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'")
17 #include "util/record.h"
18 #include <traceevent/event-parse.h>
19 #include <api/fs/tracing_path.h>
21 #include "util/bpf_map.h"
22 #include "util/rlimit.h"
24 #include "util/cgroup.h"
25 #include "util/color.h"
26 #include "util/config.h"
27 #include "util/debug.h"
30 #include "util/event.h"
31 #include "util/synthetic-events.h"
32 #include "util/evlist.h"
33 #include "util/evswitch.h"
34 #include <subcmd/pager.h>
35 #include <subcmd/exec-cmd.h>
36 #include "util/machine.h"
38 #include "util/symbol.h"
39 #include "util/path.h"
40 #include "util/session.h"
41 #include "util/thread.h"
42 #include <subcmd/parse-options.h>
43 #include "util/strlist.h"
44 #include "util/intlist.h"
45 #include "util/thread_map.h"
46 #include "util/stat.h"
47 #include "util/tool.h"
48 #include "util/util.h"
49 #include "trace/beauty/beauty.h"
50 #include "trace-event.h"
51 #include "util/parse-events.h"
52 #include "util/bpf-loader.h"
53 #include "callchain.h"
54 #include "print_binary.h"
56 #include "syscalltbl.h"
57 #include "rb_resort.h"
66 #include <linux/err.h>
67 #include <linux/filter.h>
68 #include <linux/kernel.h>
69 #include <linux/random.h>
70 #include <linux/stringify.h>
71 #include <linux/time64.h>
72 #include <linux/zalloc.h>
74 #include <sys/sysmacros.h>
76 #include <linux/ctype.h>
79 # define O_CLOEXEC 02000000
82 #ifndef F_LINUX_SPECIFIC_BASE
83 # define F_LINUX_SPECIFIC_BASE 1024
87 struct perf_tool tool;
88 struct syscalltbl *sctbl;
90 struct syscall *table;
92 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
93 struct bpf_map *sys_enter,
97 struct evsel *sys_enter,
101 struct bpf_program *unaugmented_prog;
106 struct record_opts opts;
107 struct evlist *evlist;
108 struct machine *host;
109 struct thread *current;
110 struct bpf_object *bpf_obj;
111 struct cgroup *cgroup;
114 unsigned long nr_events;
115 unsigned long nr_events_printed;
116 unsigned long max_events;
117 struct evswitch evswitch;
118 struct strlist *ev_qualifier;
128 double duration_filter;
134 unsigned int max_stack;
135 unsigned int min_stack;
136 int raw_augmented_syscalls_args_size;
137 bool raw_augmented_syscalls;
138 bool fd_path_disabled;
140 bool not_ev_qualifier;
144 bool multiple_threads;
150 bool show_tool_stats;
152 bool kernel_syscallchains;
158 bool show_string_prefix;
163 struct ordered_events data;
171 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
172 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
176 #define TP_UINT_FIELD(bits) \
177 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
180 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
189 #define TP_UINT_FIELD__SWAPPED(bits) \
190 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
193 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
194 return bswap_##bits(value);\
197 TP_UINT_FIELD__SWAPPED(16);
198 TP_UINT_FIELD__SWAPPED(32);
199 TP_UINT_FIELD__SWAPPED(64);
201 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
203 field->offset = offset;
207 field->integer = tp_field__u8;
210 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
213 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
216 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
225 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
227 return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
230 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
232 return sample->raw_data + field->offset;
235 static int __tp_field__init_ptr(struct tp_field *field, int offset)
237 field->offset = offset;
238 field->pointer = tp_field__ptr;
242 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
244 return __tp_field__init_ptr(field, format_field->offset);
250 struct tp_field args, ret;
254 static int perf_evsel__init_tp_uint_field(struct evsel *evsel,
255 struct tp_field *field,
258 struct tep_format_field *format_field = perf_evsel__field(evsel, name);
260 if (format_field == NULL)
263 return tp_field__init_uint(field, format_field, evsel->needs_swap);
266 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
267 ({ struct syscall_tp *sc = evsel->priv;\
268 perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
270 static int perf_evsel__init_tp_ptr_field(struct evsel *evsel,
271 struct tp_field *field,
274 struct tep_format_field *format_field = perf_evsel__field(evsel, name);
276 if (format_field == NULL)
279 return tp_field__init_ptr(field, format_field);
282 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
283 ({ struct syscall_tp *sc = evsel->priv;\
284 perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
286 static void evsel__delete_priv(struct evsel *evsel)
289 evsel__delete(evsel);
292 static int perf_evsel__init_syscall_tp(struct evsel *evsel)
294 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
296 if (evsel->priv != NULL) {
297 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
298 perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
309 static int perf_evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
311 struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
313 if (evsel->priv != NULL) {
314 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
315 if (syscall_id == NULL)
316 syscall_id = perf_evsel__field(tp, "__syscall_nr");
317 if (syscall_id == NULL)
319 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
331 static int perf_evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
333 struct syscall_tp *sc = evsel->priv;
335 return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
338 static int perf_evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
340 struct syscall_tp *sc = evsel->priv;
342 return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
345 static int perf_evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
347 evsel->priv = malloc(sizeof(struct syscall_tp));
348 if (evsel->priv != NULL) {
349 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
352 evsel->handler = handler;
363 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
365 struct evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
367 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
369 evsel = perf_evsel__newtp("syscalls", direction);
374 if (perf_evsel__init_raw_syscall_tp(evsel, handler))
380 evsel__delete_priv(evsel);
384 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
385 ({ struct syscall_tp *fields = evsel->priv; \
386 fields->name.integer(&fields->name, sample); })
388 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
389 ({ struct syscall_tp *fields = evsel->priv; \
390 fields->name.pointer(&fields->name, sample); })
392 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
394 int idx = val - sa->offset;
396 if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
397 size_t printed = scnprintf(bf, size, intfmt, val);
399 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
403 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
406 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
408 struct syscall_arg *arg)
410 return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
413 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
414 struct syscall_arg *arg)
416 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
419 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
421 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
423 return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
426 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
431 for (i = 0; i < sas->nr_entries; ++i) {
432 struct strarray *sa = sas->entries[i];
433 int idx = val - sa->offset;
435 if (idx >= 0 && idx < sa->nr_entries) {
436 if (sa->entries[idx] == NULL)
438 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
442 printed = scnprintf(bf, size, intfmt, val);
444 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
448 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
449 struct syscall_arg *arg)
451 return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
455 #define AT_FDCWD -100
458 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
459 struct syscall_arg *arg)
462 const char *prefix = "AT_FD";
465 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
467 return syscall_arg__scnprintf_fd(bf, size, arg);
470 #define SCA_FDAT syscall_arg__scnprintf_fd_at
472 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
473 struct syscall_arg *arg);
475 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
477 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
479 return scnprintf(bf, size, "%#lx", arg->val);
482 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
485 return scnprintf(bf, size, "NULL");
486 return syscall_arg__scnprintf_hex(bf, size, arg);
489 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
491 return scnprintf(bf, size, "%d", arg->val);
494 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
496 return scnprintf(bf, size, "%ld", arg->val);
499 static const char *bpf_cmd[] = {
500 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
501 "MAP_GET_NEXT_KEY", "PROG_LOAD",
503 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
505 static const char *fsmount_flags[] = {
508 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
510 #include "trace/beauty/generated/fsconfig_arrays.c"
512 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
514 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
515 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
517 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
518 static DEFINE_STRARRAY(itimers, "ITIMER_");
520 static const char *keyctl_options[] = {
521 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
522 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
523 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
524 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
525 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
527 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
529 static const char *whences[] = { "SET", "CUR", "END",
537 static DEFINE_STRARRAY(whences, "SEEK_");
539 static const char *fcntl_cmds[] = {
540 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
541 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
542 "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
545 static DEFINE_STRARRAY(fcntl_cmds, "F_");
547 static const char *fcntl_linux_specific_cmds[] = {
548 "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
549 "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
550 "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
553 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
555 static struct strarray *fcntl_cmds_arrays[] = {
556 &strarray__fcntl_cmds,
557 &strarray__fcntl_linux_specific_cmds,
560 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
562 static const char *rlimit_resources[] = {
563 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
564 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
567 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
569 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
570 static DEFINE_STRARRAY(sighow, "SIG_");
572 static const char *clockid[] = {
573 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
574 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
575 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
577 static DEFINE_STRARRAY(clockid, "CLOCK_");
579 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
580 struct syscall_arg *arg)
582 bool show_prefix = arg->show_string_prefix;
583 const char *suffix = "_OK";
587 if (mode == F_OK) /* 0 */
588 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
590 if (mode & n##_OK) { \
591 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
601 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
606 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
608 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
609 struct syscall_arg *arg);
611 #define SCA_FILENAME syscall_arg__scnprintf_filename
613 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
614 struct syscall_arg *arg)
616 bool show_prefix = arg->show_string_prefix;
617 const char *prefix = "O_";
618 int printed = 0, flags = arg->val;
621 if (flags & O_##n) { \
622 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
631 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
636 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
638 #ifndef GRND_NONBLOCK
639 #define GRND_NONBLOCK 0x0001
642 #define GRND_RANDOM 0x0002
645 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
646 struct syscall_arg *arg)
648 bool show_prefix = arg->show_string_prefix;
649 const char *prefix = "GRND_";
650 int printed = 0, flags = arg->val;
653 if (flags & GRND_##n) { \
654 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
655 flags &= ~GRND_##n; \
663 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
668 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
670 #define STRARRAY(name, array) \
671 { .scnprintf = SCA_STRARRAY, \
672 .parm = &strarray__##array, }
674 #define STRARRAY_FLAGS(name, array) \
675 { .scnprintf = SCA_STRARRAY_FLAGS, \
676 .parm = &strarray__##array, }
678 #include "trace/beauty/arch_errno_names.c"
679 #include "trace/beauty/eventfd.c"
680 #include "trace/beauty/futex_op.c"
681 #include "trace/beauty/futex_val3.c"
682 #include "trace/beauty/mmap.c"
683 #include "trace/beauty/mode_t.c"
684 #include "trace/beauty/msg_flags.c"
685 #include "trace/beauty/open_flags.c"
686 #include "trace/beauty/perf_event_open.c"
687 #include "trace/beauty/pid.c"
688 #include "trace/beauty/sched_policy.c"
689 #include "trace/beauty/seccomp.c"
690 #include "trace/beauty/signum.c"
691 #include "trace/beauty/socket_type.c"
692 #include "trace/beauty/waitid_options.c"
694 struct syscall_arg_fmt {
695 size_t (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
696 unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
702 static struct syscall_fmt {
706 const char *sys_enter,
709 struct syscall_arg_fmt arg[6];
716 .arg = { [1] = { .scnprintf = SCA_ACCMODE, /* mode */ }, }, },
717 { .name = "arch_prctl",
718 .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
719 [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
721 .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
722 [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
723 [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
725 .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
726 { .name = "brk", .hexret = true,
727 .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
728 { .name = "clock_gettime",
729 .arg = { [0] = STRARRAY(clk_id, clockid), }, },
730 { .name = "clone", .errpid = true, .nr_args = 5,
731 .arg = { [0] = { .name = "flags", .scnprintf = SCA_CLONE_FLAGS, },
732 [1] = { .name = "child_stack", .scnprintf = SCA_HEX, },
733 [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
734 [3] = { .name = "child_tidptr", .scnprintf = SCA_HEX, },
735 [4] = { .name = "tls", .scnprintf = SCA_HEX, }, }, },
737 .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
739 .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
740 [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
741 [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
742 { .name = "epoll_ctl",
743 .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
744 { .name = "eventfd2",
745 .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
746 { .name = "fchmodat",
747 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
748 { .name = "fchownat",
749 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
751 .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
752 .parm = &strarrays__fcntl_cmds_arrays,
753 .show_zero = true, },
754 [2] = { .scnprintf = SCA_FCNTL_ARG, /* arg */ }, }, },
756 .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
757 { .name = "fsconfig",
758 .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
760 .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
761 [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
763 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
764 [1] = { .scnprintf = SCA_FILENAME, /* path */ },
765 [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
766 { .name = "fstat", .alias = "newfstat", },
767 { .name = "fstatat", .alias = "newfstatat", },
769 .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
770 [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
771 { .name = "futimesat",
772 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
773 { .name = "getitimer",
774 .arg = { [0] = STRARRAY(which, itimers), }, },
775 { .name = "getpid", .errpid = true, },
776 { .name = "getpgid", .errpid = true, },
777 { .name = "getppid", .errpid = true, },
778 { .name = "getrandom",
779 .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
780 { .name = "getrlimit",
781 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
782 { .name = "gettid", .errpid = true, },
785 #if defined(__i386__) || defined(__x86_64__)
787 * FIXME: Make this available to all arches.
789 [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
790 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
792 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
794 { .name = "kcmp", .nr_args = 5,
795 .arg = { [0] = { .name = "pid1", .scnprintf = SCA_PID, },
796 [1] = { .name = "pid2", .scnprintf = SCA_PID, },
797 [2] = { .name = "type", .scnprintf = SCA_KCMP_TYPE, },
798 [3] = { .name = "idx1", .scnprintf = SCA_KCMP_IDX, },
799 [4] = { .name = "idx2", .scnprintf = SCA_KCMP_IDX, }, }, },
801 .arg = { [0] = STRARRAY(option, keyctl_options), }, },
803 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
805 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
807 .arg = { [2] = STRARRAY(whence, whences), }, },
808 { .name = "lstat", .alias = "newlstat", },
810 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
811 [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
813 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
815 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
816 { .name = "mmap", .hexret = true,
817 /* The standard mmap maps to old_mmap on s390x */
818 #if defined(__s390x__)
821 .arg = { [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
822 [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
823 [5] = { .scnprintf = SCA_HEX, /* offset */ }, }, },
825 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
826 [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
827 .mask_val = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
828 { .name = "move_mount",
829 .arg = { [0] = { .scnprintf = SCA_FDAT, /* from_dfd */ },
830 [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
831 [2] = { .scnprintf = SCA_FDAT, /* to_dfd */ },
832 [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
833 [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
834 { .name = "mprotect",
835 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
836 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, }, },
837 { .name = "mq_unlink",
838 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
839 { .name = "mremap", .hexret = true,
840 .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
841 { .name = "name_to_handle_at",
842 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
843 { .name = "newfstatat",
844 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
846 .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
847 { .name = "open_by_handle_at",
848 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
849 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
851 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
852 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
853 { .name = "perf_event_open",
854 .arg = { [2] = { .scnprintf = SCA_INT, /* cpu */ },
855 [3] = { .scnprintf = SCA_FD, /* group_fd */ },
856 [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
858 .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
859 { .name = "pkey_alloc",
860 .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS, /* access_rights */ }, }, },
861 { .name = "pkey_free",
862 .arg = { [0] = { .scnprintf = SCA_INT, /* key */ }, }, },
863 { .name = "pkey_mprotect",
864 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
865 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
866 [3] = { .scnprintf = SCA_INT, /* pkey */ }, }, },
867 { .name = "poll", .timeout = true, },
868 { .name = "ppoll", .timeout = true, },
870 .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
871 [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
872 [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
873 { .name = "pread", .alias = "pread64", },
874 { .name = "preadv", .alias = "pread", },
875 { .name = "prlimit64",
876 .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
877 { .name = "pwrite", .alias = "pwrite64", },
878 { .name = "readlinkat",
879 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
880 { .name = "recvfrom",
881 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
882 { .name = "recvmmsg",
883 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
885 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
886 { .name = "renameat",
887 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
888 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
889 { .name = "renameat2",
890 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
891 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
892 [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
893 { .name = "rt_sigaction",
894 .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
895 { .name = "rt_sigprocmask",
896 .arg = { [0] = STRARRAY(how, sighow), }, },
897 { .name = "rt_sigqueueinfo",
898 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
899 { .name = "rt_tgsigqueueinfo",
900 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
901 { .name = "sched_setscheduler",
902 .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
904 .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP, /* op */ },
905 [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
906 { .name = "select", .timeout = true, },
907 { .name = "sendfile", .alias = "sendfile64", },
908 { .name = "sendmmsg",
909 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
911 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
913 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
914 [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
915 { .name = "set_tid_address", .errpid = true, },
916 { .name = "setitimer",
917 .arg = { [0] = STRARRAY(which, itimers), }, },
918 { .name = "setrlimit",
919 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
921 .arg = { [0] = STRARRAY(family, socket_families),
922 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
923 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
924 { .name = "socketpair",
925 .arg = { [0] = STRARRAY(family, socket_families),
926 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
927 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
928 { .name = "stat", .alias = "newstat", },
930 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fdat */ },
931 [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
932 [3] = { .scnprintf = SCA_STATX_MASK, /* mask */ }, }, },
934 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
936 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
937 { .name = "symlinkat",
938 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
939 { .name = "sync_file_range",
940 .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
942 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
944 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
945 { .name = "umount2", .alias = "umount",
946 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
947 { .name = "uname", .alias = "newuname", },
948 { .name = "unlinkat",
949 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
950 { .name = "utimensat",
951 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
952 { .name = "wait4", .errpid = true,
953 .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
954 { .name = "waitid", .errpid = true,
955 .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
958 static int syscall_fmt__cmp(const void *name, const void *fmtp)
960 const struct syscall_fmt *fmt = fmtp;
961 return strcmp(name, fmt->name);
964 static struct syscall_fmt *syscall_fmt__find(const char *name)
966 const int nmemb = ARRAY_SIZE(syscall_fmts);
967 return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
970 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
972 int i, nmemb = ARRAY_SIZE(syscall_fmts);
974 for (i = 0; i < nmemb; ++i) {
975 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
976 return &syscall_fmts[i];
983 * is_exit: is this "exit" or "exit_group"?
984 * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
985 * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
986 * nonexistent: Just a hole in the syscall table, syscall id not allocated
989 struct tep_event *tp_format;
993 struct bpf_program *sys_enter,
999 struct tep_format_field *args;
1001 struct syscall_fmt *fmt;
1002 struct syscall_arg_fmt *arg_fmt;
1006 * Must match what is in the BPF program:
1008 * tools/perf/examples/bpf/augmented_raw_syscalls.c
1010 struct bpf_map_syscall_entry {
1012 u16 string_args_len[6];
1016 * We need to have this 'calculated' boolean because in some cases we really
1017 * don't know what is the duration of a syscall, for instance, when we start
1018 * a session and some threads are waiting for a syscall to finish, say 'poll',
1019 * in which case all we can do is to print "( ? ) for duration and for the
1022 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1024 double duration = (double)t / NSEC_PER_MSEC;
1025 size_t printed = fprintf(fp, "(");
1028 printed += fprintf(fp, " ");
1029 else if (duration >= 1.0)
1030 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1031 else if (duration >= 0.01)
1032 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1034 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1035 return printed + fprintf(fp, "): ");
1039 * filename.ptr: The filename char pointer that will be vfs_getname'd
1040 * filename.entry_str_pos: Where to insert the string translated from
1041 * filename.ptr by the vfs_getname tracepoint/kprobe.
1042 * ret_scnprintf: syscall args may set this to a different syscall return
1043 * formatter, for instance, fcntl may return fds, file flags, etc.
1045 struct thread_trace {
1048 unsigned long nr_events;
1049 unsigned long pfmaj, pfmin;
1052 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1055 short int entry_str_pos;
1057 unsigned int namelen;
1065 struct intlist *syscall_stats;
1068 static struct thread_trace *thread_trace__new(void)
1070 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
1073 ttrace->files.max = -1;
1074 ttrace->syscall_stats = intlist__new(NULL);
1080 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1082 struct thread_trace *ttrace;
1087 if (thread__priv(thread) == NULL)
1088 thread__set_priv(thread, thread_trace__new());
1090 if (thread__priv(thread) == NULL)
1093 ttrace = thread__priv(thread);
1094 ++ttrace->nr_events;
1098 color_fprintf(fp, PERF_COLOR_RED,
1099 "WARNING: not enough memory, dropping samples!\n");
1104 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1105 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1107 struct thread_trace *ttrace = thread__priv(arg->thread);
1109 ttrace->ret_scnprintf = ret_scnprintf;
1112 #define TRACE_PFMAJ (1 << 0)
1113 #define TRACE_PFMIN (1 << 1)
1115 static const size_t trace__entry_str_size = 2048;
1117 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1122 if (fd > ttrace->files.max) {
1123 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1128 if (ttrace->files.max != -1) {
1129 memset(nfiles + ttrace->files.max + 1, 0,
1130 (fd - ttrace->files.max) * sizeof(struct file));
1132 memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1135 ttrace->files.table = nfiles;
1136 ttrace->files.max = fd;
1139 return ttrace->files.table + fd;
1142 struct file *thread__files_entry(struct thread *thread, int fd)
1144 return thread_trace__files_entry(thread__priv(thread), fd);
1147 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1149 struct thread_trace *ttrace = thread__priv(thread);
1150 struct file *file = thread_trace__files_entry(ttrace, fd);
1154 if (stat(pathname, &st) == 0)
1155 file->dev_maj = major(st.st_rdev);
1156 file->pathname = strdup(pathname);
1164 static int thread__read_fd_path(struct thread *thread, int fd)
1166 char linkname[PATH_MAX], pathname[PATH_MAX];
1170 if (thread->pid_ == thread->tid) {
1171 scnprintf(linkname, sizeof(linkname),
1172 "/proc/%d/fd/%d", thread->pid_, fd);
1174 scnprintf(linkname, sizeof(linkname),
1175 "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1178 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1181 ret = readlink(linkname, pathname, sizeof(pathname));
1183 if (ret < 0 || ret > st.st_size)
1186 pathname[ret] = '\0';
1187 return trace__set_fd_pathname(thread, fd, pathname);
1190 static const char *thread__fd_path(struct thread *thread, int fd,
1191 struct trace *trace)
1193 struct thread_trace *ttrace = thread__priv(thread);
1195 if (ttrace == NULL || trace->fd_path_disabled)
1201 if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1204 ++trace->stats.proc_getname;
1205 if (thread__read_fd_path(thread, fd))
1209 return ttrace->files.table[fd].pathname;
1212 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1215 size_t printed = scnprintf(bf, size, "%d", fd);
1216 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1219 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1224 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1226 size_t printed = scnprintf(bf, size, "%d", fd);
1227 struct thread *thread = machine__find_thread(trace->host, pid, pid);
1230 const char *path = thread__fd_path(thread, fd, trace);
1233 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1235 thread__put(thread);
1241 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1242 struct syscall_arg *arg)
1245 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1246 struct thread_trace *ttrace = thread__priv(arg->thread);
1248 if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1249 zfree(&ttrace->files.table[fd].pathname);
1254 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1257 struct thread_trace *ttrace = thread__priv(thread);
1259 ttrace->filename.ptr = ptr;
1260 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1263 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1265 struct augmented_arg *augmented_arg = arg->augmented.args;
1266 size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1268 * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1269 * we would have two strings, each prefixed by its size.
1271 int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1273 arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1274 arg->augmented.size -= consumed;
1279 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1280 struct syscall_arg *arg)
1282 unsigned long ptr = arg->val;
1284 if (arg->augmented.args)
1285 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1287 if (!arg->trace->vfs_getname)
1288 return scnprintf(bf, size, "%#x", ptr);
1290 thread__set_filename_pos(arg->thread, bf, ptr);
1294 static bool trace__filter_duration(struct trace *trace, double t)
1296 return t < (trace->duration_filter * NSEC_PER_MSEC);
1299 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1301 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1303 return fprintf(fp, "%10.3f ", ts);
1307 * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1308 * using ttrace->entry_time for a thread that receives a sys_exit without
1309 * first having received a sys_enter ("poll" issued before tracing session
1310 * starts, lost sys_enter exit due to ring buffer overflow).
1312 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1315 return __trace__fprintf_tstamp(trace, tstamp, fp);
1317 return fprintf(fp, " ? ");
1320 static bool done = false;
1321 static bool interrupted = false;
1323 static void sig_handler(int sig)
1326 interrupted = sig == SIGINT;
1329 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1333 if (trace->multiple_threads) {
1334 if (trace->show_comm)
1335 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1336 printed += fprintf(fp, "%d ", thread->tid);
1342 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1343 u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1347 if (trace->show_tstamp)
1348 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1349 if (trace->show_duration)
1350 printed += fprintf_duration(duration, duration_calculated, fp);
1351 return printed + trace__fprintf_comm_tid(trace, thread, fp);
1354 static int trace__process_event(struct trace *trace, struct machine *machine,
1355 union perf_event *event, struct perf_sample *sample)
1359 switch (event->header.type) {
1360 case PERF_RECORD_LOST:
1361 color_fprintf(trace->output, PERF_COLOR_RED,
1362 "LOST %" PRIu64 " events!\n", event->lost.lost);
1363 ret = machine__process_lost_event(machine, event, sample);
1366 ret = machine__process_event(machine, event, sample);
1373 static int trace__tool_process(struct perf_tool *tool,
1374 union perf_event *event,
1375 struct perf_sample *sample,
1376 struct machine *machine)
1378 struct trace *trace = container_of(tool, struct trace, tool);
1379 return trace__process_event(trace, machine, event, sample);
1382 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1384 struct machine *machine = vmachine;
1386 if (machine->kptr_restrict_warned)
1389 if (symbol_conf.kptr_restrict) {
1390 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1391 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1392 "Kernel samples will not be resolved.\n");
1393 machine->kptr_restrict_warned = true;
1397 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1400 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1402 int err = symbol__init(NULL);
1407 trace->host = machine__new_host();
1408 if (trace->host == NULL)
1411 err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1415 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1416 evlist->core.threads, trace__tool_process, false,
1425 static void trace__symbols__exit(struct trace *trace)
1427 machine__exit(trace->host);
1433 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1437 if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1438 nr_args = sc->fmt->nr_args;
1440 sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1441 if (sc->arg_fmt == NULL)
1444 for (idx = 0; idx < nr_args; ++idx) {
1446 sc->arg_fmt[idx] = sc->fmt->arg[idx];
1449 sc->nr_args = nr_args;
1453 static int syscall__set_arg_fmts(struct syscall *sc)
1455 struct tep_format_field *field, *last_field = NULL;
1458 for (field = sc->args; field; field = field->next, ++idx) {
1461 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1464 len = strlen(field->name);
1466 if (strcmp(field->type, "const char *") == 0 &&
1467 ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1468 strstr(field->name, "path") != NULL))
1469 sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1470 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1471 sc->arg_fmt[idx].scnprintf = SCA_PTR;
1472 else if (strcmp(field->type, "pid_t") == 0)
1473 sc->arg_fmt[idx].scnprintf = SCA_PID;
1474 else if (strcmp(field->type, "umode_t") == 0)
1475 sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1476 else if ((strcmp(field->type, "int") == 0 ||
1477 strcmp(field->type, "unsigned int") == 0 ||
1478 strcmp(field->type, "long") == 0) &&
1479 len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1481 * /sys/kernel/tracing/events/syscalls/sys_enter*
1482 * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1487 sc->arg_fmt[idx].scnprintf = SCA_FD;
1492 sc->args_size = last_field->offset + last_field->size;
1497 static int trace__read_syscall_info(struct trace *trace, int id)
1501 const char *name = syscalltbl__name(trace->sctbl, id);
1503 if (trace->syscalls.table == NULL) {
1504 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1505 if (trace->syscalls.table == NULL)
1509 sc = trace->syscalls.table + id;
1510 if (sc->nonexistent)
1514 sc->nonexistent = true;
1519 sc->fmt = syscall_fmt__find(sc->name);
1521 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1522 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1524 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1525 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1526 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1529 if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1532 if (IS_ERR(sc->tp_format))
1533 return PTR_ERR(sc->tp_format);
1535 sc->args = sc->tp_format->format.fields;
1537 * We need to check and discard the first variable '__syscall_nr'
1538 * or 'nr' that mean the syscall number. It is needless here.
1539 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1541 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1542 sc->args = sc->args->next;
1546 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1547 sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1549 return syscall__set_arg_fmts(sc);
1552 static int intcmp(const void *a, const void *b)
1554 const int *one = a, *another = b;
1556 return *one - *another;
1559 static int trace__validate_ev_qualifier(struct trace *trace)
1562 bool printed_invalid_prefix = false;
1563 struct str_node *pos;
1564 size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1566 trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1567 sizeof(trace->ev_qualifier_ids.entries[0]));
1569 if (trace->ev_qualifier_ids.entries == NULL) {
1570 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1576 strlist__for_each_entry(pos, trace->ev_qualifier) {
1577 const char *sc = pos->s;
1578 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1581 id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1585 if (!printed_invalid_prefix) {
1586 pr_debug("Skipping unknown syscalls: ");
1587 printed_invalid_prefix = true;
1596 trace->ev_qualifier_ids.entries[nr_used++] = id;
1597 if (match_next == -1)
1601 id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1604 if (nr_allocated == nr_used) {
1608 entries = realloc(trace->ev_qualifier_ids.entries,
1609 nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1610 if (entries == NULL) {
1612 fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1615 trace->ev_qualifier_ids.entries = entries;
1617 trace->ev_qualifier_ids.entries[nr_used++] = id;
1621 trace->ev_qualifier_ids.nr = nr_used;
1622 qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1624 if (printed_invalid_prefix)
1628 zfree(&trace->ev_qualifier_ids.entries);
1629 trace->ev_qualifier_ids.nr = 0;
1633 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1635 bool in_ev_qualifier;
1637 if (trace->ev_qualifier_ids.nr == 0)
1640 in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1641 trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1643 if (in_ev_qualifier)
1644 return !trace->not_ev_qualifier;
1646 return trace->not_ev_qualifier;
1650 * args is to be interpreted as a series of longs but we need to handle
1651 * 8-byte unaligned accesses. args points to raw_data within the event
1652 * and raw_data is guaranteed to be 8-byte unaligned because it is
1653 * preceded by raw_size which is a u32. So we need to copy args to a temp
1654 * variable to read it. Most notably this avoids extended load instructions
1655 * on unaligned addresses
1657 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1660 unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1662 memcpy(&val, p, sizeof(val));
1666 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1667 struct syscall_arg *arg)
1669 if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1670 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1672 return scnprintf(bf, size, "arg%d: ", arg->idx);
1676 * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1677 * as mount 'flags' argument that needs ignoring some magic flag, see comment
1678 * in tools/perf/trace/beauty/mount_flags.c
1680 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1682 if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1683 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1688 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1689 struct syscall_arg *arg, unsigned long val)
1691 if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1693 if (sc->arg_fmt[arg->idx].parm)
1694 arg->parm = sc->arg_fmt[arg->idx].parm;
1695 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1697 return scnprintf(bf, size, "%ld", val);
1700 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1701 unsigned char *args, void *augmented_args, int augmented_args_size,
1702 struct trace *trace, struct thread *thread)
1707 struct syscall_arg arg = {
1710 .size = augmented_args_size,
1711 .args = augmented_args,
1717 .show_string_prefix = trace->show_string_prefix,
1719 struct thread_trace *ttrace = thread__priv(thread);
1722 * Things like fcntl will set this in its 'cmd' formatter to pick the
1723 * right formatter for the return value (an fd? file flags?), which is
1724 * not needed for syscalls that always return a given type, say an fd.
1726 ttrace->ret_scnprintf = NULL;
1728 if (sc->args != NULL) {
1729 struct tep_format_field *field;
1731 for (field = sc->args; field;
1732 field = field->next, ++arg.idx, bit <<= 1) {
1736 val = syscall_arg__val(&arg, arg.idx);
1738 * Some syscall args need some mask, most don't and
1739 * return val untouched.
1741 val = syscall__mask_val(sc, &arg, val);
1744 * Suppress this argument if its value is zero and
1745 * and we don't have a string associated in an
1749 !trace->show_zeros &&
1751 (sc->arg_fmt[arg.idx].show_zero ||
1752 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1753 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1754 sc->arg_fmt[arg.idx].parm))
1757 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1759 if (trace->show_arg_names)
1760 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1762 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1764 } else if (IS_ERR(sc->tp_format)) {
1766 * If we managed to read the tracepoint /format file, then we
1767 * may end up not having any args, like with gettid(), so only
1768 * print the raw args when we didn't manage to read it.
1770 while (arg.idx < sc->nr_args) {
1773 val = syscall_arg__val(&arg, arg.idx);
1775 printed += scnprintf(bf + printed, size - printed, ", ");
1776 printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1777 printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1787 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
1788 union perf_event *event,
1789 struct perf_sample *sample);
1791 static struct syscall *trace__syscall_info(struct trace *trace,
1792 struct evsel *evsel, int id)
1799 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1800 * before that, leaving at a higher verbosity level till that is
1801 * explained. Reproduced with plain ftrace with:
1803 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1804 * grep "NR -1 " /t/trace_pipe
1806 * After generating some load on the machine.
1810 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1811 id, perf_evsel__name(evsel), ++n);
1818 if (id > trace->sctbl->syscalls.max_id)
1821 if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
1822 (err = trace__read_syscall_info(trace, id)) != 0)
1825 if (trace->syscalls.table[id].name == NULL) {
1826 if (trace->syscalls.table[id].nonexistent)
1831 return &trace->syscalls.table[id];
1835 char sbuf[STRERR_BUFSIZE];
1836 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
1837 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
1838 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1839 fputs(" information\n", trace->output);
1844 static void thread__update_stats(struct thread_trace *ttrace,
1845 int id, struct perf_sample *sample)
1847 struct int_node *inode;
1848 struct stats *stats;
1851 inode = intlist__findnew(ttrace->syscall_stats, id);
1855 stats = inode->priv;
1856 if (stats == NULL) {
1857 stats = malloc(sizeof(struct stats));
1861 inode->priv = stats;
1864 if (ttrace->entry_time && sample->time > ttrace->entry_time)
1865 duration = sample->time - ttrace->entry_time;
1867 update_stats(stats, duration);
1870 static int trace__printf_interrupted_entry(struct trace *trace)
1872 struct thread_trace *ttrace;
1876 if (trace->failure_only || trace->current == NULL)
1879 ttrace = thread__priv(trace->current);
1881 if (!ttrace->entry_pending)
1884 printed = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1885 printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1887 if (len < trace->args_alignment - 4)
1888 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1890 printed += fprintf(trace->output, " ...\n");
1892 ttrace->entry_pending = false;
1893 ++trace->nr_events_printed;
1898 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
1899 struct perf_sample *sample, struct thread *thread)
1903 if (trace->print_sample) {
1904 double ts = (double)sample->time / NSEC_PER_MSEC;
1906 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1907 perf_evsel__name(evsel), ts,
1908 thread__comm_str(thread),
1909 sample->pid, sample->tid, sample->cpu);
1915 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1917 void *augmented_args = NULL;
1919 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1920 * and there we get all 6 syscall args plus the tracepoint common fields
1921 * that gets calculated at the start and the syscall_nr (another long).
1922 * So we check if that is the case and if so don't look after the
1923 * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1926 * We'll revisit this later to pass s->args_size to the BPF augmenter
1927 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1928 * copies only what we need for each syscall, like what happens when we
1929 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1930 * traffic to just what is needed for each syscall.
1932 int args_size = raw_augmented_args_size ?: sc->args_size;
1934 *augmented_args_size = sample->raw_size - args_size;
1935 if (*augmented_args_size > 0)
1936 augmented_args = sample->raw_data + args_size;
1938 return augmented_args;
1941 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
1942 union perf_event *event __maybe_unused,
1943 struct perf_sample *sample)
1948 struct thread *thread;
1949 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1950 int augmented_args_size = 0;
1951 void *augmented_args = NULL;
1952 struct syscall *sc = trace__syscall_info(trace, evsel, id);
1953 struct thread_trace *ttrace;
1958 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1959 ttrace = thread__trace(thread, trace->output);
1963 trace__fprintf_sample(trace, evsel, sample, thread);
1965 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1967 if (ttrace->entry_str == NULL) {
1968 ttrace->entry_str = malloc(trace__entry_str_size);
1969 if (!ttrace->entry_str)
1973 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1974 trace__printf_interrupted_entry(trace);
1976 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1977 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1978 * this breaks syscall__augmented_args() check for augmented args, as we calculate
1979 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1980 * so when handling, say the openat syscall, we end up getting 6 args for the
1981 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1982 * thinking that the extra 2 u64 args are the augmented filename, so just check
1983 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1985 if (evsel != trace->syscalls.events.sys_enter)
1986 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1987 ttrace->entry_time = sample->time;
1988 msg = ttrace->entry_str;
1989 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1991 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1992 args, augmented_args, augmented_args_size, trace, thread);
1995 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1998 trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1999 printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2000 if (trace->args_alignment > printed)
2001 alignment = trace->args_alignment - printed;
2002 fprintf(trace->output, "%*s= ?\n", alignment, " ");
2005 ttrace->entry_pending = true;
2006 /* See trace__vfs_getname & trace__sys_exit */
2007 ttrace->filename.pending_open = false;
2010 if (trace->current != thread) {
2011 thread__put(trace->current);
2012 trace->current = thread__get(thread);
2016 thread__put(thread);
2020 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2021 struct perf_sample *sample)
2023 struct thread_trace *ttrace;
2024 struct thread *thread;
2025 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2026 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2028 void *args, *augmented_args = NULL;
2029 int augmented_args_size;
2034 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2035 ttrace = thread__trace(thread, trace->output);
2037 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2038 * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2043 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2044 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2045 syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2046 fprintf(trace->output, "%s", msg);
2049 thread__put(thread);
2053 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2054 struct perf_sample *sample,
2055 struct callchain_cursor *cursor)
2057 struct addr_location al;
2058 int max_stack = evsel->core.attr.sample_max_stack ?
2059 evsel->core.attr.sample_max_stack :
2063 if (machine__resolve(trace->host, &al, sample) < 0)
2066 err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2067 addr_location__put(&al);
2071 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2073 /* TODO: user-configurable print_opts */
2074 const unsigned int print_opts = EVSEL__PRINT_SYM |
2076 EVSEL__PRINT_UNKNOWN_AS_ADDR;
2078 return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2081 static const char *errno_to_name(struct evsel *evsel, int err)
2083 struct perf_env *env = perf_evsel__env(evsel);
2084 const char *arch_name = perf_env__arch(env);
2086 return arch_syscalls__strerrno(arch_name, err);
2089 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2090 union perf_event *event __maybe_unused,
2091 struct perf_sample *sample)
2095 bool duration_calculated = false;
2096 struct thread *thread;
2097 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2098 int alignment = trace->args_alignment;
2099 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2100 struct thread_trace *ttrace;
2105 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2106 ttrace = thread__trace(thread, trace->output);
2110 trace__fprintf_sample(trace, evsel, sample, thread);
2113 thread__update_stats(ttrace, id, sample);
2115 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2117 if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2118 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2119 ttrace->filename.pending_open = false;
2120 ++trace->stats.vfs_getname;
2123 if (ttrace->entry_time) {
2124 duration = sample->time - ttrace->entry_time;
2125 if (trace__filter_duration(trace, duration))
2127 duration_calculated = true;
2128 } else if (trace->duration_filter)
2131 if (sample->callchain) {
2132 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2133 if (callchain_ret == 0) {
2134 if (callchain_cursor.nr < trace->min_stack)
2140 if (trace->summary_only || (ret >= 0 && trace->failure_only))
2143 trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2145 if (ttrace->entry_pending) {
2146 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2148 printed += fprintf(trace->output, " ... [");
2149 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2151 printed += fprintf(trace->output, "]: %s()", sc->name);
2154 printed++; /* the closing ')' */
2156 if (alignment > printed)
2157 alignment -= printed;
2161 fprintf(trace->output, ")%*s= ", alignment, " ");
2163 if (sc->fmt == NULL) {
2167 fprintf(trace->output, "%ld", ret);
2168 } else if (ret < 0) {
2170 char bf[STRERR_BUFSIZE];
2171 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2172 *e = errno_to_name(evsel, -ret);
2174 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2176 } else if (ret == 0 && sc->fmt->timeout)
2177 fprintf(trace->output, "0 (Timeout)");
2178 else if (ttrace->ret_scnprintf) {
2180 struct syscall_arg arg = {
2185 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2186 ttrace->ret_scnprintf = NULL;
2187 fprintf(trace->output, "%s", bf);
2188 } else if (sc->fmt->hexret)
2189 fprintf(trace->output, "%#lx", ret);
2190 else if (sc->fmt->errpid) {
2191 struct thread *child = machine__find_thread(trace->host, ret, ret);
2193 if (child != NULL) {
2194 fprintf(trace->output, "%ld", ret);
2195 if (child->comm_set)
2196 fprintf(trace->output, " (%s)", thread__comm_str(child));
2202 fputc('\n', trace->output);
2205 * We only consider an 'event' for the sake of --max-events a non-filtered
2206 * sys_enter + sys_exit and other tracepoint events.
2208 if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2211 if (callchain_ret > 0)
2212 trace__fprintf_callchain(trace, sample);
2213 else if (callchain_ret < 0)
2214 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2216 ttrace->entry_pending = false;
2219 thread__put(thread);
2223 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2224 union perf_event *event __maybe_unused,
2225 struct perf_sample *sample)
2227 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2228 struct thread_trace *ttrace;
2229 size_t filename_len, entry_str_len, to_move;
2230 ssize_t remaining_space;
2232 const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2237 ttrace = thread__priv(thread);
2241 filename_len = strlen(filename);
2242 if (filename_len == 0)
2245 if (ttrace->filename.namelen < filename_len) {
2246 char *f = realloc(ttrace->filename.name, filename_len + 1);
2251 ttrace->filename.namelen = filename_len;
2252 ttrace->filename.name = f;
2255 strcpy(ttrace->filename.name, filename);
2256 ttrace->filename.pending_open = true;
2258 if (!ttrace->filename.ptr)
2261 entry_str_len = strlen(ttrace->entry_str);
2262 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2263 if (remaining_space <= 0)
2266 if (filename_len > (size_t)remaining_space) {
2267 filename += filename_len - remaining_space;
2268 filename_len = remaining_space;
2271 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2272 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2273 memmove(pos + filename_len, pos, to_move);
2274 memcpy(pos, filename, filename_len);
2276 ttrace->filename.ptr = 0;
2277 ttrace->filename.entry_str_pos = 0;
2279 thread__put(thread);
2284 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2285 union perf_event *event __maybe_unused,
2286 struct perf_sample *sample)
2288 u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2289 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2290 struct thread *thread = machine__findnew_thread(trace->host,
2293 struct thread_trace *ttrace = thread__trace(thread, trace->output);
2298 ttrace->runtime_ms += runtime_ms;
2299 trace->runtime_ms += runtime_ms;
2301 thread__put(thread);
2305 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2307 perf_evsel__strval(evsel, sample, "comm"),
2308 (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2310 perf_evsel__intval(evsel, sample, "vruntime"));
2314 static int bpf_output__printer(enum binary_printer_ops op,
2315 unsigned int val, void *extra __maybe_unused, FILE *fp)
2317 unsigned char ch = (unsigned char)val;
2320 case BINARY_PRINT_CHAR_DATA:
2321 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2322 case BINARY_PRINT_DATA_BEGIN:
2323 case BINARY_PRINT_LINE_BEGIN:
2324 case BINARY_PRINT_ADDR:
2325 case BINARY_PRINT_NUM_DATA:
2326 case BINARY_PRINT_NUM_PAD:
2327 case BINARY_PRINT_SEP:
2328 case BINARY_PRINT_CHAR_PAD:
2329 case BINARY_PRINT_LINE_END:
2330 case BINARY_PRINT_DATA_END:
2338 static void bpf_output__fprintf(struct trace *trace,
2339 struct perf_sample *sample)
2341 binary__fprintf(sample->raw_data, sample->raw_size, 8,
2342 bpf_output__printer, NULL, trace->output);
2343 ++trace->nr_events_printed;
2346 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2347 union perf_event *event __maybe_unused,
2348 struct perf_sample *sample)
2350 struct thread *thread;
2351 int callchain_ret = 0;
2353 * Check if we called perf_evsel__disable(evsel) due to, for instance,
2354 * this event's max_events having been hit and this is an entry coming
2355 * from the ring buffer that we should discard, since the max events
2356 * have already been considered/printed.
2358 if (evsel->disabled)
2361 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2363 if (sample->callchain) {
2364 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2365 if (callchain_ret == 0) {
2366 if (callchain_cursor.nr < trace->min_stack)
2372 trace__printf_interrupted_entry(trace);
2373 trace__fprintf_tstamp(trace, sample->time, trace->output);
2375 if (trace->trace_syscalls && trace->show_duration)
2376 fprintf(trace->output, "( ): ");
2379 trace__fprintf_comm_tid(trace, thread, trace->output);
2381 if (evsel == trace->syscalls.events.augmented) {
2382 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2383 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2386 fprintf(trace->output, "%s(", sc->name);
2387 trace__fprintf_sys_enter(trace, evsel, sample);
2388 fputc(')', trace->output);
2393 * XXX: Not having the associated syscall info or not finding/adding
2394 * the thread should never happen, but if it does...
2395 * fall thru and print it as a bpf_output event.
2399 fprintf(trace->output, "%s:", evsel->name);
2401 if (perf_evsel__is_bpf_output(evsel)) {
2402 bpf_output__fprintf(trace, sample);
2403 } else if (evsel->tp_format) {
2404 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2405 trace__fprintf_sys_enter(trace, evsel, sample)) {
2406 event_format__fprintf(evsel->tp_format, sample->cpu,
2407 sample->raw_data, sample->raw_size,
2409 ++trace->nr_events_printed;
2411 if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2412 evsel__disable(evsel);
2413 evsel__close(evsel);
2419 fprintf(trace->output, "\n");
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 thread__put(thread);
2430 static void print_location(FILE *f, struct perf_sample *sample,
2431 struct addr_location *al,
2432 bool print_dso, bool print_sym)
2435 if ((verbose > 0 || print_dso) && al->map)
2436 fprintf(f, "%s@", al->map->dso->long_name);
2438 if ((verbose > 0 || print_sym) && al->sym)
2439 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2440 al->addr - al->sym->start);
2442 fprintf(f, "0x%" PRIx64, al->addr);
2444 fprintf(f, "0x%" PRIx64, sample->addr);
2447 static int trace__pgfault(struct trace *trace,
2448 struct evsel *evsel,
2449 union perf_event *event __maybe_unused,
2450 struct perf_sample *sample)
2452 struct thread *thread;
2453 struct addr_location al;
2454 char map_type = 'd';
2455 struct thread_trace *ttrace;
2457 int callchain_ret = 0;
2459 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2461 if (sample->callchain) {
2462 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2463 if (callchain_ret == 0) {
2464 if (callchain_cursor.nr < trace->min_stack)
2470 ttrace = thread__trace(thread, trace->output);
2474 if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2479 if (trace->summary_only)
2482 thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2484 trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2486 fprintf(trace->output, "%sfault [",
2487 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2490 print_location(trace->output, sample, &al, false, true);
2492 fprintf(trace->output, "] => ");
2494 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2497 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2505 print_location(trace->output, sample, &al, true, false);
2507 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2509 if (callchain_ret > 0)
2510 trace__fprintf_callchain(trace, sample);
2511 else if (callchain_ret < 0)
2512 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2514 ++trace->nr_events_printed;
2518 thread__put(thread);
2522 static void trace__set_base_time(struct trace *trace,
2523 struct evsel *evsel,
2524 struct perf_sample *sample)
2527 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2528 * and don't use sample->time unconditionally, we may end up having
2529 * some other event in the future without PERF_SAMPLE_TIME for good
2530 * reason, i.e. we may not be interested in its timestamps, just in
2531 * it taking place, picking some piece of information when it
2532 * appears in our event stream (vfs_getname comes to mind).
2534 if (trace->base_time == 0 && !trace->full_time &&
2535 (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2536 trace->base_time = sample->time;
2539 static int trace__process_sample(struct perf_tool *tool,
2540 union perf_event *event,
2541 struct perf_sample *sample,
2542 struct evsel *evsel,
2543 struct machine *machine __maybe_unused)
2545 struct trace *trace = container_of(tool, struct trace, tool);
2546 struct thread *thread;
2549 tracepoint_handler handler = evsel->handler;
2551 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2552 if (thread && thread__is_filtered(thread))
2555 trace__set_base_time(trace, evsel, sample);
2559 handler(trace, evsel, event, sample);
2562 thread__put(thread);
2566 static int trace__record(struct trace *trace, int argc, const char **argv)
2568 unsigned int rec_argc, i, j;
2569 const char **rec_argv;
2570 const char * const record_args[] = {
2577 const char * const sc_args[] = { "-e", };
2578 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2579 const char * const majpf_args[] = { "-e", "major-faults" };
2580 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2581 const char * const minpf_args[] = { "-e", "minor-faults" };
2582 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2584 /* +1 is for the event string below */
2585 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2586 majpf_args_nr + minpf_args_nr + argc;
2587 rec_argv = calloc(rec_argc + 1, sizeof(char *));
2589 if (rec_argv == NULL)
2593 for (i = 0; i < ARRAY_SIZE(record_args); i++)
2594 rec_argv[j++] = record_args[i];
2596 if (trace->trace_syscalls) {
2597 for (i = 0; i < sc_args_nr; i++)
2598 rec_argv[j++] = sc_args[i];
2600 /* event string may be different for older kernels - e.g., RHEL6 */
2601 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2602 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2603 else if (is_valid_tracepoint("syscalls:sys_enter"))
2604 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2606 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2612 if (trace->trace_pgfaults & TRACE_PFMAJ)
2613 for (i = 0; i < majpf_args_nr; i++)
2614 rec_argv[j++] = majpf_args[i];
2616 if (trace->trace_pgfaults & TRACE_PFMIN)
2617 for (i = 0; i < minpf_args_nr; i++)
2618 rec_argv[j++] = minpf_args[i];
2620 for (i = 0; i < (unsigned int)argc; i++)
2621 rec_argv[j++] = argv[i];
2623 return cmd_record(j, rec_argv);
2626 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2628 static bool evlist__add_vfs_getname(struct evlist *evlist)
2631 struct evsel *evsel, *tmp;
2632 struct parse_events_error err = { .idx = 0, };
2633 int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2638 evlist__for_each_entry_safe(evlist, evsel, tmp) {
2639 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2642 if (perf_evsel__field(evsel, "pathname")) {
2643 evsel->handler = trace__vfs_getname;
2648 list_del_init(&evsel->core.node);
2649 evsel->evlist = NULL;
2650 evsel__delete(evsel);
2656 static struct evsel *perf_evsel__new_pgfault(u64 config)
2658 struct evsel *evsel;
2659 struct perf_event_attr attr = {
2660 .type = PERF_TYPE_SOFTWARE,
2664 attr.config = config;
2665 attr.sample_period = 1;
2667 event_attr_init(&attr);
2669 evsel = evsel__new(&attr);
2671 evsel->handler = trace__pgfault;
2676 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2678 const u32 type = event->header.type;
2679 struct evsel *evsel;
2681 if (type != PERF_RECORD_SAMPLE) {
2682 trace__process_event(trace, trace->host, event, sample);
2686 evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2687 if (evsel == NULL) {
2688 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2692 if (evswitch__discard(&trace->evswitch, evsel))
2695 trace__set_base_time(trace, evsel, sample);
2697 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
2698 sample->raw_data == NULL) {
2699 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2700 perf_evsel__name(evsel), sample->tid,
2701 sample->cpu, sample->raw_size);
2703 tracepoint_handler handler = evsel->handler;
2704 handler(trace, evsel, event, sample);
2707 if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2711 static int trace__add_syscall_newtp(struct trace *trace)
2714 struct evlist *evlist = trace->evlist;
2715 struct evsel *sys_enter, *sys_exit;
2717 sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2718 if (sys_enter == NULL)
2721 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2722 goto out_delete_sys_enter;
2724 sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2725 if (sys_exit == NULL)
2726 goto out_delete_sys_enter;
2728 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2729 goto out_delete_sys_exit;
2731 perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2732 perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2734 evlist__add(evlist, sys_enter);
2735 evlist__add(evlist, sys_exit);
2737 if (callchain_param.enabled && !trace->kernel_syscallchains) {
2739 * We're interested only in the user space callchain
2740 * leading to the syscall, allow overriding that for
2741 * debugging reasons using --kernel_syscall_callchains
2743 sys_exit->core.attr.exclude_callchain_kernel = 1;
2746 trace->syscalls.events.sys_enter = sys_enter;
2747 trace->syscalls.events.sys_exit = sys_exit;
2753 out_delete_sys_exit:
2754 evsel__delete_priv(sys_exit);
2755 out_delete_sys_enter:
2756 evsel__delete_priv(sys_enter);
2760 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2763 struct evsel *sys_exit;
2764 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2765 trace->ev_qualifier_ids.nr,
2766 trace->ev_qualifier_ids.entries);
2771 if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2773 sys_exit = trace->syscalls.events.sys_exit;
2774 err = perf_evsel__append_tp_filter(sys_exit, filter);
2785 #ifdef HAVE_LIBBPF_SUPPORT
2786 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
2788 if (trace->bpf_obj == NULL)
2791 return bpf_object__find_program_by_title(trace->bpf_obj, name);
2794 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
2795 const char *prog_name, const char *type)
2797 struct bpf_program *prog;
2799 if (prog_name == NULL) {
2800 char default_prog_name[256];
2801 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
2802 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2805 if (sc->fmt && sc->fmt->alias) {
2806 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
2807 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2811 goto out_unaugmented;
2814 prog = trace__find_bpf_program_by_title(trace, prog_name);
2821 pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
2822 prog_name, type, sc->name);
2824 return trace->syscalls.unaugmented_prog;
2827 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
2829 struct syscall *sc = trace__syscall_info(trace, NULL, id);
2834 sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2835 sc->bpf_prog.sys_exit = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit : NULL, "exit");
2838 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
2840 struct syscall *sc = trace__syscall_info(trace, NULL, id);
2841 return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2844 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
2846 struct syscall *sc = trace__syscall_info(trace, NULL, id);
2847 return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2850 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2852 struct syscall *sc = trace__syscall_info(trace, NULL, id);
2858 for (; arg < sc->nr_args; ++arg) {
2859 entry->string_args_len[arg] = 0;
2860 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2861 /* Should be set like strace -s strsize */
2862 entry->string_args_len[arg] = PATH_MAX;
2866 for (; arg < 6; ++arg)
2867 entry->string_args_len[arg] = 0;
2869 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2871 int fd = bpf_map__fd(trace->syscalls.map);
2872 struct bpf_map_syscall_entry value = {
2873 .enabled = !trace->not_ev_qualifier,
2878 for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2879 int key = trace->ev_qualifier_ids.entries[i];
2881 if (value.enabled) {
2882 trace__init_bpf_map_syscall_args(trace, key, &value);
2883 trace__init_syscall_bpf_progs(trace, key);
2886 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2894 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2896 int fd = bpf_map__fd(trace->syscalls.map);
2897 struct bpf_map_syscall_entry value = {
2902 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2904 trace__init_bpf_map_syscall_args(trace, key, &value);
2906 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2914 static int trace__init_syscalls_bpf_map(struct trace *trace)
2916 bool enabled = true;
2918 if (trace->ev_qualifier_ids.nr)
2919 enabled = trace->not_ev_qualifier;
2921 return __trace__init_syscalls_bpf_map(trace, enabled);
2924 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
2926 struct tep_format_field *field, *candidate_field;
2930 * We're only interested in syscalls that have a pointer:
2932 for (field = sc->args; field; field = field->next) {
2933 if (field->flags & TEP_FIELD_IS_POINTER)
2934 goto try_to_find_pair;
2940 for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
2941 struct syscall *pair = trace__syscall_info(trace, NULL, id);
2942 struct bpf_program *pair_prog;
2943 bool is_candidate = false;
2945 if (pair == NULL || pair == sc ||
2946 pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
2949 for (field = sc->args, candidate_field = pair->args;
2950 field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
2951 bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
2952 candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
2955 if (!candidate_is_pointer) {
2956 // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
2960 if (candidate_is_pointer) {
2961 // The candidate might copy a pointer we don't have, skip it.
2962 goto next_candidate;
2967 if (strcmp(field->type, candidate_field->type))
2968 goto next_candidate;
2970 is_candidate = true;
2974 goto next_candidate;
2977 * Check if the tentative pair syscall augmenter has more pointers, if it has,
2978 * then it may be collecting that and we then can't use it, as it would collect
2979 * more than what is common to the two syscalls.
2981 if (candidate_field) {
2982 for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
2983 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
2984 goto next_candidate;
2987 pair_prog = pair->bpf_prog.sys_enter;
2989 * If the pair isn't enabled, then its bpf_prog.sys_enter will not
2990 * have been searched for, so search it here and if it returns the
2991 * unaugmented one, then ignore it, otherwise we'll reuse that BPF
2992 * program for a filtered syscall on a non-filtered one.
2994 * For instance, we have "!syscalls:sys_enter_renameat" and that is
2995 * useful for "renameat2".
2997 if (pair_prog == NULL) {
2998 pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2999 if (pair_prog == trace->syscalls.unaugmented_prog)
3000 goto next_candidate;
3003 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3012 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3014 int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3015 map_exit_fd = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3018 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3021 if (!trace__syscall_enabled(trace, key))
3024 trace__init_syscall_bpf_progs(trace, key);
3026 // It'll get at least the "!raw_syscalls:unaugmented"
3027 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3028 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3031 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3032 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3038 * Now lets do a second pass looking for enabled syscalls without
3039 * an augmenter that have a signature that is a superset of another
3040 * syscall with an augmenter so that we can auto-reuse it.
3042 * I.e. if we have an augmenter for the "open" syscall that has
3045 * int open(const char *pathname, int flags, mode_t mode);
3047 * I.e. that will collect just the first string argument, then we
3048 * can reuse it for the 'creat' syscall, that has this signature:
3050 * int creat(const char *pathname, mode_t mode);
3054 * int stat(const char *pathname, struct stat *statbuf);
3055 * int lstat(const char *pathname, struct stat *statbuf);
3057 * Because the 'open' augmenter will collect the first arg as a string,
3058 * and leave alone all the other args, which already helps with
3059 * beautifying 'stat' and 'lstat''s pathname arg.
3061 * Then, in time, when 'stat' gets an augmenter that collects both
3062 * first and second arg (this one on the raw_syscalls:sys_exit prog
3063 * array tail call, then that one will be used.
3065 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3066 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3067 struct bpf_program *pair_prog;
3070 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3074 * For now we're just reusing the sys_enter prog, and if it
3075 * already has an augmenter, we don't need to find one.
3077 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3081 * Look at all the other syscalls for one that has a signature
3082 * that is close enough that we can share:
3084 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3085 if (pair_prog == NULL)
3088 sc->bpf_prog.sys_enter = pair_prog;
3091 * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3092 * with the fd for the program we're reusing:
3094 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3095 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3104 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3109 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3114 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3115 const char *name __maybe_unused)
3120 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3124 #endif // HAVE_LIBBPF_SUPPORT
3126 static int trace__set_ev_qualifier_filter(struct trace *trace)
3128 if (trace->syscalls.map)
3129 return trace__set_ev_qualifier_bpf_filter(trace);
3130 if (trace->syscalls.events.sys_enter)
3131 return trace__set_ev_qualifier_tp_filter(trace);
3135 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3136 size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3139 #ifdef HAVE_LIBBPF_SUPPORT
3141 int map_fd = bpf_map__fd(map);
3144 for (i = 0; i < npids; ++i) {
3145 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3153 static int trace__set_filter_loop_pids(struct trace *trace)
3155 unsigned int nr = 1, err;
3159 struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3161 while (thread && nr < ARRAY_SIZE(pids)) {
3162 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3167 if (!strcmp(thread__comm_str(parent), "sshd") ||
3168 strstarts(thread__comm_str(parent), "gnome-terminal")) {
3169 pids[nr++] = parent->tid;
3175 err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
3176 if (!err && trace->filter_pids.map)
3177 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3182 static int trace__set_filter_pids(struct trace *trace)
3186 * Better not use !target__has_task() here because we need to cover the
3187 * case where no threads were specified in the command line, but a
3188 * workload was, and in that case we will fill in the thread_map when
3189 * we fork the workload in perf_evlist__prepare_workload.
3191 if (trace->filter_pids.nr > 0) {
3192 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3193 trace->filter_pids.entries);
3194 if (!err && trace->filter_pids.map) {
3195 err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3196 trace->filter_pids.entries);
3198 } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3199 err = trace__set_filter_loop_pids(trace);
3205 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3207 struct evlist *evlist = trace->evlist;
3208 struct perf_sample sample;
3211 err = perf_evlist__parse_sample(evlist, event, &sample);
3213 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3215 trace__handle_event(trace, event, &sample);
3220 static int __trace__flush_events(struct trace *trace)
3222 u64 first = ordered_events__first_time(&trace->oe.data);
3223 u64 flush = trace->oe.last - NSEC_PER_SEC;
3225 /* Is there some thing to flush.. */
3226 if (first && first < flush)
3227 return ordered_events__flush_time(&trace->oe.data, flush);
3232 static int trace__flush_events(struct trace *trace)
3234 return !trace->sort_events ? 0 : __trace__flush_events(trace);
3237 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3241 if (!trace->sort_events)
3242 return __trace__deliver_event(trace, event);
3244 err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3245 if (err && err != -1)
3248 err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3252 return trace__flush_events(trace);
3255 static int ordered_events__deliver_event(struct ordered_events *oe,
3256 struct ordered_event *event)
3258 struct trace *trace = container_of(oe, struct trace, oe.data);
3260 return __trace__deliver_event(trace, event->event);
3263 static int trace__run(struct trace *trace, int argc, const char **argv)
3265 struct evlist *evlist = trace->evlist;
3266 struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3268 unsigned long before;
3269 const bool forks = argc > 0;
3270 bool draining = false;
3274 if (!trace->raw_augmented_syscalls) {
3275 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3276 goto out_error_raw_syscalls;
3278 if (trace->trace_syscalls)
3279 trace->vfs_getname = evlist__add_vfs_getname(evlist);
3282 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3283 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3284 if (pgfault_maj == NULL)
3286 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3287 evlist__add(evlist, pgfault_maj);
3290 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3291 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3292 if (pgfault_min == NULL)
3294 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3295 evlist__add(evlist, pgfault_min);
3299 perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3300 trace__sched_stat_runtime))
3301 goto out_error_sched_stat_runtime;
3304 * If a global cgroup was set, apply it to all the events without an
3305 * explicit cgroup. I.e.:
3307 * trace -G A -e sched:*switch
3309 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3310 * _and_ sched:sched_switch to the 'A' cgroup, while:
3312 * trace -e sched:*switch -G A
3314 * will only set the sched:sched_switch event to the 'A' cgroup, all the
3315 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3316 * a cgroup (on the root cgroup, sys wide, etc).
3320 * trace -G A -e sched:*switch -G B
3322 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3323 * to the 'B' cgroup.
3325 * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3326 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3329 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3331 err = perf_evlist__create_maps(evlist, &trace->opts.target);
3333 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3334 goto out_delete_evlist;
3337 err = trace__symbols_init(trace, evlist);
3339 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3340 goto out_delete_evlist;
3343 perf_evlist__config(evlist, &trace->opts, &callchain_param);
3345 signal(SIGCHLD, sig_handler);
3346 signal(SIGINT, sig_handler);
3349 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3352 fprintf(trace->output, "Couldn't run the workload!\n");
3353 goto out_delete_evlist;
3357 err = evlist__open(evlist);
3359 goto out_error_open;
3361 err = bpf__apply_obj_config();
3363 char errbuf[BUFSIZ];
3365 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3366 pr_err("ERROR: Apply config to BPF failed: %s\n",
3368 goto out_error_open;
3371 err = trace__set_filter_pids(trace);
3375 if (trace->syscalls.map)
3376 trace__init_syscalls_bpf_map(trace);
3378 if (trace->syscalls.prog_array.sys_enter)
3379 trace__init_syscalls_bpf_prog_array_maps(trace);
3381 if (trace->ev_qualifier_ids.nr > 0) {
3382 err = trace__set_ev_qualifier_filter(trace);
3386 if (trace->syscalls.events.sys_exit) {
3387 pr_debug("event qualifier tracepoint filter: %s\n",
3388 trace->syscalls.events.sys_exit->filter);
3393 * If the "close" syscall is not traced, then we will not have the
3394 * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3395 * fd->pathname table and were ending up showing the last value set by
3396 * syscalls opening a pathname and associating it with a descriptor or
3397 * reading it from /proc/pid/fd/ in cases where that doesn't make
3400 * So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3403 trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3405 err = perf_evlist__apply_filters(evlist, &evsel);
3407 goto out_error_apply_filters;
3409 if (trace->dump.map)
3410 bpf_map__fprintf(trace->dump.map, trace->output);
3412 err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3414 goto out_error_mmap;
3416 if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3417 evlist__enable(evlist);
3420 perf_evlist__start_workload(evlist);
3422 if (trace->opts.initial_delay) {
3423 usleep(trace->opts.initial_delay * 1000);
3424 evlist__enable(evlist);
3427 trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3428 evlist->core.threads->nr > 1 ||
3429 perf_evlist__first(evlist)->core.attr.inherit;
3432 * Now that we already used evsel->core.attr to ask the kernel to setup the
3433 * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3434 * trace__resolve_callchain(), allowing per-event max-stack settings
3435 * to override an explicitly set --max-stack global setting.
3437 evlist__for_each_entry(evlist, evsel) {
3438 if (evsel__has_callchain(evsel) &&
3439 evsel->core.attr.sample_max_stack == 0)
3440 evsel->core.attr.sample_max_stack = trace->max_stack;
3443 before = trace->nr_events;
3445 for (i = 0; i < evlist->nr_mmaps; i++) {
3446 union perf_event *event;
3449 md = &evlist->mmap[i];
3450 if (perf_mmap__read_init(md) < 0)
3453 while ((event = perf_mmap__read_event(md)) != NULL) {
3456 err = trace__deliver_event(trace, event);
3460 perf_mmap__consume(md);
3465 if (done && !draining) {
3466 evlist__disable(evlist);
3470 perf_mmap__read_done(md);
3473 if (trace->nr_events == before) {
3474 int timeout = done ? 100 : -1;
3476 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3477 if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3482 if (trace__flush_events(trace))
3490 thread__zput(trace->current);
3492 evlist__disable(evlist);
3494 if (trace->sort_events)
3495 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3499 trace__fprintf_thread_summary(trace, trace->output);
3501 if (trace->show_tool_stats) {
3502 fprintf(trace->output, "Stats:\n "
3503 " vfs_getname : %" PRIu64 "\n"
3504 " proc_getname: %" PRIu64 "\n",
3505 trace->stats.vfs_getname,
3506 trace->stats.proc_getname);
3511 trace__symbols__exit(trace);
3513 evlist__delete(evlist);
3514 cgroup__put(trace->cgroup);
3515 trace->evlist = NULL;
3516 trace->live = false;
3519 char errbuf[BUFSIZ];
3521 out_error_sched_stat_runtime:
3522 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3525 out_error_raw_syscalls:
3526 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3530 perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3534 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3537 fprintf(trace->output, "%s\n", errbuf);
3538 goto out_delete_evlist;
3540 out_error_apply_filters:
3541 fprintf(trace->output,
3542 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3543 evsel->filter, perf_evsel__name(evsel), errno,
3544 str_error_r(errno, errbuf, sizeof(errbuf)));
3545 goto out_delete_evlist;
3548 fprintf(trace->output, "Not enough memory to run!\n");
3549 goto out_delete_evlist;
3552 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3553 goto out_delete_evlist;
3556 static int trace__replay(struct trace *trace)
3558 const struct evsel_str_handler handlers[] = {
3559 { "probe:vfs_getname", trace__vfs_getname, },
3561 struct perf_data data = {
3563 .mode = PERF_DATA_MODE_READ,
3564 .force = trace->force,
3566 struct perf_session *session;
3567 struct evsel *evsel;
3570 trace->tool.sample = trace__process_sample;
3571 trace->tool.mmap = perf_event__process_mmap;
3572 trace->tool.mmap2 = perf_event__process_mmap2;
3573 trace->tool.comm = perf_event__process_comm;
3574 trace->tool.exit = perf_event__process_exit;
3575 trace->tool.fork = perf_event__process_fork;
3576 trace->tool.attr = perf_event__process_attr;
3577 trace->tool.tracing_data = perf_event__process_tracing_data;
3578 trace->tool.build_id = perf_event__process_build_id;
3579 trace->tool.namespaces = perf_event__process_namespaces;
3581 trace->tool.ordered_events = true;
3582 trace->tool.ordering_requires_timestamps = true;
3584 /* add tid to output */
3585 trace->multiple_threads = true;
3587 session = perf_session__new(&data, false, &trace->tool);
3588 if (IS_ERR(session))
3589 return PTR_ERR(session);
3591 if (trace->opts.target.pid)
3592 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3594 if (trace->opts.target.tid)
3595 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3597 if (symbol__init(&session->header.env) < 0)
3600 trace->host = &session->machines.host;
3602 err = perf_session__set_tracepoints_handlers(session, handlers);
3606 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3607 "raw_syscalls:sys_enter");
3608 /* older kernels have syscalls tp versus raw_syscalls */
3610 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3611 "syscalls:sys_enter");
3614 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3615 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3616 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3620 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3621 "raw_syscalls:sys_exit");
3623 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3624 "syscalls:sys_exit");
3626 (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3627 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3628 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3632 evlist__for_each_entry(session->evlist, evsel) {
3633 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
3634 (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3635 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3636 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3637 evsel->handler = trace__pgfault;
3642 err = perf_session__process_events(session);
3644 pr_err("Failed to process events, error %d", err);
3646 else if (trace->summary)
3647 trace__fprintf_thread_summary(trace, trace->output);
3650 perf_session__delete(session);
3655 static size_t trace__fprintf_threads_header(FILE *fp)
3659 printed = fprintf(fp, "\n Summary of events:\n\n");
3664 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3665 struct stats *stats;
3670 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3671 struct stats *stats = source->priv;
3673 entry->syscall = source->i;
3674 entry->stats = stats;
3675 entry->msecs = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3678 static size_t thread__dump_stats(struct thread_trace *ttrace,
3679 struct trace *trace, FILE *fp)
3684 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3686 if (syscall_stats == NULL)
3689 printed += fprintf(fp, "\n");
3691 printed += fprintf(fp, " syscall calls total min avg max stddev\n");
3692 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
3693 printed += fprintf(fp, " --------------- -------- --------- --------- --------- --------- ------\n");
3695 resort_rb__for_each_entry(nd, syscall_stats) {
3696 struct stats *stats = syscall_stats_entry->stats;
3698 double min = (double)(stats->min) / NSEC_PER_MSEC;
3699 double max = (double)(stats->max) / NSEC_PER_MSEC;
3700 double avg = avg_stats(stats);
3702 u64 n = (u64) stats->n;
3704 pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3705 avg /= NSEC_PER_MSEC;
3707 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3708 printed += fprintf(fp, " %-15s", sc->name);
3709 printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3710 n, syscall_stats_entry->msecs, min, avg);
3711 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3715 resort_rb__delete(syscall_stats);
3716 printed += fprintf(fp, "\n\n");
3721 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3724 struct thread_trace *ttrace = thread__priv(thread);
3730 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3732 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3733 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3734 printed += fprintf(fp, "%.1f%%", ratio);
3736 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3738 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3740 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3741 else if (fputc('\n', fp) != EOF)
3744 printed += thread__dump_stats(ttrace, trace, fp);
3749 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3751 return ttrace ? ttrace->nr_events : 0;
3754 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3755 struct thread *thread;
3758 entry->thread = rb_entry(nd, struct thread, rb_node);
3761 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3763 size_t printed = trace__fprintf_threads_header(fp);
3767 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3768 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3770 if (threads == NULL) {
3771 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3775 resort_rb__for_each_entry(nd, threads)
3776 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3778 resort_rb__delete(threads);
3783 static int trace__set_duration(const struct option *opt, const char *str,
3784 int unset __maybe_unused)
3786 struct trace *trace = opt->value;
3788 trace->duration_filter = atof(str);
3792 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3793 int unset __maybe_unused)
3797 struct trace *trace = opt->value;
3799 * FIXME: introduce a intarray class, plain parse csv and create a
3800 * { int nr, int entries[] } struct...
3802 struct intlist *list = intlist__new(str);
3807 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3808 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3810 if (trace->filter_pids.entries == NULL)
3813 trace->filter_pids.entries[0] = getpid();
3815 for (i = 1; i < trace->filter_pids.nr; ++i)
3816 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3818 intlist__delete(list);
3824 static int trace__open_output(struct trace *trace, const char *filename)
3828 if (!stat(filename, &st) && st.st_size) {
3829 char oldname[PATH_MAX];
3831 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3833 rename(filename, oldname);
3836 trace->output = fopen(filename, "w");
3838 return trace->output == NULL ? -errno : 0;
3841 static int parse_pagefaults(const struct option *opt, const char *str,
3842 int unset __maybe_unused)
3844 int *trace_pgfaults = opt->value;
3846 if (strcmp(str, "all") == 0)
3847 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3848 else if (strcmp(str, "maj") == 0)
3849 *trace_pgfaults |= TRACE_PFMAJ;
3850 else if (strcmp(str, "min") == 0)
3851 *trace_pgfaults |= TRACE_PFMIN;
3858 static void evlist__set_evsel_handler(struct evlist *evlist, void *handler)
3860 struct evsel *evsel;
3862 evlist__for_each_entry(evlist, evsel)
3863 evsel->handler = handler;
3866 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
3868 struct evsel *evsel;
3870 evlist__for_each_entry(evlist, evsel) {
3871 if (evsel->priv || !evsel->tp_format)
3874 if (strcmp(evsel->tp_format->system, "syscalls"))
3877 if (perf_evsel__init_syscall_tp(evsel))
3880 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3881 struct syscall_tp *sc = evsel->priv;
3883 if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3885 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3886 struct syscall_tp *sc = evsel->priv;
3888 if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3897 * XXX: Hackish, just splitting the combined -e+--event (syscalls
3898 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3899 * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3901 * It'd be better to introduce a parse_options() variant that would return a
3902 * list with the terms it didn't match to an event...
3904 static int trace__parse_events_option(const struct option *opt, const char *str,
3905 int unset __maybe_unused)
3907 struct trace *trace = (struct trace *)opt->value;
3908 const char *s = str;
3909 char *sep = NULL, *lists[2] = { NULL, NULL, };
3910 int len = strlen(str) + 1, err = -1, list, idx;
3911 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3912 char group_name[PATH_MAX];
3913 struct syscall_fmt *fmt;
3915 if (strace_groups_dir == NULL)
3920 trace->not_ev_qualifier = true;
3924 if ((sep = strchr(s, ',')) != NULL)
3928 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3929 syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3934 fmt = syscall_fmt__find_by_alias(s);
3939 path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3940 if (access(group_name, R_OK) == 0)
3945 sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3947 lists[list] = malloc(len);
3948 if (lists[list] == NULL)
3950 strcpy(lists[list], s);
3960 if (lists[1] != NULL) {
3961 struct strlist_config slist_config = {
3962 .dirname = strace_groups_dir,
3965 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3966 if (trace->ev_qualifier == NULL) {
3967 fputs("Not enough memory to parse event qualifier", trace->output);
3971 if (trace__validate_ev_qualifier(trace))
3973 trace->trace_syscalls = true;
3979 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3980 "event selector. use 'perf list' to list available events",
3981 parse_events_option);
3982 err = parse_events_option(&o, lists[0], 0);
3991 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3993 struct trace *trace = opt->value;
3995 if (!list_empty(&trace->evlist->core.entries))
3996 return parse_cgroups(opt, str, unset);
3998 trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4003 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4005 if (trace->bpf_obj == NULL)
4008 return bpf_object__find_map_by_name(trace->bpf_obj, name);
4011 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4013 trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4016 static void trace__set_bpf_map_syscalls(struct trace *trace)
4018 trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4019 trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4020 trace->syscalls.prog_array.sys_exit = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4023 static int trace__config(const char *var, const char *value, void *arg)
4025 struct trace *trace = arg;
4028 if (!strcmp(var, "trace.add_events")) {
4029 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4030 "event selector. use 'perf list' to list available events",
4031 parse_events_option);
4033 * We can't propagate parse_event_option() return, as it is 1
4034 * for failure while perf_config() expects -1.
4036 if (parse_events_option(&o, value, 0))
4038 } else if (!strcmp(var, "trace.show_timestamp")) {
4039 trace->show_tstamp = perf_config_bool(var, value);
4040 } else if (!strcmp(var, "trace.show_duration")) {
4041 trace->show_duration = perf_config_bool(var, value);
4042 } else if (!strcmp(var, "trace.show_arg_names")) {
4043 trace->show_arg_names = perf_config_bool(var, value);
4044 if (!trace->show_arg_names)
4045 trace->show_zeros = true;
4046 } else if (!strcmp(var, "trace.show_zeros")) {
4047 bool new_show_zeros = perf_config_bool(var, value);
4048 if (!trace->show_arg_names && !new_show_zeros) {
4049 pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4052 trace->show_zeros = new_show_zeros;
4053 } else if (!strcmp(var, "trace.show_prefix")) {
4054 trace->show_string_prefix = perf_config_bool(var, value);
4055 } else if (!strcmp(var, "trace.no_inherit")) {
4056 trace->opts.no_inherit = perf_config_bool(var, value);
4057 } else if (!strcmp(var, "trace.args_alignment")) {
4058 int args_alignment = 0;
4059 if (perf_config_int(&args_alignment, var, value) == 0)
4060 trace->args_alignment = args_alignment;
4066 int cmd_trace(int argc, const char **argv)
4068 const char *trace_usage[] = {
4069 "perf trace [<options>] [<command>]",
4070 "perf trace [<options>] -- <command> [<options>]",
4071 "perf trace record [<options>] [<command>]",
4072 "perf trace record [<options>] -- <command> [<options>]",
4075 struct trace trace = {
4081 .user_freq = UINT_MAX,
4082 .user_interval = ULLONG_MAX,
4083 .no_buffering = true,
4084 .mmap_pages = UINT_MAX,
4088 .show_tstamp = true,
4089 .show_duration = true,
4090 .show_arg_names = true,
4091 .args_alignment = 70,
4092 .trace_syscalls = false,
4093 .kernel_syscallchains = false,
4094 .max_stack = UINT_MAX,
4095 .max_events = ULONG_MAX,
4097 const char *map_dump_str = NULL;
4098 const char *output_name = NULL;
4099 const struct option trace_options[] = {
4100 OPT_CALLBACK('e', "event", &trace, "event",
4101 "event/syscall selector. use 'perf list' to list available events",
4102 trace__parse_events_option),
4103 OPT_BOOLEAN(0, "comm", &trace.show_comm,
4104 "show the thread COMM next to its id"),
4105 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4106 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4107 trace__parse_events_option),
4108 OPT_STRING('o', "output", &output_name, "file", "output file name"),
4109 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4110 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4111 "trace events on existing process id"),
4112 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4113 "trace events on existing thread id"),
4114 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4115 "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4116 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4117 "system-wide collection from all CPUs"),
4118 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4119 "list of cpus to monitor"),
4120 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4121 "child tasks do not inherit counters"),
4122 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4123 "number of mmap data pages",
4124 perf_evlist__parse_mmap_pages),
4125 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4127 OPT_CALLBACK(0, "duration", &trace, "float",
4128 "show only events with duration > N.M ms",
4129 trace__set_duration),
4130 #ifdef HAVE_LIBBPF_SUPPORT
4131 OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4133 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4134 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4135 OPT_BOOLEAN('T', "time", &trace.full_time,
4136 "Show full timestamp, not time relative to first start"),
4137 OPT_BOOLEAN(0, "failure", &trace.failure_only,
4138 "Show only syscalls that failed"),
4139 OPT_BOOLEAN('s', "summary", &trace.summary_only,
4140 "Show only syscall summary with statistics"),
4141 OPT_BOOLEAN('S', "with-summary", &trace.summary,
4142 "Show all syscalls and summary with statistics"),
4143 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4144 "Trace pagefaults", parse_pagefaults, "maj"),
4145 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4146 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4147 OPT_CALLBACK(0, "call-graph", &trace.opts,
4148 "record_mode[,record_size]", record_callchain_help,
4149 &record_parse_callchain_opt),
4150 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4151 "Show the kernel callchains on the syscall exit path"),
4152 OPT_ULONG(0, "max-events", &trace.max_events,
4153 "Set the maximum number of events to print, exit after that is reached. "),
4154 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4155 "Set the minimum stack depth when parsing the callchain, "
4156 "anything below the specified depth will be ignored."),
4157 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4158 "Set the maximum stack depth when parsing the callchain, "
4159 "anything beyond the specified depth will be ignored. "
4160 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4161 OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4162 "Sort batch of events before processing, use if getting out of order events"),
4163 OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4164 "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4165 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4166 "per thread proc mmap processing timeout in ms"),
4167 OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4168 trace__parse_cgroups),
4169 OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4170 "ms to wait before starting measurement after program "
4172 OPTS_EVSWITCH(&trace.evswitch),
4175 bool __maybe_unused max_stack_user_set = true;
4176 bool mmap_pages_user_set = true;
4177 struct evsel *evsel;
4178 const char * const trace_subcommands[] = { "record", NULL };
4182 signal(SIGSEGV, sighandler_dump_stack);
4183 signal(SIGFPE, sighandler_dump_stack);
4185 trace.evlist = evlist__new();
4186 trace.sctbl = syscalltbl__new();
4188 if (trace.evlist == NULL || trace.sctbl == NULL) {
4189 pr_err("Not enough memory to run!\n");
4195 * Parsing .perfconfig may entail creating a BPF event, that may need
4196 * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4197 * is too small. This affects just this process, not touching the
4198 * global setting. If it fails we'll get something in 'perf trace -v'
4199 * to help diagnose the problem.
4201 rlimit__bump_memlock();
4203 err = perf_config(trace__config, &trace);
4207 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4208 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4210 if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4211 usage_with_options_msg(trace_usage, trace_options,
4212 "cgroup monitoring only available in system-wide mode");
4215 evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4216 if (IS_ERR(evsel)) {
4217 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4218 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4223 trace.syscalls.events.augmented = evsel;
4225 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4226 if (evsel == NULL) {
4227 pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4231 if (evsel->bpf_obj == NULL) {
4232 pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4236 trace.bpf_obj = evsel->bpf_obj;
4238 trace__set_bpf_map_filtered_pids(&trace);
4239 trace__set_bpf_map_syscalls(&trace);
4240 trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4243 err = bpf__setup_stdout(trace.evlist);
4245 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4246 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4253 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4254 if (trace.dump.map == NULL) {
4255 pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4260 if (trace.trace_pgfaults) {
4261 trace.opts.sample_address = true;
4262 trace.opts.sample_time = true;
4265 if (trace.opts.mmap_pages == UINT_MAX)
4266 mmap_pages_user_set = false;
4268 if (trace.max_stack == UINT_MAX) {
4269 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4270 max_stack_user_set = false;
4273 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4274 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4275 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4279 if (callchain_param.enabled) {
4280 if (!mmap_pages_user_set && geteuid() == 0)
4281 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4283 symbol_conf.use_callchain = true;
4286 if (trace.evlist->core.nr_entries > 0) {
4287 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
4288 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4289 perror("failed to set syscalls:* tracepoint fields");
4294 if (trace.sort_events) {
4295 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4296 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4300 * If we are augmenting syscalls, then combine what we put in the
4301 * __augmented_syscalls__ BPF map with what is in the
4302 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4303 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4305 * We'll switch to look at two BPF maps, one for sys_enter and the
4306 * other for sys_exit when we start augmenting the sys_exit paths with
4307 * buffers that are being copied from kernel to userspace, think 'read'
4310 if (trace.syscalls.events.augmented) {
4311 evlist__for_each_entry(trace.evlist, evsel) {
4312 bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4314 if (raw_syscalls_sys_exit) {
4315 trace.raw_augmented_syscalls = true;
4316 goto init_augmented_syscall_tp;
4319 if (trace.syscalls.events.augmented->priv == NULL &&
4320 strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
4321 struct evsel *augmented = trace.syscalls.events.augmented;
4322 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
4323 perf_evsel__init_augmented_syscall_tp_args(augmented))
4326 * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4327 * Above we made sure we can get from the payload the tp fields
4328 * that we get from syscalls:sys_enter tracefs format file.
4330 augmented->handler = trace__sys_enter;
4332 * Now we do the same for the *syscalls:sys_enter event so that
4333 * if we handle it directly, i.e. if the BPF prog returns 0 so
4334 * as not to filter it, then we'll handle it just like we would
4335 * for the BPF_OUTPUT one:
4337 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
4338 perf_evsel__init_augmented_syscall_tp_args(evsel))
4340 evsel->handler = trace__sys_enter;
4343 if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
4344 struct syscall_tp *sc;
4345 init_augmented_syscall_tp:
4346 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
4350 * For now with BPF raw_augmented we hook into
4351 * raw_syscalls:sys_enter and there we get all
4352 * 6 syscall args plus the tracepoint common
4353 * fields and the syscall_nr (another long).
4354 * So we check if that is the case and if so
4355 * don't look after the sc->args_size but
4356 * always after the full raw_syscalls:sys_enter
4357 * payload, which is fixed.
4359 * We'll revisit this later to pass
4360 * s->args_size to the BPF augmenter (now
4361 * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4362 * so that it copies only what we need for each
4363 * syscall, like what happens when we use
4364 * syscalls:sys_enter_NAME, so that we reduce
4365 * the kernel/userspace traffic to just what is
4366 * needed for each syscall.
4368 if (trace.raw_augmented_syscalls)
4369 trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4370 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4371 evsel->handler = trace__sys_exit;
4376 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4377 return trace__record(&trace, argc-1, &argv[1]);
4379 /* summary_only implies summary option, but don't overwrite summary if set */
4380 if (trace.summary_only)
4381 trace.summary = trace.summary_only;
4383 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4384 trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4385 trace.trace_syscalls = true;
4388 if (output_name != NULL) {
4389 err = trace__open_output(&trace, output_name);
4391 perror("failed to create output file");
4396 err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
4400 err = target__validate(&trace.opts.target);
4402 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4403 fprintf(trace.output, "%s", bf);
4407 err = target__parse_uid(&trace.opts.target);
4409 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4410 fprintf(trace.output, "%s", bf);
4414 if (!argc && target__none(&trace.opts.target))
4415 trace.opts.target.system_wide = true;
4418 err = trace__replay(&trace);
4420 err = trace__run(&trace, argc, argv);
4423 if (output_name != NULL)
4424 fclose(trace.output);