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 <api/fs/tracing_path.h>
19 #ifdef HAVE_LIBBPF_SUPPORT
21 #include <bpf/libbpf.h>
23 #include "bpf_skel/augmented_raw_syscalls.skel.h"
26 #include "util/bpf_map.h"
27 #include "util/rlimit.h"
29 #include "util/cgroup.h"
30 #include "util/color.h"
31 #include "util/config.h"
32 #include "util/debug.h"
35 #include "util/event.h"
36 #include "util/evsel.h"
37 #include "util/evsel_fprintf.h"
38 #include "util/synthetic-events.h"
39 #include "util/evlist.h"
40 #include "util/evswitch.h"
41 #include "util/mmap.h"
42 #include <subcmd/pager.h>
43 #include <subcmd/exec-cmd.h>
44 #include "util/machine.h"
46 #include "util/symbol.h"
47 #include "util/path.h"
48 #include "util/session.h"
49 #include "util/thread.h"
50 #include <subcmd/parse-options.h>
51 #include "util/strlist.h"
52 #include "util/intlist.h"
53 #include "util/thread_map.h"
54 #include "util/stat.h"
55 #include "util/tool.h"
56 #include "util/util.h"
57 #include "trace/beauty/beauty.h"
58 #include "trace-event.h"
59 #include "util/parse-events.h"
60 #include "util/tracepoint.h"
61 #include "callchain.h"
62 #include "print_binary.h"
64 #include "syscalltbl.h"
65 #include "rb_resort.h"
74 #include <linux/err.h>
75 #include <linux/filter.h>
76 #include <linux/kernel.h>
77 #include <linux/random.h>
78 #include <linux/stringify.h>
79 #include <linux/time64.h>
80 #include <linux/zalloc.h>
82 #include <sys/sysmacros.h>
84 #include <linux/ctype.h>
85 #include <perf/mmap.h>
87 #ifdef HAVE_LIBTRACEEVENT
88 #include <traceevent/event-parse.h>
92 # define O_CLOEXEC 02000000
95 #ifndef F_LINUX_SPECIFIC_BASE
96 # define F_LINUX_SPECIFIC_BASE 1024
99 #define RAW_SYSCALL_ARGS_NUM 6
102 * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
104 struct syscall_arg_fmt {
105 size_t (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
106 bool (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
107 unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
110 u16 nr_entries; // for arrays
118 const char *sys_enter,
121 struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
129 struct perf_tool tool;
130 struct syscalltbl *sctbl;
132 struct syscall *table;
134 struct evsel *sys_enter,
140 struct augmented_raw_syscalls_bpf *skel;
142 struct record_opts opts;
143 struct evlist *evlist;
144 struct machine *host;
145 struct thread *current;
146 struct cgroup *cgroup;
149 unsigned long nr_events;
150 unsigned long nr_events_printed;
151 unsigned long max_events;
152 struct evswitch evswitch;
153 struct strlist *ev_qualifier;
163 double duration_filter;
169 unsigned int max_stack;
170 unsigned int min_stack;
171 int raw_augmented_syscalls_args_size;
172 bool raw_augmented_syscalls;
173 bool fd_path_disabled;
175 bool not_ev_qualifier;
179 bool multiple_threads;
186 bool show_tool_stats;
188 bool libtraceevent_print;
189 bool kernel_syscallchains;
195 bool show_string_prefix;
199 char *perfconfig_events;
201 struct ordered_events data;
209 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
210 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
214 #define TP_UINT_FIELD(bits) \
215 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
218 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
227 #define TP_UINT_FIELD__SWAPPED(bits) \
228 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
231 memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
232 return bswap_##bits(value);\
235 TP_UINT_FIELD__SWAPPED(16);
236 TP_UINT_FIELD__SWAPPED(32);
237 TP_UINT_FIELD__SWAPPED(64);
239 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
241 field->offset = offset;
245 field->integer = tp_field__u8;
248 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
251 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
254 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
263 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
265 return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
268 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
270 return sample->raw_data + field->offset;
273 static int __tp_field__init_ptr(struct tp_field *field, int offset)
275 field->offset = offset;
276 field->pointer = tp_field__ptr;
280 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
282 return __tp_field__init_ptr(field, format_field->offset);
288 struct tp_field args, ret;
293 * The evsel->priv as used by 'perf trace'
294 * sc: for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
295 * fmt: for all the other tracepoints
298 struct syscall_tp sc;
299 struct syscall_arg_fmt *fmt;
302 static struct evsel_trace *evsel_trace__new(void)
304 return zalloc(sizeof(struct evsel_trace));
307 static void evsel_trace__delete(struct evsel_trace *et)
317 * Used with raw_syscalls:sys_{enter,exit} and with the
318 * syscalls:sys_{enter,exit}_SYSCALL tracepoints
320 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
322 struct evsel_trace *et = evsel->priv;
327 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
329 if (evsel->priv == NULL) {
330 evsel->priv = evsel_trace__new();
331 if (evsel->priv == NULL)
335 return __evsel__syscall_tp(evsel);
339 * Used with all the other tracepoints.
341 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
343 struct evsel_trace *et = evsel->priv;
348 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
350 struct evsel_trace *et = evsel->priv;
352 if (evsel->priv == NULL) {
353 et = evsel->priv = evsel_trace__new();
359 if (et->fmt == NULL) {
360 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
365 return __evsel__syscall_arg_fmt(evsel);
368 evsel_trace__delete(evsel->priv);
373 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
375 struct tep_format_field *format_field = evsel__field(evsel, name);
377 if (format_field == NULL)
380 return tp_field__init_uint(field, format_field, evsel->needs_swap);
383 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
384 ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
385 evsel__init_tp_uint_field(evsel, &sc->name, #name); })
387 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
389 struct tep_format_field *format_field = evsel__field(evsel, name);
391 if (format_field == NULL)
394 return tp_field__init_ptr(field, format_field);
397 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
398 ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
399 evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
401 static void evsel__delete_priv(struct evsel *evsel)
404 evsel__delete(evsel);
407 static int evsel__init_syscall_tp(struct evsel *evsel)
409 struct syscall_tp *sc = evsel__syscall_tp(evsel);
412 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
413 evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
422 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
424 struct syscall_tp *sc = evsel__syscall_tp(evsel);
427 struct tep_format_field *syscall_id = evsel__field(tp, "id");
428 if (syscall_id == NULL)
429 syscall_id = evsel__field(tp, "__syscall_nr");
430 if (syscall_id == NULL ||
431 __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
440 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
442 struct syscall_tp *sc = __evsel__syscall_tp(evsel);
444 return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
447 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
449 struct syscall_tp *sc = __evsel__syscall_tp(evsel);
451 return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
454 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
456 if (evsel__syscall_tp(evsel) != NULL) {
457 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
460 evsel->handler = handler;
467 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
469 struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
471 /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
473 evsel = evsel__newtp("syscalls", direction);
478 if (evsel__init_raw_syscall_tp(evsel, handler))
484 evsel__delete_priv(evsel);
488 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
489 ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
490 fields->name.integer(&fields->name, sample); })
492 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
493 ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
494 fields->name.pointer(&fields->name, sample); })
496 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
498 int idx = val - sa->offset;
500 if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
501 size_t printed = scnprintf(bf, size, intfmt, val);
503 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
507 return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
510 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
512 int idx = val - sa->offset;
514 if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
515 size_t printed = scnprintf(bf, size, intfmt, val);
517 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
521 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
524 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
526 struct syscall_arg *arg)
528 return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
531 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
532 struct syscall_arg *arg)
534 return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
537 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
539 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
541 return strarray__strtoul(arg->parm, bf, size, ret);
544 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
546 return strarray__strtoul_flags(arg->parm, bf, size, ret);
549 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
551 return strarrays__strtoul(arg->parm, bf, size, ret);
554 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
556 return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
559 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
564 for (i = 0; i < sas->nr_entries; ++i) {
565 struct strarray *sa = sas->entries[i];
566 int idx = val - sa->offset;
568 if (idx >= 0 && idx < sa->nr_entries) {
569 if (sa->entries[idx] == NULL)
571 return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
575 printed = scnprintf(bf, size, intfmt, val);
577 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
581 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
585 for (i = 0; i < sa->nr_entries; ++i) {
586 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
587 *ret = sa->offset + i;
595 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
598 char *tok = bf, *sep, *end;
605 sep = memchr(tok, '|', size);
607 size -= sep - tok + 1;
610 while (end > tok && isspace(*end))
613 toklen = end - tok + 1;
616 while (isspace(*tok))
619 if (isalpha(*tok) || *tok == '_') {
620 if (!strarray__strtoul(sa, tok, toklen, &val))
623 val = strtoul(tok, NULL, 0);
625 *ret |= (1 << (val - 1));
635 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
639 for (i = 0; i < sas->nr_entries; ++i) {
640 struct strarray *sa = sas->entries[i];
642 if (strarray__strtoul(sa, bf, size, ret))
649 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
650 struct syscall_arg *arg)
652 return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
656 #define AT_FDCWD -100
659 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
660 struct syscall_arg *arg)
663 const char *prefix = "AT_FD";
666 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
668 return syscall_arg__scnprintf_fd(bf, size, arg);
671 #define SCA_FDAT syscall_arg__scnprintf_fd_at
673 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
674 struct syscall_arg *arg);
676 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
678 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
680 return scnprintf(bf, size, "%#lx", arg->val);
683 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
686 return scnprintf(bf, size, "NULL");
687 return syscall_arg__scnprintf_hex(bf, size, arg);
690 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
692 return scnprintf(bf, size, "%d", arg->val);
695 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
697 return scnprintf(bf, size, "%ld", arg->val);
700 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
702 // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
703 // fill missing comms using thread__set_comm()...
704 // here or in a special syscall_arg__scnprintf_pid_sched_tp...
705 return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
708 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
710 static const char *bpf_cmd[] = {
711 "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
712 "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
713 "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
714 "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
715 "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
716 "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
717 "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
718 "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
719 "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
720 "LINK_DETACH", "PROG_BIND_MAP",
722 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
724 static const char *fsmount_flags[] = {
727 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
729 #include "trace/beauty/generated/fsconfig_arrays.c"
731 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
733 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
734 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
736 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
737 static DEFINE_STRARRAY(itimers, "ITIMER_");
739 static const char *keyctl_options[] = {
740 "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
741 "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
742 "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
743 "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
744 "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
746 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
748 static const char *whences[] = { "SET", "CUR", "END",
756 static DEFINE_STRARRAY(whences, "SEEK_");
758 static const char *fcntl_cmds[] = {
759 "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
760 "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
761 "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
764 static DEFINE_STRARRAY(fcntl_cmds, "F_");
766 static const char *fcntl_linux_specific_cmds[] = {
767 "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
768 "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
769 "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
772 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
774 static struct strarray *fcntl_cmds_arrays[] = {
775 &strarray__fcntl_cmds,
776 &strarray__fcntl_linux_specific_cmds,
779 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
781 static const char *rlimit_resources[] = {
782 "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
783 "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
786 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
788 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
789 static DEFINE_STRARRAY(sighow, "SIG_");
791 static const char *clockid[] = {
792 "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
793 "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
794 "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
796 static DEFINE_STRARRAY(clockid, "CLOCK_");
798 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
799 struct syscall_arg *arg)
801 bool show_prefix = arg->show_string_prefix;
802 const char *suffix = "_OK";
806 if (mode == F_OK) /* 0 */
807 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
809 if (mode & n##_OK) { \
810 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
820 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
825 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
827 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
828 struct syscall_arg *arg);
830 #define SCA_FILENAME syscall_arg__scnprintf_filename
832 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
833 struct syscall_arg *arg)
835 bool show_prefix = arg->show_string_prefix;
836 const char *prefix = "O_";
837 int printed = 0, flags = arg->val;
840 if (flags & O_##n) { \
841 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
850 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
855 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
857 #ifndef GRND_NONBLOCK
858 #define GRND_NONBLOCK 0x0001
861 #define GRND_RANDOM 0x0002
864 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
865 struct syscall_arg *arg)
867 bool show_prefix = arg->show_string_prefix;
868 const char *prefix = "GRND_";
869 int printed = 0, flags = arg->val;
872 if (flags & GRND_##n) { \
873 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
874 flags &= ~GRND_##n; \
882 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
887 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
889 #define STRARRAY(name, array) \
890 { .scnprintf = SCA_STRARRAY, \
891 .strtoul = STUL_STRARRAY, \
892 .parm = &strarray__##array, }
894 #define STRARRAY_FLAGS(name, array) \
895 { .scnprintf = SCA_STRARRAY_FLAGS, \
896 .strtoul = STUL_STRARRAY_FLAGS, \
897 .parm = &strarray__##array, }
899 #include "trace/beauty/arch_errno_names.c"
900 #include "trace/beauty/eventfd.c"
901 #include "trace/beauty/futex_op.c"
902 #include "trace/beauty/futex_val3.c"
903 #include "trace/beauty/mmap.c"
904 #include "trace/beauty/mode_t.c"
905 #include "trace/beauty/msg_flags.c"
906 #include "trace/beauty/open_flags.c"
907 #include "trace/beauty/perf_event_open.c"
908 #include "trace/beauty/pid.c"
909 #include "trace/beauty/sched_policy.c"
910 #include "trace/beauty/seccomp.c"
911 #include "trace/beauty/signum.c"
912 #include "trace/beauty/socket_type.c"
913 #include "trace/beauty/waitid_options.c"
915 static const struct syscall_fmt syscall_fmts[] = {
917 .arg = { [1] = { .scnprintf = SCA_ACCMODE, /* mode */ }, }, },
918 { .name = "arch_prctl",
919 .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
920 [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
922 .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
923 [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
924 [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
926 .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
927 { .name = "brk", .hexret = true,
928 .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
929 { .name = "clock_gettime",
930 .arg = { [0] = STRARRAY(clk_id, clockid), }, },
931 { .name = "clock_nanosleep",
932 .arg = { [2] = { .scnprintf = SCA_TIMESPEC, /* rqtp */ }, }, },
933 { .name = "clone", .errpid = true, .nr_args = 5,
934 .arg = { [0] = { .name = "flags", .scnprintf = SCA_CLONE_FLAGS, },
935 [1] = { .name = "child_stack", .scnprintf = SCA_HEX, },
936 [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
937 [3] = { .name = "child_tidptr", .scnprintf = SCA_HEX, },
938 [4] = { .name = "tls", .scnprintf = SCA_HEX, }, }, },
940 .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
942 .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
943 [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
944 [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
945 { .name = "epoll_ctl",
946 .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
947 { .name = "eventfd2",
948 .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
949 { .name = "fchmodat",
950 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
951 { .name = "fchownat",
952 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
954 .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
955 .strtoul = STUL_STRARRAYS,
956 .parm = &strarrays__fcntl_cmds_arrays,
957 .show_zero = true, },
958 [2] = { .scnprintf = SCA_FCNTL_ARG, /* arg */ }, }, },
960 .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
961 { .name = "fsconfig",
962 .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
964 .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
965 [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
967 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
968 [1] = { .scnprintf = SCA_FILENAME, /* path */ },
969 [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
970 { .name = "fstat", .alias = "newfstat", },
971 { .name = "fstatat", .alias = "newfstatat", },
973 .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
974 [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
975 { .name = "futimesat",
976 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
977 { .name = "getitimer",
978 .arg = { [0] = STRARRAY(which, itimers), }, },
979 { .name = "getpid", .errpid = true, },
980 { .name = "getpgid", .errpid = true, },
981 { .name = "getppid", .errpid = true, },
982 { .name = "getrandom",
983 .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
984 { .name = "getrlimit",
985 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
986 { .name = "getsockopt",
987 .arg = { [1] = STRARRAY(level, socket_level), }, },
988 { .name = "gettid", .errpid = true, },
991 #if defined(__i386__) || defined(__x86_64__)
993 * FIXME: Make this available to all arches.
995 [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
996 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
998 [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
1000 { .name = "kcmp", .nr_args = 5,
1001 .arg = { [0] = { .name = "pid1", .scnprintf = SCA_PID, },
1002 [1] = { .name = "pid2", .scnprintf = SCA_PID, },
1003 [2] = { .name = "type", .scnprintf = SCA_KCMP_TYPE, },
1004 [3] = { .name = "idx1", .scnprintf = SCA_KCMP_IDX, },
1005 [4] = { .name = "idx2", .scnprintf = SCA_KCMP_IDX, }, }, },
1007 .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1009 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1011 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1013 .arg = { [2] = STRARRAY(whence, whences), }, },
1014 { .name = "lstat", .alias = "newlstat", },
1015 { .name = "madvise",
1016 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
1017 [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1018 { .name = "mkdirat",
1019 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1020 { .name = "mknodat",
1021 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1022 { .name = "mmap", .hexret = true,
1023 /* The standard mmap maps to old_mmap on s390x */
1024 #if defined(__s390x__)
1025 .alias = "old_mmap",
1027 .arg = { [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
1028 [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1029 .strtoul = STUL_STRARRAY_FLAGS,
1030 .parm = &strarray__mmap_flags, },
1031 [5] = { .scnprintf = SCA_HEX, /* offset */ }, }, },
1033 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1034 [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1035 .mask_val = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1036 { .name = "move_mount",
1037 .arg = { [0] = { .scnprintf = SCA_FDAT, /* from_dfd */ },
1038 [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1039 [2] = { .scnprintf = SCA_FDAT, /* to_dfd */ },
1040 [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1041 [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1042 { .name = "mprotect",
1043 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
1044 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ }, }, },
1045 { .name = "mq_unlink",
1046 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1047 { .name = "mremap", .hexret = true,
1048 .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1049 { .name = "name_to_handle_at",
1050 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1051 { .name = "newfstatat",
1052 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1054 .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1055 { .name = "open_by_handle_at",
1056 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
1057 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1059 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ },
1060 [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1061 { .name = "perf_event_open",
1062 .arg = { [0] = { .scnprintf = SCA_PERF_ATTR, /* attr */ },
1063 [2] = { .scnprintf = SCA_INT, /* cpu */ },
1064 [3] = { .scnprintf = SCA_FD, /* group_fd */ },
1065 [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1067 .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1068 { .name = "pkey_alloc",
1069 .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS, /* access_rights */ }, }, },
1070 { .name = "pkey_free",
1071 .arg = { [0] = { .scnprintf = SCA_INT, /* key */ }, }, },
1072 { .name = "pkey_mprotect",
1073 .arg = { [0] = { .scnprintf = SCA_HEX, /* start */ },
1074 [2] = { .scnprintf = SCA_MMAP_PROT, /* prot */ },
1075 [3] = { .scnprintf = SCA_INT, /* pkey */ }, }, },
1076 { .name = "poll", .timeout = true, },
1077 { .name = "ppoll", .timeout = true, },
1079 .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1080 .strtoul = STUL_STRARRAY,
1081 .parm = &strarray__prctl_options, },
1082 [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1083 [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1084 { .name = "pread", .alias = "pread64", },
1085 { .name = "preadv", .alias = "pread", },
1086 { .name = "prlimit64",
1087 .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1088 { .name = "pwrite", .alias = "pwrite64", },
1089 { .name = "readlinkat",
1090 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1091 { .name = "recvfrom",
1092 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1093 { .name = "recvmmsg",
1094 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1095 { .name = "recvmsg",
1096 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1097 { .name = "renameat",
1098 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1099 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1100 { .name = "renameat2",
1101 .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1102 [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1103 [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1104 { .name = "rt_sigaction",
1105 .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1106 { .name = "rt_sigprocmask",
1107 .arg = { [0] = STRARRAY(how, sighow), }, },
1108 { .name = "rt_sigqueueinfo",
1109 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1110 { .name = "rt_tgsigqueueinfo",
1111 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1112 { .name = "sched_setscheduler",
1113 .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1114 { .name = "seccomp",
1115 .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP, /* op */ },
1116 [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1117 { .name = "select", .timeout = true, },
1118 { .name = "sendfile", .alias = "sendfile64", },
1119 { .name = "sendmmsg",
1120 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1121 { .name = "sendmsg",
1122 .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1124 .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1125 [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1126 { .name = "set_tid_address", .errpid = true, },
1127 { .name = "setitimer",
1128 .arg = { [0] = STRARRAY(which, itimers), }, },
1129 { .name = "setrlimit",
1130 .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1131 { .name = "setsockopt",
1132 .arg = { [1] = STRARRAY(level, socket_level), }, },
1134 .arg = { [0] = STRARRAY(family, socket_families),
1135 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1136 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1137 { .name = "socketpair",
1138 .arg = { [0] = STRARRAY(family, socket_families),
1139 [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1140 [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1141 { .name = "stat", .alias = "newstat", },
1143 .arg = { [0] = { .scnprintf = SCA_FDAT, /* fdat */ },
1144 [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1145 [3] = { .scnprintf = SCA_STATX_MASK, /* mask */ }, }, },
1146 { .name = "swapoff",
1147 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1149 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1150 { .name = "symlinkat",
1151 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1152 { .name = "sync_file_range",
1153 .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1155 .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1157 .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1158 { .name = "umount2", .alias = "umount",
1159 .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1160 { .name = "uname", .alias = "newuname", },
1161 { .name = "unlinkat",
1162 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1163 { .name = "utimensat",
1164 .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1165 { .name = "wait4", .errpid = true,
1166 .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1167 { .name = "waitid", .errpid = true,
1168 .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1171 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1173 const struct syscall_fmt *fmt = fmtp;
1174 return strcmp(name, fmt->name);
1177 static const struct syscall_fmt *__syscall_fmt__find(const struct syscall_fmt *fmts,
1181 return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1184 static const struct syscall_fmt *syscall_fmt__find(const char *name)
1186 const int nmemb = ARRAY_SIZE(syscall_fmts);
1187 return __syscall_fmt__find(syscall_fmts, nmemb, name);
1190 static const struct syscall_fmt *__syscall_fmt__find_by_alias(const struct syscall_fmt *fmts,
1191 const int nmemb, const char *alias)
1195 for (i = 0; i < nmemb; ++i) {
1196 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1203 static const struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1205 const int nmemb = ARRAY_SIZE(syscall_fmts);
1206 return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1210 * is_exit: is this "exit" or "exit_group"?
1211 * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1212 * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1213 * nonexistent: Just a hole in the syscall table, syscall id not allocated
1216 struct tep_event *tp_format;
1220 struct bpf_program *sys_enter,
1226 struct tep_format_field *args;
1228 const struct syscall_fmt *fmt;
1229 struct syscall_arg_fmt *arg_fmt;
1233 * We need to have this 'calculated' boolean because in some cases we really
1234 * don't know what is the duration of a syscall, for instance, when we start
1235 * a session and some threads are waiting for a syscall to finish, say 'poll',
1236 * in which case all we can do is to print "( ? ) for duration and for the
1239 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1241 double duration = (double)t / NSEC_PER_MSEC;
1242 size_t printed = fprintf(fp, "(");
1245 printed += fprintf(fp, " ");
1246 else if (duration >= 1.0)
1247 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1248 else if (duration >= 0.01)
1249 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1251 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1252 return printed + fprintf(fp, "): ");
1256 * filename.ptr: The filename char pointer that will be vfs_getname'd
1257 * filename.entry_str_pos: Where to insert the string translated from
1258 * filename.ptr by the vfs_getname tracepoint/kprobe.
1259 * ret_scnprintf: syscall args may set this to a different syscall return
1260 * formatter, for instance, fcntl may return fds, file flags, etc.
1262 struct thread_trace {
1265 unsigned long nr_events;
1266 unsigned long pfmaj, pfmin;
1269 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1272 short int entry_str_pos;
1274 unsigned int namelen;
1282 struct intlist *syscall_stats;
1285 static struct thread_trace *thread_trace__new(void)
1287 struct thread_trace *ttrace = zalloc(sizeof(struct thread_trace));
1290 ttrace->files.max = -1;
1291 ttrace->syscall_stats = intlist__new(NULL);
1297 static void thread_trace__free_files(struct thread_trace *ttrace);
1299 static void thread_trace__delete(void *pttrace)
1301 struct thread_trace *ttrace = pttrace;
1306 intlist__delete(ttrace->syscall_stats);
1307 ttrace->syscall_stats = NULL;
1308 thread_trace__free_files(ttrace);
1309 zfree(&ttrace->entry_str);
1313 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1315 struct thread_trace *ttrace;
1320 if (thread__priv(thread) == NULL)
1321 thread__set_priv(thread, thread_trace__new());
1323 if (thread__priv(thread) == NULL)
1326 ttrace = thread__priv(thread);
1327 ++ttrace->nr_events;
1331 color_fprintf(fp, PERF_COLOR_RED,
1332 "WARNING: not enough memory, dropping samples!\n");
1337 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1338 size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1340 struct thread_trace *ttrace = thread__priv(arg->thread);
1342 ttrace->ret_scnprintf = ret_scnprintf;
1345 #define TRACE_PFMAJ (1 << 0)
1346 #define TRACE_PFMIN (1 << 1)
1348 static const size_t trace__entry_str_size = 2048;
1350 static void thread_trace__free_files(struct thread_trace *ttrace)
1352 for (int i = 0; i < ttrace->files.max; ++i) {
1353 struct file *file = ttrace->files.table + i;
1354 zfree(&file->pathname);
1357 zfree(&ttrace->files.table);
1358 ttrace->files.max = -1;
1361 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1366 if (fd > ttrace->files.max) {
1367 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1372 if (ttrace->files.max != -1) {
1373 memset(nfiles + ttrace->files.max + 1, 0,
1374 (fd - ttrace->files.max) * sizeof(struct file));
1376 memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1379 ttrace->files.table = nfiles;
1380 ttrace->files.max = fd;
1383 return ttrace->files.table + fd;
1386 struct file *thread__files_entry(struct thread *thread, int fd)
1388 return thread_trace__files_entry(thread__priv(thread), fd);
1391 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1393 struct thread_trace *ttrace = thread__priv(thread);
1394 struct file *file = thread_trace__files_entry(ttrace, fd);
1398 if (stat(pathname, &st) == 0)
1399 file->dev_maj = major(st.st_rdev);
1400 file->pathname = strdup(pathname);
1408 static int thread__read_fd_path(struct thread *thread, int fd)
1410 char linkname[PATH_MAX], pathname[PATH_MAX];
1414 if (thread__pid(thread) == thread__tid(thread)) {
1415 scnprintf(linkname, sizeof(linkname),
1416 "/proc/%d/fd/%d", thread__pid(thread), fd);
1418 scnprintf(linkname, sizeof(linkname),
1419 "/proc/%d/task/%d/fd/%d",
1420 thread__pid(thread), thread__tid(thread), fd);
1423 if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1426 ret = readlink(linkname, pathname, sizeof(pathname));
1428 if (ret < 0 || ret > st.st_size)
1431 pathname[ret] = '\0';
1432 return trace__set_fd_pathname(thread, fd, pathname);
1435 static const char *thread__fd_path(struct thread *thread, int fd,
1436 struct trace *trace)
1438 struct thread_trace *ttrace = thread__priv(thread);
1440 if (ttrace == NULL || trace->fd_path_disabled)
1446 if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1449 ++trace->stats.proc_getname;
1450 if (thread__read_fd_path(thread, fd))
1454 return ttrace->files.table[fd].pathname;
1457 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1460 size_t printed = scnprintf(bf, size, "%d", fd);
1461 const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1464 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1469 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1471 size_t printed = scnprintf(bf, size, "%d", fd);
1472 struct thread *thread = machine__find_thread(trace->host, pid, pid);
1475 const char *path = thread__fd_path(thread, fd, trace);
1478 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1480 thread__put(thread);
1486 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1487 struct syscall_arg *arg)
1490 size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1491 struct thread_trace *ttrace = thread__priv(arg->thread);
1493 if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1494 zfree(&ttrace->files.table[fd].pathname);
1499 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1502 struct thread_trace *ttrace = thread__priv(thread);
1504 ttrace->filename.ptr = ptr;
1505 ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1508 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1510 struct augmented_arg *augmented_arg = arg->augmented.args;
1511 size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1513 * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1514 * we would have two strings, each prefixed by its size.
1516 int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1518 arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1519 arg->augmented.size -= consumed;
1524 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1525 struct syscall_arg *arg)
1527 unsigned long ptr = arg->val;
1529 if (arg->augmented.args)
1530 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1532 if (!arg->trace->vfs_getname)
1533 return scnprintf(bf, size, "%#x", ptr);
1535 thread__set_filename_pos(arg->thread, bf, ptr);
1539 static bool trace__filter_duration(struct trace *trace, double t)
1541 return t < (trace->duration_filter * NSEC_PER_MSEC);
1544 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1546 double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1548 return fprintf(fp, "%10.3f ", ts);
1552 * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1553 * using ttrace->entry_time for a thread that receives a sys_exit without
1554 * first having received a sys_enter ("poll" issued before tracing session
1555 * starts, lost sys_enter exit due to ring buffer overflow).
1557 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1560 return __trace__fprintf_tstamp(trace, tstamp, fp);
1562 return fprintf(fp, " ? ");
1565 static pid_t workload_pid = -1;
1566 static volatile sig_atomic_t done = false;
1567 static volatile sig_atomic_t interrupted = false;
1569 static void sighandler_interrupt(int sig __maybe_unused)
1571 done = interrupted = true;
1574 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1575 void *context __maybe_unused)
1577 if (info->si_pid == workload_pid)
1581 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1585 if (trace->multiple_threads) {
1586 if (trace->show_comm)
1587 printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1588 printed += fprintf(fp, "%d ", thread__tid(thread));
1594 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1595 u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1599 if (trace->show_tstamp)
1600 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1601 if (trace->show_duration)
1602 printed += fprintf_duration(duration, duration_calculated, fp);
1603 return printed + trace__fprintf_comm_tid(trace, thread, fp);
1606 static int trace__process_event(struct trace *trace, struct machine *machine,
1607 union perf_event *event, struct perf_sample *sample)
1611 switch (event->header.type) {
1612 case PERF_RECORD_LOST:
1613 color_fprintf(trace->output, PERF_COLOR_RED,
1614 "LOST %" PRIu64 " events!\n", event->lost.lost);
1615 ret = machine__process_lost_event(machine, event, sample);
1618 ret = machine__process_event(machine, event, sample);
1625 static int trace__tool_process(struct perf_tool *tool,
1626 union perf_event *event,
1627 struct perf_sample *sample,
1628 struct machine *machine)
1630 struct trace *trace = container_of(tool, struct trace, tool);
1631 return trace__process_event(trace, machine, event, sample);
1634 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1636 struct machine *machine = vmachine;
1638 if (machine->kptr_restrict_warned)
1641 if (symbol_conf.kptr_restrict) {
1642 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1643 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1644 "Kernel samples will not be resolved.\n");
1645 machine->kptr_restrict_warned = true;
1649 return machine__resolve_kernel_addr(vmachine, addrp, modp);
1652 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1654 int err = symbol__init(NULL);
1659 trace->host = machine__new_host();
1660 if (trace->host == NULL)
1663 thread__set_priv_destructor(thread_trace__delete);
1665 err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1669 err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1670 evlist->core.threads, trace__tool_process,
1679 static void trace__symbols__exit(struct trace *trace)
1681 machine__exit(trace->host);
1687 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1691 if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1692 nr_args = sc->fmt->nr_args;
1694 sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1695 if (sc->arg_fmt == NULL)
1698 for (idx = 0; idx < nr_args; ++idx) {
1700 sc->arg_fmt[idx] = sc->fmt->arg[idx];
1703 sc->nr_args = nr_args;
1707 static const struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1708 { .name = "msr", .scnprintf = SCA_X86_MSR, .strtoul = STUL_X86_MSR, },
1709 { .name = "vector", .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1712 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1714 const struct syscall_arg_fmt *fmt = fmtp;
1715 return strcmp(name, fmt->name);
1718 static const struct syscall_arg_fmt *
1719 __syscall_arg_fmt__find_by_name(const struct syscall_arg_fmt *fmts, const int nmemb,
1722 return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1725 static const struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1727 const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1728 return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1731 static struct tep_format_field *
1732 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1734 struct tep_format_field *last_field = NULL;
1737 for (; field; field = field->next, ++arg) {
1743 len = strlen(field->name);
1745 if (strcmp(field->type, "const char *") == 0 &&
1746 ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1747 strstr(field->name, "path") != NULL))
1748 arg->scnprintf = SCA_FILENAME;
1749 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1750 arg->scnprintf = SCA_PTR;
1751 else if (strcmp(field->type, "pid_t") == 0)
1752 arg->scnprintf = SCA_PID;
1753 else if (strcmp(field->type, "umode_t") == 0)
1754 arg->scnprintf = SCA_MODE_T;
1755 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1756 arg->scnprintf = SCA_CHAR_ARRAY;
1757 arg->nr_entries = field->arraylen;
1758 } else if ((strcmp(field->type, "int") == 0 ||
1759 strcmp(field->type, "unsigned int") == 0 ||
1760 strcmp(field->type, "long") == 0) &&
1761 len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1763 * /sys/kernel/tracing/events/syscalls/sys_enter*
1764 * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1769 arg->scnprintf = SCA_FD;
1771 const struct syscall_arg_fmt *fmt =
1772 syscall_arg_fmt__find_by_name(field->name);
1775 arg->scnprintf = fmt->scnprintf;
1776 arg->strtoul = fmt->strtoul;
1784 static int syscall__set_arg_fmts(struct syscall *sc)
1786 struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1789 sc->args_size = last_field->offset + last_field->size;
1794 static int trace__read_syscall_info(struct trace *trace, int id)
1798 const char *name = syscalltbl__name(trace->sctbl, id);
1800 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1801 if (trace->syscalls.table == NULL) {
1802 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1803 if (trace->syscalls.table == NULL)
1807 if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1808 // When using libaudit we don't know beforehand what is the max syscall id
1809 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1814 // Need to memset from offset 0 and +1 members if brand new
1815 if (trace->syscalls.table == NULL)
1816 memset(table, 0, (id + 1) * sizeof(*sc));
1818 memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1820 trace->syscalls.table = table;
1821 trace->sctbl->syscalls.max_id = id;
1824 sc = trace->syscalls.table + id;
1825 if (sc->nonexistent)
1829 sc->nonexistent = true;
1834 sc->fmt = syscall_fmt__find(sc->name);
1836 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1837 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1839 if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1840 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1841 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1845 * Fails to read trace point format via sysfs node, so the trace point
1846 * doesn't exist. Set the 'nonexistent' flag as true.
1848 if (IS_ERR(sc->tp_format)) {
1849 sc->nonexistent = true;
1850 return PTR_ERR(sc->tp_format);
1853 if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1854 RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1857 sc->args = sc->tp_format->format.fields;
1859 * We need to check and discard the first variable '__syscall_nr'
1860 * or 'nr' that mean the syscall number. It is needless here.
1861 * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1863 if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1864 sc->args = sc->args->next;
1868 sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1869 sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1871 return syscall__set_arg_fmts(sc);
1874 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1876 struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1879 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1886 static int intcmp(const void *a, const void *b)
1888 const int *one = a, *another = b;
1890 return *one - *another;
1893 static int trace__validate_ev_qualifier(struct trace *trace)
1896 bool printed_invalid_prefix = false;
1897 struct str_node *pos;
1898 size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1900 trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1901 sizeof(trace->ev_qualifier_ids.entries[0]));
1903 if (trace->ev_qualifier_ids.entries == NULL) {
1904 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1910 strlist__for_each_entry(pos, trace->ev_qualifier) {
1911 const char *sc = pos->s;
1912 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1915 id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1919 if (!printed_invalid_prefix) {
1920 pr_debug("Skipping unknown syscalls: ");
1921 printed_invalid_prefix = true;
1930 trace->ev_qualifier_ids.entries[nr_used++] = id;
1931 if (match_next == -1)
1935 id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1938 if (nr_allocated == nr_used) {
1942 entries = realloc(trace->ev_qualifier_ids.entries,
1943 nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1944 if (entries == NULL) {
1946 fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1949 trace->ev_qualifier_ids.entries = entries;
1951 trace->ev_qualifier_ids.entries[nr_used++] = id;
1955 trace->ev_qualifier_ids.nr = nr_used;
1956 qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1958 if (printed_invalid_prefix)
1962 zfree(&trace->ev_qualifier_ids.entries);
1963 trace->ev_qualifier_ids.nr = 0;
1967 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1969 bool in_ev_qualifier;
1971 if (trace->ev_qualifier_ids.nr == 0)
1974 in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1975 trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1977 if (in_ev_qualifier)
1978 return !trace->not_ev_qualifier;
1980 return trace->not_ev_qualifier;
1984 * args is to be interpreted as a series of longs but we need to handle
1985 * 8-byte unaligned accesses. args points to raw_data within the event
1986 * and raw_data is guaranteed to be 8-byte unaligned because it is
1987 * preceded by raw_size which is a u32. So we need to copy args to a temp
1988 * variable to read it. Most notably this avoids extended load instructions
1989 * on unaligned addresses
1991 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1994 unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1996 memcpy(&val, p, sizeof(val));
2000 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
2001 struct syscall_arg *arg)
2003 if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
2004 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
2006 return scnprintf(bf, size, "arg%d: ", arg->idx);
2010 * Check if the value is in fact zero, i.e. mask whatever needs masking, such
2011 * as mount 'flags' argument that needs ignoring some magic flag, see comment
2012 * in tools/perf/trace/beauty/mount_flags.c
2014 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
2016 if (fmt && fmt->mask_val)
2017 return fmt->mask_val(arg, val);
2022 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
2023 struct syscall_arg *arg, unsigned long val)
2025 if (fmt && fmt->scnprintf) {
2028 arg->parm = fmt->parm;
2029 return fmt->scnprintf(bf, size, arg);
2031 return scnprintf(bf, size, "%ld", val);
2034 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2035 unsigned char *args, void *augmented_args, int augmented_args_size,
2036 struct trace *trace, struct thread *thread)
2041 struct syscall_arg arg = {
2044 .size = augmented_args_size,
2045 .args = augmented_args,
2051 .show_string_prefix = trace->show_string_prefix,
2053 struct thread_trace *ttrace = thread__priv(thread);
2056 * Things like fcntl will set this in its 'cmd' formatter to pick the
2057 * right formatter for the return value (an fd? file flags?), which is
2058 * not needed for syscalls that always return a given type, say an fd.
2060 ttrace->ret_scnprintf = NULL;
2062 if (sc->args != NULL) {
2063 struct tep_format_field *field;
2065 for (field = sc->args; field;
2066 field = field->next, ++arg.idx, bit <<= 1) {
2070 arg.fmt = &sc->arg_fmt[arg.idx];
2071 val = syscall_arg__val(&arg, arg.idx);
2073 * Some syscall args need some mask, most don't and
2074 * return val untouched.
2076 val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2079 * Suppress this argument if its value is zero and
2080 * and we don't have a string associated in an
2084 !trace->show_zeros &&
2086 (sc->arg_fmt[arg.idx].show_zero ||
2087 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2088 sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2089 sc->arg_fmt[arg.idx].parm))
2092 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2094 if (trace->show_arg_names)
2095 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2097 printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2098 bf + printed, size - printed, &arg, val);
2100 } else if (IS_ERR(sc->tp_format)) {
2102 * If we managed to read the tracepoint /format file, then we
2103 * may end up not having any args, like with gettid(), so only
2104 * print the raw args when we didn't manage to read it.
2106 while (arg.idx < sc->nr_args) {
2109 val = syscall_arg__val(&arg, arg.idx);
2111 printed += scnprintf(bf + printed, size - printed, ", ");
2112 printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2113 printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2123 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2124 union perf_event *event,
2125 struct perf_sample *sample);
2127 static struct syscall *trace__syscall_info(struct trace *trace,
2128 struct evsel *evsel, int id)
2135 * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2136 * before that, leaving at a higher verbosity level till that is
2137 * explained. Reproduced with plain ftrace with:
2139 * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2140 * grep "NR -1 " /t/trace_pipe
2142 * After generating some load on the machine.
2146 fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2147 id, evsel__name(evsel), ++n);
2154 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2155 if (id > trace->sctbl->syscalls.max_id) {
2157 if (id >= trace->sctbl->syscalls.max_id) {
2159 * With libaudit we don't know beforehand what is the max_id,
2160 * so we let trace__read_syscall_info() figure that out as we
2161 * go on reading syscalls.
2163 err = trace__read_syscall_info(trace, id);
2169 if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2170 (err = trace__read_syscall_info(trace, id)) != 0)
2173 if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2176 return &trace->syscalls.table[id];
2180 char sbuf[STRERR_BUFSIZE];
2181 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2182 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2183 fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2184 fputs(" information\n", trace->output);
2189 struct syscall_stats {
2196 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2197 int id, struct perf_sample *sample, long err, bool errno_summary)
2199 struct int_node *inode;
2200 struct syscall_stats *stats;
2203 inode = intlist__findnew(ttrace->syscall_stats, id);
2207 stats = inode->priv;
2208 if (stats == NULL) {
2209 stats = zalloc(sizeof(*stats));
2213 init_stats(&stats->stats);
2214 inode->priv = stats;
2217 if (ttrace->entry_time && sample->time > ttrace->entry_time)
2218 duration = sample->time - ttrace->entry_time;
2220 update_stats(&stats->stats, duration);
2223 ++stats->nr_failures;
2229 if (err > stats->max_errno) {
2230 u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2233 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2235 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2236 thread__comm_str(thread), thread__pid(thread),
2237 thread__tid(thread));
2241 stats->errnos = new_errnos;
2242 stats->max_errno = err;
2245 ++stats->errnos[err - 1];
2249 static int trace__printf_interrupted_entry(struct trace *trace)
2251 struct thread_trace *ttrace;
2255 if (trace->failure_only || trace->current == NULL)
2258 ttrace = thread__priv(trace->current);
2260 if (!ttrace->entry_pending)
2263 printed = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2264 printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2266 if (len < trace->args_alignment - 4)
2267 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2269 printed += fprintf(trace->output, " ...\n");
2271 ttrace->entry_pending = false;
2272 ++trace->nr_events_printed;
2277 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2278 struct perf_sample *sample, struct thread *thread)
2282 if (trace->print_sample) {
2283 double ts = (double)sample->time / NSEC_PER_MSEC;
2285 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2286 evsel__name(evsel), ts,
2287 thread__comm_str(thread),
2288 sample->pid, sample->tid, sample->cpu);
2294 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2296 void *augmented_args = NULL;
2298 * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2299 * and there we get all 6 syscall args plus the tracepoint common fields
2300 * that gets calculated at the start and the syscall_nr (another long).
2301 * So we check if that is the case and if so don't look after the
2302 * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2305 * We'll revisit this later to pass s->args_size to the BPF augmenter
2306 * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2307 * copies only what we need for each syscall, like what happens when we
2308 * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2309 * traffic to just what is needed for each syscall.
2311 int args_size = raw_augmented_args_size ?: sc->args_size;
2313 *augmented_args_size = sample->raw_size - args_size;
2314 if (*augmented_args_size > 0)
2315 augmented_args = sample->raw_data + args_size;
2317 return augmented_args;
2320 static void syscall__exit(struct syscall *sc)
2325 zfree(&sc->arg_fmt);
2328 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2329 union perf_event *event __maybe_unused,
2330 struct perf_sample *sample)
2335 struct thread *thread;
2336 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2337 int augmented_args_size = 0;
2338 void *augmented_args = NULL;
2339 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2340 struct thread_trace *ttrace;
2345 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2346 ttrace = thread__trace(thread, trace->output);
2350 trace__fprintf_sample(trace, evsel, sample, thread);
2352 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2354 if (ttrace->entry_str == NULL) {
2355 ttrace->entry_str = malloc(trace__entry_str_size);
2356 if (!ttrace->entry_str)
2360 if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2361 trace__printf_interrupted_entry(trace);
2363 * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2364 * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2365 * this breaks syscall__augmented_args() check for augmented args, as we calculate
2366 * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2367 * so when handling, say the openat syscall, we end up getting 6 args for the
2368 * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2369 * thinking that the extra 2 u64 args are the augmented filename, so just check
2370 * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2372 if (evsel != trace->syscalls.events.sys_enter)
2373 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2374 ttrace->entry_time = sample->time;
2375 msg = ttrace->entry_str;
2376 printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2378 printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2379 args, augmented_args, augmented_args_size, trace, thread);
2382 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2385 trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2386 printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2387 if (trace->args_alignment > printed)
2388 alignment = trace->args_alignment - printed;
2389 fprintf(trace->output, "%*s= ?\n", alignment, " ");
2392 ttrace->entry_pending = true;
2393 /* See trace__vfs_getname & trace__sys_exit */
2394 ttrace->filename.pending_open = false;
2397 if (trace->current != thread) {
2398 thread__put(trace->current);
2399 trace->current = thread__get(thread);
2403 thread__put(thread);
2407 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2408 struct perf_sample *sample)
2410 struct thread_trace *ttrace;
2411 struct thread *thread;
2412 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2413 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2415 void *args, *augmented_args = NULL;
2416 int augmented_args_size;
2421 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2422 ttrace = thread__trace(thread, trace->output);
2424 * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2425 * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2430 args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2431 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2432 syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2433 fprintf(trace->output, "%s", msg);
2436 thread__put(thread);
2440 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2441 struct perf_sample *sample,
2442 struct callchain_cursor *cursor)
2444 struct addr_location al;
2445 int max_stack = evsel->core.attr.sample_max_stack ?
2446 evsel->core.attr.sample_max_stack :
2450 addr_location__init(&al);
2451 if (machine__resolve(trace->host, &al, sample) < 0)
2454 err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2456 addr_location__exit(&al);
2460 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2462 /* TODO: user-configurable print_opts */
2463 const unsigned int print_opts = EVSEL__PRINT_SYM |
2465 EVSEL__PRINT_UNKNOWN_AS_ADDR;
2467 return sample__fprintf_callchain(sample, 38, print_opts, get_tls_callchain_cursor(), symbol_conf.bt_stop_list, trace->output);
2470 static const char *errno_to_name(struct evsel *evsel, int err)
2472 struct perf_env *env = evsel__env(evsel);
2474 return perf_env__arch_strerrno(env, err);
2477 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2478 union perf_event *event __maybe_unused,
2479 struct perf_sample *sample)
2483 bool duration_calculated = false;
2484 struct thread *thread;
2485 int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2486 int alignment = trace->args_alignment;
2487 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2488 struct thread_trace *ttrace;
2493 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2494 ttrace = thread__trace(thread, trace->output);
2498 trace__fprintf_sample(trace, evsel, sample, thread);
2500 ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2503 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2505 if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2506 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2507 ttrace->filename.pending_open = false;
2508 ++trace->stats.vfs_getname;
2511 if (ttrace->entry_time) {
2512 duration = sample->time - ttrace->entry_time;
2513 if (trace__filter_duration(trace, duration))
2515 duration_calculated = true;
2516 } else if (trace->duration_filter)
2519 if (sample->callchain) {
2520 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2522 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2523 if (callchain_ret == 0) {
2524 if (cursor->nr < trace->min_stack)
2530 if (trace->summary_only || (ret >= 0 && trace->failure_only))
2533 trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2535 if (ttrace->entry_pending) {
2536 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2538 printed += fprintf(trace->output, " ... [");
2539 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2541 printed += fprintf(trace->output, "]: %s()", sc->name);
2544 printed++; /* the closing ')' */
2546 if (alignment > printed)
2547 alignment -= printed;
2551 fprintf(trace->output, ")%*s= ", alignment, " ");
2553 if (sc->fmt == NULL) {
2557 fprintf(trace->output, "%ld", ret);
2558 } else if (ret < 0) {
2560 char bf[STRERR_BUFSIZE];
2561 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2562 *e = errno_to_name(evsel, -ret);
2564 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2566 } else if (ret == 0 && sc->fmt->timeout)
2567 fprintf(trace->output, "0 (Timeout)");
2568 else if (ttrace->ret_scnprintf) {
2570 struct syscall_arg arg = {
2575 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2576 ttrace->ret_scnprintf = NULL;
2577 fprintf(trace->output, "%s", bf);
2578 } else if (sc->fmt->hexret)
2579 fprintf(trace->output, "%#lx", ret);
2580 else if (sc->fmt->errpid) {
2581 struct thread *child = machine__find_thread(trace->host, ret, ret);
2583 if (child != NULL) {
2584 fprintf(trace->output, "%ld", ret);
2585 if (thread__comm_set(child))
2586 fprintf(trace->output, " (%s)", thread__comm_str(child));
2592 fputc('\n', trace->output);
2595 * We only consider an 'event' for the sake of --max-events a non-filtered
2596 * sys_enter + sys_exit and other tracepoint events.
2598 if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2601 if (callchain_ret > 0)
2602 trace__fprintf_callchain(trace, sample);
2603 else if (callchain_ret < 0)
2604 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2606 ttrace->entry_pending = false;
2609 thread__put(thread);
2613 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2614 union perf_event *event __maybe_unused,
2615 struct perf_sample *sample)
2617 struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2618 struct thread_trace *ttrace;
2619 size_t filename_len, entry_str_len, to_move;
2620 ssize_t remaining_space;
2622 const char *filename = evsel__rawptr(evsel, sample, "pathname");
2627 ttrace = thread__priv(thread);
2631 filename_len = strlen(filename);
2632 if (filename_len == 0)
2635 if (ttrace->filename.namelen < filename_len) {
2636 char *f = realloc(ttrace->filename.name, filename_len + 1);
2641 ttrace->filename.namelen = filename_len;
2642 ttrace->filename.name = f;
2645 strcpy(ttrace->filename.name, filename);
2646 ttrace->filename.pending_open = true;
2648 if (!ttrace->filename.ptr)
2651 entry_str_len = strlen(ttrace->entry_str);
2652 remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2653 if (remaining_space <= 0)
2656 if (filename_len > (size_t)remaining_space) {
2657 filename += filename_len - remaining_space;
2658 filename_len = remaining_space;
2661 to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2662 pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2663 memmove(pos + filename_len, pos, to_move);
2664 memcpy(pos, filename, filename_len);
2666 ttrace->filename.ptr = 0;
2667 ttrace->filename.entry_str_pos = 0;
2669 thread__put(thread);
2674 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2675 union perf_event *event __maybe_unused,
2676 struct perf_sample *sample)
2678 u64 runtime = evsel__intval(evsel, sample, "runtime");
2679 double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2680 struct thread *thread = machine__findnew_thread(trace->host,
2683 struct thread_trace *ttrace = thread__trace(thread, trace->output);
2688 ttrace->runtime_ms += runtime_ms;
2689 trace->runtime_ms += runtime_ms;
2691 thread__put(thread);
2695 fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2697 evsel__strval(evsel, sample, "comm"),
2698 (pid_t)evsel__intval(evsel, sample, "pid"),
2700 evsel__intval(evsel, sample, "vruntime"));
2704 static int bpf_output__printer(enum binary_printer_ops op,
2705 unsigned int val, void *extra __maybe_unused, FILE *fp)
2707 unsigned char ch = (unsigned char)val;
2710 case BINARY_PRINT_CHAR_DATA:
2711 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2712 case BINARY_PRINT_DATA_BEGIN:
2713 case BINARY_PRINT_LINE_BEGIN:
2714 case BINARY_PRINT_ADDR:
2715 case BINARY_PRINT_NUM_DATA:
2716 case BINARY_PRINT_NUM_PAD:
2717 case BINARY_PRINT_SEP:
2718 case BINARY_PRINT_CHAR_PAD:
2719 case BINARY_PRINT_LINE_END:
2720 case BINARY_PRINT_DATA_END:
2728 static void bpf_output__fprintf(struct trace *trace,
2729 struct perf_sample *sample)
2731 binary__fprintf(sample->raw_data, sample->raw_size, 8,
2732 bpf_output__printer, NULL, trace->output);
2733 ++trace->nr_events_printed;
2736 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2737 struct thread *thread, void *augmented_args, int augmented_args_size)
2740 size_t size = sizeof(bf);
2741 struct tep_format_field *field = evsel->tp_format->format.fields;
2742 struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2746 struct syscall_arg syscall_arg = {
2748 .size = augmented_args_size,
2749 .args = augmented_args,
2755 .show_string_prefix = trace->show_string_prefix,
2758 for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2759 if (syscall_arg.mask & bit)
2762 syscall_arg.len = 0;
2763 syscall_arg.fmt = arg;
2764 if (field->flags & TEP_FIELD_IS_ARRAY) {
2765 int offset = field->offset;
2767 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2768 offset = format_field__intval(field, sample, evsel->needs_swap);
2769 syscall_arg.len = offset >> 16;
2771 if (tep_field_is_relative(field->flags))
2772 offset += field->offset + field->size;
2775 val = (uintptr_t)(sample->raw_data + offset);
2777 val = format_field__intval(field, sample, evsel->needs_swap);
2779 * Some syscall args need some mask, most don't and
2780 * return val untouched.
2782 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2785 * Suppress this argument if its value is zero and
2786 * we don't have a string associated in an
2790 !trace->show_zeros &&
2791 !((arg->show_zero ||
2792 arg->scnprintf == SCA_STRARRAY ||
2793 arg->scnprintf == SCA_STRARRAYS) &&
2797 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2799 if (trace->show_arg_names)
2800 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2802 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2805 return printed + fprintf(trace->output, "%s", bf);
2808 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2809 union perf_event *event __maybe_unused,
2810 struct perf_sample *sample)
2812 struct thread *thread;
2813 int callchain_ret = 0;
2815 * Check if we called perf_evsel__disable(evsel) due to, for instance,
2816 * this event's max_events having been hit and this is an entry coming
2817 * from the ring buffer that we should discard, since the max events
2818 * have already been considered/printed.
2820 if (evsel->disabled)
2823 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2825 if (sample->callchain) {
2826 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2828 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2829 if (callchain_ret == 0) {
2830 if (cursor->nr < trace->min_stack)
2836 trace__printf_interrupted_entry(trace);
2837 trace__fprintf_tstamp(trace, sample->time, trace->output);
2839 if (trace->trace_syscalls && trace->show_duration)
2840 fprintf(trace->output, "( ): ");
2843 trace__fprintf_comm_tid(trace, thread, trace->output);
2845 if (evsel == trace->syscalls.events.bpf_output) {
2846 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2847 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2850 fprintf(trace->output, "%s(", sc->name);
2851 trace__fprintf_sys_enter(trace, evsel, sample);
2852 fputc(')', trace->output);
2857 * XXX: Not having the associated syscall info or not finding/adding
2858 * the thread should never happen, but if it does...
2859 * fall thru and print it as a bpf_output event.
2863 fprintf(trace->output, "%s(", evsel->name);
2865 if (evsel__is_bpf_output(evsel)) {
2866 bpf_output__fprintf(trace, sample);
2867 } else if (evsel->tp_format) {
2868 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2869 trace__fprintf_sys_enter(trace, evsel, sample)) {
2870 if (trace->libtraceevent_print) {
2871 event_format__fprintf(evsel->tp_format, sample->cpu,
2872 sample->raw_data, sample->raw_size,
2875 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2881 fprintf(trace->output, ")\n");
2883 if (callchain_ret > 0)
2884 trace__fprintf_callchain(trace, sample);
2885 else if (callchain_ret < 0)
2886 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2888 ++trace->nr_events_printed;
2890 if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2891 evsel__disable(evsel);
2892 evsel__close(evsel);
2895 thread__put(thread);
2899 static void print_location(FILE *f, struct perf_sample *sample,
2900 struct addr_location *al,
2901 bool print_dso, bool print_sym)
2904 if ((verbose > 0 || print_dso) && al->map)
2905 fprintf(f, "%s@", map__dso(al->map)->long_name);
2907 if ((verbose > 0 || print_sym) && al->sym)
2908 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2909 al->addr - al->sym->start);
2911 fprintf(f, "0x%" PRIx64, al->addr);
2913 fprintf(f, "0x%" PRIx64, sample->addr);
2916 static int trace__pgfault(struct trace *trace,
2917 struct evsel *evsel,
2918 union perf_event *event __maybe_unused,
2919 struct perf_sample *sample)
2921 struct thread *thread;
2922 struct addr_location al;
2923 char map_type = 'd';
2924 struct thread_trace *ttrace;
2926 int callchain_ret = 0;
2928 addr_location__init(&al);
2929 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2931 if (sample->callchain) {
2932 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2934 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2935 if (callchain_ret == 0) {
2936 if (cursor->nr < trace->min_stack)
2942 ttrace = thread__trace(thread, trace->output);
2946 if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2951 if (trace->summary_only)
2954 thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2956 trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2958 fprintf(trace->output, "%sfault [",
2959 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2962 print_location(trace->output, sample, &al, false, true);
2964 fprintf(trace->output, "] => ");
2966 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2969 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2977 print_location(trace->output, sample, &al, true, false);
2979 fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2981 if (callchain_ret > 0)
2982 trace__fprintf_callchain(trace, sample);
2983 else if (callchain_ret < 0)
2984 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2986 ++trace->nr_events_printed;
2990 thread__put(thread);
2991 addr_location__exit(&al);
2995 static void trace__set_base_time(struct trace *trace,
2996 struct evsel *evsel,
2997 struct perf_sample *sample)
3000 * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
3001 * and don't use sample->time unconditionally, we may end up having
3002 * some other event in the future without PERF_SAMPLE_TIME for good
3003 * reason, i.e. we may not be interested in its timestamps, just in
3004 * it taking place, picking some piece of information when it
3005 * appears in our event stream (vfs_getname comes to mind).
3007 if (trace->base_time == 0 && !trace->full_time &&
3008 (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
3009 trace->base_time = sample->time;
3012 static int trace__process_sample(struct perf_tool *tool,
3013 union perf_event *event,
3014 struct perf_sample *sample,
3015 struct evsel *evsel,
3016 struct machine *machine __maybe_unused)
3018 struct trace *trace = container_of(tool, struct trace, tool);
3019 struct thread *thread;
3022 tracepoint_handler handler = evsel->handler;
3024 thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
3025 if (thread && thread__is_filtered(thread))
3028 trace__set_base_time(trace, evsel, sample);
3032 handler(trace, evsel, event, sample);
3035 thread__put(thread);
3039 static int trace__record(struct trace *trace, int argc, const char **argv)
3041 unsigned int rec_argc, i, j;
3042 const char **rec_argv;
3043 const char * const record_args[] = {
3049 pid_t pid = getpid();
3050 char *filter = asprintf__tp_filter_pids(1, &pid);
3051 const char * const sc_args[] = { "-e", };
3052 unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3053 const char * const majpf_args[] = { "-e", "major-faults" };
3054 unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3055 const char * const minpf_args[] = { "-e", "minor-faults" };
3056 unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3059 /* +3 is for the event string below and the pid filter */
3060 rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3061 majpf_args_nr + minpf_args_nr + argc;
3062 rec_argv = calloc(rec_argc + 1, sizeof(char *));
3064 if (rec_argv == NULL || filter == NULL)
3068 for (i = 0; i < ARRAY_SIZE(record_args); i++)
3069 rec_argv[j++] = record_args[i];
3071 if (trace->trace_syscalls) {
3072 for (i = 0; i < sc_args_nr; i++)
3073 rec_argv[j++] = sc_args[i];
3075 /* event string may be different for older kernels - e.g., RHEL6 */
3076 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3077 rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3078 else if (is_valid_tracepoint("syscalls:sys_enter"))
3079 rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3081 pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3086 rec_argv[j++] = "--filter";
3087 rec_argv[j++] = filter;
3089 if (trace->trace_pgfaults & TRACE_PFMAJ)
3090 for (i = 0; i < majpf_args_nr; i++)
3091 rec_argv[j++] = majpf_args[i];
3093 if (trace->trace_pgfaults & TRACE_PFMIN)
3094 for (i = 0; i < minpf_args_nr; i++)
3095 rec_argv[j++] = minpf_args[i];
3097 for (i = 0; i < (unsigned int)argc; i++)
3098 rec_argv[j++] = argv[i];
3100 err = cmd_record(j, rec_argv);
3107 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3109 static bool evlist__add_vfs_getname(struct evlist *evlist)
3112 struct evsel *evsel, *tmp;
3113 struct parse_events_error err;
3116 parse_events_error__init(&err);
3117 ret = parse_events(evlist, "probe:vfs_getname*", &err);
3118 parse_events_error__exit(&err);
3122 evlist__for_each_entry_safe(evlist, evsel, tmp) {
3123 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3126 if (evsel__field(evsel, "pathname")) {
3127 evsel->handler = trace__vfs_getname;
3132 list_del_init(&evsel->core.node);
3133 evsel->evlist = NULL;
3134 evsel__delete(evsel);
3140 static struct evsel *evsel__new_pgfault(u64 config)
3142 struct evsel *evsel;
3143 struct perf_event_attr attr = {
3144 .type = PERF_TYPE_SOFTWARE,
3148 attr.config = config;
3149 attr.sample_period = 1;
3151 event_attr_init(&attr);
3153 evsel = evsel__new(&attr);
3155 evsel->handler = trace__pgfault;
3160 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3162 struct evsel *evsel;
3164 evlist__for_each_entry(evlist, evsel) {
3165 evsel_trace__delete(evsel->priv);
3170 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3172 const u32 type = event->header.type;
3173 struct evsel *evsel;
3175 if (type != PERF_RECORD_SAMPLE) {
3176 trace__process_event(trace, trace->host, event, sample);
3180 evsel = evlist__id2evsel(trace->evlist, sample->id);
3181 if (evsel == NULL) {
3182 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3186 if (evswitch__discard(&trace->evswitch, evsel))
3189 trace__set_base_time(trace, evsel, sample);
3191 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3192 sample->raw_data == NULL) {
3193 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3194 evsel__name(evsel), sample->tid,
3195 sample->cpu, sample->raw_size);
3197 tracepoint_handler handler = evsel->handler;
3198 handler(trace, evsel, event, sample);
3201 if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3205 static int trace__add_syscall_newtp(struct trace *trace)
3208 struct evlist *evlist = trace->evlist;
3209 struct evsel *sys_enter, *sys_exit;
3211 sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3212 if (sys_enter == NULL)
3215 if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3216 goto out_delete_sys_enter;
3218 sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3219 if (sys_exit == NULL)
3220 goto out_delete_sys_enter;
3222 if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3223 goto out_delete_sys_exit;
3225 evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3226 evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3228 evlist__add(evlist, sys_enter);
3229 evlist__add(evlist, sys_exit);
3231 if (callchain_param.enabled && !trace->kernel_syscallchains) {
3233 * We're interested only in the user space callchain
3234 * leading to the syscall, allow overriding that for
3235 * debugging reasons using --kernel_syscall_callchains
3237 sys_exit->core.attr.exclude_callchain_kernel = 1;
3240 trace->syscalls.events.sys_enter = sys_enter;
3241 trace->syscalls.events.sys_exit = sys_exit;
3247 out_delete_sys_exit:
3248 evsel__delete_priv(sys_exit);
3249 out_delete_sys_enter:
3250 evsel__delete_priv(sys_enter);
3254 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3257 struct evsel *sys_exit;
3258 char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3259 trace->ev_qualifier_ids.nr,
3260 trace->ev_qualifier_ids.entries);
3265 if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3266 sys_exit = trace->syscalls.events.sys_exit;
3267 err = evsel__append_tp_filter(sys_exit, filter);
3278 #ifdef HAVE_BPF_SKEL
3279 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3281 struct bpf_program *pos, *prog = NULL;
3282 const char *sec_name;
3284 if (trace->skel->obj == NULL)
3287 bpf_object__for_each_program(pos, trace->skel->obj) {
3288 sec_name = bpf_program__section_name(pos);
3289 if (sec_name && !strcmp(sec_name, name)) {
3298 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3299 const char *prog_name, const char *type)
3301 struct bpf_program *prog;
3303 if (prog_name == NULL) {
3304 char default_prog_name[256];
3305 scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->name);
3306 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3309 if (sc->fmt && sc->fmt->alias) {
3310 scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->fmt->alias);
3311 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3315 goto out_unaugmented;
3318 prog = trace__find_bpf_program_by_title(trace, prog_name);
3325 pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3326 prog_name, type, sc->name);
3328 return trace->skel->progs.syscall_unaugmented;
3331 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3333 struct syscall *sc = trace__syscall_info(trace, NULL, id);
3338 sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3339 sc->bpf_prog.sys_exit = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit : NULL, "exit");
3342 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3344 struct syscall *sc = trace__syscall_info(trace, NULL, id);
3345 return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3348 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3350 struct syscall *sc = trace__syscall_info(trace, NULL, id);
3351 return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3354 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3356 struct tep_format_field *field, *candidate_field;
3360 * We're only interested in syscalls that have a pointer:
3362 for (field = sc->args; field; field = field->next) {
3363 if (field->flags & TEP_FIELD_IS_POINTER)
3364 goto try_to_find_pair;
3370 for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3371 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3372 struct bpf_program *pair_prog;
3373 bool is_candidate = false;
3375 if (pair == NULL || pair == sc ||
3376 pair->bpf_prog.sys_enter == trace->skel->progs.syscall_unaugmented)
3379 for (field = sc->args, candidate_field = pair->args;
3380 field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3381 bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3382 candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3385 if (!candidate_is_pointer) {
3386 // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3390 if (candidate_is_pointer) {
3391 // The candidate might copy a pointer we don't have, skip it.
3392 goto next_candidate;
3397 if (strcmp(field->type, candidate_field->type))
3398 goto next_candidate;
3401 * This is limited in the BPF program but sys_write
3402 * uses "const char *" for its "buf" arg so we need to
3403 * use some heuristic that is kinda future proof...
3405 if (strcmp(field->type, "const char *") == 0 &&
3406 !(strstr(field->name, "name") ||
3407 strstr(field->name, "path") ||
3408 strstr(field->name, "file") ||
3409 strstr(field->name, "root") ||
3410 strstr(field->name, "description")))
3411 goto next_candidate;
3413 is_candidate = true;
3417 goto next_candidate;
3420 * Check if the tentative pair syscall augmenter has more pointers, if it has,
3421 * then it may be collecting that and we then can't use it, as it would collect
3422 * more than what is common to the two syscalls.
3424 if (candidate_field) {
3425 for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3426 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3427 goto next_candidate;
3430 pair_prog = pair->bpf_prog.sys_enter;
3432 * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3433 * have been searched for, so search it here and if it returns the
3434 * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3435 * program for a filtered syscall on a non-filtered one.
3437 * For instance, we have "!syscalls:sys_enter_renameat" and that is
3438 * useful for "renameat2".
3440 if (pair_prog == NULL) {
3441 pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3442 if (pair_prog == trace->skel->progs.syscall_unaugmented)
3443 goto next_candidate;
3446 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3455 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3457 int map_enter_fd = bpf_map__fd(trace->skel->maps.syscalls_sys_enter);
3458 int map_exit_fd = bpf_map__fd(trace->skel->maps.syscalls_sys_exit);
3461 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3464 if (!trace__syscall_enabled(trace, key))
3467 trace__init_syscall_bpf_progs(trace, key);
3469 // It'll get at least the "!raw_syscalls:unaugmented"
3470 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3471 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3474 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3475 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3481 * Now lets do a second pass looking for enabled syscalls without
3482 * an augmenter that have a signature that is a superset of another
3483 * syscall with an augmenter so that we can auto-reuse it.
3485 * I.e. if we have an augmenter for the "open" syscall that has
3488 * int open(const char *pathname, int flags, mode_t mode);
3490 * I.e. that will collect just the first string argument, then we
3491 * can reuse it for the 'creat' syscall, that has this signature:
3493 * int creat(const char *pathname, mode_t mode);
3497 * int stat(const char *pathname, struct stat *statbuf);
3498 * int lstat(const char *pathname, struct stat *statbuf);
3500 * Because the 'open' augmenter will collect the first arg as a string,
3501 * and leave alone all the other args, which already helps with
3502 * beautifying 'stat' and 'lstat''s pathname arg.
3504 * Then, in time, when 'stat' gets an augmenter that collects both
3505 * first and second arg (this one on the raw_syscalls:sys_exit prog
3506 * array tail call, then that one will be used.
3508 for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3509 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3510 struct bpf_program *pair_prog;
3513 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3517 * For now we're just reusing the sys_enter prog, and if it
3518 * already has an augmenter, we don't need to find one.
3520 if (sc->bpf_prog.sys_enter != trace->skel->progs.syscall_unaugmented)
3524 * Look at all the other syscalls for one that has a signature
3525 * that is close enough that we can share:
3527 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3528 if (pair_prog == NULL)
3531 sc->bpf_prog.sys_enter = pair_prog;
3534 * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3535 * with the fd for the program we're reusing:
3537 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3538 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3545 #endif // HAVE_BPF_SKEL
3547 static int trace__set_ev_qualifier_filter(struct trace *trace)
3549 if (trace->syscalls.events.sys_enter)
3550 return trace__set_ev_qualifier_tp_filter(trace);
3554 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3555 size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3558 #ifdef HAVE_LIBBPF_SUPPORT
3560 int map_fd = bpf_map__fd(map);
3563 for (i = 0; i < npids; ++i) {
3564 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3572 static int trace__set_filter_loop_pids(struct trace *trace)
3574 unsigned int nr = 1, err;
3578 struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3580 while (thread && nr < ARRAY_SIZE(pids)) {
3581 struct thread *parent = machine__find_thread(trace->host,
3582 thread__ppid(thread),
3583 thread__ppid(thread));
3588 if (!strcmp(thread__comm_str(parent), "sshd") ||
3589 strstarts(thread__comm_str(parent), "gnome-terminal")) {
3590 pids[nr++] = thread__tid(parent);
3596 err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3597 if (!err && trace->filter_pids.map)
3598 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3603 static int trace__set_filter_pids(struct trace *trace)
3607 * Better not use !target__has_task() here because we need to cover the
3608 * case where no threads were specified in the command line, but a
3609 * workload was, and in that case we will fill in the thread_map when
3610 * we fork the workload in evlist__prepare_workload.
3612 if (trace->filter_pids.nr > 0) {
3613 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3614 trace->filter_pids.entries);
3615 if (!err && trace->filter_pids.map) {
3616 err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3617 trace->filter_pids.entries);
3619 } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3620 err = trace__set_filter_loop_pids(trace);
3626 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3628 struct evlist *evlist = trace->evlist;
3629 struct perf_sample sample;
3630 int err = evlist__parse_sample(evlist, event, &sample);
3633 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3635 trace__handle_event(trace, event, &sample);
3640 static int __trace__flush_events(struct trace *trace)
3642 u64 first = ordered_events__first_time(&trace->oe.data);
3643 u64 flush = trace->oe.last - NSEC_PER_SEC;
3645 /* Is there some thing to flush.. */
3646 if (first && first < flush)
3647 return ordered_events__flush_time(&trace->oe.data, flush);
3652 static int trace__flush_events(struct trace *trace)
3654 return !trace->sort_events ? 0 : __trace__flush_events(trace);
3657 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3661 if (!trace->sort_events)
3662 return __trace__deliver_event(trace, event);
3664 err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3665 if (err && err != -1)
3668 err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3672 return trace__flush_events(trace);
3675 static int ordered_events__deliver_event(struct ordered_events *oe,
3676 struct ordered_event *event)
3678 struct trace *trace = container_of(oe, struct trace, oe.data);
3680 return __trace__deliver_event(trace, event->event);
3683 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3685 struct tep_format_field *field;
3686 struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3688 if (evsel->tp_format == NULL || fmt == NULL)
3691 for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3692 if (strcmp(field->name, arg) == 0)
3698 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3700 char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3702 while ((tok = strpbrk(left, "=<>!")) != NULL) {
3703 char *right = tok + 1, *right_end;
3708 while (isspace(*right))
3714 while (!isalpha(*left))
3715 if (++left == tok) {
3717 * Bail out, can't find the name of the argument that is being
3718 * used in the filter, let it try to set this filter, will fail later.
3723 right_end = right + 1;
3724 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3727 if (isalpha(*right)) {
3728 struct syscall_arg_fmt *fmt;
3729 int left_size = tok - left,
3730 right_size = right_end - right;
3733 while (isspace(left[left_size - 1]))
3736 scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3738 fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3740 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3741 arg, evsel->name, evsel->filter);
3745 pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3746 arg, (int)(right - tok), tok, right_size, right);
3750 struct syscall_arg syscall_arg = {
3754 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3755 char *n, expansion[19];
3756 int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3757 int expansion_offset = right - new_filter;
3759 pr_debug("%s", expansion);
3761 if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3762 pr_debug(" out of memory!\n");
3766 if (new_filter != evsel->filter)
3768 left = n + expansion_offset + expansion_lenght;
3771 pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3772 right_size, right, arg, evsel->name, evsel->filter);
3776 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3777 arg, evsel->name, evsel->filter);
3787 if (new_filter != evsel->filter) {
3788 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3789 evsel__set_filter(evsel, new_filter);
3796 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3798 struct evlist *evlist = trace->evlist;
3799 struct evsel *evsel;
3801 evlist__for_each_entry(evlist, evsel) {
3802 if (evsel->filter == NULL)
3805 if (trace__expand_filter(trace, evsel)) {
3814 static int trace__run(struct trace *trace, int argc, const char **argv)
3816 struct evlist *evlist = trace->evlist;
3817 struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3819 unsigned long before;
3820 const bool forks = argc > 0;
3821 bool draining = false;
3825 if (!trace->raw_augmented_syscalls) {
3826 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3827 goto out_error_raw_syscalls;
3829 if (trace->trace_syscalls)
3830 trace->vfs_getname = evlist__add_vfs_getname(evlist);
3833 if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3834 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3835 if (pgfault_maj == NULL)
3837 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3838 evlist__add(evlist, pgfault_maj);
3841 if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3842 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3843 if (pgfault_min == NULL)
3845 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3846 evlist__add(evlist, pgfault_min);
3849 /* Enable ignoring missing threads when -u/-p option is defined. */
3850 trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3853 evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3854 goto out_error_sched_stat_runtime;
3856 * If a global cgroup was set, apply it to all the events without an
3857 * explicit cgroup. I.e.:
3859 * trace -G A -e sched:*switch
3861 * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3862 * _and_ sched:sched_switch to the 'A' cgroup, while:
3864 * trace -e sched:*switch -G A
3866 * will only set the sched:sched_switch event to the 'A' cgroup, all the
3867 * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3868 * a cgroup (on the root cgroup, sys wide, etc).
3872 * trace -G A -e sched:*switch -G B
3874 * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3875 * to the 'B' cgroup.
3877 * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3878 * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3881 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3883 err = evlist__create_maps(evlist, &trace->opts.target);
3885 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3886 goto out_delete_evlist;
3889 err = trace__symbols_init(trace, evlist);
3891 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3892 goto out_delete_evlist;
3895 evlist__config(evlist, &trace->opts, &callchain_param);
3898 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3900 fprintf(trace->output, "Couldn't run the workload!\n");
3901 goto out_delete_evlist;
3903 workload_pid = evlist->workload.pid;
3906 err = evlist__open(evlist);
3908 goto out_error_open;
3909 #ifdef HAVE_BPF_SKEL
3910 if (trace->syscalls.events.bpf_output) {
3911 struct perf_cpu cpu;
3914 * Set up the __augmented_syscalls__ BPF map to hold for each
3915 * CPU the bpf-output event's file descriptor.
3917 perf_cpu_map__for_each_cpu(cpu, i, trace->syscalls.events.bpf_output->core.cpus) {
3918 bpf_map__update_elem(trace->skel->maps.__augmented_syscalls__,
3919 &cpu.cpu, sizeof(int),
3920 xyarray__entry(trace->syscalls.events.bpf_output->core.fd,
3922 sizeof(__u32), BPF_ANY);
3926 err = trace__set_filter_pids(trace);
3930 #ifdef HAVE_BPF_SKEL
3931 if (trace->skel && trace->skel->progs.sys_enter)
3932 trace__init_syscalls_bpf_prog_array_maps(trace);
3935 if (trace->ev_qualifier_ids.nr > 0) {
3936 err = trace__set_ev_qualifier_filter(trace);
3940 if (trace->syscalls.events.sys_exit) {
3941 pr_debug("event qualifier tracepoint filter: %s\n",
3942 trace->syscalls.events.sys_exit->filter);
3947 * If the "close" syscall is not traced, then we will not have the
3948 * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3949 * fd->pathname table and were ending up showing the last value set by
3950 * syscalls opening a pathname and associating it with a descriptor or
3951 * reading it from /proc/pid/fd/ in cases where that doesn't make
3954 * So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3957 trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3959 err = trace__expand_filters(trace, &evsel);
3961 goto out_delete_evlist;
3962 err = evlist__apply_filters(evlist, &evsel);
3964 goto out_error_apply_filters;
3966 err = evlist__mmap(evlist, trace->opts.mmap_pages);
3968 goto out_error_mmap;
3970 if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
3971 evlist__enable(evlist);
3974 evlist__start_workload(evlist);
3976 if (trace->opts.target.initial_delay) {
3977 usleep(trace->opts.target.initial_delay * 1000);
3978 evlist__enable(evlist);
3981 trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3982 perf_thread_map__nr(evlist->core.threads) > 1 ||
3983 evlist__first(evlist)->core.attr.inherit;
3986 * Now that we already used evsel->core.attr to ask the kernel to setup the
3987 * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3988 * trace__resolve_callchain(), allowing per-event max-stack settings
3989 * to override an explicitly set --max-stack global setting.
3991 evlist__for_each_entry(evlist, evsel) {
3992 if (evsel__has_callchain(evsel) &&
3993 evsel->core.attr.sample_max_stack == 0)
3994 evsel->core.attr.sample_max_stack = trace->max_stack;
3997 before = trace->nr_events;
3999 for (i = 0; i < evlist->core.nr_mmaps; i++) {
4000 union perf_event *event;
4003 md = &evlist->mmap[i];
4004 if (perf_mmap__read_init(&md->core) < 0)
4007 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4010 err = trace__deliver_event(trace, event);
4014 perf_mmap__consume(&md->core);
4019 if (done && !draining) {
4020 evlist__disable(evlist);
4024 perf_mmap__read_done(&md->core);
4027 if (trace->nr_events == before) {
4028 int timeout = done ? 100 : -1;
4030 if (!draining && evlist__poll(evlist, timeout) > 0) {
4031 if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4036 if (trace__flush_events(trace))
4044 thread__zput(trace->current);
4046 evlist__disable(evlist);
4048 if (trace->sort_events)
4049 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4053 trace__fprintf_thread_summary(trace, trace->output);
4055 if (trace->show_tool_stats) {
4056 fprintf(trace->output, "Stats:\n "
4057 " vfs_getname : %" PRIu64 "\n"
4058 " proc_getname: %" PRIu64 "\n",
4059 trace->stats.vfs_getname,
4060 trace->stats.proc_getname);
4065 trace__symbols__exit(trace);
4066 evlist__free_syscall_tp_fields(evlist);
4067 evlist__delete(evlist);
4068 cgroup__put(trace->cgroup);
4069 trace->evlist = NULL;
4070 trace->live = false;
4073 char errbuf[BUFSIZ];
4075 out_error_sched_stat_runtime:
4076 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4079 out_error_raw_syscalls:
4080 tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4084 evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4088 evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4091 fprintf(trace->output, "%s\n", errbuf);
4092 goto out_delete_evlist;
4094 out_error_apply_filters:
4095 fprintf(trace->output,
4096 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4097 evsel->filter, evsel__name(evsel), errno,
4098 str_error_r(errno, errbuf, sizeof(errbuf)));
4099 goto out_delete_evlist;
4102 fprintf(trace->output, "Not enough memory to run!\n");
4103 goto out_delete_evlist;
4106 fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4107 goto out_delete_evlist;
4110 static int trace__replay(struct trace *trace)
4112 const struct evsel_str_handler handlers[] = {
4113 { "probe:vfs_getname", trace__vfs_getname, },
4115 struct perf_data data = {
4117 .mode = PERF_DATA_MODE_READ,
4118 .force = trace->force,
4120 struct perf_session *session;
4121 struct evsel *evsel;
4124 trace->tool.sample = trace__process_sample;
4125 trace->tool.mmap = perf_event__process_mmap;
4126 trace->tool.mmap2 = perf_event__process_mmap2;
4127 trace->tool.comm = perf_event__process_comm;
4128 trace->tool.exit = perf_event__process_exit;
4129 trace->tool.fork = perf_event__process_fork;
4130 trace->tool.attr = perf_event__process_attr;
4131 trace->tool.tracing_data = perf_event__process_tracing_data;
4132 trace->tool.build_id = perf_event__process_build_id;
4133 trace->tool.namespaces = perf_event__process_namespaces;
4135 trace->tool.ordered_events = true;
4136 trace->tool.ordering_requires_timestamps = true;
4138 /* add tid to output */
4139 trace->multiple_threads = true;
4141 session = perf_session__new(&data, &trace->tool);
4142 if (IS_ERR(session))
4143 return PTR_ERR(session);
4145 if (trace->opts.target.pid)
4146 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4148 if (trace->opts.target.tid)
4149 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4151 if (symbol__init(&session->header.env) < 0)
4154 trace->host = &session->machines.host;
4156 err = perf_session__set_tracepoints_handlers(session, handlers);
4160 evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4161 trace->syscalls.events.sys_enter = evsel;
4162 /* older kernels have syscalls tp versus raw_syscalls */
4164 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4167 (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4168 perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4169 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4173 evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4174 trace->syscalls.events.sys_exit = evsel;
4176 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4178 (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4179 perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4180 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4184 evlist__for_each_entry(session->evlist, evsel) {
4185 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4186 (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4187 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4188 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4189 evsel->handler = trace__pgfault;
4194 err = perf_session__process_events(session);
4196 pr_err("Failed to process events, error %d", err);
4198 else if (trace->summary)
4199 trace__fprintf_thread_summary(trace, trace->output);
4202 perf_session__delete(session);
4207 static size_t trace__fprintf_threads_header(FILE *fp)
4211 printed = fprintf(fp, "\n Summary of events:\n\n");
4216 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4217 struct syscall_stats *stats;
4222 struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4223 struct syscall_stats *stats = source->priv;
4225 entry->syscall = source->i;
4226 entry->stats = stats;
4227 entry->msecs = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4230 static size_t thread__dump_stats(struct thread_trace *ttrace,
4231 struct trace *trace, FILE *fp)
4236 DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4238 if (syscall_stats == NULL)
4241 printed += fprintf(fp, "\n");
4243 printed += fprintf(fp, " syscall calls errors total min avg max stddev\n");
4244 printed += fprintf(fp, " (msec) (msec) (msec) (msec) (%%)\n");
4245 printed += fprintf(fp, " --------------- -------- ------ -------- --------- --------- --------- ------\n");
4247 resort_rb__for_each_entry(nd, syscall_stats) {
4248 struct syscall_stats *stats = syscall_stats_entry->stats;
4250 double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4251 double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4252 double avg = avg_stats(&stats->stats);
4254 u64 n = (u64)stats->stats.n;
4256 pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4257 avg /= NSEC_PER_MSEC;
4259 sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4260 printed += fprintf(fp, " %-15s", sc->name);
4261 printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4262 n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4263 printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4265 if (trace->errno_summary && stats->nr_failures) {
4268 for (e = 0; e < stats->max_errno; ++e) {
4269 if (stats->errnos[e] != 0)
4270 fprintf(fp, "\t\t\t\t%s: %d\n", perf_env__arch_strerrno(trace->host->env, e + 1), stats->errnos[e]);
4276 resort_rb__delete(syscall_stats);
4277 printed += fprintf(fp, "\n\n");
4282 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4285 struct thread_trace *ttrace = thread__priv(thread);
4291 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4293 printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread__tid(thread));
4294 printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4295 printed += fprintf(fp, "%.1f%%", ratio);
4297 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4299 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4301 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4302 else if (fputc('\n', fp) != EOF)
4305 printed += thread__dump_stats(ttrace, trace, fp);
4310 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4312 return ttrace ? ttrace->nr_events : 0;
4315 DEFINE_RESORT_RB(threads,
4316 (thread__nr_events(thread__priv(a->thread)) <
4317 thread__nr_events(thread__priv(b->thread))),
4318 struct thread *thread;
4321 entry->thread = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
4324 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4326 size_t printed = trace__fprintf_threads_header(fp);
4330 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4331 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4333 if (threads == NULL) {
4334 fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4338 resort_rb__for_each_entry(nd, threads)
4339 printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4341 resort_rb__delete(threads);
4346 static int trace__set_duration(const struct option *opt, const char *str,
4347 int unset __maybe_unused)
4349 struct trace *trace = opt->value;
4351 trace->duration_filter = atof(str);
4355 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4356 int unset __maybe_unused)
4360 struct trace *trace = opt->value;
4362 * FIXME: introduce a intarray class, plain parse csv and create a
4363 * { int nr, int entries[] } struct...
4365 struct intlist *list = intlist__new(str);
4370 i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4371 trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4373 if (trace->filter_pids.entries == NULL)
4376 trace->filter_pids.entries[0] = getpid();
4378 for (i = 1; i < trace->filter_pids.nr; ++i)
4379 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4381 intlist__delete(list);
4387 static int trace__open_output(struct trace *trace, const char *filename)
4391 if (!stat(filename, &st) && st.st_size) {
4392 char oldname[PATH_MAX];
4394 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4396 rename(filename, oldname);
4399 trace->output = fopen(filename, "w");
4401 return trace->output == NULL ? -errno : 0;
4404 static int parse_pagefaults(const struct option *opt, const char *str,
4405 int unset __maybe_unused)
4407 int *trace_pgfaults = opt->value;
4409 if (strcmp(str, "all") == 0)
4410 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4411 else if (strcmp(str, "maj") == 0)
4412 *trace_pgfaults |= TRACE_PFMAJ;
4413 else if (strcmp(str, "min") == 0)
4414 *trace_pgfaults |= TRACE_PFMIN;
4421 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4423 struct evsel *evsel;
4425 evlist__for_each_entry(evlist, evsel) {
4426 if (evsel->handler == NULL)
4427 evsel->handler = handler;
4431 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4433 struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4436 const struct syscall_fmt *scfmt = syscall_fmt__find(name);
4441 if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4442 strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4445 memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4450 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4452 struct evsel *evsel;
4454 evlist__for_each_entry(evlist, evsel) {
4455 if (evsel->priv || !evsel->tp_format)
4458 if (strcmp(evsel->tp_format->system, "syscalls")) {
4459 evsel__init_tp_arg_scnprintf(evsel);
4463 if (evsel__init_syscall_tp(evsel))
4466 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4467 struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4469 if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4472 evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4473 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4474 struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4476 if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4479 evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4487 * XXX: Hackish, just splitting the combined -e+--event (syscalls
4488 * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4489 * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4491 * It'd be better to introduce a parse_options() variant that would return a
4492 * list with the terms it didn't match to an event...
4494 static int trace__parse_events_option(const struct option *opt, const char *str,
4495 int unset __maybe_unused)
4497 struct trace *trace = (struct trace *)opt->value;
4498 const char *s = str;
4499 char *sep = NULL, *lists[2] = { NULL, NULL, };
4500 int len = strlen(str) + 1, err = -1, list, idx;
4501 char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4502 char group_name[PATH_MAX];
4503 const struct syscall_fmt *fmt;
4505 if (strace_groups_dir == NULL)
4510 trace->not_ev_qualifier = true;
4514 if ((sep = strchr(s, ',')) != NULL)
4518 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4519 syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4524 fmt = syscall_fmt__find_by_alias(s);
4529 path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4530 if (access(group_name, R_OK) == 0)
4535 sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4537 lists[list] = malloc(len);
4538 if (lists[list] == NULL)
4540 strcpy(lists[list], s);
4550 if (lists[1] != NULL) {
4551 struct strlist_config slist_config = {
4552 .dirname = strace_groups_dir,
4555 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4556 if (trace->ev_qualifier == NULL) {
4557 fputs("Not enough memory to parse event qualifier", trace->output);
4561 if (trace__validate_ev_qualifier(trace))
4563 trace->trace_syscalls = true;
4569 struct parse_events_option_args parse_events_option_args = {
4570 .evlistp = &trace->evlist,
4573 .value = &parse_events_option_args,
4575 err = parse_events_option(&o, lists[0], 0);
4578 free(strace_groups_dir);
4587 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4589 struct trace *trace = opt->value;
4591 if (!list_empty(&trace->evlist->core.entries)) {
4593 .value = &trace->evlist,
4595 return parse_cgroups(&o, str, unset);
4597 trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4602 static int trace__config(const char *var, const char *value, void *arg)
4604 struct trace *trace = arg;
4607 if (!strcmp(var, "trace.add_events")) {
4608 trace->perfconfig_events = strdup(value);
4609 if (trace->perfconfig_events == NULL) {
4610 pr_err("Not enough memory for %s\n", "trace.add_events");
4613 } else if (!strcmp(var, "trace.show_timestamp")) {
4614 trace->show_tstamp = perf_config_bool(var, value);
4615 } else if (!strcmp(var, "trace.show_duration")) {
4616 trace->show_duration = perf_config_bool(var, value);
4617 } else if (!strcmp(var, "trace.show_arg_names")) {
4618 trace->show_arg_names = perf_config_bool(var, value);
4619 if (!trace->show_arg_names)
4620 trace->show_zeros = true;
4621 } else if (!strcmp(var, "trace.show_zeros")) {
4622 bool new_show_zeros = perf_config_bool(var, value);
4623 if (!trace->show_arg_names && !new_show_zeros) {
4624 pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4627 trace->show_zeros = new_show_zeros;
4628 } else if (!strcmp(var, "trace.show_prefix")) {
4629 trace->show_string_prefix = perf_config_bool(var, value);
4630 } else if (!strcmp(var, "trace.no_inherit")) {
4631 trace->opts.no_inherit = perf_config_bool(var, value);
4632 } else if (!strcmp(var, "trace.args_alignment")) {
4633 int args_alignment = 0;
4634 if (perf_config_int(&args_alignment, var, value) == 0)
4635 trace->args_alignment = args_alignment;
4636 } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4637 if (strcasecmp(value, "libtraceevent") == 0)
4638 trace->libtraceevent_print = true;
4639 else if (strcasecmp(value, "libbeauty") == 0)
4640 trace->libtraceevent_print = false;
4646 static void trace__exit(struct trace *trace)
4650 strlist__delete(trace->ev_qualifier);
4651 zfree(&trace->ev_qualifier_ids.entries);
4652 if (trace->syscalls.table) {
4653 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4654 syscall__exit(&trace->syscalls.table[i]);
4655 zfree(&trace->syscalls.table);
4657 syscalltbl__delete(trace->sctbl);
4658 zfree(&trace->perfconfig_events);
4661 #ifdef HAVE_BPF_SKEL
4662 static int bpf__setup_bpf_output(struct evlist *evlist)
4664 int err = parse_event(evlist, "bpf-output/no-inherit=1,name=__augmented_syscalls__/");
4667 pr_debug("ERROR: failed to create the \"__augmented_syscalls__\" bpf-output event\n");
4673 int cmd_trace(int argc, const char **argv)
4675 const char *trace_usage[] = {
4676 "perf trace [<options>] [<command>]",
4677 "perf trace [<options>] -- <command> [<options>]",
4678 "perf trace record [<options>] [<command>]",
4679 "perf trace record [<options>] -- <command> [<options>]",
4682 struct trace trace = {
4688 .user_freq = UINT_MAX,
4689 .user_interval = ULLONG_MAX,
4690 .no_buffering = true,
4691 .mmap_pages = UINT_MAX,
4695 .show_tstamp = true,
4696 .show_duration = true,
4697 .show_arg_names = true,
4698 .args_alignment = 70,
4699 .trace_syscalls = false,
4700 .kernel_syscallchains = false,
4701 .max_stack = UINT_MAX,
4702 .max_events = ULONG_MAX,
4704 const char *output_name = NULL;
4705 const struct option trace_options[] = {
4706 OPT_CALLBACK('e', "event", &trace, "event",
4707 "event/syscall selector. use 'perf list' to list available events",
4708 trace__parse_events_option),
4709 OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4710 "event filter", parse_filter),
4711 OPT_BOOLEAN(0, "comm", &trace.show_comm,
4712 "show the thread COMM next to its id"),
4713 OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4714 OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4715 trace__parse_events_option),
4716 OPT_STRING('o', "output", &output_name, "file", "output file name"),
4717 OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4718 OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4719 "trace events on existing process id"),
4720 OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4721 "trace events on existing thread id"),
4722 OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4723 "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4724 OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4725 "system-wide collection from all CPUs"),
4726 OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4727 "list of cpus to monitor"),
4728 OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4729 "child tasks do not inherit counters"),
4730 OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4731 "number of mmap data pages", evlist__parse_mmap_pages),
4732 OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4734 OPT_CALLBACK(0, "duration", &trace, "float",
4735 "show only events with duration > N.M ms",
4736 trace__set_duration),
4737 OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4738 OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4739 OPT_BOOLEAN('T', "time", &trace.full_time,
4740 "Show full timestamp, not time relative to first start"),
4741 OPT_BOOLEAN(0, "failure", &trace.failure_only,
4742 "Show only syscalls that failed"),
4743 OPT_BOOLEAN('s', "summary", &trace.summary_only,
4744 "Show only syscall summary with statistics"),
4745 OPT_BOOLEAN('S', "with-summary", &trace.summary,
4746 "Show all syscalls and summary with statistics"),
4747 OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4748 "Show errno stats per syscall, use with -s or -S"),
4749 OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4750 "Trace pagefaults", parse_pagefaults, "maj"),
4751 OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4752 OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4753 OPT_CALLBACK(0, "call-graph", &trace.opts,
4754 "record_mode[,record_size]", record_callchain_help,
4755 &record_parse_callchain_opt),
4756 OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4757 "Use libtraceevent to print the tracepoint arguments."),
4758 OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4759 "Show the kernel callchains on the syscall exit path"),
4760 OPT_ULONG(0, "max-events", &trace.max_events,
4761 "Set the maximum number of events to print, exit after that is reached. "),
4762 OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4763 "Set the minimum stack depth when parsing the callchain, "
4764 "anything below the specified depth will be ignored."),
4765 OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4766 "Set the maximum stack depth when parsing the callchain, "
4767 "anything beyond the specified depth will be ignored. "
4768 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4769 OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4770 "Sort batch of events before processing, use if getting out of order events"),
4771 OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4772 "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4773 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4774 "per thread proc mmap processing timeout in ms"),
4775 OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4776 trace__parse_cgroups),
4777 OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4778 "ms to wait before starting measurement after program "
4780 OPTS_EVSWITCH(&trace.evswitch),
4783 bool __maybe_unused max_stack_user_set = true;
4784 bool mmap_pages_user_set = true;
4785 struct evsel *evsel;
4786 const char * const trace_subcommands[] = { "record", NULL };
4789 struct sigaction sigchld_act;
4791 signal(SIGSEGV, sighandler_dump_stack);
4792 signal(SIGFPE, sighandler_dump_stack);
4793 signal(SIGINT, sighandler_interrupt);
4795 memset(&sigchld_act, 0, sizeof(sigchld_act));
4796 sigchld_act.sa_flags = SA_SIGINFO;
4797 sigchld_act.sa_sigaction = sighandler_chld;
4798 sigaction(SIGCHLD, &sigchld_act, NULL);
4800 trace.evlist = evlist__new();
4801 trace.sctbl = syscalltbl__new();
4803 if (trace.evlist == NULL || trace.sctbl == NULL) {
4804 pr_err("Not enough memory to run!\n");
4810 * Parsing .perfconfig may entail creating a BPF event, that may need
4811 * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4812 * is too small. This affects just this process, not touching the
4813 * global setting. If it fails we'll get something in 'perf trace -v'
4814 * to help diagnose the problem.
4816 rlimit__bump_memlock();
4818 err = perf_config(trace__config, &trace);
4822 argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4823 trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4826 * Here we already passed thru trace__parse_events_option() and it has
4827 * already figured out if -e syscall_name, if not but if --event
4828 * foo:bar was used, the user is interested _just_ in those, say,
4829 * tracepoint events, not in the strace-like syscall-name-based mode.
4831 * This is important because we need to check if strace-like mode is
4832 * needed to decided if we should filter out the eBPF
4833 * __augmented_syscalls__ code, if it is in the mix, say, via
4834 * .perfconfig trace.add_events, and filter those out.
4836 if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4837 trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4838 trace.trace_syscalls = true;
4841 * Now that we have --verbose figured out, lets see if we need to parse
4842 * events from .perfconfig, so that if those events fail parsing, say some
4843 * BPF program fails, then we'll be able to use --verbose to see what went
4844 * wrong in more detail.
4846 if (trace.perfconfig_events != NULL) {
4847 struct parse_events_error parse_err;
4849 parse_events_error__init(&parse_err);
4850 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4852 parse_events_error__print(&parse_err, trace.perfconfig_events);
4853 parse_events_error__exit(&parse_err);
4858 if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4859 usage_with_options_msg(trace_usage, trace_options,
4860 "cgroup monitoring only available in system-wide mode");
4863 #ifdef HAVE_BPF_SKEL
4864 if (!trace.trace_syscalls)
4865 goto skip_augmentation;
4867 trace.skel = augmented_raw_syscalls_bpf__open();
4869 pr_debug("Failed to open augmented syscalls BPF skeleton");
4872 * Disable attaching the BPF programs except for sys_enter and
4873 * sys_exit that tail call into this as necessary.
4875 struct bpf_program *prog;
4877 bpf_object__for_each_program(prog, trace.skel->obj) {
4878 if (prog != trace.skel->progs.sys_enter && prog != trace.skel->progs.sys_exit)
4879 bpf_program__set_autoattach(prog, /*autoattach=*/false);
4882 err = augmented_raw_syscalls_bpf__load(trace.skel);
4885 libbpf_strerror(err, bf, sizeof(bf));
4886 pr_debug("Failed to load augmented syscalls BPF skeleton: %s\n", bf);
4888 augmented_raw_syscalls_bpf__attach(trace.skel);
4889 trace__add_syscall_newtp(&trace);
4893 err = bpf__setup_bpf_output(trace.evlist);
4895 libbpf_strerror(err, bf, sizeof(bf));
4896 pr_err("ERROR: Setup BPF output event failed: %s\n", bf);
4899 trace.syscalls.events.bpf_output = evlist__last(trace.evlist);
4900 assert(!strcmp(evsel__name(trace.syscalls.events.bpf_output), "__augmented_syscalls__"));
4905 if (trace.trace_pgfaults) {
4906 trace.opts.sample_address = true;
4907 trace.opts.sample_time = true;
4910 if (trace.opts.mmap_pages == UINT_MAX)
4911 mmap_pages_user_set = false;
4913 if (trace.max_stack == UINT_MAX) {
4914 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4915 max_stack_user_set = false;
4918 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4919 if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4920 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4924 if (callchain_param.enabled) {
4925 if (!mmap_pages_user_set && geteuid() == 0)
4926 trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4928 symbol_conf.use_callchain = true;
4931 if (trace.evlist->core.nr_entries > 0) {
4932 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4933 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4934 perror("failed to set syscalls:* tracepoint fields");
4939 if (trace.sort_events) {
4940 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4941 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4945 * If we are augmenting syscalls, then combine what we put in the
4946 * __augmented_syscalls__ BPF map with what is in the
4947 * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4948 * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4950 * We'll switch to look at two BPF maps, one for sys_enter and the
4951 * other for sys_exit when we start augmenting the sys_exit paths with
4952 * buffers that are being copied from kernel to userspace, think 'read'
4955 if (trace.syscalls.events.bpf_output) {
4956 evlist__for_each_entry(trace.evlist, evsel) {
4957 bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4959 if (raw_syscalls_sys_exit) {
4960 trace.raw_augmented_syscalls = true;
4961 goto init_augmented_syscall_tp;
4964 if (trace.syscalls.events.bpf_output->priv == NULL &&
4965 strstr(evsel__name(evsel), "syscalls:sys_enter")) {
4966 struct evsel *augmented = trace.syscalls.events.bpf_output;
4967 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
4968 evsel__init_augmented_syscall_tp_args(augmented))
4971 * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4972 * Above we made sure we can get from the payload the tp fields
4973 * that we get from syscalls:sys_enter tracefs format file.
4975 augmented->handler = trace__sys_enter;
4977 * Now we do the same for the *syscalls:sys_enter event so that
4978 * if we handle it directly, i.e. if the BPF prog returns 0 so
4979 * as not to filter it, then we'll handle it just like we would
4980 * for the BPF_OUTPUT one:
4982 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
4983 evsel__init_augmented_syscall_tp_args(evsel))
4985 evsel->handler = trace__sys_enter;
4988 if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
4989 struct syscall_tp *sc;
4990 init_augmented_syscall_tp:
4991 if (evsel__init_augmented_syscall_tp(evsel, evsel))
4993 sc = __evsel__syscall_tp(evsel);
4995 * For now with BPF raw_augmented we hook into
4996 * raw_syscalls:sys_enter and there we get all
4997 * 6 syscall args plus the tracepoint common
4998 * fields and the syscall_nr (another long).
4999 * So we check if that is the case and if so
5000 * don't look after the sc->args_size but
5001 * always after the full raw_syscalls:sys_enter
5002 * payload, which is fixed.
5004 * We'll revisit this later to pass
5005 * s->args_size to the BPF augmenter (now
5006 * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5007 * so that it copies only what we need for each
5008 * syscall, like what happens when we use
5009 * syscalls:sys_enter_NAME, so that we reduce
5010 * the kernel/userspace traffic to just what is
5011 * needed for each syscall.
5013 if (trace.raw_augmented_syscalls)
5014 trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5015 evsel__init_augmented_syscall_tp_ret(evsel);
5016 evsel->handler = trace__sys_exit;
5021 if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5022 return trace__record(&trace, argc-1, &argv[1]);
5024 /* Using just --errno-summary will trigger --summary */
5025 if (trace.errno_summary && !trace.summary && !trace.summary_only)
5026 trace.summary_only = true;
5028 /* summary_only implies summary option, but don't overwrite summary if set */
5029 if (trace.summary_only)
5030 trace.summary = trace.summary_only;
5032 if (output_name != NULL) {
5033 err = trace__open_output(&trace, output_name);
5035 perror("failed to create output file");
5040 err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5044 err = target__validate(&trace.opts.target);
5046 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5047 fprintf(trace.output, "%s", bf);
5051 err = target__parse_uid(&trace.opts.target);
5053 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5054 fprintf(trace.output, "%s", bf);
5058 if (!argc && target__none(&trace.opts.target))
5059 trace.opts.target.system_wide = true;
5062 err = trace__replay(&trace);
5064 err = trace__run(&trace, argc, argv);
5067 if (output_name != NULL)
5068 fclose(trace.output);
5070 trace__exit(&trace);
5071 #ifdef HAVE_BPF_SKEL
5072 augmented_raw_syscalls_bpf__destroy(trace.skel);