4 * Builtin 'trace' command:
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.
10 * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
12 * Initially based on the 'trace' prototype by Thomas Gleixner:
14 * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
16 * Released under the GPL v2. (and only v2, not any later version)
19 #include <traceevent/event-parse.h>
20 #include <api/fs/tracing_path.h>
22 #include "util/color.h"
23 #include "util/debug.h"
24 #include "util/evlist.h"
25 #include <subcmd/exec-cmd.h>
26 #include "util/machine.h"
27 #include "util/session.h"
28 #include "util/thread.h"
29 #include <subcmd/parse-options.h>
30 #include "util/strlist.h"
31 #include "util/intlist.h"
32 #include "util/thread_map.h"
33 #include "util/stat.h"
34 #include "trace-event.h"
35 #include "util/parse-events.h"
36 #include "util/bpf-loader.h"
37 #include "callchain.h"
38 #include "syscalltbl.h"
39 #include "rb_resort.h"
41 #include <libaudit.h> /* FIXME: Still needed for audit_errno_to_name */
43 #include <linux/err.h>
44 #include <linux/filter.h>
45 #include <linux/audit.h>
46 #include <sys/ptrace.h>
47 #include <linux/random.h>
48 #include <linux/stringify.h>
51 # define O_CLOEXEC 02000000
55 struct perf_tool tool;
56 struct syscalltbl *sctbl;
59 struct syscall *table;
61 struct perf_evsel *sys_enter,
65 struct record_opts opts;
66 struct perf_evlist *evlist;
68 struct thread *current;
71 unsigned long nr_events;
72 struct strlist *ev_qualifier;
77 struct intlist *tid_list;
78 struct intlist *pid_list;
83 double duration_filter;
89 unsigned int max_stack;
90 unsigned int min_stack;
91 bool not_ev_qualifier;
95 bool multiple_threads;
101 bool kernel_syscallchains;
111 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
112 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
116 #define TP_UINT_FIELD(bits) \
117 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
120 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
129 #define TP_UINT_FIELD__SWAPPED(bits) \
130 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
133 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
134 return bswap_##bits(value);\
137 TP_UINT_FIELD__SWAPPED(16);
138 TP_UINT_FIELD__SWAPPED(32);
139 TP_UINT_FIELD__SWAPPED(64);
141 static int tp_field__init_uint(struct tp_field *field,
142 struct format_field *format_field,
145 field->offset = format_field->offset;
147 switch (format_field->size) {
149 field->integer = tp_field__u8;
152 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
155 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
158 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
167 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
169 return sample->raw_data + field->offset;
172 static int tp_field__init_ptr(struct tp_field *field, struct format_field *format_field)
174 field->offset = format_field->offset;
175 field->pointer = tp_field__ptr;
182 struct tp_field args, ret;
186 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
187 struct tp_field *field,
190 struct format_field *format_field = perf_evsel__field(evsel, name);
192 if (format_field == NULL)
195 return tp_field__init_uint(field, format_field, evsel->needs_swap);
198 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
199 ({ struct syscall_tp *sc = evsel->priv;\
200 perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
202 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
203 struct tp_field *field,
206 struct format_field *format_field = perf_evsel__field(evsel, name);
208 if (format_field == NULL)
211 return tp_field__init_ptr(field, format_field);
214 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
215 ({ struct syscall_tp *sc = evsel->priv;\
216 perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
218 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
221 perf_evsel__delete(evsel);
224 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel, void *handler)
226 evsel->priv = malloc(sizeof(struct syscall_tp));
227 if (evsel->priv != NULL) {
228 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
231 evsel->handler = handler;
242 static struct perf_evsel *perf_evsel__syscall_newtp(const char *direction, void *handler)
244 struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
246 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
248 evsel = perf_evsel__newtp("syscalls", direction);
253 if (perf_evsel__init_syscall_tp(evsel, handler))
259 perf_evsel__delete_priv(evsel);
263 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
264 ({ struct syscall_tp *fields = evsel->priv; \
265 fields->name.integer(&fields->name, sample); })
267 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
268 ({ struct syscall_tp *fields = evsel->priv; \
269 fields->name.pointer(&fields->name, sample); })
273 struct thread *thread;
283 const char **entries;
286 #define DEFINE_STRARRAY(array) struct strarray strarray__##array = { \
287 .nr_entries = ARRAY_SIZE(array), \
291 #define DEFINE_STRARRAY_OFFSET(array, off) struct strarray strarray__##array = { \
293 .nr_entries = ARRAY_SIZE(array), \
297 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
299 struct syscall_arg *arg)
301 struct strarray *sa = arg->parm;
302 int idx = arg->val - sa->offset;
304 if (idx < 0 || idx >= sa->nr_entries)
305 return scnprintf(bf, size, intfmt, arg->val);
307 return scnprintf(bf, size, "%s", sa->entries[idx]);
310 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
311 struct syscall_arg *arg)
313 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
316 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
318 #if defined(__i386__) || defined(__x86_64__)
320 * FIXME: Make this available to all arches as soon as the ioctl beautifier
321 * gets rewritten to support all arches.
323 static size_t syscall_arg__scnprintf_strhexarray(char *bf, size_t size,
324 struct syscall_arg *arg)
326 return __syscall_arg__scnprintf_strarray(bf, size, "%#x", arg);
329 #define SCA_STRHEXARRAY syscall_arg__scnprintf_strhexarray
330 #endif /* defined(__i386__) || defined(__x86_64__) */
332 static size_t syscall_arg__scnprintf_fd(char *bf, size_t size,
333 struct syscall_arg *arg);
335 #define SCA_FD syscall_arg__scnprintf_fd
337 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
338 struct syscall_arg *arg)
343 return scnprintf(bf, size, "CWD");
345 return syscall_arg__scnprintf_fd(bf, size, arg);
348 #define SCA_FDAT syscall_arg__scnprintf_fd_at
350 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
351 struct syscall_arg *arg);
353 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
355 static size_t syscall_arg__scnprintf_hex(char *bf, size_t size,
356 struct syscall_arg *arg)
358 return scnprintf(bf, size, "%#lx", arg->val);
361 #define SCA_HEX syscall_arg__scnprintf_hex
363 static size_t syscall_arg__scnprintf_int(char *bf, size_t size,
364 struct syscall_arg *arg)
366 return scnprintf(bf, size, "%d", arg->val);
369 #define SCA_INT syscall_arg__scnprintf_int
371 static const char *bpf_cmd[] = {
372 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
373 "MAP_GET_NEXT_KEY", "PROG_LOAD",
375 static DEFINE_STRARRAY(bpf_cmd);
377 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
378 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, 1);
380 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
381 static DEFINE_STRARRAY(itimers);
383 static const char *keyctl_options[] = {
384 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
385 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
386 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
387 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
388 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
390 static DEFINE_STRARRAY(keyctl_options);
392 static const char *whences[] = { "SET", "CUR", "END",
400 static DEFINE_STRARRAY(whences);
402 static const char *fcntl_cmds[] = {
403 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
404 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "F_GETLK64",
405 "F_SETLK64", "F_SETLKW64", "F_SETOWN_EX", "F_GETOWN_EX",
408 static DEFINE_STRARRAY(fcntl_cmds);
410 static const char *rlimit_resources[] = {
411 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
412 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
415 static DEFINE_STRARRAY(rlimit_resources);
417 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
418 static DEFINE_STRARRAY(sighow);
420 static const char *clockid[] = {
421 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
422 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
423 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
425 static DEFINE_STRARRAY(clockid);
427 static const char *socket_families[] = {
428 "UNSPEC", "LOCAL", "INET", "AX25", "IPX", "APPLETALK", "NETROM",
429 "BRIDGE", "ATMPVC", "X25", "INET6", "ROSE", "DECnet", "NETBEUI",
430 "SECURITY", "KEY", "NETLINK", "PACKET", "ASH", "ECONET", "ATMSVC",
431 "RDS", "SNA", "IRDA", "PPPOX", "WANPIPE", "LLC", "IB", "CAN", "TIPC",
432 "BLUETOOTH", "IUCV", "RXRPC", "ISDN", "PHONET", "IEEE802154", "CAIF",
433 "ALG", "NFC", "VSOCK",
435 static DEFINE_STRARRAY(socket_families);
437 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
438 struct syscall_arg *arg)
443 if (mode == F_OK) /* 0 */
444 return scnprintf(bf, size, "F");
446 if (mode & n##_OK) { \
447 printed += scnprintf(bf + printed, size - printed, "%s", #n); \
457 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
462 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
464 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
465 struct syscall_arg *arg);
467 #define SCA_FILENAME syscall_arg__scnprintf_filename
469 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
470 struct syscall_arg *arg)
472 int printed = 0, flags = arg->val;
475 if (flags & O_##n) { \
476 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
485 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
490 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
492 #if defined(__i386__) || defined(__x86_64__)
494 * FIXME: Make this available to all arches.
496 #define TCGETS 0x5401
498 static const char *tioctls[] = {
499 "TCGETS", "TCSETS", "TCSETSW", "TCSETSF", "TCGETA", "TCSETA", "TCSETAW",
500 "TCSETAF", "TCSBRK", "TCXONC", "TCFLSH", "TIOCEXCL", "TIOCNXCL",
501 "TIOCSCTTY", "TIOCGPGRP", "TIOCSPGRP", "TIOCOUTQ", "TIOCSTI",
502 "TIOCGWINSZ", "TIOCSWINSZ", "TIOCMGET", "TIOCMBIS", "TIOCMBIC",
503 "TIOCMSET", "TIOCGSOFTCAR", "TIOCSSOFTCAR", "FIONREAD", "TIOCLINUX",
504 "TIOCCONS", "TIOCGSERIAL", "TIOCSSERIAL", "TIOCPKT", "FIONBIO",
505 "TIOCNOTTY", "TIOCSETD", "TIOCGETD", "TCSBRKP", [0x27] = "TIOCSBRK",
506 "TIOCCBRK", "TIOCGSID", "TCGETS2", "TCSETS2", "TCSETSW2", "TCSETSF2",
507 "TIOCGRS485", "TIOCSRS485", "TIOCGPTN", "TIOCSPTLCK",
508 "TIOCGDEV||TCGETX", "TCSETX", "TCSETXF", "TCSETXW", "TIOCSIG",
509 "TIOCVHANGUP", "TIOCGPKT", "TIOCGPTLCK", "TIOCGEXCL",
510 [0x50] = "FIONCLEX", "FIOCLEX", "FIOASYNC", "TIOCSERCONFIG",
511 "TIOCSERGWILD", "TIOCSERSWILD", "TIOCGLCKTRMIOS", "TIOCSLCKTRMIOS",
512 "TIOCSERGSTRUCT", "TIOCSERGETLSR", "TIOCSERGETMULTI", "TIOCSERSETMULTI",
513 "TIOCMIWAIT", "TIOCGICOUNT", [0x60] = "FIOQSIZE",
516 static DEFINE_STRARRAY_OFFSET(tioctls, 0x5401);
517 #endif /* defined(__i386__) || defined(__x86_64__) */
519 #ifndef GRND_NONBLOCK
520 #define GRND_NONBLOCK 0x0001
523 #define GRND_RANDOM 0x0002
526 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
527 struct syscall_arg *arg)
529 int printed = 0, flags = arg->val;
532 if (flags & GRND_##n) { \
533 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
534 flags &= ~GRND_##n; \
542 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
547 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
549 #define STRARRAY(arg, name, array) \
550 .arg_scnprintf = { [arg] = SCA_STRARRAY, }, \
551 .arg_parm = { [arg] = &strarray__##array, }
553 #include "trace/beauty/eventfd.c"
554 #include "trace/beauty/flock.c"
555 #include "trace/beauty/futex_op.c"
556 #include "trace/beauty/mmap.c"
557 #include "trace/beauty/mode_t.c"
558 #include "trace/beauty/msg_flags.c"
559 #include "trace/beauty/open_flags.c"
560 #include "trace/beauty/perf_event_open.c"
561 #include "trace/beauty/pid.c"
562 #include "trace/beauty/sched_policy.c"
563 #include "trace/beauty/seccomp.c"
564 #include "trace/beauty/signum.c"
565 #include "trace/beauty/socket_type.c"
566 #include "trace/beauty/waitid_options.c"
568 static struct syscall_fmt {
571 size_t (*arg_scnprintf[6])(char *bf, size_t size, struct syscall_arg *arg);
578 { .name = "access", .errmsg = true,
579 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */
580 [1] = SCA_ACCMODE, /* mode */ }, },
581 { .name = "arch_prctl", .errmsg = true, .alias = "prctl", },
582 { .name = "bpf", .errmsg = true, STRARRAY(0, cmd, bpf_cmd), },
583 { .name = "brk", .hexret = true,
584 .arg_scnprintf = { [0] = SCA_HEX, /* brk */ }, },
585 { .name = "chdir", .errmsg = true,
586 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
587 { .name = "chmod", .errmsg = true,
588 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
589 { .name = "chroot", .errmsg = true,
590 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
591 { .name = "clock_gettime", .errmsg = true, STRARRAY(0, clk_id, clockid), },
592 { .name = "clone", .errpid = true, },
593 { .name = "close", .errmsg = true,
594 .arg_scnprintf = { [0] = SCA_CLOSE_FD, /* fd */ }, },
595 { .name = "connect", .errmsg = true, },
596 { .name = "creat", .errmsg = true,
597 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
598 { .name = "dup", .errmsg = true, },
599 { .name = "dup2", .errmsg = true, },
600 { .name = "dup3", .errmsg = true, },
601 { .name = "epoll_ctl", .errmsg = true, STRARRAY(1, op, epoll_ctl_ops), },
602 { .name = "eventfd2", .errmsg = true,
603 .arg_scnprintf = { [1] = SCA_EFD_FLAGS, /* flags */ }, },
604 { .name = "faccessat", .errmsg = true,
605 .arg_scnprintf = { [1] = SCA_FILENAME, /* filename */ }, },
606 { .name = "fadvise64", .errmsg = true, },
607 { .name = "fallocate", .errmsg = true, },
608 { .name = "fchdir", .errmsg = true, },
609 { .name = "fchmod", .errmsg = true, },
610 { .name = "fchmodat", .errmsg = true,
611 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */
612 [1] = SCA_FILENAME, /* filename */ }, },
613 { .name = "fchown", .errmsg = true, },
614 { .name = "fchownat", .errmsg = true,
615 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */
616 [1] = SCA_FILENAME, /* filename */ }, },
617 { .name = "fcntl", .errmsg = true,
618 .arg_scnprintf = { [1] = SCA_STRARRAY, /* cmd */ },
619 .arg_parm = { [1] = &strarray__fcntl_cmds, /* cmd */ }, },
620 { .name = "fdatasync", .errmsg = true, },
621 { .name = "flock", .errmsg = true,
622 .arg_scnprintf = { [1] = SCA_FLOCK, /* cmd */ }, },
623 { .name = "fsetxattr", .errmsg = true, },
624 { .name = "fstat", .errmsg = true, .alias = "newfstat", },
625 { .name = "fstatat", .errmsg = true, .alias = "newfstatat",
626 .arg_scnprintf = { [1] = SCA_FILENAME, /* filename */ }, },
627 { .name = "fstatfs", .errmsg = true, },
628 { .name = "fsync", .errmsg = true, },
629 { .name = "ftruncate", .errmsg = true, },
630 { .name = "futex", .errmsg = true,
631 .arg_scnprintf = { [1] = SCA_FUTEX_OP, /* op */ }, },
632 { .name = "futimesat", .errmsg = true,
633 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */
634 [1] = SCA_FILENAME, /* filename */ }, },
635 { .name = "getdents", .errmsg = true, },
636 { .name = "getdents64", .errmsg = true, },
637 { .name = "getitimer", .errmsg = true, STRARRAY(0, which, itimers), },
638 { .name = "getpid", .errpid = true, },
639 { .name = "getpgid", .errpid = true, },
640 { .name = "getppid", .errpid = true, },
641 { .name = "getrandom", .errmsg = true,
642 .arg_scnprintf = { [2] = SCA_GETRANDOM_FLAGS, /* flags */ }, },
643 { .name = "getrlimit", .errmsg = true, STRARRAY(0, resource, rlimit_resources), },
644 { .name = "getxattr", .errmsg = true,
645 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
646 { .name = "inotify_add_watch", .errmsg = true,
647 .arg_scnprintf = { [1] = SCA_FILENAME, /* pathname */ }, },
648 { .name = "ioctl", .errmsg = true,
650 #if defined(__i386__) || defined(__x86_64__)
652 * FIXME: Make this available to all arches.
654 [1] = SCA_STRHEXARRAY, /* cmd */
655 [2] = SCA_HEX, /* arg */ },
656 .arg_parm = { [1] = &strarray__tioctls, /* cmd */ }, },
658 [2] = SCA_HEX, /* arg */ }, },
660 { .name = "keyctl", .errmsg = true, STRARRAY(0, option, keyctl_options), },
661 { .name = "kill", .errmsg = true,
662 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
663 { .name = "lchown", .errmsg = true,
664 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
665 { .name = "lgetxattr", .errmsg = true,
666 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
667 { .name = "linkat", .errmsg = true,
668 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
669 { .name = "listxattr", .errmsg = true,
670 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
671 { .name = "llistxattr", .errmsg = true,
672 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
673 { .name = "lremovexattr", .errmsg = true,
674 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
675 { .name = "lseek", .errmsg = true,
676 .arg_scnprintf = { [2] = SCA_STRARRAY, /* whence */ },
677 .arg_parm = { [2] = &strarray__whences, /* whence */ }, },
678 { .name = "lsetxattr", .errmsg = true,
679 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
680 { .name = "lstat", .errmsg = true, .alias = "newlstat",
681 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
682 { .name = "lsxattr", .errmsg = true,
683 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
684 { .name = "madvise", .errmsg = true,
685 .arg_scnprintf = { [0] = SCA_HEX, /* start */
686 [2] = SCA_MADV_BHV, /* behavior */ }, },
687 { .name = "mkdir", .errmsg = true,
688 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
689 { .name = "mkdirat", .errmsg = true,
690 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */
691 [1] = SCA_FILENAME, /* pathname */ }, },
692 { .name = "mknod", .errmsg = true,
693 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
694 { .name = "mknodat", .errmsg = true,
695 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */
696 [1] = SCA_FILENAME, /* filename */ }, },
697 { .name = "mlock", .errmsg = true,
698 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
699 { .name = "mlockall", .errmsg = true,
700 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
701 { .name = "mmap", .hexret = true,
702 .arg_scnprintf = { [0] = SCA_HEX, /* addr */
703 [2] = SCA_MMAP_PROT, /* prot */
704 [3] = SCA_MMAP_FLAGS, /* flags */ }, },
705 { .name = "mprotect", .errmsg = true,
706 .arg_scnprintf = { [0] = SCA_HEX, /* start */
707 [2] = SCA_MMAP_PROT, /* prot */ }, },
708 { .name = "mq_unlink", .errmsg = true,
709 .arg_scnprintf = { [0] = SCA_FILENAME, /* u_name */ }, },
710 { .name = "mremap", .hexret = true,
711 .arg_scnprintf = { [0] = SCA_HEX, /* addr */
712 [3] = SCA_MREMAP_FLAGS, /* flags */
713 [4] = SCA_HEX, /* new_addr */ }, },
714 { .name = "munlock", .errmsg = true,
715 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
716 { .name = "munmap", .errmsg = true,
717 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
718 { .name = "name_to_handle_at", .errmsg = true,
719 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
720 { .name = "newfstatat", .errmsg = true,
721 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
722 [1] = SCA_FILENAME, /* filename */ }, },
723 { .name = "open", .errmsg = true,
724 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */
725 [1] = SCA_OPEN_FLAGS, /* flags */ }, },
726 { .name = "open_by_handle_at", .errmsg = true,
727 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
728 [2] = SCA_OPEN_FLAGS, /* flags */ }, },
729 { .name = "openat", .errmsg = true,
730 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
731 [1] = SCA_FILENAME, /* filename */
732 [2] = SCA_OPEN_FLAGS, /* flags */ }, },
733 { .name = "perf_event_open", .errmsg = true,
734 .arg_scnprintf = { [2] = SCA_INT, /* cpu */
735 [3] = SCA_FD, /* group_fd */
736 [4] = SCA_PERF_FLAGS, /* flags */ }, },
737 { .name = "pipe2", .errmsg = true,
738 .arg_scnprintf = { [1] = SCA_PIPE_FLAGS, /* flags */ }, },
739 { .name = "poll", .errmsg = true, .timeout = true, },
740 { .name = "ppoll", .errmsg = true, .timeout = true, },
741 { .name = "pread", .errmsg = true, .alias = "pread64", },
742 { .name = "preadv", .errmsg = true, .alias = "pread", },
743 { .name = "prlimit64", .errmsg = true, STRARRAY(1, resource, rlimit_resources), },
744 { .name = "pwrite", .errmsg = true, .alias = "pwrite64", },
745 { .name = "pwritev", .errmsg = true, },
746 { .name = "read", .errmsg = true, },
747 { .name = "readlink", .errmsg = true,
748 .arg_scnprintf = { [0] = SCA_FILENAME, /* path */ }, },
749 { .name = "readlinkat", .errmsg = true,
750 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
751 [1] = SCA_FILENAME, /* pathname */ }, },
752 { .name = "readv", .errmsg = true, },
753 { .name = "recvfrom", .errmsg = true,
754 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
755 { .name = "recvmmsg", .errmsg = true,
756 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
757 { .name = "recvmsg", .errmsg = true,
758 .arg_scnprintf = { [2] = SCA_MSG_FLAGS, /* flags */ }, },
759 { .name = "removexattr", .errmsg = true,
760 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
761 { .name = "renameat", .errmsg = true,
762 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
763 { .name = "rmdir", .errmsg = true,
764 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
765 { .name = "rt_sigaction", .errmsg = true,
766 .arg_scnprintf = { [0] = SCA_SIGNUM, /* sig */ }, },
767 { .name = "rt_sigprocmask", .errmsg = true, STRARRAY(0, how, sighow), },
768 { .name = "rt_sigqueueinfo", .errmsg = true,
769 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
770 { .name = "rt_tgsigqueueinfo", .errmsg = true,
771 .arg_scnprintf = { [2] = SCA_SIGNUM, /* sig */ }, },
772 { .name = "sched_setscheduler", .errmsg = true,
773 .arg_scnprintf = { [1] = SCA_SCHED_POLICY, /* policy */ }, },
774 { .name = "seccomp", .errmsg = true,
775 .arg_scnprintf = { [0] = SCA_SECCOMP_OP, /* op */
776 [1] = SCA_SECCOMP_FLAGS, /* flags */ }, },
777 { .name = "select", .errmsg = true, .timeout = true, },
778 { .name = "sendmmsg", .errmsg = true,
779 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
780 { .name = "sendmsg", .errmsg = true,
781 .arg_scnprintf = { [2] = SCA_MSG_FLAGS, /* flags */ }, },
782 { .name = "sendto", .errmsg = true,
783 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
784 { .name = "set_tid_address", .errpid = true, },
785 { .name = "setitimer", .errmsg = true, STRARRAY(0, which, itimers), },
786 { .name = "setpgid", .errmsg = true, },
787 { .name = "setrlimit", .errmsg = true, STRARRAY(0, resource, rlimit_resources), },
788 { .name = "setxattr", .errmsg = true,
789 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
790 { .name = "shutdown", .errmsg = true, },
791 { .name = "socket", .errmsg = true,
792 .arg_scnprintf = { [0] = SCA_STRARRAY, /* family */
793 [1] = SCA_SK_TYPE, /* type */ },
794 .arg_parm = { [0] = &strarray__socket_families, /* family */ }, },
795 { .name = "socketpair", .errmsg = true,
796 .arg_scnprintf = { [0] = SCA_STRARRAY, /* family */
797 [1] = SCA_SK_TYPE, /* type */ },
798 .arg_parm = { [0] = &strarray__socket_families, /* family */ }, },
799 { .name = "stat", .errmsg = true, .alias = "newstat",
800 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
801 { .name = "statfs", .errmsg = true,
802 .arg_scnprintf = { [0] = SCA_FILENAME, /* pathname */ }, },
803 { .name = "swapoff", .errmsg = true,
804 .arg_scnprintf = { [0] = SCA_FILENAME, /* specialfile */ }, },
805 { .name = "swapon", .errmsg = true,
806 .arg_scnprintf = { [0] = SCA_FILENAME, /* specialfile */ }, },
807 { .name = "symlinkat", .errmsg = true,
808 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
809 { .name = "tgkill", .errmsg = true,
810 .arg_scnprintf = { [2] = SCA_SIGNUM, /* sig */ }, },
811 { .name = "tkill", .errmsg = true,
812 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
813 { .name = "truncate", .errmsg = true,
814 .arg_scnprintf = { [0] = SCA_FILENAME, /* path */ }, },
815 { .name = "uname", .errmsg = true, .alias = "newuname", },
816 { .name = "unlinkat", .errmsg = true,
817 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
818 [1] = SCA_FILENAME, /* pathname */ }, },
819 { .name = "utime", .errmsg = true,
820 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
821 { .name = "utimensat", .errmsg = true,
822 .arg_scnprintf = { [0] = SCA_FDAT, /* dirfd */
823 [1] = SCA_FILENAME, /* filename */ }, },
824 { .name = "utimes", .errmsg = true,
825 .arg_scnprintf = { [0] = SCA_FILENAME, /* filename */ }, },
826 { .name = "vmsplice", .errmsg = true, },
827 { .name = "wait4", .errpid = true,
828 .arg_scnprintf = { [2] = SCA_WAITID_OPTIONS, /* options */ }, },
829 { .name = "waitid", .errpid = true,
830 .arg_scnprintf = { [3] = SCA_WAITID_OPTIONS, /* options */ }, },
831 { .name = "write", .errmsg = true, },
832 { .name = "writev", .errmsg = true, },
835 static int syscall_fmt__cmp(const void *name, const void *fmtp)
837 const struct syscall_fmt *fmt = fmtp;
838 return strcmp(name, fmt->name);
841 static struct syscall_fmt *syscall_fmt__find(const char *name)
843 const int nmemb = ARRAY_SIZE(syscall_fmts);
844 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
848 struct event_format *tp_format;
850 struct format_field *args;
853 struct syscall_fmt *fmt;
854 size_t (**arg_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
858 static size_t fprintf_duration(unsigned long t, FILE *fp)
860 double duration = (double)t / NSEC_PER_MSEC;
861 size_t printed = fprintf(fp, "(");
864 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
865 else if (duration >= 0.01)
866 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
868 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
869 return printed + fprintf(fp, "): ");
873 * filename.ptr: The filename char pointer that will be vfs_getname'd
874 * filename.entry_str_pos: Where to insert the string translated from
875 * filename.ptr by the vfs_getname tracepoint/kprobe.
877 struct thread_trace {
881 unsigned long nr_events;
882 unsigned long pfmaj, pfmin;
887 short int entry_str_pos;
889 unsigned int namelen;
897 struct intlist *syscall_stats;
900 static struct thread_trace *thread_trace__new(void)
902 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
905 ttrace->paths.max = -1;
907 ttrace->syscall_stats = intlist__new(NULL);
912 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
914 struct thread_trace *ttrace;
919 if (thread__priv(thread) == NULL)
920 thread__set_priv(thread, thread_trace__new());
922 if (thread__priv(thread) == NULL)
925 ttrace = thread__priv(thread);
930 color_fprintf(fp, PERF_COLOR_RED,
931 "WARNING: not enough memory, dropping samples!\n");
935 #define TRACE_PFMAJ (1 << 0)
936 #define TRACE_PFMIN (1 << 1)
938 static const size_t trace__entry_str_size = 2048;
940 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
942 struct thread_trace *ttrace = thread__priv(thread);
944 if (fd > ttrace->paths.max) {
945 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
950 if (ttrace->paths.max != -1) {
951 memset(npath + ttrace->paths.max + 1, 0,
952 (fd - ttrace->paths.max) * sizeof(char *));
954 memset(npath, 0, (fd + 1) * sizeof(char *));
957 ttrace->paths.table = npath;
958 ttrace->paths.max = fd;
961 ttrace->paths.table[fd] = strdup(pathname);
963 return ttrace->paths.table[fd] != NULL ? 0 : -1;
966 static int thread__read_fd_path(struct thread *thread, int fd)
968 char linkname[PATH_MAX], pathname[PATH_MAX];
972 if (thread->pid_ == thread->tid) {
973 scnprintf(linkname, sizeof(linkname),
974 "/proc/%d/fd/%d", thread->pid_, fd);
976 scnprintf(linkname, sizeof(linkname),
977 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
980 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
983 ret = readlink(linkname, pathname, sizeof(pathname));
985 if (ret < 0 || ret > st.st_size)
988 pathname[ret] = '\0';
989 return trace__set_fd_pathname(thread, fd, pathname);
992 static const char *thread__fd_path(struct thread *thread, int fd,
995 struct thread_trace *ttrace = thread__priv(thread);
1003 if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
1006 ++trace->stats.proc_getname;
1007 if (thread__read_fd_path(thread, fd))
1011 return ttrace->paths.table[fd];
1014 static size_t syscall_arg__scnprintf_fd(char *bf, size_t size,
1015 struct syscall_arg *arg)
1018 size_t printed = scnprintf(bf, size, "%d", fd);
1019 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1022 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1027 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1028 struct syscall_arg *arg)
1031 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1032 struct thread_trace *ttrace = thread__priv(arg->thread);
1034 if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1035 zfree(&ttrace->paths.table[fd]);
1040 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1043 struct thread_trace *ttrace = thread__priv(thread);
1045 ttrace->filename.ptr = ptr;
1046 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1049 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1050 struct syscall_arg *arg)
1052 unsigned long ptr = arg->val;
1054 if (!arg->trace->vfs_getname)
1055 return scnprintf(bf, size, "%#x", ptr);
1057 thread__set_filename_pos(arg->thread, bf, ptr);
1061 static bool trace__filter_duration(struct trace *trace, double t)
1063 return t < (trace->duration_filter * NSEC_PER_MSEC);
1066 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1068 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1070 return fprintf(fp, "%10.3f ", ts);
1073 static bool done = false;
1074 static bool interrupted = false;
1076 static void sig_handler(int sig)
1079 interrupted = sig == SIGINT;
1082 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1083 u64 duration, u64 tstamp, FILE *fp)
1085 size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
1086 printed += fprintf_duration(duration, fp);
1088 if (trace->multiple_threads) {
1089 if (trace->show_comm)
1090 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1091 printed += fprintf(fp, "%d ", thread->tid);
1097 static int trace__process_event(struct trace *trace, struct machine *machine,
1098 union perf_event *event, struct perf_sample *sample)
1102 switch (event->header.type) {
1103 case PERF_RECORD_LOST:
1104 color_fprintf(trace->output, PERF_COLOR_RED,
1105 "LOST %" PRIu64 " events!\n", event->lost.lost);
1106 ret = machine__process_lost_event(machine, event, sample);
1109 ret = machine__process_event(machine, event, sample);
1116 static int trace__tool_process(struct perf_tool *tool,
1117 union perf_event *event,
1118 struct perf_sample *sample,
1119 struct machine *machine)
1121 struct trace *trace = container_of(tool, struct trace, tool);
1122 return trace__process_event(trace, machine, event, sample);
1125 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1127 struct machine *machine = vmachine;
1129 if (machine->kptr_restrict_warned)
1132 if (symbol_conf.kptr_restrict) {
1133 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1134 "Check /proc/sys/kernel/kptr_restrict.\n\n"
1135 "Kernel samples will not be resolved.\n");
1136 machine->kptr_restrict_warned = true;
1140 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1143 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1145 int err = symbol__init(NULL);
1150 trace->host = machine__new_host();
1151 if (trace->host == NULL)
1154 if (trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr) < 0)
1157 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1158 evlist->threads, trace__tool_process, false,
1159 trace->opts.proc_map_timeout);
1166 static int syscall__set_arg_fmts(struct syscall *sc)
1168 struct format_field *field;
1171 sc->arg_scnprintf = calloc(sc->nr_args, sizeof(void *));
1172 if (sc->arg_scnprintf == NULL)
1176 sc->arg_parm = sc->fmt->arg_parm;
1178 for (field = sc->args; field; field = field->next) {
1179 if (sc->fmt && sc->fmt->arg_scnprintf[idx])
1180 sc->arg_scnprintf[idx] = sc->fmt->arg_scnprintf[idx];
1181 else if (field->flags & FIELD_IS_POINTER)
1182 sc->arg_scnprintf[idx] = syscall_arg__scnprintf_hex;
1183 else if (strcmp(field->type, "pid_t") == 0)
1184 sc->arg_scnprintf[idx] = SCA_PID;
1185 else if (strcmp(field->type, "umode_t") == 0)
1186 sc->arg_scnprintf[idx] = SCA_MODE_T;
1187 else if ((strcmp(field->type, "int") == 0 ||
1188 strcmp(field->type, "unsigned int") == 0 ||
1189 strcmp(field->type, "long") == 0) &&
1190 (len = strlen(field->name)) >= 2 &&
1191 strcmp(field->name + len - 2, "fd") == 0) {
1193 * /sys/kernel/tracing/events/syscalls/sys_enter*
1194 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1199 sc->arg_scnprintf[idx] = SCA_FD;
1207 static int trace__read_syscall_info(struct trace *trace, int id)
1211 const char *name = syscalltbl__name(trace->sctbl, id);
1216 if (id > trace->syscalls.max) {
1217 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1219 if (nsyscalls == NULL)
1222 if (trace->syscalls.max != -1) {
1223 memset(nsyscalls + trace->syscalls.max + 1, 0,
1224 (id - trace->syscalls.max) * sizeof(*sc));
1226 memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1229 trace->syscalls.table = nsyscalls;
1230 trace->syscalls.max = id;
1233 sc = trace->syscalls.table + id;
1236 sc->fmt = syscall_fmt__find(sc->name);
1238 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1239 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1241 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1242 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1243 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1246 if (IS_ERR(sc->tp_format))
1249 sc->args = sc->tp_format->format.fields;
1250 sc->nr_args = sc->tp_format->format.nr_fields;
1252 * We need to check and discard the first variable '__syscall_nr'
1253 * or 'nr' that mean the syscall number. It is needless here.
1254 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1256 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1257 sc->args = sc->args->next;
1261 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1263 return syscall__set_arg_fmts(sc);
1266 static int trace__validate_ev_qualifier(struct trace *trace)
1269 struct str_node *pos;
1271 trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1272 trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1273 sizeof(trace->ev_qualifier_ids.entries[0]));
1275 if (trace->ev_qualifier_ids.entries == NULL) {
1276 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1284 strlist__for_each(pos, trace->ev_qualifier) {
1285 const char *sc = pos->s;
1286 int id = syscalltbl__id(trace->sctbl, sc);
1290 fputs("Error:\tInvalid syscall ", trace->output);
1293 fputs(", ", trace->output);
1296 fputs(sc, trace->output);
1299 trace->ev_qualifier_ids.entries[i++] = id;
1303 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1304 "\nHint:\tand: 'man syscalls'\n", trace->output);
1305 zfree(&trace->ev_qualifier_ids.entries);
1306 trace->ev_qualifier_ids.nr = 0;
1313 * args is to be interpreted as a series of longs but we need to handle
1314 * 8-byte unaligned accesses. args points to raw_data within the event
1315 * and raw_data is guaranteed to be 8-byte unaligned because it is
1316 * preceded by raw_size which is a u32. So we need to copy args to a temp
1317 * variable to read it. Most notably this avoids extended load instructions
1318 * on unaligned addresses
1321 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1322 unsigned char *args, struct trace *trace,
1323 struct thread *thread)
1329 if (sc->args != NULL) {
1330 struct format_field *field;
1332 struct syscall_arg arg = {
1339 for (field = sc->args; field;
1340 field = field->next, ++arg.idx, bit <<= 1) {
1344 /* special care for unaligned accesses */
1345 p = args + sizeof(unsigned long) * arg.idx;
1346 memcpy(&val, p, sizeof(val));
1349 * Suppress this argument if its value is zero and
1350 * and we don't have a string associated in an
1354 !(sc->arg_scnprintf &&
1355 sc->arg_scnprintf[arg.idx] == SCA_STRARRAY &&
1356 sc->arg_parm[arg.idx]))
1359 printed += scnprintf(bf + printed, size - printed,
1360 "%s%s: ", printed ? ", " : "", field->name);
1361 if (sc->arg_scnprintf && sc->arg_scnprintf[arg.idx]) {
1364 arg.parm = sc->arg_parm[arg.idx];
1365 printed += sc->arg_scnprintf[arg.idx](bf + printed,
1366 size - printed, &arg);
1368 printed += scnprintf(bf + printed, size - printed,
1372 } else if (IS_ERR(sc->tp_format)) {
1374 * If we managed to read the tracepoint /format file, then we
1375 * may end up not having any args, like with gettid(), so only
1376 * print the raw args when we didn't manage to read it.
1381 /* special care for unaligned accesses */
1382 p = args + sizeof(unsigned long) * i;
1383 memcpy(&val, p, sizeof(val));
1384 printed += scnprintf(bf + printed, size - printed,
1386 printed ? ", " : "", i, val);
1394 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1395 union perf_event *event,
1396 struct perf_sample *sample);
1398 static struct syscall *trace__syscall_info(struct trace *trace,
1399 struct perf_evsel *evsel, int id)
1405 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1406 * before that, leaving at a higher verbosity level till that is
1407 * explained. Reproduced with plain ftrace with:
1409 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1410 * grep "NR -1 " /t/trace_pipe
1412 * After generating some load on the machine.
1416 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1417 id, perf_evsel__name(evsel), ++n);
1422 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1423 trace__read_syscall_info(trace, id))
1426 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1429 return &trace->syscalls.table[id];
1433 fprintf(trace->output, "Problems reading syscall %d", id);
1434 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1435 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1436 fputs(" information\n", trace->output);
1441 static void thread__update_stats(struct thread_trace *ttrace,
1442 int id, struct perf_sample *sample)
1444 struct int_node *inode;
1445 struct stats *stats;
1448 inode = intlist__findnew(ttrace->syscall_stats, id);
1452 stats = inode->priv;
1453 if (stats == NULL) {
1454 stats = malloc(sizeof(struct stats));
1458 inode->priv = stats;
1461 if (ttrace->entry_time && sample->time > ttrace->entry_time)
1462 duration = sample->time - ttrace->entry_time;
1464 update_stats(stats, duration);
1467 static int trace__printf_interrupted_entry(struct trace *trace, struct perf_sample *sample)
1469 struct thread_trace *ttrace;
1473 if (trace->current == NULL)
1476 ttrace = thread__priv(trace->current);
1478 if (!ttrace->entry_pending)
1481 duration = sample->time - ttrace->entry_time;
1483 printed = trace__fprintf_entry_head(trace, trace->current, duration, sample->time, trace->output);
1484 printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str);
1485 ttrace->entry_pending = false;
1490 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1491 union perf_event *event __maybe_unused,
1492 struct perf_sample *sample)
1497 struct thread *thread;
1498 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1499 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1500 struct thread_trace *ttrace;
1505 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1506 ttrace = thread__trace(thread, trace->output);
1510 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1512 if (ttrace->entry_str == NULL) {
1513 ttrace->entry_str = malloc(trace__entry_str_size);
1514 if (!ttrace->entry_str)
1518 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1519 trace__printf_interrupted_entry(trace, sample);
1521 ttrace->entry_time = sample->time;
1522 msg = ttrace->entry_str;
1523 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1525 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1526 args, trace, thread);
1529 if (!(trace->duration_filter || trace->summary_only || trace->min_stack)) {
1530 trace__fprintf_entry_head(trace, thread, 1, sample->time, trace->output);
1531 fprintf(trace->output, "%-70s)\n", ttrace->entry_str);
1534 ttrace->entry_pending = true;
1535 /* See trace__vfs_getname & trace__sys_exit */
1536 ttrace->filename.pending_open = false;
1539 if (trace->current != thread) {
1540 thread__put(trace->current);
1541 trace->current = thread__get(thread);
1545 thread__put(thread);
1549 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1550 struct perf_sample *sample,
1551 struct callchain_cursor *cursor)
1553 struct addr_location al;
1555 if (machine__resolve(trace->host, &al, sample) < 0 ||
1556 thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, trace->max_stack))
1562 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1564 /* TODO: user-configurable print_opts */
1565 const unsigned int print_opts = EVSEL__PRINT_SYM |
1567 EVSEL__PRINT_UNKNOWN_AS_ADDR;
1569 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1572 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1573 union perf_event *event __maybe_unused,
1574 struct perf_sample *sample)
1578 struct thread *thread;
1579 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0;
1580 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1581 struct thread_trace *ttrace;
1586 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1587 ttrace = thread__trace(thread, trace->output);
1592 thread__update_stats(ttrace, id, sample);
1594 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1596 if (id == trace->open_id && ret >= 0 && ttrace->filename.pending_open) {
1597 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1598 ttrace->filename.pending_open = false;
1599 ++trace->stats.vfs_getname;
1602 ttrace->exit_time = sample->time;
1604 if (ttrace->entry_time) {
1605 duration = sample->time - ttrace->entry_time;
1606 if (trace__filter_duration(trace, duration))
1608 } else if (trace->duration_filter)
1611 if (sample->callchain) {
1612 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1613 if (callchain_ret == 0) {
1614 if (callchain_cursor.nr < trace->min_stack)
1620 if (trace->summary_only)
1623 trace__fprintf_entry_head(trace, thread, duration, sample->time, trace->output);
1625 if (ttrace->entry_pending) {
1626 fprintf(trace->output, "%-70s", ttrace->entry_str);
1628 fprintf(trace->output, " ... [");
1629 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
1630 fprintf(trace->output, "]: %s()", sc->name);
1633 if (sc->fmt == NULL) {
1635 fprintf(trace->output, ") = %ld", ret);
1636 } else if (ret < 0 && (sc->fmt->errmsg || sc->fmt->errpid)) {
1637 char bf[STRERR_BUFSIZE];
1638 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
1639 *e = audit_errno_to_name(-ret);
1641 fprintf(trace->output, ") = -1 %s %s", e, emsg);
1642 } else if (ret == 0 && sc->fmt->timeout)
1643 fprintf(trace->output, ") = 0 Timeout");
1644 else if (sc->fmt->hexret)
1645 fprintf(trace->output, ") = %#lx", ret);
1646 else if (sc->fmt->errpid) {
1647 struct thread *child = machine__find_thread(trace->host, ret, ret);
1649 if (child != NULL) {
1650 fprintf(trace->output, ") = %ld", ret);
1651 if (child->comm_set)
1652 fprintf(trace->output, " (%s)", thread__comm_str(child));
1658 fputc('\n', trace->output);
1660 if (callchain_ret > 0)
1661 trace__fprintf_callchain(trace, sample);
1662 else if (callchain_ret < 0)
1663 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1665 ttrace->entry_pending = false;
1668 thread__put(thread);
1672 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
1673 union perf_event *event __maybe_unused,
1674 struct perf_sample *sample)
1676 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1677 struct thread_trace *ttrace;
1678 size_t filename_len, entry_str_len, to_move;
1679 ssize_t remaining_space;
1681 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
1686 ttrace = thread__priv(thread);
1690 filename_len = strlen(filename);
1692 if (ttrace->filename.namelen < filename_len) {
1693 char *f = realloc(ttrace->filename.name, filename_len + 1);
1698 ttrace->filename.namelen = filename_len;
1699 ttrace->filename.name = f;
1702 strcpy(ttrace->filename.name, filename);
1703 ttrace->filename.pending_open = true;
1705 if (!ttrace->filename.ptr)
1708 entry_str_len = strlen(ttrace->entry_str);
1709 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
1710 if (remaining_space <= 0)
1713 if (filename_len > (size_t)remaining_space) {
1714 filename += filename_len - remaining_space;
1715 filename_len = remaining_space;
1718 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
1719 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
1720 memmove(pos + filename_len, pos, to_move);
1721 memcpy(pos, filename, filename_len);
1723 ttrace->filename.ptr = 0;
1724 ttrace->filename.entry_str_pos = 0;
1729 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
1730 union perf_event *event __maybe_unused,
1731 struct perf_sample *sample)
1733 u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1734 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
1735 struct thread *thread = machine__findnew_thread(trace->host,
1738 struct thread_trace *ttrace = thread__trace(thread, trace->output);
1743 ttrace->runtime_ms += runtime_ms;
1744 trace->runtime_ms += runtime_ms;
1745 thread__put(thread);
1749 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
1751 perf_evsel__strval(evsel, sample, "comm"),
1752 (pid_t)perf_evsel__intval(evsel, sample, "pid"),
1754 perf_evsel__intval(evsel, sample, "vruntime"));
1755 thread__put(thread);
1759 static void bpf_output__printer(enum binary_printer_ops op,
1760 unsigned int val, void *extra)
1762 FILE *output = extra;
1763 unsigned char ch = (unsigned char)val;
1766 case BINARY_PRINT_CHAR_DATA:
1767 fprintf(output, "%c", isprint(ch) ? ch : '.');
1769 case BINARY_PRINT_DATA_BEGIN:
1770 case BINARY_PRINT_LINE_BEGIN:
1771 case BINARY_PRINT_ADDR:
1772 case BINARY_PRINT_NUM_DATA:
1773 case BINARY_PRINT_NUM_PAD:
1774 case BINARY_PRINT_SEP:
1775 case BINARY_PRINT_CHAR_PAD:
1776 case BINARY_PRINT_LINE_END:
1777 case BINARY_PRINT_DATA_END:
1783 static void bpf_output__fprintf(struct trace *trace,
1784 struct perf_sample *sample)
1786 print_binary(sample->raw_data, sample->raw_size, 8,
1787 bpf_output__printer, trace->output);
1790 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
1791 union perf_event *event __maybe_unused,
1792 struct perf_sample *sample)
1794 int callchain_ret = 0;
1796 if (sample->callchain) {
1797 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1798 if (callchain_ret == 0) {
1799 if (callchain_cursor.nr < trace->min_stack)
1805 trace__printf_interrupted_entry(trace, sample);
1806 trace__fprintf_tstamp(trace, sample->time, trace->output);
1808 if (trace->trace_syscalls)
1809 fprintf(trace->output, "( ): ");
1811 fprintf(trace->output, "%s:", evsel->name);
1813 if (perf_evsel__is_bpf_output(evsel)) {
1814 bpf_output__fprintf(trace, sample);
1815 } else if (evsel->tp_format) {
1816 event_format__fprintf(evsel->tp_format, sample->cpu,
1817 sample->raw_data, sample->raw_size,
1821 fprintf(trace->output, ")\n");
1823 if (callchain_ret > 0)
1824 trace__fprintf_callchain(trace, sample);
1825 else if (callchain_ret < 0)
1826 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1831 static void print_location(FILE *f, struct perf_sample *sample,
1832 struct addr_location *al,
1833 bool print_dso, bool print_sym)
1836 if ((verbose || print_dso) && al->map)
1837 fprintf(f, "%s@", al->map->dso->long_name);
1839 if ((verbose || print_sym) && al->sym)
1840 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
1841 al->addr - al->sym->start);
1843 fprintf(f, "0x%" PRIx64, al->addr);
1845 fprintf(f, "0x%" PRIx64, sample->addr);
1848 static int trace__pgfault(struct trace *trace,
1849 struct perf_evsel *evsel,
1850 union perf_event *event __maybe_unused,
1851 struct perf_sample *sample)
1853 struct thread *thread;
1854 struct addr_location al;
1855 char map_type = 'd';
1856 struct thread_trace *ttrace;
1858 int callchain_ret = 0;
1860 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1862 if (sample->callchain) {
1863 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1864 if (callchain_ret == 0) {
1865 if (callchain_cursor.nr < trace->min_stack)
1871 ttrace = thread__trace(thread, trace->output);
1875 if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
1880 if (trace->summary_only)
1883 thread__find_addr_location(thread, sample->cpumode, MAP__FUNCTION,
1886 trace__fprintf_entry_head(trace, thread, 0, sample->time, trace->output);
1888 fprintf(trace->output, "%sfault [",
1889 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
1892 print_location(trace->output, sample, &al, false, true);
1894 fprintf(trace->output, "] => ");
1896 thread__find_addr_location(thread, sample->cpumode, MAP__VARIABLE,
1900 thread__find_addr_location(thread, sample->cpumode,
1901 MAP__FUNCTION, sample->addr, &al);
1909 print_location(trace->output, sample, &al, true, false);
1911 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
1913 if (callchain_ret > 0)
1914 trace__fprintf_callchain(trace, sample);
1915 else if (callchain_ret < 0)
1916 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1920 thread__put(thread);
1924 static bool skip_sample(struct trace *trace, struct perf_sample *sample)
1926 if ((trace->pid_list && intlist__find(trace->pid_list, sample->pid)) ||
1927 (trace->tid_list && intlist__find(trace->tid_list, sample->tid)))
1930 if (trace->pid_list || trace->tid_list)
1936 static void trace__set_base_time(struct trace *trace,
1937 struct perf_evsel *evsel,
1938 struct perf_sample *sample)
1941 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
1942 * and don't use sample->time unconditionally, we may end up having
1943 * some other event in the future without PERF_SAMPLE_TIME for good
1944 * reason, i.e. we may not be interested in its timestamps, just in
1945 * it taking place, picking some piece of information when it
1946 * appears in our event stream (vfs_getname comes to mind).
1948 if (trace->base_time == 0 && !trace->full_time &&
1949 (evsel->attr.sample_type & PERF_SAMPLE_TIME))
1950 trace->base_time = sample->time;
1953 static int trace__process_sample(struct perf_tool *tool,
1954 union perf_event *event,
1955 struct perf_sample *sample,
1956 struct perf_evsel *evsel,
1957 struct machine *machine __maybe_unused)
1959 struct trace *trace = container_of(tool, struct trace, tool);
1962 tracepoint_handler handler = evsel->handler;
1964 if (skip_sample(trace, sample))
1967 trace__set_base_time(trace, evsel, sample);
1971 handler(trace, evsel, event, sample);
1977 static int parse_target_str(struct trace *trace)
1979 if (trace->opts.target.pid) {
1980 trace->pid_list = intlist__new(trace->opts.target.pid);
1981 if (trace->pid_list == NULL) {
1982 pr_err("Error parsing process id string\n");
1987 if (trace->opts.target.tid) {
1988 trace->tid_list = intlist__new(trace->opts.target.tid);
1989 if (trace->tid_list == NULL) {
1990 pr_err("Error parsing thread id string\n");
1998 static int trace__record(struct trace *trace, int argc, const char **argv)
2000 unsigned int rec_argc, i, j;
2001 const char **rec_argv;
2002 const char * const record_args[] = {
2009 const char * const sc_args[] = { "-e", };
2010 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2011 const char * const majpf_args[] = { "-e", "major-faults" };
2012 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2013 const char * const minpf_args[] = { "-e", "minor-faults" };
2014 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2016 /* +1 is for the event string below */
2017 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2018 majpf_args_nr + minpf_args_nr + argc;
2019 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2021 if (rec_argv == NULL)
2025 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2026 rec_argv[j++] = record_args[i];
2028 if (trace->trace_syscalls) {
2029 for (i = 0; i < sc_args_nr; i++)
2030 rec_argv[j++] = sc_args[i];
2032 /* event string may be different for older kernels - e.g., RHEL6 */
2033 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2034 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2035 else if (is_valid_tracepoint("syscalls:sys_enter"))
2036 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2038 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2043 if (trace->trace_pgfaults & TRACE_PFMAJ)
2044 for (i = 0; i < majpf_args_nr; i++)
2045 rec_argv[j++] = majpf_args[i];
2047 if (trace->trace_pgfaults & TRACE_PFMIN)
2048 for (i = 0; i < minpf_args_nr; i++)
2049 rec_argv[j++] = minpf_args[i];
2051 for (i = 0; i < (unsigned int)argc; i++)
2052 rec_argv[j++] = argv[i];
2054 return cmd_record(j, rec_argv, NULL);
2057 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2059 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2061 struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2066 if (perf_evsel__field(evsel, "pathname") == NULL) {
2067 perf_evsel__delete(evsel);
2071 evsel->handler = trace__vfs_getname;
2072 perf_evlist__add(evlist, evsel);
2076 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2078 struct perf_evsel *evsel;
2079 struct perf_event_attr attr = {
2080 .type = PERF_TYPE_SOFTWARE,
2084 attr.config = config;
2085 attr.sample_period = 1;
2087 event_attr_init(&attr);
2089 evsel = perf_evsel__new(&attr);
2091 evsel->handler = trace__pgfault;
2096 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2098 const u32 type = event->header.type;
2099 struct perf_evsel *evsel;
2101 if (type != PERF_RECORD_SAMPLE) {
2102 trace__process_event(trace, trace->host, event, sample);
2106 evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2107 if (evsel == NULL) {
2108 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2112 trace__set_base_time(trace, evsel, sample);
2114 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2115 sample->raw_data == NULL) {
2116 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2117 perf_evsel__name(evsel), sample->tid,
2118 sample->cpu, sample->raw_size);
2120 tracepoint_handler handler = evsel->handler;
2121 handler(trace, evsel, event, sample);
2125 static int trace__add_syscall_newtp(struct trace *trace)
2128 struct perf_evlist *evlist = trace->evlist;
2129 struct perf_evsel *sys_enter, *sys_exit;
2131 sys_enter = perf_evsel__syscall_newtp("sys_enter", trace__sys_enter);
2132 if (sys_enter == NULL)
2135 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2136 goto out_delete_sys_enter;
2138 sys_exit = perf_evsel__syscall_newtp("sys_exit", trace__sys_exit);
2139 if (sys_exit == NULL)
2140 goto out_delete_sys_enter;
2142 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2143 goto out_delete_sys_exit;
2145 perf_evlist__add(evlist, sys_enter);
2146 perf_evlist__add(evlist, sys_exit);
2148 if (callchain_param.enabled && !trace->kernel_syscallchains) {
2150 * We're interested only in the user space callchain
2151 * leading to the syscall, allow overriding that for
2152 * debugging reasons using --kernel_syscall_callchains
2154 sys_exit->attr.exclude_callchain_kernel = 1;
2157 trace->syscalls.events.sys_enter = sys_enter;
2158 trace->syscalls.events.sys_exit = sys_exit;
2164 out_delete_sys_exit:
2165 perf_evsel__delete_priv(sys_exit);
2166 out_delete_sys_enter:
2167 perf_evsel__delete_priv(sys_enter);
2171 static int trace__set_ev_qualifier_filter(struct trace *trace)
2174 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2175 trace->ev_qualifier_ids.nr,
2176 trace->ev_qualifier_ids.entries);
2181 if (!perf_evsel__append_filter(trace->syscalls.events.sys_enter, "&&", filter))
2182 err = perf_evsel__append_filter(trace->syscalls.events.sys_exit, "&&", filter);
2192 static int trace__run(struct trace *trace, int argc, const char **argv)
2194 struct perf_evlist *evlist = trace->evlist;
2195 struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2197 unsigned long before;
2198 const bool forks = argc > 0;
2199 bool draining = false;
2203 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2204 goto out_error_raw_syscalls;
2206 if (trace->trace_syscalls)
2207 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2209 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2210 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2211 if (pgfault_maj == NULL)
2213 perf_evlist__add(evlist, pgfault_maj);
2216 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2217 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2218 if (pgfault_min == NULL)
2220 perf_evlist__add(evlist, pgfault_min);
2224 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2225 trace__sched_stat_runtime))
2226 goto out_error_sched_stat_runtime;
2228 err = perf_evlist__create_maps(evlist, &trace->opts.target);
2230 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2231 goto out_delete_evlist;
2234 err = trace__symbols_init(trace, evlist);
2236 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2237 goto out_delete_evlist;
2240 perf_evlist__config(evlist, &trace->opts, NULL);
2242 if (callchain_param.enabled) {
2243 bool use_identifier = false;
2245 if (trace->syscalls.events.sys_exit) {
2246 perf_evsel__config_callchain(trace->syscalls.events.sys_exit,
2247 &trace->opts, &callchain_param);
2248 use_identifier = true;
2252 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2253 use_identifier = true;
2257 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2258 use_identifier = true;
2261 if (use_identifier) {
2263 * Now we have evsels with different sample_ids, use
2264 * PERF_SAMPLE_IDENTIFIER to map from sample to evsel
2265 * from a fixed position in each ring buffer record.
2267 * As of this the changeset introducing this comment, this
2268 * isn't strictly needed, as the fields that can come before
2269 * PERF_SAMPLE_ID are all used, but we'll probably disable
2270 * some of those for things like copying the payload of
2271 * pointer syscall arguments, and for vfs_getname we don't
2272 * need PERF_SAMPLE_ADDR and PERF_SAMPLE_IP, so do this
2273 * here as a warning we need to use PERF_SAMPLE_IDENTIFIER.
2275 perf_evlist__set_sample_bit(evlist, IDENTIFIER);
2276 perf_evlist__reset_sample_bit(evlist, ID);
2280 signal(SIGCHLD, sig_handler);
2281 signal(SIGINT, sig_handler);
2284 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2287 fprintf(trace->output, "Couldn't run the workload!\n");
2288 goto out_delete_evlist;
2292 err = perf_evlist__open(evlist);
2294 goto out_error_open;
2296 err = bpf__apply_obj_config();
2298 char errbuf[BUFSIZ];
2300 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2301 pr_err("ERROR: Apply config to BPF failed: %s\n",
2303 goto out_error_open;
2307 * Better not use !target__has_task() here because we need to cover the
2308 * case where no threads were specified in the command line, but a
2309 * workload was, and in that case we will fill in the thread_map when
2310 * we fork the workload in perf_evlist__prepare_workload.
2312 if (trace->filter_pids.nr > 0)
2313 err = perf_evlist__set_filter_pids(evlist, trace->filter_pids.nr, trace->filter_pids.entries);
2314 else if (thread_map__pid(evlist->threads, 0) == -1)
2315 err = perf_evlist__set_filter_pid(evlist, getpid());
2320 if (trace->ev_qualifier_ids.nr > 0) {
2321 err = trace__set_ev_qualifier_filter(trace);
2325 pr_debug("event qualifier tracepoint filter: %s\n",
2326 trace->syscalls.events.sys_exit->filter);
2329 err = perf_evlist__apply_filters(evlist, &evsel);
2331 goto out_error_apply_filters;
2333 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages, false);
2335 goto out_error_mmap;
2337 if (!target__none(&trace->opts.target))
2338 perf_evlist__enable(evlist);
2341 perf_evlist__start_workload(evlist);
2343 trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2344 evlist->threads->nr > 1 ||
2345 perf_evlist__first(evlist)->attr.inherit;
2347 before = trace->nr_events;
2349 for (i = 0; i < evlist->nr_mmaps; i++) {
2350 union perf_event *event;
2352 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
2353 struct perf_sample sample;
2357 err = perf_evlist__parse_sample(evlist, event, &sample);
2359 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2363 trace__handle_event(trace, event, &sample);
2365 perf_evlist__mmap_consume(evlist, i);
2370 if (done && !draining) {
2371 perf_evlist__disable(evlist);
2377 if (trace->nr_events == before) {
2378 int timeout = done ? 100 : -1;
2380 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
2381 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP) == 0)
2391 thread__zput(trace->current);
2393 perf_evlist__disable(evlist);
2397 trace__fprintf_thread_summary(trace, trace->output);
2399 if (trace->show_tool_stats) {
2400 fprintf(trace->output, "Stats:\n "
2401 " vfs_getname : %" PRIu64 "\n"
2402 " proc_getname: %" PRIu64 "\n",
2403 trace->stats.vfs_getname,
2404 trace->stats.proc_getname);
2409 perf_evlist__delete(evlist);
2410 trace->evlist = NULL;
2411 trace->live = false;
2414 char errbuf[BUFSIZ];
2416 out_error_sched_stat_runtime:
2417 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
2420 out_error_raw_syscalls:
2421 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
2425 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
2429 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
2432 fprintf(trace->output, "%s\n", errbuf);
2433 goto out_delete_evlist;
2435 out_error_apply_filters:
2436 fprintf(trace->output,
2437 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
2438 evsel->filter, perf_evsel__name(evsel), errno,
2439 strerror_r(errno, errbuf, sizeof(errbuf)));
2440 goto out_delete_evlist;
2443 fprintf(trace->output, "Not enough memory to run!\n");
2444 goto out_delete_evlist;
2447 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
2448 goto out_delete_evlist;
2451 static int trace__replay(struct trace *trace)
2453 const struct perf_evsel_str_handler handlers[] = {
2454 { "probe:vfs_getname", trace__vfs_getname, },
2456 struct perf_data_file file = {
2458 .mode = PERF_DATA_MODE_READ,
2459 .force = trace->force,
2461 struct perf_session *session;
2462 struct perf_evsel *evsel;
2465 trace->tool.sample = trace__process_sample;
2466 trace->tool.mmap = perf_event__process_mmap;
2467 trace->tool.mmap2 = perf_event__process_mmap2;
2468 trace->tool.comm = perf_event__process_comm;
2469 trace->tool.exit = perf_event__process_exit;
2470 trace->tool.fork = perf_event__process_fork;
2471 trace->tool.attr = perf_event__process_attr;
2472 trace->tool.tracing_data = perf_event__process_tracing_data;
2473 trace->tool.build_id = perf_event__process_build_id;
2475 trace->tool.ordered_events = true;
2476 trace->tool.ordering_requires_timestamps = true;
2478 /* add tid to output */
2479 trace->multiple_threads = true;
2481 session = perf_session__new(&file, false, &trace->tool);
2482 if (session == NULL)
2485 if (symbol__init(&session->header.env) < 0)
2488 trace->host = &session->machines.host;
2490 err = perf_session__set_tracepoints_handlers(session, handlers);
2494 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2495 "raw_syscalls:sys_enter");
2496 /* older kernels have syscalls tp versus raw_syscalls */
2498 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2499 "syscalls:sys_enter");
2502 (perf_evsel__init_syscall_tp(evsel, trace__sys_enter) < 0 ||
2503 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
2504 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
2508 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2509 "raw_syscalls:sys_exit");
2511 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2512 "syscalls:sys_exit");
2514 (perf_evsel__init_syscall_tp(evsel, trace__sys_exit) < 0 ||
2515 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
2516 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
2520 evlist__for_each(session->evlist, evsel) {
2521 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
2522 (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
2523 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
2524 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
2525 evsel->handler = trace__pgfault;
2528 err = parse_target_str(trace);
2534 err = perf_session__process_events(session);
2536 pr_err("Failed to process events, error %d", err);
2538 else if (trace->summary)
2539 trace__fprintf_thread_summary(trace, trace->output);
2542 perf_session__delete(session);
2547 static size_t trace__fprintf_threads_header(FILE *fp)
2551 printed = fprintf(fp, "\n Summary of events:\n\n");
2556 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
2557 struct stats *stats;
2562 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
2563 struct stats *stats = source->priv;
2565 entry->syscall = source->i;
2566 entry->stats = stats;
2567 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
2570 static size_t thread__dump_stats(struct thread_trace *ttrace,
2571 struct trace *trace, FILE *fp)
2576 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
2578 if (syscall_stats == NULL)
2581 printed += fprintf(fp, "\n");
2583 printed += fprintf(fp, " syscall calls total min avg max stddev\n");
2584 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
2585 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n");
2587 resort_rb__for_each(nd, syscall_stats) {
2588 struct stats *stats = syscall_stats_entry->stats;
2590 double min = (double)(stats->min) / NSEC_PER_MSEC;
2591 double max = (double)(stats->max) / NSEC_PER_MSEC;
2592 double avg = avg_stats(stats);
2594 u64 n = (u64) stats->n;
2596 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
2597 avg /= NSEC_PER_MSEC;
2599 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
2600 printed += fprintf(fp, " %-15s", sc->name);
2601 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
2602 n, syscall_stats_entry->msecs, min, avg);
2603 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
2607 resort_rb__delete(syscall_stats);
2608 printed += fprintf(fp, "\n\n");
2613 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
2616 struct thread_trace *ttrace = thread__priv(thread);
2622 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
2624 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
2625 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
2626 printed += fprintf(fp, "%.1f%%", ratio);
2628 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
2630 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
2632 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
2633 else if (fputc('\n', fp) != EOF)
2636 printed += thread__dump_stats(ttrace, trace, fp);
2641 static unsigned long thread__nr_events(struct thread_trace *ttrace)
2643 return ttrace ? ttrace->nr_events : 0;
2646 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
2647 struct thread *thread;
2650 entry->thread = rb_entry(nd, struct thread, rb_node);
2653 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
2655 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host);
2656 size_t printed = trace__fprintf_threads_header(fp);
2659 if (threads == NULL) {
2660 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
2664 resort_rb__for_each(nd, threads)
2665 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
2667 resort_rb__delete(threads);
2672 static int trace__set_duration(const struct option *opt, const char *str,
2673 int unset __maybe_unused)
2675 struct trace *trace = opt->value;
2677 trace->duration_filter = atof(str);
2681 static int trace__set_filter_pids(const struct option *opt, const char *str,
2682 int unset __maybe_unused)
2686 struct trace *trace = opt->value;
2688 * FIXME: introduce a intarray class, plain parse csv and create a
2689 * { int nr, int entries[] } struct...
2691 struct intlist *list = intlist__new(str);
2696 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
2697 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
2699 if (trace->filter_pids.entries == NULL)
2702 trace->filter_pids.entries[0] = getpid();
2704 for (i = 1; i < trace->filter_pids.nr; ++i)
2705 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
2707 intlist__delete(list);
2713 static int trace__open_output(struct trace *trace, const char *filename)
2717 if (!stat(filename, &st) && st.st_size) {
2718 char oldname[PATH_MAX];
2720 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
2722 rename(filename, oldname);
2725 trace->output = fopen(filename, "w");
2727 return trace->output == NULL ? -errno : 0;
2730 static int parse_pagefaults(const struct option *opt, const char *str,
2731 int unset __maybe_unused)
2733 int *trace_pgfaults = opt->value;
2735 if (strcmp(str, "all") == 0)
2736 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
2737 else if (strcmp(str, "maj") == 0)
2738 *trace_pgfaults |= TRACE_PFMAJ;
2739 else if (strcmp(str, "min") == 0)
2740 *trace_pgfaults |= TRACE_PFMIN;
2747 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
2749 struct perf_evsel *evsel;
2751 evlist__for_each(evlist, evsel)
2752 evsel->handler = handler;
2755 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
2757 const char *trace_usage[] = {
2758 "perf trace [<options>] [<command>]",
2759 "perf trace [<options>] -- <command> [<options>]",
2760 "perf trace record [<options>] [<command>]",
2761 "perf trace record [<options>] -- <command> [<options>]",
2764 struct trace trace = {
2773 .user_freq = UINT_MAX,
2774 .user_interval = ULLONG_MAX,
2775 .no_buffering = true,
2776 .mmap_pages = UINT_MAX,
2777 .proc_map_timeout = 500,
2781 .trace_syscalls = true,
2782 .kernel_syscallchains = false,
2783 .max_stack = UINT_MAX,
2785 const char *output_name = NULL;
2786 const char *ev_qualifier_str = NULL;
2787 const struct option trace_options[] = {
2788 OPT_CALLBACK(0, "event", &trace.evlist, "event",
2789 "event selector. use 'perf list' to list available events",
2790 parse_events_option),
2791 OPT_BOOLEAN(0, "comm", &trace.show_comm,
2792 "show the thread COMM next to its id"),
2793 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
2794 OPT_STRING('e', "expr", &ev_qualifier_str, "expr", "list of syscalls to trace"),
2795 OPT_STRING('o', "output", &output_name, "file", "output file name"),
2796 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
2797 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
2798 "trace events on existing process id"),
2799 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
2800 "trace events on existing thread id"),
2801 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
2802 "pids to filter (by the kernel)", trace__set_filter_pids),
2803 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
2804 "system-wide collection from all CPUs"),
2805 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
2806 "list of cpus to monitor"),
2807 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
2808 "child tasks do not inherit counters"),
2809 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
2810 "number of mmap data pages",
2811 perf_evlist__parse_mmap_pages),
2812 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
2814 OPT_CALLBACK(0, "duration", &trace, "float",
2815 "show only events with duration > N.M ms",
2816 trace__set_duration),
2817 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
2818 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
2819 OPT_BOOLEAN('T', "time", &trace.full_time,
2820 "Show full timestamp, not time relative to first start"),
2821 OPT_BOOLEAN('s', "summary", &trace.summary_only,
2822 "Show only syscall summary with statistics"),
2823 OPT_BOOLEAN('S', "with-summary", &trace.summary,
2824 "Show all syscalls and summary with statistics"),
2825 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
2826 "Trace pagefaults", parse_pagefaults, "maj"),
2827 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
2828 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
2829 OPT_CALLBACK(0, "call-graph", &trace.opts,
2830 "record_mode[,record_size]", record_callchain_help,
2831 &record_parse_callchain_opt),
2832 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
2833 "Show the kernel callchains on the syscall exit path"),
2834 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
2835 "Set the minimum stack depth when parsing the callchain, "
2836 "anything below the specified depth will be ignored."),
2837 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
2838 "Set the maximum stack depth when parsing the callchain, "
2839 "anything beyond the specified depth will be ignored. "
2840 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
2841 OPT_UINTEGER(0, "proc-map-timeout", &trace.opts.proc_map_timeout,
2842 "per thread proc mmap processing timeout in ms"),
2845 bool __maybe_unused max_stack_user_set = true;
2846 bool mmap_pages_user_set = true;
2847 const char * const trace_subcommands[] = { "record", NULL };
2851 signal(SIGSEGV, sighandler_dump_stack);
2852 signal(SIGFPE, sighandler_dump_stack);
2854 trace.evlist = perf_evlist__new();
2855 trace.sctbl = syscalltbl__new();
2857 if (trace.evlist == NULL || trace.sctbl == NULL) {
2858 pr_err("Not enough memory to run!\n");
2863 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
2864 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2866 err = bpf__setup_stdout(trace.evlist);
2868 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
2869 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
2875 if (trace.trace_pgfaults) {
2876 trace.opts.sample_address = true;
2877 trace.opts.sample_time = true;
2880 if (trace.opts.mmap_pages == UINT_MAX)
2881 mmap_pages_user_set = false;
2883 if (trace.max_stack == UINT_MAX) {
2884 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl_perf_event_max_stack;
2885 max_stack_user_set = false;
2888 #ifdef HAVE_DWARF_UNWIND_SUPPORT
2889 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled && trace.trace_syscalls)
2890 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
2893 if (callchain_param.enabled) {
2894 if (!mmap_pages_user_set && geteuid() == 0)
2895 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
2897 symbol_conf.use_callchain = true;
2900 if (trace.evlist->nr_entries > 0)
2901 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
2903 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
2904 return trace__record(&trace, argc-1, &argv[1]);
2906 /* summary_only implies summary option, but don't overwrite summary if set */
2907 if (trace.summary_only)
2908 trace.summary = trace.summary_only;
2910 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
2911 trace.evlist->nr_entries == 0 /* Was --events used? */) {
2912 pr_err("Please specify something to trace.\n");
2916 if (!trace.trace_syscalls && ev_qualifier_str) {
2917 pr_err("The -e option can't be used with --no-syscalls.\n");
2921 if (output_name != NULL) {
2922 err = trace__open_output(&trace, output_name);
2924 perror("failed to create output file");
2929 trace.open_id = syscalltbl__id(trace.sctbl, "open");
2931 if (ev_qualifier_str != NULL) {
2932 const char *s = ev_qualifier_str;
2933 struct strlist_config slist_config = {
2934 .dirname = system_path(STRACE_GROUPS_DIR),
2937 trace.not_ev_qualifier = *s == '!';
2938 if (trace.not_ev_qualifier)
2940 trace.ev_qualifier = strlist__new(s, &slist_config);
2941 if (trace.ev_qualifier == NULL) {
2942 fputs("Not enough memory to parse event qualifier",
2948 err = trace__validate_ev_qualifier(&trace);
2953 err = target__validate(&trace.opts.target);
2955 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
2956 fprintf(trace.output, "%s", bf);
2960 err = target__parse_uid(&trace.opts.target);
2962 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
2963 fprintf(trace.output, "%s", bf);
2967 if (!argc && target__none(&trace.opts.target))
2968 trace.opts.target.system_wide = true;
2971 err = trace__replay(&trace);
2973 err = trace__run(&trace, argc, argv);
2976 if (output_name != NULL)
2977 fclose(trace.output);