)
3155{
3156	return evsel__has_leader(evsel, evsel);
3157}
3158
3159void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3160{
3161	evsel->core.leader = &leader->core;
3162}
3163
3164int evsel__source_count(const struct evsel *evsel)
3165{
3166	struct evsel *pos;
3167	int count = 0;
3168
3169	evlist__for_each_entry(evsel->evlist, pos) {
3170		if (pos->metric_leader == evsel)
3171			count++;
3172	}
3173	return count;
3174}
3175
3176bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3177{
3178	return false;
3179}
3180
3181/*
3182 * Remove an event from a given group (leader).
3183 * Some events, e.g., perf metrics Topdown events,
3184 * must always be grouped. Ignore the events.
3185 */
3186void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3187{
3188	if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3189		evsel__set_leader(evsel, evsel);
3190		evsel->core.nr_members = 0;
3191		leader->core.nr_members--;
3192	}
3193}