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