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 #include "trace/beauty/generated/fsconfig_arrays.c"
 486
 487 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 488
 489 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 490 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 491
 492 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 493 static DEFINE_STRARRAY(itimers, "ITIMER_");
 494
 495 static const char *keyctl_options[] = {
 496         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 497         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 498         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 499         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 500         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 501 };
 502 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 503
 504 static const char *whences[] = { "SET", "CUR", "END",
 505 #ifdef SEEK_DATA
 506 "DATA",
 507 #endif
 508 #ifdef SEEK_HOLE
 509 "HOLE",
 510 #endif
 511 };
 512 static DEFINE_STRARRAY(whences, "SEEK_");
 513
 514 static const char *fcntl_cmds[] = {
 515         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 516         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 517         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 518         "GETOWNER_UIDS",
 519 };
 520 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 521
 522 static const char *fcntl_linux_specific_cmds[] = {
 523         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 524         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 525         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 526 };
 527
 528 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 529
 530 static struct strarray *fcntl_cmds_arrays[] = {
 531         &strarray__fcntl_cmds,
 532         &strarray__fcntl_linux_specific_cmds,
 533 };
 534
 535 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 536
 537 static const char *rlimit_resources[] = {
 538         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 539         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 540         "RTTIME",
 541 };
 542 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 543
 544 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 545 static DEFINE_STRARRAY(sighow, "SIG_");
 546
 547 static const char *clockid[] = {
 548         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 549         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 550         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 551 };
 552 static DEFINE_STRARRAY(clockid, "CLOCK_");
 553
 554 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 555                                                  struct syscall_arg *arg)
 556 {
 557         bool show_prefix = arg->show_string_prefix;
 558         const char *suffix = "_OK";
 559         size_t printed = 0;
 560         int mode = arg->val;
 561
 562         if (mode == F_OK) /* 0 */
 563                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 564 #define P_MODE(n) \
 565         if (mode & n##_OK) { \
 566                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 567                 mode &= ~n##_OK; \
 568         }
 569
 570         P_MODE(R);
 571         P_MODE(W);
 572         P_MODE(X);
 573 #undef P_MODE
 574
 575         if (mode)
 576                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 577
 578         return printed;
 579 }
 580
 581 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 582
 583 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 584                                               struct syscall_arg *arg);
 585
 586 #define SCA_FILENAME syscall_arg__scnprintf_filename
 587
 588 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 589                                                 struct syscall_arg *arg)
 590 {
 591         bool show_prefix = arg->show_string_prefix;
 592         const char *prefix = "O_";
 593         int printed = 0, flags = arg->val;
 594
 595 #define P_FLAG(n) \
 596         if (flags & O_##n) { \
 597                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 598                 flags &= ~O_##n; \
 599         }
 600
 601         P_FLAG(CLOEXEC);
 602         P_FLAG(NONBLOCK);
 603 #undef P_FLAG
 604
 605         if (flags)
 606                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 607
 608         return printed;
 609 }
 610
 611 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 612
 613 #ifndef GRND_NONBLOCK
 614 #define GRND_NONBLOCK   0x0001
 615 #endif
 616 #ifndef GRND_RANDOM
 617 #define GRND_RANDOM     0x0002
 618 #endif
 619
 620 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 621                                                    struct syscall_arg *arg)
 622 {
 623         bool show_prefix = arg->show_string_prefix;
 624         const char *prefix = "GRND_";
 625         int printed = 0, flags = arg->val;
 626
 627 #define P_FLAG(n) \
 628         if (flags & GRND_##n) { \
 629                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 630                 flags &= ~GRND_##n; \
 631         }
 632
 633         P_FLAG(RANDOM);
 634         P_FLAG(NONBLOCK);
 635 #undef P_FLAG
 636
 637         if (flags)
 638                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 639
 640         return printed;
 641 }
 642
 643 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 644
 645 #define STRARRAY(name, array) \
 646           { .scnprintf  = SCA_STRARRAY, \
 647             .parm       = &strarray__##array, }
 648
 649 #include "trace/beauty/arch_errno_names.c"
 650 #include "trace/beauty/eventfd.c"
 651 #include "trace/beauty/futex_op.c"
 652 #include "trace/beauty/futex_val3.c"
 653 #include "trace/beauty/mmap.c"
 654 #include "trace/beauty/mode_t.c"
 655 #include "trace/beauty/msg_flags.c"
 656 #include "trace/beauty/open_flags.c"
 657 #include "trace/beauty/perf_event_open.c"
 658 #include "trace/beauty/pid.c"
 659 #include "trace/beauty/sched_policy.c"
 660 #include "trace/beauty/seccomp.c"
 661 #include "trace/beauty/signum.c"
 662 #include "trace/beauty/socket_type.c"
 663 #include "trace/beauty/waitid_options.c"
 664
 665 struct syscall_arg_fmt {
 666         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 667         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 668         void       *parm;
 669         const char *name;
 670         bool       show_zero;
 671 };
 672
 673 static struct syscall_fmt {
 674         const char *name;
 675         const char *alias;
 676         struct syscall_arg_fmt arg[6];
 677         u8         nr_args;
 678         bool       errpid;
 679         bool       timeout;
 680         bool       hexret;
 681 } syscall_fmts[] = {
 682         { .name     = "access",
 683           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 684         { .name     = "arch_prctl",
 685           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 686                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 687         { .name     = "bind",
 688           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 689         { .name     = "bpf",
 690           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 691         { .name     = "brk",        .hexret = true,
 692           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 693         { .name     = "clock_gettime",
 694           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 695         { .name     = "clone",      .errpid = true, .nr_args = 5,
 696           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 697                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 698                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 699                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 700                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 701         { .name     = "close",
 702           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 703         { .name     = "connect",
 704           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 705         { .name     = "epoll_ctl",
 706           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 707         { .name     = "eventfd2",
 708           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 709         { .name     = "fchmodat",
 710           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 711         { .name     = "fchownat",
 712           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 713         { .name     = "fcntl",
 714           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 715                            .parm      = &strarrays__fcntl_cmds_arrays,
 716                            .show_zero = true, },
 717                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 718         { .name     = "flock",
 719           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 720         { .name     = "fsconfig",
 721           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 722         { .name     = "fspick",
 723           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 724                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 725                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 726         { .name     = "fstat", .alias = "newfstat", },
 727         { .name     = "fstatat", .alias = "newfstatat", },
 728         { .name     = "futex",
 729           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 730                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 731         { .name     = "futimesat",
 732           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 733         { .name     = "getitimer",
 734           .arg = { [0] = STRARRAY(which, itimers), }, },
 735         { .name     = "getpid",     .errpid = true, },
 736         { .name     = "getpgid",    .errpid = true, },
 737         { .name     = "getppid",    .errpid = true, },
 738         { .name     = "getrandom",
 739           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 740         { .name     = "getrlimit",
 741           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 742         { .name     = "gettid",     .errpid = true, },
 743         { .name     = "ioctl",
 744           .arg = {
 745 #if defined(__i386__) || defined(__x86_64__)
 746 /*
 747  * FIXME: Make this available to all arches.
 748  */
 749                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 750                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 751 #else
 752                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 753 #endif
 754         { .name     = "kcmp",       .nr_args = 5,
 755           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 756                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 757                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 758                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 759                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 760         { .name     = "keyctl",
 761           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 762         { .name     = "kill",
 763           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 764         { .name     = "linkat",
 765           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 766         { .name     = "lseek",
 767           .arg = { [2] = STRARRAY(whence, whences), }, },
 768         { .name     = "lstat", .alias = "newlstat", },
 769         { .name     = "madvise",
 770           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 771                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 772         { .name     = "mkdirat",
 773           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 774         { .name     = "mknodat",
 775           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 776         { .name     = "mmap",       .hexret = true,
 777 /* The standard mmap maps to old_mmap on s390x */
 778 #if defined(__s390x__)
 779         .alias = "old_mmap",
 780 #endif
 781           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 782                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 783                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 784         { .name     = "mount",
 785           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 786                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 787                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 788         { .name     = "move_mount",
 789           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 790                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 791                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 792                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 793                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 794         { .name     = "mprotect",
 795           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 796                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 797         { .name     = "mq_unlink",
 798           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 799         { .name     = "mremap",     .hexret = true,
 800           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 801         { .name     = "name_to_handle_at",
 802           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 803         { .name     = "newfstatat",
 804           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 805         { .name     = "open",
 806           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 807         { .name     = "open_by_handle_at",
 808           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 809                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 810         { .name     = "openat",
 811           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 812                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 813         { .name     = "perf_event_open",
 814           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 815                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 816                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 817         { .name     = "pipe2",
 818           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 819         { .name     = "pkey_alloc",
 820           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 821         { .name     = "pkey_free",
 822           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 823         { .name     = "pkey_mprotect",
 824           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 825                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 826                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 827         { .name     = "poll", .timeout = true, },
 828         { .name     = "ppoll", .timeout = true, },
 829         { .name     = "prctl",
 830           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 831                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 832                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 833         { .name     = "pread", .alias = "pread64", },
 834         { .name     = "preadv", .alias = "pread", },
 835         { .name     = "prlimit64",
 836           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 837         { .name     = "pwrite", .alias = "pwrite64", },
 838         { .name     = "readlinkat",
 839           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 840         { .name     = "recvfrom",
 841           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 842         { .name     = "recvmmsg",
 843           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 844         { .name     = "recvmsg",
 845           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 846         { .name     = "renameat",
 847           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 848                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 849         { .name     = "renameat2",
 850           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 851                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 852                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 853         { .name     = "rt_sigaction",
 854           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 855         { .name     = "rt_sigprocmask",
 856           .arg = { [0] = STRARRAY(how, sighow), }, },
 857         { .name     = "rt_sigqueueinfo",
 858           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 859         { .name     = "rt_tgsigqueueinfo",
 860           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 861         { .name     = "sched_setscheduler",
 862           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 863         { .name     = "seccomp",
 864           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 865                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 866         { .name     = "select", .timeout = true, },
 867         { .name     = "sendmmsg",
 868           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 869         { .name     = "sendmsg",
 870           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 871         { .name     = "sendto",
 872           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 873                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 874         { .name     = "set_tid_address", .errpid = true, },
 875         { .name     = "setitimer",
 876           .arg = { [0] = STRARRAY(which, itimers), }, },
 877         { .name     = "setrlimit",
 878           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 879         { .name     = "socket",
 880           .arg = { [0] = STRARRAY(family, socket_families),
 881                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 882                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 883         { .name     = "socketpair",
 884           .arg = { [0] = STRARRAY(family, socket_families),
 885                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 886                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 887         { .name     = "stat", .alias = "newstat", },
 888         { .name     = "statx",
 889           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 890                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 891                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 892         { .name     = "swapoff",
 893           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 894         { .name     = "swapon",
 895           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 896         { .name     = "symlinkat",
 897           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 898         { .name     = "tgkill",
 899           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 900         { .name     = "tkill",
 901           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 902         { .name     = "umount2", .alias = "umount",
 903           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 904         { .name     = "uname", .alias = "newuname", },
 905         { .name     = "unlinkat",
 906           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 907         { .name     = "utimensat",
 908           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 909         { .name     = "wait4",      .errpid = true,
 910           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 911         { .name     = "waitid",     .errpid = true,
 912           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 913 };
 914
 915 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 916 {
 917         const struct syscall_fmt *fmt = fmtp;
 918         return strcmp(name, fmt->name);
 919 }
 920
 921 static struct syscall_fmt *syscall_fmt__find(const char *name)
 922 {
 923         const int nmemb = ARRAY_SIZE(syscall_fmts);
 924         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 925 }
 926
 927 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 928 {
 929         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 930
 931         for (i = 0; i < nmemb; ++i) {
 932                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 933                         return &syscall_fmts[i];
 934         }
 935
 936         return NULL;
 937 }
 938
 939 /*
 940  * is_exit: is this "exit" or "exit_group"?
 941  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 942  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 943  */
 944 struct syscall {
 945         struct tep_event    *tp_format;
 946         int                 nr_args;
 947         int                 args_size;
 948         bool                is_exit;
 949         bool                is_open;
 950         struct tep_format_field *args;
 951         const char          *name;
 952         struct syscall_fmt  *fmt;
 953         struct syscall_arg_fmt *arg_fmt;
 954 };
 955
 956 struct bpf_map_syscall_entry {
 957         bool    enabled;
 958 };
 959
 960 /*
 961  * We need to have this 'calculated' boolean because in some cases we really
 962  * don't know what is the duration of a syscall, for instance, when we start
 963  * a session and some threads are waiting for a syscall to finish, say 'poll',
 964  * in which case all we can do is to print "( ? ) for duration and for the
 965  * start timestamp.
 966  */
 967 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 968 {
 969         double duration = (double)t / NSEC_PER_MSEC;
 970         size_t printed = fprintf(fp, "(");
 971
 972         if (!calculated)
 973                 printed += fprintf(fp, "         ");
 974         else if (duration >= 1.0)
 975                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 976         else if (duration >= 0.01)
 977                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 978         else
 979                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 980         return printed + fprintf(fp, "): ");
 981 }
 982
 983 /**
 984  * filename.ptr: The filename char pointer that will be vfs_getname'd
 985  * filename.entry_str_pos: Where to insert the string translated from
 986  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 987  * ret_scnprintf: syscall args may set this to a different syscall return
 988  *                formatter, for instance, fcntl may return fds, file flags, etc.
 989  */
 990 struct thread_trace {
 991         u64               entry_time;
 992         bool              entry_pending;
 993         unsigned long     nr_events;
 994         unsigned long     pfmaj, pfmin;
 995         char              *entry_str;
 996         double            runtime_ms;
 997         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 998         struct {
 999                 unsigned long ptr;
1000                 short int     entry_str_pos;
1001                 bool          pending_open;
1002                 unsigned int  namelen;
1003                 char          *name;
1004         } filename;
1005         struct {
1006                 int           max;
1007                 struct file   *table;
1008         } files;
1009
1010         struct intlist *syscall_stats;
1011 };
1012
1013 static struct thread_trace *thread_trace__new(void)
1014 {
1015         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1016
1017         if (ttrace)
1018                 ttrace->files.max = -1;
1019
1020         ttrace->syscall_stats = intlist__new(NULL);
1021
1022         return ttrace;
1023 }
1024
1025 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1026 {
1027         struct thread_trace *ttrace;
1028
1029         if (thread == NULL)
1030                 goto fail;
1031
1032         if (thread__priv(thread) == NULL)
1033                 thread__set_priv(thread, thread_trace__new());
1034
1035         if (thread__priv(thread) == NULL)
1036                 goto fail;
1037
1038         ttrace = thread__priv(thread);
1039         ++ttrace->nr_events;
1040
1041         return ttrace;
1042 fail:
1043         color_fprintf(fp, PERF_COLOR_RED,
1044                       "WARNING: not enough memory, dropping samples!\n");
1045         return NULL;
1046 }
1047
1048
1049 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1050                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1051 {
1052         struct thread_trace *ttrace = thread__priv(arg->thread);
1053
1054         ttrace->ret_scnprintf = ret_scnprintf;
1055 }
1056
1057 #define TRACE_PFMAJ             (1 << 0)
1058 #define TRACE_PFMIN             (1 << 1)
1059
1060 static const size_t trace__entry_str_size = 2048;
1061
1062 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1063 {
1064         if (fd < 0)
1065                 return NULL;
1066
1067         if (fd > ttrace->files.max) {
1068                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1069
1070                 if (nfiles == NULL)
1071                         return NULL;
1072
1073                 if (ttrace->files.max != -1) {
1074                         memset(nfiles + ttrace->files.max + 1, 0,
1075                                (fd - ttrace->files.max) * sizeof(struct file));
1076                 } else {
1077                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1078                 }
1079
1080                 ttrace->files.table = nfiles;
1081                 ttrace->files.max   = fd;
1082         }
1083
1084         return ttrace->files.table + fd;
1085 }
1086
1087 struct file *thread__files_entry(struct thread *thread, int fd)
1088 {
1089         return thread_trace__files_entry(thread__priv(thread), fd);
1090 }
1091
1092 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1093 {
1094         struct thread_trace *ttrace = thread__priv(thread);
1095         struct file *file = thread_trace__files_entry(ttrace, fd);
1096
1097         if (file != NULL) {
1098                 struct stat st;
1099                 if (stat(pathname, &st) == 0)
1100                         file->dev_maj = major(st.st_rdev);
1101                 file->pathname = strdup(pathname);
1102                 if (file->pathname)
1103                         return 0;
1104         }
1105
1106         return -1;
1107 }
1108
1109 static int thread__read_fd_path(struct thread *thread, int fd)
1110 {
1111         char linkname[PATH_MAX], pathname[PATH_MAX];
1112         struct stat st;
1113         int ret;
1114
1115         if (thread->pid_ == thread->tid) {
1116                 scnprintf(linkname, sizeof(linkname),
1117                           "/proc/%d/fd/%d", thread->pid_, fd);
1118         } else {
1119                 scnprintf(linkname, sizeof(linkname),
1120                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1121         }
1122
1123         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1124                 return -1;
1125
1126         ret = readlink(linkname, pathname, sizeof(pathname));
1127
1128         if (ret < 0 || ret > st.st_size)
1129                 return -1;
1130
1131         pathname[ret] = '\0';
1132         return trace__set_fd_pathname(thread, fd, pathname);
1133 }
1134
1135 static const char *thread__fd_path(struct thread *thread, int fd,
1136                                    struct trace *trace)
1137 {
1138         struct thread_trace *ttrace = thread__priv(thread);
1139
1140         if (ttrace == NULL)
1141                 return NULL;
1142
1143         if (fd < 0)
1144                 return NULL;
1145
1146         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1147                 if (!trace->live)
1148                         return NULL;
1149                 ++trace->stats.proc_getname;
1150                 if (thread__read_fd_path(thread, fd))
1151                         return NULL;
1152         }
1153
1154         return ttrace->files.table[fd].pathname;
1155 }
1156
1157 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1158 {
1159         int fd = arg->val;
1160         size_t printed = scnprintf(bf, size, "%d", fd);
1161         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1162
1163         if (path)
1164                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1165
1166         return printed;
1167 }
1168
1169 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1170 {
1171         size_t printed = scnprintf(bf, size, "%d", fd);
1172         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1173
1174         if (thread) {
1175                 const char *path = thread__fd_path(thread, fd, trace);
1176
1177                 if (path)
1178                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1179
1180                 thread__put(thread);
1181         }
1182
1183         return printed;
1184 }
1185
1186 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1187                                               struct syscall_arg *arg)
1188 {
1189         int fd = arg->val;
1190         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1191         struct thread_trace *ttrace = thread__priv(arg->thread);
1192
1193         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1194                 zfree(&ttrace->files.table[fd].pathname);
1195
1196         return printed;
1197 }
1198
1199 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1200                                      unsigned long ptr)
1201 {
1202         struct thread_trace *ttrace = thread__priv(thread);
1203
1204         ttrace->filename.ptr = ptr;
1205         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1206 }
1207
1208 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1209 {
1210         struct augmented_arg *augmented_arg = arg->augmented.args;
1211
1212         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1213 }
1214
1215 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1216                                               struct syscall_arg *arg)
1217 {
1218         unsigned long ptr = arg->val;
1219
1220         if (arg->augmented.args)
1221                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1222
1223         if (!arg->trace->vfs_getname)
1224                 return scnprintf(bf, size, "%#x", ptr);
1225
1226         thread__set_filename_pos(arg->thread, bf, ptr);
1227         return 0;
1228 }
1229
1230 static bool trace__filter_duration(struct trace *trace, double t)
1231 {
1232         return t < (trace->duration_filter * NSEC_PER_MSEC);
1233 }
1234
1235 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1236 {
1237         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1238
1239         return fprintf(fp, "%10.3f ", ts);
1240 }
1241
1242 /*
1243  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1244  * using ttrace->entry_time for a thread that receives a sys_exit without
1245  * first having received a sys_enter ("poll" issued before tracing session
1246  * starts, lost sys_enter exit due to ring buffer overflow).
1247  */
1248 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1249 {
1250         if (tstamp > 0)
1251                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1252
1253         return fprintf(fp, "         ? ");
1254 }
1255
1256 static bool done = false;
1257 static bool interrupted = false;
1258
1259 static void sig_handler(int sig)
1260 {
1261         done = true;
1262         interrupted = sig == SIGINT;
1263 }
1264
1265 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1266 {
1267         size_t printed = 0;
1268
1269         if (trace->multiple_threads) {
1270                 if (trace->show_comm)
1271                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1272                 printed += fprintf(fp, "%d ", thread->tid);
1273         }
1274
1275         return printed;
1276 }
1277
1278 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1279                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1280 {
1281         size_t printed = 0;
1282
1283         if (trace->show_tstamp)
1284                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1285         if (trace->show_duration)
1286                 printed += fprintf_duration(duration, duration_calculated, fp);
1287         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1288 }
1289
1290 static int trace__process_event(struct trace *trace, struct machine *machine,
1291                                 union perf_event *event, struct perf_sample *sample)
1292 {
1293         int ret = 0;
1294
1295         switch (event->header.type) {
1296         case PERF_RECORD_LOST:
1297                 color_fprintf(trace->output, PERF_COLOR_RED,
1298                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1299                 ret = machine__process_lost_event(machine, event, sample);
1300                 break;
1301         default:
1302                 ret = machine__process_event(machine, event, sample);
1303                 break;
1304         }
1305
1306         return ret;
1307 }
1308
1309 static int trace__tool_process(struct perf_tool *tool,
1310                                union perf_event *event,
1311                                struct perf_sample *sample,
1312                                struct machine *machine)
1313 {
1314         struct trace *trace = container_of(tool, struct trace, tool);
1315         return trace__process_event(trace, machine, event, sample);
1316 }
1317
1318 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1319 {
1320         struct machine *machine = vmachine;
1321
1322         if (machine->kptr_restrict_warned)
1323                 return NULL;
1324
1325         if (symbol_conf.kptr_restrict) {
1326                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1327                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1328                            "Kernel samples will not be resolved.\n");
1329                 machine->kptr_restrict_warned = true;
1330                 return NULL;
1331         }
1332
1333         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1334 }
1335
1336 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1337 {
1338         int err = symbol__init(NULL);
1339
1340         if (err)
1341                 return err;
1342
1343         trace->host = machine__new_host();
1344         if (trace->host == NULL)
1345                 return -ENOMEM;
1346
1347         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1348         if (err < 0)
1349                 goto out;
1350
1351         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1352                                             evlist->threads, trace__tool_process, false,
1353                                             1);
1354 out:
1355         if (err)
1356                 symbol__exit();
1357
1358         return err;
1359 }
1360
1361 static void trace__symbols__exit(struct trace *trace)
1362 {
1363         machine__exit(trace->host);
1364         trace->host = NULL;
1365
1366         symbol__exit();
1367 }
1368
1369 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1370 {
1371         int idx;
1372
1373         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1374                 nr_args = sc->fmt->nr_args;
1375
1376         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1377         if (sc->arg_fmt == NULL)
1378                 return -1;
1379
1380         for (idx = 0; idx < nr_args; ++idx) {
1381                 if (sc->fmt)
1382                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1383         }
1384
1385         sc->nr_args = nr_args;
1386         return 0;
1387 }
1388
1389 static int syscall__set_arg_fmts(struct syscall *sc)
1390 {
1391         struct tep_format_field *field, *last_field = NULL;
1392         int idx = 0, len;
1393
1394         for (field = sc->args; field; field = field->next, ++idx) {
1395                 last_field = field;
1396
1397                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1398                         continue;
1399
1400                 if (strcmp(field->type, "const char *") == 0 &&
1401                          (strcmp(field->name, "filename") == 0 ||
1402                           strcmp(field->name, "path") == 0 ||
1403                           strcmp(field->name, "pathname") == 0))
1404                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1405                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1406                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1407                 else if (strcmp(field->type, "pid_t") == 0)
1408                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1409                 else if (strcmp(field->type, "umode_t") == 0)
1410                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1411                 else if ((strcmp(field->type, "int") == 0 ||
1412                           strcmp(field->type, "unsigned int") == 0 ||
1413                           strcmp(field->type, "long") == 0) &&
1414                          (len = strlen(field->name)) >= 2 &&
1415                          strcmp(field->name + len - 2, "fd") == 0) {
1416                         /*
1417                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1418                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1419                          * 65 int
1420                          * 23 unsigned int
1421                          * 7 unsigned long
1422                          */
1423                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1424                 }
1425         }
1426
1427         if (last_field)
1428                 sc->args_size = last_field->offset + last_field->size;
1429
1430         return 0;
1431 }
1432
1433 static int trace__read_syscall_info(struct trace *trace, int id)
1434 {
1435         char tp_name[128];
1436         struct syscall *sc;
1437         const char *name = syscalltbl__name(trace->sctbl, id);
1438
1439         if (name == NULL)
1440                 return -1;
1441
1442         if (id > trace->syscalls.max) {
1443                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1444
1445                 if (nsyscalls == NULL)
1446                         return -1;
1447
1448                 if (trace->syscalls.max != -1) {
1449                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1450                                (id - trace->syscalls.max) * sizeof(*sc));
1451                 } else {
1452                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1453                 }
1454
1455                 trace->syscalls.table = nsyscalls;
1456                 trace->syscalls.max   = id;
1457         }
1458
1459         sc = trace->syscalls.table + id;
1460         sc->name = name;
1461
1462         sc->fmt  = syscall_fmt__find(sc->name);
1463
1464         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1465         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1466
1467         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1468                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1469                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1470         }
1471
1472         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1473                 return -1;
1474
1475         if (IS_ERR(sc->tp_format))
1476                 return -1;
1477
1478         sc->args = sc->tp_format->format.fields;
1479         /*
1480          * We need to check and discard the first variable '__syscall_nr'
1481          * or 'nr' that mean the syscall number. It is needless here.
1482          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1483          */
1484         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1485                 sc->args = sc->args->next;
1486                 --sc->nr_args;
1487         }
1488
1489         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1490         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1491
1492         return syscall__set_arg_fmts(sc);
1493 }
1494
1495 static int trace__validate_ev_qualifier(struct trace *trace)
1496 {
1497         int err = 0, i;
1498         size_t nr_allocated;
1499         struct str_node *pos;
1500
1501         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1502         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1503                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1504
1505         if (trace->ev_qualifier_ids.entries == NULL) {
1506                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1507                        trace->output);
1508                 err = -EINVAL;
1509                 goto out;
1510         }
1511
1512         nr_allocated = trace->ev_qualifier_ids.nr;
1513         i = 0;
1514
1515         strlist__for_each_entry(pos, trace->ev_qualifier) {
1516                 const char *sc = pos->s;
1517                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1518
1519                 if (id < 0) {
1520                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1521                         if (id >= 0)
1522                                 goto matches;
1523
1524                         if (err == 0) {
1525                                 fputs("Error:\tInvalid syscall ", trace->output);
1526                                 err = -EINVAL;
1527                         } else {
1528                                 fputs(", ", trace->output);
1529                         }
1530
1531                         fputs(sc, trace->output);
1532                 }
1533 matches:
1534                 trace->ev_qualifier_ids.entries[i++] = id;
1535                 if (match_next == -1)
1536                         continue;
1537
1538                 while (1) {
1539                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1540                         if (id < 0)
1541                                 break;
1542                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1543                                 void *entries;
1544
1545                                 nr_allocated += 8;
1546                                 entries = realloc(trace->ev_qualifier_ids.entries,
1547                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1548                                 if (entries == NULL) {
1549                                         err = -ENOMEM;
1550                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1551                                         goto out_free;
1552                                 }
1553                                 trace->ev_qualifier_ids.entries = entries;
1554                         }
1555                         trace->ev_qualifier_ids.nr++;
1556                         trace->ev_qualifier_ids.entries[i++] = id;
1557                 }
1558         }
1559
1560         if (err < 0) {
1561                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1562                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1563 out_free:
1564                 zfree(&trace->ev_qualifier_ids.entries);
1565                 trace->ev_qualifier_ids.nr = 0;
1566         }
1567 out:
1568         return err;
1569 }
1570
1571 /*
1572  * args is to be interpreted as a series of longs but we need to handle
1573  * 8-byte unaligned accesses. args points to raw_data within the event
1574  * and raw_data is guaranteed to be 8-byte unaligned because it is
1575  * preceded by raw_size which is a u32. So we need to copy args to a temp
1576  * variable to read it. Most notably this avoids extended load instructions
1577  * on unaligned addresses
1578  */
1579 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1580 {
1581         unsigned long val;
1582         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1583
1584         memcpy(&val, p, sizeof(val));
1585         return val;
1586 }
1587
1588 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1589                                       struct syscall_arg *arg)
1590 {
1591         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1592                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1593
1594         return scnprintf(bf, size, "arg%d: ", arg->idx);
1595 }
1596
1597 /*
1598  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1599  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1600  * in tools/perf/trace/beauty/mount_flags.c
1601  */
1602 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1603 {
1604         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1605                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1606
1607         return val;
1608 }
1609
1610 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1611                                      struct syscall_arg *arg, unsigned long val)
1612 {
1613         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1614                 arg->val = val;
1615                 if (sc->arg_fmt[arg->idx].parm)
1616                         arg->parm = sc->arg_fmt[arg->idx].parm;
1617                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1618         }
1619         return scnprintf(bf, size, "%ld", val);
1620 }
1621
1622 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1623                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1624                                       struct trace *trace, struct thread *thread)
1625 {
1626         size_t printed = 0;
1627         unsigned long val;
1628         u8 bit = 1;
1629         struct syscall_arg arg = {
1630                 .args   = args,
1631                 .augmented = {
1632                         .size = augmented_args_size,
1633                         .args = augmented_args,
1634                 },
1635                 .idx    = 0,
1636                 .mask   = 0,
1637                 .trace  = trace,
1638                 .thread = thread,
1639                 .show_string_prefix = trace->show_string_prefix,
1640         };
1641         struct thread_trace *ttrace = thread__priv(thread);
1642
1643         /*
1644          * Things like fcntl will set this in its 'cmd' formatter to pick the
1645          * right formatter for the return value (an fd? file flags?), which is
1646          * not needed for syscalls that always return a given type, say an fd.
1647          */
1648         ttrace->ret_scnprintf = NULL;
1649
1650         if (sc->args != NULL) {
1651                 struct tep_format_field *field;
1652
1653                 for (field = sc->args; field;
1654                      field = field->next, ++arg.idx, bit <<= 1) {
1655                         if (arg.mask & bit)
1656                                 continue;
1657
1658                         val = syscall_arg__val(&arg, arg.idx);
1659                         /*
1660                          * Some syscall args need some mask, most don't and
1661                          * return val untouched.
1662                          */
1663                         val = syscall__mask_val(sc, &arg, val);
1664
1665                         /*
1666                          * Suppress this argument if its value is zero and
1667                          * and we don't have a string associated in an
1668                          * strarray for it.
1669                          */
1670                         if (val == 0 &&
1671                             !trace->show_zeros &&
1672                             !(sc->arg_fmt &&
1673                               (sc->arg_fmt[arg.idx].show_zero ||
1674                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1675                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1676                               sc->arg_fmt[arg.idx].parm))
1677                                 continue;
1678
1679                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1680
1681                         if (trace->show_arg_names)
1682                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1683
1684                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1685                 }
1686         } else if (IS_ERR(sc->tp_format)) {
1687                 /*
1688                  * If we managed to read the tracepoint /format file, then we
1689                  * may end up not having any args, like with gettid(), so only
1690                  * print the raw args when we didn't manage to read it.
1691                  */
1692                 while (arg.idx < sc->nr_args) {
1693                         if (arg.mask & bit)
1694                                 goto next_arg;
1695                         val = syscall_arg__val(&arg, arg.idx);
1696                         if (printed)
1697                                 printed += scnprintf(bf + printed, size - printed, ", ");
1698                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1699                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1700 next_arg:
1701                         ++arg.idx;
1702                         bit <<= 1;
1703                 }
1704         }
1705
1706         return printed;
1707 }
1708
1709 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1710                                   union perf_event *event,
1711                                   struct perf_sample *sample);
1712
1713 static struct syscall *trace__syscall_info(struct trace *trace,
1714                                            struct perf_evsel *evsel, int id)
1715 {
1716
1717         if (id < 0) {
1718
1719                 /*
1720                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1721                  * before that, leaving at a higher verbosity level till that is
1722                  * explained. Reproduced with plain ftrace with:
1723                  *
1724                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1725                  * grep "NR -1 " /t/trace_pipe
1726                  *
1727                  * After generating some load on the machine.
1728                  */
1729                 if (verbose > 1) {
1730                         static u64 n;
1731                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1732                                 id, perf_evsel__name(evsel), ++n);
1733                 }
1734                 return NULL;
1735         }
1736
1737         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1738             trace__read_syscall_info(trace, id))
1739                 goto out_cant_read;
1740
1741         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1742                 goto out_cant_read;
1743
1744         return &trace->syscalls.table[id];
1745
1746 out_cant_read:
1747         if (verbose > 0) {
1748                 fprintf(trace->output, "Problems reading syscall %d", id);
1749                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1750                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1751                 fputs(" information\n", trace->output);
1752         }
1753         return NULL;
1754 }
1755
1756 static void thread__update_stats(struct thread_trace *ttrace,
1757                                  int id, struct perf_sample *sample)
1758 {
1759         struct int_node *inode;
1760         struct stats *stats;
1761         u64 duration = 0;
1762
1763         inode = intlist__findnew(ttrace->syscall_stats, id);
1764         if (inode == NULL)
1765                 return;
1766
1767         stats = inode->priv;
1768         if (stats == NULL) {
1769                 stats = malloc(sizeof(struct stats));
1770                 if (stats == NULL)
1771                         return;
1772                 init_stats(stats);
1773                 inode->priv = stats;
1774         }
1775
1776         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1777                 duration = sample->time - ttrace->entry_time;
1778
1779         update_stats(stats, duration);
1780 }
1781
1782 static int trace__printf_interrupted_entry(struct trace *trace)
1783 {
1784         struct thread_trace *ttrace;
1785         size_t printed;
1786         int len;
1787
1788         if (trace->failure_only || trace->current == NULL)
1789                 return 0;
1790
1791         ttrace = thread__priv(trace->current);
1792
1793         if (!ttrace->entry_pending)
1794                 return 0;
1795
1796         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1797         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1798
1799         if (len < trace->args_alignment - 4)
1800                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1801
1802         printed += fprintf(trace->output, " ...\n");
1803
1804         ttrace->entry_pending = false;
1805         ++trace->nr_events_printed;
1806
1807         return printed;
1808 }
1809
1810 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1811                                  struct perf_sample *sample, struct thread *thread)
1812 {
1813         int printed = 0;
1814
1815         if (trace->print_sample) {
1816                 double ts = (double)sample->time / NSEC_PER_MSEC;
1817
1818                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1819                                    perf_evsel__name(evsel), ts,
1820                                    thread__comm_str(thread),
1821                                    sample->pid, sample->tid, sample->cpu);
1822         }
1823
1824         return printed;
1825 }
1826
1827 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1828 {
1829         void *augmented_args = NULL;
1830         /*
1831          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1832          * and there we get all 6 syscall args plus the tracepoint common fields
1833          * that gets calculated at the start and the syscall_nr (another long).
1834          * So we check if that is the case and if so don't look after the
1835          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1836          * which is fixed.
1837          *
1838          * We'll revisit this later to pass s->args_size to the BPF augmenter
1839          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1840          * copies only what we need for each syscall, like what happens when we
1841          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1842          * traffic to just what is needed for each syscall.
1843          */
1844         int args_size = raw_augmented_args_size ?: sc->args_size;
1845
1846         *augmented_args_size = sample->raw_size - args_size;
1847         if (*augmented_args_size > 0)
1848                 augmented_args = sample->raw_data + args_size;
1849
1850         return augmented_args;
1851 }
1852
1853 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1854                             union perf_event *event __maybe_unused,
1855                             struct perf_sample *sample)
1856 {
1857         char *msg;
1858         void *args;
1859         int printed = 0;
1860         struct thread *thread;
1861         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1862         int augmented_args_size = 0;
1863         void *augmented_args = NULL;
1864         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1865         struct thread_trace *ttrace;
1866
1867         if (sc == NULL)
1868                 return -1;
1869
1870         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1871         ttrace = thread__trace(thread, trace->output);
1872         if (ttrace == NULL)
1873                 goto out_put;
1874
1875         trace__fprintf_sample(trace, evsel, sample, thread);
1876
1877         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1878
1879         if (ttrace->entry_str == NULL) {
1880                 ttrace->entry_str = malloc(trace__entry_str_size);
1881                 if (!ttrace->entry_str)
1882                         goto out_put;
1883         }
1884
1885         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1886                 trace__printf_interrupted_entry(trace);
1887         /*
1888          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1889          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1890          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1891          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1892          * so when handling, say the openat syscall, we end up getting 6 args for the
1893          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1894          * thinking that the extra 2 u64 args are the augmented filename, so just check
1895          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1896          */
1897         if (evsel != trace->syscalls.events.sys_enter)
1898                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1899         ttrace->entry_time = sample->time;
1900         msg = ttrace->entry_str;
1901         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1902
1903         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1904                                            args, augmented_args, augmented_args_size, trace, thread);
1905
1906         if (sc->is_exit) {
1907                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1908                         int alignment = 0;
1909
1910                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1911                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1912                         if (trace->args_alignment > printed)
1913                                 alignment = trace->args_alignment - printed;
1914                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1915                 }
1916         } else {
1917                 ttrace->entry_pending = true;
1918                 /* See trace__vfs_getname & trace__sys_exit */
1919                 ttrace->filename.pending_open = false;
1920         }
1921
1922         if (trace->current != thread) {
1923                 thread__put(trace->current);
1924                 trace->current = thread__get(thread);
1925         }
1926         err = 0;
1927 out_put:
1928         thread__put(thread);
1929         return err;
1930 }
1931
1932 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1933                                     struct perf_sample *sample)
1934 {
1935         struct thread_trace *ttrace;
1936         struct thread *thread;
1937         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1938         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1939         char msg[1024];
1940         void *args, *augmented_args = NULL;
1941         int augmented_args_size;
1942
1943         if (sc == NULL)
1944                 return -1;
1945
1946         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1947         ttrace = thread__trace(thread, trace->output);
1948         /*
1949          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1950          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1951          */
1952         if (ttrace == NULL)
1953                 goto out_put;
1954
1955         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1956         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1957         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1958         fprintf(trace->output, "%s", msg);
1959         err = 0;
1960 out_put:
1961         thread__put(thread);
1962         return err;
1963 }
1964
1965 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1966                                     struct perf_sample *sample,
1967                                     struct callchain_cursor *cursor)
1968 {
1969         struct addr_location al;
1970         int max_stack = evsel->attr.sample_max_stack ?
1971                         evsel->attr.sample_max_stack :
1972                         trace->max_stack;
1973         int err;
1974
1975         if (machine__resolve(trace->host, &al, sample) < 0)
1976                 return -1;
1977
1978         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1979         addr_location__put(&al);
1980         return err;
1981 }
1982
1983 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1984 {
1985         /* TODO: user-configurable print_opts */
1986         const unsigned int print_opts = EVSEL__PRINT_SYM |
1987                                         EVSEL__PRINT_DSO |
1988                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1989
1990         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1991 }
1992
1993 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1994 {
1995         struct perf_env *env = perf_evsel__env(evsel);
1996         const char *arch_name = perf_env__arch(env);
1997
1998         return arch_syscalls__strerrno(arch_name, err);
1999 }
2000
2001 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
2002                            union perf_event *event __maybe_unused,
2003                            struct perf_sample *sample)
2004 {
2005         long ret;
2006         u64 duration = 0;
2007         bool duration_calculated = false;
2008         struct thread *thread;
2009         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2010         int alignment = trace->args_alignment;
2011         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2012         struct thread_trace *ttrace;
2013
2014         if (sc == NULL)
2015                 return -1;
2016
2017         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2018         ttrace = thread__trace(thread, trace->output);
2019         if (ttrace == NULL)
2020                 goto out_put;
2021
2022         trace__fprintf_sample(trace, evsel, sample, thread);
2023
2024         if (trace->summary)
2025                 thread__update_stats(ttrace, id, sample);
2026
2027         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2028
2029         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2030                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2031                 ttrace->filename.pending_open = false;
2032                 ++trace->stats.vfs_getname;
2033         }
2034
2035         if (ttrace->entry_time) {
2036                 duration = sample->time - ttrace->entry_time;
2037                 if (trace__filter_duration(trace, duration))
2038                         goto out;
2039                 duration_calculated = true;
2040         } else if (trace->duration_filter)
2041                 goto out;
2042
2043         if (sample->callchain) {
2044                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2045                 if (callchain_ret == 0) {
2046                         if (callchain_cursor.nr < trace->min_stack)
2047                                 goto out;
2048                         callchain_ret = 1;
2049                 }
2050         }
2051
2052         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2053                 goto out;
2054
2055         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2056
2057         if (ttrace->entry_pending) {
2058                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2059         } else {
2060                 printed += fprintf(trace->output, " ... [");
2061                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2062                 printed += 9;
2063                 printed += fprintf(trace->output, "]: %s()", sc->name);
2064         }
2065
2066         printed++; /* the closing ')' */
2067
2068         if (alignment > printed)
2069                 alignment -= printed;
2070         else
2071                 alignment = 0;
2072
2073         fprintf(trace->output, ")%*s= ", alignment, " ");
2074
2075         if (sc->fmt == NULL) {
2076                 if (ret < 0)
2077                         goto errno_print;
2078 signed_print:
2079                 fprintf(trace->output, "%ld", ret);
2080         } else if (ret < 0) {
2081 errno_print: {
2082                 char bf[STRERR_BUFSIZE];
2083                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2084                            *e = errno_to_name(evsel, -ret);
2085
2086                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2087         }
2088         } else if (ret == 0 && sc->fmt->timeout)
2089                 fprintf(trace->output, "0 (Timeout)");
2090         else if (ttrace->ret_scnprintf) {
2091                 char bf[1024];
2092                 struct syscall_arg arg = {
2093                         .val    = ret,
2094                         .thread = thread,
2095                         .trace  = trace,
2096                 };
2097                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2098                 ttrace->ret_scnprintf = NULL;
2099                 fprintf(trace->output, "%s", bf);
2100         } else if (sc->fmt->hexret)
2101                 fprintf(trace->output, "%#lx", ret);
2102         else if (sc->fmt->errpid) {
2103                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2104
2105                 if (child != NULL) {
2106                         fprintf(trace->output, "%ld", ret);
2107                         if (child->comm_set)
2108                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2109                         thread__put(child);
2110                 }
2111         } else
2112                 goto signed_print;
2113
2114         fputc('\n', trace->output);
2115
2116         /*
2117          * We only consider an 'event' for the sake of --max-events a non-filtered
2118          * sys_enter + sys_exit and other tracepoint events.
2119          */
2120         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2121                 interrupted = true;
2122
2123         if (callchain_ret > 0)
2124                 trace__fprintf_callchain(trace, sample);
2125         else if (callchain_ret < 0)
2126                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2127 out:
2128         ttrace->entry_pending = false;
2129         err = 0;
2130 out_put:
2131         thread__put(thread);
2132         return err;
2133 }
2134
2135 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2136                               union perf_event *event __maybe_unused,
2137                               struct perf_sample *sample)
2138 {
2139         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2140         struct thread_trace *ttrace;
2141         size_t filename_len, entry_str_len, to_move;
2142         ssize_t remaining_space;
2143         char *pos;
2144         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2145
2146         if (!thread)
2147                 goto out;
2148
2149         ttrace = thread__priv(thread);
2150         if (!ttrace)
2151                 goto out_put;
2152
2153         filename_len = strlen(filename);
2154         if (filename_len == 0)
2155                 goto out_put;
2156
2157         if (ttrace->filename.namelen < filename_len) {
2158                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2159
2160                 if (f == NULL)
2161                         goto out_put;
2162
2163                 ttrace->filename.namelen = filename_len;
2164                 ttrace->filename.name = f;
2165         }
2166
2167         strcpy(ttrace->filename.name, filename);
2168         ttrace->filename.pending_open = true;
2169
2170         if (!ttrace->filename.ptr)
2171                 goto out_put;
2172
2173         entry_str_len = strlen(ttrace->entry_str);
2174         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2175         if (remaining_space <= 0)
2176                 goto out_put;
2177
2178         if (filename_len > (size_t)remaining_space) {
2179                 filename += filename_len - remaining_space;
2180                 filename_len = remaining_space;
2181         }
2182
2183         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2184         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2185         memmove(pos + filename_len, pos, to_move);
2186         memcpy(pos, filename, filename_len);
2187
2188         ttrace->filename.ptr = 0;
2189         ttrace->filename.entry_str_pos = 0;
2190 out_put:
2191         thread__put(thread);
2192 out:
2193         return 0;
2194 }
2195
2196 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2197                                      union perf_event *event __maybe_unused,
2198                                      struct perf_sample *sample)
2199 {
2200         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2201         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2202         struct thread *thread = machine__findnew_thread(trace->host,
2203                                                         sample->pid,
2204                                                         sample->tid);
2205         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2206
2207         if (ttrace == NULL)
2208                 goto out_dump;
2209
2210         ttrace->runtime_ms += runtime_ms;
2211         trace->runtime_ms += runtime_ms;
2212 out_put:
2213         thread__put(thread);
2214         return 0;
2215
2216 out_dump:
2217         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2218                evsel->name,
2219                perf_evsel__strval(evsel, sample, "comm"),
2220                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2221                runtime,
2222                perf_evsel__intval(evsel, sample, "vruntime"));
2223         goto out_put;
2224 }
2225
2226 static int bpf_output__printer(enum binary_printer_ops op,
2227                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2228 {
2229         unsigned char ch = (unsigned char)val;
2230
2231         switch (op) {
2232         case BINARY_PRINT_CHAR_DATA:
2233                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2234         case BINARY_PRINT_DATA_BEGIN:
2235         case BINARY_PRINT_LINE_BEGIN:
2236         case BINARY_PRINT_ADDR:
2237         case BINARY_PRINT_NUM_DATA:
2238         case BINARY_PRINT_NUM_PAD:
2239         case BINARY_PRINT_SEP:
2240         case BINARY_PRINT_CHAR_PAD:
2241         case BINARY_PRINT_LINE_END:
2242         case BINARY_PRINT_DATA_END:
2243         default:
2244                 break;
2245         }
2246
2247         return 0;
2248 }
2249
2250 static void bpf_output__fprintf(struct trace *trace,
2251                                 struct perf_sample *sample)
2252 {
2253         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2254                         bpf_output__printer, NULL, trace->output);
2255         ++trace->nr_events_printed;
2256 }
2257
2258 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2259                                 union perf_event *event __maybe_unused,
2260                                 struct perf_sample *sample)
2261 {
2262         struct thread *thread;
2263         int callchain_ret = 0;
2264         /*
2265          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2266          * this event's max_events having been hit and this is an entry coming
2267          * from the ring buffer that we should discard, since the max events
2268          * have already been considered/printed.
2269          */
2270         if (evsel->disabled)
2271                 return 0;
2272
2273         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2274
2275         if (sample->callchain) {
2276                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2277                 if (callchain_ret == 0) {
2278                         if (callchain_cursor.nr < trace->min_stack)
2279                                 goto out;
2280                         callchain_ret = 1;
2281                 }
2282         }
2283
2284         trace__printf_interrupted_entry(trace);
2285         trace__fprintf_tstamp(trace, sample->time, trace->output);
2286
2287         if (trace->trace_syscalls && trace->show_duration)
2288                 fprintf(trace->output, "(         ): ");
2289
2290         if (thread)
2291                 trace__fprintf_comm_tid(trace, thread, trace->output);
2292
2293         if (evsel == trace->syscalls.events.augmented) {
2294                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2295                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2296
2297                 if (sc) {
2298                         fprintf(trace->output, "%s(", sc->name);
2299                         trace__fprintf_sys_enter(trace, evsel, sample);
2300                         fputc(')', trace->output);
2301                         goto newline;
2302                 }
2303
2304                 /*
2305                  * XXX: Not having the associated syscall info or not finding/adding
2306                  *      the thread should never happen, but if it does...
2307                  *      fall thru and print it as a bpf_output event.
2308                  */
2309         }
2310
2311         fprintf(trace->output, "%s:", evsel->name);
2312
2313         if (perf_evsel__is_bpf_output(evsel)) {
2314                 bpf_output__fprintf(trace, sample);
2315         } else if (evsel->tp_format) {
2316                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2317                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2318                         event_format__fprintf(evsel->tp_format, sample->cpu,
2319                                               sample->raw_data, sample->raw_size,
2320                                               trace->output);
2321                         ++trace->nr_events_printed;
2322
2323                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2324                                 perf_evsel__disable(evsel);
2325                                 perf_evsel__close(evsel);
2326                         }
2327                 }
2328         }
2329
2330 newline:
2331         fprintf(trace->output, "\n");
2332
2333         if (callchain_ret > 0)
2334                 trace__fprintf_callchain(trace, sample);
2335         else if (callchain_ret < 0)
2336                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2337 out:
2338         thread__put(thread);
2339         return 0;
2340 }
2341
2342 static void print_location(FILE *f, struct perf_sample *sample,
2343                            struct addr_location *al,
2344                            bool print_dso, bool print_sym)
2345 {
2346
2347         if ((verbose > 0 || print_dso) && al->map)
2348                 fprintf(f, "%s@", al->map->dso->long_name);
2349
2350         if ((verbose > 0 || print_sym) && al->sym)
2351                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2352                         al->addr - al->sym->start);
2353         else if (al->map)
2354                 fprintf(f, "0x%" PRIx64, al->addr);
2355         else
2356                 fprintf(f, "0x%" PRIx64, sample->addr);
2357 }
2358
2359 static int trace__pgfault(struct trace *trace,
2360                           struct perf_evsel *evsel,
2361                           union perf_event *event __maybe_unused,
2362                           struct perf_sample *sample)
2363 {
2364         struct thread *thread;
2365         struct addr_location al;
2366         char map_type = 'd';
2367         struct thread_trace *ttrace;
2368         int err = -1;
2369         int callchain_ret = 0;
2370
2371         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2372
2373         if (sample->callchain) {
2374                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2375                 if (callchain_ret == 0) {
2376                         if (callchain_cursor.nr < trace->min_stack)
2377                                 goto out_put;
2378                         callchain_ret = 1;
2379                 }
2380         }
2381
2382         ttrace = thread__trace(thread, trace->output);
2383         if (ttrace == NULL)
2384                 goto out_put;
2385
2386         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2387                 ttrace->pfmaj++;
2388         else
2389                 ttrace->pfmin++;
2390
2391         if (trace->summary_only)
2392                 goto out;
2393
2394         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2395
2396         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2397
2398         fprintf(trace->output, "%sfault [",
2399                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2400                 "maj" : "min");
2401
2402         print_location(trace->output, sample, &al, false, true);
2403
2404         fprintf(trace->output, "] => ");
2405
2406         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2407
2408         if (!al.map) {
2409                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2410
2411                 if (al.map)
2412                         map_type = 'x';
2413                 else
2414                         map_type = '?';
2415         }
2416
2417         print_location(trace->output, sample, &al, true, false);
2418
2419         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2420
2421         if (callchain_ret > 0)
2422                 trace__fprintf_callchain(trace, sample);
2423         else if (callchain_ret < 0)
2424                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2425
2426         ++trace->nr_events_printed;
2427 out:
2428         err = 0;
2429 out_put:
2430         thread__put(thread);
2431         return err;
2432 }
2433
2434 static void trace__set_base_time(struct trace *trace,
2435                                  struct perf_evsel *evsel,
2436                                  struct perf_sample *sample)
2437 {
2438         /*
2439          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2440          * and don't use sample->time unconditionally, we may end up having
2441          * some other event in the future without PERF_SAMPLE_TIME for good
2442          * reason, i.e. we may not be interested in its timestamps, just in
2443          * it taking place, picking some piece of information when it
2444          * appears in our event stream (vfs_getname comes to mind).
2445          */
2446         if (trace->base_time == 0 && !trace->full_time &&
2447             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2448                 trace->base_time = sample->time;
2449 }
2450
2451 static int trace__process_sample(struct perf_tool *tool,
2452                                  union perf_event *event,
2453                                  struct perf_sample *sample,
2454                                  struct perf_evsel *evsel,
2455                                  struct machine *machine __maybe_unused)
2456 {
2457         struct trace *trace = container_of(tool, struct trace, tool);
2458         struct thread *thread;
2459         int err = 0;
2460
2461         tracepoint_handler handler = evsel->handler;
2462
2463         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2464         if (thread && thread__is_filtered(thread))
2465                 goto out;
2466
2467         trace__set_base_time(trace, evsel, sample);
2468
2469         if (handler) {
2470                 ++trace->nr_events;
2471                 handler(trace, evsel, event, sample);
2472         }
2473 out:
2474         thread__put(thread);
2475         return err;
2476 }
2477
2478 static int trace__record(struct trace *trace, int argc, const char **argv)
2479 {
2480         unsigned int rec_argc, i, j;
2481         const char **rec_argv;
2482         const char * const record_args[] = {
2483                 "record",
2484                 "-R",
2485                 "-m", "1024",
2486                 "-c", "1",
2487         };
2488
2489         const char * const sc_args[] = { "-e", };
2490         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2491         const char * const majpf_args[] = { "-e", "major-faults" };
2492         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2493         const char * const minpf_args[] = { "-e", "minor-faults" };
2494         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2495
2496         /* +1 is for the event string below */
2497         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2498                 majpf_args_nr + minpf_args_nr + argc;
2499         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2500
2501         if (rec_argv == NULL)
2502                 return -ENOMEM;
2503
2504         j = 0;
2505         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2506                 rec_argv[j++] = record_args[i];
2507
2508         if (trace->trace_syscalls) {
2509                 for (i = 0; i < sc_args_nr; i++)
2510                         rec_argv[j++] = sc_args[i];
2511
2512                 /* event string may be different for older kernels - e.g., RHEL6 */
2513                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2514                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2515                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2516                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2517                 else {
2518                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2519                         free(rec_argv);
2520                         return -1;
2521                 }
2522         }
2523
2524         if (trace->trace_pgfaults & TRACE_PFMAJ)
2525                 for (i = 0; i < majpf_args_nr; i++)
2526                         rec_argv[j++] = majpf_args[i];
2527
2528         if (trace->trace_pgfaults & TRACE_PFMIN)
2529                 for (i = 0; i < minpf_args_nr; i++)
2530                         rec_argv[j++] = minpf_args[i];
2531
2532         for (i = 0; i < (unsigned int)argc; i++)
2533                 rec_argv[j++] = argv[i];
2534
2535         return cmd_record(j, rec_argv);
2536 }
2537
2538 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2539
2540 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2541 {
2542         bool found = false;
2543         struct perf_evsel *evsel, *tmp;
2544         struct parse_events_error err = { .idx = 0, };
2545         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2546
2547         if (ret)
2548                 return false;
2549
2550         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2551                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2552                         continue;
2553
2554                 if (perf_evsel__field(evsel, "pathname")) {
2555                         evsel->handler = trace__vfs_getname;
2556                         found = true;
2557                         continue;
2558                 }
2559
2560                 list_del_init(&evsel->node);
2561                 evsel->evlist = NULL;
2562                 perf_evsel__delete(evsel);
2563         }
2564
2565         return found;
2566 }
2567
2568 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2569 {
2570         struct perf_evsel *evsel;
2571         struct perf_event_attr attr = {
2572                 .type = PERF_TYPE_SOFTWARE,
2573                 .mmap_data = 1,
2574         };
2575
2576         attr.config = config;
2577         attr.sample_period = 1;
2578
2579         event_attr_init(&attr);
2580
2581         evsel = perf_evsel__new(&attr);
2582         if (evsel)
2583                 evsel->handler = trace__pgfault;
2584
2585         return evsel;
2586 }
2587
2588 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2589 {
2590         const u32 type = event->header.type;
2591         struct perf_evsel *evsel;
2592
2593         if (type != PERF_RECORD_SAMPLE) {
2594                 trace__process_event(trace, trace->host, event, sample);
2595                 return;
2596         }
2597
2598         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2599         if (evsel == NULL) {
2600                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2601                 return;
2602         }
2603
2604         trace__set_base_time(trace, evsel, sample);
2605
2606         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2607             sample->raw_data == NULL) {
2608                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2609                        perf_evsel__name(evsel), sample->tid,
2610                        sample->cpu, sample->raw_size);
2611         } else {
2612                 tracepoint_handler handler = evsel->handler;
2613                 handler(trace, evsel, event, sample);
2614         }
2615
2616         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2617                 interrupted = true;
2618 }
2619
2620 static int trace__add_syscall_newtp(struct trace *trace)
2621 {
2622         int ret = -1;
2623         struct perf_evlist *evlist = trace->evlist;
2624         struct perf_evsel *sys_enter, *sys_exit;
2625
2626         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2627         if (sys_enter == NULL)
2628                 goto out;
2629
2630         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2631                 goto out_delete_sys_enter;
2632
2633         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2634         if (sys_exit == NULL)
2635                 goto out_delete_sys_enter;
2636
2637         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2638                 goto out_delete_sys_exit;
2639
2640         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2641         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2642
2643         perf_evlist__add(evlist, sys_enter);
2644         perf_evlist__add(evlist, sys_exit);
2645
2646         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2647                 /*
2648                  * We're interested only in the user space callchain
2649                  * leading to the syscall, allow overriding that for
2650                  * debugging reasons using --kernel_syscall_callchains
2651                  */
2652                 sys_exit->attr.exclude_callchain_kernel = 1;
2653         }
2654
2655         trace->syscalls.events.sys_enter = sys_enter;
2656         trace->syscalls.events.sys_exit  = sys_exit;
2657
2658         ret = 0;
2659 out:
2660         return ret;
2661
2662 out_delete_sys_exit:
2663         perf_evsel__delete_priv(sys_exit);
2664 out_delete_sys_enter:
2665         perf_evsel__delete_priv(sys_enter);
2666         goto out;
2667 }
2668
2669 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2670 {
2671         int err = -1;
2672         struct perf_evsel *sys_exit;
2673         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2674                                                 trace->ev_qualifier_ids.nr,
2675                                                 trace->ev_qualifier_ids.entries);
2676
2677         if (filter == NULL)
2678                 goto out_enomem;
2679
2680         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2681                                           filter)) {
2682                 sys_exit = trace->syscalls.events.sys_exit;
2683                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2684         }
2685
2686         free(filter);
2687 out:
2688         return err;
2689 out_enomem:
2690         errno = ENOMEM;
2691         goto out;
2692 }
2693
2694 #ifdef HAVE_LIBBPF_SUPPORT
2695 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2696 {
2697         int fd = bpf_map__fd(trace->syscalls.map);
2698         struct bpf_map_syscall_entry value = {
2699                 .enabled = !trace->not_ev_qualifier,
2700         };
2701         int err = 0;
2702         size_t i;
2703
2704         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2705                 int key = trace->ev_qualifier_ids.entries[i];
2706
2707                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2708                 if (err)
2709                         break;
2710         }
2711
2712         return err;
2713 }
2714
2715 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2716 {
2717         int fd = bpf_map__fd(trace->syscalls.map);
2718         struct bpf_map_syscall_entry value = {
2719                 .enabled = enabled,
2720         };
2721         int err = 0, key;
2722
2723         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2724                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2725                 if (err)
2726                         break;
2727         }
2728
2729         return err;
2730 }
2731
2732 static int trace__init_syscalls_bpf_map(struct trace *trace)
2733 {
2734         bool enabled = true;
2735
2736         if (trace->ev_qualifier_ids.nr)
2737                 enabled = trace->not_ev_qualifier;
2738
2739         return __trace__init_syscalls_bpf_map(trace, enabled);
2740 }
2741 #else
2742 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2743 {
2744         return 0;
2745 }
2746
2747 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2748 {
2749         return 0;
2750 }
2751 #endif // HAVE_LIBBPF_SUPPORT
2752
2753 static int trace__set_ev_qualifier_filter(struct trace *trace)
2754 {
2755         if (trace->syscalls.map)
2756                 return trace__set_ev_qualifier_bpf_filter(trace);
2757         if (trace->syscalls.events.sys_enter)
2758                 return trace__set_ev_qualifier_tp_filter(trace);
2759         return 0;
2760 }
2761
2762 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2763                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2764 {
2765         int err = 0;
2766 #ifdef HAVE_LIBBPF_SUPPORT
2767         bool value = true;
2768         int map_fd = bpf_map__fd(map);
2769         size_t i;
2770
2771         for (i = 0; i < npids; ++i) {
2772                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2773                 if (err)
2774                         break;
2775         }
2776 #endif
2777         return err;
2778 }
2779
2780 static int trace__set_filter_loop_pids(struct trace *trace)
2781 {
2782         unsigned int nr = 1, err;
2783         pid_t pids[32] = {
2784                 getpid(),
2785         };
2786         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2787
2788         while (thread && nr < ARRAY_SIZE(pids)) {
2789                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2790
2791                 if (parent == NULL)
2792                         break;
2793
2794                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2795                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2796                         pids[nr++] = parent->tid;
2797                         break;
2798                 }
2799                 thread = parent;
2800         }
2801
2802         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2803         if (!err && trace->filter_pids.map)
2804                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2805
2806         return err;
2807 }
2808
2809 static int trace__set_filter_pids(struct trace *trace)
2810 {
2811         int err = 0;
2812         /*
2813          * Better not use !target__has_task() here because we need to cover the
2814          * case where no threads were specified in the command line, but a
2815          * workload was, and in that case we will fill in the thread_map when
2816          * we fork the workload in perf_evlist__prepare_workload.
2817          */
2818         if (trace->filter_pids.nr > 0) {
2819                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2820                                                       trace->filter_pids.entries);
2821                 if (!err && trace->filter_pids.map) {
2822                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2823                                                        trace->filter_pids.entries);
2824                 }
2825         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2826                 err = trace__set_filter_loop_pids(trace);
2827         }
2828
2829         return err;
2830 }
2831
2832 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2833 {
2834         struct perf_evlist *evlist = trace->evlist;
2835         struct perf_sample sample;
2836         int err;
2837
2838         err = perf_evlist__parse_sample(evlist, event, &sample);
2839         if (err)
2840                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2841         else
2842                 trace__handle_event(trace, event, &sample);
2843
2844         return 0;
2845 }
2846
2847 static int __trace__flush_events(struct trace *trace)
2848 {
2849         u64 first = ordered_events__first_time(&trace->oe.data);
2850         u64 flush = trace->oe.last - NSEC_PER_SEC;
2851
2852         /* Is there some thing to flush.. */
2853         if (first && first < flush)
2854                 return ordered_events__flush_time(&trace->oe.data, flush);
2855
2856         return 0;
2857 }
2858
2859 static int trace__flush_events(struct trace *trace)
2860 {
2861         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2862 }
2863
2864 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2865 {
2866         int err;
2867
2868         if (!trace->sort_events)
2869                 return __trace__deliver_event(trace, event);
2870
2871         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2872         if (err && err != -1)
2873                 return err;
2874
2875         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2876         if (err)
2877                 return err;
2878
2879         return trace__flush_events(trace);
2880 }
2881
2882 static int ordered_events__deliver_event(struct ordered_events *oe,
2883                                          struct ordered_event *event)
2884 {
2885         struct trace *trace = container_of(oe, struct trace, oe.data);
2886
2887         return __trace__deliver_event(trace, event->event);
2888 }
2889
2890 static int trace__run(struct trace *trace, int argc, const char **argv)
2891 {
2892         struct perf_evlist *evlist = trace->evlist;
2893         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2894         int err = -1, i;
2895         unsigned long before;
2896         const bool forks = argc > 0;
2897         bool draining = false;
2898
2899         trace->live = true;
2900
2901         if (!trace->raw_augmented_syscalls) {
2902                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2903                         goto out_error_raw_syscalls;
2904
2905                 if (trace->trace_syscalls)
2906                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2907         }
2908
2909         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2910                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2911                 if (pgfault_maj == NULL)
2912                         goto out_error_mem;
2913                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2914                 perf_evlist__add(evlist, pgfault_maj);
2915         }
2916
2917         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2918                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2919                 if (pgfault_min == NULL)
2920                         goto out_error_mem;
2921                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2922                 perf_evlist__add(evlist, pgfault_min);
2923         }
2924
2925         if (trace->sched &&
2926             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2927                                    trace__sched_stat_runtime))
2928                 goto out_error_sched_stat_runtime;
2929
2930         /*
2931          * If a global cgroup was set, apply it to all the events without an
2932          * explicit cgroup. I.e.:
2933          *
2934          *      trace -G A -e sched:*switch
2935          *
2936          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2937          * _and_ sched:sched_switch to the 'A' cgroup, while:
2938          *
2939          * trace -e sched:*switch -G A
2940          *
2941          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2942          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2943          * a cgroup (on the root cgroup, sys wide, etc).
2944          *
2945          * Multiple cgroups:
2946          *
2947          * trace -G A -e sched:*switch -G B
2948          *
2949          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2950          * to the 'B' cgroup.
2951          *
2952          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2953          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2954          */
2955         if (trace->cgroup)
2956                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2957
2958         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2959         if (err < 0) {
2960                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2961                 goto out_delete_evlist;
2962         }
2963
2964         err = trace__symbols_init(trace, evlist);
2965         if (err < 0) {
2966                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2967                 goto out_delete_evlist;
2968         }
2969
2970         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2971
2972         signal(SIGCHLD, sig_handler);
2973         signal(SIGINT, sig_handler);
2974
2975         if (forks) {
2976                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2977                                                     argv, false, NULL);
2978                 if (err < 0) {
2979                         fprintf(trace->output, "Couldn't run the workload!\n");
2980                         goto out_delete_evlist;
2981                 }
2982         }
2983
2984         err = perf_evlist__open(evlist);
2985         if (err < 0)
2986                 goto out_error_open;
2987
2988         err = bpf__apply_obj_config();
2989         if (err) {
2990                 char errbuf[BUFSIZ];
2991
2992                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2993                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2994                          errbuf);
2995                 goto out_error_open;
2996         }
2997
2998         err = trace__set_filter_pids(trace);
2999         if (err < 0)
3000                 goto out_error_mem;
3001
3002         if (trace->syscalls.map)
3003                 trace__init_syscalls_bpf_map(trace);
3004
3005         if (trace->ev_qualifier_ids.nr > 0) {
3006                 err = trace__set_ev_qualifier_filter(trace);
3007                 if (err < 0)
3008                         goto out_errno;
3009
3010                 if (trace->syscalls.events.sys_exit) {
3011                         pr_debug("event qualifier tracepoint filter: %s\n",
3012                                  trace->syscalls.events.sys_exit->filter);
3013                 }
3014         }
3015
3016         err = perf_evlist__apply_filters(evlist, &evsel);
3017         if (err < 0)
3018                 goto out_error_apply_filters;
3019
3020         if (trace->dump.map)
3021                 bpf_map__fprintf(trace->dump.map, trace->output);
3022
3023         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3024         if (err < 0)
3025                 goto out_error_mmap;
3026
3027         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3028                 perf_evlist__enable(evlist);
3029
3030         if (forks)
3031                 perf_evlist__start_workload(evlist);
3032
3033         if (trace->opts.initial_delay) {
3034                 usleep(trace->opts.initial_delay * 1000);
3035                 perf_evlist__enable(evlist);
3036         }
3037
3038         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3039                                   evlist->threads->nr > 1 ||
3040                                   perf_evlist__first(evlist)->attr.inherit;
3041
3042         /*
3043          * Now that we already used evsel->attr to ask the kernel to setup the
3044          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3045          * trace__resolve_callchain(), allowing per-event max-stack settings
3046          * to override an explicitly set --max-stack global setting.
3047          */
3048         evlist__for_each_entry(evlist, evsel) {
3049                 if (evsel__has_callchain(evsel) &&
3050                     evsel->attr.sample_max_stack == 0)
3051                         evsel->attr.sample_max_stack = trace->max_stack;
3052         }
3053 again:
3054         before = trace->nr_events;
3055
3056         for (i = 0; i < evlist->nr_mmaps; i++) {
3057                 union perf_event *event;
3058                 struct perf_mmap *md;
3059
3060                 md = &evlist->mmap[i];
3061                 if (perf_mmap__read_init(md) < 0)
3062                         continue;
3063
3064                 while ((event = perf_mmap__read_event(md)) != NULL) {
3065                         ++trace->nr_events;
3066
3067                         err = trace__deliver_event(trace, event);
3068                         if (err)
3069                                 goto out_disable;
3070
3071                         perf_mmap__consume(md);
3072
3073                         if (interrupted)
3074                                 goto out_disable;
3075
3076                         if (done && !draining) {
3077                                 perf_evlist__disable(evlist);
3078                                 draining = true;
3079                         }
3080                 }
3081                 perf_mmap__read_done(md);
3082         }
3083
3084         if (trace->nr_events == before) {
3085                 int timeout = done ? 100 : -1;
3086
3087                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3088                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3089                                 draining = true;
3090
3091                         goto again;
3092                 } else {
3093                         if (trace__flush_events(trace))
3094                                 goto out_disable;
3095                 }
3096         } else {
3097                 goto again;
3098         }
3099
3100 out_disable:
3101         thread__zput(trace->current);
3102
3103         perf_evlist__disable(evlist);
3104
3105         if (trace->sort_events)
3106                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3107
3108         if (!err) {
3109                 if (trace->summary)
3110                         trace__fprintf_thread_summary(trace, trace->output);
3111
3112                 if (trace->show_tool_stats) {
3113                         fprintf(trace->output, "Stats:\n "
3114                                                " vfs_getname : %" PRIu64 "\n"
3115                                                " proc_getname: %" PRIu64 "\n",
3116                                 trace->stats.vfs_getname,
3117                                 trace->stats.proc_getname);
3118                 }
3119         }
3120
3121 out_delete_evlist:
3122         trace__symbols__exit(trace);
3123
3124         perf_evlist__delete(evlist);
3125         cgroup__put(trace->cgroup);
3126         trace->evlist = NULL;
3127         trace->live = false;
3128         return err;
3129 {
3130         char errbuf[BUFSIZ];
3131
3132 out_error_sched_stat_runtime:
3133         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3134         goto out_error;
3135
3136 out_error_raw_syscalls:
3137         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3138         goto out_error;
3139
3140 out_error_mmap:
3141         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3142         goto out_error;
3143
3144 out_error_open:
3145         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3146
3147 out_error:
3148         fprintf(trace->output, "%s\n", errbuf);
3149         goto out_delete_evlist;
3150
3151 out_error_apply_filters:
3152         fprintf(trace->output,
3153                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3154                 evsel->filter, perf_evsel__name(evsel), errno,
3155                 str_error_r(errno, errbuf, sizeof(errbuf)));
3156         goto out_delete_evlist;
3157 }
3158 out_error_mem:
3159         fprintf(trace->output, "Not enough memory to run!\n");
3160         goto out_delete_evlist;
3161
3162 out_errno:
3163         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3164         goto out_delete_evlist;
3165 }
3166
3167 static int trace__replay(struct trace *trace)
3168 {
3169         const struct perf_evsel_str_handler handlers[] = {
3170                 { "probe:vfs_getname",       trace__vfs_getname, },
3171         };
3172         struct perf_data data = {
3173                 .path  = input_name,
3174                 .mode  = PERF_DATA_MODE_READ,
3175                 .force = trace->force,
3176         };
3177         struct perf_session *session;
3178         struct perf_evsel *evsel;
3179         int err = -1;
3180
3181         trace->tool.sample        = trace__process_sample;
3182         trace->tool.mmap          = perf_event__process_mmap;
3183         trace->tool.mmap2         = perf_event__process_mmap2;
3184         trace->tool.comm          = perf_event__process_comm;
3185         trace->tool.exit          = perf_event__process_exit;
3186         trace->tool.fork          = perf_event__process_fork;
3187         trace->tool.attr          = perf_event__process_attr;
3188         trace->tool.tracing_data  = perf_event__process_tracing_data;
3189         trace->tool.build_id      = perf_event__process_build_id;
3190         trace->tool.namespaces    = perf_event__process_namespaces;
3191
3192         trace->tool.ordered_events = true;
3193         trace->tool.ordering_requires_timestamps = true;
3194
3195         /* add tid to output */
3196         trace->multiple_threads = true;
3197
3198         session = perf_session__new(&data, false, &trace->tool);
3199         if (session == NULL)
3200                 return -1;
3201
3202         if (trace->opts.target.pid)
3203                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3204
3205         if (trace->opts.target.tid)
3206                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3207
3208         if (symbol__init(&session->header.env) < 0)
3209                 goto out;
3210
3211         trace->host = &session->machines.host;
3212
3213         err = perf_session__set_tracepoints_handlers(session, handlers);
3214         if (err)
3215                 goto out;
3216
3217         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3218                                                      "raw_syscalls:sys_enter");
3219         /* older kernels have syscalls tp versus raw_syscalls */
3220         if (evsel == NULL)
3221                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3222                                                              "syscalls:sys_enter");
3223
3224         if (evsel &&
3225             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3226             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3227                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3228                 goto out;
3229         }
3230
3231         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3232                                                      "raw_syscalls:sys_exit");
3233         if (evsel == NULL)
3234                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3235                                                              "syscalls:sys_exit");
3236         if (evsel &&
3237             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3238             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3239                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3240                 goto out;
3241         }
3242
3243         evlist__for_each_entry(session->evlist, evsel) {
3244                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3245                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3246                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3247                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3248                         evsel->handler = trace__pgfault;
3249         }
3250
3251         setup_pager();
3252
3253         err = perf_session__process_events(session);
3254         if (err)
3255                 pr_err("Failed to process events, error %d", err);
3256
3257         else if (trace->summary)
3258                 trace__fprintf_thread_summary(trace, trace->output);
3259
3260 out:
3261         perf_session__delete(session);
3262
3263         return err;
3264 }
3265
3266 static size_t trace__fprintf_threads_header(FILE *fp)
3267 {
3268         size_t printed;
3269
3270         printed  = fprintf(fp, "\n Summary of events:\n\n");
3271
3272         return printed;
3273 }
3274
3275 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3276         struct stats    *stats;
3277         double          msecs;
3278         int             syscall;
3279 )
3280 {
3281         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3282         struct stats *stats = source->priv;
3283
3284         entry->syscall = source->i;
3285         entry->stats   = stats;
3286         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3287 }
3288
3289 static size_t thread__dump_stats(struct thread_trace *ttrace,
3290                                  struct trace *trace, FILE *fp)
3291 {
3292         size_t printed = 0;
3293         struct syscall *sc;
3294         struct rb_node *nd;
3295         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3296
3297         if (syscall_stats == NULL)
3298                 return 0;
3299
3300         printed += fprintf(fp, "\n");
3301
3302         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3303         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3304         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3305
3306         resort_rb__for_each_entry(nd, syscall_stats) {
3307                 struct stats *stats = syscall_stats_entry->stats;
3308                 if (stats) {
3309                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3310                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3311                         double avg = avg_stats(stats);
3312                         double pct;
3313                         u64 n = (u64) stats->n;
3314
3315                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3316                         avg /= NSEC_PER_MSEC;
3317
3318                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3319                         printed += fprintf(fp, "   %-15s", sc->name);
3320                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3321                                            n, syscall_stats_entry->msecs, min, avg);
3322                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3323                 }
3324         }
3325
3326         resort_rb__delete(syscall_stats);
3327         printed += fprintf(fp, "\n\n");
3328
3329         return printed;
3330 }
3331
3332 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3333 {
3334         size_t printed = 0;
3335         struct thread_trace *ttrace = thread__priv(thread);
3336         double ratio;
3337
3338         if (ttrace == NULL)
3339                 return 0;
3340
3341         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3342
3343         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3344         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3345         printed += fprintf(fp, "%.1f%%", ratio);
3346         if (ttrace->pfmaj)
3347                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3348         if (ttrace->pfmin)
3349                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3350         if (trace->sched)
3351                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3352         else if (fputc('\n', fp) != EOF)
3353                 ++printed;
3354
3355         printed += thread__dump_stats(ttrace, trace, fp);
3356
3357         return printed;
3358 }
3359
3360 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3361 {
3362         return ttrace ? ttrace->nr_events : 0;
3363 }
3364
3365 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3366         struct thread *thread;
3367 )
3368 {
3369         entry->thread = rb_entry(nd, struct thread, rb_node);
3370 }
3371
3372 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3373 {
3374         size_t printed = trace__fprintf_threads_header(fp);
3375         struct rb_node *nd;
3376         int i;
3377
3378         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3379                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3380
3381                 if (threads == NULL) {
3382                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3383                         return 0;
3384                 }
3385
3386                 resort_rb__for_each_entry(nd, threads)
3387                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3388
3389                 resort_rb__delete(threads);
3390         }
3391         return printed;
3392 }
3393
3394 static int trace__set_duration(const struct option *opt, const char *str,
3395                                int unset __maybe_unused)
3396 {
3397         struct trace *trace = opt->value;
3398
3399         trace->duration_filter = atof(str);
3400         return 0;
3401 }
3402
3403 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3404                                               int unset __maybe_unused)
3405 {
3406         int ret = -1;
3407         size_t i;
3408         struct trace *trace = opt->value;
3409         /*
3410          * FIXME: introduce a intarray class, plain parse csv and create a
3411          * { int nr, int entries[] } struct...
3412          */
3413         struct intlist *list = intlist__new(str);
3414
3415         if (list == NULL)
3416                 return -1;
3417
3418         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3419         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3420
3421         if (trace->filter_pids.entries == NULL)
3422                 goto out;
3423
3424         trace->filter_pids.entries[0] = getpid();
3425
3426         for (i = 1; i < trace->filter_pids.nr; ++i)
3427                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3428
3429         intlist__delete(list);
3430         ret = 0;
3431 out:
3432         return ret;
3433 }
3434
3435 static int trace__open_output(struct trace *trace, const char *filename)
3436 {
3437         struct stat st;
3438
3439         if (!stat(filename, &st) && st.st_size) {
3440                 char oldname[PATH_MAX];
3441
3442                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3443                 unlink(oldname);
3444                 rename(filename, oldname);
3445         }
3446
3447         trace->output = fopen(filename, "w");
3448
3449         return trace->output == NULL ? -errno : 0;
3450 }
3451
3452 static int parse_pagefaults(const struct option *opt, const char *str,
3453                             int unset __maybe_unused)
3454 {
3455         int *trace_pgfaults = opt->value;
3456
3457         if (strcmp(str, "all") == 0)
3458                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3459         else if (strcmp(str, "maj") == 0)
3460                 *trace_pgfaults |= TRACE_PFMAJ;
3461         else if (strcmp(str, "min") == 0)
3462                 *trace_pgfaults |= TRACE_PFMIN;
3463         else
3464                 return -1;
3465
3466         return 0;
3467 }
3468
3469 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3470 {
3471         struct perf_evsel *evsel;
3472
3473         evlist__for_each_entry(evlist, evsel)
3474                 evsel->handler = handler;
3475 }
3476
3477 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3478 {
3479         struct perf_evsel *evsel;
3480
3481         evlist__for_each_entry(evlist, evsel) {
3482                 if (evsel->priv || !evsel->tp_format)
3483                         continue;
3484
3485                 if (strcmp(evsel->tp_format->system, "syscalls"))
3486                         continue;
3487
3488                 if (perf_evsel__init_syscall_tp(evsel))
3489                         return -1;
3490
3491                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3492                         struct syscall_tp *sc = evsel->priv;
3493
3494                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3495                                 return -1;
3496                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3497                         struct syscall_tp *sc = evsel->priv;
3498
3499                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3500                                 return -1;
3501                 }
3502         }
3503
3504         return 0;
3505 }
3506
3507 /*
3508  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3509  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3510  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3511  *
3512  * It'd be better to introduce a parse_options() variant that would return a
3513  * list with the terms it didn't match to an event...
3514  */
3515 static int trace__parse_events_option(const struct option *opt, const char *str,
3516                                       int unset __maybe_unused)
3517 {
3518         struct trace *trace = (struct trace *)opt->value;
3519         const char *s = str;
3520         char *sep = NULL, *lists[2] = { NULL, NULL, };
3521         int len = strlen(str) + 1, err = -1, list, idx;
3522         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3523         char group_name[PATH_MAX];
3524         struct syscall_fmt *fmt;
3525
3526         if (strace_groups_dir == NULL)
3527                 return -1;
3528
3529         if (*s == '!') {
3530                 ++s;
3531                 trace->not_ev_qualifier = true;
3532         }
3533
3534         while (1) {
3535                 if ((sep = strchr(s, ',')) != NULL)
3536                         *sep = '\0';
3537
3538                 list = 0;
3539                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3540                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3541                         list = 1;
3542                         goto do_concat;
3543                 }
3544
3545                 fmt = syscall_fmt__find_by_alias(s);
3546                 if (fmt != NULL) {
3547                         list = 1;
3548                         s = fmt->name;
3549                 } else {
3550                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3551                         if (access(group_name, R_OK) == 0)
3552                                 list = 1;
3553                 }
3554 do_concat:
3555                 if (lists[list]) {
3556                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3557                 } else {
3558                         lists[list] = malloc(len);
3559                         if (lists[list] == NULL)
3560                                 goto out;
3561                         strcpy(lists[list], s);
3562                 }
3563
3564                 if (!sep)
3565                         break;
3566
3567                 *sep = ',';
3568                 s = sep + 1;
3569         }
3570
3571         if (lists[1] != NULL) {
3572                 struct strlist_config slist_config = {
3573                         .dirname = strace_groups_dir,
3574                 };
3575
3576                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3577                 if (trace->ev_qualifier == NULL) {
3578                         fputs("Not enough memory to parse event qualifier", trace->output);
3579                         goto out;
3580                 }
3581
3582                 if (trace__validate_ev_qualifier(trace))
3583                         goto out;
3584                 trace->trace_syscalls = true;
3585         }
3586
3587         err = 0;
3588
3589         if (lists[0]) {
3590                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3591                                                "event selector. use 'perf list' to list available events",
3592                                                parse_events_option);
3593                 err = parse_events_option(&o, lists[0], 0);
3594         }
3595 out:
3596         if (sep)
3597                 *sep = ',';
3598
3599         return err;
3600 }
3601
3602 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3603 {
3604         struct trace *trace = opt->value;
3605
3606         if (!list_empty(&trace->evlist->entries))
3607                 return parse_cgroups(opt, str, unset);
3608
3609         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3610
3611         return 0;
3612 }
3613
3614 static struct bpf_map *bpf__find_map_by_name(const char *name)
3615 {
3616         struct bpf_object *obj, *tmp;
3617
3618         bpf_object__for_each_safe(obj, tmp) {
3619                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3620                 if (map)
3621                         return map;
3622
3623         }
3624
3625         return NULL;
3626 }
3627
3628 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3629 {
3630         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3631 }
3632
3633 static void trace__set_bpf_map_syscalls(struct trace *trace)
3634 {
3635         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3636 }
3637
3638 static int trace__config(const char *var, const char *value, void *arg)
3639 {
3640         struct trace *trace = arg;
3641         int err = 0;
3642
3643         if (!strcmp(var, "trace.add_events")) {
3644                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3645                                                "event selector. use 'perf list' to list available events",
3646                                                parse_events_option);
3647                 err = parse_events_option(&o, value, 0);
3648         } else if (!strcmp(var, "trace.show_timestamp")) {
3649                 trace->show_tstamp = perf_config_bool(var, value);
3650         } else if (!strcmp(var, "trace.show_duration")) {
3651                 trace->show_duration = perf_config_bool(var, value);
3652         } else if (!strcmp(var, "trace.show_arg_names")) {
3653                 trace->show_arg_names = perf_config_bool(var, value);
3654                 if (!trace->show_arg_names)
3655                         trace->show_zeros = true;
3656         } else if (!strcmp(var, "trace.show_zeros")) {
3657                 bool new_show_zeros = perf_config_bool(var, value);
3658                 if (!trace->show_arg_names && !new_show_zeros) {
3659                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3660                         goto out;
3661                 }
3662                 trace->show_zeros = new_show_zeros;
3663         } else if (!strcmp(var, "trace.show_prefix")) {
3664                 trace->show_string_prefix = perf_config_bool(var, value);
3665         } else if (!strcmp(var, "trace.no_inherit")) {
3666                 trace->opts.no_inherit = perf_config_bool(var, value);
3667         } else if (!strcmp(var, "trace.args_alignment")) {
3668                 int args_alignment = 0;
3669                 if (perf_config_int(&args_alignment, var, value) == 0)
3670                         trace->args_alignment = args_alignment;
3671         }
3672 out:
3673         return err;
3674 }
3675
3676 int cmd_trace(int argc, const char **argv)
3677 {
3678         const char *trace_usage[] = {
3679                 "perf trace [<options>] [<command>]",
3680                 "perf trace [<options>] -- <command> [<options>]",
3681                 "perf trace record [<options>] [<command>]",
3682                 "perf trace record [<options>] -- <command> [<options>]",
3683                 NULL
3684         };
3685         struct trace trace = {
3686                 .syscalls = {
3687                         . max = -1,
3688                 },
3689                 .opts = {
3690                         .target = {
3691                                 .uid       = UINT_MAX,
3692                                 .uses_mmap = true,
3693                         },
3694                         .user_freq     = UINT_MAX,
3695                         .user_interval = ULLONG_MAX,
3696                         .no_buffering  = true,
3697                         .mmap_pages    = UINT_MAX,
3698                 },
3699                 .output = stderr,
3700                 .show_comm = true,
3701                 .show_tstamp = true,
3702                 .show_duration = true,
3703                 .show_arg_names = true,
3704                 .args_alignment = 70,
3705                 .trace_syscalls = false,
3706                 .kernel_syscallchains = false,
3707                 .max_stack = UINT_MAX,
3708                 .max_events = ULONG_MAX,
3709         };
3710         const char *map_dump_str = NULL;
3711         const char *output_name = NULL;
3712         const struct option trace_options[] = {
3713         OPT_CALLBACK('e', "event", &trace, "event",
3714                      "event/syscall selector. use 'perf list' to list available events",
3715                      trace__parse_events_option),
3716         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3717                     "show the thread COMM next to its id"),
3718         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3719         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3720                      trace__parse_events_option),
3721         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3722         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3723         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3724                     "trace events on existing process id"),
3725         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3726                     "trace events on existing thread id"),
3727         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3728                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3729         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3730                     "system-wide collection from all CPUs"),
3731         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3732                     "list of cpus to monitor"),
3733         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3734                     "child tasks do not inherit counters"),
3735         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3736                      "number of mmap data pages",
3737                      perf_evlist__parse_mmap_pages),
3738         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3739                    "user to profile"),
3740         OPT_CALLBACK(0, "duration", &trace, "float",
3741                      "show only events with duration > N.M ms",
3742                      trace__set_duration),
3743 #ifdef HAVE_LIBBPF_SUPPORT
3744         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3745 #endif
3746         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3747         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3748         OPT_BOOLEAN('T', "time", &trace.full_time,
3749                     "Show full timestamp, not time relative to first start"),
3750         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3751                     "Show only syscalls that failed"),
3752         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3753                     "Show only syscall summary with statistics"),
3754         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3755                     "Show all syscalls and summary with statistics"),
3756         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3757                      "Trace pagefaults", parse_pagefaults, "maj"),
3758         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3759         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3760         OPT_CALLBACK(0, "call-graph", &trace.opts,
3761                      "record_mode[,record_size]", record_callchain_help,
3762                      &record_parse_callchain_opt),
3763         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3764                     "Show the kernel callchains on the syscall exit path"),
3765         OPT_ULONG(0, "max-events", &trace.max_events,
3766                 "Set the maximum number of events to print, exit after that is reached. "),
3767         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3768                      "Set the minimum stack depth when parsing the callchain, "
3769                      "anything below the specified depth will be ignored."),
3770         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3771                      "Set the maximum stack depth when parsing the callchain, "
3772                      "anything beyond the specified depth will be ignored. "
3773                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3774         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3775                         "Sort batch of events before processing, use if getting out of order events"),
3776         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3777                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3778         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3779                         "per thread proc mmap processing timeout in ms"),
3780         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3781                      trace__parse_cgroups),
3782         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3783                      "ms to wait before starting measurement after program "
3784                      "start"),
3785         OPT_END()
3786         };
3787         bool __maybe_unused max_stack_user_set = true;
3788         bool mmap_pages_user_set = true;
3789         struct perf_evsel *evsel;
3790         const char * const trace_subcommands[] = { "record", NULL };
3791         int err = -1;
3792         char bf[BUFSIZ];
3793
3794         signal(SIGSEGV, sighandler_dump_stack);
3795         signal(SIGFPE, sighandler_dump_stack);
3796
3797         trace.evlist = perf_evlist__new();
3798         trace.sctbl = syscalltbl__new();
3799
3800         if (trace.evlist == NULL || trace.sctbl == NULL) {
3801                 pr_err("Not enough memory to run!\n");
3802                 err = -ENOMEM;
3803                 goto out;
3804         }
3805
3806         err = perf_config(trace__config, &trace);
3807         if (err)
3808                 goto out;
3809
3810         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3811                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3812
3813         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3814                 usage_with_options_msg(trace_usage, trace_options,
3815                                        "cgroup monitoring only available in system-wide mode");
3816         }
3817
3818         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3819         if (IS_ERR(evsel)) {
3820                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3821                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3822                 goto out;
3823         }
3824
3825         if (evsel) {
3826                 trace.syscalls.events.augmented = evsel;
3827                 trace__set_bpf_map_filtered_pids(&trace);
3828                 trace__set_bpf_map_syscalls(&trace);
3829         }
3830
3831         err = bpf__setup_stdout(trace.evlist);
3832         if (err) {
3833                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3834                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3835                 goto out;
3836         }
3837
3838         err = -1;
3839
3840         if (map_dump_str) {
3841                 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3842                 if (trace.dump.map == NULL) {
3843                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3844                         goto out;
3845                 }
3846         }
3847
3848         if (trace.trace_pgfaults) {
3849                 trace.opts.sample_address = true;
3850                 trace.opts.sample_time = true;
3851         }
3852
3853         if (trace.opts.mmap_pages == UINT_MAX)
3854                 mmap_pages_user_set = false;
3855
3856         if (trace.max_stack == UINT_MAX) {
3857                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3858                 max_stack_user_set = false;
3859         }
3860
3861 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3862         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3863                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3864         }
3865 #endif
3866
3867         if (callchain_param.enabled) {
3868                 if (!mmap_pages_user_set && geteuid() == 0)
3869                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3870
3871                 symbol_conf.use_callchain = true;
3872         }
3873
3874         if (trace.evlist->nr_entries > 0) {
3875                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3876                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3877                         perror("failed to set syscalls:* tracepoint fields");
3878                         goto out;
3879                 }
3880         }
3881
3882         if (trace.sort_events) {
3883                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3884                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3885         }
3886
3887         /*
3888          * If we are augmenting syscalls, then combine what we put in the
3889          * __augmented_syscalls__ BPF map with what is in the
3890          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3891          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3892          *
3893          * We'll switch to look at two BPF maps, one for sys_enter and the
3894          * other for sys_exit when we start augmenting the sys_exit paths with
3895          * buffers that are being copied from kernel to userspace, think 'read'
3896          * syscall.
3897          */
3898         if (trace.syscalls.events.augmented) {
3899                 evlist__for_each_entry(trace.evlist, evsel) {
3900                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3901
3902                         if (raw_syscalls_sys_exit) {
3903                                 trace.raw_augmented_syscalls = true;
3904                                 goto init_augmented_syscall_tp;
3905                         }
3906
3907                         if (trace.syscalls.events.augmented->priv == NULL &&
3908                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3909                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3910                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3911                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3912                                         goto out;
3913                                 augmented->handler = trace__sys_enter;
3914                         }
3915
3916                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3917                                 struct syscall_tp *sc;
3918 init_augmented_syscall_tp:
3919                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3920                                         goto out;
3921                                 sc = evsel->priv;
3922                                 /*
3923                                  * For now with BPF raw_augmented we hook into
3924                                  * raw_syscalls:sys_enter and there we get all
3925                                  * 6 syscall args plus the tracepoint common
3926                                  * fields and the syscall_nr (another long).
3927                                  * So we check if that is the case and if so
3928                                  * don't look after the sc->args_size but
3929                                  * always after the full raw_syscalls:sys_enter
3930                                  * payload, which is fixed.
3931                                  *
3932                                  * We'll revisit this later to pass
3933                                  * s->args_size to the BPF augmenter (now
3934                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3935                                  * so that it copies only what we need for each
3936                                  * syscall, like what happens when we use
3937                                  * syscalls:sys_enter_NAME, so that we reduce
3938                                  * the kernel/userspace traffic to just what is
3939                                  * needed for each syscall.
3940                                  */
3941                                 if (trace.raw_augmented_syscalls)
3942                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3943                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3944                                 evsel->handler = trace__sys_exit;
3945                         }
3946                 }
3947         }
3948
3949         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3950                 return trace__record(&trace, argc-1, &argv[1]);
3951
3952         /* summary_only implies summary option, but don't overwrite summary if set */
3953         if (trace.summary_only)
3954                 trace.summary = trace.summary_only;
3955
3956         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3957             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3958                 trace.trace_syscalls = true;
3959         }
3960
3961         if (output_name != NULL) {
3962                 err = trace__open_output(&trace, output_name);
3963                 if (err < 0) {
3964                         perror("failed to create output file");
3965                         goto out;
3966                 }
3967         }
3968
3969         err = target__validate(&trace.opts.target);
3970         if (err) {
3971                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3972                 fprintf(trace.output, "%s", bf);
3973                 goto out_close;
3974         }
3975
3976         err = target__parse_uid(&trace.opts.target);
3977         if (err) {
3978                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3979                 fprintf(trace.output, "%s", bf);
3980                 goto out_close;
3981         }
3982
3983         if (!argc && target__none(&trace.opts.target))
3984                 trace.opts.target.system_wide = true;
3985
3986         if (input_name)
3987                 err = trace__replay(&trace);
3988         else
3989                 err = trace__run(&trace, argc, argv);
3990
3991 out_close:
3992         if (output_name != NULL)
3993                 fclose(trace.output);
3994 out:
3995         return err;
3996 }