tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  *
  16  * Released under the GPL v2. (and only v2, not any later version)
  17  */
  18
  19 #include <traceevent/event-parse.h>
  20 #include <api/fs/tracing_path.h>
  21 #include <bpf/bpf.h>
  22 #include "util/bpf_map.h"
  23 #include "builtin.h"
  24 #include "util/cgroup.h"
  25 #include "util/color.h"
  26 #include "util/config.h"
  27 #include "util/debug.h"
  28 #include "util/env.h"
  29 #include "util/event.h"
  30 #include "util/evlist.h"
  31 #include <subcmd/exec-cmd.h>
  32 #include "util/machine.h"
  33 #include "util/map.h"
  34 #include "util/symbol.h"
  35 #include "util/path.h"
  36 #include "util/session.h"
  37 #include "util/thread.h"
  38 #include <subcmd/parse-options.h>
  39 #include "util/strlist.h"
  40 #include "util/intlist.h"
  41 #include "util/thread_map.h"
  42 #include "util/stat.h"
  43 #include "trace/beauty/beauty.h"
  44 #include "trace-event.h"
  45 #include "util/parse-events.h"
  46 #include "util/bpf-loader.h"
  47 #include "callchain.h"
  48 #include "print_binary.h"
  49 #include "string2.h"
  50 #include "syscalltbl.h"
  51 #include "rb_resort.h"
  52
  53 #include <errno.h>
  54 #include <inttypes.h>
  55 #include <poll.h>
  56 #include <signal.h>
  57 #include <stdlib.h>
  58 #include <string.h>
  59 #include <linux/err.h>
  60 #include <linux/filter.h>
  61 #include <linux/kernel.h>
  62 #include <linux/random.h>
  63 #include <linux/stringify.h>
  64 #include <linux/time64.h>
  65 #include <fcntl.h>
  66 #include <sys/sysmacros.h>
  67
  68 #include "sane_ctype.h"
  69
  70 #ifndef O_CLOEXEC
  71 # define O_CLOEXEC              02000000
  72 #endif
  73
  74 #ifndef F_LINUX_SPECIFIC_BASE
  75 # define F_LINUX_SPECIFIC_BASE  1024
  76 #endif
  77
  78 struct trace {
  79         struct perf_tool        tool;
  80         struct syscalltbl       *sctbl;
  81         struct {
  82                 int             max;
  83                 struct syscall  *table;
  84                 struct bpf_map  *map;
  85                 struct {
  86                         struct perf_evsel *sys_enter,
  87                                           *sys_exit,
  88                                           *augmented;
  89                 }               events;
  90         } syscalls;
  91         struct {
  92                 struct bpf_map *map;
  93         } dump;
  94         struct record_opts      opts;
  95         struct perf_evlist      *evlist;
  96         struct machine          *host;
  97         struct thread           *current;
  98         struct cgroup           *cgroup;
  99         u64                     base_time;
 100         FILE                    *output;
 101         unsigned long           nr_events;
 102         unsigned long           nr_events_printed;
 103         unsigned long           max_events;
 104         struct strlist          *ev_qualifier;
 105         struct {
 106                 size_t          nr;
 107                 int             *entries;
 108         }                       ev_qualifier_ids;
 109         struct {
 110                 size_t          nr;
 111                 pid_t           *entries;
 112                 struct bpf_map  *map;
 113         }                       filter_pids;
 114         double                  duration_filter;
 115         double                  runtime_ms;
 116         struct {
 117                 u64             vfs_getname,
 118                                 proc_getname;
 119         } stats;
 120         unsigned int            max_stack;
 121         unsigned int            min_stack;
 122         int                     raw_augmented_syscalls_args_size;
 123         bool                    raw_augmented_syscalls;
 124         bool                    sort_events;
 125         bool                    not_ev_qualifier;
 126         bool                    live;
 127         bool                    full_time;
 128         bool                    sched;
 129         bool                    multiple_threads;
 130         bool                    summary;
 131         bool                    summary_only;
 132         bool                    failure_only;
 133         bool                    show_comm;
 134         bool                    print_sample;
 135         bool                    show_tool_stats;
 136         bool                    trace_syscalls;
 137         bool                    kernel_syscallchains;
 138         s16                     args_alignment;
 139         bool                    show_tstamp;
 140         bool                    show_duration;
 141         bool                    show_zeros;
 142         bool                    show_arg_names;
 143         bool                    show_string_prefix;
 144         bool                    force;
 145         bool                    vfs_getname;
 146         int                     trace_pgfaults;
 147         struct {
 148                 struct ordered_events   data;
 149                 u64                     last;
 150         } oe;
 151 };
 152
 153 struct tp_field {
 154         int offset;
 155         union {
 156                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 157                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 158         };
 159 };
 160
 161 #define TP_UINT_FIELD(bits) \
 162 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 163 { \
 164         u##bits value; \
 165         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 166         return value;  \
 167 }
 168
 169 TP_UINT_FIELD(8);
 170 TP_UINT_FIELD(16);
 171 TP_UINT_FIELD(32);
 172 TP_UINT_FIELD(64);
 173
 174 #define TP_UINT_FIELD__SWAPPED(bits) \
 175 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 176 { \
 177         u##bits value; \
 178         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 179         return bswap_##bits(value);\
 180 }
 181
 182 TP_UINT_FIELD__SWAPPED(16);
 183 TP_UINT_FIELD__SWAPPED(32);
 184 TP_UINT_FIELD__SWAPPED(64);
 185
 186 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 187 {
 188         field->offset = offset;
 189
 190         switch (size) {
 191         case 1:
 192                 field->integer = tp_field__u8;
 193                 break;
 194         case 2:
 195                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 196                 break;
 197         case 4:
 198                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 199                 break;
 200         case 8:
 201                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 202                 break;
 203         default:
 204                 return -1;
 205         }
 206
 207         return 0;
 208 }
 209
 210 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 211 {
 212         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 213 }
 214
 215 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 216 {
 217         return sample->raw_data + field->offset;
 218 }
 219
 220 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 221 {
 222         field->offset = offset;
 223         field->pointer = tp_field__ptr;
 224         return 0;
 225 }
 226
 227 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 228 {
 229         return __tp_field__init_ptr(field, format_field->offset);
 230 }
 231
 232 struct syscall_tp {
 233         struct tp_field id;
 234         union {
 235                 struct tp_field args, ret;
 236         };
 237 };
 238
 239 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 240                                           struct tp_field *field,
 241                                           const char *name)
 242 {
 243         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 244
 245         if (format_field == NULL)
 246                 return -1;
 247
 248         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 249 }
 250
 251 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 252         ({ struct syscall_tp *sc = evsel->priv;\
 253            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 254
 255 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 256                                          struct tp_field *field,
 257                                          const char *name)
 258 {
 259         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 260
 261         if (format_field == NULL)
 262                 return -1;
 263
 264         return tp_field__init_ptr(field, format_field);
 265 }
 266
 267 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 268         ({ struct syscall_tp *sc = evsel->priv;\
 269            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 270
 271 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 272 {
 273         zfree(&evsel->priv);
 274         perf_evsel__delete(evsel);
 275 }
 276
 277 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 278 {
 279         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 280
 281         if (evsel->priv != NULL) {
 282                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 283                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 284                         goto out_delete;
 285                 return 0;
 286         }
 287
 288         return -ENOMEM;
 289 out_delete:
 290         zfree(&evsel->priv);
 291         return -ENOENT;
 292 }
 293
 294 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 295 {
 296         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 297
 298         if (evsel->priv != NULL) {
 299                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 300                 if (syscall_id == NULL)
 301                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 302                 if (syscall_id == NULL)
 303                         goto out_delete;
 304                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 305                         goto out_delete;
 306
 307                 return 0;
 308         }
 309
 310         return -ENOMEM;
 311 out_delete:
 312         zfree(&evsel->priv);
 313         return -EINVAL;
 314 }
 315
 316 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 317 {
 318         struct syscall_tp *sc = evsel->priv;
 319
 320         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 321 }
 322
 323 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 324 {
 325         struct syscall_tp *sc = evsel->priv;
 326
 327         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 328 }
 329
 330 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 331 {
 332         evsel->priv = malloc(sizeof(struct syscall_tp));
 333         if (evsel->priv != NULL) {
 334                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 335                         goto out_delete;
 336
 337                 evsel->handler = handler;
 338                 return 0;
 339         }
 340
 341         return -ENOMEM;
 342
 343 out_delete:
 344         zfree(&evsel->priv);
 345         return -ENOENT;
 346 }
 347
 348 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 349 {
 350         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 351
 352         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 353         if (IS_ERR(evsel))
 354                 evsel = perf_evsel__newtp("syscalls", direction);
 355
 356         if (IS_ERR(evsel))
 357                 return NULL;
 358
 359         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 360                 goto out_delete;
 361
 362         return evsel;
 363
 364 out_delete:
 365         perf_evsel__delete_priv(evsel);
 366         return NULL;
 367 }
 368
 369 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 370         ({ struct syscall_tp *fields = evsel->priv; \
 371            fields->name.integer(&fields->name, sample); })
 372
 373 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 374         ({ struct syscall_tp *fields = evsel->priv; \
 375            fields->name.pointer(&fields->name, sample); })
 376
 377 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 378 {
 379         int idx = val - sa->offset;
 380
 381         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 382                 size_t printed = scnprintf(bf, size, intfmt, val);
 383                 if (show_prefix)
 384                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 385                 return printed;
 386         }
 387
 388         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 389 }
 390
 391 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 392                                                 const char *intfmt,
 393                                                 struct syscall_arg *arg)
 394 {
 395         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 396 }
 397
 398 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 399                                               struct syscall_arg *arg)
 400 {
 401         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 402 }
 403
 404 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 405
 406 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 407 {
 408         size_t printed;
 409         int i;
 410
 411         for (i = 0; i < sas->nr_entries; ++i) {
 412                 struct strarray *sa = sas->entries[i];
 413                 int idx = val - sa->offset;
 414
 415                 if (idx >= 0 && idx < sa->nr_entries) {
 416                         if (sa->entries[idx] == NULL)
 417                                 break;
 418                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 419                 }
 420         }
 421
 422         printed = scnprintf(bf, size, intfmt, val);
 423         if (show_prefix)
 424                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 425         return printed;
 426 }
 427
 428 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 429                                         struct syscall_arg *arg)
 430 {
 431         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 432 }
 433
 434 #ifndef AT_FDCWD
 435 #define AT_FDCWD        -100
 436 #endif
 437
 438 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 439                                            struct syscall_arg *arg)
 440 {
 441         int fd = arg->val;
 442         const char *prefix = "AT_FD";
 443
 444         if (fd == AT_FDCWD)
 445                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 446
 447         return syscall_arg__scnprintf_fd(bf, size, arg);
 448 }
 449
 450 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 451
 452 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 453                                               struct syscall_arg *arg);
 454
 455 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 456
 457 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 458 {
 459         return scnprintf(bf, size, "%#lx", arg->val);
 460 }
 461
 462 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 463 {
 464         if (arg->val == 0)
 465                 return scnprintf(bf, size, "NULL");
 466         return syscall_arg__scnprintf_hex(bf, size, arg);
 467 }
 468
 469 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 470 {
 471         return scnprintf(bf, size, "%d", arg->val);
 472 }
 473
 474 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 475 {
 476         return scnprintf(bf, size, "%ld", arg->val);
 477 }
 478
 479 static const char *bpf_cmd[] = {
 480         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 481         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 482 };
 483 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 484
 485 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 486 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 487
 488 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 489 static DEFINE_STRARRAY(itimers, "ITIMER_");
 490
 491 static const char *keyctl_options[] = {
 492         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 493         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 494         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 495         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 496         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 497 };
 498 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 499
 500 static const char *whences[] = { "SET", "CUR", "END",
 501 #ifdef SEEK_DATA
 502 "DATA",
 503 #endif
 504 #ifdef SEEK_HOLE
 505 "HOLE",
 506 #endif
 507 };
 508 static DEFINE_STRARRAY(whences, "SEEK_");
 509
 510 static const char *fcntl_cmds[] = {
 511         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 512         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 513         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 514         "GETOWNER_UIDS",
 515 };
 516 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 517
 518 static const char *fcntl_linux_specific_cmds[] = {
 519         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 520         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 521         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 522 };
 523
 524 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 525
 526 static struct strarray *fcntl_cmds_arrays[] = {
 527         &strarray__fcntl_cmds,
 528         &strarray__fcntl_linux_specific_cmds,
 529 };
 530
 531 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 532
 533 static const char *rlimit_resources[] = {
 534         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 535         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 536         "RTTIME",
 537 };
 538 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 539
 540 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 541 static DEFINE_STRARRAY(sighow, "SIG_");
 542
 543 static const char *clockid[] = {
 544         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 545         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 546         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 547 };
 548 static DEFINE_STRARRAY(clockid, "CLOCK_");
 549
 550 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 551                                                  struct syscall_arg *arg)
 552 {
 553         bool show_prefix = arg->show_string_prefix;
 554         const char *suffix = "_OK";
 555         size_t printed = 0;
 556         int mode = arg->val;
 557
 558         if (mode == F_OK) /* 0 */
 559                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 560 #define P_MODE(n) \
 561         if (mode & n##_OK) { \
 562                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 563                 mode &= ~n##_OK; \
 564         }
 565
 566         P_MODE(R);
 567         P_MODE(W);
 568         P_MODE(X);
 569 #undef P_MODE
 570
 571         if (mode)
 572                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 573
 574         return printed;
 575 }
 576
 577 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 578
 579 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 580                                               struct syscall_arg *arg);
 581
 582 #define SCA_FILENAME syscall_arg__scnprintf_filename
 583
 584 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 585                                                 struct syscall_arg *arg)
 586 {
 587         bool show_prefix = arg->show_string_prefix;
 588         const char *prefix = "O_";
 589         int printed = 0, flags = arg->val;
 590
 591 #define P_FLAG(n) \
 592         if (flags & O_##n) { \
 593                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 594                 flags &= ~O_##n; \
 595         }
 596
 597         P_FLAG(CLOEXEC);
 598         P_FLAG(NONBLOCK);
 599 #undef P_FLAG
 600
 601         if (flags)
 602                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 603
 604         return printed;
 605 }
 606
 607 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 608
 609 #ifndef GRND_NONBLOCK
 610 #define GRND_NONBLOCK   0x0001
 611 #endif
 612 #ifndef GRND_RANDOM
 613 #define GRND_RANDOM     0x0002
 614 #endif
 615
 616 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 617                                                    struct syscall_arg *arg)
 618 {
 619         bool show_prefix = arg->show_string_prefix;
 620         const char *prefix = "GRND_";
 621         int printed = 0, flags = arg->val;
 622
 623 #define P_FLAG(n) \
 624         if (flags & GRND_##n) { \
 625                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 626                 flags &= ~GRND_##n; \
 627         }
 628
 629         P_FLAG(RANDOM);
 630         P_FLAG(NONBLOCK);
 631 #undef P_FLAG
 632
 633         if (flags)
 634                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 635
 636         return printed;
 637 }
 638
 639 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 640
 641 #define STRARRAY(name, array) \
 642           { .scnprintf  = SCA_STRARRAY, \
 643             .parm       = &strarray__##array, }
 644
 645 #include "trace/beauty/arch_errno_names.c"
 646 #include "trace/beauty/eventfd.c"
 647 #include "trace/beauty/futex_op.c"
 648 #include "trace/beauty/futex_val3.c"
 649 #include "trace/beauty/mmap.c"
 650 #include "trace/beauty/mode_t.c"
 651 #include "trace/beauty/msg_flags.c"
 652 #include "trace/beauty/open_flags.c"
 653 #include "trace/beauty/perf_event_open.c"
 654 #include "trace/beauty/pid.c"
 655 #include "trace/beauty/sched_policy.c"
 656 #include "trace/beauty/seccomp.c"
 657 #include "trace/beauty/signum.c"
 658 #include "trace/beauty/socket_type.c"
 659 #include "trace/beauty/waitid_options.c"
 660
 661 struct syscall_arg_fmt {
 662         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 663         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 664         void       *parm;
 665         const char *name;
 666         bool       show_zero;
 667 };
 668
 669 static struct syscall_fmt {
 670         const char *name;
 671         const char *alias;
 672         struct syscall_arg_fmt arg[6];
 673         u8         nr_args;
 674         bool       errpid;
 675         bool       timeout;
 676         bool       hexret;
 677 } syscall_fmts[] = {
 678         { .name     = "access",
 679           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 680         { .name     = "arch_prctl",
 681           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 682                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 683         { .name     = "bind",
 684           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 685         { .name     = "bpf",
 686           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 687         { .name     = "brk",        .hexret = true,
 688           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 689         { .name     = "clock_gettime",
 690           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 691         { .name     = "clone",      .errpid = true, .nr_args = 5,
 692           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 693                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 694                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 695                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 696                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 697         { .name     = "close",
 698           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 699         { .name     = "connect",
 700           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 701         { .name     = "epoll_ctl",
 702           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 703         { .name     = "eventfd2",
 704           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 705         { .name     = "fchmodat",
 706           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 707         { .name     = "fchownat",
 708           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 709         { .name     = "fcntl",
 710           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 711                            .parm      = &strarrays__fcntl_cmds_arrays,
 712                            .show_zero = true, },
 713                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 714         { .name     = "flock",
 715           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 716         { .name     = "fstat", .alias = "newfstat", },
 717         { .name     = "fstatat", .alias = "newfstatat", },
 718         { .name     = "futex",
 719           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 720                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 721         { .name     = "futimesat",
 722           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 723         { .name     = "getitimer",
 724           .arg = { [0] = STRARRAY(which, itimers), }, },
 725         { .name     = "getpid",     .errpid = true, },
 726         { .name     = "getpgid",    .errpid = true, },
 727         { .name     = "getppid",    .errpid = true, },
 728         { .name     = "getrandom",
 729           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 730         { .name     = "getrlimit",
 731           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 732         { .name     = "gettid",     .errpid = true, },
 733         { .name     = "ioctl",
 734           .arg = {
 735 #if defined(__i386__) || defined(__x86_64__)
 736 /*
 737  * FIXME: Make this available to all arches.
 738  */
 739                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 740                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 741 #else
 742                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 743 #endif
 744         { .name     = "kcmp",       .nr_args = 5,
 745           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 746                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 747                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 748                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 749                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 750         { .name     = "keyctl",
 751           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 752         { .name     = "kill",
 753           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 754         { .name     = "linkat",
 755           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 756         { .name     = "lseek",
 757           .arg = { [2] = STRARRAY(whence, whences), }, },
 758         { .name     = "lstat", .alias = "newlstat", },
 759         { .name     = "madvise",
 760           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 761                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 762         { .name     = "mkdirat",
 763           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 764         { .name     = "mknodat",
 765           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 766         { .name     = "mmap",       .hexret = true,
 767 /* The standard mmap maps to old_mmap on s390x */
 768 #if defined(__s390x__)
 769         .alias = "old_mmap",
 770 #endif
 771           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 772                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 773                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 774         { .name     = "mount",
 775           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 776                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 777                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 778         { .name     = "mprotect",
 779           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 780                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 781         { .name     = "mq_unlink",
 782           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 783         { .name     = "mremap",     .hexret = true,
 784           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 785         { .name     = "name_to_handle_at",
 786           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 787         { .name     = "newfstatat",
 788           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 789         { .name     = "open",
 790           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 791         { .name     = "open_by_handle_at",
 792           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 793                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 794         { .name     = "openat",
 795           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 796                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 797         { .name     = "perf_event_open",
 798           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 799                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 800                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 801         { .name     = "pipe2",
 802           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 803         { .name     = "pkey_alloc",
 804           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 805         { .name     = "pkey_free",
 806           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 807         { .name     = "pkey_mprotect",
 808           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 809                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 810                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 811         { .name     = "poll", .timeout = true, },
 812         { .name     = "ppoll", .timeout = true, },
 813         { .name     = "prctl",
 814           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 815                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 816                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 817         { .name     = "pread", .alias = "pread64", },
 818         { .name     = "preadv", .alias = "pread", },
 819         { .name     = "prlimit64",
 820           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 821         { .name     = "pwrite", .alias = "pwrite64", },
 822         { .name     = "readlinkat",
 823           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 824         { .name     = "recvfrom",
 825           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 826         { .name     = "recvmmsg",
 827           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 828         { .name     = "recvmsg",
 829           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 830         { .name     = "renameat",
 831           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 832                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 833         { .name     = "renameat2",
 834           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 835                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 836                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 837         { .name     = "rt_sigaction",
 838           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 839         { .name     = "rt_sigprocmask",
 840           .arg = { [0] = STRARRAY(how, sighow), }, },
 841         { .name     = "rt_sigqueueinfo",
 842           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 843         { .name     = "rt_tgsigqueueinfo",
 844           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 845         { .name     = "sched_setscheduler",
 846           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 847         { .name     = "seccomp",
 848           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 849                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 850         { .name     = "select", .timeout = true, },
 851         { .name     = "sendmmsg",
 852           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 853         { .name     = "sendmsg",
 854           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 855         { .name     = "sendto",
 856           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 857                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 858         { .name     = "set_tid_address", .errpid = true, },
 859         { .name     = "setitimer",
 860           .arg = { [0] = STRARRAY(which, itimers), }, },
 861         { .name     = "setrlimit",
 862           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 863         { .name     = "socket",
 864           .arg = { [0] = STRARRAY(family, socket_families),
 865                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 866                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 867         { .name     = "socketpair",
 868           .arg = { [0] = STRARRAY(family, socket_families),
 869                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 870                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 871         { .name     = "stat", .alias = "newstat", },
 872         { .name     = "statx",
 873           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 874                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 875                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 876         { .name     = "swapoff",
 877           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 878         { .name     = "swapon",
 879           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 880         { .name     = "symlinkat",
 881           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 882         { .name     = "tgkill",
 883           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 884         { .name     = "tkill",
 885           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 886         { .name     = "umount2", .alias = "umount",
 887           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 888         { .name     = "uname", .alias = "newuname", },
 889         { .name     = "unlinkat",
 890           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 891         { .name     = "utimensat",
 892           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 893         { .name     = "wait4",      .errpid = true,
 894           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 895         { .name     = "waitid",     .errpid = true,
 896           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 897 };
 898
 899 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 900 {
 901         const struct syscall_fmt *fmt = fmtp;
 902         return strcmp(name, fmt->name);
 903 }
 904
 905 static struct syscall_fmt *syscall_fmt__find(const char *name)
 906 {
 907         const int nmemb = ARRAY_SIZE(syscall_fmts);
 908         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 909 }
 910
 911 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 912 {
 913         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 914
 915         for (i = 0; i < nmemb; ++i) {
 916                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 917                         return &syscall_fmts[i];
 918         }
 919
 920         return NULL;
 921 }
 922
 923 /*
 924  * is_exit: is this "exit" or "exit_group"?
 925  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 926  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 927  */
 928 struct syscall {
 929         struct tep_event    *tp_format;
 930         int                 nr_args;
 931         int                 args_size;
 932         bool                is_exit;
 933         bool                is_open;
 934         struct tep_format_field *args;
 935         const char          *name;
 936         struct syscall_fmt  *fmt;
 937         struct syscall_arg_fmt *arg_fmt;
 938 };
 939
 940 struct bpf_map_syscall_entry {
 941         bool    enabled;
 942 };
 943
 944 /*
 945  * We need to have this 'calculated' boolean because in some cases we really
 946  * don't know what is the duration of a syscall, for instance, when we start
 947  * a session and some threads are waiting for a syscall to finish, say 'poll',
 948  * in which case all we can do is to print "( ? ) for duration and for the
 949  * start timestamp.
 950  */
 951 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 952 {
 953         double duration = (double)t / NSEC_PER_MSEC;
 954         size_t printed = fprintf(fp, "(");
 955
 956         if (!calculated)
 957                 printed += fprintf(fp, "         ");
 958         else if (duration >= 1.0)
 959                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 960         else if (duration >= 0.01)
 961                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 962         else
 963                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 964         return printed + fprintf(fp, "): ");
 965 }
 966
 967 /**
 968  * filename.ptr: The filename char pointer that will be vfs_getname'd
 969  * filename.entry_str_pos: Where to insert the string translated from
 970  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 971  * ret_scnprintf: syscall args may set this to a different syscall return
 972  *                formatter, for instance, fcntl may return fds, file flags, etc.
 973  */
 974 struct thread_trace {
 975         u64               entry_time;
 976         bool              entry_pending;
 977         unsigned long     nr_events;
 978         unsigned long     pfmaj, pfmin;
 979         char              *entry_str;
 980         double            runtime_ms;
 981         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 982         struct {
 983                 unsigned long ptr;
 984                 short int     entry_str_pos;
 985                 bool          pending_open;
 986                 unsigned int  namelen;
 987                 char          *name;
 988         } filename;
 989         struct {
 990                 int           max;
 991                 struct file   *table;
 992         } files;
 993
 994         struct intlist *syscall_stats;
 995 };
 996
 997 static struct thread_trace *thread_trace__new(void)
 998 {
 999         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1000
1001         if (ttrace)
1002                 ttrace->files.max = -1;
1003
1004         ttrace->syscall_stats = intlist__new(NULL);
1005
1006         return ttrace;
1007 }
1008
1009 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1010 {
1011         struct thread_trace *ttrace;
1012
1013         if (thread == NULL)
1014                 goto fail;
1015
1016         if (thread__priv(thread) == NULL)
1017                 thread__set_priv(thread, thread_trace__new());
1018
1019         if (thread__priv(thread) == NULL)
1020                 goto fail;
1021
1022         ttrace = thread__priv(thread);
1023         ++ttrace->nr_events;
1024
1025         return ttrace;
1026 fail:
1027         color_fprintf(fp, PERF_COLOR_RED,
1028                       "WARNING: not enough memory, dropping samples!\n");
1029         return NULL;
1030 }
1031
1032
1033 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1034                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1035 {
1036         struct thread_trace *ttrace = thread__priv(arg->thread);
1037
1038         ttrace->ret_scnprintf = ret_scnprintf;
1039 }
1040
1041 #define TRACE_PFMAJ             (1 << 0)
1042 #define TRACE_PFMIN             (1 << 1)
1043
1044 static const size_t trace__entry_str_size = 2048;
1045
1046 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1047 {
1048         if (fd < 0)
1049                 return NULL;
1050
1051         if (fd > ttrace->files.max) {
1052                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1053
1054                 if (nfiles == NULL)
1055                         return NULL;
1056
1057                 if (ttrace->files.max != -1) {
1058                         memset(nfiles + ttrace->files.max + 1, 0,
1059                                (fd - ttrace->files.max) * sizeof(struct file));
1060                 } else {
1061                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1062                 }
1063
1064                 ttrace->files.table = nfiles;
1065                 ttrace->files.max   = fd;
1066         }
1067
1068         return ttrace->files.table + fd;
1069 }
1070
1071 struct file *thread__files_entry(struct thread *thread, int fd)
1072 {
1073         return thread_trace__files_entry(thread__priv(thread), fd);
1074 }
1075
1076 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1077 {
1078         struct thread_trace *ttrace = thread__priv(thread);
1079         struct file *file = thread_trace__files_entry(ttrace, fd);
1080
1081         if (file != NULL) {
1082                 struct stat st;
1083                 if (stat(pathname, &st) == 0)
1084                         file->dev_maj = major(st.st_rdev);
1085                 file->pathname = strdup(pathname);
1086                 if (file->pathname)
1087                         return 0;
1088         }
1089
1090         return -1;
1091 }
1092
1093 static int thread__read_fd_path(struct thread *thread, int fd)
1094 {
1095         char linkname[PATH_MAX], pathname[PATH_MAX];
1096         struct stat st;
1097         int ret;
1098
1099         if (thread->pid_ == thread->tid) {
1100                 scnprintf(linkname, sizeof(linkname),
1101                           "/proc/%d/fd/%d", thread->pid_, fd);
1102         } else {
1103                 scnprintf(linkname, sizeof(linkname),
1104                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1105         }
1106
1107         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1108                 return -1;
1109
1110         ret = readlink(linkname, pathname, sizeof(pathname));
1111
1112         if (ret < 0 || ret > st.st_size)
1113                 return -1;
1114
1115         pathname[ret] = '\0';
1116         return trace__set_fd_pathname(thread, fd, pathname);
1117 }
1118
1119 static const char *thread__fd_path(struct thread *thread, int fd,
1120                                    struct trace *trace)
1121 {
1122         struct thread_trace *ttrace = thread__priv(thread);
1123
1124         if (ttrace == NULL)
1125                 return NULL;
1126
1127         if (fd < 0)
1128                 return NULL;
1129
1130         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1131                 if (!trace->live)
1132                         return NULL;
1133                 ++trace->stats.proc_getname;
1134                 if (thread__read_fd_path(thread, fd))
1135                         return NULL;
1136         }
1137
1138         return ttrace->files.table[fd].pathname;
1139 }
1140
1141 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1142 {
1143         int fd = arg->val;
1144         size_t printed = scnprintf(bf, size, "%d", fd);
1145         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1146
1147         if (path)
1148                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1149
1150         return printed;
1151 }
1152
1153 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1154 {
1155         size_t printed = scnprintf(bf, size, "%d", fd);
1156         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1157
1158         if (thread) {
1159                 const char *path = thread__fd_path(thread, fd, trace);
1160
1161                 if (path)
1162                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1163
1164                 thread__put(thread);
1165         }
1166
1167         return printed;
1168 }
1169
1170 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1171                                               struct syscall_arg *arg)
1172 {
1173         int fd = arg->val;
1174         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1175         struct thread_trace *ttrace = thread__priv(arg->thread);
1176
1177         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1178                 zfree(&ttrace->files.table[fd].pathname);
1179
1180         return printed;
1181 }
1182
1183 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1184                                      unsigned long ptr)
1185 {
1186         struct thread_trace *ttrace = thread__priv(thread);
1187
1188         ttrace->filename.ptr = ptr;
1189         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1190 }
1191
1192 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1193 {
1194         struct augmented_arg *augmented_arg = arg->augmented.args;
1195
1196         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1197 }
1198
1199 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1200                                               struct syscall_arg *arg)
1201 {
1202         unsigned long ptr = arg->val;
1203
1204         if (arg->augmented.args)
1205                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1206
1207         if (!arg->trace->vfs_getname)
1208                 return scnprintf(bf, size, "%#x", ptr);
1209
1210         thread__set_filename_pos(arg->thread, bf, ptr);
1211         return 0;
1212 }
1213
1214 static bool trace__filter_duration(struct trace *trace, double t)
1215 {
1216         return t < (trace->duration_filter * NSEC_PER_MSEC);
1217 }
1218
1219 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1220 {
1221         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1222
1223         return fprintf(fp, "%10.3f ", ts);
1224 }
1225
1226 /*
1227  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1228  * using ttrace->entry_time for a thread that receives a sys_exit without
1229  * first having received a sys_enter ("poll" issued before tracing session
1230  * starts, lost sys_enter exit due to ring buffer overflow).
1231  */
1232 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1233 {
1234         if (tstamp > 0)
1235                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1236
1237         return fprintf(fp, "         ? ");
1238 }
1239
1240 static bool done = false;
1241 static bool interrupted = false;
1242
1243 static void sig_handler(int sig)
1244 {
1245         done = true;
1246         interrupted = sig == SIGINT;
1247 }
1248
1249 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1250 {
1251         size_t printed = 0;
1252
1253         if (trace->multiple_threads) {
1254                 if (trace->show_comm)
1255                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1256                 printed += fprintf(fp, "%d ", thread->tid);
1257         }
1258
1259         return printed;
1260 }
1261
1262 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1263                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1264 {
1265         size_t printed = 0;
1266
1267         if (trace->show_tstamp)
1268                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1269         if (trace->show_duration)
1270                 printed += fprintf_duration(duration, duration_calculated, fp);
1271         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1272 }
1273
1274 static int trace__process_event(struct trace *trace, struct machine *machine,
1275                                 union perf_event *event, struct perf_sample *sample)
1276 {
1277         int ret = 0;
1278
1279         switch (event->header.type) {
1280         case PERF_RECORD_LOST:
1281                 color_fprintf(trace->output, PERF_COLOR_RED,
1282                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1283                 ret = machine__process_lost_event(machine, event, sample);
1284                 break;
1285         default:
1286                 ret = machine__process_event(machine, event, sample);
1287                 break;
1288         }
1289
1290         return ret;
1291 }
1292
1293 static int trace__tool_process(struct perf_tool *tool,
1294                                union perf_event *event,
1295                                struct perf_sample *sample,
1296                                struct machine *machine)
1297 {
1298         struct trace *trace = container_of(tool, struct trace, tool);
1299         return trace__process_event(trace, machine, event, sample);
1300 }
1301
1302 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1303 {
1304         struct machine *machine = vmachine;
1305
1306         if (machine->kptr_restrict_warned)
1307                 return NULL;
1308
1309         if (symbol_conf.kptr_restrict) {
1310                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1311                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1312                            "Kernel samples will not be resolved.\n");
1313                 machine->kptr_restrict_warned = true;
1314                 return NULL;
1315         }
1316
1317         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1318 }
1319
1320 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1321 {
1322         int err = symbol__init(NULL);
1323
1324         if (err)
1325                 return err;
1326
1327         trace->host = machine__new_host();
1328         if (trace->host == NULL)
1329                 return -ENOMEM;
1330
1331         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1332         if (err < 0)
1333                 goto out;
1334
1335         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1336                                             evlist->threads, trace__tool_process, false,
1337                                             1);
1338 out:
1339         if (err)
1340                 symbol__exit();
1341
1342         return err;
1343 }
1344
1345 static void trace__symbols__exit(struct trace *trace)
1346 {
1347         machine__exit(trace->host);
1348         trace->host = NULL;
1349
1350         symbol__exit();
1351 }
1352
1353 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1354 {
1355         int idx;
1356
1357         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1358                 nr_args = sc->fmt->nr_args;
1359
1360         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1361         if (sc->arg_fmt == NULL)
1362                 return -1;
1363
1364         for (idx = 0; idx < nr_args; ++idx) {
1365                 if (sc->fmt)
1366                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1367         }
1368
1369         sc->nr_args = nr_args;
1370         return 0;
1371 }
1372
1373 static int syscall__set_arg_fmts(struct syscall *sc)
1374 {
1375         struct tep_format_field *field, *last_field = NULL;
1376         int idx = 0, len;
1377
1378         for (field = sc->args; field; field = field->next, ++idx) {
1379                 last_field = field;
1380
1381                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1382                         continue;
1383
1384                 if (strcmp(field->type, "const char *") == 0 &&
1385                          (strcmp(field->name, "filename") == 0 ||
1386                           strcmp(field->name, "path") == 0 ||
1387                           strcmp(field->name, "pathname") == 0))
1388                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1389                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1390                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1391                 else if (strcmp(field->type, "pid_t") == 0)
1392                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1393                 else if (strcmp(field->type, "umode_t") == 0)
1394                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1395                 else if ((strcmp(field->type, "int") == 0 ||
1396                           strcmp(field->type, "unsigned int") == 0 ||
1397                           strcmp(field->type, "long") == 0) &&
1398                          (len = strlen(field->name)) >= 2 &&
1399                          strcmp(field->name + len - 2, "fd") == 0) {
1400                         /*
1401                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1402                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1403                          * 65 int
1404                          * 23 unsigned int
1405                          * 7 unsigned long
1406                          */
1407                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1408                 }
1409         }
1410
1411         if (last_field)
1412                 sc->args_size = last_field->offset + last_field->size;
1413
1414         return 0;
1415 }
1416
1417 static int trace__read_syscall_info(struct trace *trace, int id)
1418 {
1419         char tp_name[128];
1420         struct syscall *sc;
1421         const char *name = syscalltbl__name(trace->sctbl, id);
1422
1423         if (name == NULL)
1424                 return -1;
1425
1426         if (id > trace->syscalls.max) {
1427                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1428
1429                 if (nsyscalls == NULL)
1430                         return -1;
1431
1432                 if (trace->syscalls.max != -1) {
1433                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1434                                (id - trace->syscalls.max) * sizeof(*sc));
1435                 } else {
1436                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1437                 }
1438
1439                 trace->syscalls.table = nsyscalls;
1440                 trace->syscalls.max   = id;
1441         }
1442
1443         sc = trace->syscalls.table + id;
1444         sc->name = name;
1445
1446         sc->fmt  = syscall_fmt__find(sc->name);
1447
1448         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1449         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1450
1451         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1452                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1453                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1454         }
1455
1456         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1457                 return -1;
1458
1459         if (IS_ERR(sc->tp_format))
1460                 return -1;
1461
1462         sc->args = sc->tp_format->format.fields;
1463         /*
1464          * We need to check and discard the first variable '__syscall_nr'
1465          * or 'nr' that mean the syscall number. It is needless here.
1466          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1467          */
1468         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1469                 sc->args = sc->args->next;
1470                 --sc->nr_args;
1471         }
1472
1473         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1474         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1475
1476         return syscall__set_arg_fmts(sc);
1477 }
1478
1479 static int trace__validate_ev_qualifier(struct trace *trace)
1480 {
1481         int err = 0, i;
1482         size_t nr_allocated;
1483         struct str_node *pos;
1484
1485         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1486         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1487                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1488
1489         if (trace->ev_qualifier_ids.entries == NULL) {
1490                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1491                        trace->output);
1492                 err = -EINVAL;
1493                 goto out;
1494         }
1495
1496         nr_allocated = trace->ev_qualifier_ids.nr;
1497         i = 0;
1498
1499         strlist__for_each_entry(pos, trace->ev_qualifier) {
1500                 const char *sc = pos->s;
1501                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1502
1503                 if (id < 0) {
1504                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1505                         if (id >= 0)
1506                                 goto matches;
1507
1508                         if (err == 0) {
1509                                 fputs("Error:\tInvalid syscall ", trace->output);
1510                                 err = -EINVAL;
1511                         } else {
1512                                 fputs(", ", trace->output);
1513                         }
1514
1515                         fputs(sc, trace->output);
1516                 }
1517 matches:
1518                 trace->ev_qualifier_ids.entries[i++] = id;
1519                 if (match_next == -1)
1520                         continue;
1521
1522                 while (1) {
1523                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1524                         if (id < 0)
1525                                 break;
1526                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1527                                 void *entries;
1528
1529                                 nr_allocated += 8;
1530                                 entries = realloc(trace->ev_qualifier_ids.entries,
1531                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1532                                 if (entries == NULL) {
1533                                         err = -ENOMEM;
1534                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1535                                         goto out_free;
1536                                 }
1537                                 trace->ev_qualifier_ids.entries = entries;
1538                         }
1539                         trace->ev_qualifier_ids.nr++;
1540                         trace->ev_qualifier_ids.entries[i++] = id;
1541                 }
1542         }
1543
1544         if (err < 0) {
1545                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1546                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1547 out_free:
1548                 zfree(&trace->ev_qualifier_ids.entries);
1549                 trace->ev_qualifier_ids.nr = 0;
1550         }
1551 out:
1552         return err;
1553 }
1554
1555 /*
1556  * args is to be interpreted as a series of longs but we need to handle
1557  * 8-byte unaligned accesses. args points to raw_data within the event
1558  * and raw_data is guaranteed to be 8-byte unaligned because it is
1559  * preceded by raw_size which is a u32. So we need to copy args to a temp
1560  * variable to read it. Most notably this avoids extended load instructions
1561  * on unaligned addresses
1562  */
1563 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1564 {
1565         unsigned long val;
1566         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1567
1568         memcpy(&val, p, sizeof(val));
1569         return val;
1570 }
1571
1572 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1573                                       struct syscall_arg *arg)
1574 {
1575         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1576                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1577
1578         return scnprintf(bf, size, "arg%d: ", arg->idx);
1579 }
1580
1581 /*
1582  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1583  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1584  * in tools/perf/trace/beauty/mount_flags.c
1585  */
1586 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1587 {
1588         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1589                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1590
1591         return val;
1592 }
1593
1594 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1595                                      struct syscall_arg *arg, unsigned long val)
1596 {
1597         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1598                 arg->val = val;
1599                 if (sc->arg_fmt[arg->idx].parm)
1600                         arg->parm = sc->arg_fmt[arg->idx].parm;
1601                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1602         }
1603         return scnprintf(bf, size, "%ld", val);
1604 }
1605
1606 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1607                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1608                                       struct trace *trace, struct thread *thread)
1609 {
1610         size_t printed = 0;
1611         unsigned long val;
1612         u8 bit = 1;
1613         struct syscall_arg arg = {
1614                 .args   = args,
1615                 .augmented = {
1616                         .size = augmented_args_size,
1617                         .args = augmented_args,
1618                 },
1619                 .idx    = 0,
1620                 .mask   = 0,
1621                 .trace  = trace,
1622                 .thread = thread,
1623                 .show_string_prefix = trace->show_string_prefix,
1624         };
1625         struct thread_trace *ttrace = thread__priv(thread);
1626
1627         /*
1628          * Things like fcntl will set this in its 'cmd' formatter to pick the
1629          * right formatter for the return value (an fd? file flags?), which is
1630          * not needed for syscalls that always return a given type, say an fd.
1631          */
1632         ttrace->ret_scnprintf = NULL;
1633
1634         if (sc->args != NULL) {
1635                 struct tep_format_field *field;
1636
1637                 for (field = sc->args; field;
1638                      field = field->next, ++arg.idx, bit <<= 1) {
1639                         if (arg.mask & bit)
1640                                 continue;
1641
1642                         val = syscall_arg__val(&arg, arg.idx);
1643                         /*
1644                          * Some syscall args need some mask, most don't and
1645                          * return val untouched.
1646                          */
1647                         val = syscall__mask_val(sc, &arg, val);
1648
1649                         /*
1650                          * Suppress this argument if its value is zero and
1651                          * and we don't have a string associated in an
1652                          * strarray for it.
1653                          */
1654                         if (val == 0 &&
1655                             !trace->show_zeros &&
1656                             !(sc->arg_fmt &&
1657                               (sc->arg_fmt[arg.idx].show_zero ||
1658                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1659                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1660                               sc->arg_fmt[arg.idx].parm))
1661                                 continue;
1662
1663                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1664
1665                         if (trace->show_arg_names)
1666                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1667
1668                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1669                 }
1670         } else if (IS_ERR(sc->tp_format)) {
1671                 /*
1672                  * If we managed to read the tracepoint /format file, then we
1673                  * may end up not having any args, like with gettid(), so only
1674                  * print the raw args when we didn't manage to read it.
1675                  */
1676                 while (arg.idx < sc->nr_args) {
1677                         if (arg.mask & bit)
1678                                 goto next_arg;
1679                         val = syscall_arg__val(&arg, arg.idx);
1680                         if (printed)
1681                                 printed += scnprintf(bf + printed, size - printed, ", ");
1682                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1683                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1684 next_arg:
1685                         ++arg.idx;
1686                         bit <<= 1;
1687                 }
1688         }
1689
1690         return printed;
1691 }
1692
1693 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1694                                   union perf_event *event,
1695                                   struct perf_sample *sample);
1696
1697 static struct syscall *trace__syscall_info(struct trace *trace,
1698                                            struct perf_evsel *evsel, int id)
1699 {
1700
1701         if (id < 0) {
1702
1703                 /*
1704                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1705                  * before that, leaving at a higher verbosity level till that is
1706                  * explained. Reproduced with plain ftrace with:
1707                  *
1708                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1709                  * grep "NR -1 " /t/trace_pipe
1710                  *
1711                  * After generating some load on the machine.
1712                  */
1713                 if (verbose > 1) {
1714                         static u64 n;
1715                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1716                                 id, perf_evsel__name(evsel), ++n);
1717                 }
1718                 return NULL;
1719         }
1720
1721         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1722             trace__read_syscall_info(trace, id))
1723                 goto out_cant_read;
1724
1725         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1726                 goto out_cant_read;
1727
1728         return &trace->syscalls.table[id];
1729
1730 out_cant_read:
1731         if (verbose > 0) {
1732                 fprintf(trace->output, "Problems reading syscall %d", id);
1733                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1734                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1735                 fputs(" information\n", trace->output);
1736         }
1737         return NULL;
1738 }
1739
1740 static void thread__update_stats(struct thread_trace *ttrace,
1741                                  int id, struct perf_sample *sample)
1742 {
1743         struct int_node *inode;
1744         struct stats *stats;
1745         u64 duration = 0;
1746
1747         inode = intlist__findnew(ttrace->syscall_stats, id);
1748         if (inode == NULL)
1749                 return;
1750
1751         stats = inode->priv;
1752         if (stats == NULL) {
1753                 stats = malloc(sizeof(struct stats));
1754                 if (stats == NULL)
1755                         return;
1756                 init_stats(stats);
1757                 inode->priv = stats;
1758         }
1759
1760         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1761                 duration = sample->time - ttrace->entry_time;
1762
1763         update_stats(stats, duration);
1764 }
1765
1766 static int trace__printf_interrupted_entry(struct trace *trace)
1767 {
1768         struct thread_trace *ttrace;
1769         size_t printed;
1770         int len;
1771
1772         if (trace->failure_only || trace->current == NULL)
1773                 return 0;
1774
1775         ttrace = thread__priv(trace->current);
1776
1777         if (!ttrace->entry_pending)
1778                 return 0;
1779
1780         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1781         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1782
1783         if (len < trace->args_alignment - 4)
1784                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1785
1786         printed += fprintf(trace->output, " ...\n");
1787
1788         ttrace->entry_pending = false;
1789         ++trace->nr_events_printed;
1790
1791         return printed;
1792 }
1793
1794 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1795                                  struct perf_sample *sample, struct thread *thread)
1796 {
1797         int printed = 0;
1798
1799         if (trace->print_sample) {
1800                 double ts = (double)sample->time / NSEC_PER_MSEC;
1801
1802                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1803                                    perf_evsel__name(evsel), ts,
1804                                    thread__comm_str(thread),
1805                                    sample->pid, sample->tid, sample->cpu);
1806         }
1807
1808         return printed;
1809 }
1810
1811 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1812 {
1813         void *augmented_args = NULL;
1814         /*
1815          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1816          * and there we get all 6 syscall args plus the tracepoint common fields
1817          * that gets calculated at the start and the syscall_nr (another long).
1818          * So we check if that is the case and if so don't look after the
1819          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1820          * which is fixed.
1821          *
1822          * We'll revisit this later to pass s->args_size to the BPF augmenter
1823          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1824          * copies only what we need for each syscall, like what happens when we
1825          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1826          * traffic to just what is needed for each syscall.
1827          */
1828         int args_size = raw_augmented_args_size ?: sc->args_size;
1829
1830         *augmented_args_size = sample->raw_size - args_size;
1831         if (*augmented_args_size > 0)
1832                 augmented_args = sample->raw_data + args_size;
1833
1834         return augmented_args;
1835 }
1836
1837 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1838                             union perf_event *event __maybe_unused,
1839                             struct perf_sample *sample)
1840 {
1841         char *msg;
1842         void *args;
1843         int printed = 0;
1844         struct thread *thread;
1845         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1846         int augmented_args_size = 0;
1847         void *augmented_args = NULL;
1848         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1849         struct thread_trace *ttrace;
1850
1851         if (sc == NULL)
1852                 return -1;
1853
1854         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1855         ttrace = thread__trace(thread, trace->output);
1856         if (ttrace == NULL)
1857                 goto out_put;
1858
1859         trace__fprintf_sample(trace, evsel, sample, thread);
1860
1861         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1862
1863         if (ttrace->entry_str == NULL) {
1864                 ttrace->entry_str = malloc(trace__entry_str_size);
1865                 if (!ttrace->entry_str)
1866                         goto out_put;
1867         }
1868
1869         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1870                 trace__printf_interrupted_entry(trace);
1871         /*
1872          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1873          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1874          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1875          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1876          * so when handling, say the openat syscall, we end up getting 6 args for the
1877          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1878          * thinking that the extra 2 u64 args are the augmented filename, so just check
1879          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1880          */
1881         if (evsel != trace->syscalls.events.sys_enter)
1882                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1883         ttrace->entry_time = sample->time;
1884         msg = ttrace->entry_str;
1885         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1886
1887         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1888                                            args, augmented_args, augmented_args_size, trace, thread);
1889
1890         if (sc->is_exit) {
1891                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1892                         int alignment = 0;
1893
1894                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1895                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1896                         if (trace->args_alignment > printed)
1897                                 alignment = trace->args_alignment - printed;
1898                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1899                 }
1900         } else {
1901                 ttrace->entry_pending = true;
1902                 /* See trace__vfs_getname & trace__sys_exit */
1903                 ttrace->filename.pending_open = false;
1904         }
1905
1906         if (trace->current != thread) {
1907                 thread__put(trace->current);
1908                 trace->current = thread__get(thread);
1909         }
1910         err = 0;
1911 out_put:
1912         thread__put(thread);
1913         return err;
1914 }
1915
1916 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1917                                     struct perf_sample *sample)
1918 {
1919         struct thread_trace *ttrace;
1920         struct thread *thread;
1921         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1922         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1923         char msg[1024];
1924         void *args, *augmented_args = NULL;
1925         int augmented_args_size;
1926
1927         if (sc == NULL)
1928                 return -1;
1929
1930         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1931         ttrace = thread__trace(thread, trace->output);
1932         /*
1933          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1934          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1935          */
1936         if (ttrace == NULL)
1937                 goto out_put;
1938
1939         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1940         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1941         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1942         fprintf(trace->output, "%s", msg);
1943         err = 0;
1944 out_put:
1945         thread__put(thread);
1946         return err;
1947 }
1948
1949 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1950                                     struct perf_sample *sample,
1951                                     struct callchain_cursor *cursor)
1952 {
1953         struct addr_location al;
1954         int max_stack = evsel->attr.sample_max_stack ?
1955                         evsel->attr.sample_max_stack :
1956                         trace->max_stack;
1957         int err;
1958
1959         if (machine__resolve(trace->host, &al, sample) < 0)
1960                 return -1;
1961
1962         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1963         addr_location__put(&al);
1964         return err;
1965 }
1966
1967 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1968 {
1969         /* TODO: user-configurable print_opts */
1970         const unsigned int print_opts = EVSEL__PRINT_SYM |
1971                                         EVSEL__PRINT_DSO |
1972                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1973
1974         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1975 }
1976
1977 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1978 {
1979         struct perf_env *env = perf_evsel__env(evsel);
1980         const char *arch_name = perf_env__arch(env);
1981
1982         return arch_syscalls__strerrno(arch_name, err);
1983 }
1984
1985 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1986                            union perf_event *event __maybe_unused,
1987                            struct perf_sample *sample)
1988 {
1989         long ret;
1990         u64 duration = 0;
1991         bool duration_calculated = false;
1992         struct thread *thread;
1993         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
1994         int alignment = trace->args_alignment;
1995         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1996         struct thread_trace *ttrace;
1997
1998         if (sc == NULL)
1999                 return -1;
2000
2001         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2002         ttrace = thread__trace(thread, trace->output);
2003         if (ttrace == NULL)
2004                 goto out_put;
2005
2006         trace__fprintf_sample(trace, evsel, sample, thread);
2007
2008         if (trace->summary)
2009                 thread__update_stats(ttrace, id, sample);
2010
2011         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2012
2013         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2014                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2015                 ttrace->filename.pending_open = false;
2016                 ++trace->stats.vfs_getname;
2017         }
2018
2019         if (ttrace->entry_time) {
2020                 duration = sample->time - ttrace->entry_time;
2021                 if (trace__filter_duration(trace, duration))
2022                         goto out;
2023                 duration_calculated = true;
2024         } else if (trace->duration_filter)
2025                 goto out;
2026
2027         if (sample->callchain) {
2028                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2029                 if (callchain_ret == 0) {
2030                         if (callchain_cursor.nr < trace->min_stack)
2031                                 goto out;
2032                         callchain_ret = 1;
2033                 }
2034         }
2035
2036         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2037                 goto out;
2038
2039         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2040
2041         if (ttrace->entry_pending) {
2042                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2043         } else {
2044                 printed += fprintf(trace->output, " ... [");
2045                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2046                 printed += 9;
2047                 printed += fprintf(trace->output, "]: %s()", sc->name);
2048         }
2049
2050         printed++; /* the closing ')' */
2051
2052         if (alignment > printed)
2053                 alignment -= printed;
2054         else
2055                 alignment = 0;
2056
2057         fprintf(trace->output, ")%*s= ", alignment, " ");
2058
2059         if (sc->fmt == NULL) {
2060                 if (ret < 0)
2061                         goto errno_print;
2062 signed_print:
2063                 fprintf(trace->output, "%ld", ret);
2064         } else if (ret < 0) {
2065 errno_print: {
2066                 char bf[STRERR_BUFSIZE];
2067                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2068                            *e = errno_to_name(evsel, -ret);
2069
2070                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2071         }
2072         } else if (ret == 0 && sc->fmt->timeout)
2073                 fprintf(trace->output, "0 (Timeout)");
2074         else if (ttrace->ret_scnprintf) {
2075                 char bf[1024];
2076                 struct syscall_arg arg = {
2077                         .val    = ret,
2078                         .thread = thread,
2079                         .trace  = trace,
2080                 };
2081                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2082                 ttrace->ret_scnprintf = NULL;
2083                 fprintf(trace->output, "%s", bf);
2084         } else if (sc->fmt->hexret)
2085                 fprintf(trace->output, "%#lx", ret);
2086         else if (sc->fmt->errpid) {
2087                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2088
2089                 if (child != NULL) {
2090                         fprintf(trace->output, "%ld", ret);
2091                         if (child->comm_set)
2092                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2093                         thread__put(child);
2094                 }
2095         } else
2096                 goto signed_print;
2097
2098         fputc('\n', trace->output);
2099
2100         /*
2101          * We only consider an 'event' for the sake of --max-events a non-filtered
2102          * sys_enter + sys_exit and other tracepoint events.
2103          */
2104         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2105                 interrupted = true;
2106
2107         if (callchain_ret > 0)
2108                 trace__fprintf_callchain(trace, sample);
2109         else if (callchain_ret < 0)
2110                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2111 out:
2112         ttrace->entry_pending = false;
2113         err = 0;
2114 out_put:
2115         thread__put(thread);
2116         return err;
2117 }
2118
2119 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2120                               union perf_event *event __maybe_unused,
2121                               struct perf_sample *sample)
2122 {
2123         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2124         struct thread_trace *ttrace;
2125         size_t filename_len, entry_str_len, to_move;
2126         ssize_t remaining_space;
2127         char *pos;
2128         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2129
2130         if (!thread)
2131                 goto out;
2132
2133         ttrace = thread__priv(thread);
2134         if (!ttrace)
2135                 goto out_put;
2136
2137         filename_len = strlen(filename);
2138         if (filename_len == 0)
2139                 goto out_put;
2140
2141         if (ttrace->filename.namelen < filename_len) {
2142                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2143
2144                 if (f == NULL)
2145                         goto out_put;
2146
2147                 ttrace->filename.namelen = filename_len;
2148                 ttrace->filename.name = f;
2149         }
2150
2151         strcpy(ttrace->filename.name, filename);
2152         ttrace->filename.pending_open = true;
2153
2154         if (!ttrace->filename.ptr)
2155                 goto out_put;
2156
2157         entry_str_len = strlen(ttrace->entry_str);
2158         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2159         if (remaining_space <= 0)
2160                 goto out_put;
2161
2162         if (filename_len > (size_t)remaining_space) {
2163                 filename += filename_len - remaining_space;
2164                 filename_len = remaining_space;
2165         }
2166
2167         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2168         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2169         memmove(pos + filename_len, pos, to_move);
2170         memcpy(pos, filename, filename_len);
2171
2172         ttrace->filename.ptr = 0;
2173         ttrace->filename.entry_str_pos = 0;
2174 out_put:
2175         thread__put(thread);
2176 out:
2177         return 0;
2178 }
2179
2180 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2181                                      union perf_event *event __maybe_unused,
2182                                      struct perf_sample *sample)
2183 {
2184         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2185         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2186         struct thread *thread = machine__findnew_thread(trace->host,
2187                                                         sample->pid,
2188                                                         sample->tid);
2189         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2190
2191         if (ttrace == NULL)
2192                 goto out_dump;
2193
2194         ttrace->runtime_ms += runtime_ms;
2195         trace->runtime_ms += runtime_ms;
2196 out_put:
2197         thread__put(thread);
2198         return 0;
2199
2200 out_dump:
2201         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2202                evsel->name,
2203                perf_evsel__strval(evsel, sample, "comm"),
2204                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2205                runtime,
2206                perf_evsel__intval(evsel, sample, "vruntime"));
2207         goto out_put;
2208 }
2209
2210 static int bpf_output__printer(enum binary_printer_ops op,
2211                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2212 {
2213         unsigned char ch = (unsigned char)val;
2214
2215         switch (op) {
2216         case BINARY_PRINT_CHAR_DATA:
2217                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2218         case BINARY_PRINT_DATA_BEGIN:
2219         case BINARY_PRINT_LINE_BEGIN:
2220         case BINARY_PRINT_ADDR:
2221         case BINARY_PRINT_NUM_DATA:
2222         case BINARY_PRINT_NUM_PAD:
2223         case BINARY_PRINT_SEP:
2224         case BINARY_PRINT_CHAR_PAD:
2225         case BINARY_PRINT_LINE_END:
2226         case BINARY_PRINT_DATA_END:
2227         default:
2228                 break;
2229         }
2230
2231         return 0;
2232 }
2233
2234 static void bpf_output__fprintf(struct trace *trace,
2235                                 struct perf_sample *sample)
2236 {
2237         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2238                         bpf_output__printer, NULL, trace->output);
2239         ++trace->nr_events_printed;
2240 }
2241
2242 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2243                                 union perf_event *event __maybe_unused,
2244                                 struct perf_sample *sample)
2245 {
2246         struct thread *thread;
2247         int callchain_ret = 0;
2248         /*
2249          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2250          * this event's max_events having been hit and this is an entry coming
2251          * from the ring buffer that we should discard, since the max events
2252          * have already been considered/printed.
2253          */
2254         if (evsel->disabled)
2255                 return 0;
2256
2257         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2258
2259         if (sample->callchain) {
2260                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2261                 if (callchain_ret == 0) {
2262                         if (callchain_cursor.nr < trace->min_stack)
2263                                 goto out;
2264                         callchain_ret = 1;
2265                 }
2266         }
2267
2268         trace__printf_interrupted_entry(trace);
2269         trace__fprintf_tstamp(trace, sample->time, trace->output);
2270
2271         if (trace->trace_syscalls && trace->show_duration)
2272                 fprintf(trace->output, "(         ): ");
2273
2274         if (thread)
2275                 trace__fprintf_comm_tid(trace, thread, trace->output);
2276
2277         if (evsel == trace->syscalls.events.augmented) {
2278                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2279                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2280
2281                 if (sc) {
2282                         fprintf(trace->output, "%s(", sc->name);
2283                         trace__fprintf_sys_enter(trace, evsel, sample);
2284                         fputc(')', trace->output);
2285                         goto newline;
2286                 }
2287
2288                 /*
2289                  * XXX: Not having the associated syscall info or not finding/adding
2290                  *      the thread should never happen, but if it does...
2291                  *      fall thru and print it as a bpf_output event.
2292                  */
2293         }
2294
2295         fprintf(trace->output, "%s:", evsel->name);
2296
2297         if (perf_evsel__is_bpf_output(evsel)) {
2298                 bpf_output__fprintf(trace, sample);
2299         } else if (evsel->tp_format) {
2300                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2301                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2302                         event_format__fprintf(evsel->tp_format, sample->cpu,
2303                                               sample->raw_data, sample->raw_size,
2304                                               trace->output);
2305                         ++trace->nr_events_printed;
2306
2307                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2308                                 perf_evsel__disable(evsel);
2309                                 perf_evsel__close(evsel);
2310                         }
2311                 }
2312         }
2313
2314 newline:
2315         fprintf(trace->output, "\n");
2316
2317         if (callchain_ret > 0)
2318                 trace__fprintf_callchain(trace, sample);
2319         else if (callchain_ret < 0)
2320                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2321 out:
2322         thread__put(thread);
2323         return 0;
2324 }
2325
2326 static void print_location(FILE *f, struct perf_sample *sample,
2327                            struct addr_location *al,
2328                            bool print_dso, bool print_sym)
2329 {
2330
2331         if ((verbose > 0 || print_dso) && al->map)
2332                 fprintf(f, "%s@", al->map->dso->long_name);
2333
2334         if ((verbose > 0 || print_sym) && al->sym)
2335                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2336                         al->addr - al->sym->start);
2337         else if (al->map)
2338                 fprintf(f, "0x%" PRIx64, al->addr);
2339         else
2340                 fprintf(f, "0x%" PRIx64, sample->addr);
2341 }
2342
2343 static int trace__pgfault(struct trace *trace,
2344                           struct perf_evsel *evsel,
2345                           union perf_event *event __maybe_unused,
2346                           struct perf_sample *sample)
2347 {
2348         struct thread *thread;
2349         struct addr_location al;
2350         char map_type = 'd';
2351         struct thread_trace *ttrace;
2352         int err = -1;
2353         int callchain_ret = 0;
2354
2355         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2356
2357         if (sample->callchain) {
2358                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2359                 if (callchain_ret == 0) {
2360                         if (callchain_cursor.nr < trace->min_stack)
2361                                 goto out_put;
2362                         callchain_ret = 1;
2363                 }
2364         }
2365
2366         ttrace = thread__trace(thread, trace->output);
2367         if (ttrace == NULL)
2368                 goto out_put;
2369
2370         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2371                 ttrace->pfmaj++;
2372         else
2373                 ttrace->pfmin++;
2374
2375         if (trace->summary_only)
2376                 goto out;
2377
2378         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2379
2380         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2381
2382         fprintf(trace->output, "%sfault [",
2383                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2384                 "maj" : "min");
2385
2386         print_location(trace->output, sample, &al, false, true);
2387
2388         fprintf(trace->output, "] => ");
2389
2390         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2391
2392         if (!al.map) {
2393                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2394
2395                 if (al.map)
2396                         map_type = 'x';
2397                 else
2398                         map_type = '?';
2399         }
2400
2401         print_location(trace->output, sample, &al, true, false);
2402
2403         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2404
2405         if (callchain_ret > 0)
2406                 trace__fprintf_callchain(trace, sample);
2407         else if (callchain_ret < 0)
2408                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2409
2410         ++trace->nr_events_printed;
2411 out:
2412         err = 0;
2413 out_put:
2414         thread__put(thread);
2415         return err;
2416 }
2417
2418 static void trace__set_base_time(struct trace *trace,
2419                                  struct perf_evsel *evsel,
2420                                  struct perf_sample *sample)
2421 {
2422         /*
2423          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2424          * and don't use sample->time unconditionally, we may end up having
2425          * some other event in the future without PERF_SAMPLE_TIME for good
2426          * reason, i.e. we may not be interested in its timestamps, just in
2427          * it taking place, picking some piece of information when it
2428          * appears in our event stream (vfs_getname comes to mind).
2429          */
2430         if (trace->base_time == 0 && !trace->full_time &&
2431             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2432                 trace->base_time = sample->time;
2433 }
2434
2435 static int trace__process_sample(struct perf_tool *tool,
2436                                  union perf_event *event,
2437                                  struct perf_sample *sample,
2438                                  struct perf_evsel *evsel,
2439                                  struct machine *machine __maybe_unused)
2440 {
2441         struct trace *trace = container_of(tool, struct trace, tool);
2442         struct thread *thread;
2443         int err = 0;
2444
2445         tracepoint_handler handler = evsel->handler;
2446
2447         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2448         if (thread && thread__is_filtered(thread))
2449                 goto out;
2450
2451         trace__set_base_time(trace, evsel, sample);
2452
2453         if (handler) {
2454                 ++trace->nr_events;
2455                 handler(trace, evsel, event, sample);
2456         }
2457 out:
2458         thread__put(thread);
2459         return err;
2460 }
2461
2462 static int trace__record(struct trace *trace, int argc, const char **argv)
2463 {
2464         unsigned int rec_argc, i, j;
2465         const char **rec_argv;
2466         const char * const record_args[] = {
2467                 "record",
2468                 "-R",
2469                 "-m", "1024",
2470                 "-c", "1",
2471         };
2472
2473         const char * const sc_args[] = { "-e", };
2474         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2475         const char * const majpf_args[] = { "-e", "major-faults" };
2476         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2477         const char * const minpf_args[] = { "-e", "minor-faults" };
2478         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2479
2480         /* +1 is for the event string below */
2481         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2482                 majpf_args_nr + minpf_args_nr + argc;
2483         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2484
2485         if (rec_argv == NULL)
2486                 return -ENOMEM;
2487
2488         j = 0;
2489         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2490                 rec_argv[j++] = record_args[i];
2491
2492         if (trace->trace_syscalls) {
2493                 for (i = 0; i < sc_args_nr; i++)
2494                         rec_argv[j++] = sc_args[i];
2495
2496                 /* event string may be different for older kernels - e.g., RHEL6 */
2497                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2498                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2499                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2500                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2501                 else {
2502                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2503                         free(rec_argv);
2504                         return -1;
2505                 }
2506         }
2507
2508         if (trace->trace_pgfaults & TRACE_PFMAJ)
2509                 for (i = 0; i < majpf_args_nr; i++)
2510                         rec_argv[j++] = majpf_args[i];
2511
2512         if (trace->trace_pgfaults & TRACE_PFMIN)
2513                 for (i = 0; i < minpf_args_nr; i++)
2514                         rec_argv[j++] = minpf_args[i];
2515
2516         for (i = 0; i < (unsigned int)argc; i++)
2517                 rec_argv[j++] = argv[i];
2518
2519         return cmd_record(j, rec_argv);
2520 }
2521
2522 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2523
2524 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2525 {
2526         bool found = false;
2527         struct perf_evsel *evsel, *tmp;
2528         struct parse_events_error err = { .idx = 0, };
2529         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2530
2531         if (ret)
2532                 return false;
2533
2534         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2535                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2536                         continue;
2537
2538                 if (perf_evsel__field(evsel, "pathname")) {
2539                         evsel->handler = trace__vfs_getname;
2540                         found = true;
2541                         continue;
2542                 }
2543
2544                 list_del_init(&evsel->node);
2545                 evsel->evlist = NULL;
2546                 perf_evsel__delete(evsel);
2547         }
2548
2549         return found;
2550 }
2551
2552 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2553 {
2554         struct perf_evsel *evsel;
2555         struct perf_event_attr attr = {
2556                 .type = PERF_TYPE_SOFTWARE,
2557                 .mmap_data = 1,
2558         };
2559
2560         attr.config = config;
2561         attr.sample_period = 1;
2562
2563         event_attr_init(&attr);
2564
2565         evsel = perf_evsel__new(&attr);
2566         if (evsel)
2567                 evsel->handler = trace__pgfault;
2568
2569         return evsel;
2570 }
2571
2572 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2573 {
2574         const u32 type = event->header.type;
2575         struct perf_evsel *evsel;
2576
2577         if (type != PERF_RECORD_SAMPLE) {
2578                 trace__process_event(trace, trace->host, event, sample);
2579                 return;
2580         }
2581
2582         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2583         if (evsel == NULL) {
2584                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2585                 return;
2586         }
2587
2588         trace__set_base_time(trace, evsel, sample);
2589
2590         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2591             sample->raw_data == NULL) {
2592                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2593                        perf_evsel__name(evsel), sample->tid,
2594                        sample->cpu, sample->raw_size);
2595         } else {
2596                 tracepoint_handler handler = evsel->handler;
2597                 handler(trace, evsel, event, sample);
2598         }
2599
2600         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2601                 interrupted = true;
2602 }
2603
2604 static int trace__add_syscall_newtp(struct trace *trace)
2605 {
2606         int ret = -1;
2607         struct perf_evlist *evlist = trace->evlist;
2608         struct perf_evsel *sys_enter, *sys_exit;
2609
2610         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2611         if (sys_enter == NULL)
2612                 goto out;
2613
2614         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2615                 goto out_delete_sys_enter;
2616
2617         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2618         if (sys_exit == NULL)
2619                 goto out_delete_sys_enter;
2620
2621         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2622                 goto out_delete_sys_exit;
2623
2624         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2625         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2626
2627         perf_evlist__add(evlist, sys_enter);
2628         perf_evlist__add(evlist, sys_exit);
2629
2630         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2631                 /*
2632                  * We're interested only in the user space callchain
2633                  * leading to the syscall, allow overriding that for
2634                  * debugging reasons using --kernel_syscall_callchains
2635                  */
2636                 sys_exit->attr.exclude_callchain_kernel = 1;
2637         }
2638
2639         trace->syscalls.events.sys_enter = sys_enter;
2640         trace->syscalls.events.sys_exit  = sys_exit;
2641
2642         ret = 0;
2643 out:
2644         return ret;
2645
2646 out_delete_sys_exit:
2647         perf_evsel__delete_priv(sys_exit);
2648 out_delete_sys_enter:
2649         perf_evsel__delete_priv(sys_enter);
2650         goto out;
2651 }
2652
2653 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2654 {
2655         int err = -1;
2656         struct perf_evsel *sys_exit;
2657         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2658                                                 trace->ev_qualifier_ids.nr,
2659                                                 trace->ev_qualifier_ids.entries);
2660
2661         if (filter == NULL)
2662                 goto out_enomem;
2663
2664         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2665                                           filter)) {
2666                 sys_exit = trace->syscalls.events.sys_exit;
2667                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2668         }
2669
2670         free(filter);
2671 out:
2672         return err;
2673 out_enomem:
2674         errno = ENOMEM;
2675         goto out;
2676 }
2677
2678 #ifdef HAVE_LIBBPF_SUPPORT
2679 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2680 {
2681         int fd = bpf_map__fd(trace->syscalls.map);
2682         struct bpf_map_syscall_entry value = {
2683                 .enabled = !trace->not_ev_qualifier,
2684         };
2685         int err = 0;
2686         size_t i;
2687
2688         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2689                 int key = trace->ev_qualifier_ids.entries[i];
2690
2691                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2692                 if (err)
2693                         break;
2694         }
2695
2696         return err;
2697 }
2698
2699 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2700 {
2701         int fd = bpf_map__fd(trace->syscalls.map);
2702         struct bpf_map_syscall_entry value = {
2703                 .enabled = enabled,
2704         };
2705         int err = 0, key;
2706
2707         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2708                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2709                 if (err)
2710                         break;
2711         }
2712
2713         return err;
2714 }
2715
2716 static int trace__init_syscalls_bpf_map(struct trace *trace)
2717 {
2718         bool enabled = true;
2719
2720         if (trace->ev_qualifier_ids.nr)
2721                 enabled = trace->not_ev_qualifier;
2722
2723         return __trace__init_syscalls_bpf_map(trace, enabled);
2724 }
2725 #else
2726 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2727 {
2728         return 0;
2729 }
2730
2731 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2732 {
2733         return 0;
2734 }
2735 #endif // HAVE_LIBBPF_SUPPORT
2736
2737 static int trace__set_ev_qualifier_filter(struct trace *trace)
2738 {
2739         if (trace->syscalls.map)
2740                 return trace__set_ev_qualifier_bpf_filter(trace);
2741         if (trace->syscalls.events.sys_enter)
2742                 return trace__set_ev_qualifier_tp_filter(trace);
2743         return 0;
2744 }
2745
2746 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2747                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2748 {
2749         int err = 0;
2750 #ifdef HAVE_LIBBPF_SUPPORT
2751         bool value = true;
2752         int map_fd = bpf_map__fd(map);
2753         size_t i;
2754
2755         for (i = 0; i < npids; ++i) {
2756                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2757                 if (err)
2758                         break;
2759         }
2760 #endif
2761         return err;
2762 }
2763
2764 static int trace__set_filter_loop_pids(struct trace *trace)
2765 {
2766         unsigned int nr = 1, err;
2767         pid_t pids[32] = {
2768                 getpid(),
2769         };
2770         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2771
2772         while (thread && nr < ARRAY_SIZE(pids)) {
2773                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2774
2775                 if (parent == NULL)
2776                         break;
2777
2778                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2779                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2780                         pids[nr++] = parent->tid;
2781                         break;
2782                 }
2783                 thread = parent;
2784         }
2785
2786         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2787         if (!err && trace->filter_pids.map)
2788                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2789
2790         return err;
2791 }
2792
2793 static int trace__set_filter_pids(struct trace *trace)
2794 {
2795         int err = 0;
2796         /*
2797          * Better not use !target__has_task() here because we need to cover the
2798          * case where no threads were specified in the command line, but a
2799          * workload was, and in that case we will fill in the thread_map when
2800          * we fork the workload in perf_evlist__prepare_workload.
2801          */
2802         if (trace->filter_pids.nr > 0) {
2803                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2804                                                       trace->filter_pids.entries);
2805                 if (!err && trace->filter_pids.map) {
2806                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2807                                                        trace->filter_pids.entries);
2808                 }
2809         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2810                 err = trace__set_filter_loop_pids(trace);
2811         }
2812
2813         return err;
2814 }
2815
2816 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2817 {
2818         struct perf_evlist *evlist = trace->evlist;
2819         struct perf_sample sample;
2820         int err;
2821
2822         err = perf_evlist__parse_sample(evlist, event, &sample);
2823         if (err)
2824                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2825         else
2826                 trace__handle_event(trace, event, &sample);
2827
2828         return 0;
2829 }
2830
2831 static int __trace__flush_events(struct trace *trace)
2832 {
2833         u64 first = ordered_events__first_time(&trace->oe.data);
2834         u64 flush = trace->oe.last - NSEC_PER_SEC;
2835
2836         /* Is there some thing to flush.. */
2837         if (first && first < flush)
2838                 return ordered_events__flush_time(&trace->oe.data, flush);
2839
2840         return 0;
2841 }
2842
2843 static int trace__flush_events(struct trace *trace)
2844 {
2845         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2846 }
2847
2848 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2849 {
2850         int err;
2851
2852         if (!trace->sort_events)
2853                 return __trace__deliver_event(trace, event);
2854
2855         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2856         if (err && err != -1)
2857                 return err;
2858
2859         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2860         if (err)
2861                 return err;
2862
2863         return trace__flush_events(trace);
2864 }
2865
2866 static int ordered_events__deliver_event(struct ordered_events *oe,
2867                                          struct ordered_event *event)
2868 {
2869         struct trace *trace = container_of(oe, struct trace, oe.data);
2870
2871         return __trace__deliver_event(trace, event->event);
2872 }
2873
2874 static int trace__run(struct trace *trace, int argc, const char **argv)
2875 {
2876         struct perf_evlist *evlist = trace->evlist;
2877         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2878         int err = -1, i;
2879         unsigned long before;
2880         const bool forks = argc > 0;
2881         bool draining = false;
2882
2883         trace->live = true;
2884
2885         if (!trace->raw_augmented_syscalls) {
2886                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2887                         goto out_error_raw_syscalls;
2888
2889                 if (trace->trace_syscalls)
2890                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2891         }
2892
2893         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2894                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2895                 if (pgfault_maj == NULL)
2896                         goto out_error_mem;
2897                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2898                 perf_evlist__add(evlist, pgfault_maj);
2899         }
2900
2901         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2902                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2903                 if (pgfault_min == NULL)
2904                         goto out_error_mem;
2905                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2906                 perf_evlist__add(evlist, pgfault_min);
2907         }
2908
2909         if (trace->sched &&
2910             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2911                                    trace__sched_stat_runtime))
2912                 goto out_error_sched_stat_runtime;
2913
2914         /*
2915          * If a global cgroup was set, apply it to all the events without an
2916          * explicit cgroup. I.e.:
2917          *
2918          *      trace -G A -e sched:*switch
2919          *
2920          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2921          * _and_ sched:sched_switch to the 'A' cgroup, while:
2922          *
2923          * trace -e sched:*switch -G A
2924          *
2925          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2926          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2927          * a cgroup (on the root cgroup, sys wide, etc).
2928          *
2929          * Multiple cgroups:
2930          *
2931          * trace -G A -e sched:*switch -G B
2932          *
2933          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2934          * to the 'B' cgroup.
2935          *
2936          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2937          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2938          */
2939         if (trace->cgroup)
2940                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2941
2942         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2943         if (err < 0) {
2944                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2945                 goto out_delete_evlist;
2946         }
2947
2948         err = trace__symbols_init(trace, evlist);
2949         if (err < 0) {
2950                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2951                 goto out_delete_evlist;
2952         }
2953
2954         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2955
2956         signal(SIGCHLD, sig_handler);
2957         signal(SIGINT, sig_handler);
2958
2959         if (forks) {
2960                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2961                                                     argv, false, NULL);
2962                 if (err < 0) {
2963                         fprintf(trace->output, "Couldn't run the workload!\n");
2964                         goto out_delete_evlist;
2965                 }
2966         }
2967
2968         err = perf_evlist__open(evlist);
2969         if (err < 0)
2970                 goto out_error_open;
2971
2972         err = bpf__apply_obj_config();
2973         if (err) {
2974                 char errbuf[BUFSIZ];
2975
2976                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2977                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2978                          errbuf);
2979                 goto out_error_open;
2980         }
2981
2982         err = trace__set_filter_pids(trace);
2983         if (err < 0)
2984                 goto out_error_mem;
2985
2986         if (trace->syscalls.map)
2987                 trace__init_syscalls_bpf_map(trace);
2988
2989         if (trace->ev_qualifier_ids.nr > 0) {
2990                 err = trace__set_ev_qualifier_filter(trace);
2991                 if (err < 0)
2992                         goto out_errno;
2993
2994                 if (trace->syscalls.events.sys_exit) {
2995                         pr_debug("event qualifier tracepoint filter: %s\n",
2996                                  trace->syscalls.events.sys_exit->filter);
2997                 }
2998         }
2999
3000         err = perf_evlist__apply_filters(evlist, &evsel);
3001         if (err < 0)
3002                 goto out_error_apply_filters;
3003
3004         if (trace->dump.map)
3005                 bpf_map__fprintf(trace->dump.map, trace->output);
3006
3007         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3008         if (err < 0)
3009                 goto out_error_mmap;
3010
3011         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3012                 perf_evlist__enable(evlist);
3013
3014         if (forks)
3015                 perf_evlist__start_workload(evlist);
3016
3017         if (trace->opts.initial_delay) {
3018                 usleep(trace->opts.initial_delay * 1000);
3019                 perf_evlist__enable(evlist);
3020         }
3021
3022         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3023                                   evlist->threads->nr > 1 ||
3024                                   perf_evlist__first(evlist)->attr.inherit;
3025
3026         /*
3027          * Now that we already used evsel->attr to ask the kernel to setup the
3028          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3029          * trace__resolve_callchain(), allowing per-event max-stack settings
3030          * to override an explicitly set --max-stack global setting.
3031          */
3032         evlist__for_each_entry(evlist, evsel) {
3033                 if (evsel__has_callchain(evsel) &&
3034                     evsel->attr.sample_max_stack == 0)
3035                         evsel->attr.sample_max_stack = trace->max_stack;
3036         }
3037 again:
3038         before = trace->nr_events;
3039
3040         for (i = 0; i < evlist->nr_mmaps; i++) {
3041                 union perf_event *event;
3042                 struct perf_mmap *md;
3043
3044                 md = &evlist->mmap[i];
3045                 if (perf_mmap__read_init(md) < 0)
3046                         continue;
3047
3048                 while ((event = perf_mmap__read_event(md)) != NULL) {
3049                         ++trace->nr_events;
3050
3051                         err = trace__deliver_event(trace, event);
3052                         if (err)
3053                                 goto out_disable;
3054
3055                         perf_mmap__consume(md);
3056
3057                         if (interrupted)
3058                                 goto out_disable;
3059
3060                         if (done && !draining) {
3061                                 perf_evlist__disable(evlist);
3062                                 draining = true;
3063                         }
3064                 }
3065                 perf_mmap__read_done(md);
3066         }
3067
3068         if (trace->nr_events == before) {
3069                 int timeout = done ? 100 : -1;
3070
3071                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3072                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3073                                 draining = true;
3074
3075                         goto again;
3076                 } else {
3077                         if (trace__flush_events(trace))
3078                                 goto out_disable;
3079                 }
3080         } else {
3081                 goto again;
3082         }
3083
3084 out_disable:
3085         thread__zput(trace->current);
3086
3087         perf_evlist__disable(evlist);
3088
3089         if (trace->sort_events)
3090                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3091
3092         if (!err) {
3093                 if (trace->summary)
3094                         trace__fprintf_thread_summary(trace, trace->output);
3095
3096                 if (trace->show_tool_stats) {
3097                         fprintf(trace->output, "Stats:\n "
3098                                                " vfs_getname : %" PRIu64 "\n"
3099                                                " proc_getname: %" PRIu64 "\n",
3100                                 trace->stats.vfs_getname,
3101                                 trace->stats.proc_getname);
3102                 }
3103         }
3104
3105 out_delete_evlist:
3106         trace__symbols__exit(trace);
3107
3108         perf_evlist__delete(evlist);
3109         cgroup__put(trace->cgroup);
3110         trace->evlist = NULL;
3111         trace->live = false;
3112         return err;
3113 {
3114         char errbuf[BUFSIZ];
3115
3116 out_error_sched_stat_runtime:
3117         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3118         goto out_error;
3119
3120 out_error_raw_syscalls:
3121         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3122         goto out_error;
3123
3124 out_error_mmap:
3125         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3126         goto out_error;
3127
3128 out_error_open:
3129         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3130
3131 out_error:
3132         fprintf(trace->output, "%s\n", errbuf);
3133         goto out_delete_evlist;
3134
3135 out_error_apply_filters:
3136         fprintf(trace->output,
3137                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3138                 evsel->filter, perf_evsel__name(evsel), errno,
3139                 str_error_r(errno, errbuf, sizeof(errbuf)));
3140         goto out_delete_evlist;
3141 }
3142 out_error_mem:
3143         fprintf(trace->output, "Not enough memory to run!\n");
3144         goto out_delete_evlist;
3145
3146 out_errno:
3147         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3148         goto out_delete_evlist;
3149 }
3150
3151 static int trace__replay(struct trace *trace)
3152 {
3153         const struct perf_evsel_str_handler handlers[] = {
3154                 { "probe:vfs_getname",       trace__vfs_getname, },
3155         };
3156         struct perf_data data = {
3157                 .path  = input_name,
3158                 .mode  = PERF_DATA_MODE_READ,
3159                 .force = trace->force,
3160         };
3161         struct perf_session *session;
3162         struct perf_evsel *evsel;
3163         int err = -1;
3164
3165         trace->tool.sample        = trace__process_sample;
3166         trace->tool.mmap          = perf_event__process_mmap;
3167         trace->tool.mmap2         = perf_event__process_mmap2;
3168         trace->tool.comm          = perf_event__process_comm;
3169         trace->tool.exit          = perf_event__process_exit;
3170         trace->tool.fork          = perf_event__process_fork;
3171         trace->tool.attr          = perf_event__process_attr;
3172         trace->tool.tracing_data  = perf_event__process_tracing_data;
3173         trace->tool.build_id      = perf_event__process_build_id;
3174         trace->tool.namespaces    = perf_event__process_namespaces;
3175
3176         trace->tool.ordered_events = true;
3177         trace->tool.ordering_requires_timestamps = true;
3178
3179         /* add tid to output */
3180         trace->multiple_threads = true;
3181
3182         session = perf_session__new(&data, false, &trace->tool);
3183         if (session == NULL)
3184                 return -1;
3185
3186         if (trace->opts.target.pid)
3187                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3188
3189         if (trace->opts.target.tid)
3190                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3191
3192         if (symbol__init(&session->header.env) < 0)
3193                 goto out;
3194
3195         trace->host = &session->machines.host;
3196
3197         err = perf_session__set_tracepoints_handlers(session, handlers);
3198         if (err)
3199                 goto out;
3200
3201         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3202                                                      "raw_syscalls:sys_enter");
3203         /* older kernels have syscalls tp versus raw_syscalls */
3204         if (evsel == NULL)
3205                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3206                                                              "syscalls:sys_enter");
3207
3208         if (evsel &&
3209             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3210             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3211                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3212                 goto out;
3213         }
3214
3215         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3216                                                      "raw_syscalls:sys_exit");
3217         if (evsel == NULL)
3218                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3219                                                              "syscalls:sys_exit");
3220         if (evsel &&
3221             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3222             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3223                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3224                 goto out;
3225         }
3226
3227         evlist__for_each_entry(session->evlist, evsel) {
3228                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3229                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3230                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3231                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3232                         evsel->handler = trace__pgfault;
3233         }
3234
3235         setup_pager();
3236
3237         err = perf_session__process_events(session);
3238         if (err)
3239                 pr_err("Failed to process events, error %d", err);
3240
3241         else if (trace->summary)
3242                 trace__fprintf_thread_summary(trace, trace->output);
3243
3244 out:
3245         perf_session__delete(session);
3246
3247         return err;
3248 }
3249
3250 static size_t trace__fprintf_threads_header(FILE *fp)
3251 {
3252         size_t printed;
3253
3254         printed  = fprintf(fp, "\n Summary of events:\n\n");
3255
3256         return printed;
3257 }
3258
3259 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3260         struct stats    *stats;
3261         double          msecs;
3262         int             syscall;
3263 )
3264 {
3265         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3266         struct stats *stats = source->priv;
3267
3268         entry->syscall = source->i;
3269         entry->stats   = stats;
3270         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3271 }
3272
3273 static size_t thread__dump_stats(struct thread_trace *ttrace,
3274                                  struct trace *trace, FILE *fp)
3275 {
3276         size_t printed = 0;
3277         struct syscall *sc;
3278         struct rb_node *nd;
3279         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3280
3281         if (syscall_stats == NULL)
3282                 return 0;
3283
3284         printed += fprintf(fp, "\n");
3285
3286         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3287         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3288         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3289
3290         resort_rb__for_each_entry(nd, syscall_stats) {
3291                 struct stats *stats = syscall_stats_entry->stats;
3292                 if (stats) {
3293                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3294                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3295                         double avg = avg_stats(stats);
3296                         double pct;
3297                         u64 n = (u64) stats->n;
3298
3299                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3300                         avg /= NSEC_PER_MSEC;
3301
3302                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3303                         printed += fprintf(fp, "   %-15s", sc->name);
3304                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3305                                            n, syscall_stats_entry->msecs, min, avg);
3306                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3307                 }
3308         }
3309
3310         resort_rb__delete(syscall_stats);
3311         printed += fprintf(fp, "\n\n");
3312
3313         return printed;
3314 }
3315
3316 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3317 {
3318         size_t printed = 0;
3319         struct thread_trace *ttrace = thread__priv(thread);
3320         double ratio;
3321
3322         if (ttrace == NULL)
3323                 return 0;
3324
3325         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3326
3327         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3328         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3329         printed += fprintf(fp, "%.1f%%", ratio);
3330         if (ttrace->pfmaj)
3331                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3332         if (ttrace->pfmin)
3333                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3334         if (trace->sched)
3335                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3336         else if (fputc('\n', fp) != EOF)
3337                 ++printed;
3338
3339         printed += thread__dump_stats(ttrace, trace, fp);
3340
3341         return printed;
3342 }
3343
3344 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3345 {
3346         return ttrace ? ttrace->nr_events : 0;
3347 }
3348
3349 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3350         struct thread *thread;
3351 )
3352 {
3353         entry->thread = rb_entry(nd, struct thread, rb_node);
3354 }
3355
3356 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3357 {
3358         size_t printed = trace__fprintf_threads_header(fp);
3359         struct rb_node *nd;
3360         int i;
3361
3362         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3363                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3364
3365                 if (threads == NULL) {
3366                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3367                         return 0;
3368                 }
3369
3370                 resort_rb__for_each_entry(nd, threads)
3371                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3372
3373                 resort_rb__delete(threads);
3374         }
3375         return printed;
3376 }
3377
3378 static int trace__set_duration(const struct option *opt, const char *str,
3379                                int unset __maybe_unused)
3380 {
3381         struct trace *trace = opt->value;
3382
3383         trace->duration_filter = atof(str);
3384         return 0;
3385 }
3386
3387 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3388                                               int unset __maybe_unused)
3389 {
3390         int ret = -1;
3391         size_t i;
3392         struct trace *trace = opt->value;
3393         /*
3394          * FIXME: introduce a intarray class, plain parse csv and create a
3395          * { int nr, int entries[] } struct...
3396          */
3397         struct intlist *list = intlist__new(str);
3398
3399         if (list == NULL)
3400                 return -1;
3401
3402         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3403         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3404
3405         if (trace->filter_pids.entries == NULL)
3406                 goto out;
3407
3408         trace->filter_pids.entries[0] = getpid();
3409
3410         for (i = 1; i < trace->filter_pids.nr; ++i)
3411                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3412
3413         intlist__delete(list);
3414         ret = 0;
3415 out:
3416         return ret;
3417 }
3418
3419 static int trace__open_output(struct trace *trace, const char *filename)
3420 {
3421         struct stat st;
3422
3423         if (!stat(filename, &st) && st.st_size) {
3424                 char oldname[PATH_MAX];
3425
3426                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3427                 unlink(oldname);
3428                 rename(filename, oldname);
3429         }
3430
3431         trace->output = fopen(filename, "w");
3432
3433         return trace->output == NULL ? -errno : 0;
3434 }
3435
3436 static int parse_pagefaults(const struct option *opt, const char *str,
3437                             int unset __maybe_unused)
3438 {
3439         int *trace_pgfaults = opt->value;
3440
3441         if (strcmp(str, "all") == 0)
3442                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3443         else if (strcmp(str, "maj") == 0)
3444                 *trace_pgfaults |= TRACE_PFMAJ;
3445         else if (strcmp(str, "min") == 0)
3446                 *trace_pgfaults |= TRACE_PFMIN;
3447         else
3448                 return -1;
3449
3450         return 0;
3451 }
3452
3453 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3454 {
3455         struct perf_evsel *evsel;
3456
3457         evlist__for_each_entry(evlist, evsel)
3458                 evsel->handler = handler;
3459 }
3460
3461 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3462 {
3463         struct perf_evsel *evsel;
3464
3465         evlist__for_each_entry(evlist, evsel) {
3466                 if (evsel->priv || !evsel->tp_format)
3467                         continue;
3468
3469                 if (strcmp(evsel->tp_format->system, "syscalls"))
3470                         continue;
3471
3472                 if (perf_evsel__init_syscall_tp(evsel))
3473                         return -1;
3474
3475                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3476                         struct syscall_tp *sc = evsel->priv;
3477
3478                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3479                                 return -1;
3480                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3481                         struct syscall_tp *sc = evsel->priv;
3482
3483                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3484                                 return -1;
3485                 }
3486         }
3487
3488         return 0;
3489 }
3490
3491 /*
3492  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3493  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3494  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3495  *
3496  * It'd be better to introduce a parse_options() variant that would return a
3497  * list with the terms it didn't match to an event...
3498  */
3499 static int trace__parse_events_option(const struct option *opt, const char *str,
3500                                       int unset __maybe_unused)
3501 {
3502         struct trace *trace = (struct trace *)opt->value;
3503         const char *s = str;
3504         char *sep = NULL, *lists[2] = { NULL, NULL, };
3505         int len = strlen(str) + 1, err = -1, list, idx;
3506         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3507         char group_name[PATH_MAX];
3508         struct syscall_fmt *fmt;
3509
3510         if (strace_groups_dir == NULL)
3511                 return -1;
3512
3513         if (*s == '!') {
3514                 ++s;
3515                 trace->not_ev_qualifier = true;
3516         }
3517
3518         while (1) {
3519                 if ((sep = strchr(s, ',')) != NULL)
3520                         *sep = '\0';
3521
3522                 list = 0;
3523                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3524                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3525                         list = 1;
3526                         goto do_concat;
3527                 }
3528
3529                 fmt = syscall_fmt__find_by_alias(s);
3530                 if (fmt != NULL) {
3531                         list = 1;
3532                         s = fmt->name;
3533                 } else {
3534                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3535                         if (access(group_name, R_OK) == 0)
3536                                 list = 1;
3537                 }
3538 do_concat:
3539                 if (lists[list]) {
3540                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3541                 } else {
3542                         lists[list] = malloc(len);
3543                         if (lists[list] == NULL)
3544                                 goto out;
3545                         strcpy(lists[list], s);
3546                 }
3547
3548                 if (!sep)
3549                         break;
3550
3551                 *sep = ',';
3552                 s = sep + 1;
3553         }
3554
3555         if (lists[1] != NULL) {
3556                 struct strlist_config slist_config = {
3557                         .dirname = strace_groups_dir,
3558                 };
3559
3560                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3561                 if (trace->ev_qualifier == NULL) {
3562                         fputs("Not enough memory to parse event qualifier", trace->output);
3563                         goto out;
3564                 }
3565
3566                 if (trace__validate_ev_qualifier(trace))
3567                         goto out;
3568                 trace->trace_syscalls = true;
3569         }
3570
3571         err = 0;
3572
3573         if (lists[0]) {
3574                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3575                                                "event selector. use 'perf list' to list available events",
3576                                                parse_events_option);
3577                 err = parse_events_option(&o, lists[0], 0);
3578         }
3579 out:
3580         if (sep)
3581                 *sep = ',';
3582
3583         return err;
3584 }
3585
3586 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3587 {
3588         struct trace *trace = opt->value;
3589
3590         if (!list_empty(&trace->evlist->entries))
3591                 return parse_cgroups(opt, str, unset);
3592
3593         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3594
3595         return 0;
3596 }
3597
3598 static struct bpf_map *bpf__find_map_by_name(const char *name)
3599 {
3600         struct bpf_object *obj, *tmp;
3601
3602         bpf_object__for_each_safe(obj, tmp) {
3603                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3604                 if (map)
3605                         return map;
3606
3607         }
3608
3609         return NULL;
3610 }
3611
3612 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3613 {
3614         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3615 }
3616
3617 static void trace__set_bpf_map_syscalls(struct trace *trace)
3618 {
3619         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3620 }
3621
3622 static int trace__config(const char *var, const char *value, void *arg)
3623 {
3624         struct trace *trace = arg;
3625         int err = 0;
3626
3627         if (!strcmp(var, "trace.add_events")) {
3628                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3629                                                "event selector. use 'perf list' to list available events",
3630                                                parse_events_option);
3631                 err = parse_events_option(&o, value, 0);
3632         } else if (!strcmp(var, "trace.show_timestamp")) {
3633                 trace->show_tstamp = perf_config_bool(var, value);
3634         } else if (!strcmp(var, "trace.show_duration")) {
3635                 trace->show_duration = perf_config_bool(var, value);
3636         } else if (!strcmp(var, "trace.show_arg_names")) {
3637                 trace->show_arg_names = perf_config_bool(var, value);
3638                 if (!trace->show_arg_names)
3639                         trace->show_zeros = true;
3640         } else if (!strcmp(var, "trace.show_zeros")) {
3641                 bool new_show_zeros = perf_config_bool(var, value);
3642                 if (!trace->show_arg_names && !new_show_zeros) {
3643                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3644                         goto out;
3645                 }
3646                 trace->show_zeros = new_show_zeros;
3647         } else if (!strcmp(var, "trace.show_prefix")) {
3648                 trace->show_string_prefix = perf_config_bool(var, value);
3649         } else if (!strcmp(var, "trace.no_inherit")) {
3650                 trace->opts.no_inherit = perf_config_bool(var, value);
3651         } else if (!strcmp(var, "trace.args_alignment")) {
3652                 int args_alignment = 0;
3653                 if (perf_config_int(&args_alignment, var, value) == 0)
3654                         trace->args_alignment = args_alignment;
3655         }
3656 out:
3657         return err;
3658 }
3659
3660 int cmd_trace(int argc, const char **argv)
3661 {
3662         const char *trace_usage[] = {
3663                 "perf trace [<options>] [<command>]",
3664                 "perf trace [<options>] -- <command> [<options>]",
3665                 "perf trace record [<options>] [<command>]",
3666                 "perf trace record [<options>] -- <command> [<options>]",
3667                 NULL
3668         };
3669         struct trace trace = {
3670                 .syscalls = {
3671                         . max = -1,
3672                 },
3673                 .opts = {
3674                         .target = {
3675                                 .uid       = UINT_MAX,
3676                                 .uses_mmap = true,
3677                         },
3678                         .user_freq     = UINT_MAX,
3679                         .user_interval = ULLONG_MAX,
3680                         .no_buffering  = true,
3681                         .mmap_pages    = UINT_MAX,
3682                 },
3683                 .output = stderr,
3684                 .show_comm = true,
3685                 .show_tstamp = true,
3686                 .show_duration = true,
3687                 .show_arg_names = true,
3688                 .args_alignment = 70,
3689                 .trace_syscalls = false,
3690                 .kernel_syscallchains = false,
3691                 .max_stack = UINT_MAX,
3692                 .max_events = ULONG_MAX,
3693         };
3694         const char *map_dump_str = NULL;
3695         const char *output_name = NULL;
3696         const struct option trace_options[] = {
3697         OPT_CALLBACK('e', "event", &trace, "event",
3698                      "event/syscall selector. use 'perf list' to list available events",
3699                      trace__parse_events_option),
3700         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3701                     "show the thread COMM next to its id"),
3702         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3703         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3704                      trace__parse_events_option),
3705         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3706         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3707         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3708                     "trace events on existing process id"),
3709         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3710                     "trace events on existing thread id"),
3711         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3712                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3713         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3714                     "system-wide collection from all CPUs"),
3715         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3716                     "list of cpus to monitor"),
3717         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3718                     "child tasks do not inherit counters"),
3719         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3720                      "number of mmap data pages",
3721                      perf_evlist__parse_mmap_pages),
3722         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3723                    "user to profile"),
3724         OPT_CALLBACK(0, "duration", &trace, "float",
3725                      "show only events with duration > N.M ms",
3726                      trace__set_duration),
3727 #ifdef HAVE_LIBBPF_SUPPORT
3728         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3729 #endif
3730         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3731         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3732         OPT_BOOLEAN('T', "time", &trace.full_time,
3733                     "Show full timestamp, not time relative to first start"),
3734         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3735                     "Show only syscalls that failed"),
3736         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3737                     "Show only syscall summary with statistics"),
3738         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3739                     "Show all syscalls and summary with statistics"),
3740         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3741                      "Trace pagefaults", parse_pagefaults, "maj"),
3742         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3743         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3744         OPT_CALLBACK(0, "call-graph", &trace.opts,
3745                      "record_mode[,record_size]", record_callchain_help,
3746                      &record_parse_callchain_opt),
3747         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3748                     "Show the kernel callchains on the syscall exit path"),
3749         OPT_ULONG(0, "max-events", &trace.max_events,
3750                 "Set the maximum number of events to print, exit after that is reached. "),
3751         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3752                      "Set the minimum stack depth when parsing the callchain, "
3753                      "anything below the specified depth will be ignored."),
3754         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3755                      "Set the maximum stack depth when parsing the callchain, "
3756                      "anything beyond the specified depth will be ignored. "
3757                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3758         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3759                         "Sort batch of events before processing, use if getting out of order events"),
3760         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3761                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3762         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3763                         "per thread proc mmap processing timeout in ms"),
3764         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3765                      trace__parse_cgroups),
3766         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3767                      "ms to wait before starting measurement after program "
3768                      "start"),
3769         OPT_END()
3770         };
3771         bool __maybe_unused max_stack_user_set = true;
3772         bool mmap_pages_user_set = true;
3773         struct perf_evsel *evsel;
3774         const char * const trace_subcommands[] = { "record", NULL };
3775         int err = -1;
3776         char bf[BUFSIZ];
3777
3778         signal(SIGSEGV, sighandler_dump_stack);
3779         signal(SIGFPE, sighandler_dump_stack);
3780
3781         trace.evlist = perf_evlist__new();
3782         trace.sctbl = syscalltbl__new();
3783
3784         if (trace.evlist == NULL || trace.sctbl == NULL) {
3785                 pr_err("Not enough memory to run!\n");
3786                 err = -ENOMEM;
3787                 goto out;
3788         }
3789
3790         err = perf_config(trace__config, &trace);
3791         if (err)
3792                 goto out;
3793
3794         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3795                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3796
3797         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3798                 usage_with_options_msg(trace_usage, trace_options,
3799                                        "cgroup monitoring only available in system-wide mode");
3800         }
3801
3802         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3803         if (IS_ERR(evsel)) {
3804                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3805                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3806                 goto out;
3807         }
3808
3809         if (evsel) {
3810                 trace.syscalls.events.augmented = evsel;
3811                 trace__set_bpf_map_filtered_pids(&trace);
3812                 trace__set_bpf_map_syscalls(&trace);
3813         }
3814
3815         err = bpf__setup_stdout(trace.evlist);
3816         if (err) {
3817                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3818                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3819                 goto out;
3820         }
3821
3822         err = -1;
3823
3824         if (map_dump_str) {
3825                 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3826                 if (trace.dump.map == NULL) {
3827                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3828                         goto out;
3829                 }
3830         }
3831
3832         if (trace.trace_pgfaults) {
3833                 trace.opts.sample_address = true;
3834                 trace.opts.sample_time = true;
3835         }
3836
3837         if (trace.opts.mmap_pages == UINT_MAX)
3838                 mmap_pages_user_set = false;
3839
3840         if (trace.max_stack == UINT_MAX) {
3841                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3842                 max_stack_user_set = false;
3843         }
3844
3845 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3846         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3847                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3848         }
3849 #endif
3850
3851         if (callchain_param.enabled) {
3852                 if (!mmap_pages_user_set && geteuid() == 0)
3853                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3854
3855                 symbol_conf.use_callchain = true;
3856         }
3857
3858         if (trace.evlist->nr_entries > 0) {
3859                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3860                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3861                         perror("failed to set syscalls:* tracepoint fields");
3862                         goto out;
3863                 }
3864         }
3865
3866         if (trace.sort_events) {
3867                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3868                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3869         }
3870
3871         /*
3872          * If we are augmenting syscalls, then combine what we put in the
3873          * __augmented_syscalls__ BPF map with what is in the
3874          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3875          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3876          *
3877          * We'll switch to look at two BPF maps, one for sys_enter and the
3878          * other for sys_exit when we start augmenting the sys_exit paths with
3879          * buffers that are being copied from kernel to userspace, think 'read'
3880          * syscall.
3881          */
3882         if (trace.syscalls.events.augmented) {
3883                 evlist__for_each_entry(trace.evlist, evsel) {
3884                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3885
3886                         if (raw_syscalls_sys_exit) {
3887                                 trace.raw_augmented_syscalls = true;
3888                                 goto init_augmented_syscall_tp;
3889                         }
3890
3891                         if (trace.syscalls.events.augmented->priv == NULL &&
3892                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3893                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3894                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3895                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3896                                         goto out;
3897                                 augmented->handler = trace__sys_enter;
3898                         }
3899
3900                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3901                                 struct syscall_tp *sc;
3902 init_augmented_syscall_tp:
3903                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3904                                         goto out;
3905                                 sc = evsel->priv;
3906                                 /*
3907                                  * For now with BPF raw_augmented we hook into
3908                                  * raw_syscalls:sys_enter and there we get all
3909                                  * 6 syscall args plus the tracepoint common
3910                                  * fields and the syscall_nr (another long).
3911                                  * So we check if that is the case and if so
3912                                  * don't look after the sc->args_size but
3913                                  * always after the full raw_syscalls:sys_enter
3914                                  * payload, which is fixed.
3915                                  *
3916                                  * We'll revisit this later to pass
3917                                  * s->args_size to the BPF augmenter (now
3918                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3919                                  * so that it copies only what we need for each
3920                                  * syscall, like what happens when we use
3921                                  * syscalls:sys_enter_NAME, so that we reduce
3922                                  * the kernel/userspace traffic to just what is
3923                                  * needed for each syscall.
3924                                  */
3925                                 if (trace.raw_augmented_syscalls)
3926                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3927                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3928                                 evsel->handler = trace__sys_exit;
3929                         }
3930                 }
3931         }
3932
3933         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3934                 return trace__record(&trace, argc-1, &argv[1]);
3935
3936         /* summary_only implies summary option, but don't overwrite summary if set */
3937         if (trace.summary_only)
3938                 trace.summary = trace.summary_only;
3939
3940         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3941             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3942                 trace.trace_syscalls = true;
3943         }
3944
3945         if (output_name != NULL) {
3946                 err = trace__open_output(&trace, output_name);
3947                 if (err < 0) {
3948                         perror("failed to create output file");
3949                         goto out;
3950                 }
3951         }
3952
3953         err = target__validate(&trace.opts.target);
3954         if (err) {
3955                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3956                 fprintf(trace.output, "%s", bf);
3957                 goto out_close;
3958         }
3959
3960         err = target__parse_uid(&trace.opts.target);
3961         if (err) {
3962                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3963                 fprintf(trace.output, "%s", bf);
3964                 goto out_close;
3965         }
3966
3967         if (!argc && target__none(&trace.opts.target))
3968                 trace.opts.target.system_wide = true;
3969
3970         if (input_name)
3971                 err = trace__replay(&trace);
3972         else
3973                 err = trace__run(&trace, argc, argv);
3974
3975 out_close:
3976         if (output_name != NULL)
3977                 fclose(trace.output);
3978 out:
3979         return err;
3980 }