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