1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin 'trace' command:
7 * Display a continuously updated trace of any workload, CPU, specific PID,
8 * system wide, etc. Default format is loosely strace like, but any other
9 * event may be specified using --event.
11 * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
13 * Initially based on the 'trace' prototype by Thomas Gleixner:
15 * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
18 #include <traceevent/event-parse.h>
19 #include <api/fs/tracing_path.h>
21 #include "util/bpf_map.h"
23 #include "util/cgroup.h"
24 #include "util/color.h"
25 #include "util/config.h"
26 #include "util/debug.h"
28 #include "util/event.h"
29 #include "util/evlist.h"
30 #include <subcmd/exec-cmd.h>
31 #include "util/machine.h"
33 #include "util/symbol.h"
34 #include "util/path.h"
35 #include "util/session.h"
36 #include "util/thread.h"
37 #include <subcmd/parse-options.h>
38 #include "util/strlist.h"
39 #include "util/intlist.h"
40 #include "util/thread_map.h"
41 #include "util/stat.h"
42 #include "trace/beauty/beauty.h"
43 #include "trace-event.h"
44 #include "util/parse-events.h"
45 #include "util/bpf-loader.h"
46 #include "callchain.h"
47 #include "print_binary.h"
49 #include "syscalltbl.h"
50 #include "rb_resort.h"
58 #include <linux/err.h>
59 #include <linux/filter.h>
60 #include <linux/kernel.h>
61 #include <linux/random.h>
62 #include <linux/stringify.h>
63 #include <linux/time64.h>
64 #include <linux/zalloc.h>
66 #include <sys/sysmacros.h>
68 #include <linux/ctype.h>
71 # define O_CLOEXEC 02000000
74 #ifndef F_LINUX_SPECIFIC_BASE
75 # define F_LINUX_SPECIFIC_BASE 1024
79 struct perf_tool tool;
80 struct syscalltbl *sctbl;
83 struct syscall *table;
86 struct perf_evsel *sys_enter,
94 struct record_opts opts;
95 struct perf_evlist *evlist;
97 struct thread *current;
98 struct cgroup *cgroup;
101 unsigned long nr_events;
102 unsigned long nr_events_printed;
103 unsigned long max_events;
104 struct strlist *ev_qualifier;
114 double duration_filter;
120 unsigned int max_stack;
121 unsigned int min_stack;
122 int raw_augmented_syscalls_args_size;
123 bool raw_augmented_syscalls;
125 bool not_ev_qualifier;
129 bool multiple_threads;
135 bool show_tool_stats;
137 bool kernel_syscallchains;
143 bool show_string_prefix;
148 struct ordered_events data;
156 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
157 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
161 #define TP_UINT_FIELD(bits) \
162 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
165 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
174 #define TP_UINT_FIELD__SWAPPED(bits) \
175 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
178 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
179 return bswap_##bits(value);\
182 TP_UINT_FIELD__SWAPPED(16);
183 TP_UINT_FIELD__SWAPPED(32);
184 TP_UINT_FIELD__SWAPPED(64);
186 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
188 field->offset = offset;
192 field->integer = tp_field__u8;
195 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
198 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
201 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
210 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
212 return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
215 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
217 return sample->raw_data + field->offset;
220 static int __tp_field__init_ptr(struct tp_field *field, int offset)
222 field->offset = offset;
223 field->pointer = tp_field__ptr;
227 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
229 return __tp_field__init_ptr(field, format_field->offset);
235 struct tp_field args, ret;
239 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
240 struct tp_field *field,
243 struct tep_format_field *format_field = perf_evsel__field(evsel, name);
245 if (format_field == NULL)
248 return tp_field__init_uint(field, format_field, evsel->needs_swap);
251 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
252 ({ struct syscall_tp *sc = evsel->priv;\
253 perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
255 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
256 struct tp_field *field,
259 struct tep_format_field *format_field = perf_evsel__field(evsel, name);
261 if (format_field == NULL)
264 return tp_field__init_ptr(field, format_field);
267 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
268 ({ struct syscall_tp *sc = evsel->priv;\
269 perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
271 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
274 perf_evsel__delete(evsel);
277 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
279 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
281 if (evsel->priv != NULL) {
282 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
283 perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
294 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
296 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
298 if (evsel->priv != NULL) {
299 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
300 if (syscall_id == NULL)
301 syscall_id = perf_evsel__field(tp, "__syscall_nr");
302 if (syscall_id == NULL)
304 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
316 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
318 struct syscall_tp *sc = evsel->priv;
320 return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
323 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
325 struct syscall_tp *sc = evsel->priv;
327 return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
330 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
332 evsel->priv = malloc(sizeof(struct syscall_tp));
333 if (evsel->priv != NULL) {
334 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
337 evsel->handler = handler;
348 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
350 struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
352 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
354 evsel = perf_evsel__newtp("syscalls", direction);
359 if (perf_evsel__init_raw_syscall_tp(evsel, handler))
365 perf_evsel__delete_priv(evsel);
369 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
370 ({ struct syscall_tp *fields = evsel->priv; \
371 fields->name.integer(&fields->name, sample); })
373 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
374 ({ struct syscall_tp *fields = evsel->priv; \
375 fields->name.pointer(&fields->name, sample); })
377 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
379 int idx = val - sa->offset;
381 if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
382 size_t printed = scnprintf(bf, size, intfmt, val);
384 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
388 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
391 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
393 struct syscall_arg *arg)
395 return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
398 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
399 struct syscall_arg *arg)
401 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
404 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
406 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
408 return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
411 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
416 for (i = 0; i < sas->nr_entries; ++i) {
417 struct strarray *sa = sas->entries[i];
418 int idx = val - sa->offset;
420 if (idx >= 0 && idx < sa->nr_entries) {
421 if (sa->entries[idx] == NULL)
423 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
427 printed = scnprintf(bf, size, intfmt, val);
429 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
433 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
434 struct syscall_arg *arg)
436 return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
440 #define AT_FDCWD -100
443 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
444 struct syscall_arg *arg)
447 const char *prefix = "AT_FD";
450 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
452 return syscall_arg__scnprintf_fd(bf, size, arg);
455 #define SCA_FDAT syscall_arg__scnprintf_fd_at
457 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
458 struct syscall_arg *arg);
460 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
462 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
464 return scnprintf(bf, size, "%#lx", arg->val);
467 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
470 return scnprintf(bf, size, "NULL");
471 return syscall_arg__scnprintf_hex(bf, size, arg);
474 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
476 return scnprintf(bf, size, "%d", arg->val);
479 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
481 return scnprintf(bf, size, "%ld", arg->val);
484 static const char *bpf_cmd[] = {
485 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
486 "MAP_GET_NEXT_KEY", "PROG_LOAD",
488 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
490 static const char *fsmount_flags[] = {
493 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
495 #include "trace/beauty/generated/fsconfig_arrays.c"
497 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
499 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
500 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
502 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
503 static DEFINE_STRARRAY(itimers, "ITIMER_");
505 static const char *keyctl_options[] = {
506 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
507 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
508 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
509 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
510 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
512 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
514 static const char *whences[] = { "SET", "CUR", "END",
522 static DEFINE_STRARRAY(whences, "SEEK_");
524 static const char *fcntl_cmds[] = {
525 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
526 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
527 "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
530 static DEFINE_STRARRAY(fcntl_cmds, "F_");
532 static const char *fcntl_linux_specific_cmds[] = {
533 "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
534 "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
535 "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
538 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
540 static struct strarray *fcntl_cmds_arrays[] = {
541 &strarray__fcntl_cmds,
542 &strarray__fcntl_linux_specific_cmds,
545 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
547 static const char *rlimit_resources[] = {
548 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
549 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
552 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
554 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
555 static DEFINE_STRARRAY(sighow, "SIG_");
557 static const char *clockid[] = {
558 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
559 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
560 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
562 static DEFINE_STRARRAY(clockid, "CLOCK_");
564 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
565 struct syscall_arg *arg)
567 bool show_prefix = arg->show_string_prefix;
568 const char *suffix = "_OK";
572 if (mode == F_OK) /* 0 */
573 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
575 if (mode & n##_OK) { \
576 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
586 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
591 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
593 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
594 struct syscall_arg *arg);
596 #define SCA_FILENAME syscall_arg__scnprintf_filename
598 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
599 struct syscall_arg *arg)
601 bool show_prefix = arg->show_string_prefix;
602 const char *prefix = "O_";
603 int printed = 0, flags = arg->val;
606 if (flags & O_##n) { \
607 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
616 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
621 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
623 #ifndef GRND_NONBLOCK
624 #define GRND_NONBLOCK 0x0001
627 #define GRND_RANDOM 0x0002
630 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
631 struct syscall_arg *arg)
633 bool show_prefix = arg->show_string_prefix;
634 const char *prefix = "GRND_";
635 int printed = 0, flags = arg->val;
638 if (flags & GRND_##n) { \
639 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
640 flags &= ~GRND_##n; \
648 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
653 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
655 #define STRARRAY(name, array) \
656 { .scnprintf = SCA_STRARRAY, \
657 .parm = &strarray__##array, }
659 #define STRARRAY_FLAGS(name, array) \
660 { .scnprintf = SCA_STRARRAY_FLAGS, \
661 .parm = &strarray__##array, }
663 #include "trace/beauty/arch_errno_names.c"
664 #include "trace/beauty/eventfd.c"
665 #include "trace/beauty/futex_op.c"
666 #include "trace/beauty/futex_val3.c"
667 #include "trace/beauty/mmap.c"
668 #include "trace/beauty/mode_t.c"
669 #include "trace/beauty/msg_flags.c"
670 #include "trace/beauty/open_flags.c"
671 #include "trace/beauty/perf_event_open.c"
672 #include "trace/beauty/pid.c"
673 #include "trace/beauty/sched_policy.c"
674 #include "trace/beauty/seccomp.c"
675 #include "trace/beauty/signum.c"
676 #include "trace/beauty/socket_type.c"
677 #include "trace/beauty/waitid_options.c"
679 struct syscall_arg_fmt {
680 size_t (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
681 unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
687 static struct syscall_fmt {
690 struct syscall_arg_fmt arg[6];
697 .arg = { [1] = { .scnprintf = SCA_ACCMODE, /* mode */ }, }, },
698 { .name = "arch_prctl",
699 .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
700 [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
702 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
704 .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
705 { .name = "brk", .hexret = true,
706 .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
707 { .name = "clock_gettime",
708 .arg = { [0] = STRARRAY(clk_id, clockid), }, },
709 { .name = "clone", .errpid = true, .nr_args = 5,
710 .arg = { [0] = { .name = "flags", .scnprintf = SCA_CLONE_FLAGS, },
711 [1] = { .name = "child_stack", .scnprintf = SCA_HEX, },
712 [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
713 [3] = { .name = "child_tidptr", .scnprintf = SCA_HEX, },
714 [4] = { .name = "tls", .scnprintf = SCA_HEX, }, }, },
716 .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
718 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
719 { .name = "epoll_ctl",
720 .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
721 { .name = "eventfd2",
722 .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
723 { .name = "fchmodat",
724 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
725 { .name = "fchownat",
726 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
728 .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
729 .parm = &strarrays__fcntl_cmds_arrays,
730 .show_zero = true, },
731 [2] = { .scnprintf = SCA_FCNTL_ARG, /* arg */ }, }, },
733 .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
734 { .name = "fsconfig",
735 .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
737 .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
738 [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
740 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
741 [1] = { .scnprintf = SCA_FILENAME, /* path */ },
742 [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
743 { .name = "fstat", .alias = "newfstat", },
744 { .name = "fstatat", .alias = "newfstatat", },
746 .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
747 [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
748 { .name = "futimesat",
749 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
750 { .name = "getitimer",
751 .arg = { [0] = STRARRAY(which, itimers), }, },
752 { .name = "getpid", .errpid = true, },
753 { .name = "getpgid", .errpid = true, },
754 { .name = "getppid", .errpid = true, },
755 { .name = "getrandom",
756 .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
757 { .name = "getrlimit",
758 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
759 { .name = "gettid", .errpid = true, },
762 #if defined(__i386__) || defined(__x86_64__)
764 * FIXME: Make this available to all arches.
766 [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
767 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
769 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
771 { .name = "kcmp", .nr_args = 5,
772 .arg = { [0] = { .name = "pid1", .scnprintf = SCA_PID, },
773 [1] = { .name = "pid2", .scnprintf = SCA_PID, },
774 [2] = { .name = "type", .scnprintf = SCA_KCMP_TYPE, },
775 [3] = { .name = "idx1", .scnprintf = SCA_KCMP_IDX, },
776 [4] = { .name = "idx2", .scnprintf = SCA_KCMP_IDX, }, }, },
778 .arg = { [0] = STRARRAY(option, keyctl_options), }, },
780 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
782 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
784 .arg = { [2] = STRARRAY(whence, whences), }, },
785 { .name = "lstat", .alias = "newlstat", },
787 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
788 [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
790 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
792 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
793 { .name = "mmap", .hexret = true,
794 /* The standard mmap maps to old_mmap on s390x */
795 #if defined(__s390x__)
798 .arg = { [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
799 [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
800 [5] = { .scnprintf = SCA_HEX, /* offset */ }, }, },
802 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
803 [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
804 .mask_val = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
805 { .name = "move_mount",
806 .arg = { [0] = { .scnprintf = SCA_FDAT, /* from_dfd */ },
807 [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
808 [2] = { .scnprintf = SCA_FDAT, /* to_dfd */ },
809 [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
810 [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
811 { .name = "mprotect",
812 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
813 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, }, },
814 { .name = "mq_unlink",
815 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
816 { .name = "mremap", .hexret = true,
817 .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
818 { .name = "name_to_handle_at",
819 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
820 { .name = "newfstatat",
821 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
823 .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
824 { .name = "open_by_handle_at",
825 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
826 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
828 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
829 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
830 { .name = "perf_event_open",
831 .arg = { [2] = { .scnprintf = SCA_INT, /* cpu */ },
832 [3] = { .scnprintf = SCA_FD, /* group_fd */ },
833 [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
835 .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
836 { .name = "pkey_alloc",
837 .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS, /* access_rights */ }, }, },
838 { .name = "pkey_free",
839 .arg = { [0] = { .scnprintf = SCA_INT, /* key */ }, }, },
840 { .name = "pkey_mprotect",
841 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
842 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
843 [3] = { .scnprintf = SCA_INT, /* pkey */ }, }, },
844 { .name = "poll", .timeout = true, },
845 { .name = "ppoll", .timeout = true, },
847 .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
848 [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
849 [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
850 { .name = "pread", .alias = "pread64", },
851 { .name = "preadv", .alias = "pread", },
852 { .name = "prlimit64",
853 .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
854 { .name = "pwrite", .alias = "pwrite64", },
855 { .name = "readlinkat",
856 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
857 { .name = "recvfrom",
858 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
859 { .name = "recvmmsg",
860 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
862 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
863 { .name = "renameat",
864 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
865 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
866 { .name = "renameat2",
867 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
868 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
869 [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
870 { .name = "rt_sigaction",
871 .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
872 { .name = "rt_sigprocmask",
873 .arg = { [0] = STRARRAY(how, sighow), }, },
874 { .name = "rt_sigqueueinfo",
875 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
876 { .name = "rt_tgsigqueueinfo",
877 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
878 { .name = "sched_setscheduler",
879 .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
881 .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP, /* op */ },
882 [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
883 { .name = "select", .timeout = true, },
884 { .name = "sendmmsg",
885 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
887 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
889 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
890 [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
891 { .name = "set_tid_address", .errpid = true, },
892 { .name = "setitimer",
893 .arg = { [0] = STRARRAY(which, itimers), }, },
894 { .name = "setrlimit",
895 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
897 .arg = { [0] = STRARRAY(family, socket_families),
898 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
899 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
900 { .name = "socketpair",
901 .arg = { [0] = STRARRAY(family, socket_families),
902 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
903 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
904 { .name = "stat", .alias = "newstat", },
906 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fdat */ },
907 [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
908 [3] = { .scnprintf = SCA_STATX_MASK, /* mask */ }, }, },
910 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
912 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
913 { .name = "symlinkat",
914 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
915 { .name = "sync_file_range",
916 .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
918 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
920 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
921 { .name = "umount2", .alias = "umount",
922 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
923 { .name = "uname", .alias = "newuname", },
924 { .name = "unlinkat",
925 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
926 { .name = "utimensat",
927 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
928 { .name = "wait4", .errpid = true,
929 .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
930 { .name = "waitid", .errpid = true,
931 .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
934 static int syscall_fmt__cmp(const void *name, const void *fmtp)
936 const struct syscall_fmt *fmt = fmtp;
937 return strcmp(name, fmt->name);
940 static struct syscall_fmt *syscall_fmt__find(const char *name)
942 const int nmemb = ARRAY_SIZE(syscall_fmts);
943 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
946 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
948 int i, nmemb = ARRAY_SIZE(syscall_fmts);
950 for (i = 0; i < nmemb; ++i) {
951 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
952 return &syscall_fmts[i];
959 * is_exit: is this "exit" or "exit_group"?
960 * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
961 * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
964 struct tep_event *tp_format;
969 struct tep_format_field *args;
971 struct syscall_fmt *fmt;
972 struct syscall_arg_fmt *arg_fmt;
976 * Must match what is in the BPF program:
978 * tools/perf/examples/bpf/augmented_raw_syscalls.c
980 struct bpf_map_syscall_entry {
982 u16 string_args_len[6];
986 * We need to have this 'calculated' boolean because in some cases we really
987 * don't know what is the duration of a syscall, for instance, when we start
988 * a session and some threads are waiting for a syscall to finish, say 'poll',
989 * in which case all we can do is to print "( ? ) for duration and for the
992 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
994 double duration = (double)t / NSEC_PER_MSEC;
995 size_t printed = fprintf(fp, "(");
998 printed += fprintf(fp, " ");
999 else if (duration >= 1.0)
1000 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1001 else if (duration >= 0.01)
1002 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1004 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1005 return printed + fprintf(fp, "): ");
1009 * filename.ptr: The filename char pointer that will be vfs_getname'd
1010 * filename.entry_str_pos: Where to insert the string translated from
1011 * filename.ptr by the vfs_getname tracepoint/kprobe.
1012 * ret_scnprintf: syscall args may set this to a different syscall return
1013 * formatter, for instance, fcntl may return fds, file flags, etc.
1015 struct thread_trace {
1018 unsigned long nr_events;
1019 unsigned long pfmaj, pfmin;
1022 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1025 short int entry_str_pos;
1027 unsigned int namelen;
1035 struct intlist *syscall_stats;
1038 static struct thread_trace *thread_trace__new(void)
1040 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
1043 ttrace->files.max = -1;
1044 ttrace->syscall_stats = intlist__new(NULL);
1050 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1052 struct thread_trace *ttrace;
1057 if (thread__priv(thread) == NULL)
1058 thread__set_priv(thread, thread_trace__new());
1060 if (thread__priv(thread) == NULL)
1063 ttrace = thread__priv(thread);
1064 ++ttrace->nr_events;
1068 color_fprintf(fp, PERF_COLOR_RED,
1069 "WARNING: not enough memory, dropping samples!\n");
1074 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1075 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1077 struct thread_trace *ttrace = thread__priv(arg->thread);
1079 ttrace->ret_scnprintf = ret_scnprintf;
1082 #define TRACE_PFMAJ (1 << 0)
1083 #define TRACE_PFMIN (1 << 1)
1085 static const size_t trace__entry_str_size = 2048;
1087 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1092 if (fd > ttrace->files.max) {
1093 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1098 if (ttrace->files.max != -1) {
1099 memset(nfiles + ttrace->files.max + 1, 0,
1100 (fd - ttrace->files.max) * sizeof(struct file));
1102 memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1105 ttrace->files.table = nfiles;
1106 ttrace->files.max = fd;
1109 return ttrace->files.table + fd;
1112 struct file *thread__files_entry(struct thread *thread, int fd)
1114 return thread_trace__files_entry(thread__priv(thread), fd);
1117 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1119 struct thread_trace *ttrace = thread__priv(thread);
1120 struct file *file = thread_trace__files_entry(ttrace, fd);
1124 if (stat(pathname, &st) == 0)
1125 file->dev_maj = major(st.st_rdev);
1126 file->pathname = strdup(pathname);
1134 static int thread__read_fd_path(struct thread *thread, int fd)
1136 char linkname[PATH_MAX], pathname[PATH_MAX];
1140 if (thread->pid_ == thread->tid) {
1141 scnprintf(linkname, sizeof(linkname),
1142 "/proc/%d/fd/%d", thread->pid_, fd);
1144 scnprintf(linkname, sizeof(linkname),
1145 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1148 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1151 ret = readlink(linkname, pathname, sizeof(pathname));
1153 if (ret < 0 || ret > st.st_size)
1156 pathname[ret] = '\0';
1157 return trace__set_fd_pathname(thread, fd, pathname);
1160 static const char *thread__fd_path(struct thread *thread, int fd,
1161 struct trace *trace)
1163 struct thread_trace *ttrace = thread__priv(thread);
1171 if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1174 ++trace->stats.proc_getname;
1175 if (thread__read_fd_path(thread, fd))
1179 return ttrace->files.table[fd].pathname;
1182 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1185 size_t printed = scnprintf(bf, size, "%d", fd);
1186 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1189 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1194 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1196 size_t printed = scnprintf(bf, size, "%d", fd);
1197 struct thread *thread = machine__find_thread(trace->host, pid, pid);
1200 const char *path = thread__fd_path(thread, fd, trace);
1203 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1205 thread__put(thread);
1211 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1212 struct syscall_arg *arg)
1215 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1216 struct thread_trace *ttrace = thread__priv(arg->thread);
1218 if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1219 zfree(&ttrace->files.table[fd].pathname);
1224 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1227 struct thread_trace *ttrace = thread__priv(thread);
1229 ttrace->filename.ptr = ptr;
1230 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1233 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1235 struct augmented_arg *augmented_arg = arg->augmented.args;
1236 size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1238 * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1239 * we would have two strings, each prefixed by its size.
1241 int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1243 arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1244 arg->augmented.size -= consumed;
1249 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1250 struct syscall_arg *arg)
1252 unsigned long ptr = arg->val;
1254 if (arg->augmented.args)
1255 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1257 if (!arg->trace->vfs_getname)
1258 return scnprintf(bf, size, "%#x", ptr);
1260 thread__set_filename_pos(arg->thread, bf, ptr);
1264 static bool trace__filter_duration(struct trace *trace, double t)
1266 return t < (trace->duration_filter * NSEC_PER_MSEC);
1269 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1271 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1273 return fprintf(fp, "%10.3f ", ts);
1277 * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1278 * using ttrace->entry_time for a thread that receives a sys_exit without
1279 * first having received a sys_enter ("poll" issued before tracing session
1280 * starts, lost sys_enter exit due to ring buffer overflow).
1282 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1285 return __trace__fprintf_tstamp(trace, tstamp, fp);
1287 return fprintf(fp, " ? ");
1290 static bool done = false;
1291 static bool interrupted = false;
1293 static void sig_handler(int sig)
1296 interrupted = sig == SIGINT;
1299 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1303 if (trace->multiple_threads) {
1304 if (trace->show_comm)
1305 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1306 printed += fprintf(fp, "%d ", thread->tid);
1312 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1313 u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1317 if (trace->show_tstamp)
1318 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1319 if (trace->show_duration)
1320 printed += fprintf_duration(duration, duration_calculated, fp);
1321 return printed + trace__fprintf_comm_tid(trace, thread, fp);
1324 static int trace__process_event(struct trace *trace, struct machine *machine,
1325 union perf_event *event, struct perf_sample *sample)
1329 switch (event->header.type) {
1330 case PERF_RECORD_LOST:
1331 color_fprintf(trace->output, PERF_COLOR_RED,
1332 "LOST %" PRIu64 " events!\n", event->lost.lost);
1333 ret = machine__process_lost_event(machine, event, sample);
1336 ret = machine__process_event(machine, event, sample);
1343 static int trace__tool_process(struct perf_tool *tool,
1344 union perf_event *event,
1345 struct perf_sample *sample,
1346 struct machine *machine)
1348 struct trace *trace = container_of(tool, struct trace, tool);
1349 return trace__process_event(trace, machine, event, sample);
1352 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1354 struct machine *machine = vmachine;
1356 if (machine->kptr_restrict_warned)
1359 if (symbol_conf.kptr_restrict) {
1360 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1361 "Check /proc/sys/kernel/kptr_restrict.\n\n"
1362 "Kernel samples will not be resolved.\n");
1363 machine->kptr_restrict_warned = true;
1367 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1370 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1372 int err = symbol__init(NULL);
1377 trace->host = machine__new_host();
1378 if (trace->host == NULL)
1381 err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1385 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1386 evlist->threads, trace__tool_process, false,
1395 static void trace__symbols__exit(struct trace *trace)
1397 machine__exit(trace->host);
1403 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1407 if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1408 nr_args = sc->fmt->nr_args;
1410 sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1411 if (sc->arg_fmt == NULL)
1414 for (idx = 0; idx < nr_args; ++idx) {
1416 sc->arg_fmt[idx] = sc->fmt->arg[idx];
1419 sc->nr_args = nr_args;
1423 static int syscall__set_arg_fmts(struct syscall *sc)
1425 struct tep_format_field *field, *last_field = NULL;
1428 for (field = sc->args; field; field = field->next, ++idx) {
1431 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1434 len = strlen(field->name);
1436 if (strcmp(field->type, "const char *") == 0 &&
1437 ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1438 strstr(field->name, "path") != NULL))
1439 sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1440 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1441 sc->arg_fmt[idx].scnprintf = SCA_PTR;
1442 else if (strcmp(field->type, "pid_t") == 0)
1443 sc->arg_fmt[idx].scnprintf = SCA_PID;
1444 else if (strcmp(field->type, "umode_t") == 0)
1445 sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1446 else if ((strcmp(field->type, "int") == 0 ||
1447 strcmp(field->type, "unsigned int") == 0 ||
1448 strcmp(field->type, "long") == 0) &&
1449 len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1451 * /sys/kernel/tracing/events/syscalls/sys_enter*
1452 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1457 sc->arg_fmt[idx].scnprintf = SCA_FD;
1462 sc->args_size = last_field->offset + last_field->size;
1467 static int trace__read_syscall_info(struct trace *trace, int id)
1471 const char *name = syscalltbl__name(trace->sctbl, id);
1476 if (id > trace->syscalls.max) {
1477 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1479 if (nsyscalls == NULL)
1482 if (trace->syscalls.max != -1) {
1483 memset(nsyscalls + trace->syscalls.max + 1, 0,
1484 (id - trace->syscalls.max) * sizeof(*sc));
1486 memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1489 trace->syscalls.table = nsyscalls;
1490 trace->syscalls.max = id;
1493 sc = trace->syscalls.table + id;
1496 sc->fmt = syscall_fmt__find(sc->name);
1498 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1499 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1501 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1502 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1503 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1506 if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1509 if (IS_ERR(sc->tp_format))
1512 sc->args = sc->tp_format->format.fields;
1514 * We need to check and discard the first variable '__syscall_nr'
1515 * or 'nr' that mean the syscall number. It is needless here.
1516 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1518 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1519 sc->args = sc->args->next;
1523 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1524 sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1526 return syscall__set_arg_fmts(sc);
1529 static int trace__validate_ev_qualifier(struct trace *trace)
1532 bool printed_invalid_prefix = false;
1533 struct str_node *pos;
1534 size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1536 trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1537 sizeof(trace->ev_qualifier_ids.entries[0]));
1539 if (trace->ev_qualifier_ids.entries == NULL) {
1540 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1546 strlist__for_each_entry(pos, trace->ev_qualifier) {
1547 const char *sc = pos->s;
1548 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1551 id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1555 if (!printed_invalid_prefix) {
1556 pr_debug("Skipping unknown syscalls: ");
1557 printed_invalid_prefix = true;
1566 trace->ev_qualifier_ids.entries[nr_used++] = id;
1567 if (match_next == -1)
1571 id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1574 if (nr_allocated == nr_used) {
1578 entries = realloc(trace->ev_qualifier_ids.entries,
1579 nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1580 if (entries == NULL) {
1582 fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1585 trace->ev_qualifier_ids.entries = entries;
1587 trace->ev_qualifier_ids.entries[nr_used++] = id;
1591 trace->ev_qualifier_ids.nr = nr_used;
1593 if (printed_invalid_prefix)
1597 zfree(&trace->ev_qualifier_ids.entries);
1598 trace->ev_qualifier_ids.nr = 0;
1603 * args is to be interpreted as a series of longs but we need to handle
1604 * 8-byte unaligned accesses. args points to raw_data within the event
1605 * and raw_data is guaranteed to be 8-byte unaligned because it is
1606 * preceded by raw_size which is a u32. So we need to copy args to a temp
1607 * variable to read it. Most notably this avoids extended load instructions
1608 * on unaligned addresses
1610 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1613 unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1615 memcpy(&val, p, sizeof(val));
1619 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1620 struct syscall_arg *arg)
1622 if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1623 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1625 return scnprintf(bf, size, "arg%d: ", arg->idx);
1629 * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1630 * as mount 'flags' argument that needs ignoring some magic flag, see comment
1631 * in tools/perf/trace/beauty/mount_flags.c
1633 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1635 if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1636 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1641 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1642 struct syscall_arg *arg, unsigned long val)
1644 if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1646 if (sc->arg_fmt[arg->idx].parm)
1647 arg->parm = sc->arg_fmt[arg->idx].parm;
1648 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1650 return scnprintf(bf, size, "%ld", val);
1653 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1654 unsigned char *args, void *augmented_args, int augmented_args_size,
1655 struct trace *trace, struct thread *thread)
1660 struct syscall_arg arg = {
1663 .size = augmented_args_size,
1664 .args = augmented_args,
1670 .show_string_prefix = trace->show_string_prefix,
1672 struct thread_trace *ttrace = thread__priv(thread);
1675 * Things like fcntl will set this in its 'cmd' formatter to pick the
1676 * right formatter for the return value (an fd? file flags?), which is
1677 * not needed for syscalls that always return a given type, say an fd.
1679 ttrace->ret_scnprintf = NULL;
1681 if (sc->args != NULL) {
1682 struct tep_format_field *field;
1684 for (field = sc->args; field;
1685 field = field->next, ++arg.idx, bit <<= 1) {
1689 val = syscall_arg__val(&arg, arg.idx);
1691 * Some syscall args need some mask, most don't and
1692 * return val untouched.
1694 val = syscall__mask_val(sc, &arg, val);
1697 * Suppress this argument if its value is zero and
1698 * and we don't have a string associated in an
1702 !trace->show_zeros &&
1704 (sc->arg_fmt[arg.idx].show_zero ||
1705 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1706 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1707 sc->arg_fmt[arg.idx].parm))
1710 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1712 if (trace->show_arg_names)
1713 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1715 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1717 } else if (IS_ERR(sc->tp_format)) {
1719 * If we managed to read the tracepoint /format file, then we
1720 * may end up not having any args, like with gettid(), so only
1721 * print the raw args when we didn't manage to read it.
1723 while (arg.idx < sc->nr_args) {
1726 val = syscall_arg__val(&arg, arg.idx);
1728 printed += scnprintf(bf + printed, size - printed, ", ");
1729 printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1730 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1740 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1741 union perf_event *event,
1742 struct perf_sample *sample);
1744 static struct syscall *trace__syscall_info(struct trace *trace,
1745 struct perf_evsel *evsel, int id)
1751 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1752 * before that, leaving at a higher verbosity level till that is
1753 * explained. Reproduced with plain ftrace with:
1755 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1756 * grep "NR -1 " /t/trace_pipe
1758 * After generating some load on the machine.
1762 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1763 id, perf_evsel__name(evsel), ++n);
1768 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1769 trace__read_syscall_info(trace, id))
1772 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1775 return &trace->syscalls.table[id];
1779 fprintf(trace->output, "Problems reading syscall %d", id);
1780 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1781 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1782 fputs(" information\n", trace->output);
1787 static void thread__update_stats(struct thread_trace *ttrace,
1788 int id, struct perf_sample *sample)
1790 struct int_node *inode;
1791 struct stats *stats;
1794 inode = intlist__findnew(ttrace->syscall_stats, id);
1798 stats = inode->priv;
1799 if (stats == NULL) {
1800 stats = malloc(sizeof(struct stats));
1804 inode->priv = stats;
1807 if (ttrace->entry_time && sample->time > ttrace->entry_time)
1808 duration = sample->time - ttrace->entry_time;
1810 update_stats(stats, duration);
1813 static int trace__printf_interrupted_entry(struct trace *trace)
1815 struct thread_trace *ttrace;
1819 if (trace->failure_only || trace->current == NULL)
1822 ttrace = thread__priv(trace->current);
1824 if (!ttrace->entry_pending)
1827 printed = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1828 printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1830 if (len < trace->args_alignment - 4)
1831 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1833 printed += fprintf(trace->output, " ...\n");
1835 ttrace->entry_pending = false;
1836 ++trace->nr_events_printed;
1841 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1842 struct perf_sample *sample, struct thread *thread)
1846 if (trace->print_sample) {
1847 double ts = (double)sample->time / NSEC_PER_MSEC;
1849 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1850 perf_evsel__name(evsel), ts,
1851 thread__comm_str(thread),
1852 sample->pid, sample->tid, sample->cpu);
1858 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1860 void *augmented_args = NULL;
1862 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1863 * and there we get all 6 syscall args plus the tracepoint common fields
1864 * that gets calculated at the start and the syscall_nr (another long).
1865 * So we check if that is the case and if so don't look after the
1866 * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1869 * We'll revisit this later to pass s->args_size to the BPF augmenter
1870 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1871 * copies only what we need for each syscall, like what happens when we
1872 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1873 * traffic to just what is needed for each syscall.
1875 int args_size = raw_augmented_args_size ?: sc->args_size;
1877 *augmented_args_size = sample->raw_size - args_size;
1878 if (*augmented_args_size > 0)
1879 augmented_args = sample->raw_data + args_size;
1881 return augmented_args;
1884 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1885 union perf_event *event __maybe_unused,
1886 struct perf_sample *sample)
1891 struct thread *thread;
1892 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1893 int augmented_args_size = 0;
1894 void *augmented_args = NULL;
1895 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1896 struct thread_trace *ttrace;
1901 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1902 ttrace = thread__trace(thread, trace->output);
1906 trace__fprintf_sample(trace, evsel, sample, thread);
1908 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1910 if (ttrace->entry_str == NULL) {
1911 ttrace->entry_str = malloc(trace__entry_str_size);
1912 if (!ttrace->entry_str)
1916 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1917 trace__printf_interrupted_entry(trace);
1919 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1920 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1921 * this breaks syscall__augmented_args() check for augmented args, as we calculate
1922 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1923 * so when handling, say the openat syscall, we end up getting 6 args for the
1924 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1925 * thinking that the extra 2 u64 args are the augmented filename, so just check
1926 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1928 if (evsel != trace->syscalls.events.sys_enter)
1929 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1930 ttrace->entry_time = sample->time;
1931 msg = ttrace->entry_str;
1932 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1934 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1935 args, augmented_args, augmented_args_size, trace, thread);
1938 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1941 trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1942 printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1943 if (trace->args_alignment > printed)
1944 alignment = trace->args_alignment - printed;
1945 fprintf(trace->output, "%*s= ?\n", alignment, " ");
1948 ttrace->entry_pending = true;
1949 /* See trace__vfs_getname & trace__sys_exit */
1950 ttrace->filename.pending_open = false;
1953 if (trace->current != thread) {
1954 thread__put(trace->current);
1955 trace->current = thread__get(thread);
1959 thread__put(thread);
1963 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1964 struct perf_sample *sample)
1966 struct thread_trace *ttrace;
1967 struct thread *thread;
1968 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1969 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1971 void *args, *augmented_args = NULL;
1972 int augmented_args_size;
1977 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1978 ttrace = thread__trace(thread, trace->output);
1980 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1981 * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1986 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1987 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1988 syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1989 fprintf(trace->output, "%s", msg);
1992 thread__put(thread);
1996 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1997 struct perf_sample *sample,
1998 struct callchain_cursor *cursor)
2000 struct addr_location al;
2001 int max_stack = evsel->attr.sample_max_stack ?
2002 evsel->attr.sample_max_stack :
2006 if (machine__resolve(trace->host, &al, sample) < 0)
2009 err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2010 addr_location__put(&al);
2014 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2016 /* TODO: user-configurable print_opts */
2017 const unsigned int print_opts = EVSEL__PRINT_SYM |
2019 EVSEL__PRINT_UNKNOWN_AS_ADDR;
2021 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2024 static const char *errno_to_name(struct perf_evsel *evsel, int err)
2026 struct perf_env *env = perf_evsel__env(evsel);
2027 const char *arch_name = perf_env__arch(env);
2029 return arch_syscalls__strerrno(arch_name, err);
2032 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
2033 union perf_event *event __maybe_unused,
2034 struct perf_sample *sample)
2038 bool duration_calculated = false;
2039 struct thread *thread;
2040 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2041 int alignment = trace->args_alignment;
2042 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2043 struct thread_trace *ttrace;
2048 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2049 ttrace = thread__trace(thread, trace->output);
2053 trace__fprintf_sample(trace, evsel, sample, thread);
2056 thread__update_stats(ttrace, id, sample);
2058 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2060 if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2061 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2062 ttrace->filename.pending_open = false;
2063 ++trace->stats.vfs_getname;
2066 if (ttrace->entry_time) {
2067 duration = sample->time - ttrace->entry_time;
2068 if (trace__filter_duration(trace, duration))
2070 duration_calculated = true;
2071 } else if (trace->duration_filter)
2074 if (sample->callchain) {
2075 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2076 if (callchain_ret == 0) {
2077 if (callchain_cursor.nr < trace->min_stack)
2083 if (trace->summary_only || (ret >= 0 && trace->failure_only))
2086 trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2088 if (ttrace->entry_pending) {
2089 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2091 printed += fprintf(trace->output, " ... [");
2092 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2094 printed += fprintf(trace->output, "]: %s()", sc->name);
2097 printed++; /* the closing ')' */
2099 if (alignment > printed)
2100 alignment -= printed;
2104 fprintf(trace->output, ")%*s= ", alignment, " ");
2106 if (sc->fmt == NULL) {
2110 fprintf(trace->output, "%ld", ret);
2111 } else if (ret < 0) {
2113 char bf[STRERR_BUFSIZE];
2114 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2115 *e = errno_to_name(evsel, -ret);
2117 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2119 } else if (ret == 0 && sc->fmt->timeout)
2120 fprintf(trace->output, "0 (Timeout)");
2121 else if (ttrace->ret_scnprintf) {
2123 struct syscall_arg arg = {
2128 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2129 ttrace->ret_scnprintf = NULL;
2130 fprintf(trace->output, "%s", bf);
2131 } else if (sc->fmt->hexret)
2132 fprintf(trace->output, "%#lx", ret);
2133 else if (sc->fmt->errpid) {
2134 struct thread *child = machine__find_thread(trace->host, ret, ret);
2136 if (child != NULL) {
2137 fprintf(trace->output, "%ld", ret);
2138 if (child->comm_set)
2139 fprintf(trace->output, " (%s)", thread__comm_str(child));
2145 fputc('\n', trace->output);
2148 * We only consider an 'event' for the sake of --max-events a non-filtered
2149 * sys_enter + sys_exit and other tracepoint events.
2151 if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2154 if (callchain_ret > 0)
2155 trace__fprintf_callchain(trace, sample);
2156 else if (callchain_ret < 0)
2157 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2159 ttrace->entry_pending = false;
2162 thread__put(thread);
2166 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2167 union perf_event *event __maybe_unused,
2168 struct perf_sample *sample)
2170 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2171 struct thread_trace *ttrace;
2172 size_t filename_len, entry_str_len, to_move;
2173 ssize_t remaining_space;
2175 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2180 ttrace = thread__priv(thread);
2184 filename_len = strlen(filename);
2185 if (filename_len == 0)
2188 if (ttrace->filename.namelen < filename_len) {
2189 char *f = realloc(ttrace->filename.name, filename_len + 1);
2194 ttrace->filename.namelen = filename_len;
2195 ttrace->filename.name = f;
2198 strcpy(ttrace->filename.name, filename);
2199 ttrace->filename.pending_open = true;
2201 if (!ttrace->filename.ptr)
2204 entry_str_len = strlen(ttrace->entry_str);
2205 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2206 if (remaining_space <= 0)
2209 if (filename_len > (size_t)remaining_space) {
2210 filename += filename_len - remaining_space;
2211 filename_len = remaining_space;
2214 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2215 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2216 memmove(pos + filename_len, pos, to_move);
2217 memcpy(pos, filename, filename_len);
2219 ttrace->filename.ptr = 0;
2220 ttrace->filename.entry_str_pos = 0;
2222 thread__put(thread);
2227 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2228 union perf_event *event __maybe_unused,
2229 struct perf_sample *sample)
2231 u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2232 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2233 struct thread *thread = machine__findnew_thread(trace->host,
2236 struct thread_trace *ttrace = thread__trace(thread, trace->output);
2241 ttrace->runtime_ms += runtime_ms;
2242 trace->runtime_ms += runtime_ms;
2244 thread__put(thread);
2248 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2250 perf_evsel__strval(evsel, sample, "comm"),
2251 (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2253 perf_evsel__intval(evsel, sample, "vruntime"));
2257 static int bpf_output__printer(enum binary_printer_ops op,
2258 unsigned int val, void *extra __maybe_unused, FILE *fp)
2260 unsigned char ch = (unsigned char)val;
2263 case BINARY_PRINT_CHAR_DATA:
2264 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2265 case BINARY_PRINT_DATA_BEGIN:
2266 case BINARY_PRINT_LINE_BEGIN:
2267 case BINARY_PRINT_ADDR:
2268 case BINARY_PRINT_NUM_DATA:
2269 case BINARY_PRINT_NUM_PAD:
2270 case BINARY_PRINT_SEP:
2271 case BINARY_PRINT_CHAR_PAD:
2272 case BINARY_PRINT_LINE_END:
2273 case BINARY_PRINT_DATA_END:
2281 static void bpf_output__fprintf(struct trace *trace,
2282 struct perf_sample *sample)
2284 binary__fprintf(sample->raw_data, sample->raw_size, 8,
2285 bpf_output__printer, NULL, trace->output);
2286 ++trace->nr_events_printed;
2289 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2290 union perf_event *event __maybe_unused,
2291 struct perf_sample *sample)
2293 struct thread *thread;
2294 int callchain_ret = 0;
2296 * Check if we called perf_evsel__disable(evsel) due to, for instance,
2297 * this event's max_events having been hit and this is an entry coming
2298 * from the ring buffer that we should discard, since the max events
2299 * have already been considered/printed.
2301 if (evsel->disabled)
2304 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2306 if (sample->callchain) {
2307 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2308 if (callchain_ret == 0) {
2309 if (callchain_cursor.nr < trace->min_stack)
2315 trace__printf_interrupted_entry(trace);
2316 trace__fprintf_tstamp(trace, sample->time, trace->output);
2318 if (trace->trace_syscalls && trace->show_duration)
2319 fprintf(trace->output, "( ): ");
2322 trace__fprintf_comm_tid(trace, thread, trace->output);
2324 if (evsel == trace->syscalls.events.augmented) {
2325 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2326 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2329 fprintf(trace->output, "%s(", sc->name);
2330 trace__fprintf_sys_enter(trace, evsel, sample);
2331 fputc(')', trace->output);
2336 * XXX: Not having the associated syscall info or not finding/adding
2337 * the thread should never happen, but if it does...
2338 * fall thru and print it as a bpf_output event.
2342 fprintf(trace->output, "%s:", evsel->name);
2344 if (perf_evsel__is_bpf_output(evsel)) {
2345 bpf_output__fprintf(trace, sample);
2346 } else if (evsel->tp_format) {
2347 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2348 trace__fprintf_sys_enter(trace, evsel, sample)) {
2349 event_format__fprintf(evsel->tp_format, sample->cpu,
2350 sample->raw_data, sample->raw_size,
2352 ++trace->nr_events_printed;
2354 if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2355 perf_evsel__disable(evsel);
2356 perf_evsel__close(evsel);
2362 fprintf(trace->output, "\n");
2364 if (callchain_ret > 0)
2365 trace__fprintf_callchain(trace, sample);
2366 else if (callchain_ret < 0)
2367 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2369 thread__put(thread);
2373 static void print_location(FILE *f, struct perf_sample *sample,
2374 struct addr_location *al,
2375 bool print_dso, bool print_sym)
2378 if ((verbose > 0 || print_dso) && al->map)
2379 fprintf(f, "%s@", al->map->dso->long_name);
2381 if ((verbose > 0 || print_sym) && al->sym)
2382 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2383 al->addr - al->sym->start);
2385 fprintf(f, "0x%" PRIx64, al->addr);
2387 fprintf(f, "0x%" PRIx64, sample->addr);
2390 static int trace__pgfault(struct trace *trace,
2391 struct perf_evsel *evsel,
2392 union perf_event *event __maybe_unused,
2393 struct perf_sample *sample)
2395 struct thread *thread;
2396 struct addr_location al;
2397 char map_type = 'd';
2398 struct thread_trace *ttrace;
2400 int callchain_ret = 0;
2402 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2404 if (sample->callchain) {
2405 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2406 if (callchain_ret == 0) {
2407 if (callchain_cursor.nr < trace->min_stack)
2413 ttrace = thread__trace(thread, trace->output);
2417 if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2422 if (trace->summary_only)
2425 thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2427 trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2429 fprintf(trace->output, "%sfault [",
2430 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2433 print_location(trace->output, sample, &al, false, true);
2435 fprintf(trace->output, "] => ");
2437 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2440 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2448 print_location(trace->output, sample, &al, true, false);
2450 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2452 if (callchain_ret > 0)
2453 trace__fprintf_callchain(trace, sample);
2454 else if (callchain_ret < 0)
2455 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2457 ++trace->nr_events_printed;
2461 thread__put(thread);
2465 static void trace__set_base_time(struct trace *trace,
2466 struct perf_evsel *evsel,
2467 struct perf_sample *sample)
2470 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2471 * and don't use sample->time unconditionally, we may end up having
2472 * some other event in the future without PERF_SAMPLE_TIME for good
2473 * reason, i.e. we may not be interested in its timestamps, just in
2474 * it taking place, picking some piece of information when it
2475 * appears in our event stream (vfs_getname comes to mind).
2477 if (trace->base_time == 0 && !trace->full_time &&
2478 (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2479 trace->base_time = sample->time;
2482 static int trace__process_sample(struct perf_tool *tool,
2483 union perf_event *event,
2484 struct perf_sample *sample,
2485 struct perf_evsel *evsel,
2486 struct machine *machine __maybe_unused)
2488 struct trace *trace = container_of(tool, struct trace, tool);
2489 struct thread *thread;
2492 tracepoint_handler handler = evsel->handler;
2494 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2495 if (thread && thread__is_filtered(thread))
2498 trace__set_base_time(trace, evsel, sample);
2502 handler(trace, evsel, event, sample);
2505 thread__put(thread);
2509 static int trace__record(struct trace *trace, int argc, const char **argv)
2511 unsigned int rec_argc, i, j;
2512 const char **rec_argv;
2513 const char * const record_args[] = {
2520 const char * const sc_args[] = { "-e", };
2521 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2522 const char * const majpf_args[] = { "-e", "major-faults" };
2523 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2524 const char * const minpf_args[] = { "-e", "minor-faults" };
2525 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2527 /* +1 is for the event string below */
2528 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2529 majpf_args_nr + minpf_args_nr + argc;
2530 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2532 if (rec_argv == NULL)
2536 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2537 rec_argv[j++] = record_args[i];
2539 if (trace->trace_syscalls) {
2540 for (i = 0; i < sc_args_nr; i++)
2541 rec_argv[j++] = sc_args[i];
2543 /* event string may be different for older kernels - e.g., RHEL6 */
2544 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2545 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2546 else if (is_valid_tracepoint("syscalls:sys_enter"))
2547 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2549 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2555 if (trace->trace_pgfaults & TRACE_PFMAJ)
2556 for (i = 0; i < majpf_args_nr; i++)
2557 rec_argv[j++] = majpf_args[i];
2559 if (trace->trace_pgfaults & TRACE_PFMIN)
2560 for (i = 0; i < minpf_args_nr; i++)
2561 rec_argv[j++] = minpf_args[i];
2563 for (i = 0; i < (unsigned int)argc; i++)
2564 rec_argv[j++] = argv[i];
2566 return cmd_record(j, rec_argv);
2569 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2571 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2574 struct perf_evsel *evsel, *tmp;
2575 struct parse_events_error err = { .idx = 0, };
2576 int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2581 evlist__for_each_entry_safe(evlist, evsel, tmp) {
2582 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2585 if (perf_evsel__field(evsel, "pathname")) {
2586 evsel->handler = trace__vfs_getname;
2591 list_del_init(&evsel->node);
2592 evsel->evlist = NULL;
2593 perf_evsel__delete(evsel);
2599 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2601 struct perf_evsel *evsel;
2602 struct perf_event_attr attr = {
2603 .type = PERF_TYPE_SOFTWARE,
2607 attr.config = config;
2608 attr.sample_period = 1;
2610 event_attr_init(&attr);
2612 evsel = perf_evsel__new(&attr);
2614 evsel->handler = trace__pgfault;
2619 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2621 const u32 type = event->header.type;
2622 struct perf_evsel *evsel;
2624 if (type != PERF_RECORD_SAMPLE) {
2625 trace__process_event(trace, trace->host, event, sample);
2629 evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2630 if (evsel == NULL) {
2631 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2635 trace__set_base_time(trace, evsel, sample);
2637 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2638 sample->raw_data == NULL) {
2639 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2640 perf_evsel__name(evsel), sample->tid,
2641 sample->cpu, sample->raw_size);
2643 tracepoint_handler handler = evsel->handler;
2644 handler(trace, evsel, event, sample);
2647 if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2651 static int trace__add_syscall_newtp(struct trace *trace)
2654 struct perf_evlist *evlist = trace->evlist;
2655 struct perf_evsel *sys_enter, *sys_exit;
2657 sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2658 if (sys_enter == NULL)
2661 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2662 goto out_delete_sys_enter;
2664 sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2665 if (sys_exit == NULL)
2666 goto out_delete_sys_enter;
2668 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2669 goto out_delete_sys_exit;
2671 perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2672 perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2674 perf_evlist__add(evlist, sys_enter);
2675 perf_evlist__add(evlist, sys_exit);
2677 if (callchain_param.enabled && !trace->kernel_syscallchains) {
2679 * We're interested only in the user space callchain
2680 * leading to the syscall, allow overriding that for
2681 * debugging reasons using --kernel_syscall_callchains
2683 sys_exit->attr.exclude_callchain_kernel = 1;
2686 trace->syscalls.events.sys_enter = sys_enter;
2687 trace->syscalls.events.sys_exit = sys_exit;
2693 out_delete_sys_exit:
2694 perf_evsel__delete_priv(sys_exit);
2695 out_delete_sys_enter:
2696 perf_evsel__delete_priv(sys_enter);
2700 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2703 struct perf_evsel *sys_exit;
2704 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2705 trace->ev_qualifier_ids.nr,
2706 trace->ev_qualifier_ids.entries);
2711 if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2713 sys_exit = trace->syscalls.events.sys_exit;
2714 err = perf_evsel__append_tp_filter(sys_exit, filter);
2725 #ifdef HAVE_LIBBPF_SUPPORT
2726 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2728 struct syscall *sc = trace__syscall_info(trace, NULL, id);
2734 for (; arg < sc->nr_args; ++arg) {
2735 entry->string_args_len[arg] = 0;
2736 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2737 /* Should be set like strace -s strsize */
2738 entry->string_args_len[arg] = PATH_MAX;
2742 for (; arg < 6; ++arg)
2743 entry->string_args_len[arg] = 0;
2745 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2747 int fd = bpf_map__fd(trace->syscalls.map);
2748 struct bpf_map_syscall_entry value = {
2749 .enabled = !trace->not_ev_qualifier,
2754 for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2755 int key = trace->ev_qualifier_ids.entries[i];
2758 trace__init_bpf_map_syscall_args(trace, key, &value);
2760 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2768 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2770 int fd = bpf_map__fd(trace->syscalls.map);
2771 struct bpf_map_syscall_entry value = {
2776 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2778 trace__init_bpf_map_syscall_args(trace, key, &value);
2780 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2788 static int trace__init_syscalls_bpf_map(struct trace *trace)
2790 bool enabled = true;
2792 if (trace->ev_qualifier_ids.nr)
2793 enabled = trace->not_ev_qualifier;
2795 return __trace__init_syscalls_bpf_map(trace, enabled);
2798 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2803 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2807 #endif // HAVE_LIBBPF_SUPPORT
2809 static int trace__set_ev_qualifier_filter(struct trace *trace)
2811 if (trace->syscalls.map)
2812 return trace__set_ev_qualifier_bpf_filter(trace);
2813 if (trace->syscalls.events.sys_enter)
2814 return trace__set_ev_qualifier_tp_filter(trace);
2818 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2819 size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2822 #ifdef HAVE_LIBBPF_SUPPORT
2824 int map_fd = bpf_map__fd(map);
2827 for (i = 0; i < npids; ++i) {
2828 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2836 static int trace__set_filter_loop_pids(struct trace *trace)
2838 unsigned int nr = 1, err;
2842 struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2844 while (thread && nr < ARRAY_SIZE(pids)) {
2845 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2850 if (!strcmp(thread__comm_str(parent), "sshd") ||
2851 strstarts(thread__comm_str(parent), "gnome-terminal")) {
2852 pids[nr++] = parent->tid;
2858 err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2859 if (!err && trace->filter_pids.map)
2860 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2865 static int trace__set_filter_pids(struct trace *trace)
2869 * Better not use !target__has_task() here because we need to cover the
2870 * case where no threads were specified in the command line, but a
2871 * workload was, and in that case we will fill in the thread_map when
2872 * we fork the workload in perf_evlist__prepare_workload.
2874 if (trace->filter_pids.nr > 0) {
2875 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2876 trace->filter_pids.entries);
2877 if (!err && trace->filter_pids.map) {
2878 err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2879 trace->filter_pids.entries);
2881 } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2882 err = trace__set_filter_loop_pids(trace);
2888 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2890 struct perf_evlist *evlist = trace->evlist;
2891 struct perf_sample sample;
2894 err = perf_evlist__parse_sample(evlist, event, &sample);
2896 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2898 trace__handle_event(trace, event, &sample);
2903 static int __trace__flush_events(struct trace *trace)
2905 u64 first = ordered_events__first_time(&trace->oe.data);
2906 u64 flush = trace->oe.last - NSEC_PER_SEC;
2908 /* Is there some thing to flush.. */
2909 if (first && first < flush)
2910 return ordered_events__flush_time(&trace->oe.data, flush);
2915 static int trace__flush_events(struct trace *trace)
2917 return !trace->sort_events ? 0 : __trace__flush_events(trace);
2920 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2924 if (!trace->sort_events)
2925 return __trace__deliver_event(trace, event);
2927 err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2928 if (err && err != -1)
2931 err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2935 return trace__flush_events(trace);
2938 static int ordered_events__deliver_event(struct ordered_events *oe,
2939 struct ordered_event *event)
2941 struct trace *trace = container_of(oe, struct trace, oe.data);
2943 return __trace__deliver_event(trace, event->event);
2946 static int trace__run(struct trace *trace, int argc, const char **argv)
2948 struct perf_evlist *evlist = trace->evlist;
2949 struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2951 unsigned long before;
2952 const bool forks = argc > 0;
2953 bool draining = false;
2957 if (!trace->raw_augmented_syscalls) {
2958 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2959 goto out_error_raw_syscalls;
2961 if (trace->trace_syscalls)
2962 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2965 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2966 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2967 if (pgfault_maj == NULL)
2969 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2970 perf_evlist__add(evlist, pgfault_maj);
2973 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2974 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2975 if (pgfault_min == NULL)
2977 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2978 perf_evlist__add(evlist, pgfault_min);
2982 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2983 trace__sched_stat_runtime))
2984 goto out_error_sched_stat_runtime;
2987 * If a global cgroup was set, apply it to all the events without an
2988 * explicit cgroup. I.e.:
2990 * trace -G A -e sched:*switch
2992 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2993 * _and_ sched:sched_switch to the 'A' cgroup, while:
2995 * trace -e sched:*switch -G A
2997 * will only set the sched:sched_switch event to the 'A' cgroup, all the
2998 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2999 * a cgroup (on the root cgroup, sys wide, etc).
3003 * trace -G A -e sched:*switch -G B
3005 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3006 * to the 'B' cgroup.
3008 * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3009 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3012 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3014 err = perf_evlist__create_maps(evlist, &trace->opts.target);
3016 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3017 goto out_delete_evlist;
3020 err = trace__symbols_init(trace, evlist);
3022 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3023 goto out_delete_evlist;
3026 perf_evlist__config(evlist, &trace->opts, &callchain_param);
3028 signal(SIGCHLD, sig_handler);
3029 signal(SIGINT, sig_handler);
3032 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3035 fprintf(trace->output, "Couldn't run the workload!\n");
3036 goto out_delete_evlist;
3040 err = perf_evlist__open(evlist);
3042 goto out_error_open;
3044 err = bpf__apply_obj_config();
3046 char errbuf[BUFSIZ];
3048 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3049 pr_err("ERROR: Apply config to BPF failed: %s\n",
3051 goto out_error_open;
3054 err = trace__set_filter_pids(trace);
3058 if (trace->syscalls.map)
3059 trace__init_syscalls_bpf_map(trace);
3061 if (trace->ev_qualifier_ids.nr > 0) {
3062 err = trace__set_ev_qualifier_filter(trace);
3066 if (trace->syscalls.events.sys_exit) {
3067 pr_debug("event qualifier tracepoint filter: %s\n",
3068 trace->syscalls.events.sys_exit->filter);
3072 err = perf_evlist__apply_filters(evlist, &evsel);
3074 goto out_error_apply_filters;
3076 if (trace->dump.map)
3077 bpf_map__fprintf(trace->dump.map, trace->output);
3079 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3081 goto out_error_mmap;
3083 if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3084 perf_evlist__enable(evlist);
3087 perf_evlist__start_workload(evlist);
3089 if (trace->opts.initial_delay) {
3090 usleep(trace->opts.initial_delay * 1000);
3091 perf_evlist__enable(evlist);
3094 trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3095 evlist->threads->nr > 1 ||
3096 perf_evlist__first(evlist)->attr.inherit;
3099 * Now that we already used evsel->attr to ask the kernel to setup the
3100 * events, lets reuse evsel->attr.sample_max_stack as the limit in
3101 * trace__resolve_callchain(), allowing per-event max-stack settings
3102 * to override an explicitly set --max-stack global setting.
3104 evlist__for_each_entry(evlist, evsel) {
3105 if (evsel__has_callchain(evsel) &&
3106 evsel->attr.sample_max_stack == 0)
3107 evsel->attr.sample_max_stack = trace->max_stack;
3110 before = trace->nr_events;
3112 for (i = 0; i < evlist->nr_mmaps; i++) {
3113 union perf_event *event;
3114 struct perf_mmap *md;
3116 md = &evlist->mmap[i];
3117 if (perf_mmap__read_init(md) < 0)
3120 while ((event = perf_mmap__read_event(md)) != NULL) {
3123 err = trace__deliver_event(trace, event);
3127 perf_mmap__consume(md);
3132 if (done && !draining) {
3133 perf_evlist__disable(evlist);
3137 perf_mmap__read_done(md);
3140 if (trace->nr_events == before) {
3141 int timeout = done ? 100 : -1;
3143 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3144 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3149 if (trace__flush_events(trace))
3157 thread__zput(trace->current);
3159 perf_evlist__disable(evlist);
3161 if (trace->sort_events)
3162 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3166 trace__fprintf_thread_summary(trace, trace->output);
3168 if (trace->show_tool_stats) {
3169 fprintf(trace->output, "Stats:\n "
3170 " vfs_getname : %" PRIu64 "\n"
3171 " proc_getname: %" PRIu64 "\n",
3172 trace->stats.vfs_getname,
3173 trace->stats.proc_getname);
3178 trace__symbols__exit(trace);
3180 perf_evlist__delete(evlist);
3181 cgroup__put(trace->cgroup);
3182 trace->evlist = NULL;
3183 trace->live = false;
3186 char errbuf[BUFSIZ];
3188 out_error_sched_stat_runtime:
3189 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3192 out_error_raw_syscalls:
3193 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3197 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3201 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3204 fprintf(trace->output, "%s\n", errbuf);
3205 goto out_delete_evlist;
3207 out_error_apply_filters:
3208 fprintf(trace->output,
3209 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3210 evsel->filter, perf_evsel__name(evsel), errno,
3211 str_error_r(errno, errbuf, sizeof(errbuf)));
3212 goto out_delete_evlist;
3215 fprintf(trace->output, "Not enough memory to run!\n");
3216 goto out_delete_evlist;
3219 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3220 goto out_delete_evlist;
3223 static int trace__replay(struct trace *trace)
3225 const struct perf_evsel_str_handler handlers[] = {
3226 { "probe:vfs_getname", trace__vfs_getname, },
3228 struct perf_data data = {
3230 .mode = PERF_DATA_MODE_READ,
3231 .force = trace->force,
3233 struct perf_session *session;
3234 struct perf_evsel *evsel;
3237 trace->tool.sample = trace__process_sample;
3238 trace->tool.mmap = perf_event__process_mmap;
3239 trace->tool.mmap2 = perf_event__process_mmap2;
3240 trace->tool.comm = perf_event__process_comm;
3241 trace->tool.exit = perf_event__process_exit;
3242 trace->tool.fork = perf_event__process_fork;
3243 trace->tool.attr = perf_event__process_attr;
3244 trace->tool.tracing_data = perf_event__process_tracing_data;
3245 trace->tool.build_id = perf_event__process_build_id;
3246 trace->tool.namespaces = perf_event__process_namespaces;
3248 trace->tool.ordered_events = true;
3249 trace->tool.ordering_requires_timestamps = true;
3251 /* add tid to output */
3252 trace->multiple_threads = true;
3254 session = perf_session__new(&data, false, &trace->tool);
3255 if (session == NULL)
3258 if (trace->opts.target.pid)
3259 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3261 if (trace->opts.target.tid)
3262 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3264 if (symbol__init(&session->header.env) < 0)
3267 trace->host = &session->machines.host;
3269 err = perf_session__set_tracepoints_handlers(session, handlers);
3273 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3274 "raw_syscalls:sys_enter");
3275 /* older kernels have syscalls tp versus raw_syscalls */
3277 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3278 "syscalls:sys_enter");
3281 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3282 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3283 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3287 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3288 "raw_syscalls:sys_exit");
3290 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3291 "syscalls:sys_exit");
3293 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3294 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3295 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3299 evlist__for_each_entry(session->evlist, evsel) {
3300 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3301 (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3302 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3303 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3304 evsel->handler = trace__pgfault;
3309 err = perf_session__process_events(session);
3311 pr_err("Failed to process events, error %d", err);
3313 else if (trace->summary)
3314 trace__fprintf_thread_summary(trace, trace->output);
3317 perf_session__delete(session);
3322 static size_t trace__fprintf_threads_header(FILE *fp)
3326 printed = fprintf(fp, "\n Summary of events:\n\n");
3331 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3332 struct stats *stats;
3337 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3338 struct stats *stats = source->priv;
3340 entry->syscall = source->i;
3341 entry->stats = stats;
3342 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3345 static size_t thread__dump_stats(struct thread_trace *ttrace,
3346 struct trace *trace, FILE *fp)
3351 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3353 if (syscall_stats == NULL)
3356 printed += fprintf(fp, "\n");
3358 printed += fprintf(fp, " syscall calls total min avg max stddev\n");
3359 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
3360 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n");
3362 resort_rb__for_each_entry(nd, syscall_stats) {
3363 struct stats *stats = syscall_stats_entry->stats;
3365 double min = (double)(stats->min) / NSEC_PER_MSEC;
3366 double max = (double)(stats->max) / NSEC_PER_MSEC;
3367 double avg = avg_stats(stats);
3369 u64 n = (u64) stats->n;
3371 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3372 avg /= NSEC_PER_MSEC;
3374 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3375 printed += fprintf(fp, " %-15s", sc->name);
3376 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3377 n, syscall_stats_entry->msecs, min, avg);
3378 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3382 resort_rb__delete(syscall_stats);
3383 printed += fprintf(fp, "\n\n");
3388 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3391 struct thread_trace *ttrace = thread__priv(thread);
3397 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3399 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3400 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3401 printed += fprintf(fp, "%.1f%%", ratio);
3403 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3405 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3407 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3408 else if (fputc('\n', fp) != EOF)
3411 printed += thread__dump_stats(ttrace, trace, fp);
3416 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3418 return ttrace ? ttrace->nr_events : 0;
3421 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3422 struct thread *thread;
3425 entry->thread = rb_entry(nd, struct thread, rb_node);
3428 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3430 size_t printed = trace__fprintf_threads_header(fp);
3434 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3435 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3437 if (threads == NULL) {
3438 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3442 resort_rb__for_each_entry(nd, threads)
3443 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3445 resort_rb__delete(threads);
3450 static int trace__set_duration(const struct option *opt, const char *str,
3451 int unset __maybe_unused)
3453 struct trace *trace = opt->value;
3455 trace->duration_filter = atof(str);
3459 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3460 int unset __maybe_unused)
3464 struct trace *trace = opt->value;
3466 * FIXME: introduce a intarray class, plain parse csv and create a
3467 * { int nr, int entries[] } struct...
3469 struct intlist *list = intlist__new(str);
3474 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3475 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3477 if (trace->filter_pids.entries == NULL)
3480 trace->filter_pids.entries[0] = getpid();
3482 for (i = 1; i < trace->filter_pids.nr; ++i)
3483 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3485 intlist__delete(list);
3491 static int trace__open_output(struct trace *trace, const char *filename)
3495 if (!stat(filename, &st) && st.st_size) {
3496 char oldname[PATH_MAX];
3498 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3500 rename(filename, oldname);
3503 trace->output = fopen(filename, "w");
3505 return trace->output == NULL ? -errno : 0;
3508 static int parse_pagefaults(const struct option *opt, const char *str,
3509 int unset __maybe_unused)
3511 int *trace_pgfaults = opt->value;
3513 if (strcmp(str, "all") == 0)
3514 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3515 else if (strcmp(str, "maj") == 0)
3516 *trace_pgfaults |= TRACE_PFMAJ;
3517 else if (strcmp(str, "min") == 0)
3518 *trace_pgfaults |= TRACE_PFMIN;
3525 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3527 struct perf_evsel *evsel;
3529 evlist__for_each_entry(evlist, evsel)
3530 evsel->handler = handler;
3533 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3535 struct perf_evsel *evsel;
3537 evlist__for_each_entry(evlist, evsel) {
3538 if (evsel->priv || !evsel->tp_format)
3541 if (strcmp(evsel->tp_format->system, "syscalls"))
3544 if (perf_evsel__init_syscall_tp(evsel))
3547 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3548 struct syscall_tp *sc = evsel->priv;
3550 if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3552 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3553 struct syscall_tp *sc = evsel->priv;
3555 if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3564 * XXX: Hackish, just splitting the combined -e+--event (syscalls
3565 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3566 * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3568 * It'd be better to introduce a parse_options() variant that would return a
3569 * list with the terms it didn't match to an event...
3571 static int trace__parse_events_option(const struct option *opt, const char *str,
3572 int unset __maybe_unused)
3574 struct trace *trace = (struct trace *)opt->value;
3575 const char *s = str;
3576 char *sep = NULL, *lists[2] = { NULL, NULL, };
3577 int len = strlen(str) + 1, err = -1, list, idx;
3578 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3579 char group_name[PATH_MAX];
3580 struct syscall_fmt *fmt;
3582 if (strace_groups_dir == NULL)
3587 trace->not_ev_qualifier = true;
3591 if ((sep = strchr(s, ',')) != NULL)
3595 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3596 syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3601 fmt = syscall_fmt__find_by_alias(s);
3606 path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3607 if (access(group_name, R_OK) == 0)
3612 sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3614 lists[list] = malloc(len);
3615 if (lists[list] == NULL)
3617 strcpy(lists[list], s);
3627 if (lists[1] != NULL) {
3628 struct strlist_config slist_config = {
3629 .dirname = strace_groups_dir,
3632 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3633 if (trace->ev_qualifier == NULL) {
3634 fputs("Not enough memory to parse event qualifier", trace->output);
3638 if (trace__validate_ev_qualifier(trace))
3640 trace->trace_syscalls = true;
3646 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3647 "event selector. use 'perf list' to list available events",
3648 parse_events_option);
3649 err = parse_events_option(&o, lists[0], 0);
3658 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3660 struct trace *trace = opt->value;
3662 if (!list_empty(&trace->evlist->entries))
3663 return parse_cgroups(opt, str, unset);
3665 trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3670 static struct bpf_map *bpf__find_map_by_name(const char *name)
3672 struct bpf_object *obj, *tmp;
3674 bpf_object__for_each_safe(obj, tmp) {
3675 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3684 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3686 trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3689 static void trace__set_bpf_map_syscalls(struct trace *trace)
3691 trace->syscalls.map = bpf__find_map_by_name("syscalls");
3694 static int trace__config(const char *var, const char *value, void *arg)
3696 struct trace *trace = arg;
3699 if (!strcmp(var, "trace.add_events")) {
3700 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3701 "event selector. use 'perf list' to list available events",
3702 parse_events_option);
3704 * We can't propagate parse_event_option() return, as it is 1
3705 * for failure while perf_config() expects -1.
3707 if (parse_events_option(&o, value, 0))
3709 } else if (!strcmp(var, "trace.show_timestamp")) {
3710 trace->show_tstamp = perf_config_bool(var, value);
3711 } else if (!strcmp(var, "trace.show_duration")) {
3712 trace->show_duration = perf_config_bool(var, value);
3713 } else if (!strcmp(var, "trace.show_arg_names")) {
3714 trace->show_arg_names = perf_config_bool(var, value);
3715 if (!trace->show_arg_names)
3716 trace->show_zeros = true;
3717 } else if (!strcmp(var, "trace.show_zeros")) {
3718 bool new_show_zeros = perf_config_bool(var, value);
3719 if (!trace->show_arg_names && !new_show_zeros) {
3720 pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3723 trace->show_zeros = new_show_zeros;
3724 } else if (!strcmp(var, "trace.show_prefix")) {
3725 trace->show_string_prefix = perf_config_bool(var, value);
3726 } else if (!strcmp(var, "trace.no_inherit")) {
3727 trace->opts.no_inherit = perf_config_bool(var, value);
3728 } else if (!strcmp(var, "trace.args_alignment")) {
3729 int args_alignment = 0;
3730 if (perf_config_int(&args_alignment, var, value) == 0)
3731 trace->args_alignment = args_alignment;
3737 int cmd_trace(int argc, const char **argv)
3739 const char *trace_usage[] = {
3740 "perf trace [<options>] [<command>]",
3741 "perf trace [<options>] -- <command> [<options>]",
3742 "perf trace record [<options>] [<command>]",
3743 "perf trace record [<options>] -- <command> [<options>]",
3746 struct trace trace = {
3755 .user_freq = UINT_MAX,
3756 .user_interval = ULLONG_MAX,
3757 .no_buffering = true,
3758 .mmap_pages = UINT_MAX,
3762 .show_tstamp = true,
3763 .show_duration = true,
3764 .show_arg_names = true,
3765 .args_alignment = 70,
3766 .trace_syscalls = false,
3767 .kernel_syscallchains = false,
3768 .max_stack = UINT_MAX,
3769 .max_events = ULONG_MAX,
3771 const char *map_dump_str = NULL;
3772 const char *output_name = NULL;
3773 const struct option trace_options[] = {
3774 OPT_CALLBACK('e', "event", &trace, "event",
3775 "event/syscall selector. use 'perf list' to list available events",
3776 trace__parse_events_option),
3777 OPT_BOOLEAN(0, "comm", &trace.show_comm,
3778 "show the thread COMM next to its id"),
3779 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3780 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3781 trace__parse_events_option),
3782 OPT_STRING('o', "output", &output_name, "file", "output file name"),
3783 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3784 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3785 "trace events on existing process id"),
3786 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3787 "trace events on existing thread id"),
3788 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3789 "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3790 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3791 "system-wide collection from all CPUs"),
3792 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3793 "list of cpus to monitor"),
3794 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3795 "child tasks do not inherit counters"),
3796 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3797 "number of mmap data pages",
3798 perf_evlist__parse_mmap_pages),
3799 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3801 OPT_CALLBACK(0, "duration", &trace, "float",
3802 "show only events with duration > N.M ms",
3803 trace__set_duration),
3804 #ifdef HAVE_LIBBPF_SUPPORT
3805 OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3807 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3808 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3809 OPT_BOOLEAN('T', "time", &trace.full_time,
3810 "Show full timestamp, not time relative to first start"),
3811 OPT_BOOLEAN(0, "failure", &trace.failure_only,
3812 "Show only syscalls that failed"),
3813 OPT_BOOLEAN('s', "summary", &trace.summary_only,
3814 "Show only syscall summary with statistics"),
3815 OPT_BOOLEAN('S', "with-summary", &trace.summary,
3816 "Show all syscalls and summary with statistics"),
3817 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3818 "Trace pagefaults", parse_pagefaults, "maj"),
3819 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3820 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3821 OPT_CALLBACK(0, "call-graph", &trace.opts,
3822 "record_mode[,record_size]", record_callchain_help,
3823 &record_parse_callchain_opt),
3824 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3825 "Show the kernel callchains on the syscall exit path"),
3826 OPT_ULONG(0, "max-events", &trace.max_events,
3827 "Set the maximum number of events to print, exit after that is reached. "),
3828 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3829 "Set the minimum stack depth when parsing the callchain, "
3830 "anything below the specified depth will be ignored."),
3831 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3832 "Set the maximum stack depth when parsing the callchain, "
3833 "anything beyond the specified depth will be ignored. "
3834 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3835 OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3836 "Sort batch of events before processing, use if getting out of order events"),
3837 OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3838 "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3839 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3840 "per thread proc mmap processing timeout in ms"),
3841 OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3842 trace__parse_cgroups),
3843 OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3844 "ms to wait before starting measurement after program "
3848 bool __maybe_unused max_stack_user_set = true;
3849 bool mmap_pages_user_set = true;
3850 struct perf_evsel *evsel;
3851 const char * const trace_subcommands[] = { "record", NULL };
3855 signal(SIGSEGV, sighandler_dump_stack);
3856 signal(SIGFPE, sighandler_dump_stack);
3858 trace.evlist = perf_evlist__new();
3859 trace.sctbl = syscalltbl__new();
3861 if (trace.evlist == NULL || trace.sctbl == NULL) {
3862 pr_err("Not enough memory to run!\n");
3867 err = perf_config(trace__config, &trace);
3871 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3872 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3874 if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3875 usage_with_options_msg(trace_usage, trace_options,
3876 "cgroup monitoring only available in system-wide mode");
3879 evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3880 if (IS_ERR(evsel)) {
3881 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3882 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3887 trace.syscalls.events.augmented = evsel;
3888 trace__set_bpf_map_filtered_pids(&trace);
3889 trace__set_bpf_map_syscalls(&trace);
3892 err = bpf__setup_stdout(trace.evlist);
3894 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3895 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3902 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3903 if (trace.dump.map == NULL) {
3904 pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3909 if (trace.trace_pgfaults) {
3910 trace.opts.sample_address = true;
3911 trace.opts.sample_time = true;
3914 if (trace.opts.mmap_pages == UINT_MAX)
3915 mmap_pages_user_set = false;
3917 if (trace.max_stack == UINT_MAX) {
3918 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3919 max_stack_user_set = false;
3922 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3923 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3924 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3928 if (callchain_param.enabled) {
3929 if (!mmap_pages_user_set && geteuid() == 0)
3930 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3932 symbol_conf.use_callchain = true;
3935 if (trace.evlist->nr_entries > 0) {
3936 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3937 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3938 perror("failed to set syscalls:* tracepoint fields");
3943 if (trace.sort_events) {
3944 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3945 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3949 * If we are augmenting syscalls, then combine what we put in the
3950 * __augmented_syscalls__ BPF map with what is in the
3951 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3952 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3954 * We'll switch to look at two BPF maps, one for sys_enter and the
3955 * other for sys_exit when we start augmenting the sys_exit paths with
3956 * buffers that are being copied from kernel to userspace, think 'read'
3959 if (trace.syscalls.events.augmented) {
3960 evlist__for_each_entry(trace.evlist, evsel) {
3961 bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3963 if (raw_syscalls_sys_exit) {
3964 trace.raw_augmented_syscalls = true;
3965 goto init_augmented_syscall_tp;
3968 if (trace.syscalls.events.augmented->priv == NULL &&
3969 strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3970 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3971 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3972 perf_evsel__init_augmented_syscall_tp_args(augmented))
3974 augmented->handler = trace__sys_enter;
3977 if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3978 struct syscall_tp *sc;
3979 init_augmented_syscall_tp:
3980 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3984 * For now with BPF raw_augmented we hook into
3985 * raw_syscalls:sys_enter and there we get all
3986 * 6 syscall args plus the tracepoint common
3987 * fields and the syscall_nr (another long).
3988 * So we check if that is the case and if so
3989 * don't look after the sc->args_size but
3990 * always after the full raw_syscalls:sys_enter
3991 * payload, which is fixed.
3993 * We'll revisit this later to pass
3994 * s->args_size to the BPF augmenter (now
3995 * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3996 * so that it copies only what we need for each
3997 * syscall, like what happens when we use
3998 * syscalls:sys_enter_NAME, so that we reduce
3999 * the kernel/userspace traffic to just what is
4000 * needed for each syscall.
4002 if (trace.raw_augmented_syscalls)
4003 trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4004 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4005 evsel->handler = trace__sys_exit;
4010 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4011 return trace__record(&trace, argc-1, &argv[1]);
4013 /* summary_only implies summary option, but don't overwrite summary if set */
4014 if (trace.summary_only)
4015 trace.summary = trace.summary_only;
4017 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4018 trace.evlist->nr_entries == 0 /* Was --events used? */) {
4019 trace.trace_syscalls = true;
4022 if (output_name != NULL) {
4023 err = trace__open_output(&trace, output_name);
4025 perror("failed to create output file");
4030 err = target__validate(&trace.opts.target);
4032 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4033 fprintf(trace.output, "%s", bf);
4037 err = target__parse_uid(&trace.opts.target);
4039 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4040 fprintf(trace.output, "%s", bf);
4044 if (!argc && target__none(&trace.opts.target))
4045 trace.opts.target.system_wide = true;
4048 err = trace__replay(&trace);
4050 err = trace__run(&trace, argc, argv);
4053 if (output_name != NULL)
4054 fclose(trace.output);