4 * Builtin 'trace' command:
6 * Display a continuously updated trace of any workload, CPU, specific PID,
7 * system wide, etc. Default format is loosely strace like, but any other
8 * event may be specified using --event.
10 * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
12 * Initially based on the 'trace' prototype by Thomas Gleixner:
14 * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
16 * Released under the GPL v2. (and only v2, not any later version)
19 #include <traceevent/event-parse.h>
20 #include <api/fs/tracing_path.h>
22 #include "util/bpf_map.h"
24 #include "util/cgroup.h"
25 #include "util/color.h"
26 #include "util/config.h"
27 #include "util/debug.h"
29 #include "util/event.h"
30 #include "util/evlist.h"
31 #include <subcmd/exec-cmd.h>
32 #include "util/machine.h"
34 #include "util/symbol.h"
35 #include "util/path.h"
36 #include "util/session.h"
37 #include "util/thread.h"
38 #include <subcmd/parse-options.h>
39 #include "util/strlist.h"
40 #include "util/intlist.h"
41 #include "util/thread_map.h"
42 #include "util/stat.h"
43 #include "trace/beauty/beauty.h"
44 #include "trace-event.h"
45 #include "util/parse-events.h"
46 #include "util/bpf-loader.h"
47 #include "callchain.h"
48 #include "print_binary.h"
50 #include "syscalltbl.h"
51 #include "rb_resort.h"
59 #include <linux/err.h>
60 #include <linux/filter.h>
61 #include <linux/kernel.h>
62 #include <linux/random.h>
63 #include <linux/stringify.h>
64 #include <linux/time64.h>
66 #include <sys/sysmacros.h>
68 #include "sane_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 strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
411 for (i = 0; i < sas->nr_entries; ++i) {
412 struct strarray *sa = sas->entries[i];
413 int idx = val - sa->offset;
415 if (idx >= 0 && idx < sa->nr_entries) {
416 if (sa->entries[idx] == NULL)
418 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
422 printed = scnprintf(bf, size, intfmt, val);
424 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
428 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
429 struct syscall_arg *arg)
431 return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
435 #define AT_FDCWD -100
438 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
439 struct syscall_arg *arg)
442 const char *prefix = "AT_FD";
445 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
447 return syscall_arg__scnprintf_fd(bf, size, arg);
450 #define SCA_FDAT syscall_arg__scnprintf_fd_at
452 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
453 struct syscall_arg *arg);
455 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
457 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
459 return scnprintf(bf, size, "%#lx", arg->val);
462 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
465 return scnprintf(bf, size, "NULL");
466 return syscall_arg__scnprintf_hex(bf, size, arg);
469 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
471 return scnprintf(bf, size, "%d", arg->val);
474 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
476 return scnprintf(bf, size, "%ld", arg->val);
479 static const char *bpf_cmd[] = {
480 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
481 "MAP_GET_NEXT_KEY", "PROG_LOAD",
483 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
485 #include "trace/beauty/generated/fsconfig_arrays.c"
487 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
489 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
490 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
492 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
493 static DEFINE_STRARRAY(itimers, "ITIMER_");
495 static const char *keyctl_options[] = {
496 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
497 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
498 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
499 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
500 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
502 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
504 static const char *whences[] = { "SET", "CUR", "END",
512 static DEFINE_STRARRAY(whences, "SEEK_");
514 static const char *fcntl_cmds[] = {
515 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
516 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
517 "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
520 static DEFINE_STRARRAY(fcntl_cmds, "F_");
522 static const char *fcntl_linux_specific_cmds[] = {
523 "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
524 "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
525 "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
528 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
530 static struct strarray *fcntl_cmds_arrays[] = {
531 &strarray__fcntl_cmds,
532 &strarray__fcntl_linux_specific_cmds,
535 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
537 static const char *rlimit_resources[] = {
538 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
539 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
542 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
544 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
545 static DEFINE_STRARRAY(sighow, "SIG_");
547 static const char *clockid[] = {
548 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
549 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
550 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
552 static DEFINE_STRARRAY(clockid, "CLOCK_");
554 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
555 struct syscall_arg *arg)
557 bool show_prefix = arg->show_string_prefix;
558 const char *suffix = "_OK";
562 if (mode == F_OK) /* 0 */
563 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
565 if (mode & n##_OK) { \
566 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
576 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
581 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
583 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
584 struct syscall_arg *arg);
586 #define SCA_FILENAME syscall_arg__scnprintf_filename
588 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
589 struct syscall_arg *arg)
591 bool show_prefix = arg->show_string_prefix;
592 const char *prefix = "O_";
593 int printed = 0, flags = arg->val;
596 if (flags & O_##n) { \
597 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
606 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
611 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
613 #ifndef GRND_NONBLOCK
614 #define GRND_NONBLOCK 0x0001
617 #define GRND_RANDOM 0x0002
620 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
621 struct syscall_arg *arg)
623 bool show_prefix = arg->show_string_prefix;
624 const char *prefix = "GRND_";
625 int printed = 0, flags = arg->val;
628 if (flags & GRND_##n) { \
629 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
630 flags &= ~GRND_##n; \
638 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
643 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
645 #define STRARRAY(name, array) \
646 { .scnprintf = SCA_STRARRAY, \
647 .parm = &strarray__##array, }
649 #include "trace/beauty/arch_errno_names.c"
650 #include "trace/beauty/eventfd.c"
651 #include "trace/beauty/futex_op.c"
652 #include "trace/beauty/futex_val3.c"
653 #include "trace/beauty/mmap.c"
654 #include "trace/beauty/mode_t.c"
655 #include "trace/beauty/msg_flags.c"
656 #include "trace/beauty/open_flags.c"
657 #include "trace/beauty/perf_event_open.c"
658 #include "trace/beauty/pid.c"
659 #include "trace/beauty/sched_policy.c"
660 #include "trace/beauty/seccomp.c"
661 #include "trace/beauty/signum.c"
662 #include "trace/beauty/socket_type.c"
663 #include "trace/beauty/waitid_options.c"
665 struct syscall_arg_fmt {
666 size_t (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
667 unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
673 static struct syscall_fmt {
676 struct syscall_arg_fmt arg[6];
683 .arg = { [1] = { .scnprintf = SCA_ACCMODE, /* mode */ }, }, },
684 { .name = "arch_prctl",
685 .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
686 [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
688 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
690 .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
691 { .name = "brk", .hexret = true,
692 .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
693 { .name = "clock_gettime",
694 .arg = { [0] = STRARRAY(clk_id, clockid), }, },
695 { .name = "clone", .errpid = true, .nr_args = 5,
696 .arg = { [0] = { .name = "flags", .scnprintf = SCA_CLONE_FLAGS, },
697 [1] = { .name = "child_stack", .scnprintf = SCA_HEX, },
698 [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
699 [3] = { .name = "child_tidptr", .scnprintf = SCA_HEX, },
700 [4] = { .name = "tls", .scnprintf = SCA_HEX, }, }, },
702 .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
704 .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
705 { .name = "epoll_ctl",
706 .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
707 { .name = "eventfd2",
708 .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
709 { .name = "fchmodat",
710 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
711 { .name = "fchownat",
712 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
714 .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
715 .parm = &strarrays__fcntl_cmds_arrays,
716 .show_zero = true, },
717 [2] = { .scnprintf = SCA_FCNTL_ARG, /* arg */ }, }, },
719 .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
720 { .name = "fsconfig",
721 .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
723 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
724 [1] = { .scnprintf = SCA_FILENAME, /* path */ },
725 [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
726 { .name = "fstat", .alias = "newfstat", },
727 { .name = "fstatat", .alias = "newfstatat", },
729 .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
730 [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
731 { .name = "futimesat",
732 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
733 { .name = "getitimer",
734 .arg = { [0] = STRARRAY(which, itimers), }, },
735 { .name = "getpid", .errpid = true, },
736 { .name = "getpgid", .errpid = true, },
737 { .name = "getppid", .errpid = true, },
738 { .name = "getrandom",
739 .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
740 { .name = "getrlimit",
741 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
742 { .name = "gettid", .errpid = true, },
745 #if defined(__i386__) || defined(__x86_64__)
747 * FIXME: Make this available to all arches.
749 [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
750 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
752 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
754 { .name = "kcmp", .nr_args = 5,
755 .arg = { [0] = { .name = "pid1", .scnprintf = SCA_PID, },
756 [1] = { .name = "pid2", .scnprintf = SCA_PID, },
757 [2] = { .name = "type", .scnprintf = SCA_KCMP_TYPE, },
758 [3] = { .name = "idx1", .scnprintf = SCA_KCMP_IDX, },
759 [4] = { .name = "idx2", .scnprintf = SCA_KCMP_IDX, }, }, },
761 .arg = { [0] = STRARRAY(option, keyctl_options), }, },
763 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
765 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
767 .arg = { [2] = STRARRAY(whence, whences), }, },
768 { .name = "lstat", .alias = "newlstat", },
770 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
771 [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
773 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
775 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
776 { .name = "mmap", .hexret = true,
777 /* The standard mmap maps to old_mmap on s390x */
778 #if defined(__s390x__)
781 .arg = { [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
782 [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
783 [5] = { .scnprintf = SCA_HEX, /* offset */ }, }, },
785 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
786 [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
787 .mask_val = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
788 { .name = "move_mount",
789 .arg = { [0] = { .scnprintf = SCA_FDAT, /* from_dfd */ },
790 [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
791 [2] = { .scnprintf = SCA_FDAT, /* to_dfd */ },
792 [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
793 [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
794 { .name = "mprotect",
795 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
796 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, }, },
797 { .name = "mq_unlink",
798 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
799 { .name = "mremap", .hexret = true,
800 .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
801 { .name = "name_to_handle_at",
802 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
803 { .name = "newfstatat",
804 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
806 .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
807 { .name = "open_by_handle_at",
808 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
809 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
811 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
812 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
813 { .name = "perf_event_open",
814 .arg = { [2] = { .scnprintf = SCA_INT, /* cpu */ },
815 [3] = { .scnprintf = SCA_FD, /* group_fd */ },
816 [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
818 .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
819 { .name = "pkey_alloc",
820 .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS, /* access_rights */ }, }, },
821 { .name = "pkey_free",
822 .arg = { [0] = { .scnprintf = SCA_INT, /* key */ }, }, },
823 { .name = "pkey_mprotect",
824 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
825 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
826 [3] = { .scnprintf = SCA_INT, /* pkey */ }, }, },
827 { .name = "poll", .timeout = true, },
828 { .name = "ppoll", .timeout = true, },
830 .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
831 [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
832 [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
833 { .name = "pread", .alias = "pread64", },
834 { .name = "preadv", .alias = "pread", },
835 { .name = "prlimit64",
836 .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
837 { .name = "pwrite", .alias = "pwrite64", },
838 { .name = "readlinkat",
839 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
840 { .name = "recvfrom",
841 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
842 { .name = "recvmmsg",
843 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
845 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
846 { .name = "renameat",
847 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
848 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
849 { .name = "renameat2",
850 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
851 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
852 [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
853 { .name = "rt_sigaction",
854 .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
855 { .name = "rt_sigprocmask",
856 .arg = { [0] = STRARRAY(how, sighow), }, },
857 { .name = "rt_sigqueueinfo",
858 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
859 { .name = "rt_tgsigqueueinfo",
860 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
861 { .name = "sched_setscheduler",
862 .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
864 .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP, /* op */ },
865 [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
866 { .name = "select", .timeout = true, },
867 { .name = "sendmmsg",
868 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
870 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
872 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
873 [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
874 { .name = "set_tid_address", .errpid = true, },
875 { .name = "setitimer",
876 .arg = { [0] = STRARRAY(which, itimers), }, },
877 { .name = "setrlimit",
878 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
880 .arg = { [0] = STRARRAY(family, socket_families),
881 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
882 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
883 { .name = "socketpair",
884 .arg = { [0] = STRARRAY(family, socket_families),
885 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
886 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
887 { .name = "stat", .alias = "newstat", },
889 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fdat */ },
890 [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
891 [3] = { .scnprintf = SCA_STATX_MASK, /* mask */ }, }, },
893 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
895 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
896 { .name = "symlinkat",
897 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
899 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
901 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
902 { .name = "umount2", .alias = "umount",
903 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
904 { .name = "uname", .alias = "newuname", },
905 { .name = "unlinkat",
906 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
907 { .name = "utimensat",
908 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
909 { .name = "wait4", .errpid = true,
910 .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
911 { .name = "waitid", .errpid = true,
912 .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
915 static int syscall_fmt__cmp(const void *name, const void *fmtp)
917 const struct syscall_fmt *fmt = fmtp;
918 return strcmp(name, fmt->name);
921 static struct syscall_fmt *syscall_fmt__find(const char *name)
923 const int nmemb = ARRAY_SIZE(syscall_fmts);
924 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
927 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
929 int i, nmemb = ARRAY_SIZE(syscall_fmts);
931 for (i = 0; i < nmemb; ++i) {
932 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
933 return &syscall_fmts[i];
940 * is_exit: is this "exit" or "exit_group"?
941 * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
942 * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
945 struct tep_event *tp_format;
950 struct tep_format_field *args;
952 struct syscall_fmt *fmt;
953 struct syscall_arg_fmt *arg_fmt;
956 struct bpf_map_syscall_entry {
961 * We need to have this 'calculated' boolean because in some cases we really
962 * don't know what is the duration of a syscall, for instance, when we start
963 * a session and some threads are waiting for a syscall to finish, say 'poll',
964 * in which case all we can do is to print "( ? ) for duration and for the
967 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
969 double duration = (double)t / NSEC_PER_MSEC;
970 size_t printed = fprintf(fp, "(");
973 printed += fprintf(fp, " ");
974 else if (duration >= 1.0)
975 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
976 else if (duration >= 0.01)
977 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
979 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
980 return printed + fprintf(fp, "): ");
984 * filename.ptr: The filename char pointer that will be vfs_getname'd
985 * filename.entry_str_pos: Where to insert the string translated from
986 * filename.ptr by the vfs_getname tracepoint/kprobe.
987 * ret_scnprintf: syscall args may set this to a different syscall return
988 * formatter, for instance, fcntl may return fds, file flags, etc.
990 struct thread_trace {
993 unsigned long nr_events;
994 unsigned long pfmaj, pfmin;
997 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1000 short int entry_str_pos;
1002 unsigned int namelen;
1010 struct intlist *syscall_stats;
1013 static struct thread_trace *thread_trace__new(void)
1015 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
1018 ttrace->files.max = -1;
1020 ttrace->syscall_stats = intlist__new(NULL);
1025 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1027 struct thread_trace *ttrace;
1032 if (thread__priv(thread) == NULL)
1033 thread__set_priv(thread, thread_trace__new());
1035 if (thread__priv(thread) == NULL)
1038 ttrace = thread__priv(thread);
1039 ++ttrace->nr_events;
1043 color_fprintf(fp, PERF_COLOR_RED,
1044 "WARNING: not enough memory, dropping samples!\n");
1049 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1050 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1052 struct thread_trace *ttrace = thread__priv(arg->thread);
1054 ttrace->ret_scnprintf = ret_scnprintf;
1057 #define TRACE_PFMAJ (1 << 0)
1058 #define TRACE_PFMIN (1 << 1)
1060 static const size_t trace__entry_str_size = 2048;
1062 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1067 if (fd > ttrace->files.max) {
1068 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1073 if (ttrace->files.max != -1) {
1074 memset(nfiles + ttrace->files.max + 1, 0,
1075 (fd - ttrace->files.max) * sizeof(struct file));
1077 memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1080 ttrace->files.table = nfiles;
1081 ttrace->files.max = fd;
1084 return ttrace->files.table + fd;
1087 struct file *thread__files_entry(struct thread *thread, int fd)
1089 return thread_trace__files_entry(thread__priv(thread), fd);
1092 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1094 struct thread_trace *ttrace = thread__priv(thread);
1095 struct file *file = thread_trace__files_entry(ttrace, fd);
1099 if (stat(pathname, &st) == 0)
1100 file->dev_maj = major(st.st_rdev);
1101 file->pathname = strdup(pathname);
1109 static int thread__read_fd_path(struct thread *thread, int fd)
1111 char linkname[PATH_MAX], pathname[PATH_MAX];
1115 if (thread->pid_ == thread->tid) {
1116 scnprintf(linkname, sizeof(linkname),
1117 "/proc/%d/fd/%d", thread->pid_, fd);
1119 scnprintf(linkname, sizeof(linkname),
1120 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1123 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1126 ret = readlink(linkname, pathname, sizeof(pathname));
1128 if (ret < 0 || ret > st.st_size)
1131 pathname[ret] = '\0';
1132 return trace__set_fd_pathname(thread, fd, pathname);
1135 static const char *thread__fd_path(struct thread *thread, int fd,
1136 struct trace *trace)
1138 struct thread_trace *ttrace = thread__priv(thread);
1146 if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1149 ++trace->stats.proc_getname;
1150 if (thread__read_fd_path(thread, fd))
1154 return ttrace->files.table[fd].pathname;
1157 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1160 size_t printed = scnprintf(bf, size, "%d", fd);
1161 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1164 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1169 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1171 size_t printed = scnprintf(bf, size, "%d", fd);
1172 struct thread *thread = machine__find_thread(trace->host, pid, pid);
1175 const char *path = thread__fd_path(thread, fd, trace);
1178 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1180 thread__put(thread);
1186 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1187 struct syscall_arg *arg)
1190 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1191 struct thread_trace *ttrace = thread__priv(arg->thread);
1193 if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1194 zfree(&ttrace->files.table[fd].pathname);
1199 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1202 struct thread_trace *ttrace = thread__priv(thread);
1204 ttrace->filename.ptr = ptr;
1205 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1208 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1210 struct augmented_arg *augmented_arg = arg->augmented.args;
1212 return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1215 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1216 struct syscall_arg *arg)
1218 unsigned long ptr = arg->val;
1220 if (arg->augmented.args)
1221 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1223 if (!arg->trace->vfs_getname)
1224 return scnprintf(bf, size, "%#x", ptr);
1226 thread__set_filename_pos(arg->thread, bf, ptr);
1230 static bool trace__filter_duration(struct trace *trace, double t)
1232 return t < (trace->duration_filter * NSEC_PER_MSEC);
1235 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1237 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1239 return fprintf(fp, "%10.3f ", ts);
1243 * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1244 * using ttrace->entry_time for a thread that receives a sys_exit without
1245 * first having received a sys_enter ("poll" issued before tracing session
1246 * starts, lost sys_enter exit due to ring buffer overflow).
1248 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1251 return __trace__fprintf_tstamp(trace, tstamp, fp);
1253 return fprintf(fp, " ? ");
1256 static bool done = false;
1257 static bool interrupted = false;
1259 static void sig_handler(int sig)
1262 interrupted = sig == SIGINT;
1265 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1269 if (trace->multiple_threads) {
1270 if (trace->show_comm)
1271 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1272 printed += fprintf(fp, "%d ", thread->tid);
1278 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1279 u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1283 if (trace->show_tstamp)
1284 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1285 if (trace->show_duration)
1286 printed += fprintf_duration(duration, duration_calculated, fp);
1287 return printed + trace__fprintf_comm_tid(trace, thread, fp);
1290 static int trace__process_event(struct trace *trace, struct machine *machine,
1291 union perf_event *event, struct perf_sample *sample)
1295 switch (event->header.type) {
1296 case PERF_RECORD_LOST:
1297 color_fprintf(trace->output, PERF_COLOR_RED,
1298 "LOST %" PRIu64 " events!\n", event->lost.lost);
1299 ret = machine__process_lost_event(machine, event, sample);
1302 ret = machine__process_event(machine, event, sample);
1309 static int trace__tool_process(struct perf_tool *tool,
1310 union perf_event *event,
1311 struct perf_sample *sample,
1312 struct machine *machine)
1314 struct trace *trace = container_of(tool, struct trace, tool);
1315 return trace__process_event(trace, machine, event, sample);
1318 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1320 struct machine *machine = vmachine;
1322 if (machine->kptr_restrict_warned)
1325 if (symbol_conf.kptr_restrict) {
1326 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1327 "Check /proc/sys/kernel/kptr_restrict.\n\n"
1328 "Kernel samples will not be resolved.\n");
1329 machine->kptr_restrict_warned = true;
1333 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1336 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1338 int err = symbol__init(NULL);
1343 trace->host = machine__new_host();
1344 if (trace->host == NULL)
1347 err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1351 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1352 evlist->threads, trace__tool_process, false,
1361 static void trace__symbols__exit(struct trace *trace)
1363 machine__exit(trace->host);
1369 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1373 if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1374 nr_args = sc->fmt->nr_args;
1376 sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1377 if (sc->arg_fmt == NULL)
1380 for (idx = 0; idx < nr_args; ++idx) {
1382 sc->arg_fmt[idx] = sc->fmt->arg[idx];
1385 sc->nr_args = nr_args;
1389 static int syscall__set_arg_fmts(struct syscall *sc)
1391 struct tep_format_field *field, *last_field = NULL;
1394 for (field = sc->args; field; field = field->next, ++idx) {
1397 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1400 if (strcmp(field->type, "const char *") == 0 &&
1401 (strcmp(field->name, "filename") == 0 ||
1402 strcmp(field->name, "path") == 0 ||
1403 strcmp(field->name, "pathname") == 0))
1404 sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1405 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1406 sc->arg_fmt[idx].scnprintf = SCA_PTR;
1407 else if (strcmp(field->type, "pid_t") == 0)
1408 sc->arg_fmt[idx].scnprintf = SCA_PID;
1409 else if (strcmp(field->type, "umode_t") == 0)
1410 sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1411 else if ((strcmp(field->type, "int") == 0 ||
1412 strcmp(field->type, "unsigned int") == 0 ||
1413 strcmp(field->type, "long") == 0) &&
1414 (len = strlen(field->name)) >= 2 &&
1415 strcmp(field->name + len - 2, "fd") == 0) {
1417 * /sys/kernel/tracing/events/syscalls/sys_enter*
1418 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1423 sc->arg_fmt[idx].scnprintf = SCA_FD;
1428 sc->args_size = last_field->offset + last_field->size;
1433 static int trace__read_syscall_info(struct trace *trace, int id)
1437 const char *name = syscalltbl__name(trace->sctbl, id);
1442 if (id > trace->syscalls.max) {
1443 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1445 if (nsyscalls == NULL)
1448 if (trace->syscalls.max != -1) {
1449 memset(nsyscalls + trace->syscalls.max + 1, 0,
1450 (id - trace->syscalls.max) * sizeof(*sc));
1452 memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1455 trace->syscalls.table = nsyscalls;
1456 trace->syscalls.max = id;
1459 sc = trace->syscalls.table + id;
1462 sc->fmt = syscall_fmt__find(sc->name);
1464 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1465 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1467 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1468 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1469 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1472 if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1475 if (IS_ERR(sc->tp_format))
1478 sc->args = sc->tp_format->format.fields;
1480 * We need to check and discard the first variable '__syscall_nr'
1481 * or 'nr' that mean the syscall number. It is needless here.
1482 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1484 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1485 sc->args = sc->args->next;
1489 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1490 sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1492 return syscall__set_arg_fmts(sc);
1495 static int trace__validate_ev_qualifier(struct trace *trace)
1498 size_t nr_allocated;
1499 struct str_node *pos;
1501 trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1502 trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1503 sizeof(trace->ev_qualifier_ids.entries[0]));
1505 if (trace->ev_qualifier_ids.entries == NULL) {
1506 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1512 nr_allocated = trace->ev_qualifier_ids.nr;
1515 strlist__for_each_entry(pos, trace->ev_qualifier) {
1516 const char *sc = pos->s;
1517 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1520 id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1525 fputs("Error:\tInvalid syscall ", trace->output);
1528 fputs(", ", trace->output);
1531 fputs(sc, trace->output);
1534 trace->ev_qualifier_ids.entries[i++] = id;
1535 if (match_next == -1)
1539 id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1542 if (nr_allocated == trace->ev_qualifier_ids.nr) {
1546 entries = realloc(trace->ev_qualifier_ids.entries,
1547 nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1548 if (entries == NULL) {
1550 fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1553 trace->ev_qualifier_ids.entries = entries;
1555 trace->ev_qualifier_ids.nr++;
1556 trace->ev_qualifier_ids.entries[i++] = id;
1561 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1562 "\nHint:\tand: 'man syscalls'\n", trace->output);
1564 zfree(&trace->ev_qualifier_ids.entries);
1565 trace->ev_qualifier_ids.nr = 0;
1572 * args is to be interpreted as a series of longs but we need to handle
1573 * 8-byte unaligned accesses. args points to raw_data within the event
1574 * and raw_data is guaranteed to be 8-byte unaligned because it is
1575 * preceded by raw_size which is a u32. So we need to copy args to a temp
1576 * variable to read it. Most notably this avoids extended load instructions
1577 * on unaligned addresses
1579 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1582 unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1584 memcpy(&val, p, sizeof(val));
1588 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1589 struct syscall_arg *arg)
1591 if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1592 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1594 return scnprintf(bf, size, "arg%d: ", arg->idx);
1598 * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1599 * as mount 'flags' argument that needs ignoring some magic flag, see comment
1600 * in tools/perf/trace/beauty/mount_flags.c
1602 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1604 if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1605 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1610 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1611 struct syscall_arg *arg, unsigned long val)
1613 if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1615 if (sc->arg_fmt[arg->idx].parm)
1616 arg->parm = sc->arg_fmt[arg->idx].parm;
1617 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1619 return scnprintf(bf, size, "%ld", val);
1622 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1623 unsigned char *args, void *augmented_args, int augmented_args_size,
1624 struct trace *trace, struct thread *thread)
1629 struct syscall_arg arg = {
1632 .size = augmented_args_size,
1633 .args = augmented_args,
1639 .show_string_prefix = trace->show_string_prefix,
1641 struct thread_trace *ttrace = thread__priv(thread);
1644 * Things like fcntl will set this in its 'cmd' formatter to pick the
1645 * right formatter for the return value (an fd? file flags?), which is
1646 * not needed for syscalls that always return a given type, say an fd.
1648 ttrace->ret_scnprintf = NULL;
1650 if (sc->args != NULL) {
1651 struct tep_format_field *field;
1653 for (field = sc->args; field;
1654 field = field->next, ++arg.idx, bit <<= 1) {
1658 val = syscall_arg__val(&arg, arg.idx);
1660 * Some syscall args need some mask, most don't and
1661 * return val untouched.
1663 val = syscall__mask_val(sc, &arg, val);
1666 * Suppress this argument if its value is zero and
1667 * and we don't have a string associated in an
1671 !trace->show_zeros &&
1673 (sc->arg_fmt[arg.idx].show_zero ||
1674 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1675 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1676 sc->arg_fmt[arg.idx].parm))
1679 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1681 if (trace->show_arg_names)
1682 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1684 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1686 } else if (IS_ERR(sc->tp_format)) {
1688 * If we managed to read the tracepoint /format file, then we
1689 * may end up not having any args, like with gettid(), so only
1690 * print the raw args when we didn't manage to read it.
1692 while (arg.idx < sc->nr_args) {
1695 val = syscall_arg__val(&arg, arg.idx);
1697 printed += scnprintf(bf + printed, size - printed, ", ");
1698 printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1699 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1709 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1710 union perf_event *event,
1711 struct perf_sample *sample);
1713 static struct syscall *trace__syscall_info(struct trace *trace,
1714 struct perf_evsel *evsel, int id)
1720 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1721 * before that, leaving at a higher verbosity level till that is
1722 * explained. Reproduced with plain ftrace with:
1724 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1725 * grep "NR -1 " /t/trace_pipe
1727 * After generating some load on the machine.
1731 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1732 id, perf_evsel__name(evsel), ++n);
1737 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1738 trace__read_syscall_info(trace, id))
1741 if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1744 return &trace->syscalls.table[id];
1748 fprintf(trace->output, "Problems reading syscall %d", id);
1749 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1750 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1751 fputs(" information\n", trace->output);
1756 static void thread__update_stats(struct thread_trace *ttrace,
1757 int id, struct perf_sample *sample)
1759 struct int_node *inode;
1760 struct stats *stats;
1763 inode = intlist__findnew(ttrace->syscall_stats, id);
1767 stats = inode->priv;
1768 if (stats == NULL) {
1769 stats = malloc(sizeof(struct stats));
1773 inode->priv = stats;
1776 if (ttrace->entry_time && sample->time > ttrace->entry_time)
1777 duration = sample->time - ttrace->entry_time;
1779 update_stats(stats, duration);
1782 static int trace__printf_interrupted_entry(struct trace *trace)
1784 struct thread_trace *ttrace;
1788 if (trace->failure_only || trace->current == NULL)
1791 ttrace = thread__priv(trace->current);
1793 if (!ttrace->entry_pending)
1796 printed = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1797 printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1799 if (len < trace->args_alignment - 4)
1800 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1802 printed += fprintf(trace->output, " ...\n");
1804 ttrace->entry_pending = false;
1805 ++trace->nr_events_printed;
1810 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1811 struct perf_sample *sample, struct thread *thread)
1815 if (trace->print_sample) {
1816 double ts = (double)sample->time / NSEC_PER_MSEC;
1818 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1819 perf_evsel__name(evsel), ts,
1820 thread__comm_str(thread),
1821 sample->pid, sample->tid, sample->cpu);
1827 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1829 void *augmented_args = NULL;
1831 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1832 * and there we get all 6 syscall args plus the tracepoint common fields
1833 * that gets calculated at the start and the syscall_nr (another long).
1834 * So we check if that is the case and if so don't look after the
1835 * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1838 * We'll revisit this later to pass s->args_size to the BPF augmenter
1839 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1840 * copies only what we need for each syscall, like what happens when we
1841 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1842 * traffic to just what is needed for each syscall.
1844 int args_size = raw_augmented_args_size ?: sc->args_size;
1846 *augmented_args_size = sample->raw_size - args_size;
1847 if (*augmented_args_size > 0)
1848 augmented_args = sample->raw_data + args_size;
1850 return augmented_args;
1853 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1854 union perf_event *event __maybe_unused,
1855 struct perf_sample *sample)
1860 struct thread *thread;
1861 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1862 int augmented_args_size = 0;
1863 void *augmented_args = NULL;
1864 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1865 struct thread_trace *ttrace;
1870 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1871 ttrace = thread__trace(thread, trace->output);
1875 trace__fprintf_sample(trace, evsel, sample, thread);
1877 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1879 if (ttrace->entry_str == NULL) {
1880 ttrace->entry_str = malloc(trace__entry_str_size);
1881 if (!ttrace->entry_str)
1885 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1886 trace__printf_interrupted_entry(trace);
1888 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1889 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1890 * this breaks syscall__augmented_args() check for augmented args, as we calculate
1891 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1892 * so when handling, say the openat syscall, we end up getting 6 args for the
1893 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1894 * thinking that the extra 2 u64 args are the augmented filename, so just check
1895 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1897 if (evsel != trace->syscalls.events.sys_enter)
1898 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1899 ttrace->entry_time = sample->time;
1900 msg = ttrace->entry_str;
1901 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1903 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1904 args, augmented_args, augmented_args_size, trace, thread);
1907 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1910 trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1911 printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1912 if (trace->args_alignment > printed)
1913 alignment = trace->args_alignment - printed;
1914 fprintf(trace->output, "%*s= ?\n", alignment, " ");
1917 ttrace->entry_pending = true;
1918 /* See trace__vfs_getname & trace__sys_exit */
1919 ttrace->filename.pending_open = false;
1922 if (trace->current != thread) {
1923 thread__put(trace->current);
1924 trace->current = thread__get(thread);
1928 thread__put(thread);
1932 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1933 struct perf_sample *sample)
1935 struct thread_trace *ttrace;
1936 struct thread *thread;
1937 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1938 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1940 void *args, *augmented_args = NULL;
1941 int augmented_args_size;
1946 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1947 ttrace = thread__trace(thread, trace->output);
1949 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1950 * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1955 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1956 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1957 syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1958 fprintf(trace->output, "%s", msg);
1961 thread__put(thread);
1965 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1966 struct perf_sample *sample,
1967 struct callchain_cursor *cursor)
1969 struct addr_location al;
1970 int max_stack = evsel->attr.sample_max_stack ?
1971 evsel->attr.sample_max_stack :
1975 if (machine__resolve(trace->host, &al, sample) < 0)
1978 err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1979 addr_location__put(&al);
1983 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1985 /* TODO: user-configurable print_opts */
1986 const unsigned int print_opts = EVSEL__PRINT_SYM |
1988 EVSEL__PRINT_UNKNOWN_AS_ADDR;
1990 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1993 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1995 struct perf_env *env = perf_evsel__env(evsel);
1996 const char *arch_name = perf_env__arch(env);
1998 return arch_syscalls__strerrno(arch_name, err);
2001 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
2002 union perf_event *event __maybe_unused,
2003 struct perf_sample *sample)
2007 bool duration_calculated = false;
2008 struct thread *thread;
2009 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2010 int alignment = trace->args_alignment;
2011 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2012 struct thread_trace *ttrace;
2017 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2018 ttrace = thread__trace(thread, trace->output);
2022 trace__fprintf_sample(trace, evsel, sample, thread);
2025 thread__update_stats(ttrace, id, sample);
2027 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2029 if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2030 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2031 ttrace->filename.pending_open = false;
2032 ++trace->stats.vfs_getname;
2035 if (ttrace->entry_time) {
2036 duration = sample->time - ttrace->entry_time;
2037 if (trace__filter_duration(trace, duration))
2039 duration_calculated = true;
2040 } else if (trace->duration_filter)
2043 if (sample->callchain) {
2044 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2045 if (callchain_ret == 0) {
2046 if (callchain_cursor.nr < trace->min_stack)
2052 if (trace->summary_only || (ret >= 0 && trace->failure_only))
2055 trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2057 if (ttrace->entry_pending) {
2058 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2060 printed += fprintf(trace->output, " ... [");
2061 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2063 printed += fprintf(trace->output, "]: %s()", sc->name);
2066 printed++; /* the closing ')' */
2068 if (alignment > printed)
2069 alignment -= printed;
2073 fprintf(trace->output, ")%*s= ", alignment, " ");
2075 if (sc->fmt == NULL) {
2079 fprintf(trace->output, "%ld", ret);
2080 } else if (ret < 0) {
2082 char bf[STRERR_BUFSIZE];
2083 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2084 *e = errno_to_name(evsel, -ret);
2086 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2088 } else if (ret == 0 && sc->fmt->timeout)
2089 fprintf(trace->output, "0 (Timeout)");
2090 else if (ttrace->ret_scnprintf) {
2092 struct syscall_arg arg = {
2097 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2098 ttrace->ret_scnprintf = NULL;
2099 fprintf(trace->output, "%s", bf);
2100 } else if (sc->fmt->hexret)
2101 fprintf(trace->output, "%#lx", ret);
2102 else if (sc->fmt->errpid) {
2103 struct thread *child = machine__find_thread(trace->host, ret, ret);
2105 if (child != NULL) {
2106 fprintf(trace->output, "%ld", ret);
2107 if (child->comm_set)
2108 fprintf(trace->output, " (%s)", thread__comm_str(child));
2114 fputc('\n', trace->output);
2117 * We only consider an 'event' for the sake of --max-events a non-filtered
2118 * sys_enter + sys_exit and other tracepoint events.
2120 if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2123 if (callchain_ret > 0)
2124 trace__fprintf_callchain(trace, sample);
2125 else if (callchain_ret < 0)
2126 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2128 ttrace->entry_pending = false;
2131 thread__put(thread);
2135 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2136 union perf_event *event __maybe_unused,
2137 struct perf_sample *sample)
2139 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2140 struct thread_trace *ttrace;
2141 size_t filename_len, entry_str_len, to_move;
2142 ssize_t remaining_space;
2144 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2149 ttrace = thread__priv(thread);
2153 filename_len = strlen(filename);
2154 if (filename_len == 0)
2157 if (ttrace->filename.namelen < filename_len) {
2158 char *f = realloc(ttrace->filename.name, filename_len + 1);
2163 ttrace->filename.namelen = filename_len;
2164 ttrace->filename.name = f;
2167 strcpy(ttrace->filename.name, filename);
2168 ttrace->filename.pending_open = true;
2170 if (!ttrace->filename.ptr)
2173 entry_str_len = strlen(ttrace->entry_str);
2174 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2175 if (remaining_space <= 0)
2178 if (filename_len > (size_t)remaining_space) {
2179 filename += filename_len - remaining_space;
2180 filename_len = remaining_space;
2183 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2184 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2185 memmove(pos + filename_len, pos, to_move);
2186 memcpy(pos, filename, filename_len);
2188 ttrace->filename.ptr = 0;
2189 ttrace->filename.entry_str_pos = 0;
2191 thread__put(thread);
2196 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2197 union perf_event *event __maybe_unused,
2198 struct perf_sample *sample)
2200 u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2201 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2202 struct thread *thread = machine__findnew_thread(trace->host,
2205 struct thread_trace *ttrace = thread__trace(thread, trace->output);
2210 ttrace->runtime_ms += runtime_ms;
2211 trace->runtime_ms += runtime_ms;
2213 thread__put(thread);
2217 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2219 perf_evsel__strval(evsel, sample, "comm"),
2220 (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2222 perf_evsel__intval(evsel, sample, "vruntime"));
2226 static int bpf_output__printer(enum binary_printer_ops op,
2227 unsigned int val, void *extra __maybe_unused, FILE *fp)
2229 unsigned char ch = (unsigned char)val;
2232 case BINARY_PRINT_CHAR_DATA:
2233 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2234 case BINARY_PRINT_DATA_BEGIN:
2235 case BINARY_PRINT_LINE_BEGIN:
2236 case BINARY_PRINT_ADDR:
2237 case BINARY_PRINT_NUM_DATA:
2238 case BINARY_PRINT_NUM_PAD:
2239 case BINARY_PRINT_SEP:
2240 case BINARY_PRINT_CHAR_PAD:
2241 case BINARY_PRINT_LINE_END:
2242 case BINARY_PRINT_DATA_END:
2250 static void bpf_output__fprintf(struct trace *trace,
2251 struct perf_sample *sample)
2253 binary__fprintf(sample->raw_data, sample->raw_size, 8,
2254 bpf_output__printer, NULL, trace->output);
2255 ++trace->nr_events_printed;
2258 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2259 union perf_event *event __maybe_unused,
2260 struct perf_sample *sample)
2262 struct thread *thread;
2263 int callchain_ret = 0;
2265 * Check if we called perf_evsel__disable(evsel) due to, for instance,
2266 * this event's max_events having been hit and this is an entry coming
2267 * from the ring buffer that we should discard, since the max events
2268 * have already been considered/printed.
2270 if (evsel->disabled)
2273 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2275 if (sample->callchain) {
2276 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2277 if (callchain_ret == 0) {
2278 if (callchain_cursor.nr < trace->min_stack)
2284 trace__printf_interrupted_entry(trace);
2285 trace__fprintf_tstamp(trace, sample->time, trace->output);
2287 if (trace->trace_syscalls && trace->show_duration)
2288 fprintf(trace->output, "( ): ");
2291 trace__fprintf_comm_tid(trace, thread, trace->output);
2293 if (evsel == trace->syscalls.events.augmented) {
2294 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2295 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2298 fprintf(trace->output, "%s(", sc->name);
2299 trace__fprintf_sys_enter(trace, evsel, sample);
2300 fputc(')', trace->output);
2305 * XXX: Not having the associated syscall info or not finding/adding
2306 * the thread should never happen, but if it does...
2307 * fall thru and print it as a bpf_output event.
2311 fprintf(trace->output, "%s:", evsel->name);
2313 if (perf_evsel__is_bpf_output(evsel)) {
2314 bpf_output__fprintf(trace, sample);
2315 } else if (evsel->tp_format) {
2316 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2317 trace__fprintf_sys_enter(trace, evsel, sample)) {
2318 event_format__fprintf(evsel->tp_format, sample->cpu,
2319 sample->raw_data, sample->raw_size,
2321 ++trace->nr_events_printed;
2323 if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2324 perf_evsel__disable(evsel);
2325 perf_evsel__close(evsel);
2331 fprintf(trace->output, "\n");
2333 if (callchain_ret > 0)
2334 trace__fprintf_callchain(trace, sample);
2335 else if (callchain_ret < 0)
2336 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2338 thread__put(thread);
2342 static void print_location(FILE *f, struct perf_sample *sample,
2343 struct addr_location *al,
2344 bool print_dso, bool print_sym)
2347 if ((verbose > 0 || print_dso) && al->map)
2348 fprintf(f, "%s@", al->map->dso->long_name);
2350 if ((verbose > 0 || print_sym) && al->sym)
2351 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2352 al->addr - al->sym->start);
2354 fprintf(f, "0x%" PRIx64, al->addr);
2356 fprintf(f, "0x%" PRIx64, sample->addr);
2359 static int trace__pgfault(struct trace *trace,
2360 struct perf_evsel *evsel,
2361 union perf_event *event __maybe_unused,
2362 struct perf_sample *sample)
2364 struct thread *thread;
2365 struct addr_location al;
2366 char map_type = 'd';
2367 struct thread_trace *ttrace;
2369 int callchain_ret = 0;
2371 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2373 if (sample->callchain) {
2374 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2375 if (callchain_ret == 0) {
2376 if (callchain_cursor.nr < trace->min_stack)
2382 ttrace = thread__trace(thread, trace->output);
2386 if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2391 if (trace->summary_only)
2394 thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2396 trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2398 fprintf(trace->output, "%sfault [",
2399 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2402 print_location(trace->output, sample, &al, false, true);
2404 fprintf(trace->output, "] => ");
2406 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2409 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2417 print_location(trace->output, sample, &al, true, false);
2419 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2421 if (callchain_ret > 0)
2422 trace__fprintf_callchain(trace, sample);
2423 else if (callchain_ret < 0)
2424 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2426 ++trace->nr_events_printed;
2430 thread__put(thread);
2434 static void trace__set_base_time(struct trace *trace,
2435 struct perf_evsel *evsel,
2436 struct perf_sample *sample)
2439 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2440 * and don't use sample->time unconditionally, we may end up having
2441 * some other event in the future without PERF_SAMPLE_TIME for good
2442 * reason, i.e. we may not be interested in its timestamps, just in
2443 * it taking place, picking some piece of information when it
2444 * appears in our event stream (vfs_getname comes to mind).
2446 if (trace->base_time == 0 && !trace->full_time &&
2447 (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2448 trace->base_time = sample->time;
2451 static int trace__process_sample(struct perf_tool *tool,
2452 union perf_event *event,
2453 struct perf_sample *sample,
2454 struct perf_evsel *evsel,
2455 struct machine *machine __maybe_unused)
2457 struct trace *trace = container_of(tool, struct trace, tool);
2458 struct thread *thread;
2461 tracepoint_handler handler = evsel->handler;
2463 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2464 if (thread && thread__is_filtered(thread))
2467 trace__set_base_time(trace, evsel, sample);
2471 handler(trace, evsel, event, sample);
2474 thread__put(thread);
2478 static int trace__record(struct trace *trace, int argc, const char **argv)
2480 unsigned int rec_argc, i, j;
2481 const char **rec_argv;
2482 const char * const record_args[] = {
2489 const char * const sc_args[] = { "-e", };
2490 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2491 const char * const majpf_args[] = { "-e", "major-faults" };
2492 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2493 const char * const minpf_args[] = { "-e", "minor-faults" };
2494 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2496 /* +1 is for the event string below */
2497 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2498 majpf_args_nr + minpf_args_nr + argc;
2499 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2501 if (rec_argv == NULL)
2505 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2506 rec_argv[j++] = record_args[i];
2508 if (trace->trace_syscalls) {
2509 for (i = 0; i < sc_args_nr; i++)
2510 rec_argv[j++] = sc_args[i];
2512 /* event string may be different for older kernels - e.g., RHEL6 */
2513 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2514 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2515 else if (is_valid_tracepoint("syscalls:sys_enter"))
2516 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2518 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2524 if (trace->trace_pgfaults & TRACE_PFMAJ)
2525 for (i = 0; i < majpf_args_nr; i++)
2526 rec_argv[j++] = majpf_args[i];
2528 if (trace->trace_pgfaults & TRACE_PFMIN)
2529 for (i = 0; i < minpf_args_nr; i++)
2530 rec_argv[j++] = minpf_args[i];
2532 for (i = 0; i < (unsigned int)argc; i++)
2533 rec_argv[j++] = argv[i];
2535 return cmd_record(j, rec_argv);
2538 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2540 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2543 struct perf_evsel *evsel, *tmp;
2544 struct parse_events_error err = { .idx = 0, };
2545 int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2550 evlist__for_each_entry_safe(evlist, evsel, tmp) {
2551 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2554 if (perf_evsel__field(evsel, "pathname")) {
2555 evsel->handler = trace__vfs_getname;
2560 list_del_init(&evsel->node);
2561 evsel->evlist = NULL;
2562 perf_evsel__delete(evsel);
2568 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2570 struct perf_evsel *evsel;
2571 struct perf_event_attr attr = {
2572 .type = PERF_TYPE_SOFTWARE,
2576 attr.config = config;
2577 attr.sample_period = 1;
2579 event_attr_init(&attr);
2581 evsel = perf_evsel__new(&attr);
2583 evsel->handler = trace__pgfault;
2588 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2590 const u32 type = event->header.type;
2591 struct perf_evsel *evsel;
2593 if (type != PERF_RECORD_SAMPLE) {
2594 trace__process_event(trace, trace->host, event, sample);
2598 evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2599 if (evsel == NULL) {
2600 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2604 trace__set_base_time(trace, evsel, sample);
2606 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2607 sample->raw_data == NULL) {
2608 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2609 perf_evsel__name(evsel), sample->tid,
2610 sample->cpu, sample->raw_size);
2612 tracepoint_handler handler = evsel->handler;
2613 handler(trace, evsel, event, sample);
2616 if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2620 static int trace__add_syscall_newtp(struct trace *trace)
2623 struct perf_evlist *evlist = trace->evlist;
2624 struct perf_evsel *sys_enter, *sys_exit;
2626 sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2627 if (sys_enter == NULL)
2630 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2631 goto out_delete_sys_enter;
2633 sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2634 if (sys_exit == NULL)
2635 goto out_delete_sys_enter;
2637 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2638 goto out_delete_sys_exit;
2640 perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2641 perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2643 perf_evlist__add(evlist, sys_enter);
2644 perf_evlist__add(evlist, sys_exit);
2646 if (callchain_param.enabled && !trace->kernel_syscallchains) {
2648 * We're interested only in the user space callchain
2649 * leading to the syscall, allow overriding that for
2650 * debugging reasons using --kernel_syscall_callchains
2652 sys_exit->attr.exclude_callchain_kernel = 1;
2655 trace->syscalls.events.sys_enter = sys_enter;
2656 trace->syscalls.events.sys_exit = sys_exit;
2662 out_delete_sys_exit:
2663 perf_evsel__delete_priv(sys_exit);
2664 out_delete_sys_enter:
2665 perf_evsel__delete_priv(sys_enter);
2669 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2672 struct perf_evsel *sys_exit;
2673 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2674 trace->ev_qualifier_ids.nr,
2675 trace->ev_qualifier_ids.entries);
2680 if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2682 sys_exit = trace->syscalls.events.sys_exit;
2683 err = perf_evsel__append_tp_filter(sys_exit, filter);
2694 #ifdef HAVE_LIBBPF_SUPPORT
2695 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2697 int fd = bpf_map__fd(trace->syscalls.map);
2698 struct bpf_map_syscall_entry value = {
2699 .enabled = !trace->not_ev_qualifier,
2704 for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2705 int key = trace->ev_qualifier_ids.entries[i];
2707 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2715 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2717 int fd = bpf_map__fd(trace->syscalls.map);
2718 struct bpf_map_syscall_entry value = {
2723 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2724 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2732 static int trace__init_syscalls_bpf_map(struct trace *trace)
2734 bool enabled = true;
2736 if (trace->ev_qualifier_ids.nr)
2737 enabled = trace->not_ev_qualifier;
2739 return __trace__init_syscalls_bpf_map(trace, enabled);
2742 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2747 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2751 #endif // HAVE_LIBBPF_SUPPORT
2753 static int trace__set_ev_qualifier_filter(struct trace *trace)
2755 if (trace->syscalls.map)
2756 return trace__set_ev_qualifier_bpf_filter(trace);
2757 if (trace->syscalls.events.sys_enter)
2758 return trace__set_ev_qualifier_tp_filter(trace);
2762 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2763 size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2766 #ifdef HAVE_LIBBPF_SUPPORT
2768 int map_fd = bpf_map__fd(map);
2771 for (i = 0; i < npids; ++i) {
2772 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2780 static int trace__set_filter_loop_pids(struct trace *trace)
2782 unsigned int nr = 1, err;
2786 struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2788 while (thread && nr < ARRAY_SIZE(pids)) {
2789 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2794 if (!strcmp(thread__comm_str(parent), "sshd") ||
2795 strstarts(thread__comm_str(parent), "gnome-terminal")) {
2796 pids[nr++] = parent->tid;
2802 err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2803 if (!err && trace->filter_pids.map)
2804 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2809 static int trace__set_filter_pids(struct trace *trace)
2813 * Better not use !target__has_task() here because we need to cover the
2814 * case where no threads were specified in the command line, but a
2815 * workload was, and in that case we will fill in the thread_map when
2816 * we fork the workload in perf_evlist__prepare_workload.
2818 if (trace->filter_pids.nr > 0) {
2819 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2820 trace->filter_pids.entries);
2821 if (!err && trace->filter_pids.map) {
2822 err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2823 trace->filter_pids.entries);
2825 } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2826 err = trace__set_filter_loop_pids(trace);
2832 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2834 struct perf_evlist *evlist = trace->evlist;
2835 struct perf_sample sample;
2838 err = perf_evlist__parse_sample(evlist, event, &sample);
2840 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2842 trace__handle_event(trace, event, &sample);
2847 static int __trace__flush_events(struct trace *trace)
2849 u64 first = ordered_events__first_time(&trace->oe.data);
2850 u64 flush = trace->oe.last - NSEC_PER_SEC;
2852 /* Is there some thing to flush.. */
2853 if (first && first < flush)
2854 return ordered_events__flush_time(&trace->oe.data, flush);
2859 static int trace__flush_events(struct trace *trace)
2861 return !trace->sort_events ? 0 : __trace__flush_events(trace);
2864 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2868 if (!trace->sort_events)
2869 return __trace__deliver_event(trace, event);
2871 err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2872 if (err && err != -1)
2875 err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2879 return trace__flush_events(trace);
2882 static int ordered_events__deliver_event(struct ordered_events *oe,
2883 struct ordered_event *event)
2885 struct trace *trace = container_of(oe, struct trace, oe.data);
2887 return __trace__deliver_event(trace, event->event);
2890 static int trace__run(struct trace *trace, int argc, const char **argv)
2892 struct perf_evlist *evlist = trace->evlist;
2893 struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2895 unsigned long before;
2896 const bool forks = argc > 0;
2897 bool draining = false;
2901 if (!trace->raw_augmented_syscalls) {
2902 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2903 goto out_error_raw_syscalls;
2905 if (trace->trace_syscalls)
2906 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2909 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2910 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2911 if (pgfault_maj == NULL)
2913 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2914 perf_evlist__add(evlist, pgfault_maj);
2917 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2918 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2919 if (pgfault_min == NULL)
2921 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2922 perf_evlist__add(evlist, pgfault_min);
2926 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2927 trace__sched_stat_runtime))
2928 goto out_error_sched_stat_runtime;
2931 * If a global cgroup was set, apply it to all the events without an
2932 * explicit cgroup. I.e.:
2934 * trace -G A -e sched:*switch
2936 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2937 * _and_ sched:sched_switch to the 'A' cgroup, while:
2939 * trace -e sched:*switch -G A
2941 * will only set the sched:sched_switch event to the 'A' cgroup, all the
2942 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2943 * a cgroup (on the root cgroup, sys wide, etc).
2947 * trace -G A -e sched:*switch -G B
2949 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2950 * to the 'B' cgroup.
2952 * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2953 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2956 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2958 err = perf_evlist__create_maps(evlist, &trace->opts.target);
2960 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2961 goto out_delete_evlist;
2964 err = trace__symbols_init(trace, evlist);
2966 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2967 goto out_delete_evlist;
2970 perf_evlist__config(evlist, &trace->opts, &callchain_param);
2972 signal(SIGCHLD, sig_handler);
2973 signal(SIGINT, sig_handler);
2976 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2979 fprintf(trace->output, "Couldn't run the workload!\n");
2980 goto out_delete_evlist;
2984 err = perf_evlist__open(evlist);
2986 goto out_error_open;
2988 err = bpf__apply_obj_config();
2990 char errbuf[BUFSIZ];
2992 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2993 pr_err("ERROR: Apply config to BPF failed: %s\n",
2995 goto out_error_open;
2998 err = trace__set_filter_pids(trace);
3002 if (trace->syscalls.map)
3003 trace__init_syscalls_bpf_map(trace);
3005 if (trace->ev_qualifier_ids.nr > 0) {
3006 err = trace__set_ev_qualifier_filter(trace);
3010 if (trace->syscalls.events.sys_exit) {
3011 pr_debug("event qualifier tracepoint filter: %s\n",
3012 trace->syscalls.events.sys_exit->filter);
3016 err = perf_evlist__apply_filters(evlist, &evsel);
3018 goto out_error_apply_filters;
3020 if (trace->dump.map)
3021 bpf_map__fprintf(trace->dump.map, trace->output);
3023 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3025 goto out_error_mmap;
3027 if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3028 perf_evlist__enable(evlist);
3031 perf_evlist__start_workload(evlist);
3033 if (trace->opts.initial_delay) {
3034 usleep(trace->opts.initial_delay * 1000);
3035 perf_evlist__enable(evlist);
3038 trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3039 evlist->threads->nr > 1 ||
3040 perf_evlist__first(evlist)->attr.inherit;
3043 * Now that we already used evsel->attr to ask the kernel to setup the
3044 * events, lets reuse evsel->attr.sample_max_stack as the limit in
3045 * trace__resolve_callchain(), allowing per-event max-stack settings
3046 * to override an explicitly set --max-stack global setting.
3048 evlist__for_each_entry(evlist, evsel) {
3049 if (evsel__has_callchain(evsel) &&
3050 evsel->attr.sample_max_stack == 0)
3051 evsel->attr.sample_max_stack = trace->max_stack;
3054 before = trace->nr_events;
3056 for (i = 0; i < evlist->nr_mmaps; i++) {
3057 union perf_event *event;
3058 struct perf_mmap *md;
3060 md = &evlist->mmap[i];
3061 if (perf_mmap__read_init(md) < 0)
3064 while ((event = perf_mmap__read_event(md)) != NULL) {
3067 err = trace__deliver_event(trace, event);
3071 perf_mmap__consume(md);
3076 if (done && !draining) {
3077 perf_evlist__disable(evlist);
3081 perf_mmap__read_done(md);
3084 if (trace->nr_events == before) {
3085 int timeout = done ? 100 : -1;
3087 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3088 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3093 if (trace__flush_events(trace))
3101 thread__zput(trace->current);
3103 perf_evlist__disable(evlist);
3105 if (trace->sort_events)
3106 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3110 trace__fprintf_thread_summary(trace, trace->output);
3112 if (trace->show_tool_stats) {
3113 fprintf(trace->output, "Stats:\n "
3114 " vfs_getname : %" PRIu64 "\n"
3115 " proc_getname: %" PRIu64 "\n",
3116 trace->stats.vfs_getname,
3117 trace->stats.proc_getname);
3122 trace__symbols__exit(trace);
3124 perf_evlist__delete(evlist);
3125 cgroup__put(trace->cgroup);
3126 trace->evlist = NULL;
3127 trace->live = false;
3130 char errbuf[BUFSIZ];
3132 out_error_sched_stat_runtime:
3133 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3136 out_error_raw_syscalls:
3137 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3141 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3145 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3148 fprintf(trace->output, "%s\n", errbuf);
3149 goto out_delete_evlist;
3151 out_error_apply_filters:
3152 fprintf(trace->output,
3153 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3154 evsel->filter, perf_evsel__name(evsel), errno,
3155 str_error_r(errno, errbuf, sizeof(errbuf)));
3156 goto out_delete_evlist;
3159 fprintf(trace->output, "Not enough memory to run!\n");
3160 goto out_delete_evlist;
3163 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3164 goto out_delete_evlist;
3167 static int trace__replay(struct trace *trace)
3169 const struct perf_evsel_str_handler handlers[] = {
3170 { "probe:vfs_getname", trace__vfs_getname, },
3172 struct perf_data data = {
3174 .mode = PERF_DATA_MODE_READ,
3175 .force = trace->force,
3177 struct perf_session *session;
3178 struct perf_evsel *evsel;
3181 trace->tool.sample = trace__process_sample;
3182 trace->tool.mmap = perf_event__process_mmap;
3183 trace->tool.mmap2 = perf_event__process_mmap2;
3184 trace->tool.comm = perf_event__process_comm;
3185 trace->tool.exit = perf_event__process_exit;
3186 trace->tool.fork = perf_event__process_fork;
3187 trace->tool.attr = perf_event__process_attr;
3188 trace->tool.tracing_data = perf_event__process_tracing_data;
3189 trace->tool.build_id = perf_event__process_build_id;
3190 trace->tool.namespaces = perf_event__process_namespaces;
3192 trace->tool.ordered_events = true;
3193 trace->tool.ordering_requires_timestamps = true;
3195 /* add tid to output */
3196 trace->multiple_threads = true;
3198 session = perf_session__new(&data, false, &trace->tool);
3199 if (session == NULL)
3202 if (trace->opts.target.pid)
3203 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3205 if (trace->opts.target.tid)
3206 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3208 if (symbol__init(&session->header.env) < 0)
3211 trace->host = &session->machines.host;
3213 err = perf_session__set_tracepoints_handlers(session, handlers);
3217 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3218 "raw_syscalls:sys_enter");
3219 /* older kernels have syscalls tp versus raw_syscalls */
3221 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3222 "syscalls:sys_enter");
3225 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3226 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3227 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3231 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3232 "raw_syscalls:sys_exit");
3234 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3235 "syscalls:sys_exit");
3237 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3238 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3239 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3243 evlist__for_each_entry(session->evlist, evsel) {
3244 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3245 (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3246 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3247 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3248 evsel->handler = trace__pgfault;
3253 err = perf_session__process_events(session);
3255 pr_err("Failed to process events, error %d", err);
3257 else if (trace->summary)
3258 trace__fprintf_thread_summary(trace, trace->output);
3261 perf_session__delete(session);
3266 static size_t trace__fprintf_threads_header(FILE *fp)
3270 printed = fprintf(fp, "\n Summary of events:\n\n");
3275 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3276 struct stats *stats;
3281 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3282 struct stats *stats = source->priv;
3284 entry->syscall = source->i;
3285 entry->stats = stats;
3286 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3289 static size_t thread__dump_stats(struct thread_trace *ttrace,
3290 struct trace *trace, FILE *fp)
3295 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3297 if (syscall_stats == NULL)
3300 printed += fprintf(fp, "\n");
3302 printed += fprintf(fp, " syscall calls total min avg max stddev\n");
3303 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
3304 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n");
3306 resort_rb__for_each_entry(nd, syscall_stats) {
3307 struct stats *stats = syscall_stats_entry->stats;
3309 double min = (double)(stats->min) / NSEC_PER_MSEC;
3310 double max = (double)(stats->max) / NSEC_PER_MSEC;
3311 double avg = avg_stats(stats);
3313 u64 n = (u64) stats->n;
3315 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3316 avg /= NSEC_PER_MSEC;
3318 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3319 printed += fprintf(fp, " %-15s", sc->name);
3320 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3321 n, syscall_stats_entry->msecs, min, avg);
3322 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3326 resort_rb__delete(syscall_stats);
3327 printed += fprintf(fp, "\n\n");
3332 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3335 struct thread_trace *ttrace = thread__priv(thread);
3341 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3343 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3344 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3345 printed += fprintf(fp, "%.1f%%", ratio);
3347 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3349 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3351 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3352 else if (fputc('\n', fp) != EOF)
3355 printed += thread__dump_stats(ttrace, trace, fp);
3360 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3362 return ttrace ? ttrace->nr_events : 0;
3365 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3366 struct thread *thread;
3369 entry->thread = rb_entry(nd, struct thread, rb_node);
3372 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3374 size_t printed = trace__fprintf_threads_header(fp);
3378 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3379 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3381 if (threads == NULL) {
3382 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3386 resort_rb__for_each_entry(nd, threads)
3387 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3389 resort_rb__delete(threads);
3394 static int trace__set_duration(const struct option *opt, const char *str,
3395 int unset __maybe_unused)
3397 struct trace *trace = opt->value;
3399 trace->duration_filter = atof(str);
3403 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3404 int unset __maybe_unused)
3408 struct trace *trace = opt->value;
3410 * FIXME: introduce a intarray class, plain parse csv and create a
3411 * { int nr, int entries[] } struct...
3413 struct intlist *list = intlist__new(str);
3418 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3419 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3421 if (trace->filter_pids.entries == NULL)
3424 trace->filter_pids.entries[0] = getpid();
3426 for (i = 1; i < trace->filter_pids.nr; ++i)
3427 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3429 intlist__delete(list);
3435 static int trace__open_output(struct trace *trace, const char *filename)
3439 if (!stat(filename, &st) && st.st_size) {
3440 char oldname[PATH_MAX];
3442 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3444 rename(filename, oldname);
3447 trace->output = fopen(filename, "w");
3449 return trace->output == NULL ? -errno : 0;
3452 static int parse_pagefaults(const struct option *opt, const char *str,
3453 int unset __maybe_unused)
3455 int *trace_pgfaults = opt->value;
3457 if (strcmp(str, "all") == 0)
3458 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3459 else if (strcmp(str, "maj") == 0)
3460 *trace_pgfaults |= TRACE_PFMAJ;
3461 else if (strcmp(str, "min") == 0)
3462 *trace_pgfaults |= TRACE_PFMIN;
3469 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3471 struct perf_evsel *evsel;
3473 evlist__for_each_entry(evlist, evsel)
3474 evsel->handler = handler;
3477 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3479 struct perf_evsel *evsel;
3481 evlist__for_each_entry(evlist, evsel) {
3482 if (evsel->priv || !evsel->tp_format)
3485 if (strcmp(evsel->tp_format->system, "syscalls"))
3488 if (perf_evsel__init_syscall_tp(evsel))
3491 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3492 struct syscall_tp *sc = evsel->priv;
3494 if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3496 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3497 struct syscall_tp *sc = evsel->priv;
3499 if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3508 * XXX: Hackish, just splitting the combined -e+--event (syscalls
3509 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3510 * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3512 * It'd be better to introduce a parse_options() variant that would return a
3513 * list with the terms it didn't match to an event...
3515 static int trace__parse_events_option(const struct option *opt, const char *str,
3516 int unset __maybe_unused)
3518 struct trace *trace = (struct trace *)opt->value;
3519 const char *s = str;
3520 char *sep = NULL, *lists[2] = { NULL, NULL, };
3521 int len = strlen(str) + 1, err = -1, list, idx;
3522 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3523 char group_name[PATH_MAX];
3524 struct syscall_fmt *fmt;
3526 if (strace_groups_dir == NULL)
3531 trace->not_ev_qualifier = true;
3535 if ((sep = strchr(s, ',')) != NULL)
3539 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3540 syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3545 fmt = syscall_fmt__find_by_alias(s);
3550 path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3551 if (access(group_name, R_OK) == 0)
3556 sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3558 lists[list] = malloc(len);
3559 if (lists[list] == NULL)
3561 strcpy(lists[list], s);
3571 if (lists[1] != NULL) {
3572 struct strlist_config slist_config = {
3573 .dirname = strace_groups_dir,
3576 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3577 if (trace->ev_qualifier == NULL) {
3578 fputs("Not enough memory to parse event qualifier", trace->output);
3582 if (trace__validate_ev_qualifier(trace))
3584 trace->trace_syscalls = true;
3590 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3591 "event selector. use 'perf list' to list available events",
3592 parse_events_option);
3593 err = parse_events_option(&o, lists[0], 0);
3602 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3604 struct trace *trace = opt->value;
3606 if (!list_empty(&trace->evlist->entries))
3607 return parse_cgroups(opt, str, unset);
3609 trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3614 static struct bpf_map *bpf__find_map_by_name(const char *name)
3616 struct bpf_object *obj, *tmp;
3618 bpf_object__for_each_safe(obj, tmp) {
3619 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3628 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3630 trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3633 static void trace__set_bpf_map_syscalls(struct trace *trace)
3635 trace->syscalls.map = bpf__find_map_by_name("syscalls");
3638 static int trace__config(const char *var, const char *value, void *arg)
3640 struct trace *trace = arg;
3643 if (!strcmp(var, "trace.add_events")) {
3644 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3645 "event selector. use 'perf list' to list available events",
3646 parse_events_option);
3647 err = parse_events_option(&o, value, 0);
3648 } else if (!strcmp(var, "trace.show_timestamp")) {
3649 trace->show_tstamp = perf_config_bool(var, value);
3650 } else if (!strcmp(var, "trace.show_duration")) {
3651 trace->show_duration = perf_config_bool(var, value);
3652 } else if (!strcmp(var, "trace.show_arg_names")) {
3653 trace->show_arg_names = perf_config_bool(var, value);
3654 if (!trace->show_arg_names)
3655 trace->show_zeros = true;
3656 } else if (!strcmp(var, "trace.show_zeros")) {
3657 bool new_show_zeros = perf_config_bool(var, value);
3658 if (!trace->show_arg_names && !new_show_zeros) {
3659 pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3662 trace->show_zeros = new_show_zeros;
3663 } else if (!strcmp(var, "trace.show_prefix")) {
3664 trace->show_string_prefix = perf_config_bool(var, value);
3665 } else if (!strcmp(var, "trace.no_inherit")) {
3666 trace->opts.no_inherit = perf_config_bool(var, value);
3667 } else if (!strcmp(var, "trace.args_alignment")) {
3668 int args_alignment = 0;
3669 if (perf_config_int(&args_alignment, var, value) == 0)
3670 trace->args_alignment = args_alignment;
3676 int cmd_trace(int argc, const char **argv)
3678 const char *trace_usage[] = {
3679 "perf trace [<options>] [<command>]",
3680 "perf trace [<options>] -- <command> [<options>]",
3681 "perf trace record [<options>] [<command>]",
3682 "perf trace record [<options>] -- <command> [<options>]",
3685 struct trace trace = {
3694 .user_freq = UINT_MAX,
3695 .user_interval = ULLONG_MAX,
3696 .no_buffering = true,
3697 .mmap_pages = UINT_MAX,
3701 .show_tstamp = true,
3702 .show_duration = true,
3703 .show_arg_names = true,
3704 .args_alignment = 70,
3705 .trace_syscalls = false,
3706 .kernel_syscallchains = false,
3707 .max_stack = UINT_MAX,
3708 .max_events = ULONG_MAX,
3710 const char *map_dump_str = NULL;
3711 const char *output_name = NULL;
3712 const struct option trace_options[] = {
3713 OPT_CALLBACK('e', "event", &trace, "event",
3714 "event/syscall selector. use 'perf list' to list available events",
3715 trace__parse_events_option),
3716 OPT_BOOLEAN(0, "comm", &trace.show_comm,
3717 "show the thread COMM next to its id"),
3718 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3719 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3720 trace__parse_events_option),
3721 OPT_STRING('o', "output", &output_name, "file", "output file name"),
3722 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3723 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3724 "trace events on existing process id"),
3725 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3726 "trace events on existing thread id"),
3727 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3728 "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3729 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3730 "system-wide collection from all CPUs"),
3731 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3732 "list of cpus to monitor"),
3733 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3734 "child tasks do not inherit counters"),
3735 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3736 "number of mmap data pages",
3737 perf_evlist__parse_mmap_pages),
3738 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3740 OPT_CALLBACK(0, "duration", &trace, "float",
3741 "show only events with duration > N.M ms",
3742 trace__set_duration),
3743 #ifdef HAVE_LIBBPF_SUPPORT
3744 OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3746 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3747 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3748 OPT_BOOLEAN('T', "time", &trace.full_time,
3749 "Show full timestamp, not time relative to first start"),
3750 OPT_BOOLEAN(0, "failure", &trace.failure_only,
3751 "Show only syscalls that failed"),
3752 OPT_BOOLEAN('s', "summary", &trace.summary_only,
3753 "Show only syscall summary with statistics"),
3754 OPT_BOOLEAN('S', "with-summary", &trace.summary,
3755 "Show all syscalls and summary with statistics"),
3756 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3757 "Trace pagefaults", parse_pagefaults, "maj"),
3758 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3759 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3760 OPT_CALLBACK(0, "call-graph", &trace.opts,
3761 "record_mode[,record_size]", record_callchain_help,
3762 &record_parse_callchain_opt),
3763 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3764 "Show the kernel callchains on the syscall exit path"),
3765 OPT_ULONG(0, "max-events", &trace.max_events,
3766 "Set the maximum number of events to print, exit after that is reached. "),
3767 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3768 "Set the minimum stack depth when parsing the callchain, "
3769 "anything below the specified depth will be ignored."),
3770 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3771 "Set the maximum stack depth when parsing the callchain, "
3772 "anything beyond the specified depth will be ignored. "
3773 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3774 OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3775 "Sort batch of events before processing, use if getting out of order events"),
3776 OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3777 "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3778 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3779 "per thread proc mmap processing timeout in ms"),
3780 OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3781 trace__parse_cgroups),
3782 OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3783 "ms to wait before starting measurement after program "
3787 bool __maybe_unused max_stack_user_set = true;
3788 bool mmap_pages_user_set = true;
3789 struct perf_evsel *evsel;
3790 const char * const trace_subcommands[] = { "record", NULL };
3794 signal(SIGSEGV, sighandler_dump_stack);
3795 signal(SIGFPE, sighandler_dump_stack);
3797 trace.evlist = perf_evlist__new();
3798 trace.sctbl = syscalltbl__new();
3800 if (trace.evlist == NULL || trace.sctbl == NULL) {
3801 pr_err("Not enough memory to run!\n");
3806 err = perf_config(trace__config, &trace);
3810 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3811 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3813 if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3814 usage_with_options_msg(trace_usage, trace_options,
3815 "cgroup monitoring only available in system-wide mode");
3818 evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3819 if (IS_ERR(evsel)) {
3820 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3821 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3826 trace.syscalls.events.augmented = evsel;
3827 trace__set_bpf_map_filtered_pids(&trace);
3828 trace__set_bpf_map_syscalls(&trace);
3831 err = bpf__setup_stdout(trace.evlist);
3833 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3834 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3841 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3842 if (trace.dump.map == NULL) {
3843 pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3848 if (trace.trace_pgfaults) {
3849 trace.opts.sample_address = true;
3850 trace.opts.sample_time = true;
3853 if (trace.opts.mmap_pages == UINT_MAX)
3854 mmap_pages_user_set = false;
3856 if (trace.max_stack == UINT_MAX) {
3857 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3858 max_stack_user_set = false;
3861 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3862 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3863 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3867 if (callchain_param.enabled) {
3868 if (!mmap_pages_user_set && geteuid() == 0)
3869 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3871 symbol_conf.use_callchain = true;
3874 if (trace.evlist->nr_entries > 0) {
3875 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3876 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3877 perror("failed to set syscalls:* tracepoint fields");
3882 if (trace.sort_events) {
3883 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3884 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3888 * If we are augmenting syscalls, then combine what we put in the
3889 * __augmented_syscalls__ BPF map with what is in the
3890 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3891 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3893 * We'll switch to look at two BPF maps, one for sys_enter and the
3894 * other for sys_exit when we start augmenting the sys_exit paths with
3895 * buffers that are being copied from kernel to userspace, think 'read'
3898 if (trace.syscalls.events.augmented) {
3899 evlist__for_each_entry(trace.evlist, evsel) {
3900 bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3902 if (raw_syscalls_sys_exit) {
3903 trace.raw_augmented_syscalls = true;
3904 goto init_augmented_syscall_tp;
3907 if (trace.syscalls.events.augmented->priv == NULL &&
3908 strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3909 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3910 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3911 perf_evsel__init_augmented_syscall_tp_args(augmented))
3913 augmented->handler = trace__sys_enter;
3916 if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3917 struct syscall_tp *sc;
3918 init_augmented_syscall_tp:
3919 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3923 * For now with BPF raw_augmented we hook into
3924 * raw_syscalls:sys_enter and there we get all
3925 * 6 syscall args plus the tracepoint common
3926 * fields and the syscall_nr (another long).
3927 * So we check if that is the case and if so
3928 * don't look after the sc->args_size but
3929 * always after the full raw_syscalls:sys_enter
3930 * payload, which is fixed.
3932 * We'll revisit this later to pass
3933 * s->args_size to the BPF augmenter (now
3934 * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3935 * so that it copies only what we need for each
3936 * syscall, like what happens when we use
3937 * syscalls:sys_enter_NAME, so that we reduce
3938 * the kernel/userspace traffic to just what is
3939 * needed for each syscall.
3941 if (trace.raw_augmented_syscalls)
3942 trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3943 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3944 evsel->handler = trace__sys_exit;
3949 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3950 return trace__record(&trace, argc-1, &argv[1]);
3952 /* summary_only implies summary option, but don't overwrite summary if set */
3953 if (trace.summary_only)
3954 trace.summary = trace.summary_only;
3956 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3957 trace.evlist->nr_entries == 0 /* Was --events used? */) {
3958 trace.trace_syscalls = true;
3961 if (output_name != NULL) {
3962 err = trace__open_output(&trace, output_name);
3964 perror("failed to create output file");
3969 err = target__validate(&trace.opts.target);
3971 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3972 fprintf(trace.output, "%s", bf);
3976 err = target__parse_uid(&trace.opts.target);
3978 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3979 fprintf(trace.output, "%s", bf);
3983 if (!argc && target__none(&trace.opts.target))
3984 trace.opts.target.system_wide = true;
3987 err = trace__replay(&trace);
3989 err = trace__run(&trace, argc, argv);
3992 if (output_name != NULL)
3993 fclose(trace.output);