git.rot13.org Git - linux/blob - tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   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.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  *
  16  * Released under the GPL v2. (and only v2, not any later version)
  17  */
  18
  19 #include <traceevent/event-parse.h>
  20 #include <api/fs/tracing_path.h>
  21 #include <bpf/bpf.h>
  22 #include "builtin.h"
  23 #include "util/cgroup.h"
  24 #include "util/color.h"
  25 #include "util/config.h"
  26 #include "util/debug.h"
  27 #include "util/env.h"
  28 #include "util/event.h"
  29 #include "util/evlist.h"
  30 #include <subcmd/exec-cmd.h>
  31 #include "util/machine.h"
  32 #include "util/path.h"
  33 #include "util/session.h"
  34 #include "util/thread.h"
  35 #include <subcmd/parse-options.h>
  36 #include "util/strlist.h"
  37 #include "util/intlist.h"
  38 #include "util/thread_map.h"
  39 #include "util/stat.h"
  40 #include "trace/beauty/beauty.h"
  41 #include "trace-event.h"
  42 #include "util/parse-events.h"
  43 #include "util/bpf-loader.h"
  44 #include "callchain.h"
  45 #include "print_binary.h"
  46 #include "string2.h"
  47 #include "syscalltbl.h"
  48 #include "rb_resort.h"
  49
  50 #include <errno.h>
  51 #include <inttypes.h>
  52 #include <poll.h>
  53 #include <signal.h>
  54 #include <stdlib.h>
  55 #include <string.h>
  56 #include <linux/err.h>
  57 #include <linux/filter.h>
  58 #include <linux/kernel.h>
  59 #include <linux/random.h>
  60 #include <linux/stringify.h>
  61 #include <linux/time64.h>
  62 #include <fcntl.h>
  63
  64 #include "sane_ctype.h"
  65
  66 #ifndef O_CLOEXEC
  67 # define O_CLOEXEC              02000000
  68 #endif
  69
  70 #ifndef F_LINUX_SPECIFIC_BASE
  71 # define F_LINUX_SPECIFIC_BASE  1024
  72 #endif
  73
  74 struct trace {
  75         struct perf_tool        tool;
  76         struct syscalltbl       *sctbl;
  77         struct {
  78                 int             max;
  79                 struct syscall  *table;
  80                 struct bpf_map  *map;
  81                 struct {
  82                         struct perf_evsel *sys_enter,
  83                                           *sys_exit,
  84                                           *augmented;
  85                 }               events;
  86         } syscalls;
  87         struct record_opts      opts;
  88         struct perf_evlist      *evlist;
  89         struct machine          *host;
  90         struct thread           *current;
  91         struct cgroup           *cgroup;
  92         u64                     base_time;
  93         FILE                    *output;
  94         unsigned long           nr_events;
  95         unsigned long           nr_events_printed;
  96         unsigned long           max_events;
  97         struct strlist          *ev_qualifier;
  98         struct {
  99                 size_t          nr;
 100                 int             *entries;
 101         }                       ev_qualifier_ids;
 102         struct {
 103                 size_t          nr;
 104                 pid_t           *entries;
 105                 struct bpf_map  *map;
 106         }                       filter_pids;
 107         double                  duration_filter;
 108         double                  runtime_ms;
 109         struct {
 110                 u64             vfs_getname,
 111                                 proc_getname;
 112         } stats;
 113         unsigned int            max_stack;
 114         unsigned int            min_stack;
 115         bool                    sort_events;
 116         bool                    raw_augmented_syscalls;
 117         bool                    not_ev_qualifier;
 118         bool                    live;
 119         bool                    full_time;
 120         bool                    sched;
 121         bool                    multiple_threads;
 122         bool                    summary;
 123         bool                    summary_only;
 124         bool                    failure_only;
 125         bool                    show_comm;
 126         bool                    print_sample;
 127         bool                    show_tool_stats;
 128         bool                    trace_syscalls;
 129         bool                    kernel_syscallchains;
 130         s16                     args_alignment;
 131         bool                    show_tstamp;
 132         bool                    show_duration;
 133         bool                    show_zeros;
 134         bool                    show_arg_names;
 135         bool                    show_string_prefix;
 136         bool                    force;
 137         bool                    vfs_getname;
 138         int                     trace_pgfaults;
 139         struct {
 140                 struct ordered_events   data;
 141                 u64                     last;
 142         } oe;
 143 };
 144
 145 struct tp_field {
 146         int offset;
 147         union {
 148                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 149                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 150         };
 151 };
 152
 153 #define TP_UINT_FIELD(bits) \
 154 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 155 { \
 156         u##bits value; \
 157         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 158         return value;  \
 159 }
 160
 161 TP_UINT_FIELD(8);
 162 TP_UINT_FIELD(16);
 163 TP_UINT_FIELD(32);
 164 TP_UINT_FIELD(64);
 165
 166 #define TP_UINT_FIELD__SWAPPED(bits) \
 167 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 168 { \
 169         u##bits value; \
 170         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 171         return bswap_##bits(value);\
 172 }
 173
 174 TP_UINT_FIELD__SWAPPED(16);
 175 TP_UINT_FIELD__SWAPPED(32);
 176 TP_UINT_FIELD__SWAPPED(64);
 177
 178 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 179 {
 180         field->offset = offset;
 181
 182         switch (size) {
 183         case 1:
 184                 field->integer = tp_field__u8;
 185                 break;
 186         case 2:
 187                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 188                 break;
 189         case 4:
 190                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 191                 break;
 192         case 8:
 193                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 194                 break;
 195         default:
 196                 return -1;
 197         }
 198
 199         return 0;
 200 }
 201
 202 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 203 {
 204         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 205 }
 206
 207 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 208 {
 209         return sample->raw_data + field->offset;
 210 }
 211
 212 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 213 {
 214         field->offset = offset;
 215         field->pointer = tp_field__ptr;
 216         return 0;
 217 }
 218
 219 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 220 {
 221         return __tp_field__init_ptr(field, format_field->offset);
 222 }
 223
 224 struct syscall_tp {
 225         struct tp_field id;
 226         union {
 227                 struct tp_field args, ret;
 228         };
 229 };
 230
 231 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 232                                           struct tp_field *field,
 233                                           const char *name)
 234 {
 235         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 236
 237         if (format_field == NULL)
 238                 return -1;
 239
 240         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 241 }
 242
 243 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 244         ({ struct syscall_tp *sc = evsel->priv;\
 245            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 246
 247 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 248                                          struct tp_field *field,
 249                                          const char *name)
 250 {
 251         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 252
 253         if (format_field == NULL)
 254                 return -1;
 255
 256         return tp_field__init_ptr(field, format_field);
 257 }
 258
 259 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 260         ({ struct syscall_tp *sc = evsel->priv;\
 261            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 262
 263 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 264 {
 265         zfree(&evsel->priv);
 266         perf_evsel__delete(evsel);
 267 }
 268
 269 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 270 {
 271         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 272
 273         if (evsel->priv != NULL) {
 274                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 275                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 276                         goto out_delete;
 277                 return 0;
 278         }
 279
 280         return -ENOMEM;
 281 out_delete:
 282         zfree(&evsel->priv);
 283         return -ENOENT;
 284 }
 285
 286 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel)
 287 {
 288         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 289
 290         if (evsel->priv != NULL) {       /* field, sizeof_field, offsetof_field */
 291                 if (__tp_field__init_uint(&sc->id, sizeof(long), sizeof(long long), evsel->needs_swap))
 292                         goto out_delete;
 293
 294                 return 0;
 295         }
 296
 297         return -ENOMEM;
 298 out_delete:
 299         zfree(&evsel->priv);
 300         return -EINVAL;
 301 }
 302
 303 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 304 {
 305         struct syscall_tp *sc = evsel->priv;
 306
 307         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 308 }
 309
 310 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 311 {
 312         struct syscall_tp *sc = evsel->priv;
 313
 314         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 315 }
 316
 317 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 318 {
 319         evsel->priv = malloc(sizeof(struct syscall_tp));
 320         if (evsel->priv != NULL) {
 321                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 322                         goto out_delete;
 323
 324                 evsel->handler = handler;
 325                 return 0;
 326         }
 327
 328         return -ENOMEM;
 329
 330 out_delete:
 331         zfree(&evsel->priv);
 332         return -ENOENT;
 333 }
 334
 335 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 336 {
 337         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 338
 339         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 340         if (IS_ERR(evsel))
 341                 evsel = perf_evsel__newtp("syscalls", direction);
 342
 343         if (IS_ERR(evsel))
 344                 return NULL;
 345
 346         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 347                 goto out_delete;
 348
 349         return evsel;
 350
 351 out_delete:
 352         perf_evsel__delete_priv(evsel);
 353         return NULL;
 354 }
 355
 356 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 357         ({ struct syscall_tp *fields = evsel->priv; \
 358            fields->name.integer(&fields->name, sample); })
 359
 360 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 361         ({ struct syscall_tp *fields = evsel->priv; \
 362            fields->name.pointer(&fields->name, sample); })
 363
 364 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 365 {
 366         int idx = val - sa->offset;
 367
 368         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 369                 size_t printed = scnprintf(bf, size, intfmt, val);
 370                 if (show_prefix)
 371                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 372                 return printed;
 373         }
 374
 375         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 376 }
 377
 378 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 379                                                 const char *intfmt,
 380                                                 struct syscall_arg *arg)
 381 {
 382         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 383 }
 384
 385 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 386                                               struct syscall_arg *arg)
 387 {
 388         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 389 }
 390
 391 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 392
 393 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 394 {
 395         size_t printed;
 396         int i;
 397
 398         for (i = 0; i < sas->nr_entries; ++i) {
 399                 struct strarray *sa = sas->entries[i];
 400                 int idx = val - sa->offset;
 401
 402                 if (idx >= 0 && idx < sa->nr_entries) {
 403                         if (sa->entries[idx] == NULL)
 404                                 break;
 405                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 406                 }
 407         }
 408
 409         printed = scnprintf(bf, size, intfmt, val);
 410         if (show_prefix)
 411                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 412         return printed;
 413 }
 414
 415 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 416                                         struct syscall_arg *arg)
 417 {
 418         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 419 }
 420
 421 #ifndef AT_FDCWD
 422 #define AT_FDCWD        -100
 423 #endif
 424
 425 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 426                                            struct syscall_arg *arg)
 427 {
 428         int fd = arg->val;
 429         const char *prefix = "AT_FD";
 430
 431         if (fd == AT_FDCWD)
 432                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 433
 434         return syscall_arg__scnprintf_fd(bf, size, arg);
 435 }
 436
 437 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 438
 439 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 440                                               struct syscall_arg *arg);
 441
 442 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 443
 444 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 445 {
 446         return scnprintf(bf, size, "%#lx", arg->val);
 447 }
 448
 449 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 450 {
 451         if (arg->val == 0)
 452                 return scnprintf(bf, size, "NULL");
 453         return syscall_arg__scnprintf_hex(bf, size, arg);
 454 }
 455
 456 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 457 {
 458         return scnprintf(bf, size, "%d", arg->val);
 459 }
 460
 461 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 462 {
 463         return scnprintf(bf, size, "%ld", arg->val);
 464 }
 465
 466 static const char *bpf_cmd[] = {
 467         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 468         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 469 };
 470 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 471
 472 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 473 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 474
 475 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 476 static DEFINE_STRARRAY(itimers, "ITIMER_");
 477
 478 static const char *keyctl_options[] = {
 479         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 480         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 481         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 482         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 483         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 484 };
 485 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 486
 487 static const char *whences[] = { "SET", "CUR", "END",
 488 #ifdef SEEK_DATA
 489 "DATA",
 490 #endif
 491 #ifdef SEEK_HOLE
 492 "HOLE",
 493 #endif
 494 };
 495 static DEFINE_STRARRAY(whences, "SEEK_");
 496
 497 static const char *fcntl_cmds[] = {
 498         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 499         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 500         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 501         "GETOWNER_UIDS",
 502 };
 503 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 504
 505 static const char *fcntl_linux_specific_cmds[] = {
 506         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 507         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 508         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 509 };
 510
 511 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 512
 513 static struct strarray *fcntl_cmds_arrays[] = {
 514         &strarray__fcntl_cmds,
 515         &strarray__fcntl_linux_specific_cmds,
 516 };
 517
 518 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 519
 520 static const char *rlimit_resources[] = {
 521         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 522         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 523         "RTTIME",
 524 };
 525 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 526
 527 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 528 static DEFINE_STRARRAY(sighow, "SIG_");
 529
 530 static const char *clockid[] = {
 531         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 532         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 533         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 534 };
 535 static DEFINE_STRARRAY(clockid, "CLOCK_");
 536
 537 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 538                                                  struct syscall_arg *arg)
 539 {
 540         bool show_prefix = arg->show_string_prefix;
 541         const char *suffix = "_OK";
 542         size_t printed = 0;
 543         int mode = arg->val;
 544
 545         if (mode == F_OK) /* 0 */
 546                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 547 #define P_MODE(n) \
 548         if (mode & n##_OK) { \
 549                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 550                 mode &= ~n##_OK; \
 551         }
 552
 553         P_MODE(R);
 554         P_MODE(W);
 555         P_MODE(X);
 556 #undef P_MODE
 557
 558         if (mode)
 559                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 560
 561         return printed;
 562 }
 563
 564 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 565
 566 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 567                                               struct syscall_arg *arg);
 568
 569 #define SCA_FILENAME syscall_arg__scnprintf_filename
 570
 571 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 572                                                 struct syscall_arg *arg)
 573 {
 574         bool show_prefix = arg->show_string_prefix;
 575         const char *prefix = "O_";
 576         int printed = 0, flags = arg->val;
 577
 578 #define P_FLAG(n) \
 579         if (flags & O_##n) { \
 580                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 581                 flags &= ~O_##n; \
 582         }
 583
 584         P_FLAG(CLOEXEC);
 585         P_FLAG(NONBLOCK);
 586 #undef P_FLAG
 587
 588         if (flags)
 589                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 590
 591         return printed;
 592 }
 593
 594 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 595
 596 #ifndef GRND_NONBLOCK
 597 #define GRND_NONBLOCK   0x0001
 598 #endif
 599 #ifndef GRND_RANDOM
 600 #define GRND_RANDOM     0x0002
 601 #endif
 602
 603 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 604                                                    struct syscall_arg *arg)
 605 {
 606         bool show_prefix = arg->show_string_prefix;
 607         const char *prefix = "GRND_";
 608         int printed = 0, flags = arg->val;
 609
 610 #define P_FLAG(n) \
 611         if (flags & GRND_##n) { \
 612                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 613                 flags &= ~GRND_##n; \
 614         }
 615
 616         P_FLAG(RANDOM);
 617         P_FLAG(NONBLOCK);
 618 #undef P_FLAG
 619
 620         if (flags)
 621                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 622
 623         return printed;
 624 }
 625
 626 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 627
 628 #define STRARRAY(name, array) \
 629           { .scnprintf  = SCA_STRARRAY, \
 630             .parm       = &strarray__##array, }
 631
 632 #include "trace/beauty/arch_errno_names.c"
 633 #include "trace/beauty/eventfd.c"
 634 #include "trace/beauty/futex_op.c"
 635 #include "trace/beauty/futex_val3.c"
 636 #include "trace/beauty/mmap.c"
 637 #include "trace/beauty/mode_t.c"
 638 #include "trace/beauty/msg_flags.c"
 639 #include "trace/beauty/open_flags.c"
 640 #include "trace/beauty/perf_event_open.c"
 641 #include "trace/beauty/pid.c"
 642 #include "trace/beauty/sched_policy.c"
 643 #include "trace/beauty/seccomp.c"
 644 #include "trace/beauty/signum.c"
 645 #include "trace/beauty/socket_type.c"
 646 #include "trace/beauty/waitid_options.c"
 647
 648 struct syscall_arg_fmt {
 649         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 650         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 651         void       *parm;
 652         const char *name;
 653         bool       show_zero;
 654 };
 655
 656 static struct syscall_fmt {
 657         const char *name;
 658         const char *alias;
 659         struct syscall_arg_fmt arg[6];
 660         u8         nr_args;
 661         bool       errpid;
 662         bool       timeout;
 663         bool       hexret;
 664 } syscall_fmts[] = {
 665         { .name     = "access",
 666           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 667         { .name     = "arch_prctl",
 668           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 669                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 670         { .name     = "bind",
 671           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 672         { .name     = "bpf",
 673           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 674         { .name     = "brk",        .hexret = true,
 675           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 676         { .name     = "clock_gettime",
 677           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 678         { .name     = "clone",      .errpid = true, .nr_args = 5,
 679           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 680                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 681                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 682                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 683                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 684         { .name     = "close",
 685           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 686         { .name     = "connect",
 687           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 688         { .name     = "epoll_ctl",
 689           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 690         { .name     = "eventfd2",
 691           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 692         { .name     = "fchmodat",
 693           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 694         { .name     = "fchownat",
 695           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 696         { .name     = "fcntl",
 697           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 698                            .parm      = &strarrays__fcntl_cmds_arrays,
 699                            .show_zero = true, },
 700                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 701         { .name     = "flock",
 702           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 703         { .name     = "fstat", .alias = "newfstat", },
 704         { .name     = "fstatat", .alias = "newfstatat", },
 705         { .name     = "futex",
 706           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 707                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 708         { .name     = "futimesat",
 709           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 710         { .name     = "getitimer",
 711           .arg = { [0] = STRARRAY(which, itimers), }, },
 712         { .name     = "getpid",     .errpid = true, },
 713         { .name     = "getpgid",    .errpid = true, },
 714         { .name     = "getppid",    .errpid = true, },
 715         { .name     = "getrandom",
 716           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 717         { .name     = "getrlimit",
 718           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 719         { .name     = "gettid",     .errpid = true, },
 720         { .name     = "ioctl",
 721           .arg = {
 722 #if defined(__i386__) || defined(__x86_64__)
 723 /*
 724  * FIXME: Make this available to all arches.
 725  */
 726                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 727                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 728 #else
 729                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 730 #endif
 731         { .name     = "kcmp",       .nr_args = 5,
 732           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 733                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 734                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 735                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 736                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 737         { .name     = "keyctl",
 738           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 739         { .name     = "kill",
 740           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 741         { .name     = "linkat",
 742           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 743         { .name     = "lseek",
 744           .arg = { [2] = STRARRAY(whence, whences), }, },
 745         { .name     = "lstat", .alias = "newlstat", },
 746         { .name     = "madvise",
 747           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 748                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 749         { .name     = "mkdirat",
 750           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 751         { .name     = "mknodat",
 752           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 753         { .name     = "mmap",       .hexret = true,
 754 /* The standard mmap maps to old_mmap on s390x */
 755 #if defined(__s390x__)
 756         .alias = "old_mmap",
 757 #endif
 758           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 759                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 760                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 761         { .name     = "mount",
 762           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 763                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 764                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 765         { .name     = "mprotect",
 766           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 767                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 768         { .name     = "mq_unlink",
 769           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 770         { .name     = "mremap",     .hexret = true,
 771           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 772         { .name     = "name_to_handle_at",
 773           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 774         { .name     = "newfstatat",
 775           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 776         { .name     = "open",
 777           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 778         { .name     = "open_by_handle_at",
 779           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 780                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 781         { .name     = "openat",
 782           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 783                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 784         { .name     = "perf_event_open",
 785           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 786                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 787                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 788         { .name     = "pipe2",
 789           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 790         { .name     = "pkey_alloc",
 791           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 792         { .name     = "pkey_free",
 793           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 794         { .name     = "pkey_mprotect",
 795           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 796                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 797                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 798         { .name     = "poll", .timeout = true, },
 799         { .name     = "ppoll", .timeout = true, },
 800         { .name     = "prctl",
 801           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 802                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 803                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 804         { .name     = "pread", .alias = "pread64", },
 805         { .name     = "preadv", .alias = "pread", },
 806         { .name     = "prlimit64",
 807           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 808         { .name     = "pwrite", .alias = "pwrite64", },
 809         { .name     = "readlinkat",
 810           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 811         { .name     = "recvfrom",
 812           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 813         { .name     = "recvmmsg",
 814           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 815         { .name     = "recvmsg",
 816           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 817         { .name     = "renameat",
 818           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 819                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 820         { .name     = "renameat2",
 821           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 822                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 823                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 824         { .name     = "rt_sigaction",
 825           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 826         { .name     = "rt_sigprocmask",
 827           .arg = { [0] = STRARRAY(how, sighow), }, },
 828         { .name     = "rt_sigqueueinfo",
 829           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 830         { .name     = "rt_tgsigqueueinfo",
 831           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 832         { .name     = "sched_setscheduler",
 833           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 834         { .name     = "seccomp",
 835           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 836                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 837         { .name     = "select", .timeout = true, },
 838         { .name     = "sendmmsg",
 839           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 840         { .name     = "sendmsg",
 841           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 842         { .name     = "sendto",
 843           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 844                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 845         { .name     = "set_tid_address", .errpid = true, },
 846         { .name     = "setitimer",
 847           .arg = { [0] = STRARRAY(which, itimers), }, },
 848         { .name     = "setrlimit",
 849           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 850         { .name     = "socket",
 851           .arg = { [0] = STRARRAY(family, socket_families),
 852                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 853                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 854         { .name     = "socketpair",
 855           .arg = { [0] = STRARRAY(family, socket_families),
 856                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 857                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 858         { .name     = "stat", .alias = "newstat", },
 859         { .name     = "statx",
 860           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 861                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 862                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 863         { .name     = "swapoff",
 864           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 865         { .name     = "swapon",
 866           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 867         { .name     = "symlinkat",
 868           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 869         { .name     = "tgkill",
 870           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 871         { .name     = "tkill",
 872           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 873         { .name     = "umount2", .alias = "umount",
 874           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 875         { .name     = "uname", .alias = "newuname", },
 876         { .name     = "unlinkat",
 877           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 878         { .name     = "utimensat",
 879           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 880         { .name     = "wait4",      .errpid = true,
 881           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 882         { .name     = "waitid",     .errpid = true,
 883           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 884 };
 885
 886 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 887 {
 888         const struct syscall_fmt *fmt = fmtp;
 889         return strcmp(name, fmt->name);
 890 }
 891
 892 static struct syscall_fmt *syscall_fmt__find(const char *name)
 893 {
 894         const int nmemb = ARRAY_SIZE(syscall_fmts);
 895         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 896 }
 897
 898 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 899 {
 900         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 901
 902         for (i = 0; i < nmemb; ++i) {
 903                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 904                         return &syscall_fmts[i];
 905         }
 906
 907         return NULL;
 908 }
 909
 910 /*
 911  * is_exit: is this "exit" or "exit_group"?
 912  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 913  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 914  */
 915 struct syscall {
 916         struct tep_event    *tp_format;
 917         int                 nr_args;
 918         int                 args_size;
 919         bool                is_exit;
 920         bool                is_open;
 921         struct tep_format_field *args;
 922         const char          *name;
 923         struct syscall_fmt  *fmt;
 924         struct syscall_arg_fmt *arg_fmt;
 925 };
 926
 927 struct bpf_map_syscall_entry {
 928         bool    enabled;
 929 };
 930
 931 /*
 932  * We need to have this 'calculated' boolean because in some cases we really
 933  * don't know what is the duration of a syscall, for instance, when we start
 934  * a session and some threads are waiting for a syscall to finish, say 'poll',
 935  * in which case all we can do is to print "( ? ) for duration and for the
 936  * start timestamp.
 937  */
 938 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 939 {
 940         double duration = (double)t / NSEC_PER_MSEC;
 941         size_t printed = fprintf(fp, "(");
 942
 943         if (!calculated)
 944                 printed += fprintf(fp, "         ");
 945         else if (duration >= 1.0)
 946                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 947         else if (duration >= 0.01)
 948                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 949         else
 950                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 951         return printed + fprintf(fp, "): ");
 952 }
 953
 954 /**
 955  * filename.ptr: The filename char pointer that will be vfs_getname'd
 956  * filename.entry_str_pos: Where to insert the string translated from
 957  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 958  * ret_scnprintf: syscall args may set this to a different syscall return
 959  *                formatter, for instance, fcntl may return fds, file flags, etc.
 960  */
 961 struct thread_trace {
 962         u64               entry_time;
 963         bool              entry_pending;
 964         unsigned long     nr_events;
 965         unsigned long     pfmaj, pfmin;
 966         char              *entry_str;
 967         double            runtime_ms;
 968         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 969         struct {
 970                 unsigned long ptr;
 971                 short int     entry_str_pos;
 972                 bool          pending_open;
 973                 unsigned int  namelen;
 974                 char          *name;
 975         } filename;
 976         struct {
 977                 int       max;
 978                 char      **table;
 979         } paths;
 980
 981         struct intlist *syscall_stats;
 982 };
 983
 984 static struct thread_trace *thread_trace__new(void)
 985 {
 986         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 987
 988         if (ttrace)
 989                 ttrace->paths.max = -1;
 990
 991         ttrace->syscall_stats = intlist__new(NULL);
 992
 993         return ttrace;
 994 }
 995
 996 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 997 {
 998         struct thread_trace *ttrace;
 999
1000         if (thread == NULL)
1001                 goto fail;
1002
1003         if (thread__priv(thread) == NULL)
1004                 thread__set_priv(thread, thread_trace__new());
1005
1006         if (thread__priv(thread) == NULL)
1007                 goto fail;
1008
1009         ttrace = thread__priv(thread);
1010         ++ttrace->nr_events;
1011
1012         return ttrace;
1013 fail:
1014         color_fprintf(fp, PERF_COLOR_RED,
1015                       "WARNING: not enough memory, dropping samples!\n");
1016         return NULL;
1017 }
1018
1019
1020 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1021                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1022 {
1023         struct thread_trace *ttrace = thread__priv(arg->thread);
1024
1025         ttrace->ret_scnprintf = ret_scnprintf;
1026 }
1027
1028 #define TRACE_PFMAJ             (1 << 0)
1029 #define TRACE_PFMIN             (1 << 1)
1030
1031 static const size_t trace__entry_str_size = 2048;
1032
1033 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1034 {
1035         struct thread_trace *ttrace = thread__priv(thread);
1036
1037         if (fd > ttrace->paths.max) {
1038                 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
1039
1040                 if (npath == NULL)
1041                         return -1;
1042
1043                 if (ttrace->paths.max != -1) {
1044                         memset(npath + ttrace->paths.max + 1, 0,
1045                                (fd - ttrace->paths.max) * sizeof(char *));
1046                 } else {
1047                         memset(npath, 0, (fd + 1) * sizeof(char *));
1048                 }
1049
1050                 ttrace->paths.table = npath;
1051                 ttrace->paths.max   = fd;
1052         }
1053
1054         ttrace->paths.table[fd] = strdup(pathname);
1055
1056         return ttrace->paths.table[fd] != NULL ? 0 : -1;
1057 }
1058
1059 static int thread__read_fd_path(struct thread *thread, int fd)
1060 {
1061         char linkname[PATH_MAX], pathname[PATH_MAX];
1062         struct stat st;
1063         int ret;
1064
1065         if (thread->pid_ == thread->tid) {
1066                 scnprintf(linkname, sizeof(linkname),
1067                           "/proc/%d/fd/%d", thread->pid_, fd);
1068         } else {
1069                 scnprintf(linkname, sizeof(linkname),
1070                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1071         }
1072
1073         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1074                 return -1;
1075
1076         ret = readlink(linkname, pathname, sizeof(pathname));
1077
1078         if (ret < 0 || ret > st.st_size)
1079                 return -1;
1080
1081         pathname[ret] = '\0';
1082         return trace__set_fd_pathname(thread, fd, pathname);
1083 }
1084
1085 static const char *thread__fd_path(struct thread *thread, int fd,
1086                                    struct trace *trace)
1087 {
1088         struct thread_trace *ttrace = thread__priv(thread);
1089
1090         if (ttrace == NULL)
1091                 return NULL;
1092
1093         if (fd < 0)
1094                 return NULL;
1095
1096         if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
1097                 if (!trace->live)
1098                         return NULL;
1099                 ++trace->stats.proc_getname;
1100                 if (thread__read_fd_path(thread, fd))
1101                         return NULL;
1102         }
1103
1104         return ttrace->paths.table[fd];
1105 }
1106
1107 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1108 {
1109         int fd = arg->val;
1110         size_t printed = scnprintf(bf, size, "%d", fd);
1111         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1112
1113         if (path)
1114                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1115
1116         return printed;
1117 }
1118
1119 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1120 {
1121         size_t printed = scnprintf(bf, size, "%d", fd);
1122         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1123
1124         if (thread) {
1125                 const char *path = thread__fd_path(thread, fd, trace);
1126
1127                 if (path)
1128                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1129
1130                 thread__put(thread);
1131         }
1132
1133         return printed;
1134 }
1135
1136 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1137                                               struct syscall_arg *arg)
1138 {
1139         int fd = arg->val;
1140         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1141         struct thread_trace *ttrace = thread__priv(arg->thread);
1142
1143         if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1144                 zfree(&ttrace->paths.table[fd]);
1145
1146         return printed;
1147 }
1148
1149 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1150                                      unsigned long ptr)
1151 {
1152         struct thread_trace *ttrace = thread__priv(thread);
1153
1154         ttrace->filename.ptr = ptr;
1155         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1156 }
1157
1158 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1159 {
1160         struct augmented_arg *augmented_arg = arg->augmented.args;
1161
1162         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1163 }
1164
1165 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1166                                               struct syscall_arg *arg)
1167 {
1168         unsigned long ptr = arg->val;
1169
1170         if (arg->augmented.args)
1171                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1172
1173         if (!arg->trace->vfs_getname)
1174                 return scnprintf(bf, size, "%#x", ptr);
1175
1176         thread__set_filename_pos(arg->thread, bf, ptr);
1177         return 0;
1178 }
1179
1180 static bool trace__filter_duration(struct trace *trace, double t)
1181 {
1182         return t < (trace->duration_filter * NSEC_PER_MSEC);
1183 }
1184
1185 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1186 {
1187         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1188
1189         return fprintf(fp, "%10.3f ", ts);
1190 }
1191
1192 /*
1193  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1194  * using ttrace->entry_time for a thread that receives a sys_exit without
1195  * first having received a sys_enter ("poll" issued before tracing session
1196  * starts, lost sys_enter exit due to ring buffer overflow).
1197  */
1198 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1199 {
1200         if (tstamp > 0)
1201                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1202
1203         return fprintf(fp, "         ? ");
1204 }
1205
1206 static bool done = false;
1207 static bool interrupted = false;
1208
1209 static void sig_handler(int sig)
1210 {
1211         done = true;
1212         interrupted = sig == SIGINT;
1213 }
1214
1215 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1216 {
1217         size_t printed = 0;
1218
1219         if (trace->multiple_threads) {
1220                 if (trace->show_comm)
1221                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1222                 printed += fprintf(fp, "%d ", thread->tid);
1223         }
1224
1225         return printed;
1226 }
1227
1228 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1229                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1230 {
1231         size_t printed = 0;
1232
1233         if (trace->show_tstamp)
1234                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1235         if (trace->show_duration)
1236                 printed += fprintf_duration(duration, duration_calculated, fp);
1237         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1238 }
1239
1240 static int trace__process_event(struct trace *trace, struct machine *machine,
1241                                 union perf_event *event, struct perf_sample *sample)
1242 {
1243         int ret = 0;
1244
1245         switch (event->header.type) {
1246         case PERF_RECORD_LOST:
1247                 color_fprintf(trace->output, PERF_COLOR_RED,
1248                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1249                 ret = machine__process_lost_event(machine, event, sample);
1250                 break;
1251         default:
1252                 ret = machine__process_event(machine, event, sample);
1253                 break;
1254         }
1255
1256         return ret;
1257 }
1258
1259 static int trace__tool_process(struct perf_tool *tool,
1260                                union perf_event *event,
1261                                struct perf_sample *sample,
1262                                struct machine *machine)
1263 {
1264         struct trace *trace = container_of(tool, struct trace, tool);
1265         return trace__process_event(trace, machine, event, sample);
1266 }
1267
1268 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1269 {
1270         struct machine *machine = vmachine;
1271
1272         if (machine->kptr_restrict_warned)
1273                 return NULL;
1274
1275         if (symbol_conf.kptr_restrict) {
1276                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1277                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1278                            "Kernel samples will not be resolved.\n");
1279                 machine->kptr_restrict_warned = true;
1280                 return NULL;
1281         }
1282
1283         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1284 }
1285
1286 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1287 {
1288         int err = symbol__init(NULL);
1289
1290         if (err)
1291                 return err;
1292
1293         trace->host = machine__new_host();
1294         if (trace->host == NULL)
1295                 return -ENOMEM;
1296
1297         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1298         if (err < 0)
1299                 goto out;
1300
1301         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1302                                             evlist->threads, trace__tool_process, false,
1303                                             1);
1304 out:
1305         if (err)
1306                 symbol__exit();
1307
1308         return err;
1309 }
1310
1311 static void trace__symbols__exit(struct trace *trace)
1312 {
1313         machine__exit(trace->host);
1314         trace->host = NULL;
1315
1316         symbol__exit();
1317 }
1318
1319 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1320 {
1321         int idx;
1322
1323         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1324                 nr_args = sc->fmt->nr_args;
1325
1326         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1327         if (sc->arg_fmt == NULL)
1328                 return -1;
1329
1330         for (idx = 0; idx < nr_args; ++idx) {
1331                 if (sc->fmt)
1332                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1333         }
1334
1335         sc->nr_args = nr_args;
1336         return 0;
1337 }
1338
1339 static int syscall__set_arg_fmts(struct syscall *sc)
1340 {
1341         struct tep_format_field *field, *last_field = NULL;
1342         int idx = 0, len;
1343
1344         for (field = sc->args; field; field = field->next, ++idx) {
1345                 last_field = field;
1346
1347                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1348                         continue;
1349
1350                 if (strcmp(field->type, "const char *") == 0 &&
1351                          (strcmp(field->name, "filename") == 0 ||
1352                           strcmp(field->name, "path") == 0 ||
1353                           strcmp(field->name, "pathname") == 0))
1354                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1355                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1356                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1357                 else if (strcmp(field->type, "pid_t") == 0)
1358                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1359                 else if (strcmp(field->type, "umode_t") == 0)
1360                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1361                 else if ((strcmp(field->type, "int") == 0 ||
1362                           strcmp(field->type, "unsigned int") == 0 ||
1363                           strcmp(field->type, "long") == 0) &&
1364                          (len = strlen(field->name)) >= 2 &&
1365                          strcmp(field->name + len - 2, "fd") == 0) {
1366                         /*
1367                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1368                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1369                          * 65 int
1370                          * 23 unsigned int
1371                          * 7 unsigned long
1372                          */
1373                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1374                 }
1375         }
1376
1377         if (last_field)
1378                 sc->args_size = last_field->offset + last_field->size;
1379
1380         return 0;
1381 }
1382
1383 static int trace__read_syscall_info(struct trace *trace, int id)
1384 {
1385         char tp_name[128];
1386         struct syscall *sc;
1387         const char *name = syscalltbl__name(trace->sctbl, id);
1388
1389         if (name == NULL)
1390                 return -1;
1391
1392         if (id > trace->syscalls.max) {
1393                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1394
1395                 if (nsyscalls == NULL)
1396                         return -1;
1397
1398                 if (trace->syscalls.max != -1) {
1399                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1400                                (id - trace->syscalls.max) * sizeof(*sc));
1401                 } else {
1402                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1403                 }
1404
1405                 trace->syscalls.table = nsyscalls;
1406                 trace->syscalls.max   = id;
1407         }
1408
1409         sc = trace->syscalls.table + id;
1410         sc->name = name;
1411
1412         sc->fmt  = syscall_fmt__find(sc->name);
1413
1414         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1415         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1416
1417         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1418                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1419                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1420         }
1421
1422         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1423                 return -1;
1424
1425         if (IS_ERR(sc->tp_format))
1426                 return -1;
1427
1428         sc->args = sc->tp_format->format.fields;
1429         /*
1430          * We need to check and discard the first variable '__syscall_nr'
1431          * or 'nr' that mean the syscall number. It is needless here.
1432          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1433          */
1434         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1435                 sc->args = sc->args->next;
1436                 --sc->nr_args;
1437         }
1438
1439         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1440         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1441
1442         return syscall__set_arg_fmts(sc);
1443 }
1444
1445 static int trace__validate_ev_qualifier(struct trace *trace)
1446 {
1447         int err = 0, i;
1448         size_t nr_allocated;
1449         struct str_node *pos;
1450
1451         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1452         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1453                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1454
1455         if (trace->ev_qualifier_ids.entries == NULL) {
1456                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1457                        trace->output);
1458                 err = -EINVAL;
1459                 goto out;
1460         }
1461
1462         nr_allocated = trace->ev_qualifier_ids.nr;
1463         i = 0;
1464
1465         strlist__for_each_entry(pos, trace->ev_qualifier) {
1466                 const char *sc = pos->s;
1467                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1468
1469                 if (id < 0) {
1470                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1471                         if (id >= 0)
1472                                 goto matches;
1473
1474                         if (err == 0) {
1475                                 fputs("Error:\tInvalid syscall ", trace->output);
1476                                 err = -EINVAL;
1477                         } else {
1478                                 fputs(", ", trace->output);
1479                         }
1480
1481                         fputs(sc, trace->output);
1482                 }
1483 matches:
1484                 trace->ev_qualifier_ids.entries[i++] = id;
1485                 if (match_next == -1)
1486                         continue;
1487
1488                 while (1) {
1489                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1490                         if (id < 0)
1491                                 break;
1492                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1493                                 void *entries;
1494
1495                                 nr_allocated += 8;
1496                                 entries = realloc(trace->ev_qualifier_ids.entries,
1497                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1498                                 if (entries == NULL) {
1499                                         err = -ENOMEM;
1500                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1501                                         goto out_free;
1502                                 }
1503                                 trace->ev_qualifier_ids.entries = entries;
1504                         }
1505                         trace->ev_qualifier_ids.nr++;
1506                         trace->ev_qualifier_ids.entries[i++] = id;
1507                 }
1508         }
1509
1510         if (err < 0) {
1511                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1512                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1513 out_free:
1514                 zfree(&trace->ev_qualifier_ids.entries);
1515                 trace->ev_qualifier_ids.nr = 0;
1516         }
1517 out:
1518         return err;
1519 }
1520
1521 /*
1522  * args is to be interpreted as a series of longs but we need to handle
1523  * 8-byte unaligned accesses. args points to raw_data within the event
1524  * and raw_data is guaranteed to be 8-byte unaligned because it is
1525  * preceded by raw_size which is a u32. So we need to copy args to a temp
1526  * variable to read it. Most notably this avoids extended load instructions
1527  * on unaligned addresses
1528  */
1529 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1530 {
1531         unsigned long val;
1532         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1533
1534         memcpy(&val, p, sizeof(val));
1535         return val;
1536 }
1537
1538 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1539                                       struct syscall_arg *arg)
1540 {
1541         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1542                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1543
1544         return scnprintf(bf, size, "arg%d: ", arg->idx);
1545 }
1546
1547 /*
1548  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1549  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1550  * in tools/perf/trace/beauty/mount_flags.c
1551  */
1552 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1553 {
1554         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1555                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1556
1557         return val;
1558 }
1559
1560 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1561                                      struct syscall_arg *arg, unsigned long val)
1562 {
1563         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1564                 arg->val = val;
1565                 if (sc->arg_fmt[arg->idx].parm)
1566                         arg->parm = sc->arg_fmt[arg->idx].parm;
1567                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1568         }
1569         return scnprintf(bf, size, "%ld", val);
1570 }
1571
1572 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1573                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1574                                       struct trace *trace, struct thread *thread)
1575 {
1576         size_t printed = 0;
1577         unsigned long val;
1578         u8 bit = 1;
1579         struct syscall_arg arg = {
1580                 .args   = args,
1581                 .augmented = {
1582                         .size = augmented_args_size,
1583                         .args = augmented_args,
1584                 },
1585                 .idx    = 0,
1586                 .mask   = 0,
1587                 .trace  = trace,
1588                 .thread = thread,
1589                 .show_string_prefix = trace->show_string_prefix,
1590         };
1591         struct thread_trace *ttrace = thread__priv(thread);
1592
1593         /*
1594          * Things like fcntl will set this in its 'cmd' formatter to pick the
1595          * right formatter for the return value (an fd? file flags?), which is
1596          * not needed for syscalls that always return a given type, say an fd.
1597          */
1598         ttrace->ret_scnprintf = NULL;
1599
1600         if (sc->args != NULL) {
1601                 struct tep_format_field *field;
1602
1603                 for (field = sc->args; field;
1604                      field = field->next, ++arg.idx, bit <<= 1) {
1605                         if (arg.mask & bit)
1606                                 continue;
1607
1608                         val = syscall_arg__val(&arg, arg.idx);
1609                         /*
1610                          * Some syscall args need some mask, most don't and
1611                          * return val untouched.
1612                          */
1613                         val = syscall__mask_val(sc, &arg, val);
1614
1615                         /*
1616                          * Suppress this argument if its value is zero and
1617                          * and we don't have a string associated in an
1618                          * strarray for it.
1619                          */
1620                         if (val == 0 &&
1621                             !trace->show_zeros &&
1622                             !(sc->arg_fmt &&
1623                               (sc->arg_fmt[arg.idx].show_zero ||
1624                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1625                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1626                               sc->arg_fmt[arg.idx].parm))
1627                                 continue;
1628
1629                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1630
1631                         if (trace->show_arg_names)
1632                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1633
1634                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1635                 }
1636         } else if (IS_ERR(sc->tp_format)) {
1637                 /*
1638                  * If we managed to read the tracepoint /format file, then we
1639                  * may end up not having any args, like with gettid(), so only
1640                  * print the raw args when we didn't manage to read it.
1641                  */
1642                 while (arg.idx < sc->nr_args) {
1643                         if (arg.mask & bit)
1644                                 goto next_arg;
1645                         val = syscall_arg__val(&arg, arg.idx);
1646                         if (printed)
1647                                 printed += scnprintf(bf + printed, size - printed, ", ");
1648                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1649                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1650 next_arg:
1651                         ++arg.idx;
1652                         bit <<= 1;
1653                 }
1654         }
1655
1656         return printed;
1657 }
1658
1659 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1660                                   union perf_event *event,
1661                                   struct perf_sample *sample);
1662
1663 static struct syscall *trace__syscall_info(struct trace *trace,
1664                                            struct perf_evsel *evsel, int id)
1665 {
1666
1667         if (id < 0) {
1668
1669                 /*
1670                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1671                  * before that, leaving at a higher verbosity level till that is
1672                  * explained. Reproduced with plain ftrace with:
1673                  *
1674                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1675                  * grep "NR -1 " /t/trace_pipe
1676                  *
1677                  * After generating some load on the machine.
1678                  */
1679                 if (verbose > 1) {
1680                         static u64 n;
1681                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1682                                 id, perf_evsel__name(evsel), ++n);
1683                 }
1684                 return NULL;
1685         }
1686
1687         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1688             trace__read_syscall_info(trace, id))
1689                 goto out_cant_read;
1690
1691         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1692                 goto out_cant_read;
1693
1694         return &trace->syscalls.table[id];
1695
1696 out_cant_read:
1697         if (verbose > 0) {
1698                 fprintf(trace->output, "Problems reading syscall %d", id);
1699                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1700                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1701                 fputs(" information\n", trace->output);
1702         }
1703         return NULL;
1704 }
1705
1706 static void thread__update_stats(struct thread_trace *ttrace,
1707                                  int id, struct perf_sample *sample)
1708 {
1709         struct int_node *inode;
1710         struct stats *stats;
1711         u64 duration = 0;
1712
1713         inode = intlist__findnew(ttrace->syscall_stats, id);
1714         if (inode == NULL)
1715                 return;
1716
1717         stats = inode->priv;
1718         if (stats == NULL) {
1719                 stats = malloc(sizeof(struct stats));
1720                 if (stats == NULL)
1721                         return;
1722                 init_stats(stats);
1723                 inode->priv = stats;
1724         }
1725
1726         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1727                 duration = sample->time - ttrace->entry_time;
1728
1729         update_stats(stats, duration);
1730 }
1731
1732 static int trace__printf_interrupted_entry(struct trace *trace)
1733 {
1734         struct thread_trace *ttrace;
1735         size_t printed;
1736
1737         if (trace->failure_only || trace->current == NULL)
1738                 return 0;
1739
1740         ttrace = thread__priv(trace->current);
1741
1742         if (!ttrace->entry_pending)
1743                 return 0;
1744
1745         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1746         printed += fprintf(trace->output, ")%-*s ...\n", trace->args_alignment, ttrace->entry_str);
1747         ttrace->entry_pending = false;
1748
1749         ++trace->nr_events_printed;
1750
1751         return printed;
1752 }
1753
1754 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1755                                  struct perf_sample *sample, struct thread *thread)
1756 {
1757         int printed = 0;
1758
1759         if (trace->print_sample) {
1760                 double ts = (double)sample->time / NSEC_PER_MSEC;
1761
1762                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1763                                    perf_evsel__name(evsel), ts,
1764                                    thread__comm_str(thread),
1765                                    sample->pid, sample->tid, sample->cpu);
1766         }
1767
1768         return printed;
1769 }
1770
1771 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, bool raw_augmented)
1772 {
1773         void *augmented_args = NULL;
1774         /*
1775          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1776          * and there we get all 6 syscall args plus the tracepoint common
1777          * fields (sizeof(long)) and the syscall_nr (another long). So we check
1778          * if that is the case and if so don't look after the sc->args_size,
1779          * but always after the full raw_syscalls:sys_enter payload, which is
1780          * fixed.
1781          *
1782          * We'll revisit this later to pass s->args_size to the BPF augmenter
1783          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1784          * copies only what we need for each syscall, like what happens when we
1785          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1786          * traffic to just what is needed for each syscall.
1787          */
1788         int args_size = raw_augmented ? (8 * (int)sizeof(long)) : sc->args_size;
1789
1790         *augmented_args_size = sample->raw_size - args_size;
1791         if (*augmented_args_size > 0)
1792                 augmented_args = sample->raw_data + args_size;
1793
1794         return augmented_args;
1795 }
1796
1797 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1798                             union perf_event *event __maybe_unused,
1799                             struct perf_sample *sample)
1800 {
1801         char *msg;
1802         void *args;
1803         int printed = 0;
1804         struct thread *thread;
1805         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1806         int augmented_args_size = 0;
1807         void *augmented_args = NULL;
1808         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1809         struct thread_trace *ttrace;
1810
1811         if (sc == NULL)
1812                 return -1;
1813
1814         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1815         ttrace = thread__trace(thread, trace->output);
1816         if (ttrace == NULL)
1817                 goto out_put;
1818
1819         trace__fprintf_sample(trace, evsel, sample, thread);
1820
1821         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1822
1823         if (ttrace->entry_str == NULL) {
1824                 ttrace->entry_str = malloc(trace__entry_str_size);
1825                 if (!ttrace->entry_str)
1826                         goto out_put;
1827         }
1828
1829         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1830                 trace__printf_interrupted_entry(trace);
1831         /*
1832          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1833          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1834          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1835          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1836          * so when handling, say the openat syscall, we end up getting 6 args for the
1837          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1838          * thinking that the extra 2 u64 args are the augmented filename, so just check
1839          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1840          */
1841         if (evsel != trace->syscalls.events.sys_enter)
1842                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls);
1843         ttrace->entry_time = sample->time;
1844         msg = ttrace->entry_str;
1845         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1846
1847         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1848                                            args, augmented_args, augmented_args_size, trace, thread);
1849
1850         if (sc->is_exit) {
1851                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1852                         int alignment = 0;
1853
1854                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1855                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1856                         if (trace->args_alignment > printed)
1857                                 alignment = trace->args_alignment - printed;
1858                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1859                 }
1860         } else {
1861                 ttrace->entry_pending = true;
1862                 /* See trace__vfs_getname & trace__sys_exit */
1863                 ttrace->filename.pending_open = false;
1864         }
1865
1866         if (trace->current != thread) {
1867                 thread__put(trace->current);
1868                 trace->current = thread__get(thread);
1869         }
1870         err = 0;
1871 out_put:
1872         thread__put(thread);
1873         return err;
1874 }
1875
1876 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1877                                     struct perf_sample *sample)
1878 {
1879         struct thread_trace *ttrace;
1880         struct thread *thread;
1881         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1882         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1883         char msg[1024];
1884         void *args, *augmented_args = NULL;
1885         int augmented_args_size;
1886
1887         if (sc == NULL)
1888                 return -1;
1889
1890         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1891         ttrace = thread__trace(thread, trace->output);
1892         /*
1893          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1894          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1895          */
1896         if (ttrace == NULL)
1897                 goto out_put;
1898
1899         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1900         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls);
1901         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1902         fprintf(trace->output, "%s", msg);
1903         err = 0;
1904 out_put:
1905         thread__put(thread);
1906         return err;
1907 }
1908
1909 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1910                                     struct perf_sample *sample,
1911                                     struct callchain_cursor *cursor)
1912 {
1913         struct addr_location al;
1914         int max_stack = evsel->attr.sample_max_stack ?
1915                         evsel->attr.sample_max_stack :
1916                         trace->max_stack;
1917         int err;
1918
1919         if (machine__resolve(trace->host, &al, sample) < 0)
1920                 return -1;
1921
1922         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1923         addr_location__put(&al);
1924         return err;
1925 }
1926
1927 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1928 {
1929         /* TODO: user-configurable print_opts */
1930         const unsigned int print_opts = EVSEL__PRINT_SYM |
1931                                         EVSEL__PRINT_DSO |
1932                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1933
1934         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1935 }
1936
1937 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1938 {
1939         struct perf_env *env = perf_evsel__env(evsel);
1940         const char *arch_name = perf_env__arch(env);
1941
1942         return arch_syscalls__strerrno(arch_name, err);
1943 }
1944
1945 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1946                            union perf_event *event __maybe_unused,
1947                            struct perf_sample *sample)
1948 {
1949         long ret;
1950         u64 duration = 0;
1951         bool duration_calculated = false;
1952         struct thread *thread;
1953         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
1954         int alignment = trace->args_alignment;
1955         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1956         struct thread_trace *ttrace;
1957
1958         if (sc == NULL)
1959                 return -1;
1960
1961         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1962         ttrace = thread__trace(thread, trace->output);
1963         if (ttrace == NULL)
1964                 goto out_put;
1965
1966         trace__fprintf_sample(trace, evsel, sample, thread);
1967
1968         if (trace->summary)
1969                 thread__update_stats(ttrace, id, sample);
1970
1971         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1972
1973         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
1974                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1975                 ttrace->filename.pending_open = false;
1976                 ++trace->stats.vfs_getname;
1977         }
1978
1979         if (ttrace->entry_time) {
1980                 duration = sample->time - ttrace->entry_time;
1981                 if (trace__filter_duration(trace, duration))
1982                         goto out;
1983                 duration_calculated = true;
1984         } else if (trace->duration_filter)
1985                 goto out;
1986
1987         if (sample->callchain) {
1988                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1989                 if (callchain_ret == 0) {
1990                         if (callchain_cursor.nr < trace->min_stack)
1991                                 goto out;
1992                         callchain_ret = 1;
1993                 }
1994         }
1995
1996         if (trace->summary_only || (ret >= 0 && trace->failure_only))
1997                 goto out;
1998
1999         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2000
2001         if (ttrace->entry_pending) {
2002                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2003         } else {
2004                 fprintf(trace->output, " ... [");
2005                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2006                 fprintf(trace->output, "]: %s()", sc->name);
2007         }
2008
2009         printed++; /* the closing ')' */
2010
2011         if (alignment > printed)
2012                 alignment -= printed;
2013         else
2014                 alignment = 0;
2015
2016         fprintf(trace->output, ")%*s= ", alignment, " ");
2017
2018         if (sc->fmt == NULL) {
2019                 if (ret < 0)
2020                         goto errno_print;
2021 signed_print:
2022                 fprintf(trace->output, "%ld", ret);
2023         } else if (ret < 0) {
2024 errno_print: {
2025                 char bf[STRERR_BUFSIZE];
2026                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2027                            *e = errno_to_name(evsel, -ret);
2028
2029                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2030         }
2031         } else if (ret == 0 && sc->fmt->timeout)
2032                 fprintf(trace->output, "0 (Timeout)");
2033         else if (ttrace->ret_scnprintf) {
2034                 char bf[1024];
2035                 struct syscall_arg arg = {
2036                         .val    = ret,
2037                         .thread = thread,
2038                         .trace  = trace,
2039                 };
2040                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2041                 ttrace->ret_scnprintf = NULL;
2042                 fprintf(trace->output, "%s", bf);
2043         } else if (sc->fmt->hexret)
2044                 fprintf(trace->output, "%#lx", ret);
2045         else if (sc->fmt->errpid) {
2046                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2047
2048                 if (child != NULL) {
2049                         fprintf(trace->output, "%ld", ret);
2050                         if (child->comm_set)
2051                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2052                         thread__put(child);
2053                 }
2054         } else
2055                 goto signed_print;
2056
2057         fputc('\n', trace->output);
2058
2059         /*
2060          * We only consider an 'event' for the sake of --max-events a non-filtered
2061          * sys_enter + sys_exit and other tracepoint events.
2062          */
2063         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2064                 interrupted = true;
2065
2066         if (callchain_ret > 0)
2067                 trace__fprintf_callchain(trace, sample);
2068         else if (callchain_ret < 0)
2069                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2070 out:
2071         ttrace->entry_pending = false;
2072         err = 0;
2073 out_put:
2074         thread__put(thread);
2075         return err;
2076 }
2077
2078 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2079                               union perf_event *event __maybe_unused,
2080                               struct perf_sample *sample)
2081 {
2082         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2083         struct thread_trace *ttrace;
2084         size_t filename_len, entry_str_len, to_move;
2085         ssize_t remaining_space;
2086         char *pos;
2087         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2088
2089         if (!thread)
2090                 goto out;
2091
2092         ttrace = thread__priv(thread);
2093         if (!ttrace)
2094                 goto out_put;
2095
2096         filename_len = strlen(filename);
2097         if (filename_len == 0)
2098                 goto out_put;
2099
2100         if (ttrace->filename.namelen < filename_len) {
2101                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2102
2103                 if (f == NULL)
2104                         goto out_put;
2105
2106                 ttrace->filename.namelen = filename_len;
2107                 ttrace->filename.name = f;
2108         }
2109
2110         strcpy(ttrace->filename.name, filename);
2111         ttrace->filename.pending_open = true;
2112
2113         if (!ttrace->filename.ptr)
2114                 goto out_put;
2115
2116         entry_str_len = strlen(ttrace->entry_str);
2117         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2118         if (remaining_space <= 0)
2119                 goto out_put;
2120
2121         if (filename_len > (size_t)remaining_space) {
2122                 filename += filename_len - remaining_space;
2123                 filename_len = remaining_space;
2124         }
2125
2126         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2127         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2128         memmove(pos + filename_len, pos, to_move);
2129         memcpy(pos, filename, filename_len);
2130
2131         ttrace->filename.ptr = 0;
2132         ttrace->filename.entry_str_pos = 0;
2133 out_put:
2134         thread__put(thread);
2135 out:
2136         return 0;
2137 }
2138
2139 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2140                                      union perf_event *event __maybe_unused,
2141                                      struct perf_sample *sample)
2142 {
2143         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2144         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2145         struct thread *thread = machine__findnew_thread(trace->host,
2146                                                         sample->pid,
2147                                                         sample->tid);
2148         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2149
2150         if (ttrace == NULL)
2151                 goto out_dump;
2152
2153         ttrace->runtime_ms += runtime_ms;
2154         trace->runtime_ms += runtime_ms;
2155 out_put:
2156         thread__put(thread);
2157         return 0;
2158
2159 out_dump:
2160         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2161                evsel->name,
2162                perf_evsel__strval(evsel, sample, "comm"),
2163                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2164                runtime,
2165                perf_evsel__intval(evsel, sample, "vruntime"));
2166         goto out_put;
2167 }
2168
2169 static int bpf_output__printer(enum binary_printer_ops op,
2170                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2171 {
2172         unsigned char ch = (unsigned char)val;
2173
2174         switch (op) {
2175         case BINARY_PRINT_CHAR_DATA:
2176                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2177         case BINARY_PRINT_DATA_BEGIN:
2178         case BINARY_PRINT_LINE_BEGIN:
2179         case BINARY_PRINT_ADDR:
2180         case BINARY_PRINT_NUM_DATA:
2181         case BINARY_PRINT_NUM_PAD:
2182         case BINARY_PRINT_SEP:
2183         case BINARY_PRINT_CHAR_PAD:
2184         case BINARY_PRINT_LINE_END:
2185         case BINARY_PRINT_DATA_END:
2186         default:
2187                 break;
2188         }
2189
2190         return 0;
2191 }
2192
2193 static void bpf_output__fprintf(struct trace *trace,
2194                                 struct perf_sample *sample)
2195 {
2196         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2197                         bpf_output__printer, NULL, trace->output);
2198         ++trace->nr_events_printed;
2199 }
2200
2201 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2202                                 union perf_event *event __maybe_unused,
2203                                 struct perf_sample *sample)
2204 {
2205         struct thread *thread;
2206         int callchain_ret = 0;
2207         /*
2208          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2209          * this event's max_events having been hit and this is an entry coming
2210          * from the ring buffer that we should discard, since the max events
2211          * have already been considered/printed.
2212          */
2213         if (evsel->disabled)
2214                 return 0;
2215
2216         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2217
2218         if (sample->callchain) {
2219                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2220                 if (callchain_ret == 0) {
2221                         if (callchain_cursor.nr < trace->min_stack)
2222                                 goto out;
2223                         callchain_ret = 1;
2224                 }
2225         }
2226
2227         trace__printf_interrupted_entry(trace);
2228         trace__fprintf_tstamp(trace, sample->time, trace->output);
2229
2230         if (trace->trace_syscalls && trace->show_duration)
2231                 fprintf(trace->output, "(         ): ");
2232
2233         if (thread)
2234                 trace__fprintf_comm_tid(trace, thread, trace->output);
2235
2236         if (evsel == trace->syscalls.events.augmented) {
2237                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2238                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2239
2240                 if (sc) {
2241                         fprintf(trace->output, "%s(", sc->name);
2242                         trace__fprintf_sys_enter(trace, evsel, sample);
2243                         fputc(')', trace->output);
2244                         goto newline;
2245                 }
2246
2247                 /*
2248                  * XXX: Not having the associated syscall info or not finding/adding
2249                  *      the thread should never happen, but if it does...
2250                  *      fall thru and print it as a bpf_output event.
2251                  */
2252         }
2253
2254         fprintf(trace->output, "%s:", evsel->name);
2255
2256         if (perf_evsel__is_bpf_output(evsel)) {
2257                 bpf_output__fprintf(trace, sample);
2258         } else if (evsel->tp_format) {
2259                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2260                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2261                         event_format__fprintf(evsel->tp_format, sample->cpu,
2262                                               sample->raw_data, sample->raw_size,
2263                                               trace->output);
2264                         ++trace->nr_events_printed;
2265
2266                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2267                                 perf_evsel__disable(evsel);
2268                                 perf_evsel__close(evsel);
2269                         }
2270                 }
2271         }
2272
2273 newline:
2274         fprintf(trace->output, "\n");
2275
2276         if (callchain_ret > 0)
2277                 trace__fprintf_callchain(trace, sample);
2278         else if (callchain_ret < 0)
2279                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2280 out:
2281         thread__put(thread);
2282         return 0;
2283 }
2284
2285 static void print_location(FILE *f, struct perf_sample *sample,
2286                            struct addr_location *al,
2287                            bool print_dso, bool print_sym)
2288 {
2289
2290         if ((verbose > 0 || print_dso) && al->map)
2291                 fprintf(f, "%s@", al->map->dso->long_name);
2292
2293         if ((verbose > 0 || print_sym) && al->sym)
2294                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2295                         al->addr - al->sym->start);
2296         else if (al->map)
2297                 fprintf(f, "0x%" PRIx64, al->addr);
2298         else
2299                 fprintf(f, "0x%" PRIx64, sample->addr);
2300 }
2301
2302 static int trace__pgfault(struct trace *trace,
2303                           struct perf_evsel *evsel,
2304                           union perf_event *event __maybe_unused,
2305                           struct perf_sample *sample)
2306 {
2307         struct thread *thread;
2308         struct addr_location al;
2309         char map_type = 'd';
2310         struct thread_trace *ttrace;
2311         int err = -1;
2312         int callchain_ret = 0;
2313
2314         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2315
2316         if (sample->callchain) {
2317                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2318                 if (callchain_ret == 0) {
2319                         if (callchain_cursor.nr < trace->min_stack)
2320                                 goto out_put;
2321                         callchain_ret = 1;
2322                 }
2323         }
2324
2325         ttrace = thread__trace(thread, trace->output);
2326         if (ttrace == NULL)
2327                 goto out_put;
2328
2329         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2330                 ttrace->pfmaj++;
2331         else
2332                 ttrace->pfmin++;
2333
2334         if (trace->summary_only)
2335                 goto out;
2336
2337         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2338
2339         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2340
2341         fprintf(trace->output, "%sfault [",
2342                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2343                 "maj" : "min");
2344
2345         print_location(trace->output, sample, &al, false, true);
2346
2347         fprintf(trace->output, "] => ");
2348
2349         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2350
2351         if (!al.map) {
2352                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2353
2354                 if (al.map)
2355                         map_type = 'x';
2356                 else
2357                         map_type = '?';
2358         }
2359
2360         print_location(trace->output, sample, &al, true, false);
2361
2362         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2363
2364         if (callchain_ret > 0)
2365                 trace__fprintf_callchain(trace, sample);
2366         else if (callchain_ret < 0)
2367                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2368
2369         ++trace->nr_events_printed;
2370 out:
2371         err = 0;
2372 out_put:
2373         thread__put(thread);
2374         return err;
2375 }
2376
2377 static void trace__set_base_time(struct trace *trace,
2378                                  struct perf_evsel *evsel,
2379                                  struct perf_sample *sample)
2380 {
2381         /*
2382          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2383          * and don't use sample->time unconditionally, we may end up having
2384          * some other event in the future without PERF_SAMPLE_TIME for good
2385          * reason, i.e. we may not be interested in its timestamps, just in
2386          * it taking place, picking some piece of information when it
2387          * appears in our event stream (vfs_getname comes to mind).
2388          */
2389         if (trace->base_time == 0 && !trace->full_time &&
2390             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2391                 trace->base_time = sample->time;
2392 }
2393
2394 static int trace__process_sample(struct perf_tool *tool,
2395                                  union perf_event *event,
2396                                  struct perf_sample *sample,
2397                                  struct perf_evsel *evsel,
2398                                  struct machine *machine __maybe_unused)
2399 {
2400         struct trace *trace = container_of(tool, struct trace, tool);
2401         struct thread *thread;
2402         int err = 0;
2403
2404         tracepoint_handler handler = evsel->handler;
2405
2406         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2407         if (thread && thread__is_filtered(thread))
2408                 goto out;
2409
2410         trace__set_base_time(trace, evsel, sample);
2411
2412         if (handler) {
2413                 ++trace->nr_events;
2414                 handler(trace, evsel, event, sample);
2415         }
2416 out:
2417         thread__put(thread);
2418         return err;
2419 }
2420
2421 static int trace__record(struct trace *trace, int argc, const char **argv)
2422 {
2423         unsigned int rec_argc, i, j;
2424         const char **rec_argv;
2425         const char * const record_args[] = {
2426                 "record",
2427                 "-R",
2428                 "-m", "1024",
2429                 "-c", "1",
2430         };
2431
2432         const char * const sc_args[] = { "-e", };
2433         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2434         const char * const majpf_args[] = { "-e", "major-faults" };
2435         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2436         const char * const minpf_args[] = { "-e", "minor-faults" };
2437         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2438
2439         /* +1 is for the event string below */
2440         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2441                 majpf_args_nr + minpf_args_nr + argc;
2442         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2443
2444         if (rec_argv == NULL)
2445                 return -ENOMEM;
2446
2447         j = 0;
2448         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2449                 rec_argv[j++] = record_args[i];
2450
2451         if (trace->trace_syscalls) {
2452                 for (i = 0; i < sc_args_nr; i++)
2453                         rec_argv[j++] = sc_args[i];
2454
2455                 /* event string may be different for older kernels - e.g., RHEL6 */
2456                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2457                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2458                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2459                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2460                 else {
2461                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2462                         free(rec_argv);
2463                         return -1;
2464                 }
2465         }
2466
2467         if (trace->trace_pgfaults & TRACE_PFMAJ)
2468                 for (i = 0; i < majpf_args_nr; i++)
2469                         rec_argv[j++] = majpf_args[i];
2470
2471         if (trace->trace_pgfaults & TRACE_PFMIN)
2472                 for (i = 0; i < minpf_args_nr; i++)
2473                         rec_argv[j++] = minpf_args[i];
2474
2475         for (i = 0; i < (unsigned int)argc; i++)
2476                 rec_argv[j++] = argv[i];
2477
2478         return cmd_record(j, rec_argv);
2479 }
2480
2481 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2482
2483 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2484 {
2485         struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2486
2487         if (IS_ERR(evsel))
2488                 return false;
2489
2490         if (perf_evsel__field(evsel, "pathname") == NULL) {
2491                 perf_evsel__delete(evsel);
2492                 return false;
2493         }
2494
2495         evsel->handler = trace__vfs_getname;
2496         perf_evlist__add(evlist, evsel);
2497         return true;
2498 }
2499
2500 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2501 {
2502         struct perf_evsel *evsel;
2503         struct perf_event_attr attr = {
2504                 .type = PERF_TYPE_SOFTWARE,
2505                 .mmap_data = 1,
2506         };
2507
2508         attr.config = config;
2509         attr.sample_period = 1;
2510
2511         event_attr_init(&attr);
2512
2513         evsel = perf_evsel__new(&attr);
2514         if (evsel)
2515                 evsel->handler = trace__pgfault;
2516
2517         return evsel;
2518 }
2519
2520 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2521 {
2522         const u32 type = event->header.type;
2523         struct perf_evsel *evsel;
2524
2525         if (type != PERF_RECORD_SAMPLE) {
2526                 trace__process_event(trace, trace->host, event, sample);
2527                 return;
2528         }
2529
2530         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2531         if (evsel == NULL) {
2532                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2533                 return;
2534         }
2535
2536         trace__set_base_time(trace, evsel, sample);
2537
2538         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2539             sample->raw_data == NULL) {
2540                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2541                        perf_evsel__name(evsel), sample->tid,
2542                        sample->cpu, sample->raw_size);
2543         } else {
2544                 tracepoint_handler handler = evsel->handler;
2545                 handler(trace, evsel, event, sample);
2546         }
2547
2548         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2549                 interrupted = true;
2550 }
2551
2552 static int trace__add_syscall_newtp(struct trace *trace)
2553 {
2554         int ret = -1;
2555         struct perf_evlist *evlist = trace->evlist;
2556         struct perf_evsel *sys_enter, *sys_exit;
2557
2558         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2559         if (sys_enter == NULL)
2560                 goto out;
2561
2562         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2563                 goto out_delete_sys_enter;
2564
2565         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2566         if (sys_exit == NULL)
2567                 goto out_delete_sys_enter;
2568
2569         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2570                 goto out_delete_sys_exit;
2571
2572         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2573         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2574
2575         perf_evlist__add(evlist, sys_enter);
2576         perf_evlist__add(evlist, sys_exit);
2577
2578         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2579                 /*
2580                  * We're interested only in the user space callchain
2581                  * leading to the syscall, allow overriding that for
2582                  * debugging reasons using --kernel_syscall_callchains
2583                  */
2584                 sys_exit->attr.exclude_callchain_kernel = 1;
2585         }
2586
2587         trace->syscalls.events.sys_enter = sys_enter;
2588         trace->syscalls.events.sys_exit  = sys_exit;
2589
2590         ret = 0;
2591 out:
2592         return ret;
2593
2594 out_delete_sys_exit:
2595         perf_evsel__delete_priv(sys_exit);
2596 out_delete_sys_enter:
2597         perf_evsel__delete_priv(sys_enter);
2598         goto out;
2599 }
2600
2601 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2602 {
2603         int err = -1;
2604         struct perf_evsel *sys_exit;
2605         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2606                                                 trace->ev_qualifier_ids.nr,
2607                                                 trace->ev_qualifier_ids.entries);
2608
2609         if (filter == NULL)
2610                 goto out_enomem;
2611
2612         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2613                                           filter)) {
2614                 sys_exit = trace->syscalls.events.sys_exit;
2615                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2616         }
2617
2618         free(filter);
2619 out:
2620         return err;
2621 out_enomem:
2622         errno = ENOMEM;
2623         goto out;
2624 }
2625
2626 #ifdef HAVE_LIBBPF_SUPPORT
2627 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2628 {
2629         int fd = bpf_map__fd(trace->syscalls.map);
2630         struct bpf_map_syscall_entry value = {
2631                 .enabled = !trace->not_ev_qualifier,
2632         };
2633         int err = 0;
2634         size_t i;
2635
2636         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2637                 int key = trace->ev_qualifier_ids.entries[i];
2638
2639                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2640                 if (err)
2641                         break;
2642         }
2643
2644         return err;
2645 }
2646
2647 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2648 {
2649         int fd = bpf_map__fd(trace->syscalls.map);
2650         struct bpf_map_syscall_entry value = {
2651                 .enabled = enabled,
2652         };
2653         int err = 0, key;
2654
2655         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2656                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2657                 if (err)
2658                         break;
2659         }
2660
2661         return err;
2662 }
2663
2664 static int trace__init_syscalls_bpf_map(struct trace *trace)
2665 {
2666         bool enabled = true;
2667
2668         if (trace->ev_qualifier_ids.nr)
2669                 enabled = trace->not_ev_qualifier;
2670
2671         return __trace__init_syscalls_bpf_map(trace, enabled);
2672 }
2673 #else
2674 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2675 {
2676         return 0;
2677 }
2678
2679 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2680 {
2681         return 0;
2682 }
2683 #endif // HAVE_LIBBPF_SUPPORT
2684
2685 static int trace__set_ev_qualifier_filter(struct trace *trace)
2686 {
2687         if (trace->syscalls.map)
2688                 return trace__set_ev_qualifier_bpf_filter(trace);
2689         return trace__set_ev_qualifier_tp_filter(trace);
2690 }
2691
2692 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2693                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2694 {
2695         int err = 0;
2696 #ifdef HAVE_LIBBPF_SUPPORT
2697         bool value = true;
2698         int map_fd = bpf_map__fd(map);
2699         size_t i;
2700
2701         for (i = 0; i < npids; ++i) {
2702                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2703                 if (err)
2704                         break;
2705         }
2706 #endif
2707         return err;
2708 }
2709
2710 static int trace__set_filter_loop_pids(struct trace *trace)
2711 {
2712         unsigned int nr = 1, err;
2713         pid_t pids[32] = {
2714                 getpid(),
2715         };
2716         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2717
2718         while (thread && nr < ARRAY_SIZE(pids)) {
2719                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2720
2721                 if (parent == NULL)
2722                         break;
2723
2724                 if (!strcmp(thread__comm_str(parent), "sshd")) {
2725                         pids[nr++] = parent->tid;
2726                         break;
2727                 }
2728                 thread = parent;
2729         }
2730
2731         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2732         if (!err && trace->filter_pids.map)
2733                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2734
2735         return err;
2736 }
2737
2738 static int trace__set_filter_pids(struct trace *trace)
2739 {
2740         int err = 0;
2741         /*
2742          * Better not use !target__has_task() here because we need to cover the
2743          * case where no threads were specified in the command line, but a
2744          * workload was, and in that case we will fill in the thread_map when
2745          * we fork the workload in perf_evlist__prepare_workload.
2746          */
2747         if (trace->filter_pids.nr > 0) {
2748                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2749                                                       trace->filter_pids.entries);
2750                 if (!err && trace->filter_pids.map) {
2751                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2752                                                        trace->filter_pids.entries);
2753                 }
2754         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2755                 err = trace__set_filter_loop_pids(trace);
2756         }
2757
2758         return err;
2759 }
2760
2761 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2762 {
2763         struct perf_evlist *evlist = trace->evlist;
2764         struct perf_sample sample;
2765         int err;
2766
2767         err = perf_evlist__parse_sample(evlist, event, &sample);
2768         if (err)
2769                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2770         else
2771                 trace__handle_event(trace, event, &sample);
2772
2773         return 0;
2774 }
2775
2776 static int __trace__flush_events(struct trace *trace)
2777 {
2778         u64 first = ordered_events__first_time(&trace->oe.data);
2779         u64 flush = trace->oe.last - NSEC_PER_SEC;
2780
2781         /* Is there some thing to flush.. */
2782         if (first && first < flush)
2783                 return ordered_events__flush_time(&trace->oe.data, flush);
2784
2785         return 0;
2786 }
2787
2788 static int trace__flush_events(struct trace *trace)
2789 {
2790         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2791 }
2792
2793 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2794 {
2795         int err;
2796
2797         if (!trace->sort_events)
2798                 return __trace__deliver_event(trace, event);
2799
2800         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2801         if (err && err != -1)
2802                 return err;
2803
2804         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2805         if (err)
2806                 return err;
2807
2808         return trace__flush_events(trace);
2809 }
2810
2811 static int ordered_events__deliver_event(struct ordered_events *oe,
2812                                          struct ordered_event *event)
2813 {
2814         struct trace *trace = container_of(oe, struct trace, oe.data);
2815
2816         return __trace__deliver_event(trace, event->event);
2817 }
2818
2819 static int trace__run(struct trace *trace, int argc, const char **argv)
2820 {
2821         struct perf_evlist *evlist = trace->evlist;
2822         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2823         int err = -1, i;
2824         unsigned long before;
2825         const bool forks = argc > 0;
2826         bool draining = false;
2827
2828         trace->live = true;
2829
2830         if (!trace->raw_augmented_syscalls) {
2831                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2832                         goto out_error_raw_syscalls;
2833
2834                 if (trace->trace_syscalls)
2835                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2836         }
2837
2838         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2839                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2840                 if (pgfault_maj == NULL)
2841                         goto out_error_mem;
2842                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2843                 perf_evlist__add(evlist, pgfault_maj);
2844         }
2845
2846         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2847                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2848                 if (pgfault_min == NULL)
2849                         goto out_error_mem;
2850                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2851                 perf_evlist__add(evlist, pgfault_min);
2852         }
2853
2854         if (trace->sched &&
2855             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2856                                    trace__sched_stat_runtime))
2857                 goto out_error_sched_stat_runtime;
2858
2859         /*
2860          * If a global cgroup was set, apply it to all the events without an
2861          * explicit cgroup. I.e.:
2862          *
2863          *      trace -G A -e sched:*switch
2864          *
2865          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2866          * _and_ sched:sched_switch to the 'A' cgroup, while:
2867          *
2868          * trace -e sched:*switch -G A
2869          *
2870          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2871          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2872          * a cgroup (on the root cgroup, sys wide, etc).
2873          *
2874          * Multiple cgroups:
2875          *
2876          * trace -G A -e sched:*switch -G B
2877          *
2878          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2879          * to the 'B' cgroup.
2880          *
2881          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2882          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2883          */
2884         if (trace->cgroup)
2885                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2886
2887         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2888         if (err < 0) {
2889                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2890                 goto out_delete_evlist;
2891         }
2892
2893         err = trace__symbols_init(trace, evlist);
2894         if (err < 0) {
2895                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2896                 goto out_delete_evlist;
2897         }
2898
2899         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2900
2901         signal(SIGCHLD, sig_handler);
2902         signal(SIGINT, sig_handler);
2903
2904         if (forks) {
2905                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2906                                                     argv, false, NULL);
2907                 if (err < 0) {
2908                         fprintf(trace->output, "Couldn't run the workload!\n");
2909                         goto out_delete_evlist;
2910                 }
2911         }
2912
2913         err = perf_evlist__open(evlist);
2914         if (err < 0)
2915                 goto out_error_open;
2916
2917         err = bpf__apply_obj_config();
2918         if (err) {
2919                 char errbuf[BUFSIZ];
2920
2921                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2922                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2923                          errbuf);
2924                 goto out_error_open;
2925         }
2926
2927         err = trace__set_filter_pids(trace);
2928         if (err < 0)
2929                 goto out_error_mem;
2930
2931         if (trace->syscalls.map)
2932                 trace__init_syscalls_bpf_map(trace);
2933
2934         if (trace->ev_qualifier_ids.nr > 0) {
2935                 err = trace__set_ev_qualifier_filter(trace);
2936                 if (err < 0)
2937                         goto out_errno;
2938
2939                 if (trace->syscalls.events.sys_exit) {
2940                         pr_debug("event qualifier tracepoint filter: %s\n",
2941                                  trace->syscalls.events.sys_exit->filter);
2942                 }
2943         }
2944
2945         err = perf_evlist__apply_filters(evlist, &evsel);
2946         if (err < 0)
2947                 goto out_error_apply_filters;
2948
2949         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
2950         if (err < 0)
2951                 goto out_error_mmap;
2952
2953         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2954                 perf_evlist__enable(evlist);
2955
2956         if (forks)
2957                 perf_evlist__start_workload(evlist);
2958
2959         if (trace->opts.initial_delay) {
2960                 usleep(trace->opts.initial_delay * 1000);
2961                 perf_evlist__enable(evlist);
2962         }
2963
2964         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2965                                   evlist->threads->nr > 1 ||
2966                                   perf_evlist__first(evlist)->attr.inherit;
2967
2968         /*
2969          * Now that we already used evsel->attr to ask the kernel to setup the
2970          * events, lets reuse evsel->attr.sample_max_stack as the limit in
2971          * trace__resolve_callchain(), allowing per-event max-stack settings
2972          * to override an explicitly set --max-stack global setting.
2973          */
2974         evlist__for_each_entry(evlist, evsel) {
2975                 if (evsel__has_callchain(evsel) &&
2976                     evsel->attr.sample_max_stack == 0)
2977                         evsel->attr.sample_max_stack = trace->max_stack;
2978         }
2979 again:
2980         before = trace->nr_events;
2981
2982         for (i = 0; i < evlist->nr_mmaps; i++) {
2983                 union perf_event *event;
2984                 struct perf_mmap *md;
2985
2986                 md = &evlist->mmap[i];
2987                 if (perf_mmap__read_init(md) < 0)
2988                         continue;
2989
2990                 while ((event = perf_mmap__read_event(md)) != NULL) {
2991                         ++trace->nr_events;
2992
2993                         err = trace__deliver_event(trace, event);
2994                         if (err)
2995                                 goto out_disable;
2996
2997                         perf_mmap__consume(md);
2998
2999                         if (interrupted)
3000                                 goto out_disable;
3001
3002                         if (done && !draining) {
3003                                 perf_evlist__disable(evlist);
3004                                 draining = true;
3005                         }
3006                 }
3007                 perf_mmap__read_done(md);
3008         }
3009
3010         if (trace->nr_events == before) {
3011                 int timeout = done ? 100 : -1;
3012
3013                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3014                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3015                                 draining = true;
3016
3017                         goto again;
3018                 } else {
3019                         if (trace__flush_events(trace))
3020                                 goto out_disable;
3021                 }
3022         } else {
3023                 goto again;
3024         }
3025
3026 out_disable:
3027         thread__zput(trace->current);
3028
3029         perf_evlist__disable(evlist);
3030
3031         if (trace->sort_events)
3032                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3033
3034         if (!err) {
3035                 if (trace->summary)
3036                         trace__fprintf_thread_summary(trace, trace->output);
3037
3038                 if (trace->show_tool_stats) {
3039                         fprintf(trace->output, "Stats:\n "
3040                                                " vfs_getname : %" PRIu64 "\n"
3041                                                " proc_getname: %" PRIu64 "\n",
3042                                 trace->stats.vfs_getname,
3043                                 trace->stats.proc_getname);
3044                 }
3045         }
3046
3047 out_delete_evlist:
3048         trace__symbols__exit(trace);
3049
3050         perf_evlist__delete(evlist);
3051         cgroup__put(trace->cgroup);
3052         trace->evlist = NULL;
3053         trace->live = false;
3054         return err;
3055 {
3056         char errbuf[BUFSIZ];
3057
3058 out_error_sched_stat_runtime:
3059         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3060         goto out_error;
3061
3062 out_error_raw_syscalls:
3063         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3064         goto out_error;
3065
3066 out_error_mmap:
3067         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3068         goto out_error;
3069
3070 out_error_open:
3071         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3072
3073 out_error:
3074         fprintf(trace->output, "%s\n", errbuf);
3075         goto out_delete_evlist;
3076
3077 out_error_apply_filters:
3078         fprintf(trace->output,
3079                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3080                 evsel->filter, perf_evsel__name(evsel), errno,
3081                 str_error_r(errno, errbuf, sizeof(errbuf)));
3082         goto out_delete_evlist;
3083 }
3084 out_error_mem:
3085         fprintf(trace->output, "Not enough memory to run!\n");
3086         goto out_delete_evlist;
3087
3088 out_errno:
3089         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3090         goto out_delete_evlist;
3091 }
3092
3093 static int trace__replay(struct trace *trace)
3094 {
3095         const struct perf_evsel_str_handler handlers[] = {
3096                 { "probe:vfs_getname",       trace__vfs_getname, },
3097         };
3098         struct perf_data data = {
3099                 .file      = {
3100                         .path = input_name,
3101                 },
3102                 .mode      = PERF_DATA_MODE_READ,
3103                 .force     = trace->force,
3104         };
3105         struct perf_session *session;
3106         struct perf_evsel *evsel;
3107         int err = -1;
3108
3109         trace->tool.sample        = trace__process_sample;
3110         trace->tool.mmap          = perf_event__process_mmap;
3111         trace->tool.mmap2         = perf_event__process_mmap2;
3112         trace->tool.comm          = perf_event__process_comm;
3113         trace->tool.exit          = perf_event__process_exit;
3114         trace->tool.fork          = perf_event__process_fork;
3115         trace->tool.attr          = perf_event__process_attr;
3116         trace->tool.tracing_data  = perf_event__process_tracing_data;
3117         trace->tool.build_id      = perf_event__process_build_id;
3118         trace->tool.namespaces    = perf_event__process_namespaces;
3119
3120         trace->tool.ordered_events = true;
3121         trace->tool.ordering_requires_timestamps = true;
3122
3123         /* add tid to output */
3124         trace->multiple_threads = true;
3125
3126         session = perf_session__new(&data, false, &trace->tool);
3127         if (session == NULL)
3128                 return -1;
3129
3130         if (trace->opts.target.pid)
3131                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3132
3133         if (trace->opts.target.tid)
3134                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3135
3136         if (symbol__init(&session->header.env) < 0)
3137                 goto out;
3138
3139         trace->host = &session->machines.host;
3140
3141         err = perf_session__set_tracepoints_handlers(session, handlers);
3142         if (err)
3143                 goto out;
3144
3145         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3146                                                      "raw_syscalls:sys_enter");
3147         /* older kernels have syscalls tp versus raw_syscalls */
3148         if (evsel == NULL)
3149                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3150                                                              "syscalls:sys_enter");
3151
3152         if (evsel &&
3153             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3154             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3155                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3156                 goto out;
3157         }
3158
3159         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3160                                                      "raw_syscalls:sys_exit");
3161         if (evsel == NULL)
3162                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3163                                                              "syscalls:sys_exit");
3164         if (evsel &&
3165             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3166             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3167                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3168                 goto out;
3169         }
3170
3171         evlist__for_each_entry(session->evlist, evsel) {
3172                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3173                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3174                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3175                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3176                         evsel->handler = trace__pgfault;
3177         }
3178
3179         setup_pager();
3180
3181         err = perf_session__process_events(session);
3182         if (err)
3183                 pr_err("Failed to process events, error %d", err);
3184
3185         else if (trace->summary)
3186                 trace__fprintf_thread_summary(trace, trace->output);
3187
3188 out:
3189         perf_session__delete(session);
3190
3191         return err;
3192 }
3193
3194 static size_t trace__fprintf_threads_header(FILE *fp)
3195 {
3196         size_t printed;
3197
3198         printed  = fprintf(fp, "\n Summary of events:\n\n");
3199
3200         return printed;
3201 }
3202
3203 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3204         struct stats    *stats;
3205         double          msecs;
3206         int             syscall;
3207 )
3208 {
3209         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3210         struct stats *stats = source->priv;
3211
3212         entry->syscall = source->i;
3213         entry->stats   = stats;
3214         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3215 }
3216
3217 static size_t thread__dump_stats(struct thread_trace *ttrace,
3218                                  struct trace *trace, FILE *fp)
3219 {
3220         size_t printed = 0;
3221         struct syscall *sc;
3222         struct rb_node *nd;
3223         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3224
3225         if (syscall_stats == NULL)
3226                 return 0;
3227
3228         printed += fprintf(fp, "\n");
3229
3230         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3231         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3232         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3233
3234         resort_rb__for_each_entry(nd, syscall_stats) {
3235                 struct stats *stats = syscall_stats_entry->stats;
3236                 if (stats) {
3237                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3238                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3239                         double avg = avg_stats(stats);
3240                         double pct;
3241                         u64 n = (u64) stats->n;
3242
3243                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3244                         avg /= NSEC_PER_MSEC;
3245
3246                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3247                         printed += fprintf(fp, "   %-15s", sc->name);
3248                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3249                                            n, syscall_stats_entry->msecs, min, avg);
3250                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3251                 }
3252         }
3253
3254         resort_rb__delete(syscall_stats);
3255         printed += fprintf(fp, "\n\n");
3256
3257         return printed;
3258 }
3259
3260 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3261 {
3262         size_t printed = 0;
3263         struct thread_trace *ttrace = thread__priv(thread);
3264         double ratio;
3265
3266         if (ttrace == NULL)
3267                 return 0;
3268
3269         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3270
3271         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3272         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3273         printed += fprintf(fp, "%.1f%%", ratio);
3274         if (ttrace->pfmaj)
3275                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3276         if (ttrace->pfmin)
3277                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3278         if (trace->sched)
3279                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3280         else if (fputc('\n', fp) != EOF)
3281                 ++printed;
3282
3283         printed += thread__dump_stats(ttrace, trace, fp);
3284
3285         return printed;
3286 }
3287
3288 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3289 {
3290         return ttrace ? ttrace->nr_events : 0;
3291 }
3292
3293 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3294         struct thread *thread;
3295 )
3296 {
3297         entry->thread = rb_entry(nd, struct thread, rb_node);
3298 }
3299
3300 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3301 {
3302         size_t printed = trace__fprintf_threads_header(fp);
3303         struct rb_node *nd;
3304         int i;
3305
3306         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3307                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3308
3309                 if (threads == NULL) {
3310                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3311                         return 0;
3312                 }
3313
3314                 resort_rb__for_each_entry(nd, threads)
3315                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3316
3317                 resort_rb__delete(threads);
3318         }
3319         return printed;
3320 }
3321
3322 static int trace__set_duration(const struct option *opt, const char *str,
3323                                int unset __maybe_unused)
3324 {
3325         struct trace *trace = opt->value;
3326
3327         trace->duration_filter = atof(str);
3328         return 0;
3329 }
3330
3331 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3332                                               int unset __maybe_unused)
3333 {
3334         int ret = -1;
3335         size_t i;
3336         struct trace *trace = opt->value;
3337         /*
3338          * FIXME: introduce a intarray class, plain parse csv and create a
3339          * { int nr, int entries[] } struct...
3340          */
3341         struct intlist *list = intlist__new(str);
3342
3343         if (list == NULL)
3344                 return -1;
3345
3346         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3347         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3348
3349         if (trace->filter_pids.entries == NULL)
3350                 goto out;
3351
3352         trace->filter_pids.entries[0] = getpid();
3353
3354         for (i = 1; i < trace->filter_pids.nr; ++i)
3355                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3356
3357         intlist__delete(list);
3358         ret = 0;
3359 out:
3360         return ret;
3361 }
3362
3363 static int trace__open_output(struct trace *trace, const char *filename)
3364 {
3365         struct stat st;
3366
3367         if (!stat(filename, &st) && st.st_size) {
3368                 char oldname[PATH_MAX];
3369
3370                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3371                 unlink(oldname);
3372                 rename(filename, oldname);
3373         }
3374
3375         trace->output = fopen(filename, "w");
3376
3377         return trace->output == NULL ? -errno : 0;
3378 }
3379
3380 static int parse_pagefaults(const struct option *opt, const char *str,
3381                             int unset __maybe_unused)
3382 {
3383         int *trace_pgfaults = opt->value;
3384
3385         if (strcmp(str, "all") == 0)
3386                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3387         else if (strcmp(str, "maj") == 0)
3388                 *trace_pgfaults |= TRACE_PFMAJ;
3389         else if (strcmp(str, "min") == 0)
3390                 *trace_pgfaults |= TRACE_PFMIN;
3391         else
3392                 return -1;
3393
3394         return 0;
3395 }
3396
3397 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3398 {
3399         struct perf_evsel *evsel;
3400
3401         evlist__for_each_entry(evlist, evsel)
3402                 evsel->handler = handler;
3403 }
3404
3405 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3406 {
3407         struct perf_evsel *evsel;
3408
3409         evlist__for_each_entry(evlist, evsel) {
3410                 if (evsel->priv || !evsel->tp_format)
3411                         continue;
3412
3413                 if (strcmp(evsel->tp_format->system, "syscalls"))
3414                         continue;
3415
3416                 if (perf_evsel__init_syscall_tp(evsel))
3417                         return -1;
3418
3419                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3420                         struct syscall_tp *sc = evsel->priv;
3421
3422                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3423                                 return -1;
3424                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3425                         struct syscall_tp *sc = evsel->priv;
3426
3427                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3428                                 return -1;
3429                 }
3430         }
3431
3432         return 0;
3433 }
3434
3435 /*
3436  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3437  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3438  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3439  *
3440  * It'd be better to introduce a parse_options() variant that would return a
3441  * list with the terms it didn't match to an event...
3442  */
3443 static int trace__parse_events_option(const struct option *opt, const char *str,
3444                                       int unset __maybe_unused)
3445 {
3446         struct trace *trace = (struct trace *)opt->value;
3447         const char *s = str;
3448         char *sep = NULL, *lists[2] = { NULL, NULL, };
3449         int len = strlen(str) + 1, err = -1, list, idx;
3450         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3451         char group_name[PATH_MAX];
3452         struct syscall_fmt *fmt;
3453
3454         if (strace_groups_dir == NULL)
3455                 return -1;
3456
3457         if (*s == '!') {
3458                 ++s;
3459                 trace->not_ev_qualifier = true;
3460         }
3461
3462         while (1) {
3463                 if ((sep = strchr(s, ',')) != NULL)
3464                         *sep = '\0';
3465
3466                 list = 0;
3467                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3468                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3469                         list = 1;
3470                         goto do_concat;
3471                 }
3472
3473                 fmt = syscall_fmt__find_by_alias(s);
3474                 if (fmt != NULL) {
3475                         list = 1;
3476                         s = fmt->name;
3477                 } else {
3478                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3479                         if (access(group_name, R_OK) == 0)
3480                                 list = 1;
3481                 }
3482 do_concat:
3483                 if (lists[list]) {
3484                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3485                 } else {
3486                         lists[list] = malloc(len);
3487                         if (lists[list] == NULL)
3488                                 goto out;
3489                         strcpy(lists[list], s);
3490                 }
3491
3492                 if (!sep)
3493                         break;
3494
3495                 *sep = ',';
3496                 s = sep + 1;
3497         }
3498
3499         if (lists[1] != NULL) {
3500                 struct strlist_config slist_config = {
3501                         .dirname = strace_groups_dir,
3502                 };
3503
3504                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3505                 if (trace->ev_qualifier == NULL) {
3506                         fputs("Not enough memory to parse event qualifier", trace->output);
3507                         goto out;
3508                 }
3509
3510                 if (trace__validate_ev_qualifier(trace))
3511                         goto out;
3512                 trace->trace_syscalls = true;
3513         }
3514
3515         err = 0;
3516
3517         if (lists[0]) {
3518                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3519                                                "event selector. use 'perf list' to list available events",
3520                                                parse_events_option);
3521                 err = parse_events_option(&o, lists[0], 0);
3522         }
3523 out:
3524         if (sep)
3525                 *sep = ',';
3526
3527         return err;
3528 }
3529
3530 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3531 {
3532         struct trace *trace = opt->value;
3533
3534         if (!list_empty(&trace->evlist->entries))
3535                 return parse_cgroups(opt, str, unset);
3536
3537         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3538
3539         return 0;
3540 }
3541
3542 static struct bpf_map *bpf__find_map_by_name(const char *name)
3543 {
3544         struct bpf_object *obj, *tmp;
3545
3546         bpf_object__for_each_safe(obj, tmp) {
3547                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3548                 if (map)
3549                         return map;
3550
3551         }
3552
3553         return NULL;
3554 }
3555
3556 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3557 {
3558         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3559 }
3560
3561 static void trace__set_bpf_map_syscalls(struct trace *trace)
3562 {
3563         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3564 }
3565
3566 static int trace__config(const char *var, const char *value, void *arg)
3567 {
3568         struct trace *trace = arg;
3569         int err = 0;
3570
3571         if (!strcmp(var, "trace.add_events")) {
3572                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3573                                                "event selector. use 'perf list' to list available events",
3574                                                parse_events_option);
3575                 err = parse_events_option(&o, value, 0);
3576         } else if (!strcmp(var, "trace.show_timestamp")) {
3577                 trace->show_tstamp = perf_config_bool(var, value);
3578         } else if (!strcmp(var, "trace.show_duration")) {
3579                 trace->show_duration = perf_config_bool(var, value);
3580         } else if (!strcmp(var, "trace.show_arg_names")) {
3581                 trace->show_arg_names = perf_config_bool(var, value);
3582                 if (!trace->show_arg_names)
3583                         trace->show_zeros = true;
3584         } else if (!strcmp(var, "trace.show_zeros")) {
3585                 bool new_show_zeros = perf_config_bool(var, value);
3586                 if (!trace->show_arg_names && !new_show_zeros) {
3587                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3588                         goto out;
3589                 }
3590                 trace->show_zeros = new_show_zeros;
3591         } else if (!strcmp(var, "trace.show_prefix")) {
3592                 trace->show_string_prefix = perf_config_bool(var, value);
3593         } else if (!strcmp(var, "trace.no_inherit")) {
3594                 trace->opts.no_inherit = perf_config_bool(var, value);
3595         } else if (!strcmp(var, "trace.args_alignment")) {
3596                 int args_alignment = 0;
3597                 if (perf_config_int(&args_alignment, var, value) == 0)
3598                         trace->args_alignment = args_alignment;
3599         }
3600 out:
3601         return err;
3602 }
3603
3604 int cmd_trace(int argc, const char **argv)
3605 {
3606         const char *trace_usage[] = {
3607                 "perf trace [<options>] [<command>]",
3608                 "perf trace [<options>] -- <command> [<options>]",
3609                 "perf trace record [<options>] [<command>]",
3610                 "perf trace record [<options>] -- <command> [<options>]",
3611                 NULL
3612         };
3613         struct trace trace = {
3614                 .syscalls = {
3615                         . max = -1,
3616                 },
3617                 .opts = {
3618                         .target = {
3619                                 .uid       = UINT_MAX,
3620                                 .uses_mmap = true,
3621                         },
3622                         .user_freq     = UINT_MAX,
3623                         .user_interval = ULLONG_MAX,
3624                         .no_buffering  = true,
3625                         .mmap_pages    = UINT_MAX,
3626                 },
3627                 .output = stderr,
3628                 .show_comm = true,
3629                 .show_tstamp = true,
3630                 .show_duration = true,
3631                 .show_arg_names = true,
3632                 .args_alignment = 70,
3633                 .trace_syscalls = false,
3634                 .kernel_syscallchains = false,
3635                 .max_stack = UINT_MAX,
3636                 .max_events = ULONG_MAX,
3637         };
3638         const char *output_name = NULL;
3639         const struct option trace_options[] = {
3640         OPT_CALLBACK('e', "event", &trace, "event",
3641                      "event/syscall selector. use 'perf list' to list available events",
3642                      trace__parse_events_option),
3643         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3644                     "show the thread COMM next to its id"),
3645         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3646         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3647                      trace__parse_events_option),
3648         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3649         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3650         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3651                     "trace events on existing process id"),
3652         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3653                     "trace events on existing thread id"),
3654         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3655                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3656         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3657                     "system-wide collection from all CPUs"),
3658         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3659                     "list of cpus to monitor"),
3660         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3661                     "child tasks do not inherit counters"),
3662         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3663                      "number of mmap data pages",
3664                      perf_evlist__parse_mmap_pages),
3665         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3666                    "user to profile"),
3667         OPT_CALLBACK(0, "duration", &trace, "float",
3668                      "show only events with duration > N.M ms",
3669                      trace__set_duration),
3670         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3671         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3672         OPT_BOOLEAN('T', "time", &trace.full_time,
3673                     "Show full timestamp, not time relative to first start"),
3674         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3675                     "Show only syscalls that failed"),
3676         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3677                     "Show only syscall summary with statistics"),
3678         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3679                     "Show all syscalls and summary with statistics"),
3680         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3681                      "Trace pagefaults", parse_pagefaults, "maj"),
3682         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3683         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3684         OPT_CALLBACK(0, "call-graph", &trace.opts,
3685                      "record_mode[,record_size]", record_callchain_help,
3686                      &record_parse_callchain_opt),
3687         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3688                     "Show the kernel callchains on the syscall exit path"),
3689         OPT_ULONG(0, "max-events", &trace.max_events,
3690                 "Set the maximum number of events to print, exit after that is reached. "),
3691         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3692                      "Set the minimum stack depth when parsing the callchain, "
3693                      "anything below the specified depth will be ignored."),
3694         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3695                      "Set the maximum stack depth when parsing the callchain, "
3696                      "anything beyond the specified depth will be ignored. "
3697                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3698         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3699                         "Sort batch of events before processing, use if getting out of order events"),
3700         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3701                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3702         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3703                         "per thread proc mmap processing timeout in ms"),
3704         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3705                      trace__parse_cgroups),
3706         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3707                      "ms to wait before starting measurement after program "
3708                      "start"),
3709         OPT_END()
3710         };
3711         bool __maybe_unused max_stack_user_set = true;
3712         bool mmap_pages_user_set = true;
3713         struct perf_evsel *evsel;
3714         const char * const trace_subcommands[] = { "record", NULL };
3715         int err = -1;
3716         char bf[BUFSIZ];
3717
3718         signal(SIGSEGV, sighandler_dump_stack);
3719         signal(SIGFPE, sighandler_dump_stack);
3720
3721         trace.evlist = perf_evlist__new();
3722         trace.sctbl = syscalltbl__new();
3723
3724         if (trace.evlist == NULL || trace.sctbl == NULL) {
3725                 pr_err("Not enough memory to run!\n");
3726                 err = -ENOMEM;
3727                 goto out;
3728         }
3729
3730         err = perf_config(trace__config, &trace);
3731         if (err)
3732                 goto out;
3733
3734         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3735                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3736
3737         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3738                 usage_with_options_msg(trace_usage, trace_options,
3739                                        "cgroup monitoring only available in system-wide mode");
3740         }
3741
3742         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3743         if (IS_ERR(evsel)) {
3744                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3745                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3746                 goto out;
3747         }
3748
3749         if (evsel) {
3750                 trace.syscalls.events.augmented = evsel;
3751                 trace__set_bpf_map_filtered_pids(&trace);
3752                 trace__set_bpf_map_syscalls(&trace);
3753         }
3754
3755         err = bpf__setup_stdout(trace.evlist);
3756         if (err) {
3757                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3758                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3759                 goto out;
3760         }
3761
3762         err = -1;
3763
3764         if (trace.trace_pgfaults) {
3765                 trace.opts.sample_address = true;
3766                 trace.opts.sample_time = true;
3767         }
3768
3769         if (trace.opts.mmap_pages == UINT_MAX)
3770                 mmap_pages_user_set = false;
3771
3772         if (trace.max_stack == UINT_MAX) {
3773                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3774                 max_stack_user_set = false;
3775         }
3776
3777 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3778         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3779                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3780         }
3781 #endif
3782
3783         if (callchain_param.enabled) {
3784                 if (!mmap_pages_user_set && geteuid() == 0)
3785                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3786
3787                 symbol_conf.use_callchain = true;
3788         }
3789
3790         if (trace.evlist->nr_entries > 0) {
3791                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3792                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3793                         perror("failed to set syscalls:* tracepoint fields");
3794                         goto out;
3795                 }
3796         }
3797
3798         if (trace.sort_events) {
3799                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3800                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3801         }
3802
3803         /*
3804          * If we are augmenting syscalls, then combine what we put in the
3805          * __augmented_syscalls__ BPF map with what is in the
3806          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3807          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3808          *
3809          * We'll switch to look at two BPF maps, one for sys_enter and the
3810          * other for sys_exit when we start augmenting the sys_exit paths with
3811          * buffers that are being copied from kernel to userspace, think 'read'
3812          * syscall.
3813          */
3814         if (trace.syscalls.events.augmented) {
3815                 evsel = trace.syscalls.events.augmented;
3816
3817                 if (perf_evsel__init_augmented_syscall_tp(evsel) ||
3818                     perf_evsel__init_augmented_syscall_tp_args(evsel))
3819                         goto out;
3820                 evsel->handler = trace__sys_enter;
3821
3822                 evlist__for_each_entry(trace.evlist, evsel) {
3823                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3824
3825                         if (raw_syscalls_sys_exit) {
3826                                 trace.raw_augmented_syscalls = true;
3827                                 goto init_augmented_syscall_tp;
3828                         }
3829
3830                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3831 init_augmented_syscall_tp:
3832                                 perf_evsel__init_augmented_syscall_tp(evsel);
3833                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3834                                 evsel->handler = trace__sys_exit;
3835                         }
3836                 }
3837         }
3838
3839         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3840                 return trace__record(&trace, argc-1, &argv[1]);
3841
3842         /* summary_only implies summary option, but don't overwrite summary if set */
3843         if (trace.summary_only)
3844                 trace.summary = trace.summary_only;
3845
3846         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3847             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3848                 trace.trace_syscalls = true;
3849         }
3850
3851         if (output_name != NULL) {
3852                 err = trace__open_output(&trace, output_name);
3853                 if (err < 0) {
3854                         perror("failed to create output file");
3855                         goto out;
3856                 }
3857         }
3858
3859         err = target__validate(&trace.opts.target);
3860         if (err) {
3861                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3862                 fprintf(trace.output, "%s", bf);
3863                 goto out_close;
3864         }
3865
3866         err = target__parse_uid(&trace.opts.target);
3867         if (err) {
3868                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3869                 fprintf(trace.output, "%s", bf);
3870                 goto out_close;
3871         }
3872
3873         if (!argc && target__none(&trace.opts.target))
3874                 trace.opts.target.system_wide = true;
3875
3876         if (input_name)
3877                 err = trace__replay(&trace);
3878         else
3879                 err = trace__run(&trace, argc, argv);
3880
3881 out_close:
3882         if (output_name != NULL)
3883                 fclose(trace.output);
3884 out:
3885         return err;
3886 }