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
66	struct perf_missing_features perf_missing_features;
67
68	static clockid_t clockid;
69
70	static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
71	NULL,
72	"duration_time",
73	"user_time",
74	"system_time",
75	};
76
77	const 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
85	enum 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
97	static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
98	{
99	return 0;
100	}
101
102	void __weak test_attr__ready(void) { }
103
104	static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
105	{
106	}
107
108	static 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
118	int 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
130	set_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
142	int __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	*/
166	static 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	*/
199	static 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
218	void evsel__calc_id_pos(struct evsel *evsel)
219	{
220	evsel->id_pos = __perf_evsel__calc_id_pos(sample_type: evsel->core.attr.sample_type);
221	evsel->is_pos = __perf_evsel__calc_is_pos(sample_type: evsel->core.attr.sample_type);
222	}
223
224	void __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
234	void __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
244	void 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	*/
264	bool 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
274	void 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(list: &evsel->config_terms);
286	INIT_LIST_HEAD(list: &evsel->bpf_counter_list);
287	INIT_LIST_HEAD(list: &evsel->bpf_filters);
288	perf_evsel__object.init(evsel);
289	evsel->sample_size = __evsel__sample_size(sample_type: 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
300	struct 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
323	int 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(new: &tmp->list, head: dst);
341	}
342	return 0;
343	}
344
345	static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
346	{
347	return copy_config_terms(dst: &dst->config_terms, src: &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	*/
357	struct 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(attr: &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(cgroup: 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(dst: evsel, src: orig) < 0)
442	goto out_err;
443
444	return evsel;
445
446	out_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
455	struct 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
486	out_free:
487	zfree(&evsel->name);
488	free(evsel);
489	out_err:
490	return ERR_PTR(err);
491	}
492	#endif
493
494	const 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
507	char *evsel__bpf_counter_events;
508
509	bool 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
529	static 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
537	static 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(buf: bf + r, size: size - r, fmt: "%.*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
573	int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
574	{
575	return scnprintf(buf: bf, size, fmt: "%s", __evsel__hw_name(config: evsel->core.attr.config));
576	}
577
578	static 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: bf + r, size: size - r);
582	}
583
584	const 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
597	static 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
604	static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
605	{
606	int r = scnprintf(buf: bf, size, fmt: "%s", __evsel__sw_name(config: evsel->core.attr.config));
607	return r + evsel__add_modifiers(evsel, bf: bf + r, size: size - r);
608	}
609
610	static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
611	{
612	return scnprintf(buf: bf, size, fmt: "%s", perf_tool_event__to_str(ev));
613	}
614
615	static 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(buf: bf + r, size: size - r, fmt: "r");
623
624	if (type & HW_BREAKPOINT_W)
625	r += scnprintf(buf: bf + r, size: size - r, fmt: "w");
626
627	if (type & HW_BREAKPOINT_X)
628	r += scnprintf(buf: bf + r, size: size - r, fmt: "x");
629
630	return r;
631	}
632
633	static 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, addr: attr->bp_addr, type: attr->bp_type);
637	return r + evsel__add_modifiers(evsel, bf: bf + r, size: size - r);
638	}
639
640	const 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
650	const 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
656	const 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	*/
672	static 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
682	bool 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
690	int __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(buf: bf, size, fmt: "%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(buf: bf, size, fmt: "%s-%s", evsel__hw_cache[type][0],
699	evsel__hw_cache_op[op][1]);
700	}
701
702	static 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);
725	out_err:
726	return scnprintf(buf: bf, size, fmt: "%s", err);
727	}
728
729	static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
730	{
731	int ret = __evsel__hw_cache_name(config: evsel->core.attr.config, bf, size);
732	return ret + evsel__add_modifiers(evsel, bf: bf + ret, size: size - ret);
733	}
734
735	static 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: bf + ret, size: size - ret);
739	}
740
741	const 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, size: sizeof(bf));
754	break;
755
756	case PERF_TYPE_HARDWARE:
757	evsel__hw_name(evsel, bf, size: sizeof(bf));
758	break;
759
760	case PERF_TYPE_HW_CACHE:
761	evsel__hw_cache_name(evsel, bf, size: sizeof(bf));
762	break;
763
764	case PERF_TYPE_SOFTWARE:
765	if (evsel__is_tool(evsel))
766	evsel__tool_name(ev: evsel->tool_event, bf, size: sizeof(bf));
767	else
768	evsel__sw_name(evsel, bf, size: sizeof(bf));
769	break;
770
771	case PERF_TYPE_TRACEPOINT:
772	scnprintf(buf: bf, size: sizeof(bf), fmt: "%s", "unknown tracepoint");
773	break;
774
775	case PERF_TYPE_BREAKPOINT:
776	evsel__bp_name(evsel, bf, size: sizeof(bf));
777	break;
778
779	default:
780	scnprintf(buf: bf, size: sizeof(bf), fmt: "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;
789	out_unknown:
790	return "unknown";
791	}
792
793	bool evsel__name_is(struct evsel *evsel, const char *name)
794	{
795	return !strcmp(evsel__name(evsel), name);
796	}
797
798	const 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(ev: evsel->tool_event);
805
806	return "unknown";
807	}
808
809	const 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	*/
824	int 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, fmt: "%s { ", group_name);
832
833	ret += scnprintf(buf: buf + ret, size: size - ret, fmt: "%s", evsel__name(evsel));
834
835	for_each_group_member(pos, evsel)
836	ret += scnprintf(buf: buf + ret, size: size - ret, fmt: ", %s", evsel__name(evsel: pos));
837
838	if (!evsel->forced_leader)
839	ret += scnprintf(buf: buf + ret, size: size - ret, fmt: " }");
840
841	return ret;
842	}
843
844	static 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
905	void 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
912	static 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
929	static 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(arg: callgraph_buf, param: &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, param: &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: &param);
1058	}
1059	}
1060	}
1061
1062	struct 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
1074	void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1075	{
1076	evsel__set_sample_bit(evsel, WEIGHT);
1077	}
1078
1079	void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1080	struct perf_event_attr *attr __maybe_unused)
1081	{
1082	}
1083
1084	static 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
1095	static 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	*/
1128	void 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, can_sample_identifier: 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, param: 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(target: &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(target: &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(target: &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
1372	int 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
1385	static 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
1401	int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1402	{
1403	return evsel__append_filter(evsel, fmt: "(%s) && (%s)", filter);
1404	}
1405
1406	int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1407	{
1408	return evsel__append_filter(evsel, fmt: "%s,%s", filter);
1409	}
1410
1411	/* Caller has to clear disabled after going through all CPUs. */
1412	int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1413	{
1414	return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1415	}
1416
1417	int 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. */
1427	int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1428	{
1429	return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1430	}
1431
1432	int 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
1447	void 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(entry: &term->list);
1453	if (term->free_str)
1454	zfree(&term->val.str);
1455	free(term);
1456	}
1457	}
1458
1459	static void evsel__free_config_terms(struct evsel *evsel)
1460	{
1461	free_config_terms(config_terms: &evsel->config_terms);
1462	}
1463
1464	void evsel__exit(struct evsel *evsel)
1465	{
1466	assert(list_empty(head: &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(cgroup: 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
1492	void evsel__delete(struct evsel *evsel)
1493	{
1494	if (!evsel)
1495	return;
1496
1497	evsel__exit(evsel);
1498	free(evsel);
1499	}
1500
1501	void 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(counts: evsel->prev_raw_counts, cpu_map_idx, thread);
1510	*perf_counts(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
1517	static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1518	{
1519	struct perf_counts_values *count = perf_counts(counts: evsel->counts, cpu_map_idx, thread);
1520
1521	return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1522	}
1523
1524	static 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(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(counts: counter->counts, cpu_map_idx, thread, loaded: true);
1537	}
1538
1539	static 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(evlist: leader->evlist, id: 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, val: v->value, ena, run, lost);
1568	}
1569
1570	return 0;
1571	}
1572
1573	static 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(evsel: 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
1603	int 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(leader: evsel, cpu_map_idx, thread);
1609
1610	return evsel__read_one(evsel, cpu_map_idx, thread);
1611	}
1612
1613	int __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: &count);
1628	perf_counts_values__scale(&count, scale, NULL);
1629	*perf_counts(counts: evsel->counts, cpu_map_idx, thread) = count;
1630	return 0;
1631	}
1632
1633	static 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
1642	static 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(evsel: leader)) ||
1647	(!evsel__is_hybrid(evsel) && evsel__is_hybrid(evsel: leader))) {
1648	return evsel__match_other_cpu(evsel, other: leader, cpu_map_idx);
1649	}
1650
1651	return cpu_map_idx;
1652	}
1653
1654	static 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
1682	static 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
1689	static 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
1713	static 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, nr_threads: threads->nr, thread_idx: thread))
1740	return false;
1741
1742	if (thread_map__remove(threads, idx: thread))
1743	return false;
1744
1745	pr_warning("WARNING: Ignored open failure for pid %d\n",
1746	ignore_pid);
1747	return true;
1748	}
1749
1750	static 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
1756	static 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
1766	bool 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(attr: &evsel->core.attr);
1787	return true;
1788	}
1789
1790	static struct perf_cpu_map *empty_cpu_map;
1791	static struct perf_thread_map *empty_thread_map;
1792
1793	static 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(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
1833	static 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
1870	int 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
1884	bool 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
2000	static 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
2023	fallback_missing_features:
2024	evsel__disable_missing_features(evsel);
2025
2026	pr_debug3("Opening: %s\n", evsel__name(evsel));
2027	display_attr(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;
2033	retry_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, cpu_map_idx: 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(attr: &evsel->core.attr, pid,
2055	cpu: perf_cpu_map__cpu(cpus, idx).cpu,
2056	group_fd, flags: 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, cpu: idx, fd);
2069
2070	if (unlikely(test_attr__enabled)) {
2071	test_attr__open(attr: &evsel->core.attr, pid,
2072	cpu: perf_cpu_map__cpu(cpus, idx),
2073	fd, group_fd, flags: 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
2110	try_fallback:
2111	if (evsel__precise_ip_fallback(evsel))
2112	goto retry_open;
2113
2114	if (evsel__ignore_missing_thread(evsel, nr_cpus: perf_cpu_map__nr(cpus),
2115	cpu_map_idx: 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: &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;
2135	out_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
2152	int 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, start_cpu_map_idx: 0, end_cpu_map_idx: perf_cpu_map__nr(cpus));
2156	}
2157
2158	void evsel__close(struct evsel *evsel)
2159	{
2160	perf_evsel__close(&evsel->core);
2161	perf_evsel__free_id(&evsel->core);
2162	}
2163
2164	int 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, start_cpu_map_idx: 0, end_cpu_map_idx: perf_cpu_map__nr(cpus));
2168
2169	return evsel__open_cpu(evsel, cpus, NULL, start_cpu_map_idx: cpu_map_idx, end_cpu_map_idx: cpu_map_idx + 1);
2170	}
2171
2172	int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2173	{
2174	return evsel__open(evsel, NULL, threads);
2175	}
2176
2177	static 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
2238	static 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
2253	static int
2254	perf_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
2267	void __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
2274	u64 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
2328	static 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(evsel: cur)) &&
2338	(cur->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS))
2339	return true;
2340	}
2341	return false;
2342	}
2343
2344	int 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->data_src = PERF_MEM_DATA_SRC_NONE;
2367	data->vcpu = -1;
2368
2369	if (event->header.type != PERF_RECORD_SAMPLE) {
2370	if (!evsel->core.attr.sample_id_all)
2371	return 0;
2372	return perf_evsel__parse_id_sample(evsel, event, sample: data);
2373	}
2374
2375	array = event->sample.array;
2376
2377	if (perf_event__check_size(event, sample_size: evsel->sample_size))
2378	return -EFAULT;
2379
2380	if (type & PERF_SAMPLE_IDENTIFIER) {
2381	data->id = *array;
2382	array++;
2383	}
2384
2385	if (type & PERF_SAMPLE_IP) {
2386	data->ip = *array;
2387	array++;
2388	}
2389
2390	if (type & PERF_SAMPLE_TID) {
2391	u.val64 = *array;
2392	if (swapped) {
2393	/* undo swap of u64, then swap on individual u32s */
2394	u.val64 = bswap_64(u.val64);
2395	u.val32[0] = bswap_32(u.val32[0]);
2396	u.val32[1] = bswap_32(u.val32[1]);
2397	}
2398
2399	data->pid = u.val32[0];
2400	data->tid = u.val32[1];
2401	array++;
2402	}
2403
2404	if (type & PERF_SAMPLE_TIME) {
2405	data->time = *array;
2406	array++;
2407	}
2408
2409	if (type & PERF_SAMPLE_ADDR) {
2410	data->addr = *array;
2411	array++;
2412	}
2413
2414	if (type & PERF_SAMPLE_ID) {
2415	data->id = *array;
2416	array++;
2417	}
2418
2419	if (type & PERF_SAMPLE_STREAM_ID) {
2420	data->stream_id = *array;
2421	array++;
2422	}
2423
2424	if (type & PERF_SAMPLE_CPU) {
2425
2426	u.val64 = *array;
2427	if (swapped) {
2428	/* undo swap of u64, then swap on individual u32s */
2429	u.val64 = bswap_64(u.val64);
2430	u.val32[0] = bswap_32(u.val32[0]);
2431	}
2432
2433	data->cpu = u.val32[0];
2434	array++;
2435	}
2436
2437	if (type & PERF_SAMPLE_PERIOD) {
2438	data->period = *array;
2439	array++;
2440	}
2441
2442	if (type & PERF_SAMPLE_READ) {
2443	u64 read_format = evsel->core.attr.read_format;
2444
2445	OVERFLOW_CHECK_u64(array);
2446	if (read_format & PERF_FORMAT_GROUP)
2447	data->read.group.nr = *array;
2448	else
2449	data->read.one.value = *array;
2450
2451	array++;
2452
2453	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2454	OVERFLOW_CHECK_u64(array);
2455	data->read.time_enabled = *array;
2456	array++;
2457	}
2458
2459	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2460	OVERFLOW_CHECK_u64(array);
2461	data->read.time_running = *array;
2462	array++;
2463	}
2464
2465	/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2466	if (read_format & PERF_FORMAT_GROUP) {
2467	const u64 max_group_nr = UINT64_MAX /
2468	sizeof(struct sample_read_value);
2469
2470	if (data->read.group.nr > max_group_nr)
2471	return -EFAULT;
2472
2473	sz = data->read.group.nr * sample_read_value_size(read_format);
2474	OVERFLOW_CHECK(array, sz, max_size);
2475	data->read.group.values =
2476	(struct sample_read_value *)array;
2477	array = (void *)array + sz;
2478	} else {
2479	OVERFLOW_CHECK_u64(array);
2480	data->read.one.id = *array;
2481	array++;
2482
2483	if (read_format & PERF_FORMAT_LOST) {
2484	OVERFLOW_CHECK_u64(array);
2485	data->read.one.lost = *array;
2486	array++;
2487	}
2488	}
2489	}
2490
2491	if (type & PERF_SAMPLE_CALLCHAIN) {
2492	const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2493
2494	OVERFLOW_CHECK_u64(array);
2495	data->callchain = (struct ip_callchain *)array++;
2496	if (data->callchain->nr > max_callchain_nr)
2497	return -EFAULT;
2498	sz = data->callchain->nr * sizeof(u64);
2499	OVERFLOW_CHECK(array, sz, max_size);
2500	array = (void *)array + sz;
2501	}
2502
2503	if (type & PERF_SAMPLE_RAW) {
2504	OVERFLOW_CHECK_u64(array);
2505	u.val64 = *array;
2506
2507	/*
2508	* Undo swap of u64, then swap on individual u32s,
2509	* get the size of the raw area and undo all of the
2510	* swap. The pevent interface handles endianness by
2511	* itself.
2512	*/
2513	if (swapped) {
2514	u.val64 = bswap_64(u.val64);
2515	u.val32[0] = bswap_32(u.val32[0]);
2516	u.val32[1] = bswap_32(u.val32[1]);
2517	}
2518	data->raw_size = u.val32[0];
2519
2520	/*
2521	* The raw data is aligned on 64bits including the
2522	* u32 size, so it's safe to use mem_bswap_64.
2523	*/
2524	if (swapped)
2525	mem_bswap_64(src: (void *) array, byte_size: data->raw_size);
2526
2527	array = (void *)array + sizeof(u32);
2528
2529	OVERFLOW_CHECK(array, data->raw_size, max_size);
2530	data->raw_data = (void *)array;
2531	array = (void *)array + data->raw_size;
2532	}
2533
2534	if (type & PERF_SAMPLE_BRANCH_STACK) {
2535	const u64 max_branch_nr = UINT64_MAX /
2536	sizeof(struct branch_entry);
2537	struct branch_entry *e;
2538	unsigned int i;
2539
2540	OVERFLOW_CHECK_u64(array);
2541	data->branch_stack = (struct branch_stack *)array++;
2542
2543	if (data->branch_stack->nr > max_branch_nr)
2544	return -EFAULT;
2545
2546	sz = data->branch_stack->nr * sizeof(struct branch_entry);
2547	if (evsel__has_branch_hw_idx(evsel)) {
2548	sz += sizeof(u64);
2549	e = &data->branch_stack->entries[0];
2550	} else {
2551	data->no_hw_idx = true;
2552	/*
2553	* if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2554	* only nr and entries[] will be output by kernel.
2555	*/
2556	e = (struct branch_entry *)&data->branch_stack->hw_idx;
2557	}
2558
2559	if (swapped) {
2560	/*
2561	* struct branch_flag does not have endian
2562	* specific bit field definition. And bswap
2563	* will not resolve the issue, since these
2564	* are bit fields.
2565	*
2566	* evsel__bitfield_swap_branch_flags() uses a
2567	* bitfield_swap macro to swap the bit position
2568	* based on the host endians.
2569	*/
2570	for (i = 0; i < data->branch_stack->nr; i++, e++)
2571	e->flags.value = evsel__bitfield_swap_branch_flags(value: e->flags.value);
2572	}
2573
2574	OVERFLOW_CHECK(array, sz, max_size);
2575	array = (void *)array + sz;
2576
2577	if (evsel__has_branch_counters(evsel)) {
2578	OVERFLOW_CHECK_u64(array);
2579
2580	data->branch_stack_cntr = (u64 *)array;
2581	sz = data->branch_stack->nr * sizeof(u64);
2582
2583	OVERFLOW_CHECK(array, sz, max_size);
2584	array = (void *)array + sz;
2585	}
2586	}
2587
2588	if (type & PERF_SAMPLE_REGS_USER) {
2589	OVERFLOW_CHECK_u64(array);
2590	data->user_regs.abi = *array;
2591	array++;
2592
2593	if (data->user_regs.abi) {
2594	u64 mask = evsel->core.attr.sample_regs_user;
2595
2596	sz = hweight64(mask) * sizeof(u64);
2597	OVERFLOW_CHECK(array, sz, max_size);
2598	data->user_regs.mask = mask;
2599	data->user_regs.regs = (u64 *)array;
2600	array = (void *)array + sz;
2601	}
2602	}
2603
2604	if (type & PERF_SAMPLE_STACK_USER) {
2605	OVERFLOW_CHECK_u64(array);
2606	sz = *array++;
2607
2608	data->user_stack.offset = ((char *)(array - 1)
2609	- (char *) event);
2610
2611	if (!sz) {
2612	data->user_stack.size = 0;
2613	} else {
2614	OVERFLOW_CHECK(array, sz, max_size);
2615	data->user_stack.data = (char *)array;
2616	array = (void *)array + sz;
2617	OVERFLOW_CHECK_u64(array);
2618	data->user_stack.size = *array++;
2619	if (WARN_ONCE(data->user_stack.size > sz,
2620	"user stack dump failure\n"))
2621	return -EFAULT;
2622	}
2623	}
2624
2625	if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2626	OVERFLOW_CHECK_u64(array);
2627	arch_perf_parse_sample_weight(data, array, type);
2628	array++;
2629	}
2630
2631	if (type & PERF_SAMPLE_DATA_SRC) {
2632	OVERFLOW_CHECK_u64(array);
2633	data->data_src = *array;
2634	array++;
2635	}
2636
2637	if (type & PERF_SAMPLE_TRANSACTION) {
2638	OVERFLOW_CHECK_u64(array);
2639	data->transaction = *array;
2640	array++;
2641	}
2642
2643	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2644	if (type & PERF_SAMPLE_REGS_INTR) {
2645	OVERFLOW_CHECK_u64(array);
2646	data->intr_regs.abi = *array;
2647	array++;
2648
2649	if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2650	u64 mask = evsel->core.attr.sample_regs_intr;
2651
2652	sz = hweight64(mask) * sizeof(u64);
2653	OVERFLOW_CHECK(array, sz, max_size);
2654	data->intr_regs.mask = mask;
2655	data->intr_regs.regs = (u64 *)array;
2656	array = (void *)array + sz;
2657	}
2658	}
2659
2660	data->phys_addr = 0;
2661	if (type & PERF_SAMPLE_PHYS_ADDR) {
2662	data->phys_addr = *array;
2663	array++;
2664	}
2665
2666	data->cgroup = 0;
2667	if (type & PERF_SAMPLE_CGROUP) {
2668	data->cgroup = *array;
2669	array++;
2670	}
2671
2672	data->data_page_size = 0;
2673	if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2674	data->data_page_size = *array;
2675	array++;
2676	}
2677
2678	data->code_page_size = 0;
2679	if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2680	data->code_page_size = *array;
2681	array++;
2682	}
2683
2684	if (type & PERF_SAMPLE_AUX) {
2685	OVERFLOW_CHECK_u64(array);
2686	sz = *array++;
2687
2688	OVERFLOW_CHECK(array, sz, max_size);
2689	/* Undo swap of data */
2690	if (swapped)
2691	mem_bswap_64(src: (char *)array, byte_size: sz);
2692	data->aux_sample.size = sz;
2693	data->aux_sample.data = (char *)array;
2694	array = (void *)array + sz;
2695	}
2696
2697	return 0;
2698	}
2699
2700	int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2701	u64 *timestamp)
2702	{
2703	u64 type = evsel->core.attr.sample_type;
2704	const __u64 *array;
2705
2706	if (!(type & PERF_SAMPLE_TIME))
2707	return -1;
2708
2709	if (event->header.type != PERF_RECORD_SAMPLE) {
2710	struct perf_sample data = {
2711	.time = -1ULL,
2712	};
2713
2714	if (!evsel->core.attr.sample_id_all)
2715	return -1;
2716	if (perf_evsel__parse_id_sample(evsel, event, sample: &data))
2717	return -1;
2718
2719	*timestamp = data.time;
2720	return 0;
2721	}
2722
2723	array = event->sample.array;
2724
2725	if (perf_event__check_size(event, sample_size: evsel->sample_size))
2726	return -EFAULT;
2727
2728	if (type & PERF_SAMPLE_IDENTIFIER)
2729	array++;
2730
2731	if (type & PERF_SAMPLE_IP)
2732	array++;
2733
2734	if (type & PERF_SAMPLE_TID)
2735	array++;
2736
2737	if (type & PERF_SAMPLE_TIME)
2738	*timestamp = *array;
2739
2740	return 0;
2741	}
2742
2743	u16 evsel__id_hdr_size(struct evsel *evsel)
2744	{
2745	u64 sample_type = evsel->core.attr.sample_type;
2746	u16 size = 0;
2747
2748	if (sample_type & PERF_SAMPLE_TID)
2749	size += sizeof(u64);
2750
2751	if (sample_type & PERF_SAMPLE_TIME)
2752	size += sizeof(u64);
2753
2754	if (sample_type & PERF_SAMPLE_ID)
2755	size += sizeof(u64);
2756
2757	if (sample_type & PERF_SAMPLE_STREAM_ID)
2758	size += sizeof(u64);
2759
2760	if (sample_type & PERF_SAMPLE_CPU)
2761	size += sizeof(u64);
2762
2763	if (sample_type & PERF_SAMPLE_IDENTIFIER)
2764	size += sizeof(u64);
2765
2766	return size;
2767	}
2768
2769	#ifdef HAVE_LIBTRACEEVENT
2770	struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2771	{
2772	return tep_find_field(evsel->tp_format, name);
2773	}
2774
2775	struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name)
2776	{
2777	return tep_find_common_field(evsel->tp_format, name);
2778	}
2779
2780	void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2781	{
2782	struct tep_format_field *field = evsel__field(evsel, name);
2783	int offset;
2784
2785	if (!field)
2786	return NULL;
2787
2788	offset = field->offset;
2789
2790	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2791	offset = *(int *)(sample->raw_data + field->offset);
2792	offset &= 0xffff;
2793	if (tep_field_is_relative(field->flags))
2794	offset += field->offset + field->size;
2795	}
2796
2797	return sample->raw_data + offset;
2798	}
2799
2800	u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2801	bool needs_swap)
2802	{
2803	u64 value;
2804	void *ptr = sample->raw_data + field->offset;
2805
2806	switch (field->size) {
2807	case 1:
2808	return *(u8 *)ptr;
2809	case 2:
2810	value = *(u16 *)ptr;
2811	break;
2812	case 4:
2813	value = *(u32 *)ptr;
2814	break;
2815	case 8:
2816	memcpy(&value, ptr, sizeof(u64));
2817	break;
2818	default:
2819	return 0;
2820	}
2821
2822	if (!needs_swap)
2823	return value;
2824
2825	switch (field->size) {
2826	case 2:
2827	return bswap_16(value);
2828	case 4:
2829	return bswap_32(value);
2830	case 8:
2831	return bswap_64(value);
2832	default:
2833	return 0;
2834	}
2835
2836	return 0;
2837	}
2838
2839	u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2840	{
2841	struct tep_format_field *field = evsel__field(evsel, name);
2842
2843	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2844	}
2845
2846	u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name)
2847	{
2848	struct tep_format_field *field = evsel__common_field(evsel, name);
2849
2850	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2851	}
2852
2853	char evsel__taskstate(struct evsel *evsel, struct perf_sample *sample, const char *name)
2854	{
2855	static struct tep_format_field *prev_state_field;
2856	static const char *states;
2857	struct tep_format_field *field;
2858	unsigned long long val;
2859	unsigned int bit;
2860	char state = '?'; /* '?' denotes unknown task state */
2861
2862	field = evsel__field(evsel, name);
2863
2864	if (!field)
2865	return state;
2866
2867	if (!states || field != prev_state_field) {
2868	states = parse_task_states(field);
2869	if (!states)
2870	return state;
2871	prev_state_field = field;
2872	}
2873
2874	/*
2875	* Note since the kernel exposes TASK_REPORT_MAX to userspace
2876	* to denote the 'preempted' state, we might as welll report
2877	* 'R' for this case, which make senses to users as well.
2878	*
2879	* We can change this if we have a good reason in the future.
2880	*/
2881	val = evsel__intval(evsel, sample, name);
2882	bit = val ? ffs(val) : 0;
2883	state = (!bit || bit > strlen(states)) ? 'R' : states[bit-1];
2884	return state;
2885	}
2886	#endif
2887
2888	bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
2889	char *msg, size_t msgsize)
2890	{
2891	int paranoid;
2892
2893	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2894	evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2895	evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2896	/*
2897	* If it's cycles then fall back to hrtimer based cpu-clock sw
2898	* counter, which is always available even if no PMU support.
2899	*
2900	* PPC returns ENXIO until 2.6.37 (behavior changed with commit
2901	* b0a873e).
2902	*/
2903	evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2904	evsel->core.attr.config = target__has_cpu(target)
2905	? PERF_COUNT_SW_CPU_CLOCK
2906	: PERF_COUNT_SW_TASK_CLOCK;
2907	scnprintf(buf: msg, size: msgsize,
2908	fmt: "The cycles event is not supported, trying to fall back to %s",
2909	target__has_cpu(target) ? "cpu-clock" : "task-clock");
2910
2911	zfree(&evsel->name);
2912	return true;
2913	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2914	(paranoid = perf_event_paranoid()) > 1) {
2915	const char *name = evsel__name(evsel);
2916	char *new_name;
2917	const char *sep = ":";
2918
2919	/* If event has exclude user then don't exclude kernel. */
2920	if (evsel->core.attr.exclude_user)
2921	return false;
2922
2923	/* Is there already the separator in the name. */
2924	if (strchr(name, '/') ||
2925	(strchr(name, ':') && !evsel->is_libpfm_event))
2926	sep = "";
2927
2928	if (asprintf(&new_name, "%s%su", name, sep) < 0)
2929	return false;
2930
2931	free(evsel->name);
2932	evsel->name = new_name;
2933	scnprintf(buf: msg, size: msgsize, fmt: "kernel.perf_event_paranoid=%d, trying "
2934	"to fall back to excluding kernel and hypervisor "
2935	" samples", paranoid);
2936	evsel->core.attr.exclude_kernel = 1;
2937	evsel->core.attr.exclude_hv = 1;
2938
2939	return true;
2940	}
2941
2942	return false;
2943	}
2944
2945	static bool find_process(const char *name)
2946	{
2947	size_t len = strlen(name);
2948	DIR *dir;
2949	struct dirent *d;
2950	int ret = -1;
2951
2952	dir = opendir(procfs__mountpoint());
2953	if (!dir)
2954	return false;
2955
2956	/* Walk through the directory. */
2957	while (ret && (d = readdir(dir)) != NULL) {
2958	char path[PATH_MAX];
2959	char *data;
2960	size_t size;
2961
2962	if ((d->d_type != DT_DIR) ||
2963	!strcmp(".", d->d_name) ||
2964	!strcmp("..", d->d_name))
2965	continue;
2966
2967	scnprintf(buf: path, size: sizeof(path), fmt: "%s/%s/comm",
2968	procfs__mountpoint(), d->d_name);
2969
2970	if (filename__read_str(path, &data, &size))
2971	continue;
2972
2973	ret = strncmp(name, data, len);
2974	free(data);
2975	}
2976
2977	closedir(dir);
2978	return ret ? false : true;
2979	}
2980
2981	int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
2982	char *msg __maybe_unused,
2983	size_t size __maybe_unused)
2984	{
2985	return 0;
2986	}
2987
2988	int evsel__open_strerror(struct evsel *evsel, struct target *target,
2989	int err, char *msg, size_t size)
2990	{
2991	char sbuf[STRERR_BUFSIZE];
2992	int printed = 0, enforced = 0;
2993	int ret;
2994
2995	switch (err) {
2996	case EPERM:
2997	case EACCES:
2998	printed += scnprintf(buf: msg + printed, size: size - printed,
2999	fmt: "Access to performance monitoring and observability operations is limited.\n");
3000
3001	if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
3002	if (enforced) {
3003	printed += scnprintf(buf: msg + printed, size: size - printed,
3004	fmt: "Enforced MAC policy settings (SELinux) can limit access to performance\n"
3005	"monitoring and observability operations. Inspect system audit records for\n"
3006	"more perf_event access control information and adjusting the policy.\n");
3007	}
3008	}
3009
3010	if (err == EPERM)
3011	printed += scnprintf(buf: msg, size,
3012	fmt: "No permission to enable %s event.\n\n", evsel__name(evsel));
3013
3014	return scnprintf(buf: msg + printed, size: size - printed,
3015	fmt: "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3016	"access to performance monitoring and observability operations for processes\n"
3017	"without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3018	"More information can be found at 'Perf events and tool security' document:\n"
3019	"https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3020	"perf_event_paranoid setting is %d:\n"
3021	" -1: Allow use of (almost) all events by all users\n"
3022	" Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3023	">= 0: Disallow raw and ftrace function tracepoint access\n"
3024	">= 1: Disallow CPU event access\n"
3025	">= 2: Disallow kernel profiling\n"
3026	"To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3027	"in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3028	perf_event_paranoid());
3029	case ENOENT:
3030	return scnprintf(buf: msg, size, fmt: "The %s event is not supported.", evsel__name(evsel));
3031	case EMFILE:
3032	return scnprintf(buf: msg, size, fmt: "%s",
3033	"Too many events are opened.\n"
3034	"Probably the maximum number of open file descriptors has been reached.\n"
3035	"Hint: Try again after reducing the number of events.\n"
3036	"Hint: Try increasing the limit with 'ulimit -n <limit>'");
3037	case ENOMEM:
3038	if (evsel__has_callchain(evsel) &&
3039	access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3040	return scnprintf(buf: msg, size,
3041	fmt: "Not enough memory to setup event with callchain.\n"
3042	"Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3043	"Hint: Current value: %d", sysctl__max_stack());
3044	break;
3045	case ENODEV:
3046	if (target->cpu_list)
3047	return scnprintf(buf: msg, size, fmt: "%s",
3048	"No such device - did you specify an out-of-range profile CPU?");
3049	break;
3050	case EOPNOTSUPP:
3051	if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3052	return scnprintf(buf: msg, size,
3053	fmt: "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3054	evsel__name(evsel));
3055	if (evsel->core.attr.aux_output)
3056	return scnprintf(buf: msg, size,
3057	fmt: "%s: PMU Hardware doesn't support 'aux_output' feature",
3058	evsel__name(evsel));
3059	if (evsel->core.attr.sample_period != 0)
3060	return scnprintf(buf: msg, size,
3061	fmt: "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3062	evsel__name(evsel));
3063	if (evsel->core.attr.precise_ip)
3064	return scnprintf(buf: msg, size, fmt: "%s",
3065	"\'precise\' request may not be supported. Try removing 'p' modifier.");
3066	#if defined(__i386__) || defined(__x86_64__)
3067	if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3068	return scnprintf(buf: msg, size, fmt: "%s",
3069	"No hardware sampling interrupt available.\n");
3070	#endif
3071	break;
3072	case EBUSY:
3073	if (find_process(name: "oprofiled"))
3074	return scnprintf(buf: msg, size,
3075	fmt: "The PMU counters are busy/taken by another profiler.\n"
3076	"We found oprofile daemon running, please stop it and try again.");
3077	break;
3078	case EINVAL:
3079	if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3080	return scnprintf(buf: msg, size, fmt: "Asking for the code page size isn't supported by this kernel.");
3081	if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3082	return scnprintf(buf: msg, size, fmt: "Asking for the data page size isn't supported by this kernel.");
3083	if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3084	return scnprintf(buf: msg, size, fmt: "Reading from overwrite event is not supported by this kernel.");
3085	if (perf_missing_features.clockid)
3086	return scnprintf(buf: msg, size, fmt: "clockid feature not supported.");
3087	if (perf_missing_features.clockid_wrong)
3088	return scnprintf(buf: msg, size, fmt: "wrong clockid (%d).", clockid);
3089	if (perf_missing_features.aux_output)
3090	return scnprintf(buf: msg, size, fmt: "The 'aux_output' feature is not supported, update the kernel.");
3091	if (!target__has_cpu(target))
3092	return scnprintf(buf: msg, size,
3093	fmt: "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3094	evsel__name(evsel));
3095
3096	break;
3097	case ENODATA:
3098	return scnprintf(buf: msg, size, fmt: "Cannot collect data source with the load latency event alone. "
3099	"Please add an auxiliary event in front of the load latency event.");
3100	default:
3101	break;
3102	}
3103
3104	ret = arch_evsel__open_strerror(evsel, msg, size);
3105	if (ret)
3106	return ret;
3107
3108	return scnprintf(buf: msg, size,
3109	fmt: "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3110	"/bin/dmesg | grep -i perf may provide additional information.\n",
3111	err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3112	}
3113
3114	struct perf_env *evsel__env(struct evsel *evsel)
3115	{
3116	if (evsel && evsel->evlist && evsel->evlist->env)
3117	return evsel->evlist->env;
3118	return &perf_env;
3119	}
3120
3121	static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3122	{
3123	int cpu_map_idx, thread;
3124
3125	for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3126	for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3127	thread++) {
3128	int fd = FD(evsel, cpu_map_idx, thread);
3129
3130	if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3131	cpu_map_idx, thread, fd) < 0)
3132	return -1;
3133	}
3134	}
3135
3136	return 0;
3137	}
3138
3139	int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3140	{
3141	struct perf_cpu_map *cpus = evsel->core.cpus;
3142	struct perf_thread_map *threads = evsel->core.threads;
3143
3144	if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3145	return -ENOMEM;
3146
3147	return store_evsel_ids(evsel, evlist);
3148	}
3149
3150	void evsel__zero_per_pkg(struct evsel *evsel)
3151	{
3152	struct hashmap_entry *cur;
3153	size_t bkt;
3154
3155	if (evsel->per_pkg_mask) {
3156	hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3157	zfree(&cur->pkey);
3158
3159	hashmap__clear(evsel->per_pkg_mask);
3160	}
3161	}
3162
3163	/**
3164	* evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3165	* will be false on hybrid systems for hardware and legacy
3166	* cache events.
3167	*/
3168	bool evsel__is_hybrid(const struct evsel *evsel)
3169	{
3170	if (perf_pmus__num_core_pmus() == 1)
3171	return false;
3172
3173	return evsel->core.is_pmu_core;
3174	}
3175
3176	struct evsel *evsel__leader(const struct evsel *evsel)
3177	{
3178	return container_of(evsel->core.leader, struct evsel, core);
3179	}
3180
3181	bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3182	{
3183	return evsel->core.leader == &leader->core;
3184	}
3185
3186	bool evsel__is_leader(struct evsel *evsel)
3187	{
3188	return evsel__has_leader(evsel, leader: evsel);
3189	}
3190
3191	void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3192	{
3193	evsel->core.leader = &leader->core;
3194	}
3195
3196	int evsel__source_count(const struct evsel *evsel)
3197	{
3198	struct evsel *pos;
3199	int count = 0;
3200
3201	evlist__for_each_entry(evsel->evlist, pos) {
3202	if (pos->metric_leader == evsel)
3203	count++;
3204	}
3205	return count;
3206	}
3207
3208	bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3209	{
3210	return false;
3211	}
3212
3213	/*
3214	* Remove an event from a given group (leader).
3215	* Some events, e.g., perf metrics Topdown events,
3216	* must always be grouped. Ignore the events.
3217	*/
3218	void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3219	{
3220	if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3221	evsel__set_leader(evsel, leader: evsel);
3222	evsel->core.nr_members = 0;
3223	leader->core.nr_members--;
3224	}
3225	}
3226