xref: /linux/tools/perf/builtin-trace.c (revision 61d007138a44dc69265cf948eadd1759edbd3875)
1 /*
2  * builtin-trace.c
3  *
4  * Builtin 'trace' command:
5  *
6  * Display a continuously updated trace of any workload, CPU, specific PID,
7  * system wide, etc.  Default format is loosely strace like, but any other
8  * event may be specified using --event.
9  *
10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
11  *
12  * Initially based on the 'trace' prototype by Thomas Gleixner:
13  *
14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
15  *
16  * Released under the GPL v2. (and only v2, not any later version)
17  */
18 
19 #include <traceevent/event-parse.h>
20 #include <api/fs/tracing_path.h>
21 #include <bpf/bpf.h>
22 #include "builtin.h"
23 #include "util/cgroup.h"
24 #include "util/color.h"
25 #include "util/debug.h"
26 #include "util/env.h"
27 #include "util/event.h"
28 #include "util/evlist.h"
29 #include <subcmd/exec-cmd.h>
30 #include "util/machine.h"
31 #include "util/path.h"
32 #include "util/session.h"
33 #include "util/thread.h"
34 #include <subcmd/parse-options.h>
35 #include "util/strlist.h"
36 #include "util/intlist.h"
37 #include "util/thread_map.h"
38 #include "util/stat.h"
39 #include "trace/beauty/beauty.h"
40 #include "trace-event.h"
41 #include "util/parse-events.h"
42 #include "util/bpf-loader.h"
43 #include "callchain.h"
44 #include "print_binary.h"
45 #include "string2.h"
46 #include "syscalltbl.h"
47 #include "rb_resort.h"
48 
49 #include <errno.h>
50 #include <inttypes.h>
51 #include <poll.h>
52 #include <signal.h>
53 #include <stdlib.h>
54 #include <string.h>
55 #include <linux/err.h>
56 #include <linux/filter.h>
57 #include <linux/kernel.h>
58 #include <linux/random.h>
59 #include <linux/stringify.h>
60 #include <linux/time64.h>
61 #include <fcntl.h>
62 
63 #include "sane_ctype.h"
64 
65 #ifndef O_CLOEXEC
66 # define O_CLOEXEC		02000000
67 #endif
68 
69 #ifndef F_LINUX_SPECIFIC_BASE
70 # define F_LINUX_SPECIFIC_BASE	1024
71 #endif
72 
73 struct trace {
74 	struct perf_tool	tool;
75 	struct syscalltbl	*sctbl;
76 	struct {
77 		int		max;
78 		struct syscall  *table;
79 		struct bpf_map  *map;
80 		struct {
81 			struct perf_evsel *sys_enter,
82 					  *sys_exit,
83 					  *augmented;
84 		}		events;
85 	} syscalls;
86 	struct record_opts	opts;
87 	struct perf_evlist	*evlist;
88 	struct machine		*host;
89 	struct thread		*current;
90 	struct cgroup		*cgroup;
91 	u64			base_time;
92 	FILE			*output;
93 	unsigned long		nr_events;
94 	unsigned long		nr_events_printed;
95 	unsigned long		max_events;
96 	struct strlist		*ev_qualifier;
97 	struct {
98 		size_t		nr;
99 		int		*entries;
100 	}			ev_qualifier_ids;
101 	struct {
102 		size_t		nr;
103 		pid_t		*entries;
104 		struct bpf_map  *map;
105 	}			filter_pids;
106 	double			duration_filter;
107 	double			runtime_ms;
108 	struct {
109 		u64		vfs_getname,
110 				proc_getname;
111 	} stats;
112 	unsigned int		max_stack;
113 	unsigned int		min_stack;
114 	bool			sort_events;
115 	bool			raw_augmented_syscalls;
116 	bool			not_ev_qualifier;
117 	bool			live;
118 	bool			full_time;
119 	bool			sched;
120 	bool			multiple_threads;
121 	bool			summary;
122 	bool			summary_only;
123 	bool			failure_only;
124 	bool			show_comm;
125 	bool			print_sample;
126 	bool			show_tool_stats;
127 	bool			trace_syscalls;
128 	bool			kernel_syscallchains;
129 	bool			force;
130 	bool			vfs_getname;
131 	int			trace_pgfaults;
132 	struct {
133 		struct ordered_events	data;
134 		u64			last;
135 	} oe;
136 };
137 
138 struct tp_field {
139 	int offset;
140 	union {
141 		u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
142 		void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
143 	};
144 };
145 
146 #define TP_UINT_FIELD(bits) \
147 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
148 { \
149 	u##bits value; \
150 	memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
151 	return value;  \
152 }
153 
154 TP_UINT_FIELD(8);
155 TP_UINT_FIELD(16);
156 TP_UINT_FIELD(32);
157 TP_UINT_FIELD(64);
158 
159 #define TP_UINT_FIELD__SWAPPED(bits) \
160 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
161 { \
162 	u##bits value; \
163 	memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
164 	return bswap_##bits(value);\
165 }
166 
167 TP_UINT_FIELD__SWAPPED(16);
168 TP_UINT_FIELD__SWAPPED(32);
169 TP_UINT_FIELD__SWAPPED(64);
170 
171 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
172 {
173 	field->offset = offset;
174 
175 	switch (size) {
176 	case 1:
177 		field->integer = tp_field__u8;
178 		break;
179 	case 2:
180 		field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
181 		break;
182 	case 4:
183 		field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
184 		break;
185 	case 8:
186 		field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
187 		break;
188 	default:
189 		return -1;
190 	}
191 
192 	return 0;
193 }
194 
195 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
196 {
197 	return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
198 }
199 
200 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
201 {
202 	return sample->raw_data + field->offset;
203 }
204 
205 static int __tp_field__init_ptr(struct tp_field *field, int offset)
206 {
207 	field->offset = offset;
208 	field->pointer = tp_field__ptr;
209 	return 0;
210 }
211 
212 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
213 {
214 	return __tp_field__init_ptr(field, format_field->offset);
215 }
216 
217 struct syscall_tp {
218 	struct tp_field id;
219 	union {
220 		struct tp_field args, ret;
221 	};
222 };
223 
224 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
225 					  struct tp_field *field,
226 					  const char *name)
227 {
228 	struct tep_format_field *format_field = perf_evsel__field(evsel, name);
229 
230 	if (format_field == NULL)
231 		return -1;
232 
233 	return tp_field__init_uint(field, format_field, evsel->needs_swap);
234 }
235 
236 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
237 	({ struct syscall_tp *sc = evsel->priv;\
238 	   perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
239 
240 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
241 					 struct tp_field *field,
242 					 const char *name)
243 {
244 	struct tep_format_field *format_field = perf_evsel__field(evsel, name);
245 
246 	if (format_field == NULL)
247 		return -1;
248 
249 	return tp_field__init_ptr(field, format_field);
250 }
251 
252 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
253 	({ struct syscall_tp *sc = evsel->priv;\
254 	   perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
255 
256 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
257 {
258 	zfree(&evsel->priv);
259 	perf_evsel__delete(evsel);
260 }
261 
262 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
263 {
264 	struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
265 
266 	if (evsel->priv != NULL) {
267 		if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
268 		    perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
269 			goto out_delete;
270 		return 0;
271 	}
272 
273 	return -ENOMEM;
274 out_delete:
275 	zfree(&evsel->priv);
276 	return -ENOENT;
277 }
278 
279 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel)
280 {
281 	struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
282 
283 	if (evsel->priv != NULL) {       /* field, sizeof_field, offsetof_field */
284 		if (__tp_field__init_uint(&sc->id, sizeof(long), sizeof(long long), evsel->needs_swap))
285 			goto out_delete;
286 
287 		return 0;
288 	}
289 
290 	return -ENOMEM;
291 out_delete:
292 	zfree(&evsel->priv);
293 	return -EINVAL;
294 }
295 
296 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
297 {
298 	struct syscall_tp *sc = evsel->priv;
299 
300 	return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
301 }
302 
303 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
304 {
305 	struct syscall_tp *sc = evsel->priv;
306 
307 	return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
308 }
309 
310 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
311 {
312 	evsel->priv = malloc(sizeof(struct syscall_tp));
313 	if (evsel->priv != NULL) {
314 		if (perf_evsel__init_sc_tp_uint_field(evsel, id))
315 			goto out_delete;
316 
317 		evsel->handler = handler;
318 		return 0;
319 	}
320 
321 	return -ENOMEM;
322 
323 out_delete:
324 	zfree(&evsel->priv);
325 	return -ENOENT;
326 }
327 
328 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
329 {
330 	struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
331 
332 	/* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
333 	if (IS_ERR(evsel))
334 		evsel = perf_evsel__newtp("syscalls", direction);
335 
336 	if (IS_ERR(evsel))
337 		return NULL;
338 
339 	if (perf_evsel__init_raw_syscall_tp(evsel, handler))
340 		goto out_delete;
341 
342 	return evsel;
343 
344 out_delete:
345 	perf_evsel__delete_priv(evsel);
346 	return NULL;
347 }
348 
349 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
350 	({ struct syscall_tp *fields = evsel->priv; \
351 	   fields->name.integer(&fields->name, sample); })
352 
353 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
354 	({ struct syscall_tp *fields = evsel->priv; \
355 	   fields->name.pointer(&fields->name, sample); })
356 
357 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, int val)
358 {
359 	int idx = val - sa->offset;
360 
361 	if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL)
362 		return scnprintf(bf, size, intfmt, val);
363 
364 	return scnprintf(bf, size, "%s", sa->entries[idx]);
365 }
366 
367 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
368 						const char *intfmt,
369 					        struct syscall_arg *arg)
370 {
371 	return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->val);
372 }
373 
374 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
375 					      struct syscall_arg *arg)
376 {
377 	return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
378 }
379 
380 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
381 
382 struct strarrays {
383 	int		nr_entries;
384 	struct strarray **entries;
385 };
386 
387 #define DEFINE_STRARRAYS(array) struct strarrays strarrays__##array = { \
388 	.nr_entries = ARRAY_SIZE(array), \
389 	.entries = array, \
390 }
391 
392 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
393 					struct syscall_arg *arg)
394 {
395 	struct strarrays *sas = arg->parm;
396 	int i;
397 
398 	for (i = 0; i < sas->nr_entries; ++i) {
399 		struct strarray *sa = sas->entries[i];
400 		int idx = arg->val - sa->offset;
401 
402 		if (idx >= 0 && idx < sa->nr_entries) {
403 			if (sa->entries[idx] == NULL)
404 				break;
405 			return scnprintf(bf, size, "%s", sa->entries[idx]);
406 		}
407 	}
408 
409 	return scnprintf(bf, size, "%d", arg->val);
410 }
411 
412 #ifndef AT_FDCWD
413 #define AT_FDCWD	-100
414 #endif
415 
416 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
417 					   struct syscall_arg *arg)
418 {
419 	int fd = arg->val;
420 
421 	if (fd == AT_FDCWD)
422 		return scnprintf(bf, size, "CWD");
423 
424 	return syscall_arg__scnprintf_fd(bf, size, arg);
425 }
426 
427 #define SCA_FDAT syscall_arg__scnprintf_fd_at
428 
429 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
430 					      struct syscall_arg *arg);
431 
432 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
433 
434 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
435 {
436 	return scnprintf(bf, size, "%#lx", arg->val);
437 }
438 
439 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
440 {
441 	return scnprintf(bf, size, "%d", arg->val);
442 }
443 
444 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
445 {
446 	return scnprintf(bf, size, "%ld", arg->val);
447 }
448 
449 static const char *bpf_cmd[] = {
450 	"MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
451 	"MAP_GET_NEXT_KEY", "PROG_LOAD",
452 };
453 static DEFINE_STRARRAY(bpf_cmd);
454 
455 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
456 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, 1);
457 
458 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
459 static DEFINE_STRARRAY(itimers);
460 
461 static const char *keyctl_options[] = {
462 	"GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
463 	"SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
464 	"INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
465 	"ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
466 	"INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
467 };
468 static DEFINE_STRARRAY(keyctl_options);
469 
470 static const char *whences[] = { "SET", "CUR", "END",
471 #ifdef SEEK_DATA
472 "DATA",
473 #endif
474 #ifdef SEEK_HOLE
475 "HOLE",
476 #endif
477 };
478 static DEFINE_STRARRAY(whences);
479 
480 static const char *fcntl_cmds[] = {
481 	"DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
482 	"SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
483 	"SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
484 	"GETOWNER_UIDS",
485 };
486 static DEFINE_STRARRAY(fcntl_cmds);
487 
488 static const char *fcntl_linux_specific_cmds[] = {
489 	"SETLEASE", "GETLEASE", "NOTIFY", [5] =	"CANCELLK", "DUPFD_CLOEXEC",
490 	"SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
491 	"GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
492 };
493 
494 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, F_LINUX_SPECIFIC_BASE);
495 
496 static struct strarray *fcntl_cmds_arrays[] = {
497 	&strarray__fcntl_cmds,
498 	&strarray__fcntl_linux_specific_cmds,
499 };
500 
501 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
502 
503 static const char *rlimit_resources[] = {
504 	"CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
505 	"MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
506 	"RTTIME",
507 };
508 static DEFINE_STRARRAY(rlimit_resources);
509 
510 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
511 static DEFINE_STRARRAY(sighow);
512 
513 static const char *clockid[] = {
514 	"REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
515 	"MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
516 	"REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
517 };
518 static DEFINE_STRARRAY(clockid);
519 
520 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
521 						 struct syscall_arg *arg)
522 {
523 	size_t printed = 0;
524 	int mode = arg->val;
525 
526 	if (mode == F_OK) /* 0 */
527 		return scnprintf(bf, size, "F");
528 #define	P_MODE(n) \
529 	if (mode & n##_OK) { \
530 		printed += scnprintf(bf + printed, size - printed, "%s", #n); \
531 		mode &= ~n##_OK; \
532 	}
533 
534 	P_MODE(R);
535 	P_MODE(W);
536 	P_MODE(X);
537 #undef P_MODE
538 
539 	if (mode)
540 		printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
541 
542 	return printed;
543 }
544 
545 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
546 
547 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
548 					      struct syscall_arg *arg);
549 
550 #define SCA_FILENAME syscall_arg__scnprintf_filename
551 
552 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
553 						struct syscall_arg *arg)
554 {
555 	int printed = 0, flags = arg->val;
556 
557 #define	P_FLAG(n) \
558 	if (flags & O_##n) { \
559 		printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
560 		flags &= ~O_##n; \
561 	}
562 
563 	P_FLAG(CLOEXEC);
564 	P_FLAG(NONBLOCK);
565 #undef P_FLAG
566 
567 	if (flags)
568 		printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
569 
570 	return printed;
571 }
572 
573 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
574 
575 #ifndef GRND_NONBLOCK
576 #define GRND_NONBLOCK	0x0001
577 #endif
578 #ifndef GRND_RANDOM
579 #define GRND_RANDOM	0x0002
580 #endif
581 
582 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
583 						   struct syscall_arg *arg)
584 {
585 	int printed = 0, flags = arg->val;
586 
587 #define	P_FLAG(n) \
588 	if (flags & GRND_##n) { \
589 		printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
590 		flags &= ~GRND_##n; \
591 	}
592 
593 	P_FLAG(RANDOM);
594 	P_FLAG(NONBLOCK);
595 #undef P_FLAG
596 
597 	if (flags)
598 		printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
599 
600 	return printed;
601 }
602 
603 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
604 
605 #define STRARRAY(name, array) \
606 	  { .scnprintf	= SCA_STRARRAY, \
607 	    .parm	= &strarray__##array, }
608 
609 #include "trace/beauty/arch_errno_names.c"
610 #include "trace/beauty/eventfd.c"
611 #include "trace/beauty/futex_op.c"
612 #include "trace/beauty/futex_val3.c"
613 #include "trace/beauty/mmap.c"
614 #include "trace/beauty/mode_t.c"
615 #include "trace/beauty/msg_flags.c"
616 #include "trace/beauty/open_flags.c"
617 #include "trace/beauty/perf_event_open.c"
618 #include "trace/beauty/pid.c"
619 #include "trace/beauty/sched_policy.c"
620 #include "trace/beauty/seccomp.c"
621 #include "trace/beauty/signum.c"
622 #include "trace/beauty/socket_type.c"
623 #include "trace/beauty/waitid_options.c"
624 
625 struct syscall_arg_fmt {
626 	size_t	   (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
627 	unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
628 	void	   *parm;
629 	const char *name;
630 	bool	   show_zero;
631 };
632 
633 static struct syscall_fmt {
634 	const char *name;
635 	const char *alias;
636 	struct syscall_arg_fmt arg[6];
637 	u8	   nr_args;
638 	bool	   errpid;
639 	bool	   timeout;
640 	bool	   hexret;
641 } syscall_fmts[] = {
642 	{ .name	    = "access",
643 	  .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
644 	{ .name	    = "bind",
645 	  .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
646 	{ .name	    = "bpf",
647 	  .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
648 	{ .name	    = "brk",	    .hexret = true,
649 	  .arg = { [0] = { .scnprintf = SCA_HEX, /* brk */ }, }, },
650 	{ .name     = "clock_gettime",
651 	  .arg = { [0] = STRARRAY(clk_id, clockid), }, },
652 	{ .name	    = "clone",	    .errpid = true, .nr_args = 5,
653 	  .arg = { [0] = { .name = "flags",	    .scnprintf = SCA_CLONE_FLAGS, },
654 		   [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
655 		   [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
656 		   [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
657 		   [4] = { .name = "tls",	    .scnprintf = SCA_HEX, }, }, },
658 	{ .name	    = "close",
659 	  .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
660 	{ .name	    = "connect",
661 	  .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
662 	{ .name	    = "epoll_ctl",
663 	  .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
664 	{ .name	    = "eventfd2",
665 	  .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
666 	{ .name	    = "fchmodat",
667 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
668 	{ .name	    = "fchownat",
669 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
670 	{ .name	    = "fcntl",
671 	  .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
672 			   .parm      = &strarrays__fcntl_cmds_arrays,
673 			   .show_zero = true, },
674 		   [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
675 	{ .name	    = "flock",
676 	  .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
677 	{ .name	    = "fstat", .alias = "newfstat", },
678 	{ .name	    = "fstatat", .alias = "newfstatat", },
679 	{ .name	    = "futex",
680 	  .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
681 		   [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
682 	{ .name	    = "futimesat",
683 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
684 	{ .name	    = "getitimer",
685 	  .arg = { [0] = STRARRAY(which, itimers), }, },
686 	{ .name	    = "getpid",	    .errpid = true, },
687 	{ .name	    = "getpgid",    .errpid = true, },
688 	{ .name	    = "getppid",    .errpid = true, },
689 	{ .name	    = "getrandom",
690 	  .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
691 	{ .name	    = "getrlimit",
692 	  .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
693 	{ .name	    = "gettid",	    .errpid = true, },
694 	{ .name	    = "ioctl",
695 	  .arg = {
696 #if defined(__i386__) || defined(__x86_64__)
697 /*
698  * FIXME: Make this available to all arches.
699  */
700 		   [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
701 		   [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
702 #else
703 		   [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
704 #endif
705 	{ .name	    = "kcmp",	    .nr_args = 5,
706 	  .arg = { [0] = { .name = "pid1",	.scnprintf = SCA_PID, },
707 		   [1] = { .name = "pid2",	.scnprintf = SCA_PID, },
708 		   [2] = { .name = "type",	.scnprintf = SCA_KCMP_TYPE, },
709 		   [3] = { .name = "idx1",	.scnprintf = SCA_KCMP_IDX, },
710 		   [4] = { .name = "idx2",	.scnprintf = SCA_KCMP_IDX, }, }, },
711 	{ .name	    = "keyctl",
712 	  .arg = { [0] = STRARRAY(option, keyctl_options), }, },
713 	{ .name	    = "kill",
714 	  .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
715 	{ .name	    = "linkat",
716 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
717 	{ .name	    = "lseek",
718 	  .arg = { [2] = STRARRAY(whence, whences), }, },
719 	{ .name	    = "lstat", .alias = "newlstat", },
720 	{ .name     = "madvise",
721 	  .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
722 		   [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
723 	{ .name	    = "mkdirat",
724 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
725 	{ .name	    = "mknodat",
726 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
727 	{ .name	    = "mlock",
728 	  .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
729 	{ .name	    = "mlockall",
730 	  .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
731 	{ .name	    = "mmap",	    .hexret = true,
732 /* The standard mmap maps to old_mmap on s390x */
733 #if defined(__s390x__)
734 	.alias = "old_mmap",
735 #endif
736 	  .arg = { [0] = { .scnprintf = SCA_HEX,	/* addr */ },
737 		   [2] = { .scnprintf = SCA_MMAP_PROT,	/* prot */ },
738 		   [3] = { .scnprintf = SCA_MMAP_FLAGS,	/* flags */ }, }, },
739 	{ .name	    = "mount",
740 	  .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
741 		   [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
742 			   .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
743 	{ .name	    = "mprotect",
744 	  .arg = { [0] = { .scnprintf = SCA_HEX,	/* start */ },
745 		   [2] = { .scnprintf = SCA_MMAP_PROT,	/* prot */ }, }, },
746 	{ .name	    = "mq_unlink",
747 	  .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
748 	{ .name	    = "mremap",	    .hexret = true,
749 	  .arg = { [0] = { .scnprintf = SCA_HEX,	  /* addr */ },
750 		   [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ },
751 		   [4] = { .scnprintf = SCA_HEX,	  /* new_addr */ }, }, },
752 	{ .name	    = "munlock",
753 	  .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
754 	{ .name	    = "munmap",
755 	  .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
756 	{ .name	    = "name_to_handle_at",
757 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
758 	{ .name	    = "newfstatat",
759 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
760 	{ .name	    = "open",
761 	  .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
762 	{ .name	    = "open_by_handle_at",
763 	  .arg = { [0] = { .scnprintf = SCA_FDAT,	/* dfd */ },
764 		   [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
765 	{ .name	    = "openat",
766 	  .arg = { [0] = { .scnprintf = SCA_FDAT,	/* dfd */ },
767 		   [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
768 	{ .name	    = "perf_event_open",
769 	  .arg = { [2] = { .scnprintf = SCA_INT,	/* cpu */ },
770 		   [3] = { .scnprintf = SCA_FD,		/* group_fd */ },
771 		   [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
772 	{ .name	    = "pipe2",
773 	  .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
774 	{ .name	    = "pkey_alloc",
775 	  .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,	/* access_rights */ }, }, },
776 	{ .name	    = "pkey_free",
777 	  .arg = { [0] = { .scnprintf = SCA_INT,	/* key */ }, }, },
778 	{ .name	    = "pkey_mprotect",
779 	  .arg = { [0] = { .scnprintf = SCA_HEX,	/* start */ },
780 		   [2] = { .scnprintf = SCA_MMAP_PROT,	/* prot */ },
781 		   [3] = { .scnprintf = SCA_INT,	/* pkey */ }, }, },
782 	{ .name	    = "poll", .timeout = true, },
783 	{ .name	    = "ppoll", .timeout = true, },
784 	{ .name	    = "prctl", .alias = "arch_prctl",
785 	  .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
786 		   [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
787 		   [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
788 	{ .name	    = "pread", .alias = "pread64", },
789 	{ .name	    = "preadv", .alias = "pread", },
790 	{ .name	    = "prlimit64",
791 	  .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
792 	{ .name	    = "pwrite", .alias = "pwrite64", },
793 	{ .name	    = "readlinkat",
794 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
795 	{ .name	    = "recvfrom",
796 	  .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
797 	{ .name	    = "recvmmsg",
798 	  .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
799 	{ .name	    = "recvmsg",
800 	  .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
801 	{ .name	    = "renameat",
802 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
803 		   [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
804 	{ .name	    = "renameat2",
805 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
806 		   [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
807 		   [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
808 	{ .name	    = "rt_sigaction",
809 	  .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
810 	{ .name	    = "rt_sigprocmask",
811 	  .arg = { [0] = STRARRAY(how, sighow), }, },
812 	{ .name	    = "rt_sigqueueinfo",
813 	  .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
814 	{ .name	    = "rt_tgsigqueueinfo",
815 	  .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
816 	{ .name	    = "sched_setscheduler",
817 	  .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
818 	{ .name	    = "seccomp",
819 	  .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,	   /* op */ },
820 		   [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
821 	{ .name	    = "select", .timeout = true, },
822 	{ .name	    = "sendmmsg",
823 	  .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
824 	{ .name	    = "sendmsg",
825 	  .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
826 	{ .name	    = "sendto",
827 	  .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
828 		   [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
829 	{ .name	    = "set_tid_address", .errpid = true, },
830 	{ .name	    = "setitimer",
831 	  .arg = { [0] = STRARRAY(which, itimers), }, },
832 	{ .name	    = "setrlimit",
833 	  .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
834 	{ .name	    = "socket",
835 	  .arg = { [0] = STRARRAY(family, socket_families),
836 		   [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
837 		   [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
838 	{ .name	    = "socketpair",
839 	  .arg = { [0] = STRARRAY(family, socket_families),
840 		   [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
841 		   [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
842 	{ .name	    = "stat", .alias = "newstat", },
843 	{ .name	    = "statx",
844 	  .arg = { [0] = { .scnprintf = SCA_FDAT,	 /* fdat */ },
845 		   [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
846 		   [3] = { .scnprintf = SCA_STATX_MASK,	 /* mask */ }, }, },
847 	{ .name	    = "swapoff",
848 	  .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
849 	{ .name	    = "swapon",
850 	  .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
851 	{ .name	    = "symlinkat",
852 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
853 	{ .name	    = "tgkill",
854 	  .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
855 	{ .name	    = "tkill",
856 	  .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
857 	{ .name     = "umount2", .alias = "umount",
858 	  .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
859 	{ .name	    = "uname", .alias = "newuname", },
860 	{ .name	    = "unlinkat",
861 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
862 	{ .name	    = "utimensat",
863 	  .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
864 	{ .name	    = "wait4",	    .errpid = true,
865 	  .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
866 	{ .name	    = "waitid",	    .errpid = true,
867 	  .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
868 };
869 
870 static int syscall_fmt__cmp(const void *name, const void *fmtp)
871 {
872 	const struct syscall_fmt *fmt = fmtp;
873 	return strcmp(name, fmt->name);
874 }
875 
876 static struct syscall_fmt *syscall_fmt__find(const char *name)
877 {
878 	const int nmemb = ARRAY_SIZE(syscall_fmts);
879 	return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
880 }
881 
882 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
883 {
884 	int i, nmemb = ARRAY_SIZE(syscall_fmts);
885 
886 	for (i = 0; i < nmemb; ++i) {
887 		if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
888 			return &syscall_fmts[i];
889 	}
890 
891 	return NULL;
892 }
893 
894 /*
895  * is_exit: is this "exit" or "exit_group"?
896  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
897  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
898  */
899 struct syscall {
900 	struct tep_event    *tp_format;
901 	int		    nr_args;
902 	int		    args_size;
903 	bool		    is_exit;
904 	bool		    is_open;
905 	struct tep_format_field *args;
906 	const char	    *name;
907 	struct syscall_fmt  *fmt;
908 	struct syscall_arg_fmt *arg_fmt;
909 };
910 
911 /*
912  * We need to have this 'calculated' boolean because in some cases we really
913  * don't know what is the duration of a syscall, for instance, when we start
914  * a session and some threads are waiting for a syscall to finish, say 'poll',
915  * in which case all we can do is to print "( ? ) for duration and for the
916  * start timestamp.
917  */
918 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
919 {
920 	double duration = (double)t / NSEC_PER_MSEC;
921 	size_t printed = fprintf(fp, "(");
922 
923 	if (!calculated)
924 		printed += fprintf(fp, "         ");
925 	else if (duration >= 1.0)
926 		printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
927 	else if (duration >= 0.01)
928 		printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
929 	else
930 		printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
931 	return printed + fprintf(fp, "): ");
932 }
933 
934 /**
935  * filename.ptr: The filename char pointer that will be vfs_getname'd
936  * filename.entry_str_pos: Where to insert the string translated from
937  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
938  * ret_scnprintf: syscall args may set this to a different syscall return
939  *                formatter, for instance, fcntl may return fds, file flags, etc.
940  */
941 struct thread_trace {
942 	u64		  entry_time;
943 	bool		  entry_pending;
944 	unsigned long	  nr_events;
945 	unsigned long	  pfmaj, pfmin;
946 	char		  *entry_str;
947 	double		  runtime_ms;
948 	size_t		  (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
949         struct {
950 		unsigned long ptr;
951 		short int     entry_str_pos;
952 		bool	      pending_open;
953 		unsigned int  namelen;
954 		char	      *name;
955 	} filename;
956 	struct {
957 		int	  max;
958 		char	  **table;
959 	} paths;
960 
961 	struct intlist *syscall_stats;
962 };
963 
964 static struct thread_trace *thread_trace__new(void)
965 {
966 	struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
967 
968 	if (ttrace)
969 		ttrace->paths.max = -1;
970 
971 	ttrace->syscall_stats = intlist__new(NULL);
972 
973 	return ttrace;
974 }
975 
976 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
977 {
978 	struct thread_trace *ttrace;
979 
980 	if (thread == NULL)
981 		goto fail;
982 
983 	if (thread__priv(thread) == NULL)
984 		thread__set_priv(thread, thread_trace__new());
985 
986 	if (thread__priv(thread) == NULL)
987 		goto fail;
988 
989 	ttrace = thread__priv(thread);
990 	++ttrace->nr_events;
991 
992 	return ttrace;
993 fail:
994 	color_fprintf(fp, PERF_COLOR_RED,
995 		      "WARNING: not enough memory, dropping samples!\n");
996 	return NULL;
997 }
998 
999 
1000 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1001 				    size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1002 {
1003 	struct thread_trace *ttrace = thread__priv(arg->thread);
1004 
1005 	ttrace->ret_scnprintf = ret_scnprintf;
1006 }
1007 
1008 #define TRACE_PFMAJ		(1 << 0)
1009 #define TRACE_PFMIN		(1 << 1)
1010 
1011 static const size_t trace__entry_str_size = 2048;
1012 
1013 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1014 {
1015 	struct thread_trace *ttrace = thread__priv(thread);
1016 
1017 	if (fd > ttrace->paths.max) {
1018 		char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
1019 
1020 		if (npath == NULL)
1021 			return -1;
1022 
1023 		if (ttrace->paths.max != -1) {
1024 			memset(npath + ttrace->paths.max + 1, 0,
1025 			       (fd - ttrace->paths.max) * sizeof(char *));
1026 		} else {
1027 			memset(npath, 0, (fd + 1) * sizeof(char *));
1028 		}
1029 
1030 		ttrace->paths.table = npath;
1031 		ttrace->paths.max   = fd;
1032 	}
1033 
1034 	ttrace->paths.table[fd] = strdup(pathname);
1035 
1036 	return ttrace->paths.table[fd] != NULL ? 0 : -1;
1037 }
1038 
1039 static int thread__read_fd_path(struct thread *thread, int fd)
1040 {
1041 	char linkname[PATH_MAX], pathname[PATH_MAX];
1042 	struct stat st;
1043 	int ret;
1044 
1045 	if (thread->pid_ == thread->tid) {
1046 		scnprintf(linkname, sizeof(linkname),
1047 			  "/proc/%d/fd/%d", thread->pid_, fd);
1048 	} else {
1049 		scnprintf(linkname, sizeof(linkname),
1050 			  "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1051 	}
1052 
1053 	if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1054 		return -1;
1055 
1056 	ret = readlink(linkname, pathname, sizeof(pathname));
1057 
1058 	if (ret < 0 || ret > st.st_size)
1059 		return -1;
1060 
1061 	pathname[ret] = '\0';
1062 	return trace__set_fd_pathname(thread, fd, pathname);
1063 }
1064 
1065 static const char *thread__fd_path(struct thread *thread, int fd,
1066 				   struct trace *trace)
1067 {
1068 	struct thread_trace *ttrace = thread__priv(thread);
1069 
1070 	if (ttrace == NULL)
1071 		return NULL;
1072 
1073 	if (fd < 0)
1074 		return NULL;
1075 
1076 	if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
1077 		if (!trace->live)
1078 			return NULL;
1079 		++trace->stats.proc_getname;
1080 		if (thread__read_fd_path(thread, fd))
1081 			return NULL;
1082 	}
1083 
1084 	return ttrace->paths.table[fd];
1085 }
1086 
1087 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1088 {
1089 	int fd = arg->val;
1090 	size_t printed = scnprintf(bf, size, "%d", fd);
1091 	const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1092 
1093 	if (path)
1094 		printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1095 
1096 	return printed;
1097 }
1098 
1099 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1100 {
1101         size_t printed = scnprintf(bf, size, "%d", fd);
1102 	struct thread *thread = machine__find_thread(trace->host, pid, pid);
1103 
1104 	if (thread) {
1105 		const char *path = thread__fd_path(thread, fd, trace);
1106 
1107 		if (path)
1108 			printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1109 
1110 		thread__put(thread);
1111 	}
1112 
1113         return printed;
1114 }
1115 
1116 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1117 					      struct syscall_arg *arg)
1118 {
1119 	int fd = arg->val;
1120 	size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1121 	struct thread_trace *ttrace = thread__priv(arg->thread);
1122 
1123 	if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1124 		zfree(&ttrace->paths.table[fd]);
1125 
1126 	return printed;
1127 }
1128 
1129 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1130 				     unsigned long ptr)
1131 {
1132 	struct thread_trace *ttrace = thread__priv(thread);
1133 
1134 	ttrace->filename.ptr = ptr;
1135 	ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1136 }
1137 
1138 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1139 {
1140 	struct augmented_arg *augmented_arg = arg->augmented.args;
1141 
1142 	return scnprintf(bf, size, "%.*s", augmented_arg->size, augmented_arg->value);
1143 }
1144 
1145 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1146 					      struct syscall_arg *arg)
1147 {
1148 	unsigned long ptr = arg->val;
1149 
1150 	if (arg->augmented.args)
1151 		return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1152 
1153 	if (!arg->trace->vfs_getname)
1154 		return scnprintf(bf, size, "%#x", ptr);
1155 
1156 	thread__set_filename_pos(arg->thread, bf, ptr);
1157 	return 0;
1158 }
1159 
1160 static bool trace__filter_duration(struct trace *trace, double t)
1161 {
1162 	return t < (trace->duration_filter * NSEC_PER_MSEC);
1163 }
1164 
1165 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1166 {
1167 	double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1168 
1169 	return fprintf(fp, "%10.3f ", ts);
1170 }
1171 
1172 /*
1173  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1174  * using ttrace->entry_time for a thread that receives a sys_exit without
1175  * first having received a sys_enter ("poll" issued before tracing session
1176  * starts, lost sys_enter exit due to ring buffer overflow).
1177  */
1178 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1179 {
1180 	if (tstamp > 0)
1181 		return __trace__fprintf_tstamp(trace, tstamp, fp);
1182 
1183 	return fprintf(fp, "         ? ");
1184 }
1185 
1186 static bool done = false;
1187 static bool interrupted = false;
1188 
1189 static void sig_handler(int sig)
1190 {
1191 	done = true;
1192 	interrupted = sig == SIGINT;
1193 }
1194 
1195 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1196 {
1197 	size_t printed = 0;
1198 
1199 	if (trace->multiple_threads) {
1200 		if (trace->show_comm)
1201 			printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1202 		printed += fprintf(fp, "%d ", thread->tid);
1203 	}
1204 
1205 	return printed;
1206 }
1207 
1208 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1209 					u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1210 {
1211 	size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
1212 	printed += fprintf_duration(duration, duration_calculated, fp);
1213 	return printed + trace__fprintf_comm_tid(trace, thread, fp);
1214 }
1215 
1216 static int trace__process_event(struct trace *trace, struct machine *machine,
1217 				union perf_event *event, struct perf_sample *sample)
1218 {
1219 	int ret = 0;
1220 
1221 	switch (event->header.type) {
1222 	case PERF_RECORD_LOST:
1223 		color_fprintf(trace->output, PERF_COLOR_RED,
1224 			      "LOST %" PRIu64 " events!\n", event->lost.lost);
1225 		ret = machine__process_lost_event(machine, event, sample);
1226 		break;
1227 	default:
1228 		ret = machine__process_event(machine, event, sample);
1229 		break;
1230 	}
1231 
1232 	return ret;
1233 }
1234 
1235 static int trace__tool_process(struct perf_tool *tool,
1236 			       union perf_event *event,
1237 			       struct perf_sample *sample,
1238 			       struct machine *machine)
1239 {
1240 	struct trace *trace = container_of(tool, struct trace, tool);
1241 	return trace__process_event(trace, machine, event, sample);
1242 }
1243 
1244 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1245 {
1246 	struct machine *machine = vmachine;
1247 
1248 	if (machine->kptr_restrict_warned)
1249 		return NULL;
1250 
1251 	if (symbol_conf.kptr_restrict) {
1252 		pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1253 			   "Check /proc/sys/kernel/kptr_restrict.\n\n"
1254 			   "Kernel samples will not be resolved.\n");
1255 		machine->kptr_restrict_warned = true;
1256 		return NULL;
1257 	}
1258 
1259 	return machine__resolve_kernel_addr(vmachine, addrp, modp);
1260 }
1261 
1262 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1263 {
1264 	int err = symbol__init(NULL);
1265 
1266 	if (err)
1267 		return err;
1268 
1269 	trace->host = machine__new_host();
1270 	if (trace->host == NULL)
1271 		return -ENOMEM;
1272 
1273 	err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1274 	if (err < 0)
1275 		goto out;
1276 
1277 	err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1278 					    evlist->threads, trace__tool_process, false,
1279 					    1);
1280 out:
1281 	if (err)
1282 		symbol__exit();
1283 
1284 	return err;
1285 }
1286 
1287 static void trace__symbols__exit(struct trace *trace)
1288 {
1289 	machine__exit(trace->host);
1290 	trace->host = NULL;
1291 
1292 	symbol__exit();
1293 }
1294 
1295 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1296 {
1297 	int idx;
1298 
1299 	if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1300 		nr_args = sc->fmt->nr_args;
1301 
1302 	sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1303 	if (sc->arg_fmt == NULL)
1304 		return -1;
1305 
1306 	for (idx = 0; idx < nr_args; ++idx) {
1307 		if (sc->fmt)
1308 			sc->arg_fmt[idx] = sc->fmt->arg[idx];
1309 	}
1310 
1311 	sc->nr_args = nr_args;
1312 	return 0;
1313 }
1314 
1315 static int syscall__set_arg_fmts(struct syscall *sc)
1316 {
1317 	struct tep_format_field *field, *last_field = NULL;
1318 	int idx = 0, len;
1319 
1320 	for (field = sc->args; field; field = field->next, ++idx) {
1321 		last_field = field;
1322 
1323 		if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1324 			continue;
1325 
1326 		if (strcmp(field->type, "const char *") == 0 &&
1327 			 (strcmp(field->name, "filename") == 0 ||
1328 			  strcmp(field->name, "path") == 0 ||
1329 			  strcmp(field->name, "pathname") == 0))
1330 			sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1331 		else if (field->flags & TEP_FIELD_IS_POINTER)
1332 			sc->arg_fmt[idx].scnprintf = syscall_arg__scnprintf_hex;
1333 		else if (strcmp(field->type, "pid_t") == 0)
1334 			sc->arg_fmt[idx].scnprintf = SCA_PID;
1335 		else if (strcmp(field->type, "umode_t") == 0)
1336 			sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1337 		else if ((strcmp(field->type, "int") == 0 ||
1338 			  strcmp(field->type, "unsigned int") == 0 ||
1339 			  strcmp(field->type, "long") == 0) &&
1340 			 (len = strlen(field->name)) >= 2 &&
1341 			 strcmp(field->name + len - 2, "fd") == 0) {
1342 			/*
1343 			 * /sys/kernel/tracing/events/syscalls/sys_enter*
1344 			 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1345 			 * 65 int
1346 			 * 23 unsigned int
1347 			 * 7 unsigned long
1348 			 */
1349 			sc->arg_fmt[idx].scnprintf = SCA_FD;
1350 		}
1351 	}
1352 
1353 	if (last_field)
1354 		sc->args_size = last_field->offset + last_field->size;
1355 
1356 	return 0;
1357 }
1358 
1359 static int trace__read_syscall_info(struct trace *trace, int id)
1360 {
1361 	char tp_name[128];
1362 	struct syscall *sc;
1363 	const char *name = syscalltbl__name(trace->sctbl, id);
1364 
1365 	if (name == NULL)
1366 		return -1;
1367 
1368 	if (id > trace->syscalls.max) {
1369 		struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1370 
1371 		if (nsyscalls == NULL)
1372 			return -1;
1373 
1374 		if (trace->syscalls.max != -1) {
1375 			memset(nsyscalls + trace->syscalls.max + 1, 0,
1376 			       (id - trace->syscalls.max) * sizeof(*sc));
1377 		} else {
1378 			memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1379 		}
1380 
1381 		trace->syscalls.table = nsyscalls;
1382 		trace->syscalls.max   = id;
1383 	}
1384 
1385 	sc = trace->syscalls.table + id;
1386 	sc->name = name;
1387 
1388 	sc->fmt  = syscall_fmt__find(sc->name);
1389 
1390 	snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1391 	sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1392 
1393 	if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1394 		snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1395 		sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1396 	}
1397 
1398 	if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1399 		return -1;
1400 
1401 	if (IS_ERR(sc->tp_format))
1402 		return -1;
1403 
1404 	sc->args = sc->tp_format->format.fields;
1405 	/*
1406 	 * We need to check and discard the first variable '__syscall_nr'
1407 	 * or 'nr' that mean the syscall number. It is needless here.
1408 	 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1409 	 */
1410 	if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1411 		sc->args = sc->args->next;
1412 		--sc->nr_args;
1413 	}
1414 
1415 	sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1416 	sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1417 
1418 	return syscall__set_arg_fmts(sc);
1419 }
1420 
1421 static int trace__validate_ev_qualifier(struct trace *trace)
1422 {
1423 	int err = 0, i;
1424 	size_t nr_allocated;
1425 	struct str_node *pos;
1426 
1427 	trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1428 	trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1429 						 sizeof(trace->ev_qualifier_ids.entries[0]));
1430 
1431 	if (trace->ev_qualifier_ids.entries == NULL) {
1432 		fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1433 		       trace->output);
1434 		err = -EINVAL;
1435 		goto out;
1436 	}
1437 
1438 	nr_allocated = trace->ev_qualifier_ids.nr;
1439 	i = 0;
1440 
1441 	strlist__for_each_entry(pos, trace->ev_qualifier) {
1442 		const char *sc = pos->s;
1443 		int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1444 
1445 		if (id < 0) {
1446 			id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1447 			if (id >= 0)
1448 				goto matches;
1449 
1450 			if (err == 0) {
1451 				fputs("Error:\tInvalid syscall ", trace->output);
1452 				err = -EINVAL;
1453 			} else {
1454 				fputs(", ", trace->output);
1455 			}
1456 
1457 			fputs(sc, trace->output);
1458 		}
1459 matches:
1460 		trace->ev_qualifier_ids.entries[i++] = id;
1461 		if (match_next == -1)
1462 			continue;
1463 
1464 		while (1) {
1465 			id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1466 			if (id < 0)
1467 				break;
1468 			if (nr_allocated == trace->ev_qualifier_ids.nr) {
1469 				void *entries;
1470 
1471 				nr_allocated += 8;
1472 				entries = realloc(trace->ev_qualifier_ids.entries,
1473 						  nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1474 				if (entries == NULL) {
1475 					err = -ENOMEM;
1476 					fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1477 					goto out_free;
1478 				}
1479 				trace->ev_qualifier_ids.entries = entries;
1480 			}
1481 			trace->ev_qualifier_ids.nr++;
1482 			trace->ev_qualifier_ids.entries[i++] = id;
1483 		}
1484 	}
1485 
1486 	if (err < 0) {
1487 		fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1488 		      "\nHint:\tand: 'man syscalls'\n", trace->output);
1489 out_free:
1490 		zfree(&trace->ev_qualifier_ids.entries);
1491 		trace->ev_qualifier_ids.nr = 0;
1492 	}
1493 out:
1494 	return err;
1495 }
1496 
1497 /*
1498  * args is to be interpreted as a series of longs but we need to handle
1499  * 8-byte unaligned accesses. args points to raw_data within the event
1500  * and raw_data is guaranteed to be 8-byte unaligned because it is
1501  * preceded by raw_size which is a u32. So we need to copy args to a temp
1502  * variable to read it. Most notably this avoids extended load instructions
1503  * on unaligned addresses
1504  */
1505 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1506 {
1507 	unsigned long val;
1508 	unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1509 
1510 	memcpy(&val, p, sizeof(val));
1511 	return val;
1512 }
1513 
1514 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1515 				      struct syscall_arg *arg)
1516 {
1517 	if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1518 		return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1519 
1520 	return scnprintf(bf, size, "arg%d: ", arg->idx);
1521 }
1522 
1523 /*
1524  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1525  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1526  * in tools/perf/trace/beauty/mount_flags.c
1527  */
1528 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1529 {
1530 	if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1531 		return sc->arg_fmt[arg->idx].mask_val(arg, val);
1532 
1533 	return val;
1534 }
1535 
1536 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1537 				     struct syscall_arg *arg, unsigned long val)
1538 {
1539 	if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1540 		arg->val = val;
1541 		if (sc->arg_fmt[arg->idx].parm)
1542 			arg->parm = sc->arg_fmt[arg->idx].parm;
1543 		return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1544 	}
1545 	return scnprintf(bf, size, "%ld", val);
1546 }
1547 
1548 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1549 				      unsigned char *args, void *augmented_args, int augmented_args_size,
1550 				      struct trace *trace, struct thread *thread)
1551 {
1552 	size_t printed = 0;
1553 	unsigned long val;
1554 	u8 bit = 1;
1555 	struct syscall_arg arg = {
1556 		.args	= args,
1557 		.augmented = {
1558 			.size = augmented_args_size,
1559 			.args = augmented_args,
1560 		},
1561 		.idx	= 0,
1562 		.mask	= 0,
1563 		.trace  = trace,
1564 		.thread = thread,
1565 	};
1566 	struct thread_trace *ttrace = thread__priv(thread);
1567 
1568 	/*
1569 	 * Things like fcntl will set this in its 'cmd' formatter to pick the
1570 	 * right formatter for the return value (an fd? file flags?), which is
1571 	 * not needed for syscalls that always return a given type, say an fd.
1572 	 */
1573 	ttrace->ret_scnprintf = NULL;
1574 
1575 	if (sc->args != NULL) {
1576 		struct tep_format_field *field;
1577 
1578 		for (field = sc->args; field;
1579 		     field = field->next, ++arg.idx, bit <<= 1) {
1580 			if (arg.mask & bit)
1581 				continue;
1582 
1583 			val = syscall_arg__val(&arg, arg.idx);
1584 			/*
1585 			 * Some syscall args need some mask, most don't and
1586 			 * return val untouched.
1587 			 */
1588 			val = syscall__mask_val(sc, &arg, val);
1589 
1590 			/*
1591  			 * Suppress this argument if its value is zero and
1592  			 * and we don't have a string associated in an
1593  			 * strarray for it.
1594  			 */
1595 			if (val == 0 &&
1596 			    !(sc->arg_fmt &&
1597 			      (sc->arg_fmt[arg.idx].show_zero ||
1598 			       sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1599 			       sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1600 			      sc->arg_fmt[arg.idx].parm))
1601 				continue;
1602 
1603 			printed += scnprintf(bf + printed, size - printed,
1604 					     "%s%s: ", printed ? ", " : "", field->name);
1605 			printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1606 		}
1607 	} else if (IS_ERR(sc->tp_format)) {
1608 		/*
1609 		 * If we managed to read the tracepoint /format file, then we
1610 		 * may end up not having any args, like with gettid(), so only
1611 		 * print the raw args when we didn't manage to read it.
1612 		 */
1613 		while (arg.idx < sc->nr_args) {
1614 			if (arg.mask & bit)
1615 				goto next_arg;
1616 			val = syscall_arg__val(&arg, arg.idx);
1617 			if (printed)
1618 				printed += scnprintf(bf + printed, size - printed, ", ");
1619 			printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1620 			printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1621 next_arg:
1622 			++arg.idx;
1623 			bit <<= 1;
1624 		}
1625 	}
1626 
1627 	return printed;
1628 }
1629 
1630 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1631 				  union perf_event *event,
1632 				  struct perf_sample *sample);
1633 
1634 static struct syscall *trace__syscall_info(struct trace *trace,
1635 					   struct perf_evsel *evsel, int id)
1636 {
1637 
1638 	if (id < 0) {
1639 
1640 		/*
1641 		 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1642 		 * before that, leaving at a higher verbosity level till that is
1643 		 * explained. Reproduced with plain ftrace with:
1644 		 *
1645 		 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1646 		 * grep "NR -1 " /t/trace_pipe
1647 		 *
1648 		 * After generating some load on the machine.
1649  		 */
1650 		if (verbose > 1) {
1651 			static u64 n;
1652 			fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1653 				id, perf_evsel__name(evsel), ++n);
1654 		}
1655 		return NULL;
1656 	}
1657 
1658 	if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1659 	    trace__read_syscall_info(trace, id))
1660 		goto out_cant_read;
1661 
1662 	if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1663 		goto out_cant_read;
1664 
1665 	return &trace->syscalls.table[id];
1666 
1667 out_cant_read:
1668 	if (verbose > 0) {
1669 		fprintf(trace->output, "Problems reading syscall %d", id);
1670 		if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1671 			fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1672 		fputs(" information\n", trace->output);
1673 	}
1674 	return NULL;
1675 }
1676 
1677 static void thread__update_stats(struct thread_trace *ttrace,
1678 				 int id, struct perf_sample *sample)
1679 {
1680 	struct int_node *inode;
1681 	struct stats *stats;
1682 	u64 duration = 0;
1683 
1684 	inode = intlist__findnew(ttrace->syscall_stats, id);
1685 	if (inode == NULL)
1686 		return;
1687 
1688 	stats = inode->priv;
1689 	if (stats == NULL) {
1690 		stats = malloc(sizeof(struct stats));
1691 		if (stats == NULL)
1692 			return;
1693 		init_stats(stats);
1694 		inode->priv = stats;
1695 	}
1696 
1697 	if (ttrace->entry_time && sample->time > ttrace->entry_time)
1698 		duration = sample->time - ttrace->entry_time;
1699 
1700 	update_stats(stats, duration);
1701 }
1702 
1703 static int trace__printf_interrupted_entry(struct trace *trace)
1704 {
1705 	struct thread_trace *ttrace;
1706 	size_t printed;
1707 
1708 	if (trace->failure_only || trace->current == NULL)
1709 		return 0;
1710 
1711 	ttrace = thread__priv(trace->current);
1712 
1713 	if (!ttrace->entry_pending)
1714 		return 0;
1715 
1716 	printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1717 	printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str);
1718 	ttrace->entry_pending = false;
1719 
1720 	++trace->nr_events_printed;
1721 
1722 	return printed;
1723 }
1724 
1725 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1726 				 struct perf_sample *sample, struct thread *thread)
1727 {
1728 	int printed = 0;
1729 
1730 	if (trace->print_sample) {
1731 		double ts = (double)sample->time / NSEC_PER_MSEC;
1732 
1733 		printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1734 				   perf_evsel__name(evsel), ts,
1735 				   thread__comm_str(thread),
1736 				   sample->pid, sample->tid, sample->cpu);
1737 	}
1738 
1739 	return printed;
1740 }
1741 
1742 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, bool raw_augmented)
1743 {
1744 	void *augmented_args = NULL;
1745 	/*
1746 	 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1747 	 * and there we get all 6 syscall args plus the tracepoint common
1748 	 * fields (sizeof(long)) and the syscall_nr (another long). So we check
1749 	 * if that is the case and if so don't look after the sc->args_size,
1750 	 * but always after the full raw_syscalls:sys_enter payload, which is
1751 	 * fixed.
1752 	 *
1753 	 * We'll revisit this later to pass s->args_size to the BPF augmenter
1754 	 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1755 	 * copies only what we need for each syscall, like what happens when we
1756 	 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1757 	 * traffic to just what is needed for each syscall.
1758 	 */
1759 	int args_size = raw_augmented ? (8 * (int)sizeof(long)) : sc->args_size;
1760 
1761 	*augmented_args_size = sample->raw_size - args_size;
1762 	if (*augmented_args_size > 0)
1763 		augmented_args = sample->raw_data + args_size;
1764 
1765 	return augmented_args;
1766 }
1767 
1768 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1769 			    union perf_event *event __maybe_unused,
1770 			    struct perf_sample *sample)
1771 {
1772 	char *msg;
1773 	void *args;
1774 	size_t printed = 0;
1775 	struct thread *thread;
1776 	int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1777 	int augmented_args_size = 0;
1778 	void *augmented_args = NULL;
1779 	struct syscall *sc = trace__syscall_info(trace, evsel, id);
1780 	struct thread_trace *ttrace;
1781 
1782 	if (sc == NULL)
1783 		return -1;
1784 
1785 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1786 	ttrace = thread__trace(thread, trace->output);
1787 	if (ttrace == NULL)
1788 		goto out_put;
1789 
1790 	trace__fprintf_sample(trace, evsel, sample, thread);
1791 
1792 	args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1793 
1794 	if (ttrace->entry_str == NULL) {
1795 		ttrace->entry_str = malloc(trace__entry_str_size);
1796 		if (!ttrace->entry_str)
1797 			goto out_put;
1798 	}
1799 
1800 	if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1801 		trace__printf_interrupted_entry(trace);
1802 	/*
1803 	 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1804 	 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1805 	 * this breaks syscall__augmented_args() check for augmented args, as we calculate
1806 	 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1807 	 * so when handling, say the openat syscall, we end up getting 6 args for the
1808 	 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1809 	 * thinking that the extra 2 u64 args are the augmented filename, so just check
1810 	 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1811 	 */
1812 	if (evsel != trace->syscalls.events.sys_enter)
1813 		augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls);
1814 	ttrace->entry_time = sample->time;
1815 	msg = ttrace->entry_str;
1816 	printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1817 
1818 	printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1819 					   args, augmented_args, augmented_args_size, trace, thread);
1820 
1821 	if (sc->is_exit) {
1822 		if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1823 			trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1824 			fprintf(trace->output, "%-70s)\n", ttrace->entry_str);
1825 		}
1826 	} else {
1827 		ttrace->entry_pending = true;
1828 		/* See trace__vfs_getname & trace__sys_exit */
1829 		ttrace->filename.pending_open = false;
1830 	}
1831 
1832 	if (trace->current != thread) {
1833 		thread__put(trace->current);
1834 		trace->current = thread__get(thread);
1835 	}
1836 	err = 0;
1837 out_put:
1838 	thread__put(thread);
1839 	return err;
1840 }
1841 
1842 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1843 				    struct perf_sample *sample)
1844 {
1845 	struct thread_trace *ttrace;
1846 	struct thread *thread;
1847 	int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1848 	struct syscall *sc = trace__syscall_info(trace, evsel, id);
1849 	char msg[1024];
1850 	void *args, *augmented_args = NULL;
1851 	int augmented_args_size;
1852 
1853 	if (sc == NULL)
1854 		return -1;
1855 
1856 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1857 	ttrace = thread__trace(thread, trace->output);
1858 	/*
1859 	 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1860 	 * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1861 	 */
1862 	if (ttrace == NULL)
1863 		goto out_put;
1864 
1865 	args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1866 	augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls);
1867 	syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1868 	fprintf(trace->output, "%s", msg);
1869 	err = 0;
1870 out_put:
1871 	thread__put(thread);
1872 	return err;
1873 }
1874 
1875 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1876 				    struct perf_sample *sample,
1877 				    struct callchain_cursor *cursor)
1878 {
1879 	struct addr_location al;
1880 	int max_stack = evsel->attr.sample_max_stack ?
1881 			evsel->attr.sample_max_stack :
1882 			trace->max_stack;
1883 	int err;
1884 
1885 	if (machine__resolve(trace->host, &al, sample) < 0)
1886 		return -1;
1887 
1888 	err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1889 	addr_location__put(&al);
1890 	return err;
1891 }
1892 
1893 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1894 {
1895 	/* TODO: user-configurable print_opts */
1896 	const unsigned int print_opts = EVSEL__PRINT_SYM |
1897 				        EVSEL__PRINT_DSO |
1898 				        EVSEL__PRINT_UNKNOWN_AS_ADDR;
1899 
1900 	return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1901 }
1902 
1903 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1904 {
1905 	struct perf_env *env = perf_evsel__env(evsel);
1906 	const char *arch_name = perf_env__arch(env);
1907 
1908 	return arch_syscalls__strerrno(arch_name, err);
1909 }
1910 
1911 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1912 			   union perf_event *event __maybe_unused,
1913 			   struct perf_sample *sample)
1914 {
1915 	long ret;
1916 	u64 duration = 0;
1917 	bool duration_calculated = false;
1918 	struct thread *thread;
1919 	int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0;
1920 	struct syscall *sc = trace__syscall_info(trace, evsel, id);
1921 	struct thread_trace *ttrace;
1922 
1923 	if (sc == NULL)
1924 		return -1;
1925 
1926 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1927 	ttrace = thread__trace(thread, trace->output);
1928 	if (ttrace == NULL)
1929 		goto out_put;
1930 
1931 	trace__fprintf_sample(trace, evsel, sample, thread);
1932 
1933 	if (trace->summary)
1934 		thread__update_stats(ttrace, id, sample);
1935 
1936 	ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1937 
1938 	if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
1939 		trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1940 		ttrace->filename.pending_open = false;
1941 		++trace->stats.vfs_getname;
1942 	}
1943 
1944 	if (ttrace->entry_time) {
1945 		duration = sample->time - ttrace->entry_time;
1946 		if (trace__filter_duration(trace, duration))
1947 			goto out;
1948 		duration_calculated = true;
1949 	} else if (trace->duration_filter)
1950 		goto out;
1951 
1952 	if (sample->callchain) {
1953 		callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1954 		if (callchain_ret == 0) {
1955 			if (callchain_cursor.nr < trace->min_stack)
1956 				goto out;
1957 			callchain_ret = 1;
1958 		}
1959 	}
1960 
1961 	if (trace->summary_only || (ret >= 0 && trace->failure_only))
1962 		goto out;
1963 
1964 	trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
1965 
1966 	if (ttrace->entry_pending) {
1967 		fprintf(trace->output, "%-70s", ttrace->entry_str);
1968 	} else {
1969 		fprintf(trace->output, " ... [");
1970 		color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
1971 		fprintf(trace->output, "]: %s()", sc->name);
1972 	}
1973 
1974 	if (sc->fmt == NULL) {
1975 		if (ret < 0)
1976 			goto errno_print;
1977 signed_print:
1978 		fprintf(trace->output, ") = %ld", ret);
1979 	} else if (ret < 0) {
1980 errno_print: {
1981 		char bf[STRERR_BUFSIZE];
1982 		const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
1983 			   *e = errno_to_name(evsel, -ret);
1984 
1985 		fprintf(trace->output, ") = -1 %s %s", e, emsg);
1986 	}
1987 	} else if (ret == 0 && sc->fmt->timeout)
1988 		fprintf(trace->output, ") = 0 Timeout");
1989 	else if (ttrace->ret_scnprintf) {
1990 		char bf[1024];
1991 		struct syscall_arg arg = {
1992 			.val	= ret,
1993 			.thread	= thread,
1994 			.trace	= trace,
1995 		};
1996 		ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
1997 		ttrace->ret_scnprintf = NULL;
1998 		fprintf(trace->output, ") = %s", bf);
1999 	} else if (sc->fmt->hexret)
2000 		fprintf(trace->output, ") = %#lx", ret);
2001 	else if (sc->fmt->errpid) {
2002 		struct thread *child = machine__find_thread(trace->host, ret, ret);
2003 
2004 		if (child != NULL) {
2005 			fprintf(trace->output, ") = %ld", ret);
2006 			if (child->comm_set)
2007 				fprintf(trace->output, " (%s)", thread__comm_str(child));
2008 			thread__put(child);
2009 		}
2010 	} else
2011 		goto signed_print;
2012 
2013 	fputc('\n', trace->output);
2014 
2015 	/*
2016 	 * We only consider an 'event' for the sake of --max-events a non-filtered
2017 	 * sys_enter + sys_exit and other tracepoint events.
2018 	 */
2019 	if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2020 		interrupted = true;
2021 
2022 	if (callchain_ret > 0)
2023 		trace__fprintf_callchain(trace, sample);
2024 	else if (callchain_ret < 0)
2025 		pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2026 out:
2027 	ttrace->entry_pending = false;
2028 	err = 0;
2029 out_put:
2030 	thread__put(thread);
2031 	return err;
2032 }
2033 
2034 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2035 			      union perf_event *event __maybe_unused,
2036 			      struct perf_sample *sample)
2037 {
2038 	struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2039 	struct thread_trace *ttrace;
2040 	size_t filename_len, entry_str_len, to_move;
2041 	ssize_t remaining_space;
2042 	char *pos;
2043 	const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2044 
2045 	if (!thread)
2046 		goto out;
2047 
2048 	ttrace = thread__priv(thread);
2049 	if (!ttrace)
2050 		goto out_put;
2051 
2052 	filename_len = strlen(filename);
2053 	if (filename_len == 0)
2054 		goto out_put;
2055 
2056 	if (ttrace->filename.namelen < filename_len) {
2057 		char *f = realloc(ttrace->filename.name, filename_len + 1);
2058 
2059 		if (f == NULL)
2060 			goto out_put;
2061 
2062 		ttrace->filename.namelen = filename_len;
2063 		ttrace->filename.name = f;
2064 	}
2065 
2066 	strcpy(ttrace->filename.name, filename);
2067 	ttrace->filename.pending_open = true;
2068 
2069 	if (!ttrace->filename.ptr)
2070 		goto out_put;
2071 
2072 	entry_str_len = strlen(ttrace->entry_str);
2073 	remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2074 	if (remaining_space <= 0)
2075 		goto out_put;
2076 
2077 	if (filename_len > (size_t)remaining_space) {
2078 		filename += filename_len - remaining_space;
2079 		filename_len = remaining_space;
2080 	}
2081 
2082 	to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2083 	pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2084 	memmove(pos + filename_len, pos, to_move);
2085 	memcpy(pos, filename, filename_len);
2086 
2087 	ttrace->filename.ptr = 0;
2088 	ttrace->filename.entry_str_pos = 0;
2089 out_put:
2090 	thread__put(thread);
2091 out:
2092 	return 0;
2093 }
2094 
2095 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2096 				     union perf_event *event __maybe_unused,
2097 				     struct perf_sample *sample)
2098 {
2099         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2100 	double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2101 	struct thread *thread = machine__findnew_thread(trace->host,
2102 							sample->pid,
2103 							sample->tid);
2104 	struct thread_trace *ttrace = thread__trace(thread, trace->output);
2105 
2106 	if (ttrace == NULL)
2107 		goto out_dump;
2108 
2109 	ttrace->runtime_ms += runtime_ms;
2110 	trace->runtime_ms += runtime_ms;
2111 out_put:
2112 	thread__put(thread);
2113 	return 0;
2114 
2115 out_dump:
2116 	fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2117 	       evsel->name,
2118 	       perf_evsel__strval(evsel, sample, "comm"),
2119 	       (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2120 	       runtime,
2121 	       perf_evsel__intval(evsel, sample, "vruntime"));
2122 	goto out_put;
2123 }
2124 
2125 static int bpf_output__printer(enum binary_printer_ops op,
2126 			       unsigned int val, void *extra __maybe_unused, FILE *fp)
2127 {
2128 	unsigned char ch = (unsigned char)val;
2129 
2130 	switch (op) {
2131 	case BINARY_PRINT_CHAR_DATA:
2132 		return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2133 	case BINARY_PRINT_DATA_BEGIN:
2134 	case BINARY_PRINT_LINE_BEGIN:
2135 	case BINARY_PRINT_ADDR:
2136 	case BINARY_PRINT_NUM_DATA:
2137 	case BINARY_PRINT_NUM_PAD:
2138 	case BINARY_PRINT_SEP:
2139 	case BINARY_PRINT_CHAR_PAD:
2140 	case BINARY_PRINT_LINE_END:
2141 	case BINARY_PRINT_DATA_END:
2142 	default:
2143 		break;
2144 	}
2145 
2146 	return 0;
2147 }
2148 
2149 static void bpf_output__fprintf(struct trace *trace,
2150 				struct perf_sample *sample)
2151 {
2152 	binary__fprintf(sample->raw_data, sample->raw_size, 8,
2153 			bpf_output__printer, NULL, trace->output);
2154 	++trace->nr_events_printed;
2155 }
2156 
2157 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2158 				union perf_event *event __maybe_unused,
2159 				struct perf_sample *sample)
2160 {
2161 	struct thread *thread;
2162 	int callchain_ret = 0;
2163 	/*
2164 	 * Check if we called perf_evsel__disable(evsel) due to, for instance,
2165 	 * this event's max_events having been hit and this is an entry coming
2166 	 * from the ring buffer that we should discard, since the max events
2167 	 * have already been considered/printed.
2168 	 */
2169 	if (evsel->disabled)
2170 		return 0;
2171 
2172 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2173 
2174 	if (sample->callchain) {
2175 		callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2176 		if (callchain_ret == 0) {
2177 			if (callchain_cursor.nr < trace->min_stack)
2178 				goto out;
2179 			callchain_ret = 1;
2180 		}
2181 	}
2182 
2183 	trace__printf_interrupted_entry(trace);
2184 	trace__fprintf_tstamp(trace, sample->time, trace->output);
2185 
2186 	if (trace->trace_syscalls)
2187 		fprintf(trace->output, "(         ): ");
2188 
2189 	if (thread)
2190 		trace__fprintf_comm_tid(trace, thread, trace->output);
2191 
2192 	if (evsel == trace->syscalls.events.augmented) {
2193 		int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2194 		struct syscall *sc = trace__syscall_info(trace, evsel, id);
2195 
2196 		if (sc) {
2197 			fprintf(trace->output, "%s(", sc->name);
2198 			trace__fprintf_sys_enter(trace, evsel, sample);
2199 			fputc(')', trace->output);
2200 			goto newline;
2201 		}
2202 
2203 		/*
2204 		 * XXX: Not having the associated syscall info or not finding/adding
2205 		 * 	the thread should never happen, but if it does...
2206 		 * 	fall thru and print it as a bpf_output event.
2207 		 */
2208 	}
2209 
2210 	fprintf(trace->output, "%s:", evsel->name);
2211 
2212 	if (perf_evsel__is_bpf_output(evsel)) {
2213 		bpf_output__fprintf(trace, sample);
2214 	} else if (evsel->tp_format) {
2215 		if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2216 		    trace__fprintf_sys_enter(trace, evsel, sample)) {
2217 			event_format__fprintf(evsel->tp_format, sample->cpu,
2218 					      sample->raw_data, sample->raw_size,
2219 					      trace->output);
2220 			++trace->nr_events_printed;
2221 
2222 			if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2223 				perf_evsel__disable(evsel);
2224 				perf_evsel__close(evsel);
2225 			}
2226 		}
2227 	}
2228 
2229 newline:
2230 	fprintf(trace->output, "\n");
2231 
2232 	if (callchain_ret > 0)
2233 		trace__fprintf_callchain(trace, sample);
2234 	else if (callchain_ret < 0)
2235 		pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2236 out:
2237 	thread__put(thread);
2238 	return 0;
2239 }
2240 
2241 static void print_location(FILE *f, struct perf_sample *sample,
2242 			   struct addr_location *al,
2243 			   bool print_dso, bool print_sym)
2244 {
2245 
2246 	if ((verbose > 0 || print_dso) && al->map)
2247 		fprintf(f, "%s@", al->map->dso->long_name);
2248 
2249 	if ((verbose > 0 || print_sym) && al->sym)
2250 		fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2251 			al->addr - al->sym->start);
2252 	else if (al->map)
2253 		fprintf(f, "0x%" PRIx64, al->addr);
2254 	else
2255 		fprintf(f, "0x%" PRIx64, sample->addr);
2256 }
2257 
2258 static int trace__pgfault(struct trace *trace,
2259 			  struct perf_evsel *evsel,
2260 			  union perf_event *event __maybe_unused,
2261 			  struct perf_sample *sample)
2262 {
2263 	struct thread *thread;
2264 	struct addr_location al;
2265 	char map_type = 'd';
2266 	struct thread_trace *ttrace;
2267 	int err = -1;
2268 	int callchain_ret = 0;
2269 
2270 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2271 
2272 	if (sample->callchain) {
2273 		callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2274 		if (callchain_ret == 0) {
2275 			if (callchain_cursor.nr < trace->min_stack)
2276 				goto out_put;
2277 			callchain_ret = 1;
2278 		}
2279 	}
2280 
2281 	ttrace = thread__trace(thread, trace->output);
2282 	if (ttrace == NULL)
2283 		goto out_put;
2284 
2285 	if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2286 		ttrace->pfmaj++;
2287 	else
2288 		ttrace->pfmin++;
2289 
2290 	if (trace->summary_only)
2291 		goto out;
2292 
2293 	thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2294 
2295 	trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2296 
2297 	fprintf(trace->output, "%sfault [",
2298 		evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2299 		"maj" : "min");
2300 
2301 	print_location(trace->output, sample, &al, false, true);
2302 
2303 	fprintf(trace->output, "] => ");
2304 
2305 	thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2306 
2307 	if (!al.map) {
2308 		thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2309 
2310 		if (al.map)
2311 			map_type = 'x';
2312 		else
2313 			map_type = '?';
2314 	}
2315 
2316 	print_location(trace->output, sample, &al, true, false);
2317 
2318 	fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2319 
2320 	if (callchain_ret > 0)
2321 		trace__fprintf_callchain(trace, sample);
2322 	else if (callchain_ret < 0)
2323 		pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2324 
2325 	++trace->nr_events_printed;
2326 out:
2327 	err = 0;
2328 out_put:
2329 	thread__put(thread);
2330 	return err;
2331 }
2332 
2333 static void trace__set_base_time(struct trace *trace,
2334 				 struct perf_evsel *evsel,
2335 				 struct perf_sample *sample)
2336 {
2337 	/*
2338 	 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2339 	 * and don't use sample->time unconditionally, we may end up having
2340 	 * some other event in the future without PERF_SAMPLE_TIME for good
2341 	 * reason, i.e. we may not be interested in its timestamps, just in
2342 	 * it taking place, picking some piece of information when it
2343 	 * appears in our event stream (vfs_getname comes to mind).
2344 	 */
2345 	if (trace->base_time == 0 && !trace->full_time &&
2346 	    (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2347 		trace->base_time = sample->time;
2348 }
2349 
2350 static int trace__process_sample(struct perf_tool *tool,
2351 				 union perf_event *event,
2352 				 struct perf_sample *sample,
2353 				 struct perf_evsel *evsel,
2354 				 struct machine *machine __maybe_unused)
2355 {
2356 	struct trace *trace = container_of(tool, struct trace, tool);
2357 	struct thread *thread;
2358 	int err = 0;
2359 
2360 	tracepoint_handler handler = evsel->handler;
2361 
2362 	thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2363 	if (thread && thread__is_filtered(thread))
2364 		goto out;
2365 
2366 	trace__set_base_time(trace, evsel, sample);
2367 
2368 	if (handler) {
2369 		++trace->nr_events;
2370 		handler(trace, evsel, event, sample);
2371 	}
2372 out:
2373 	thread__put(thread);
2374 	return err;
2375 }
2376 
2377 static int trace__record(struct trace *trace, int argc, const char **argv)
2378 {
2379 	unsigned int rec_argc, i, j;
2380 	const char **rec_argv;
2381 	const char * const record_args[] = {
2382 		"record",
2383 		"-R",
2384 		"-m", "1024",
2385 		"-c", "1",
2386 	};
2387 
2388 	const char * const sc_args[] = { "-e", };
2389 	unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2390 	const char * const majpf_args[] = { "-e", "major-faults" };
2391 	unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2392 	const char * const minpf_args[] = { "-e", "minor-faults" };
2393 	unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2394 
2395 	/* +1 is for the event string below */
2396 	rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2397 		majpf_args_nr + minpf_args_nr + argc;
2398 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
2399 
2400 	if (rec_argv == NULL)
2401 		return -ENOMEM;
2402 
2403 	j = 0;
2404 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
2405 		rec_argv[j++] = record_args[i];
2406 
2407 	if (trace->trace_syscalls) {
2408 		for (i = 0; i < sc_args_nr; i++)
2409 			rec_argv[j++] = sc_args[i];
2410 
2411 		/* event string may be different for older kernels - e.g., RHEL6 */
2412 		if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2413 			rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2414 		else if (is_valid_tracepoint("syscalls:sys_enter"))
2415 			rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2416 		else {
2417 			pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2418 			free(rec_argv);
2419 			return -1;
2420 		}
2421 	}
2422 
2423 	if (trace->trace_pgfaults & TRACE_PFMAJ)
2424 		for (i = 0; i < majpf_args_nr; i++)
2425 			rec_argv[j++] = majpf_args[i];
2426 
2427 	if (trace->trace_pgfaults & TRACE_PFMIN)
2428 		for (i = 0; i < minpf_args_nr; i++)
2429 			rec_argv[j++] = minpf_args[i];
2430 
2431 	for (i = 0; i < (unsigned int)argc; i++)
2432 		rec_argv[j++] = argv[i];
2433 
2434 	return cmd_record(j, rec_argv);
2435 }
2436 
2437 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2438 
2439 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2440 {
2441 	struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2442 
2443 	if (IS_ERR(evsel))
2444 		return false;
2445 
2446 	if (perf_evsel__field(evsel, "pathname") == NULL) {
2447 		perf_evsel__delete(evsel);
2448 		return false;
2449 	}
2450 
2451 	evsel->handler = trace__vfs_getname;
2452 	perf_evlist__add(evlist, evsel);
2453 	return true;
2454 }
2455 
2456 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2457 {
2458 	struct perf_evsel *evsel;
2459 	struct perf_event_attr attr = {
2460 		.type = PERF_TYPE_SOFTWARE,
2461 		.mmap_data = 1,
2462 	};
2463 
2464 	attr.config = config;
2465 	attr.sample_period = 1;
2466 
2467 	event_attr_init(&attr);
2468 
2469 	evsel = perf_evsel__new(&attr);
2470 	if (evsel)
2471 		evsel->handler = trace__pgfault;
2472 
2473 	return evsel;
2474 }
2475 
2476 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2477 {
2478 	const u32 type = event->header.type;
2479 	struct perf_evsel *evsel;
2480 
2481 	if (type != PERF_RECORD_SAMPLE) {
2482 		trace__process_event(trace, trace->host, event, sample);
2483 		return;
2484 	}
2485 
2486 	evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2487 	if (evsel == NULL) {
2488 		fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2489 		return;
2490 	}
2491 
2492 	trace__set_base_time(trace, evsel, sample);
2493 
2494 	if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2495 	    sample->raw_data == NULL) {
2496 		fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2497 		       perf_evsel__name(evsel), sample->tid,
2498 		       sample->cpu, sample->raw_size);
2499 	} else {
2500 		tracepoint_handler handler = evsel->handler;
2501 		handler(trace, evsel, event, sample);
2502 	}
2503 
2504 	if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2505 		interrupted = true;
2506 }
2507 
2508 static int trace__add_syscall_newtp(struct trace *trace)
2509 {
2510 	int ret = -1;
2511 	struct perf_evlist *evlist = trace->evlist;
2512 	struct perf_evsel *sys_enter, *sys_exit;
2513 
2514 	sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2515 	if (sys_enter == NULL)
2516 		goto out;
2517 
2518 	if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2519 		goto out_delete_sys_enter;
2520 
2521 	sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2522 	if (sys_exit == NULL)
2523 		goto out_delete_sys_enter;
2524 
2525 	if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2526 		goto out_delete_sys_exit;
2527 
2528 	perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2529 	perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2530 
2531 	perf_evlist__add(evlist, sys_enter);
2532 	perf_evlist__add(evlist, sys_exit);
2533 
2534 	if (callchain_param.enabled && !trace->kernel_syscallchains) {
2535 		/*
2536 		 * We're interested only in the user space callchain
2537 		 * leading to the syscall, allow overriding that for
2538 		 * debugging reasons using --kernel_syscall_callchains
2539 		 */
2540 		sys_exit->attr.exclude_callchain_kernel = 1;
2541 	}
2542 
2543 	trace->syscalls.events.sys_enter = sys_enter;
2544 	trace->syscalls.events.sys_exit  = sys_exit;
2545 
2546 	ret = 0;
2547 out:
2548 	return ret;
2549 
2550 out_delete_sys_exit:
2551 	perf_evsel__delete_priv(sys_exit);
2552 out_delete_sys_enter:
2553 	perf_evsel__delete_priv(sys_enter);
2554 	goto out;
2555 }
2556 
2557 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2558 {
2559 	int err = -1;
2560 	struct perf_evsel *sys_exit;
2561 	char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2562 						trace->ev_qualifier_ids.nr,
2563 						trace->ev_qualifier_ids.entries);
2564 
2565 	if (filter == NULL)
2566 		goto out_enomem;
2567 
2568 	if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2569 					  filter)) {
2570 		sys_exit = trace->syscalls.events.sys_exit;
2571 		err = perf_evsel__append_tp_filter(sys_exit, filter);
2572 	}
2573 
2574 	free(filter);
2575 out:
2576 	return err;
2577 out_enomem:
2578 	errno = ENOMEM;
2579 	goto out;
2580 }
2581 
2582 #ifdef HAVE_LIBBPF_SUPPORT
2583 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2584 {
2585 	int fd = bpf_map__fd(trace->syscalls.map);
2586 	bool value = !trace->not_ev_qualifier;
2587 	int err = 0;
2588 	size_t i;
2589 
2590 	for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2591 		int key = trace->ev_qualifier_ids.entries[i];
2592 
2593 		err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2594 		if (err)
2595 			break;
2596 	}
2597 
2598 	return err;
2599 }
2600 
2601 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2602 {
2603 	int fd = bpf_map__fd(trace->syscalls.map);
2604 	int err = 0, key;
2605 
2606 	for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2607 		err = bpf_map_update_elem(fd, &key, &enabled, BPF_ANY);
2608 		if (err)
2609 			break;
2610 	}
2611 
2612 	return err;
2613 }
2614 
2615 static int trace__init_syscalls_bpf_map(struct trace *trace)
2616 {
2617 	bool enabled = true;
2618 
2619 	if (trace->ev_qualifier_ids.nr)
2620 		enabled = trace->not_ev_qualifier;
2621 
2622 	return __trace__init_syscalls_bpf_map(trace, enabled);
2623 }
2624 #else
2625 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2626 {
2627 	return 0;
2628 }
2629 
2630 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2631 {
2632 	return 0;
2633 }
2634 #endif // HAVE_LIBBPF_SUPPORT
2635 
2636 static int trace__set_ev_qualifier_filter(struct trace *trace)
2637 {
2638 	if (trace->syscalls.map)
2639 		return trace__set_ev_qualifier_bpf_filter(trace);
2640 	return trace__set_ev_qualifier_tp_filter(trace);
2641 }
2642 
2643 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2644 				    size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2645 {
2646 	int err = 0;
2647 #ifdef HAVE_LIBBPF_SUPPORT
2648 	bool value = true;
2649 	int map_fd = bpf_map__fd(map);
2650 	size_t i;
2651 
2652 	for (i = 0; i < npids; ++i) {
2653 		err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2654 		if (err)
2655 			break;
2656 	}
2657 #endif
2658 	return err;
2659 }
2660 
2661 static int trace__set_filter_loop_pids(struct trace *trace)
2662 {
2663 	unsigned int nr = 1, err;
2664 	pid_t pids[32] = {
2665 		getpid(),
2666 	};
2667 	struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2668 
2669 	while (thread && nr < ARRAY_SIZE(pids)) {
2670 		struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2671 
2672 		if (parent == NULL)
2673 			break;
2674 
2675 		if (!strcmp(thread__comm_str(parent), "sshd")) {
2676 			pids[nr++] = parent->tid;
2677 			break;
2678 		}
2679 		thread = parent;
2680 	}
2681 
2682 	err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2683 	if (!err && trace->filter_pids.map)
2684 		err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2685 
2686 	return err;
2687 }
2688 
2689 static int trace__set_filter_pids(struct trace *trace)
2690 {
2691 	int err = 0;
2692 	/*
2693 	 * Better not use !target__has_task() here because we need to cover the
2694 	 * case where no threads were specified in the command line, but a
2695 	 * workload was, and in that case we will fill in the thread_map when
2696 	 * we fork the workload in perf_evlist__prepare_workload.
2697 	 */
2698 	if (trace->filter_pids.nr > 0) {
2699 		err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2700 						      trace->filter_pids.entries);
2701 		if (!err && trace->filter_pids.map) {
2702 			err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2703 						       trace->filter_pids.entries);
2704 		}
2705 	} else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2706 		err = trace__set_filter_loop_pids(trace);
2707 	}
2708 
2709 	return err;
2710 }
2711 
2712 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2713 {
2714 	struct perf_evlist *evlist = trace->evlist;
2715 	struct perf_sample sample;
2716 	int err;
2717 
2718 	err = perf_evlist__parse_sample(evlist, event, &sample);
2719 	if (err)
2720 		fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2721 	else
2722 		trace__handle_event(trace, event, &sample);
2723 
2724 	return 0;
2725 }
2726 
2727 static int __trace__flush_events(struct trace *trace)
2728 {
2729 	u64 first = ordered_events__first_time(&trace->oe.data);
2730 	u64 flush = trace->oe.last - NSEC_PER_SEC;
2731 
2732 	/* Is there some thing to flush.. */
2733 	if (first && first < flush)
2734 		return ordered_events__flush_time(&trace->oe.data, flush);
2735 
2736 	return 0;
2737 }
2738 
2739 static int trace__flush_events(struct trace *trace)
2740 {
2741 	return !trace->sort_events ? 0 : __trace__flush_events(trace);
2742 }
2743 
2744 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2745 {
2746 	int err;
2747 
2748 	if (!trace->sort_events)
2749 		return __trace__deliver_event(trace, event);
2750 
2751 	err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2752 	if (err && err != -1)
2753 		return err;
2754 
2755 	err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2756 	if (err)
2757 		return err;
2758 
2759 	return trace__flush_events(trace);
2760 }
2761 
2762 static int ordered_events__deliver_event(struct ordered_events *oe,
2763 					 struct ordered_event *event)
2764 {
2765 	struct trace *trace = container_of(oe, struct trace, oe.data);
2766 
2767 	return __trace__deliver_event(trace, event->event);
2768 }
2769 
2770 static int trace__run(struct trace *trace, int argc, const char **argv)
2771 {
2772 	struct perf_evlist *evlist = trace->evlist;
2773 	struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2774 	int err = -1, i;
2775 	unsigned long before;
2776 	const bool forks = argc > 0;
2777 	bool draining = false;
2778 
2779 	trace->live = true;
2780 
2781 	if (!trace->raw_augmented_syscalls) {
2782 		if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2783 			goto out_error_raw_syscalls;
2784 
2785 		if (trace->trace_syscalls)
2786 			trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2787 	}
2788 
2789 	if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2790 		pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2791 		if (pgfault_maj == NULL)
2792 			goto out_error_mem;
2793 		perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2794 		perf_evlist__add(evlist, pgfault_maj);
2795 	}
2796 
2797 	if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2798 		pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2799 		if (pgfault_min == NULL)
2800 			goto out_error_mem;
2801 		perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2802 		perf_evlist__add(evlist, pgfault_min);
2803 	}
2804 
2805 	if (trace->sched &&
2806 	    perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2807 				   trace__sched_stat_runtime))
2808 		goto out_error_sched_stat_runtime;
2809 
2810 	/*
2811 	 * If a global cgroup was set, apply it to all the events without an
2812 	 * explicit cgroup. I.e.:
2813 	 *
2814 	 * 	trace -G A -e sched:*switch
2815 	 *
2816 	 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2817 	 * _and_ sched:sched_switch to the 'A' cgroup, while:
2818 	 *
2819 	 * trace -e sched:*switch -G A
2820 	 *
2821 	 * will only set the sched:sched_switch event to the 'A' cgroup, all the
2822 	 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2823 	 * a cgroup (on the root cgroup, sys wide, etc).
2824 	 *
2825 	 * Multiple cgroups:
2826 	 *
2827 	 * trace -G A -e sched:*switch -G B
2828 	 *
2829 	 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2830 	 * to the 'B' cgroup.
2831 	 *
2832 	 * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2833 	 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2834 	 */
2835 	if (trace->cgroup)
2836 		evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2837 
2838 	err = perf_evlist__create_maps(evlist, &trace->opts.target);
2839 	if (err < 0) {
2840 		fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2841 		goto out_delete_evlist;
2842 	}
2843 
2844 	err = trace__symbols_init(trace, evlist);
2845 	if (err < 0) {
2846 		fprintf(trace->output, "Problems initializing symbol libraries!\n");
2847 		goto out_delete_evlist;
2848 	}
2849 
2850 	perf_evlist__config(evlist, &trace->opts, &callchain_param);
2851 
2852 	signal(SIGCHLD, sig_handler);
2853 	signal(SIGINT, sig_handler);
2854 
2855 	if (forks) {
2856 		err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2857 						    argv, false, NULL);
2858 		if (err < 0) {
2859 			fprintf(trace->output, "Couldn't run the workload!\n");
2860 			goto out_delete_evlist;
2861 		}
2862 	}
2863 
2864 	err = perf_evlist__open(evlist);
2865 	if (err < 0)
2866 		goto out_error_open;
2867 
2868 	err = bpf__apply_obj_config();
2869 	if (err) {
2870 		char errbuf[BUFSIZ];
2871 
2872 		bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2873 		pr_err("ERROR: Apply config to BPF failed: %s\n",
2874 			 errbuf);
2875 		goto out_error_open;
2876 	}
2877 
2878 	err = trace__set_filter_pids(trace);
2879 	if (err < 0)
2880 		goto out_error_mem;
2881 
2882 	if (trace->syscalls.map)
2883 		trace__init_syscalls_bpf_map(trace);
2884 
2885 	if (trace->ev_qualifier_ids.nr > 0) {
2886 		err = trace__set_ev_qualifier_filter(trace);
2887 		if (err < 0)
2888 			goto out_errno;
2889 
2890 		if (trace->syscalls.events.sys_exit) {
2891 			pr_debug("event qualifier tracepoint filter: %s\n",
2892 				 trace->syscalls.events.sys_exit->filter);
2893 		}
2894 	}
2895 
2896 	err = perf_evlist__apply_filters(evlist, &evsel);
2897 	if (err < 0)
2898 		goto out_error_apply_filters;
2899 
2900 	err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
2901 	if (err < 0)
2902 		goto out_error_mmap;
2903 
2904 	if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2905 		perf_evlist__enable(evlist);
2906 
2907 	if (forks)
2908 		perf_evlist__start_workload(evlist);
2909 
2910 	if (trace->opts.initial_delay) {
2911 		usleep(trace->opts.initial_delay * 1000);
2912 		perf_evlist__enable(evlist);
2913 	}
2914 
2915 	trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2916 				  evlist->threads->nr > 1 ||
2917 				  perf_evlist__first(evlist)->attr.inherit;
2918 
2919 	/*
2920 	 * Now that we already used evsel->attr to ask the kernel to setup the
2921 	 * events, lets reuse evsel->attr.sample_max_stack as the limit in
2922 	 * trace__resolve_callchain(), allowing per-event max-stack settings
2923 	 * to override an explicitly set --max-stack global setting.
2924 	 */
2925 	evlist__for_each_entry(evlist, evsel) {
2926 		if (evsel__has_callchain(evsel) &&
2927 		    evsel->attr.sample_max_stack == 0)
2928 			evsel->attr.sample_max_stack = trace->max_stack;
2929 	}
2930 again:
2931 	before = trace->nr_events;
2932 
2933 	for (i = 0; i < evlist->nr_mmaps; i++) {
2934 		union perf_event *event;
2935 		struct perf_mmap *md;
2936 
2937 		md = &evlist->mmap[i];
2938 		if (perf_mmap__read_init(md) < 0)
2939 			continue;
2940 
2941 		while ((event = perf_mmap__read_event(md)) != NULL) {
2942 			++trace->nr_events;
2943 
2944 			err = trace__deliver_event(trace, event);
2945 			if (err)
2946 				goto out_disable;
2947 
2948 			perf_mmap__consume(md);
2949 
2950 			if (interrupted)
2951 				goto out_disable;
2952 
2953 			if (done && !draining) {
2954 				perf_evlist__disable(evlist);
2955 				draining = true;
2956 			}
2957 		}
2958 		perf_mmap__read_done(md);
2959 	}
2960 
2961 	if (trace->nr_events == before) {
2962 		int timeout = done ? 100 : -1;
2963 
2964 		if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
2965 			if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
2966 				draining = true;
2967 
2968 			goto again;
2969 		} else {
2970 			if (trace__flush_events(trace))
2971 				goto out_disable;
2972 		}
2973 	} else {
2974 		goto again;
2975 	}
2976 
2977 out_disable:
2978 	thread__zput(trace->current);
2979 
2980 	perf_evlist__disable(evlist);
2981 
2982 	if (trace->sort_events)
2983 		ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
2984 
2985 	if (!err) {
2986 		if (trace->summary)
2987 			trace__fprintf_thread_summary(trace, trace->output);
2988 
2989 		if (trace->show_tool_stats) {
2990 			fprintf(trace->output, "Stats:\n "
2991 					       " vfs_getname : %" PRIu64 "\n"
2992 					       " proc_getname: %" PRIu64 "\n",
2993 				trace->stats.vfs_getname,
2994 				trace->stats.proc_getname);
2995 		}
2996 	}
2997 
2998 out_delete_evlist:
2999 	trace__symbols__exit(trace);
3000 
3001 	perf_evlist__delete(evlist);
3002 	cgroup__put(trace->cgroup);
3003 	trace->evlist = NULL;
3004 	trace->live = false;
3005 	return err;
3006 {
3007 	char errbuf[BUFSIZ];
3008 
3009 out_error_sched_stat_runtime:
3010 	tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3011 	goto out_error;
3012 
3013 out_error_raw_syscalls:
3014 	tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3015 	goto out_error;
3016 
3017 out_error_mmap:
3018 	perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3019 	goto out_error;
3020 
3021 out_error_open:
3022 	perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3023 
3024 out_error:
3025 	fprintf(trace->output, "%s\n", errbuf);
3026 	goto out_delete_evlist;
3027 
3028 out_error_apply_filters:
3029 	fprintf(trace->output,
3030 		"Failed to set filter \"%s\" on event %s with %d (%s)\n",
3031 		evsel->filter, perf_evsel__name(evsel), errno,
3032 		str_error_r(errno, errbuf, sizeof(errbuf)));
3033 	goto out_delete_evlist;
3034 }
3035 out_error_mem:
3036 	fprintf(trace->output, "Not enough memory to run!\n");
3037 	goto out_delete_evlist;
3038 
3039 out_errno:
3040 	fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3041 	goto out_delete_evlist;
3042 }
3043 
3044 static int trace__replay(struct trace *trace)
3045 {
3046 	const struct perf_evsel_str_handler handlers[] = {
3047 		{ "probe:vfs_getname",	     trace__vfs_getname, },
3048 	};
3049 	struct perf_data data = {
3050 		.file      = {
3051 			.path = input_name,
3052 		},
3053 		.mode      = PERF_DATA_MODE_READ,
3054 		.force     = trace->force,
3055 	};
3056 	struct perf_session *session;
3057 	struct perf_evsel *evsel;
3058 	int err = -1;
3059 
3060 	trace->tool.sample	  = trace__process_sample;
3061 	trace->tool.mmap	  = perf_event__process_mmap;
3062 	trace->tool.mmap2	  = perf_event__process_mmap2;
3063 	trace->tool.comm	  = perf_event__process_comm;
3064 	trace->tool.exit	  = perf_event__process_exit;
3065 	trace->tool.fork	  = perf_event__process_fork;
3066 	trace->tool.attr	  = perf_event__process_attr;
3067 	trace->tool.tracing_data  = perf_event__process_tracing_data;
3068 	trace->tool.build_id	  = perf_event__process_build_id;
3069 	trace->tool.namespaces	  = perf_event__process_namespaces;
3070 
3071 	trace->tool.ordered_events = true;
3072 	trace->tool.ordering_requires_timestamps = true;
3073 
3074 	/* add tid to output */
3075 	trace->multiple_threads = true;
3076 
3077 	session = perf_session__new(&data, false, &trace->tool);
3078 	if (session == NULL)
3079 		return -1;
3080 
3081 	if (trace->opts.target.pid)
3082 		symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3083 
3084 	if (trace->opts.target.tid)
3085 		symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3086 
3087 	if (symbol__init(&session->header.env) < 0)
3088 		goto out;
3089 
3090 	trace->host = &session->machines.host;
3091 
3092 	err = perf_session__set_tracepoints_handlers(session, handlers);
3093 	if (err)
3094 		goto out;
3095 
3096 	evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3097 						     "raw_syscalls:sys_enter");
3098 	/* older kernels have syscalls tp versus raw_syscalls */
3099 	if (evsel == NULL)
3100 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3101 							     "syscalls:sys_enter");
3102 
3103 	if (evsel &&
3104 	    (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3105 	    perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3106 		pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3107 		goto out;
3108 	}
3109 
3110 	evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3111 						     "raw_syscalls:sys_exit");
3112 	if (evsel == NULL)
3113 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3114 							     "syscalls:sys_exit");
3115 	if (evsel &&
3116 	    (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3117 	    perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3118 		pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3119 		goto out;
3120 	}
3121 
3122 	evlist__for_each_entry(session->evlist, evsel) {
3123 		if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3124 		    (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3125 		     evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3126 		     evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3127 			evsel->handler = trace__pgfault;
3128 	}
3129 
3130 	setup_pager();
3131 
3132 	err = perf_session__process_events(session);
3133 	if (err)
3134 		pr_err("Failed to process events, error %d", err);
3135 
3136 	else if (trace->summary)
3137 		trace__fprintf_thread_summary(trace, trace->output);
3138 
3139 out:
3140 	perf_session__delete(session);
3141 
3142 	return err;
3143 }
3144 
3145 static size_t trace__fprintf_threads_header(FILE *fp)
3146 {
3147 	size_t printed;
3148 
3149 	printed  = fprintf(fp, "\n Summary of events:\n\n");
3150 
3151 	return printed;
3152 }
3153 
3154 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3155 	struct stats 	*stats;
3156 	double		msecs;
3157 	int		syscall;
3158 )
3159 {
3160 	struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3161 	struct stats *stats = source->priv;
3162 
3163 	entry->syscall = source->i;
3164 	entry->stats   = stats;
3165 	entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3166 }
3167 
3168 static size_t thread__dump_stats(struct thread_trace *ttrace,
3169 				 struct trace *trace, FILE *fp)
3170 {
3171 	size_t printed = 0;
3172 	struct syscall *sc;
3173 	struct rb_node *nd;
3174 	DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3175 
3176 	if (syscall_stats == NULL)
3177 		return 0;
3178 
3179 	printed += fprintf(fp, "\n");
3180 
3181 	printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3182 	printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3183 	printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3184 
3185 	resort_rb__for_each_entry(nd, syscall_stats) {
3186 		struct stats *stats = syscall_stats_entry->stats;
3187 		if (stats) {
3188 			double min = (double)(stats->min) / NSEC_PER_MSEC;
3189 			double max = (double)(stats->max) / NSEC_PER_MSEC;
3190 			double avg = avg_stats(stats);
3191 			double pct;
3192 			u64 n = (u64) stats->n;
3193 
3194 			pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3195 			avg /= NSEC_PER_MSEC;
3196 
3197 			sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3198 			printed += fprintf(fp, "   %-15s", sc->name);
3199 			printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3200 					   n, syscall_stats_entry->msecs, min, avg);
3201 			printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3202 		}
3203 	}
3204 
3205 	resort_rb__delete(syscall_stats);
3206 	printed += fprintf(fp, "\n\n");
3207 
3208 	return printed;
3209 }
3210 
3211 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3212 {
3213 	size_t printed = 0;
3214 	struct thread_trace *ttrace = thread__priv(thread);
3215 	double ratio;
3216 
3217 	if (ttrace == NULL)
3218 		return 0;
3219 
3220 	ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3221 
3222 	printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3223 	printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3224 	printed += fprintf(fp, "%.1f%%", ratio);
3225 	if (ttrace->pfmaj)
3226 		printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3227 	if (ttrace->pfmin)
3228 		printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3229 	if (trace->sched)
3230 		printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3231 	else if (fputc('\n', fp) != EOF)
3232 		++printed;
3233 
3234 	printed += thread__dump_stats(ttrace, trace, fp);
3235 
3236 	return printed;
3237 }
3238 
3239 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3240 {
3241 	return ttrace ? ttrace->nr_events : 0;
3242 }
3243 
3244 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3245 	struct thread *thread;
3246 )
3247 {
3248 	entry->thread = rb_entry(nd, struct thread, rb_node);
3249 }
3250 
3251 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3252 {
3253 	size_t printed = trace__fprintf_threads_header(fp);
3254 	struct rb_node *nd;
3255 	int i;
3256 
3257 	for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3258 		DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3259 
3260 		if (threads == NULL) {
3261 			fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3262 			return 0;
3263 		}
3264 
3265 		resort_rb__for_each_entry(nd, threads)
3266 			printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3267 
3268 		resort_rb__delete(threads);
3269 	}
3270 	return printed;
3271 }
3272 
3273 static int trace__set_duration(const struct option *opt, const char *str,
3274 			       int unset __maybe_unused)
3275 {
3276 	struct trace *trace = opt->value;
3277 
3278 	trace->duration_filter = atof(str);
3279 	return 0;
3280 }
3281 
3282 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3283 					      int unset __maybe_unused)
3284 {
3285 	int ret = -1;
3286 	size_t i;
3287 	struct trace *trace = opt->value;
3288 	/*
3289 	 * FIXME: introduce a intarray class, plain parse csv and create a
3290 	 * { int nr, int entries[] } struct...
3291 	 */
3292 	struct intlist *list = intlist__new(str);
3293 
3294 	if (list == NULL)
3295 		return -1;
3296 
3297 	i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3298 	trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3299 
3300 	if (trace->filter_pids.entries == NULL)
3301 		goto out;
3302 
3303 	trace->filter_pids.entries[0] = getpid();
3304 
3305 	for (i = 1; i < trace->filter_pids.nr; ++i)
3306 		trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3307 
3308 	intlist__delete(list);
3309 	ret = 0;
3310 out:
3311 	return ret;
3312 }
3313 
3314 static int trace__open_output(struct trace *trace, const char *filename)
3315 {
3316 	struct stat st;
3317 
3318 	if (!stat(filename, &st) && st.st_size) {
3319 		char oldname[PATH_MAX];
3320 
3321 		scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3322 		unlink(oldname);
3323 		rename(filename, oldname);
3324 	}
3325 
3326 	trace->output = fopen(filename, "w");
3327 
3328 	return trace->output == NULL ? -errno : 0;
3329 }
3330 
3331 static int parse_pagefaults(const struct option *opt, const char *str,
3332 			    int unset __maybe_unused)
3333 {
3334 	int *trace_pgfaults = opt->value;
3335 
3336 	if (strcmp(str, "all") == 0)
3337 		*trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3338 	else if (strcmp(str, "maj") == 0)
3339 		*trace_pgfaults |= TRACE_PFMAJ;
3340 	else if (strcmp(str, "min") == 0)
3341 		*trace_pgfaults |= TRACE_PFMIN;
3342 	else
3343 		return -1;
3344 
3345 	return 0;
3346 }
3347 
3348 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3349 {
3350 	struct perf_evsel *evsel;
3351 
3352 	evlist__for_each_entry(evlist, evsel)
3353 		evsel->handler = handler;
3354 }
3355 
3356 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3357 {
3358 	struct perf_evsel *evsel;
3359 
3360 	evlist__for_each_entry(evlist, evsel) {
3361 		if (evsel->priv || !evsel->tp_format)
3362 			continue;
3363 
3364 		if (strcmp(evsel->tp_format->system, "syscalls"))
3365 			continue;
3366 
3367 		if (perf_evsel__init_syscall_tp(evsel))
3368 			return -1;
3369 
3370 		if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3371 			struct syscall_tp *sc = evsel->priv;
3372 
3373 			if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3374 				return -1;
3375 		} else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3376 			struct syscall_tp *sc = evsel->priv;
3377 
3378 			if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3379 				return -1;
3380 		}
3381 	}
3382 
3383 	return 0;
3384 }
3385 
3386 /*
3387  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3388  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3389  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3390  *
3391  * It'd be better to introduce a parse_options() variant that would return a
3392  * list with the terms it didn't match to an event...
3393  */
3394 static int trace__parse_events_option(const struct option *opt, const char *str,
3395 				      int unset __maybe_unused)
3396 {
3397 	struct trace *trace = (struct trace *)opt->value;
3398 	const char *s = str;
3399 	char *sep = NULL, *lists[2] = { NULL, NULL, };
3400 	int len = strlen(str) + 1, err = -1, list, idx;
3401 	char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3402 	char group_name[PATH_MAX];
3403 	struct syscall_fmt *fmt;
3404 
3405 	if (strace_groups_dir == NULL)
3406 		return -1;
3407 
3408 	if (*s == '!') {
3409 		++s;
3410 		trace->not_ev_qualifier = true;
3411 	}
3412 
3413 	while (1) {
3414 		if ((sep = strchr(s, ',')) != NULL)
3415 			*sep = '\0';
3416 
3417 		list = 0;
3418 		if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3419 		    syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3420 			list = 1;
3421 			goto do_concat;
3422 		}
3423 
3424 		fmt = syscall_fmt__find_by_alias(s);
3425 		if (fmt != NULL) {
3426 			list = 1;
3427 			s = fmt->name;
3428 		} else {
3429 			path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3430 			if (access(group_name, R_OK) == 0)
3431 				list = 1;
3432 		}
3433 do_concat:
3434 		if (lists[list]) {
3435 			sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3436 		} else {
3437 			lists[list] = malloc(len);
3438 			if (lists[list] == NULL)
3439 				goto out;
3440 			strcpy(lists[list], s);
3441 		}
3442 
3443 		if (!sep)
3444 			break;
3445 
3446 		*sep = ',';
3447 		s = sep + 1;
3448 	}
3449 
3450 	if (lists[1] != NULL) {
3451 		struct strlist_config slist_config = {
3452 			.dirname = strace_groups_dir,
3453 		};
3454 
3455 		trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3456 		if (trace->ev_qualifier == NULL) {
3457 			fputs("Not enough memory to parse event qualifier", trace->output);
3458 			goto out;
3459 		}
3460 
3461 		if (trace__validate_ev_qualifier(trace))
3462 			goto out;
3463 		trace->trace_syscalls = true;
3464 	}
3465 
3466 	err = 0;
3467 
3468 	if (lists[0]) {
3469 		struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3470 					       "event selector. use 'perf list' to list available events",
3471 					       parse_events_option);
3472 		err = parse_events_option(&o, lists[0], 0);
3473 	}
3474 out:
3475 	if (sep)
3476 		*sep = ',';
3477 
3478 	return err;
3479 }
3480 
3481 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3482 {
3483 	struct trace *trace = opt->value;
3484 
3485 	if (!list_empty(&trace->evlist->entries))
3486 		return parse_cgroups(opt, str, unset);
3487 
3488 	trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3489 
3490 	return 0;
3491 }
3492 
3493 static struct bpf_map *bpf__find_map_by_name(const char *name)
3494 {
3495 	struct bpf_object *obj, *tmp;
3496 
3497 	bpf_object__for_each_safe(obj, tmp) {
3498 		struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3499 		if (map)
3500 			return map;
3501 
3502 	}
3503 
3504 	return NULL;
3505 }
3506 
3507 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3508 {
3509 	trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3510 }
3511 
3512 static void trace__set_bpf_map_syscalls(struct trace *trace)
3513 {
3514 	trace->syscalls.map = bpf__find_map_by_name("syscalls");
3515 }
3516 
3517 int cmd_trace(int argc, const char **argv)
3518 {
3519 	const char *trace_usage[] = {
3520 		"perf trace [<options>] [<command>]",
3521 		"perf trace [<options>] -- <command> [<options>]",
3522 		"perf trace record [<options>] [<command>]",
3523 		"perf trace record [<options>] -- <command> [<options>]",
3524 		NULL
3525 	};
3526 	struct trace trace = {
3527 		.syscalls = {
3528 			. max = -1,
3529 		},
3530 		.opts = {
3531 			.target = {
3532 				.uid	   = UINT_MAX,
3533 				.uses_mmap = true,
3534 			},
3535 			.user_freq     = UINT_MAX,
3536 			.user_interval = ULLONG_MAX,
3537 			.no_buffering  = true,
3538 			.mmap_pages    = UINT_MAX,
3539 		},
3540 		.output = stderr,
3541 		.show_comm = true,
3542 		.trace_syscalls = false,
3543 		.kernel_syscallchains = false,
3544 		.max_stack = UINT_MAX,
3545 		.max_events = ULONG_MAX,
3546 	};
3547 	const char *output_name = NULL;
3548 	const struct option trace_options[] = {
3549 	OPT_CALLBACK('e', "event", &trace, "event",
3550 		     "event/syscall selector. use 'perf list' to list available events",
3551 		     trace__parse_events_option),
3552 	OPT_BOOLEAN(0, "comm", &trace.show_comm,
3553 		    "show the thread COMM next to its id"),
3554 	OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3555 	OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3556 		     trace__parse_events_option),
3557 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
3558 	OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3559 	OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3560 		    "trace events on existing process id"),
3561 	OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3562 		    "trace events on existing thread id"),
3563 	OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3564 		     "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3565 	OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3566 		    "system-wide collection from all CPUs"),
3567 	OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3568 		    "list of cpus to monitor"),
3569 	OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3570 		    "child tasks do not inherit counters"),
3571 	OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3572 		     "number of mmap data pages",
3573 		     perf_evlist__parse_mmap_pages),
3574 	OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3575 		   "user to profile"),
3576 	OPT_CALLBACK(0, "duration", &trace, "float",
3577 		     "show only events with duration > N.M ms",
3578 		     trace__set_duration),
3579 	OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3580 	OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3581 	OPT_BOOLEAN('T', "time", &trace.full_time,
3582 		    "Show full timestamp, not time relative to first start"),
3583 	OPT_BOOLEAN(0, "failure", &trace.failure_only,
3584 		    "Show only syscalls that failed"),
3585 	OPT_BOOLEAN('s', "summary", &trace.summary_only,
3586 		    "Show only syscall summary with statistics"),
3587 	OPT_BOOLEAN('S', "with-summary", &trace.summary,
3588 		    "Show all syscalls and summary with statistics"),
3589 	OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3590 		     "Trace pagefaults", parse_pagefaults, "maj"),
3591 	OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3592 	OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3593 	OPT_CALLBACK(0, "call-graph", &trace.opts,
3594 		     "record_mode[,record_size]", record_callchain_help,
3595 		     &record_parse_callchain_opt),
3596 	OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3597 		    "Show the kernel callchains on the syscall exit path"),
3598 	OPT_ULONG(0, "max-events", &trace.max_events,
3599 		"Set the maximum number of events to print, exit after that is reached. "),
3600 	OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3601 		     "Set the minimum stack depth when parsing the callchain, "
3602 		     "anything below the specified depth will be ignored."),
3603 	OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3604 		     "Set the maximum stack depth when parsing the callchain, "
3605 		     "anything beyond the specified depth will be ignored. "
3606 		     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3607 	OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3608 			"Sort batch of events before processing, use if getting out of order events"),
3609 	OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3610 			"print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3611 	OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3612 			"per thread proc mmap processing timeout in ms"),
3613 	OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3614 		     trace__parse_cgroups),
3615 	OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3616 		     "ms to wait before starting measurement after program "
3617 		     "start"),
3618 	OPT_END()
3619 	};
3620 	bool __maybe_unused max_stack_user_set = true;
3621 	bool mmap_pages_user_set = true;
3622 	struct perf_evsel *evsel;
3623 	const char * const trace_subcommands[] = { "record", NULL };
3624 	int err = -1;
3625 	char bf[BUFSIZ];
3626 
3627 	signal(SIGSEGV, sighandler_dump_stack);
3628 	signal(SIGFPE, sighandler_dump_stack);
3629 
3630 	trace.evlist = perf_evlist__new();
3631 	trace.sctbl = syscalltbl__new();
3632 
3633 	if (trace.evlist == NULL || trace.sctbl == NULL) {
3634 		pr_err("Not enough memory to run!\n");
3635 		err = -ENOMEM;
3636 		goto out;
3637 	}
3638 
3639 	argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3640 				 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3641 
3642 	if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3643 		usage_with_options_msg(trace_usage, trace_options,
3644 				       "cgroup monitoring only available in system-wide mode");
3645 	}
3646 
3647 	evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3648 	if (IS_ERR(evsel)) {
3649 		bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3650 		pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3651 		goto out;
3652 	}
3653 
3654 	if (evsel) {
3655 		trace.syscalls.events.augmented = evsel;
3656 		trace__set_bpf_map_filtered_pids(&trace);
3657 		trace__set_bpf_map_syscalls(&trace);
3658 	}
3659 
3660 	err = bpf__setup_stdout(trace.evlist);
3661 	if (err) {
3662 		bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3663 		pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3664 		goto out;
3665 	}
3666 
3667 	err = -1;
3668 
3669 	if (trace.trace_pgfaults) {
3670 		trace.opts.sample_address = true;
3671 		trace.opts.sample_time = true;
3672 	}
3673 
3674 	if (trace.opts.mmap_pages == UINT_MAX)
3675 		mmap_pages_user_set = false;
3676 
3677 	if (trace.max_stack == UINT_MAX) {
3678 		trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3679 		max_stack_user_set = false;
3680 	}
3681 
3682 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3683 	if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3684 		record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3685 	}
3686 #endif
3687 
3688 	if (callchain_param.enabled) {
3689 		if (!mmap_pages_user_set && geteuid() == 0)
3690 			trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3691 
3692 		symbol_conf.use_callchain = true;
3693 	}
3694 
3695 	if (trace.evlist->nr_entries > 0) {
3696 		evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3697 		if (evlist__set_syscall_tp_fields(trace.evlist)) {
3698 			perror("failed to set syscalls:* tracepoint fields");
3699 			goto out;
3700 		}
3701 	}
3702 
3703 	if (trace.sort_events) {
3704 		ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3705 		ordered_events__set_copy_on_queue(&trace.oe.data, true);
3706 	}
3707 
3708 	/*
3709 	 * If we are augmenting syscalls, then combine what we put in the
3710 	 * __augmented_syscalls__ BPF map with what is in the
3711 	 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3712 	 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3713 	 *
3714 	 * We'll switch to look at two BPF maps, one for sys_enter and the
3715 	 * other for sys_exit when we start augmenting the sys_exit paths with
3716 	 * buffers that are being copied from kernel to userspace, think 'read'
3717 	 * syscall.
3718 	 */
3719 	if (trace.syscalls.events.augmented) {
3720 		evsel = trace.syscalls.events.augmented;
3721 
3722 		if (perf_evsel__init_augmented_syscall_tp(evsel) ||
3723 		    perf_evsel__init_augmented_syscall_tp_args(evsel))
3724 			goto out;
3725 		evsel->handler = trace__sys_enter;
3726 
3727 		evlist__for_each_entry(trace.evlist, evsel) {
3728 			bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3729 
3730 			if (raw_syscalls_sys_exit) {
3731 				trace.raw_augmented_syscalls = true;
3732 				goto init_augmented_syscall_tp;
3733 			}
3734 
3735 			if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3736 init_augmented_syscall_tp:
3737 				perf_evsel__init_augmented_syscall_tp(evsel);
3738 				perf_evsel__init_augmented_syscall_tp_ret(evsel);
3739 				evsel->handler = trace__sys_exit;
3740 			}
3741 		}
3742 	}
3743 
3744 	if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3745 		return trace__record(&trace, argc-1, &argv[1]);
3746 
3747 	/* summary_only implies summary option, but don't overwrite summary if set */
3748 	if (trace.summary_only)
3749 		trace.summary = trace.summary_only;
3750 
3751 	if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3752 	    trace.evlist->nr_entries == 0 /* Was --events used? */) {
3753 		trace.trace_syscalls = true;
3754 	}
3755 
3756 	if (output_name != NULL) {
3757 		err = trace__open_output(&trace, output_name);
3758 		if (err < 0) {
3759 			perror("failed to create output file");
3760 			goto out;
3761 		}
3762 	}
3763 
3764 	err = target__validate(&trace.opts.target);
3765 	if (err) {
3766 		target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3767 		fprintf(trace.output, "%s", bf);
3768 		goto out_close;
3769 	}
3770 
3771 	err = target__parse_uid(&trace.opts.target);
3772 	if (err) {
3773 		target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3774 		fprintf(trace.output, "%s", bf);
3775 		goto out_close;
3776 	}
3777 
3778 	if (!argc && target__none(&trace.opts.target))
3779 		trace.opts.target.system_wide = true;
3780 
3781 	if (input_name)
3782 		err = trace__replay(&trace);
3783 	else
3784 		err = trace__run(&trace, argc, argv);
3785 
3786 out_close:
3787 	if (output_name != NULL)
3788 		fclose(trace.output);
3789 out:
3790 	return err;
3791 }
3792