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 */
44 #include <linux/err.h>
45 #include <linux/filter.h>
46 #include <linux/audit.h>
47 #include <linux/random.h>
48 #include <linux/stringify.h>
49 #include <linux/time64.h>
52 # define O_CLOEXEC 02000000
56 struct perf_tool tool;
57 struct syscalltbl *sctbl;
60 struct syscall *table;
62 struct perf_evsel *sys_enter,
66 struct record_opts opts;
67 struct perf_evlist *evlist;
69 struct thread *current;
72 unsigned long nr_events;
73 struct strlist *ev_qualifier;
82 double duration_filter;
88 unsigned int max_stack;
89 unsigned int min_stack;
90 bool not_ev_qualifier;
94 bool multiple_threads;
100 bool kernel_syscallchains;
110 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
111 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
115 #define TP_UINT_FIELD(bits) \
116 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
119 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
128 #define TP_UINT_FIELD__SWAPPED(bits) \
129 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
132 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
133 return bswap_##bits(value);\
136 TP_UINT_FIELD__SWAPPED(16);
137 TP_UINT_FIELD__SWAPPED(32);
138 TP_UINT_FIELD__SWAPPED(64);
140 static int tp_field__init_uint(struct tp_field *field,
141 struct format_field *format_field,
144 field->offset = format_field->offset;
146 switch (format_field->size) {
148 field->integer = tp_field__u8;
151 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
154 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
157 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
166 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
168 return sample->raw_data + field->offset;
171 static int tp_field__init_ptr(struct tp_field *field, struct format_field *format_field)
173 field->offset = format_field->offset;
174 field->pointer = tp_field__ptr;
181 struct tp_field args, ret;
185 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
186 struct tp_field *field,
189 struct format_field *format_field = perf_evsel__field(evsel, name);
191 if (format_field == NULL)
194 return tp_field__init_uint(field, format_field, evsel->needs_swap);
197 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
198 ({ struct syscall_tp *sc = evsel->priv;\
199 perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
201 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
202 struct tp_field *field,
205 struct format_field *format_field = perf_evsel__field(evsel, name);
207 if (format_field == NULL)
210 return tp_field__init_ptr(field, format_field);
213 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
214 ({ struct syscall_tp *sc = evsel->priv;\
215 perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
217 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
220 perf_evsel__delete(evsel);
223 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel, void *handler)
225 evsel->priv = malloc(sizeof(struct syscall_tp));
226 if (evsel->priv != NULL) {
227 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
230 evsel->handler = handler;
241 static struct perf_evsel *perf_evsel__syscall_newtp(const char *direction, void *handler)
243 struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
245 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
247 evsel = perf_evsel__newtp("syscalls", direction);
252 if (perf_evsel__init_syscall_tp(evsel, handler))
258 perf_evsel__delete_priv(evsel);
262 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
263 ({ struct syscall_tp *fields = evsel->priv; \
264 fields->name.integer(&fields->name, sample); })
266 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
267 ({ struct syscall_tp *fields = evsel->priv; \
268 fields->name.pointer(&fields->name, sample); })
272 struct thread *thread;
282 const char **entries;
285 #define DEFINE_STRARRAY(array) struct strarray strarray__##array = { \
286 .nr_entries = ARRAY_SIZE(array), \
290 #define DEFINE_STRARRAY_OFFSET(array, off) struct strarray strarray__##array = { \
292 .nr_entries = ARRAY_SIZE(array), \
296 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
298 struct syscall_arg *arg)
300 struct strarray *sa = arg->parm;
301 int idx = arg->val - sa->offset;
303 if (idx < 0 || idx >= sa->nr_entries)
304 return scnprintf(bf, size, intfmt, arg->val);
306 return scnprintf(bf, size, "%s", sa->entries[idx]);
309 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
310 struct syscall_arg *arg)
312 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
315 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
317 #if defined(__i386__) || defined(__x86_64__)
319 * FIXME: Make this available to all arches as soon as the ioctl beautifier
320 * gets rewritten to support all arches.
322 static size_t syscall_arg__scnprintf_strhexarray(char *bf, size_t size,
323 struct syscall_arg *arg)
325 return __syscall_arg__scnprintf_strarray(bf, size, "%#x", arg);
328 #define SCA_STRHEXARRAY syscall_arg__scnprintf_strhexarray
329 #endif /* defined(__i386__) || defined(__x86_64__) */
331 static size_t syscall_arg__scnprintf_fd(char *bf, size_t size,
332 struct syscall_arg *arg);
334 #define SCA_FD syscall_arg__scnprintf_fd
337 #define AT_FDCWD -100
340 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
341 struct syscall_arg *arg)
346 return scnprintf(bf, size, "CWD");
348 return syscall_arg__scnprintf_fd(bf, size, arg);
351 #define SCA_FDAT syscall_arg__scnprintf_fd_at
353 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
354 struct syscall_arg *arg);
356 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
358 static size_t syscall_arg__scnprintf_hex(char *bf, size_t size,
359 struct syscall_arg *arg)
361 return scnprintf(bf, size, "%#lx", arg->val);
364 #define SCA_HEX syscall_arg__scnprintf_hex
366 static size_t syscall_arg__scnprintf_int(char *bf, size_t size,
367 struct syscall_arg *arg)
369 return scnprintf(bf, size, "%d", arg->val);
372 #define SCA_INT syscall_arg__scnprintf_int
374 static const char *bpf_cmd[] = {
375 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
376 "MAP_GET_NEXT_KEY", "PROG_LOAD",
378 static DEFINE_STRARRAY(bpf_cmd);
380 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
381 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, 1);
383 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
384 static DEFINE_STRARRAY(itimers);
386 static const char *keyctl_options[] = {
387 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
388 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
389 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
390 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
391 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
393 static DEFINE_STRARRAY(keyctl_options);
395 static const char *whences[] = { "SET", "CUR", "END",
403 static DEFINE_STRARRAY(whences);
405 static const char *fcntl_cmds[] = {
406 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
407 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "F_GETLK64",
408 "F_SETLK64", "F_SETLKW64", "F_SETOWN_EX", "F_GETOWN_EX",
411 static DEFINE_STRARRAY(fcntl_cmds);
413 static const char *rlimit_resources[] = {
414 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
415 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
418 static DEFINE_STRARRAY(rlimit_resources);
420 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
421 static DEFINE_STRARRAY(sighow);
423 static const char *clockid[] = {
424 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
425 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
426 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
428 static DEFINE_STRARRAY(clockid);
430 static const char *socket_families[] = {
431 "UNSPEC", "LOCAL", "INET", "AX25", "IPX", "APPLETALK", "NETROM",
432 "BRIDGE", "ATMPVC", "X25", "INET6", "ROSE", "DECnet", "NETBEUI",
433 "SECURITY", "KEY", "NETLINK", "PACKET", "ASH", "ECONET", "ATMSVC",
434 "RDS", "SNA", "IRDA", "PPPOX", "WANPIPE", "LLC", "IB", "CAN", "TIPC",
435 "BLUETOOTH", "IUCV", "RXRPC", "ISDN", "PHONET", "IEEE802154", "CAIF",
436 "ALG", "NFC", "VSOCK",
438 static DEFINE_STRARRAY(socket_families);
440 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
441 struct syscall_arg *arg)
446 if (mode == F_OK) /* 0 */
447 return scnprintf(bf, size, "F");
449 if (mode & n##_OK) { \
450 printed += scnprintf(bf + printed, size - printed, "%s", #n); \
460 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
465 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
467 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
468 struct syscall_arg *arg);
470 #define SCA_FILENAME syscall_arg__scnprintf_filename
472 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
473 struct syscall_arg *arg)
475 int printed = 0, flags = arg->val;
478 if (flags & O_##n) { \
479 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
488 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
493 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
495 #if defined(__i386__) || defined(__x86_64__)
497 * FIXME: Make this available to all arches.
499 #define TCGETS 0x5401
501 static const char *tioctls[] = {
502 "TCGETS", "TCSETS", "TCSETSW", "TCSETSF", "TCGETA", "TCSETA", "TCSETAW",
503 "TCSETAF", "TCSBRK", "TCXONC", "TCFLSH", "TIOCEXCL", "TIOCNXCL",
504 "TIOCSCTTY", "TIOCGPGRP", "TIOCSPGRP", "TIOCOUTQ", "TIOCSTI",
505 "TIOCGWINSZ", "TIOCSWINSZ", "TIOCMGET", "TIOCMBIS", "TIOCMBIC",
506 "TIOCMSET", "TIOCGSOFTCAR", "TIOCSSOFTCAR", "FIONREAD", "TIOCLINUX",
507 "TIOCCONS", "TIOCGSERIAL", "TIOCSSERIAL", "TIOCPKT", "FIONBIO",
508 "TIOCNOTTY", "TIOCSETD", "TIOCGETD", "TCSBRKP", [0x27] = "TIOCSBRK",
509 "TIOCCBRK", "TIOCGSID", "TCGETS2", "TCSETS2", "TCSETSW2", "TCSETSF2",
510 "TIOCGRS485", "TIOCSRS485", "TIOCGPTN", "TIOCSPTLCK",
511 "TIOCGDEV||TCGETX", "TCSETX", "TCSETXF", "TCSETXW", "TIOCSIG",
512 "TIOCVHANGUP", "TIOCGPKT", "TIOCGPTLCK", "TIOCGEXCL",
513 [0x50] = "FIONCLEX", "FIOCLEX", "FIOASYNC", "TIOCSERCONFIG",
514 "TIOCSERGWILD", "TIOCSERSWILD", "TIOCGLCKTRMIOS", "TIOCSLCKTRMIOS",
515 "TIOCSERGSTRUCT", "TIOCSERGETLSR", "TIOCSERGETMULTI", "TIOCSERSETMULTI",
516 "TIOCMIWAIT", "TIOCGICOUNT", [0x60] = "FIOQSIZE",
519 static DEFINE_STRARRAY_OFFSET(tioctls, 0x5401);
520 #endif /* defined(__i386__) || defined(__x86_64__) */
522 #ifndef GRND_NONBLOCK
523 #define GRND_NONBLOCK 0x0001
526 #define GRND_RANDOM 0x0002
529 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
530 struct syscall_arg *arg)
532 int printed = 0, flags = arg->val;
535 if (flags & GRND_##n) { \
536 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
537 flags &= ~GRND_##n; \
545 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
550 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
552 #define STRARRAY(arg, name, array) \
553 .arg_scnprintf = { [arg] = SCA_STRARRAY, }, \
554 .arg_parm = { [arg] = &strarray__##array, }
556 #include "trace/beauty/eventfd.c"
557 #include "trace/beauty/flock.c"
558 #include "trace/beauty/futex_op.c"
559 #include "trace/beauty/mmap.c"
560 #include "trace/beauty/mode_t.c"
561 #include "trace/beauty/msg_flags.c"
562 #include "trace/beauty/open_flags.c"
563 #include "trace/beauty/perf_event_open.c"
564 #include "trace/beauty/pid.c"
565 #include "trace/beauty/sched_policy.c"
566 #include "trace/beauty/seccomp.c"
567 #include "trace/beauty/signum.c"
568 #include "trace/beauty/socket_type.c"
569 #include "trace/beauty/waitid_options.c"
571 static struct syscall_fmt {
574 size_t (*arg_scnprintf[6])(char *bf, size_t size, struct syscall_arg *arg);
581 { .name = "access", .errmsg = true,
582 .arg_scnprintf = { [1] = SCA_ACCMODE, /* mode */ }, },
583 { .name = "arch_prctl", .errmsg = true, .alias = "prctl", },
584 { .name = "bpf", .errmsg = true, STRARRAY(0, cmd, bpf_cmd), },
585 { .name = "brk", .hexret = true,
586 .arg_scnprintf = { [0] = SCA_HEX, /* brk */ }, },
587 { .name = "chdir", .errmsg = true, },
588 { .name = "chmod", .errmsg = true, },
589 { .name = "chroot", .errmsg = true, },
590 { .name = "clock_gettime", .errmsg = true, STRARRAY(0, clk_id, clockid), },
591 { .name = "clone", .errpid = true, },
592 { .name = "close", .errmsg = true,
593 .arg_scnprintf = { [0] = SCA_CLOSE_FD, /* fd */ }, },
594 { .name = "connect", .errmsg = true, },
595 { .name = "creat", .errmsg = true, },
596 { .name = "dup", .errmsg = true, },
597 { .name = "dup2", .errmsg = true, },
598 { .name = "dup3", .errmsg = true, },
599 { .name = "epoll_ctl", .errmsg = true, STRARRAY(1, op, epoll_ctl_ops), },
600 { .name = "eventfd2", .errmsg = true,
601 .arg_scnprintf = { [1] = SCA_EFD_FLAGS, /* flags */ }, },
602 { .name = "faccessat", .errmsg = true, },
603 { .name = "fadvise64", .errmsg = true, },
604 { .name = "fallocate", .errmsg = true, },
605 { .name = "fchdir", .errmsg = true, },
606 { .name = "fchmod", .errmsg = true, },
607 { .name = "fchmodat", .errmsg = true,
608 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
609 { .name = "fchown", .errmsg = true, },
610 { .name = "fchownat", .errmsg = true,
611 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
612 { .name = "fcntl", .errmsg = true,
613 .arg_scnprintf = { [1] = SCA_STRARRAY, /* cmd */ },
614 .arg_parm = { [1] = &strarray__fcntl_cmds, /* cmd */ }, },
615 { .name = "fdatasync", .errmsg = true, },
616 { .name = "flock", .errmsg = true,
617 .arg_scnprintf = { [1] = SCA_FLOCK, /* cmd */ }, },
618 { .name = "fsetxattr", .errmsg = true, },
619 { .name = "fstat", .errmsg = true, .alias = "newfstat", },
620 { .name = "fstatat", .errmsg = true, .alias = "newfstatat", },
621 { .name = "fstatfs", .errmsg = true, },
622 { .name = "fsync", .errmsg = true, },
623 { .name = "ftruncate", .errmsg = true, },
624 { .name = "futex", .errmsg = true,
625 .arg_scnprintf = { [1] = SCA_FUTEX_OP, /* op */ }, },
626 { .name = "futimesat", .errmsg = true,
627 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
628 { .name = "getdents", .errmsg = true, },
629 { .name = "getdents64", .errmsg = true, },
630 { .name = "getitimer", .errmsg = true, STRARRAY(0, which, itimers), },
631 { .name = "getpid", .errpid = true, },
632 { .name = "getpgid", .errpid = true, },
633 { .name = "getppid", .errpid = true, },
634 { .name = "getrandom", .errmsg = true,
635 .arg_scnprintf = { [2] = SCA_GETRANDOM_FLAGS, /* flags */ }, },
636 { .name = "getrlimit", .errmsg = true, STRARRAY(0, resource, rlimit_resources), },
637 { .name = "getxattr", .errmsg = true, },
638 { .name = "inotify_add_watch", .errmsg = true, },
639 { .name = "ioctl", .errmsg = true,
641 #if defined(__i386__) || defined(__x86_64__)
643 * FIXME: Make this available to all arches.
645 [1] = SCA_STRHEXARRAY, /* cmd */
646 [2] = SCA_HEX, /* arg */ },
647 .arg_parm = { [1] = &strarray__tioctls, /* cmd */ }, },
649 [2] = SCA_HEX, /* arg */ }, },
651 { .name = "keyctl", .errmsg = true, STRARRAY(0, option, keyctl_options), },
652 { .name = "kill", .errmsg = true,
653 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
654 { .name = "lchown", .errmsg = true, },
655 { .name = "lgetxattr", .errmsg = true, },
656 { .name = "linkat", .errmsg = true,
657 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
658 { .name = "listxattr", .errmsg = true, },
659 { .name = "llistxattr", .errmsg = true, },
660 { .name = "lremovexattr", .errmsg = true, },
661 { .name = "lseek", .errmsg = true,
662 .arg_scnprintf = { [2] = SCA_STRARRAY, /* whence */ },
663 .arg_parm = { [2] = &strarray__whences, /* whence */ }, },
664 { .name = "lsetxattr", .errmsg = true, },
665 { .name = "lstat", .errmsg = true, .alias = "newlstat", },
666 { .name = "lsxattr", .errmsg = true, },
667 { .name = "madvise", .errmsg = true,
668 .arg_scnprintf = { [0] = SCA_HEX, /* start */
669 [2] = SCA_MADV_BHV, /* behavior */ }, },
670 { .name = "mkdir", .errmsg = true, },
671 { .name = "mkdirat", .errmsg = true,
672 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
673 { .name = "mknod", .errmsg = true, },
674 { .name = "mknodat", .errmsg = true,
675 .arg_scnprintf = { [0] = SCA_FDAT, /* fd */ }, },
676 { .name = "mlock", .errmsg = true,
677 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
678 { .name = "mlockall", .errmsg = true,
679 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
680 { .name = "mmap", .hexret = true,
681 .arg_scnprintf = { [0] = SCA_HEX, /* addr */
682 [2] = SCA_MMAP_PROT, /* prot */
683 [3] = SCA_MMAP_FLAGS, /* flags */ }, },
684 { .name = "mprotect", .errmsg = true,
685 .arg_scnprintf = { [0] = SCA_HEX, /* start */
686 [2] = SCA_MMAP_PROT, /* prot */ }, },
687 { .name = "mq_unlink", .errmsg = true,
688 .arg_scnprintf = { [0] = SCA_FILENAME, /* u_name */ }, },
689 { .name = "mremap", .hexret = true,
690 .arg_scnprintf = { [0] = SCA_HEX, /* addr */
691 [3] = SCA_MREMAP_FLAGS, /* flags */
692 [4] = SCA_HEX, /* new_addr */ }, },
693 { .name = "munlock", .errmsg = true,
694 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
695 { .name = "munmap", .errmsg = true,
696 .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
697 { .name = "name_to_handle_at", .errmsg = true,
698 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
699 { .name = "newfstatat", .errmsg = true,
700 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
701 { .name = "open", .errmsg = true,
702 .arg_scnprintf = { [1] = SCA_OPEN_FLAGS, /* flags */ }, },
703 { .name = "open_by_handle_at", .errmsg = true,
704 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
705 [2] = SCA_OPEN_FLAGS, /* flags */ }, },
706 { .name = "openat", .errmsg = true,
707 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */
708 [2] = SCA_OPEN_FLAGS, /* flags */ }, },
709 { .name = "perf_event_open", .errmsg = true,
710 .arg_scnprintf = { [2] = SCA_INT, /* cpu */
711 [3] = SCA_FD, /* group_fd */
712 [4] = SCA_PERF_FLAGS, /* flags */ }, },
713 { .name = "pipe2", .errmsg = true,
714 .arg_scnprintf = { [1] = SCA_PIPE_FLAGS, /* flags */ }, },
715 { .name = "poll", .errmsg = true, .timeout = true, },
716 { .name = "ppoll", .errmsg = true, .timeout = true, },
717 { .name = "pread", .errmsg = true, .alias = "pread64", },
718 { .name = "preadv", .errmsg = true, .alias = "pread", },
719 { .name = "prlimit64", .errmsg = true, STRARRAY(1, resource, rlimit_resources), },
720 { .name = "pwrite", .errmsg = true, .alias = "pwrite64", },
721 { .name = "pwritev", .errmsg = true, },
722 { .name = "read", .errmsg = true, },
723 { .name = "readlink", .errmsg = true, },
724 { .name = "readlinkat", .errmsg = true,
725 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
726 { .name = "readv", .errmsg = true, },
727 { .name = "recvfrom", .errmsg = true,
728 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
729 { .name = "recvmmsg", .errmsg = true,
730 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
731 { .name = "recvmsg", .errmsg = true,
732 .arg_scnprintf = { [2] = SCA_MSG_FLAGS, /* flags */ }, },
733 { .name = "removexattr", .errmsg = true, },
734 { .name = "renameat", .errmsg = true,
735 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
736 { .name = "rmdir", .errmsg = true, },
737 { .name = "rt_sigaction", .errmsg = true,
738 .arg_scnprintf = { [0] = SCA_SIGNUM, /* sig */ }, },
739 { .name = "rt_sigprocmask", .errmsg = true, STRARRAY(0, how, sighow), },
740 { .name = "rt_sigqueueinfo", .errmsg = true,
741 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
742 { .name = "rt_tgsigqueueinfo", .errmsg = true,
743 .arg_scnprintf = { [2] = SCA_SIGNUM, /* sig */ }, },
744 { .name = "sched_getattr", .errmsg = true, },
745 { .name = "sched_setattr", .errmsg = true, },
746 { .name = "sched_setscheduler", .errmsg = true,
747 .arg_scnprintf = { [1] = SCA_SCHED_POLICY, /* policy */ }, },
748 { .name = "seccomp", .errmsg = true,
749 .arg_scnprintf = { [0] = SCA_SECCOMP_OP, /* op */
750 [1] = SCA_SECCOMP_FLAGS, /* flags */ }, },
751 { .name = "select", .errmsg = true, .timeout = true, },
752 { .name = "sendmmsg", .errmsg = true,
753 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
754 { .name = "sendmsg", .errmsg = true,
755 .arg_scnprintf = { [2] = SCA_MSG_FLAGS, /* flags */ }, },
756 { .name = "sendto", .errmsg = true,
757 .arg_scnprintf = { [3] = SCA_MSG_FLAGS, /* flags */ }, },
758 { .name = "set_tid_address", .errpid = true, },
759 { .name = "setitimer", .errmsg = true, STRARRAY(0, which, itimers), },
760 { .name = "setpgid", .errmsg = true, },
761 { .name = "setrlimit", .errmsg = true, STRARRAY(0, resource, rlimit_resources), },
762 { .name = "setxattr", .errmsg = true, },
763 { .name = "shutdown", .errmsg = true, },
764 { .name = "socket", .errmsg = true,
765 .arg_scnprintf = { [0] = SCA_STRARRAY, /* family */
766 [1] = SCA_SK_TYPE, /* type */ },
767 .arg_parm = { [0] = &strarray__socket_families, /* family */ }, },
768 { .name = "socketpair", .errmsg = true,
769 .arg_scnprintf = { [0] = SCA_STRARRAY, /* family */
770 [1] = SCA_SK_TYPE, /* type */ },
771 .arg_parm = { [0] = &strarray__socket_families, /* family */ }, },
772 { .name = "stat", .errmsg = true, .alias = "newstat", },
773 { .name = "statfs", .errmsg = true, },
774 { .name = "swapoff", .errmsg = true,
775 .arg_scnprintf = { [0] = SCA_FILENAME, /* specialfile */ }, },
776 { .name = "swapon", .errmsg = true,
777 .arg_scnprintf = { [0] = SCA_FILENAME, /* specialfile */ }, },
778 { .name = "symlinkat", .errmsg = true,
779 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
780 { .name = "tgkill", .errmsg = true,
781 .arg_scnprintf = { [2] = SCA_SIGNUM, /* sig */ }, },
782 { .name = "tkill", .errmsg = true,
783 .arg_scnprintf = { [1] = SCA_SIGNUM, /* sig */ }, },
784 { .name = "truncate", .errmsg = true, },
785 { .name = "uname", .errmsg = true, .alias = "newuname", },
786 { .name = "unlinkat", .errmsg = true,
787 .arg_scnprintf = { [0] = SCA_FDAT, /* dfd */ }, },
788 { .name = "utime", .errmsg = true, },
789 { .name = "utimensat", .errmsg = true,
790 .arg_scnprintf = { [0] = SCA_FDAT, /* dirfd */ }, },
791 { .name = "utimes", .errmsg = true, },
792 { .name = "vmsplice", .errmsg = true, },
793 { .name = "wait4", .errpid = true,
794 .arg_scnprintf = { [2] = SCA_WAITID_OPTIONS, /* options */ }, },
795 { .name = "waitid", .errpid = true,
796 .arg_scnprintf = { [3] = SCA_WAITID_OPTIONS, /* options */ }, },
797 { .name = "write", .errmsg = true, },
798 { .name = "writev", .errmsg = true, },
801 static int syscall_fmt__cmp(const void *name, const void *fmtp)
803 const struct syscall_fmt *fmt = fmtp;
804 return strcmp(name, fmt->name);
807 static struct syscall_fmt *syscall_fmt__find(const char *name)
809 const int nmemb = ARRAY_SIZE(syscall_fmts);
810 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
814 struct event_format *tp_format;
816 struct format_field *args;
819 struct syscall_fmt *fmt;
820 size_t (**arg_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
824 static size_t fprintf_duration(unsigned long t, FILE *fp)
826 double duration = (double)t / NSEC_PER_MSEC;
827 size_t printed = fprintf(fp, "(");
830 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
831 else if (duration >= 0.01)
832 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
834 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
835 return printed + fprintf(fp, "): ");
839 * filename.ptr: The filename char pointer that will be vfs_getname'd
840 * filename.entry_str_pos: Where to insert the string translated from
841 * filename.ptr by the vfs_getname tracepoint/kprobe.
843 struct thread_trace {
846 unsigned long nr_events;
847 unsigned long pfmaj, pfmin;
852 short int entry_str_pos;
854 unsigned int namelen;
862 struct intlist *syscall_stats;
865 static struct thread_trace *thread_trace__new(void)
867 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
870 ttrace->paths.max = -1;
872 ttrace->syscall_stats = intlist__new(NULL);
877 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
879 struct thread_trace *ttrace;
884 if (thread__priv(thread) == NULL)
885 thread__set_priv(thread, thread_trace__new());
887 if (thread__priv(thread) == NULL)
890 ttrace = thread__priv(thread);
895 color_fprintf(fp, PERF_COLOR_RED,
896 "WARNING: not enough memory, dropping samples!\n");
900 #define TRACE_PFMAJ (1 << 0)
901 #define TRACE_PFMIN (1 << 1)
903 static const size_t trace__entry_str_size = 2048;
905 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
907 struct thread_trace *ttrace = thread__priv(thread);
909 if (fd > ttrace->paths.max) {
910 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
915 if (ttrace->paths.max != -1) {
916 memset(npath + ttrace->paths.max + 1, 0,
917 (fd - ttrace->paths.max) * sizeof(char *));
919 memset(npath, 0, (fd + 1) * sizeof(char *));
922 ttrace->paths.table = npath;
923 ttrace->paths.max = fd;
926 ttrace->paths.table[fd] = strdup(pathname);
928 return ttrace->paths.table[fd] != NULL ? 0 : -1;
931 static int thread__read_fd_path(struct thread *thread, int fd)
933 char linkname[PATH_MAX], pathname[PATH_MAX];
937 if (thread->pid_ == thread->tid) {
938 scnprintf(linkname, sizeof(linkname),
939 "/proc/%d/fd/%d", thread->pid_, fd);
941 scnprintf(linkname, sizeof(linkname),
942 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
945 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
948 ret = readlink(linkname, pathname, sizeof(pathname));
950 if (ret < 0 || ret > st.st_size)
953 pathname[ret] = '\0';
954 return trace__set_fd_pathname(thread, fd, pathname);
957 static const char *thread__fd_path(struct thread *thread, int fd,
960 struct thread_trace *ttrace = thread__priv(thread);
968 if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
971 ++trace->stats.proc_getname;
972 if (thread__read_fd_path(thread, fd))
976 return ttrace->paths.table[fd];
979 static size_t syscall_arg__scnprintf_fd(char *bf, size_t size,
980 struct syscall_arg *arg)
983 size_t printed = scnprintf(bf, size, "%d", fd);
984 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
987 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
992 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
993 struct syscall_arg *arg)
996 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
997 struct thread_trace *ttrace = thread__priv(arg->thread);
999 if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1000 zfree(&ttrace->paths.table[fd]);
1005 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1008 struct thread_trace *ttrace = thread__priv(thread);
1010 ttrace->filename.ptr = ptr;
1011 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1014 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1015 struct syscall_arg *arg)
1017 unsigned long ptr = arg->val;
1019 if (!arg->trace->vfs_getname)
1020 return scnprintf(bf, size, "%#x", ptr);
1022 thread__set_filename_pos(arg->thread, bf, ptr);
1026 static bool trace__filter_duration(struct trace *trace, double t)
1028 return t < (trace->duration_filter * NSEC_PER_MSEC);
1031 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1033 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1035 return fprintf(fp, "%10.3f ", ts);
1038 static bool done = false;
1039 static bool interrupted = false;
1041 static void sig_handler(int sig)
1044 interrupted = sig == SIGINT;
1047 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1048 u64 duration, u64 tstamp, FILE *fp)
1050 size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
1051 printed += fprintf_duration(duration, fp);
1053 if (trace->multiple_threads) {
1054 if (trace->show_comm)
1055 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1056 printed += fprintf(fp, "%d ", thread->tid);
1062 static int trace__process_event(struct trace *trace, struct machine *machine,
1063 union perf_event *event, struct perf_sample *sample)
1067 switch (event->header.type) {
1068 case PERF_RECORD_LOST:
1069 color_fprintf(trace->output, PERF_COLOR_RED,
1070 "LOST %" PRIu64 " events!\n", event->lost.lost);
1071 ret = machine__process_lost_event(machine, event, sample);
1074 ret = machine__process_event(machine, event, sample);
1081 static int trace__tool_process(struct perf_tool *tool,
1082 union perf_event *event,
1083 struct perf_sample *sample,
1084 struct machine *machine)
1086 struct trace *trace = container_of(tool, struct trace, tool);
1087 return trace__process_event(trace, machine, event, sample);
1090 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1092 struct machine *machine = vmachine;
1094 if (machine->kptr_restrict_warned)
1097 if (symbol_conf.kptr_restrict) {
1098 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1099 "Check /proc/sys/kernel/kptr_restrict.\n\n"
1100 "Kernel samples will not be resolved.\n");
1101 machine->kptr_restrict_warned = true;
1105 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1108 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1110 int err = symbol__init(NULL);
1115 trace->host = machine__new_host();
1116 if (trace->host == NULL)
1119 if (trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr) < 0)
1122 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1123 evlist->threads, trace__tool_process, false,
1124 trace->opts.proc_map_timeout);
1131 static int syscall__set_arg_fmts(struct syscall *sc)
1133 struct format_field *field;
1136 sc->arg_scnprintf = calloc(sc->nr_args, sizeof(void *));
1137 if (sc->arg_scnprintf == NULL)
1141 sc->arg_parm = sc->fmt->arg_parm;
1143 for (field = sc->args; field; field = field->next) {
1144 if (sc->fmt && sc->fmt->arg_scnprintf[idx])
1145 sc->arg_scnprintf[idx] = sc->fmt->arg_scnprintf[idx];
1146 else if (strcmp(field->type, "const char *") == 0 &&
1147 (strcmp(field->name, "filename") == 0 ||
1148 strcmp(field->name, "path") == 0 ||
1149 strcmp(field->name, "pathname") == 0))
1150 sc->arg_scnprintf[idx] = SCA_FILENAME;
1151 else if (field->flags & FIELD_IS_POINTER)
1152 sc->arg_scnprintf[idx] = syscall_arg__scnprintf_hex;
1153 else if (strcmp(field->type, "pid_t") == 0)
1154 sc->arg_scnprintf[idx] = SCA_PID;
1155 else if (strcmp(field->type, "umode_t") == 0)
1156 sc->arg_scnprintf[idx] = SCA_MODE_T;
1157 else if ((strcmp(field->type, "int") == 0 ||
1158 strcmp(field->type, "unsigned int") == 0 ||
1159 strcmp(field->type, "long") == 0) &&
1160 (len = strlen(field->name)) >= 2 &&
1161 strcmp(field->name + len - 2, "fd") == 0) {
1163 * /sys/kernel/tracing/events/syscalls/sys_enter*
1164 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1169 sc->arg_scnprintf[idx] = SCA_FD;
1177 static int trace__read_syscall_info(struct trace *trace, int id)
1181 const char *name = syscalltbl__name(trace->sctbl, id);
1186 if (id > trace->syscalls.max) {
1187 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1189 if (nsyscalls == NULL)
1192 if (trace->syscalls.max != -1) {
1193 memset(nsyscalls + trace->syscalls.max + 1, 0,
1194 (id - trace->syscalls.max) * sizeof(*sc));
1196 memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1199 trace->syscalls.table = nsyscalls;
1200 trace->syscalls.max = id;
1203 sc = trace->syscalls.table + id;
1206 sc->fmt = syscall_fmt__find(sc->name);
1208 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1209 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1211 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1212 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1213 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1216 if (IS_ERR(sc->tp_format))
1219 sc->args = sc->tp_format->format.fields;
1220 sc->nr_args = sc->tp_format->format.nr_fields;
1222 * We need to check and discard the first variable '__syscall_nr'
1223 * or 'nr' that mean the syscall number. It is needless here.
1224 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1226 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1227 sc->args = sc->args->next;
1231 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1233 return syscall__set_arg_fmts(sc);
1236 static int trace__validate_ev_qualifier(struct trace *trace)
1239 struct str_node *pos;
1241 trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1242 trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1243 sizeof(trace->ev_qualifier_ids.entries[0]));
1245 if (trace->ev_qualifier_ids.entries == NULL) {
1246 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1254 strlist__for_each_entry(pos, trace->ev_qualifier) {
1255 const char *sc = pos->s;
1256 int id = syscalltbl__id(trace->sctbl, sc);
1260 fputs("Error:\tInvalid syscall ", trace->output);
1263 fputs(", ", trace->output);
1266 fputs(sc, trace->output);
1269 trace->ev_qualifier_ids.entries[i++] = id;
1273 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1274 "\nHint:\tand: 'man syscalls'\n", trace->output);
1275 zfree(&trace->ev_qualifier_ids.entries);
1276 trace->ev_qualifier_ids.nr = 0;
1283 * args is to be interpreted as a series of longs but we need to handle
1284 * 8-byte unaligned accesses. args points to raw_data within the event
1285 * and raw_data is guaranteed to be 8-byte unaligned because it is
1286 * preceded by raw_size which is a u32. So we need to copy args to a temp
1287 * variable to read it. Most notably this avoids extended load instructions
1288 * on unaligned addresses
1291 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1292 unsigned char *args, struct trace *trace,
1293 struct thread *thread)
1299 if (sc->args != NULL) {
1300 struct format_field *field;
1302 struct syscall_arg arg = {
1309 for (field = sc->args; field;
1310 field = field->next, ++arg.idx, bit <<= 1) {
1314 /* special care for unaligned accesses */
1315 p = args + sizeof(unsigned long) * arg.idx;
1316 memcpy(&val, p, sizeof(val));
1319 * Suppress this argument if its value is zero and
1320 * and we don't have a string associated in an
1324 !(sc->arg_scnprintf &&
1325 sc->arg_scnprintf[arg.idx] == SCA_STRARRAY &&
1326 sc->arg_parm[arg.idx]))
1329 printed += scnprintf(bf + printed, size - printed,
1330 "%s%s: ", printed ? ", " : "", field->name);
1331 if (sc->arg_scnprintf && sc->arg_scnprintf[arg.idx]) {
1334 arg.parm = sc->arg_parm[arg.idx];
1335 printed += sc->arg_scnprintf[arg.idx](bf + printed,
1336 size - printed, &arg);
1338 printed += scnprintf(bf + printed, size - printed,
1342 } else if (IS_ERR(sc->tp_format)) {
1344 * If we managed to read the tracepoint /format file, then we
1345 * may end up not having any args, like with gettid(), so only
1346 * print the raw args when we didn't manage to read it.
1351 /* special care for unaligned accesses */
1352 p = args + sizeof(unsigned long) * i;
1353 memcpy(&val, p, sizeof(val));
1354 printed += scnprintf(bf + printed, size - printed,
1356 printed ? ", " : "", i, val);
1364 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1365 union perf_event *event,
1366 struct perf_sample *sample);
1368 static struct syscall *trace__syscall_info(struct trace *trace,
1369 struct perf_evsel *evsel, int id)
1375 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1376 * before that, leaving at a higher verbosity level till that is
1377 * explained. Reproduced with plain ftrace with:
1379 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1380 * grep "NR -1 " /t/trace_pipe
1382 * After generating some load on the machine.
1386 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1387 id, perf_evsel__name(evsel), ++n);
1392 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1393 trace__read_syscall_info(trace, id))
1396 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1399 return &trace->syscalls.table[id];
1403 fprintf(trace->output, "Problems reading syscall %d", id);
1404 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1405 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1406 fputs(" information\n", trace->output);
1411 static void thread__update_stats(struct thread_trace *ttrace,
1412 int id, struct perf_sample *sample)
1414 struct int_node *inode;
1415 struct stats *stats;
1418 inode = intlist__findnew(ttrace->syscall_stats, id);
1422 stats = inode->priv;
1423 if (stats == NULL) {
1424 stats = malloc(sizeof(struct stats));
1428 inode->priv = stats;
1431 if (ttrace->entry_time && sample->time > ttrace->entry_time)
1432 duration = sample->time - ttrace->entry_time;
1434 update_stats(stats, duration);
1437 static int trace__printf_interrupted_entry(struct trace *trace, struct perf_sample *sample)
1439 struct thread_trace *ttrace;
1443 if (trace->current == NULL)
1446 ttrace = thread__priv(trace->current);
1448 if (!ttrace->entry_pending)
1451 duration = sample->time - ttrace->entry_time;
1453 printed = trace__fprintf_entry_head(trace, trace->current, duration, ttrace->entry_time, trace->output);
1454 printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str);
1455 ttrace->entry_pending = false;
1460 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1461 union perf_event *event __maybe_unused,
1462 struct perf_sample *sample)
1467 struct thread *thread;
1468 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1469 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1470 struct thread_trace *ttrace;
1475 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1476 ttrace = thread__trace(thread, trace->output);
1480 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1482 if (ttrace->entry_str == NULL) {
1483 ttrace->entry_str = malloc(trace__entry_str_size);
1484 if (!ttrace->entry_str)
1488 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1489 trace__printf_interrupted_entry(trace, sample);
1491 ttrace->entry_time = sample->time;
1492 msg = ttrace->entry_str;
1493 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1495 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1496 args, trace, thread);
1499 if (!(trace->duration_filter || trace->summary_only || trace->min_stack)) {
1500 trace__fprintf_entry_head(trace, thread, 1, ttrace->entry_time, trace->output);
1501 fprintf(trace->output, "%-70s)\n", ttrace->entry_str);
1504 ttrace->entry_pending = true;
1505 /* See trace__vfs_getname & trace__sys_exit */
1506 ttrace->filename.pending_open = false;
1509 if (trace->current != thread) {
1510 thread__put(trace->current);
1511 trace->current = thread__get(thread);
1515 thread__put(thread);
1519 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1520 struct perf_sample *sample,
1521 struct callchain_cursor *cursor)
1523 struct addr_location al;
1525 if (machine__resolve(trace->host, &al, sample) < 0 ||
1526 thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, trace->max_stack))
1532 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1534 /* TODO: user-configurable print_opts */
1535 const unsigned int print_opts = EVSEL__PRINT_SYM |
1537 EVSEL__PRINT_UNKNOWN_AS_ADDR;
1539 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1542 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1543 union perf_event *event __maybe_unused,
1544 struct perf_sample *sample)
1548 struct thread *thread;
1549 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0;
1550 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1551 struct thread_trace *ttrace;
1556 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1557 ttrace = thread__trace(thread, trace->output);
1562 thread__update_stats(ttrace, id, sample);
1564 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1566 if (id == trace->open_id && ret >= 0 && ttrace->filename.pending_open) {
1567 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1568 ttrace->filename.pending_open = false;
1569 ++trace->stats.vfs_getname;
1572 if (ttrace->entry_time) {
1573 duration = sample->time - ttrace->entry_time;
1574 if (trace__filter_duration(trace, duration))
1576 } else if (trace->duration_filter)
1579 if (sample->callchain) {
1580 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1581 if (callchain_ret == 0) {
1582 if (callchain_cursor.nr < trace->min_stack)
1588 if (trace->summary_only)
1591 trace__fprintf_entry_head(trace, thread, duration, ttrace->entry_time, trace->output);
1593 if (ttrace->entry_pending) {
1594 fprintf(trace->output, "%-70s", ttrace->entry_str);
1596 fprintf(trace->output, " ... [");
1597 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
1598 fprintf(trace->output, "]: %s()", sc->name);
1601 if (sc->fmt == NULL) {
1603 fprintf(trace->output, ") = %ld", ret);
1604 } else if (ret < 0 && (sc->fmt->errmsg || sc->fmt->errpid)) {
1605 char bf[STRERR_BUFSIZE];
1606 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
1607 *e = audit_errno_to_name(-ret);
1609 fprintf(trace->output, ") = -1 %s %s", e, emsg);
1610 } else if (ret == 0 && sc->fmt->timeout)
1611 fprintf(trace->output, ") = 0 Timeout");
1612 else if (sc->fmt->hexret)
1613 fprintf(trace->output, ") = %#lx", ret);
1614 else if (sc->fmt->errpid) {
1615 struct thread *child = machine__find_thread(trace->host, ret, ret);
1617 if (child != NULL) {
1618 fprintf(trace->output, ") = %ld", ret);
1619 if (child->comm_set)
1620 fprintf(trace->output, " (%s)", thread__comm_str(child));
1626 fputc('\n', trace->output);
1628 if (callchain_ret > 0)
1629 trace__fprintf_callchain(trace, sample);
1630 else if (callchain_ret < 0)
1631 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1633 ttrace->entry_pending = false;
1636 thread__put(thread);
1640 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
1641 union perf_event *event __maybe_unused,
1642 struct perf_sample *sample)
1644 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1645 struct thread_trace *ttrace;
1646 size_t filename_len, entry_str_len, to_move;
1647 ssize_t remaining_space;
1649 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
1654 ttrace = thread__priv(thread);
1658 filename_len = strlen(filename);
1660 if (ttrace->filename.namelen < filename_len) {
1661 char *f = realloc(ttrace->filename.name, filename_len + 1);
1666 ttrace->filename.namelen = filename_len;
1667 ttrace->filename.name = f;
1670 strcpy(ttrace->filename.name, filename);
1671 ttrace->filename.pending_open = true;
1673 if (!ttrace->filename.ptr)
1676 entry_str_len = strlen(ttrace->entry_str);
1677 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
1678 if (remaining_space <= 0)
1681 if (filename_len > (size_t)remaining_space) {
1682 filename += filename_len - remaining_space;
1683 filename_len = remaining_space;
1686 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
1687 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
1688 memmove(pos + filename_len, pos, to_move);
1689 memcpy(pos, filename, filename_len);
1691 ttrace->filename.ptr = 0;
1692 ttrace->filename.entry_str_pos = 0;
1697 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
1698 union perf_event *event __maybe_unused,
1699 struct perf_sample *sample)
1701 u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
1702 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
1703 struct thread *thread = machine__findnew_thread(trace->host,
1706 struct thread_trace *ttrace = thread__trace(thread, trace->output);
1711 ttrace->runtime_ms += runtime_ms;
1712 trace->runtime_ms += runtime_ms;
1713 thread__put(thread);
1717 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
1719 perf_evsel__strval(evsel, sample, "comm"),
1720 (pid_t)perf_evsel__intval(evsel, sample, "pid"),
1722 perf_evsel__intval(evsel, sample, "vruntime"));
1723 thread__put(thread);
1727 static void bpf_output__printer(enum binary_printer_ops op,
1728 unsigned int val, void *extra)
1730 FILE *output = extra;
1731 unsigned char ch = (unsigned char)val;
1734 case BINARY_PRINT_CHAR_DATA:
1735 fprintf(output, "%c", isprint(ch) ? ch : '.');
1737 case BINARY_PRINT_DATA_BEGIN:
1738 case BINARY_PRINT_LINE_BEGIN:
1739 case BINARY_PRINT_ADDR:
1740 case BINARY_PRINT_NUM_DATA:
1741 case BINARY_PRINT_NUM_PAD:
1742 case BINARY_PRINT_SEP:
1743 case BINARY_PRINT_CHAR_PAD:
1744 case BINARY_PRINT_LINE_END:
1745 case BINARY_PRINT_DATA_END:
1751 static void bpf_output__fprintf(struct trace *trace,
1752 struct perf_sample *sample)
1754 print_binary(sample->raw_data, sample->raw_size, 8,
1755 bpf_output__printer, trace->output);
1758 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
1759 union perf_event *event __maybe_unused,
1760 struct perf_sample *sample)
1762 int callchain_ret = 0;
1764 if (sample->callchain) {
1765 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1766 if (callchain_ret == 0) {
1767 if (callchain_cursor.nr < trace->min_stack)
1773 trace__printf_interrupted_entry(trace, sample);
1774 trace__fprintf_tstamp(trace, sample->time, trace->output);
1776 if (trace->trace_syscalls)
1777 fprintf(trace->output, "( ): ");
1779 fprintf(trace->output, "%s:", evsel->name);
1781 if (perf_evsel__is_bpf_output(evsel)) {
1782 bpf_output__fprintf(trace, sample);
1783 } else if (evsel->tp_format) {
1784 event_format__fprintf(evsel->tp_format, sample->cpu,
1785 sample->raw_data, sample->raw_size,
1789 fprintf(trace->output, ")\n");
1791 if (callchain_ret > 0)
1792 trace__fprintf_callchain(trace, sample);
1793 else if (callchain_ret < 0)
1794 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1799 static void print_location(FILE *f, struct perf_sample *sample,
1800 struct addr_location *al,
1801 bool print_dso, bool print_sym)
1804 if ((verbose > 0 || print_dso) && al->map)
1805 fprintf(f, "%s@", al->map->dso->long_name);
1807 if ((verbose > 0 || print_sym) && al->sym)
1808 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
1809 al->addr - al->sym->start);
1811 fprintf(f, "0x%" PRIx64, al->addr);
1813 fprintf(f, "0x%" PRIx64, sample->addr);
1816 static int trace__pgfault(struct trace *trace,
1817 struct perf_evsel *evsel,
1818 union perf_event *event __maybe_unused,
1819 struct perf_sample *sample)
1821 struct thread *thread;
1822 struct addr_location al;
1823 char map_type = 'd';
1824 struct thread_trace *ttrace;
1826 int callchain_ret = 0;
1828 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1830 if (sample->callchain) {
1831 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1832 if (callchain_ret == 0) {
1833 if (callchain_cursor.nr < trace->min_stack)
1839 ttrace = thread__trace(thread, trace->output);
1843 if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
1848 if (trace->summary_only)
1851 thread__find_addr_location(thread, sample->cpumode, MAP__FUNCTION,
1854 trace__fprintf_entry_head(trace, thread, 0, sample->time, trace->output);
1856 fprintf(trace->output, "%sfault [",
1857 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
1860 print_location(trace->output, sample, &al, false, true);
1862 fprintf(trace->output, "] => ");
1864 thread__find_addr_location(thread, sample->cpumode, MAP__VARIABLE,
1868 thread__find_addr_location(thread, sample->cpumode,
1869 MAP__FUNCTION, sample->addr, &al);
1877 print_location(trace->output, sample, &al, true, false);
1879 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
1881 if (callchain_ret > 0)
1882 trace__fprintf_callchain(trace, sample);
1883 else if (callchain_ret < 0)
1884 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1888 thread__put(thread);
1892 static void trace__set_base_time(struct trace *trace,
1893 struct perf_evsel *evsel,
1894 struct perf_sample *sample)
1897 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
1898 * and don't use sample->time unconditionally, we may end up having
1899 * some other event in the future without PERF_SAMPLE_TIME for good
1900 * reason, i.e. we may not be interested in its timestamps, just in
1901 * it taking place, picking some piece of information when it
1902 * appears in our event stream (vfs_getname comes to mind).
1904 if (trace->base_time == 0 && !trace->full_time &&
1905 (evsel->attr.sample_type & PERF_SAMPLE_TIME))
1906 trace->base_time = sample->time;
1909 static int trace__process_sample(struct perf_tool *tool,
1910 union perf_event *event,
1911 struct perf_sample *sample,
1912 struct perf_evsel *evsel,
1913 struct machine *machine __maybe_unused)
1915 struct trace *trace = container_of(tool, struct trace, tool);
1916 struct thread *thread;
1919 tracepoint_handler handler = evsel->handler;
1921 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1922 if (thread && thread__is_filtered(thread))
1925 trace__set_base_time(trace, evsel, sample);
1929 handler(trace, evsel, event, sample);
1935 static int trace__record(struct trace *trace, int argc, const char **argv)
1937 unsigned int rec_argc, i, j;
1938 const char **rec_argv;
1939 const char * const record_args[] = {
1946 const char * const sc_args[] = { "-e", };
1947 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
1948 const char * const majpf_args[] = { "-e", "major-faults" };
1949 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
1950 const char * const minpf_args[] = { "-e", "minor-faults" };
1951 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
1953 /* +1 is for the event string below */
1954 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
1955 majpf_args_nr + minpf_args_nr + argc;
1956 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1958 if (rec_argv == NULL)
1962 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1963 rec_argv[j++] = record_args[i];
1965 if (trace->trace_syscalls) {
1966 for (i = 0; i < sc_args_nr; i++)
1967 rec_argv[j++] = sc_args[i];
1969 /* event string may be different for older kernels - e.g., RHEL6 */
1970 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
1971 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
1972 else if (is_valid_tracepoint("syscalls:sys_enter"))
1973 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
1975 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
1980 if (trace->trace_pgfaults & TRACE_PFMAJ)
1981 for (i = 0; i < majpf_args_nr; i++)
1982 rec_argv[j++] = majpf_args[i];
1984 if (trace->trace_pgfaults & TRACE_PFMIN)
1985 for (i = 0; i < minpf_args_nr; i++)
1986 rec_argv[j++] = minpf_args[i];
1988 for (i = 0; i < (unsigned int)argc; i++)
1989 rec_argv[j++] = argv[i];
1991 return cmd_record(j, rec_argv, NULL);
1994 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
1996 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
1998 struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2003 if (perf_evsel__field(evsel, "pathname") == NULL) {
2004 perf_evsel__delete(evsel);
2008 evsel->handler = trace__vfs_getname;
2009 perf_evlist__add(evlist, evsel);
2013 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2015 struct perf_evsel *evsel;
2016 struct perf_event_attr attr = {
2017 .type = PERF_TYPE_SOFTWARE,
2021 attr.config = config;
2022 attr.sample_period = 1;
2024 event_attr_init(&attr);
2026 evsel = perf_evsel__new(&attr);
2028 evsel->handler = trace__pgfault;
2033 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2035 const u32 type = event->header.type;
2036 struct perf_evsel *evsel;
2038 if (type != PERF_RECORD_SAMPLE) {
2039 trace__process_event(trace, trace->host, event, sample);
2043 evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2044 if (evsel == NULL) {
2045 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2049 trace__set_base_time(trace, evsel, sample);
2051 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2052 sample->raw_data == NULL) {
2053 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2054 perf_evsel__name(evsel), sample->tid,
2055 sample->cpu, sample->raw_size);
2057 tracepoint_handler handler = evsel->handler;
2058 handler(trace, evsel, event, sample);
2062 static int trace__add_syscall_newtp(struct trace *trace)
2065 struct perf_evlist *evlist = trace->evlist;
2066 struct perf_evsel *sys_enter, *sys_exit;
2068 sys_enter = perf_evsel__syscall_newtp("sys_enter", trace__sys_enter);
2069 if (sys_enter == NULL)
2072 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2073 goto out_delete_sys_enter;
2075 sys_exit = perf_evsel__syscall_newtp("sys_exit", trace__sys_exit);
2076 if (sys_exit == NULL)
2077 goto out_delete_sys_enter;
2079 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2080 goto out_delete_sys_exit;
2082 perf_evlist__add(evlist, sys_enter);
2083 perf_evlist__add(evlist, sys_exit);
2085 if (callchain_param.enabled && !trace->kernel_syscallchains) {
2087 * We're interested only in the user space callchain
2088 * leading to the syscall, allow overriding that for
2089 * debugging reasons using --kernel_syscall_callchains
2091 sys_exit->attr.exclude_callchain_kernel = 1;
2094 trace->syscalls.events.sys_enter = sys_enter;
2095 trace->syscalls.events.sys_exit = sys_exit;
2101 out_delete_sys_exit:
2102 perf_evsel__delete_priv(sys_exit);
2103 out_delete_sys_enter:
2104 perf_evsel__delete_priv(sys_enter);
2108 static int trace__set_ev_qualifier_filter(struct trace *trace)
2111 struct perf_evsel *sys_exit;
2112 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2113 trace->ev_qualifier_ids.nr,
2114 trace->ev_qualifier_ids.entries);
2119 if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2121 sys_exit = trace->syscalls.events.sys_exit;
2122 err = perf_evsel__append_tp_filter(sys_exit, filter);
2133 static int trace__run(struct trace *trace, int argc, const char **argv)
2135 struct perf_evlist *evlist = trace->evlist;
2136 struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2138 unsigned long before;
2139 const bool forks = argc > 0;
2140 bool draining = false;
2144 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2145 goto out_error_raw_syscalls;
2147 if (trace->trace_syscalls)
2148 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2150 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2151 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2152 if (pgfault_maj == NULL)
2154 perf_evlist__add(evlist, pgfault_maj);
2157 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2158 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2159 if (pgfault_min == NULL)
2161 perf_evlist__add(evlist, pgfault_min);
2165 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2166 trace__sched_stat_runtime))
2167 goto out_error_sched_stat_runtime;
2169 err = perf_evlist__create_maps(evlist, &trace->opts.target);
2171 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2172 goto out_delete_evlist;
2175 err = trace__symbols_init(trace, evlist);
2177 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2178 goto out_delete_evlist;
2181 perf_evlist__config(evlist, &trace->opts, NULL);
2183 if (callchain_param.enabled) {
2184 bool use_identifier = false;
2186 if (trace->syscalls.events.sys_exit) {
2187 perf_evsel__config_callchain(trace->syscalls.events.sys_exit,
2188 &trace->opts, &callchain_param);
2189 use_identifier = true;
2193 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2194 use_identifier = true;
2198 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2199 use_identifier = true;
2202 if (use_identifier) {
2204 * Now we have evsels with different sample_ids, use
2205 * PERF_SAMPLE_IDENTIFIER to map from sample to evsel
2206 * from a fixed position in each ring buffer record.
2208 * As of this the changeset introducing this comment, this
2209 * isn't strictly needed, as the fields that can come before
2210 * PERF_SAMPLE_ID are all used, but we'll probably disable
2211 * some of those for things like copying the payload of
2212 * pointer syscall arguments, and for vfs_getname we don't
2213 * need PERF_SAMPLE_ADDR and PERF_SAMPLE_IP, so do this
2214 * here as a warning we need to use PERF_SAMPLE_IDENTIFIER.
2216 perf_evlist__set_sample_bit(evlist, IDENTIFIER);
2217 perf_evlist__reset_sample_bit(evlist, ID);
2221 signal(SIGCHLD, sig_handler);
2222 signal(SIGINT, sig_handler);
2225 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2228 fprintf(trace->output, "Couldn't run the workload!\n");
2229 goto out_delete_evlist;
2233 err = perf_evlist__open(evlist);
2235 goto out_error_open;
2237 err = bpf__apply_obj_config();
2239 char errbuf[BUFSIZ];
2241 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2242 pr_err("ERROR: Apply config to BPF failed: %s\n",
2244 goto out_error_open;
2248 * Better not use !target__has_task() here because we need to cover the
2249 * case where no threads were specified in the command line, but a
2250 * workload was, and in that case we will fill in the thread_map when
2251 * we fork the workload in perf_evlist__prepare_workload.
2253 if (trace->filter_pids.nr > 0)
2254 err = perf_evlist__set_filter_pids(evlist, trace->filter_pids.nr, trace->filter_pids.entries);
2255 else if (thread_map__pid(evlist->threads, 0) == -1)
2256 err = perf_evlist__set_filter_pid(evlist, getpid());
2261 if (trace->ev_qualifier_ids.nr > 0) {
2262 err = trace__set_ev_qualifier_filter(trace);
2266 pr_debug("event qualifier tracepoint filter: %s\n",
2267 trace->syscalls.events.sys_exit->filter);
2270 err = perf_evlist__apply_filters(evlist, &evsel);
2272 goto out_error_apply_filters;
2274 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages, false);
2276 goto out_error_mmap;
2278 if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2279 perf_evlist__enable(evlist);
2282 perf_evlist__start_workload(evlist);
2284 if (trace->opts.initial_delay) {
2285 usleep(trace->opts.initial_delay * 1000);
2286 perf_evlist__enable(evlist);
2289 trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2290 evlist->threads->nr > 1 ||
2291 perf_evlist__first(evlist)->attr.inherit;
2293 before = trace->nr_events;
2295 for (i = 0; i < evlist->nr_mmaps; i++) {
2296 union perf_event *event;
2298 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
2299 struct perf_sample sample;
2303 err = perf_evlist__parse_sample(evlist, event, &sample);
2305 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2309 trace__handle_event(trace, event, &sample);
2311 perf_evlist__mmap_consume(evlist, i);
2316 if (done && !draining) {
2317 perf_evlist__disable(evlist);
2323 if (trace->nr_events == before) {
2324 int timeout = done ? 100 : -1;
2326 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
2327 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP) == 0)
2337 thread__zput(trace->current);
2339 perf_evlist__disable(evlist);
2343 trace__fprintf_thread_summary(trace, trace->output);
2345 if (trace->show_tool_stats) {
2346 fprintf(trace->output, "Stats:\n "
2347 " vfs_getname : %" PRIu64 "\n"
2348 " proc_getname: %" PRIu64 "\n",
2349 trace->stats.vfs_getname,
2350 trace->stats.proc_getname);
2355 perf_evlist__delete(evlist);
2356 trace->evlist = NULL;
2357 trace->live = false;
2360 char errbuf[BUFSIZ];
2362 out_error_sched_stat_runtime:
2363 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
2366 out_error_raw_syscalls:
2367 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
2371 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
2375 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
2378 fprintf(trace->output, "%s\n", errbuf);
2379 goto out_delete_evlist;
2381 out_error_apply_filters:
2382 fprintf(trace->output,
2383 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
2384 evsel->filter, perf_evsel__name(evsel), errno,
2385 str_error_r(errno, errbuf, sizeof(errbuf)));
2386 goto out_delete_evlist;
2389 fprintf(trace->output, "Not enough memory to run!\n");
2390 goto out_delete_evlist;
2393 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
2394 goto out_delete_evlist;
2397 static int trace__replay(struct trace *trace)
2399 const struct perf_evsel_str_handler handlers[] = {
2400 { "probe:vfs_getname", trace__vfs_getname, },
2402 struct perf_data_file file = {
2404 .mode = PERF_DATA_MODE_READ,
2405 .force = trace->force,
2407 struct perf_session *session;
2408 struct perf_evsel *evsel;
2411 trace->tool.sample = trace__process_sample;
2412 trace->tool.mmap = perf_event__process_mmap;
2413 trace->tool.mmap2 = perf_event__process_mmap2;
2414 trace->tool.comm = perf_event__process_comm;
2415 trace->tool.exit = perf_event__process_exit;
2416 trace->tool.fork = perf_event__process_fork;
2417 trace->tool.attr = perf_event__process_attr;
2418 trace->tool.tracing_data = perf_event__process_tracing_data;
2419 trace->tool.build_id = perf_event__process_build_id;
2421 trace->tool.ordered_events = true;
2422 trace->tool.ordering_requires_timestamps = true;
2424 /* add tid to output */
2425 trace->multiple_threads = true;
2427 session = perf_session__new(&file, false, &trace->tool);
2428 if (session == NULL)
2431 if (trace->opts.target.pid)
2432 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
2434 if (trace->opts.target.tid)
2435 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
2437 if (symbol__init(&session->header.env) < 0)
2440 trace->host = &session->machines.host;
2442 err = perf_session__set_tracepoints_handlers(session, handlers);
2446 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2447 "raw_syscalls:sys_enter");
2448 /* older kernels have syscalls tp versus raw_syscalls */
2450 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2451 "syscalls:sys_enter");
2454 (perf_evsel__init_syscall_tp(evsel, trace__sys_enter) < 0 ||
2455 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
2456 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
2460 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2461 "raw_syscalls:sys_exit");
2463 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2464 "syscalls:sys_exit");
2466 (perf_evsel__init_syscall_tp(evsel, trace__sys_exit) < 0 ||
2467 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
2468 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
2472 evlist__for_each_entry(session->evlist, evsel) {
2473 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
2474 (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
2475 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
2476 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
2477 evsel->handler = trace__pgfault;
2482 err = perf_session__process_events(session);
2484 pr_err("Failed to process events, error %d", err);
2486 else if (trace->summary)
2487 trace__fprintf_thread_summary(trace, trace->output);
2490 perf_session__delete(session);
2495 static size_t trace__fprintf_threads_header(FILE *fp)
2499 printed = fprintf(fp, "\n Summary of events:\n\n");
2504 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
2505 struct stats *stats;
2510 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
2511 struct stats *stats = source->priv;
2513 entry->syscall = source->i;
2514 entry->stats = stats;
2515 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
2518 static size_t thread__dump_stats(struct thread_trace *ttrace,
2519 struct trace *trace, FILE *fp)
2524 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
2526 if (syscall_stats == NULL)
2529 printed += fprintf(fp, "\n");
2531 printed += fprintf(fp, " syscall calls total min avg max stddev\n");
2532 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
2533 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n");
2535 resort_rb__for_each_entry(nd, syscall_stats) {
2536 struct stats *stats = syscall_stats_entry->stats;
2538 double min = (double)(stats->min) / NSEC_PER_MSEC;
2539 double max = (double)(stats->max) / NSEC_PER_MSEC;
2540 double avg = avg_stats(stats);
2542 u64 n = (u64) stats->n;
2544 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
2545 avg /= NSEC_PER_MSEC;
2547 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
2548 printed += fprintf(fp, " %-15s", sc->name);
2549 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
2550 n, syscall_stats_entry->msecs, min, avg);
2551 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
2555 resort_rb__delete(syscall_stats);
2556 printed += fprintf(fp, "\n\n");
2561 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
2564 struct thread_trace *ttrace = thread__priv(thread);
2570 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
2572 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
2573 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
2574 printed += fprintf(fp, "%.1f%%", ratio);
2576 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
2578 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
2580 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
2581 else if (fputc('\n', fp) != EOF)
2584 printed += thread__dump_stats(ttrace, trace, fp);
2589 static unsigned long thread__nr_events(struct thread_trace *ttrace)
2591 return ttrace ? ttrace->nr_events : 0;
2594 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
2595 struct thread *thread;
2598 entry->thread = rb_entry(nd, struct thread, rb_node);
2601 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
2603 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host);
2604 size_t printed = trace__fprintf_threads_header(fp);
2607 if (threads == NULL) {
2608 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
2612 resort_rb__for_each_entry(nd, threads)
2613 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
2615 resort_rb__delete(threads);
2620 static int trace__set_duration(const struct option *opt, const char *str,
2621 int unset __maybe_unused)
2623 struct trace *trace = opt->value;
2625 trace->duration_filter = atof(str);
2629 static int trace__set_filter_pids(const struct option *opt, const char *str,
2630 int unset __maybe_unused)
2634 struct trace *trace = opt->value;
2636 * FIXME: introduce a intarray class, plain parse csv and create a
2637 * { int nr, int entries[] } struct...
2639 struct intlist *list = intlist__new(str);
2644 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
2645 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
2647 if (trace->filter_pids.entries == NULL)
2650 trace->filter_pids.entries[0] = getpid();
2652 for (i = 1; i < trace->filter_pids.nr; ++i)
2653 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
2655 intlist__delete(list);
2661 static int trace__open_output(struct trace *trace, const char *filename)
2665 if (!stat(filename, &st) && st.st_size) {
2666 char oldname[PATH_MAX];
2668 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
2670 rename(filename, oldname);
2673 trace->output = fopen(filename, "w");
2675 return trace->output == NULL ? -errno : 0;
2678 static int parse_pagefaults(const struct option *opt, const char *str,
2679 int unset __maybe_unused)
2681 int *trace_pgfaults = opt->value;
2683 if (strcmp(str, "all") == 0)
2684 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
2685 else if (strcmp(str, "maj") == 0)
2686 *trace_pgfaults |= TRACE_PFMAJ;
2687 else if (strcmp(str, "min") == 0)
2688 *trace_pgfaults |= TRACE_PFMIN;
2695 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
2697 struct perf_evsel *evsel;
2699 evlist__for_each_entry(evlist, evsel)
2700 evsel->handler = handler;
2704 * XXX: Hackish, just splitting the combined -e+--event (syscalls
2705 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
2706 * existing facilities unchanged (trace->ev_qualifier + parse_options()).
2708 * It'd be better to introduce a parse_options() variant that would return a
2709 * list with the terms it didn't match to an event...
2711 static int trace__parse_events_option(const struct option *opt, const char *str,
2712 int unset __maybe_unused)
2714 struct trace *trace = (struct trace *)opt->value;
2715 const char *s = str;
2716 char *sep = NULL, *lists[2] = { NULL, NULL, };
2717 int len = strlen(str), err = -1, list;
2718 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
2719 char group_name[PATH_MAX];
2721 if (strace_groups_dir == NULL)
2726 trace->not_ev_qualifier = true;
2730 if ((sep = strchr(s, ',')) != NULL)
2734 if (syscalltbl__id(trace->sctbl, s) >= 0) {
2737 path__join(group_name, sizeof(group_name), strace_groups_dir, s);
2738 if (access(group_name, R_OK) == 0)
2743 sprintf(lists[list] + strlen(lists[list]), ",%s", s);
2745 lists[list] = malloc(len);
2746 if (lists[list] == NULL)
2748 strcpy(lists[list], s);
2758 if (lists[1] != NULL) {
2759 struct strlist_config slist_config = {
2760 .dirname = strace_groups_dir,
2763 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
2764 if (trace->ev_qualifier == NULL) {
2765 fputs("Not enough memory to parse event qualifier", trace->output);
2769 if (trace__validate_ev_qualifier(trace))
2776 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
2777 "event selector. use 'perf list' to list available events",
2778 parse_events_option);
2779 err = parse_events_option(&o, lists[0], 0);
2788 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
2790 const char *trace_usage[] = {
2791 "perf trace [<options>] [<command>]",
2792 "perf trace [<options>] -- <command> [<options>]",
2793 "perf trace record [<options>] [<command>]",
2794 "perf trace record [<options>] -- <command> [<options>]",
2797 struct trace trace = {
2806 .user_freq = UINT_MAX,
2807 .user_interval = ULLONG_MAX,
2808 .no_buffering = true,
2809 .mmap_pages = UINT_MAX,
2810 .proc_map_timeout = 500,
2814 .trace_syscalls = true,
2815 .kernel_syscallchains = false,
2816 .max_stack = UINT_MAX,
2818 const char *output_name = NULL;
2819 const struct option trace_options[] = {
2820 OPT_CALLBACK('e', "event", &trace, "event",
2821 "event/syscall selector. use 'perf list' to list available events",
2822 trace__parse_events_option),
2823 OPT_BOOLEAN(0, "comm", &trace.show_comm,
2824 "show the thread COMM next to its id"),
2825 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
2826 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
2827 trace__parse_events_option),
2828 OPT_STRING('o', "output", &output_name, "file", "output file name"),
2829 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
2830 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
2831 "trace events on existing process id"),
2832 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
2833 "trace events on existing thread id"),
2834 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
2835 "pids to filter (by the kernel)", trace__set_filter_pids),
2836 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
2837 "system-wide collection from all CPUs"),
2838 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
2839 "list of cpus to monitor"),
2840 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
2841 "child tasks do not inherit counters"),
2842 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
2843 "number of mmap data pages",
2844 perf_evlist__parse_mmap_pages),
2845 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
2847 OPT_CALLBACK(0, "duration", &trace, "float",
2848 "show only events with duration > N.M ms",
2849 trace__set_duration),
2850 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
2851 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
2852 OPT_BOOLEAN('T', "time", &trace.full_time,
2853 "Show full timestamp, not time relative to first start"),
2854 OPT_BOOLEAN('s', "summary", &trace.summary_only,
2855 "Show only syscall summary with statistics"),
2856 OPT_BOOLEAN('S', "with-summary", &trace.summary,
2857 "Show all syscalls and summary with statistics"),
2858 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
2859 "Trace pagefaults", parse_pagefaults, "maj"),
2860 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
2861 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
2862 OPT_CALLBACK(0, "call-graph", &trace.opts,
2863 "record_mode[,record_size]", record_callchain_help,
2864 &record_parse_callchain_opt),
2865 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
2866 "Show the kernel callchains on the syscall exit path"),
2867 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
2868 "Set the minimum stack depth when parsing the callchain, "
2869 "anything below the specified depth will be ignored."),
2870 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
2871 "Set the maximum stack depth when parsing the callchain, "
2872 "anything beyond the specified depth will be ignored. "
2873 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
2874 OPT_UINTEGER(0, "proc-map-timeout", &trace.opts.proc_map_timeout,
2875 "per thread proc mmap processing timeout in ms"),
2876 OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
2877 "ms to wait before starting measurement after program "
2881 bool __maybe_unused max_stack_user_set = true;
2882 bool mmap_pages_user_set = true;
2883 const char * const trace_subcommands[] = { "record", NULL };
2887 signal(SIGSEGV, sighandler_dump_stack);
2888 signal(SIGFPE, sighandler_dump_stack);
2890 trace.evlist = perf_evlist__new();
2891 trace.sctbl = syscalltbl__new();
2893 if (trace.evlist == NULL || trace.sctbl == NULL) {
2894 pr_err("Not enough memory to run!\n");
2899 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
2900 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
2902 err = bpf__setup_stdout(trace.evlist);
2904 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
2905 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
2911 if (trace.trace_pgfaults) {
2912 trace.opts.sample_address = true;
2913 trace.opts.sample_time = true;
2916 if (trace.opts.mmap_pages == UINT_MAX)
2917 mmap_pages_user_set = false;
2919 if (trace.max_stack == UINT_MAX) {
2920 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl_perf_event_max_stack;
2921 max_stack_user_set = false;
2924 #ifdef HAVE_DWARF_UNWIND_SUPPORT
2925 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled && trace.trace_syscalls)
2926 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
2929 if (callchain_param.enabled) {
2930 if (!mmap_pages_user_set && geteuid() == 0)
2931 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
2933 symbol_conf.use_callchain = true;
2936 if (trace.evlist->nr_entries > 0)
2937 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
2939 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
2940 return trace__record(&trace, argc-1, &argv[1]);
2942 /* summary_only implies summary option, but don't overwrite summary if set */
2943 if (trace.summary_only)
2944 trace.summary = trace.summary_only;
2946 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
2947 trace.evlist->nr_entries == 0 /* Was --events used? */) {
2948 pr_err("Please specify something to trace.\n");
2952 if (!trace.trace_syscalls && trace.ev_qualifier) {
2953 pr_err("The -e option can't be used with --no-syscalls.\n");
2957 if (output_name != NULL) {
2958 err = trace__open_output(&trace, output_name);
2960 perror("failed to create output file");
2965 trace.open_id = syscalltbl__id(trace.sctbl, "open");
2967 err = target__validate(&trace.opts.target);
2969 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
2970 fprintf(trace.output, "%s", bf);
2974 err = target__parse_uid(&trace.opts.target);
2976 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
2977 fprintf(trace.output, "%s", bf);
2981 if (!argc && target__none(&trace.opts.target))
2982 trace.opts.target.system_wide = true;
2985 err = trace__replay(&trace);
2987 err = trace__run(&trace, argc, argv);
2990 if (output_name != NULL)
2991 fclose(trace.output);