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
  17 #include "util/record.h"
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "util/rlimit.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/dso.h"
  29 #include "util/env.h"
  30 #include "util/event.h"
  31 #include "util/evsel.h"
  32 #include "util/evsel_fprintf.h"
  33 #include "util/synthetic-events.h"
  34 #include "util/evlist.h"
  35 #include "util/evswitch.h"
  36 #include "util/mmap.h"
  37 #include <subcmd/pager.h>
  38 #include <subcmd/exec-cmd.h>
  39 #include "util/machine.h"
  40 #include "util/map.h"
  41 #include "util/symbol.h"
  42 #include "util/path.h"
  43 #include "util/session.h"
  44 #include "util/thread.h"
  45 #include <subcmd/parse-options.h>
  46 #include "util/strlist.h"
  47 #include "util/intlist.h"
  48 #include "util/thread_map.h"
  49 #include "util/stat.h"
  50 #include "util/tool.h"
  51 #include "util/util.h"
  52 #include "trace/beauty/beauty.h"
  53 #include "trace-event.h"
  54 #include "util/parse-events.h"
  55 #include "util/bpf-loader.h"
  56 #include "callchain.h"
  57 #include "print_binary.h"
  58 #include "string2.h"
  59 #include "syscalltbl.h"
  60 #include "rb_resort.h"
  61 #include "../perf.h"
  62
  63 #include <errno.h>
  64 #include <inttypes.h>
  65 #include <poll.h>
  66 #include <signal.h>
  67 #include <stdlib.h>
  68 #include <string.h>
  69 #include <linux/err.h>
  70 #include <linux/filter.h>
  71 #include <linux/kernel.h>
  72 #include <linux/random.h>
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/zalloc.h>
  76 #include <fcntl.h>
  77 #include <sys/sysmacros.h>
  78
  79 #include <linux/ctype.h>
  80 #include <perf/mmap.h>
  81
  82 #ifndef O_CLOEXEC
  83 # define O_CLOEXEC              02000000
  84 #endif
  85
  86 #ifndef F_LINUX_SPECIFIC_BASE
  87 # define F_LINUX_SPECIFIC_BASE  1024
  88 #endif
  89
  90 /*
  91  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  92  */
  93 struct syscall_arg_fmt {
  94         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
  95         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
  96         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
  97         void       *parm;
  98         const char *name;
  99         u16        nr_entries; // for arrays
 100         bool       show_zero;
 101 };
 102
 103 struct syscall_fmt {
 104         const char *name;
 105         const char *alias;
 106         struct {
 107                 const char *sys_enter,
 108                            *sys_exit;
 109         }          bpf_prog_name;
 110         struct syscall_arg_fmt arg[6];
 111         u8         nr_args;
 112         bool       errpid;
 113         bool       timeout;
 114         bool       hexret;
 115 };
 116
 117 struct trace {
 118         struct perf_tool        tool;
 119         struct syscalltbl       *sctbl;
 120         struct {
 121                 struct syscall  *table;
 122                 struct bpf_map  *map;
 123                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 124                         struct bpf_map  *sys_enter,
 125                                         *sys_exit;
 126                 }               prog_array;
 127                 struct {
 128                         struct evsel *sys_enter,
 129                                           *sys_exit,
 130                                           *augmented;
 131                 }               events;
 132                 struct bpf_program *unaugmented_prog;
 133         } syscalls;
 134         struct {
 135                 struct bpf_map *map;
 136         } dump;
 137         struct record_opts      opts;
 138         struct evlist   *evlist;
 139         struct machine          *host;
 140         struct thread           *current;
 141         struct bpf_object       *bpf_obj;
 142         struct cgroup           *cgroup;
 143         u64                     base_time;
 144         FILE                    *output;
 145         unsigned long           nr_events;
 146         unsigned long           nr_events_printed;
 147         unsigned long           max_events;
 148         struct evswitch         evswitch;
 149         struct strlist          *ev_qualifier;
 150         struct {
 151                 size_t          nr;
 152                 int             *entries;
 153         }                       ev_qualifier_ids;
 154         struct {
 155                 size_t          nr;
 156                 pid_t           *entries;
 157                 struct bpf_map  *map;
 158         }                       filter_pids;
 159         double                  duration_filter;
 160         double                  runtime_ms;
 161         struct {
 162                 u64             vfs_getname,
 163                                 proc_getname;
 164         } stats;
 165         unsigned int            max_stack;
 166         unsigned int            min_stack;
 167         int                     raw_augmented_syscalls_args_size;
 168         bool                    raw_augmented_syscalls;
 169         bool                    fd_path_disabled;
 170         bool                    sort_events;
 171         bool                    not_ev_qualifier;
 172         bool                    live;
 173         bool                    full_time;
 174         bool                    sched;
 175         bool                    multiple_threads;
 176         bool                    summary;
 177         bool                    summary_only;
 178         bool                    errno_summary;
 179         bool                    failure_only;
 180         bool                    show_comm;
 181         bool                    print_sample;
 182         bool                    show_tool_stats;
 183         bool                    trace_syscalls;
 184         bool                    libtraceevent_print;
 185         bool                    kernel_syscallchains;
 186         s16                     args_alignment;
 187         bool                    show_tstamp;
 188         bool                    show_duration;
 189         bool                    show_zeros;
 190         bool                    show_arg_names;
 191         bool                    show_string_prefix;
 192         bool                    force;
 193         bool                    vfs_getname;
 194         int                     trace_pgfaults;
 195         char                    *perfconfig_events;
 196         struct {
 197                 struct ordered_events   data;
 198                 u64                     last;
 199         } oe;
 200 };
 201
 202 struct tp_field {
 203         int offset;
 204         union {
 205                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 206                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 207         };
 208 };
 209
 210 #define TP_UINT_FIELD(bits) \
 211 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 212 { \
 213         u##bits value; \
 214         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 215         return value;  \
 216 }
 217
 218 TP_UINT_FIELD(8);
 219 TP_UINT_FIELD(16);
 220 TP_UINT_FIELD(32);
 221 TP_UINT_FIELD(64);
 222
 223 #define TP_UINT_FIELD__SWAPPED(bits) \
 224 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 225 { \
 226         u##bits value; \
 227         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 228         return bswap_##bits(value);\
 229 }
 230
 231 TP_UINT_FIELD__SWAPPED(16);
 232 TP_UINT_FIELD__SWAPPED(32);
 233 TP_UINT_FIELD__SWAPPED(64);
 234
 235 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 236 {
 237         field->offset = offset;
 238
 239         switch (size) {
 240         case 1:
 241                 field->integer = tp_field__u8;
 242                 break;
 243         case 2:
 244                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 245                 break;
 246         case 4:
 247                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 248                 break;
 249         case 8:
 250                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 251                 break;
 252         default:
 253                 return -1;
 254         }
 255
 256         return 0;
 257 }
 258
 259 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 260 {
 261         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 262 }
 263
 264 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 265 {
 266         return sample->raw_data + field->offset;
 267 }
 268
 269 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 270 {
 271         field->offset = offset;
 272         field->pointer = tp_field__ptr;
 273         return 0;
 274 }
 275
 276 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 277 {
 278         return __tp_field__init_ptr(field, format_field->offset);
 279 }
 280
 281 struct syscall_tp {
 282         struct tp_field id;
 283         union {
 284                 struct tp_field args, ret;
 285         };
 286 };
 287
 288 /*
 289  * The evsel->priv as used by 'perf trace'
 290  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 291  * fmt: for all the other tracepoints
 292  */
 293 struct evsel_trace {
 294         struct syscall_tp       sc;
 295         struct syscall_arg_fmt  *fmt;
 296 };
 297
 298 static struct evsel_trace *evsel_trace__new(void)
 299 {
 300         return zalloc(sizeof(struct evsel_trace));
 301 }
 302
 303 static void evsel_trace__delete(struct evsel_trace *et)
 304 {
 305         if (et == NULL)
 306                 return;
 307
 308         zfree(&et->fmt);
 309         free(et);
 310 }
 311
 312 /*
 313  * Used with raw_syscalls:sys_{enter,exit} and with the
 314  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 315  */
 316 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 317 {
 318         struct evsel_trace *et = evsel->priv;
 319
 320         return &et->sc;
 321 }
 322
 323 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 324 {
 325         if (evsel->priv == NULL) {
 326                 evsel->priv = evsel_trace__new();
 327                 if (evsel->priv == NULL)
 328                         return NULL;
 329         }
 330
 331         return __evsel__syscall_tp(evsel);
 332 }
 333
 334 /*
 335  * Used with all the other tracepoints.
 336  */
 337 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 338 {
 339         struct evsel_trace *et = evsel->priv;
 340
 341         return et->fmt;
 342 }
 343
 344 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 345 {
 346         struct evsel_trace *et = evsel->priv;
 347
 348         if (evsel->priv == NULL) {
 349                 et = evsel->priv = evsel_trace__new();
 350
 351                 if (et == NULL)
 352                         return NULL;
 353         }
 354
 355         if (et->fmt == NULL) {
 356                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 357                 if (et->fmt == NULL)
 358                         goto out_delete;
 359         }
 360
 361         return __evsel__syscall_arg_fmt(evsel);
 362
 363 out_delete:
 364         evsel_trace__delete(evsel->priv);
 365         evsel->priv = NULL;
 366         return NULL;
 367 }
 368
 369 static int perf_evsel__init_tp_uint_field(struct evsel *evsel,
 370                                           struct tp_field *field,
 371                                           const char *name)
 372 {
 373         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 374
 375         if (format_field == NULL)
 376                 return -1;
 377
 378         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 379 }
 380
 381 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 382         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 383            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 384
 385 static int perf_evsel__init_tp_ptr_field(struct evsel *evsel,
 386                                          struct tp_field *field,
 387                                          const char *name)
 388 {
 389         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 390
 391         if (format_field == NULL)
 392                 return -1;
 393
 394         return tp_field__init_ptr(field, format_field);
 395 }
 396
 397 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 398         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 399            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 400
 401 static void evsel__delete_priv(struct evsel *evsel)
 402 {
 403         zfree(&evsel->priv);
 404         evsel__delete(evsel);
 405 }
 406
 407 static int perf_evsel__init_syscall_tp(struct evsel *evsel)
 408 {
 409         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 410
 411         if (sc != NULL) {
 412                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 413                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 414                         return -ENOENT;
 415                 return 0;
 416         }
 417
 418         return -ENOMEM;
 419 }
 420
 421 static int perf_evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 422 {
 423         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 424
 425         if (sc != NULL) {
 426                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 427                 if (syscall_id == NULL)
 428                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 429                 if (syscall_id == NULL ||
 430                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 431                         return -EINVAL;
 432
 433                 return 0;
 434         }
 435
 436         return -ENOMEM;
 437 }
 438
 439 static int perf_evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 440 {
 441         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 442
 443         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 444 }
 445
 446 static int perf_evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 447 {
 448         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 449
 450         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 451 }
 452
 453 static int perf_evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 454 {
 455         if (evsel__syscall_tp(evsel) != NULL) {
 456                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 457                         return -ENOENT;
 458
 459                 evsel->handler = handler;
 460                 return 0;
 461         }
 462
 463         return -ENOMEM;
 464 }
 465
 466 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 467 {
 468         struct evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 469
 470         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 471         if (IS_ERR(evsel))
 472                 evsel = perf_evsel__newtp("syscalls", direction);
 473
 474         if (IS_ERR(evsel))
 475                 return NULL;
 476
 477         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 478                 goto out_delete;
 479
 480         return evsel;
 481
 482 out_delete:
 483         evsel__delete_priv(evsel);
 484         return NULL;
 485 }
 486
 487 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 488         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 489            fields->name.integer(&fields->name, sample); })
 490
 491 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 492         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 493            fields->name.pointer(&fields->name, sample); })
 494
 495 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 496 {
 497         int idx = val - sa->offset;
 498
 499         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 500                 size_t printed = scnprintf(bf, size, intfmt, val);
 501                 if (show_suffix)
 502                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 503                 return printed;
 504         }
 505
 506         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 507 }
 508
 509 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 510 {
 511         int idx = val - sa->offset;
 512
 513         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 514                 size_t printed = scnprintf(bf, size, intfmt, val);
 515                 if (show_prefix)
 516                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 517                 return printed;
 518         }
 519
 520         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 521 }
 522
 523 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 524                                                 const char *intfmt,
 525                                                 struct syscall_arg *arg)
 526 {
 527         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 528 }
 529
 530 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 531                                               struct syscall_arg *arg)
 532 {
 533         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 534 }
 535
 536 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 537
 538 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 539 {
 540         return strarray__strtoul(arg->parm, bf, size, ret);
 541 }
 542
 543 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 544 {
 545         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 546 }
 547
 548 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 549 {
 550         return strarrays__strtoul(arg->parm, bf, size, ret);
 551 }
 552
 553 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 554 {
 555         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 556 }
 557
 558 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 559 {
 560         size_t printed;
 561         int i;
 562
 563         for (i = 0; i < sas->nr_entries; ++i) {
 564                 struct strarray *sa = sas->entries[i];
 565                 int idx = val - sa->offset;
 566
 567                 if (idx >= 0 && idx < sa->nr_entries) {
 568                         if (sa->entries[idx] == NULL)
 569                                 break;
 570                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 571                 }
 572         }
 573
 574         printed = scnprintf(bf, size, intfmt, val);
 575         if (show_prefix)
 576                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 577         return printed;
 578 }
 579
 580 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 581 {
 582         int i;
 583
 584         for (i = 0; i < sa->nr_entries; ++i) {
 585                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 586                         *ret = sa->offset + i;
 587                         return true;
 588                 }
 589         }
 590
 591         return false;
 592 }
 593
 594 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 595 {
 596         u64 val = 0;
 597         char *tok = bf, *sep, *end;
 598
 599         *ret = 0;
 600
 601         while (size != 0) {
 602                 int toklen = size;
 603
 604                 sep = memchr(tok, '|', size);
 605                 if (sep != NULL) {
 606                         size -= sep - tok + 1;
 607
 608                         end = sep - 1;
 609                         while (end > tok && isspace(*end))
 610                                 --end;
 611
 612                         toklen = end - tok + 1;
 613                 }
 614
 615                 while (isspace(*tok))
 616                         ++tok;
 617
 618                 if (isalpha(*tok) || *tok == '_') {
 619                         if (!strarray__strtoul(sa, tok, toklen, &val))
 620                                 return false;
 621                 } else {
 622                         bool is_hexa = tok[0] == 0 && (tok[1] = 'x' || tok[1] == 'X');
 623
 624                         val = strtoul(tok, NULL, is_hexa ? 16 : 0);
 625                 }
 626
 627                 *ret |= (1 << (val - 1));
 628
 629                 if (sep == NULL)
 630                         break;
 631                 tok = sep + 1;
 632         }
 633
 634         return true;
 635 }
 636
 637 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 638 {
 639         int i;
 640
 641         for (i = 0; i < sas->nr_entries; ++i) {
 642                 struct strarray *sa = sas->entries[i];
 643
 644                 if (strarray__strtoul(sa, bf, size, ret))
 645                         return true;
 646         }
 647
 648         return false;
 649 }
 650
 651 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 652                                         struct syscall_arg *arg)
 653 {
 654         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 655 }
 656
 657 #ifndef AT_FDCWD
 658 #define AT_FDCWD        -100
 659 #endif
 660
 661 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 662                                            struct syscall_arg *arg)
 663 {
 664         int fd = arg->val;
 665         const char *prefix = "AT_FD";
 666
 667         if (fd == AT_FDCWD)
 668                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 669
 670         return syscall_arg__scnprintf_fd(bf, size, arg);
 671 }
 672
 673 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 674
 675 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 676                                               struct syscall_arg *arg);
 677
 678 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 679
 680 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 681 {
 682         return scnprintf(bf, size, "%#lx", arg->val);
 683 }
 684
 685 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 686 {
 687         if (arg->val == 0)
 688                 return scnprintf(bf, size, "NULL");
 689         return syscall_arg__scnprintf_hex(bf, size, arg);
 690 }
 691
 692 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 693 {
 694         return scnprintf(bf, size, "%d", arg->val);
 695 }
 696
 697 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 698 {
 699         return scnprintf(bf, size, "%ld", arg->val);
 700 }
 701
 702 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 703 {
 704         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 705         //     fill missing comms using thread__set_comm()...
 706         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 707         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 708 }
 709
 710 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 711
 712 static const char *bpf_cmd[] = {
 713         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 714         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 715 };
 716 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 717
 718 static const char *fsmount_flags[] = {
 719         [1] = "CLOEXEC",
 720 };
 721 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 722
 723 #include "trace/beauty/generated/fsconfig_arrays.c"
 724
 725 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 726
 727 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 728 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 729
 730 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 731 static DEFINE_STRARRAY(itimers, "ITIMER_");
 732
 733 static const char *keyctl_options[] = {
 734         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 735         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 736         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 737         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 738         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 739 };
 740 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 741
 742 static const char *whences[] = { "SET", "CUR", "END",
 743 #ifdef SEEK_DATA
 744 "DATA",
 745 #endif
 746 #ifdef SEEK_HOLE
 747 "HOLE",
 748 #endif
 749 };
 750 static DEFINE_STRARRAY(whences, "SEEK_");
 751
 752 static const char *fcntl_cmds[] = {
 753         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 754         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 755         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 756         "GETOWNER_UIDS",
 757 };
 758 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 759
 760 static const char *fcntl_linux_specific_cmds[] = {
 761         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 762         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 763         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 764 };
 765
 766 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 767
 768 static struct strarray *fcntl_cmds_arrays[] = {
 769         &strarray__fcntl_cmds,
 770         &strarray__fcntl_linux_specific_cmds,
 771 };
 772
 773 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 774
 775 static const char *rlimit_resources[] = {
 776         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 777         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 778         "RTTIME",
 779 };
 780 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 781
 782 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 783 static DEFINE_STRARRAY(sighow, "SIG_");
 784
 785 static const char *clockid[] = {
 786         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 787         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 788         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 789 };
 790 static DEFINE_STRARRAY(clockid, "CLOCK_");
 791
 792 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 793                                                  struct syscall_arg *arg)
 794 {
 795         bool show_prefix = arg->show_string_prefix;
 796         const char *suffix = "_OK";
 797         size_t printed = 0;
 798         int mode = arg->val;
 799
 800         if (mode == F_OK) /* 0 */
 801                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 802 #define P_MODE(n) \
 803         if (mode & n##_OK) { \
 804                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 805                 mode &= ~n##_OK; \
 806         }
 807
 808         P_MODE(R);
 809         P_MODE(W);
 810         P_MODE(X);
 811 #undef P_MODE
 812
 813         if (mode)
 814                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 815
 816         return printed;
 817 }
 818
 819 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 820
 821 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 822                                               struct syscall_arg *arg);
 823
 824 #define SCA_FILENAME syscall_arg__scnprintf_filename
 825
 826 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 827                                                 struct syscall_arg *arg)
 828 {
 829         bool show_prefix = arg->show_string_prefix;
 830         const char *prefix = "O_";
 831         int printed = 0, flags = arg->val;
 832
 833 #define P_FLAG(n) \
 834         if (flags & O_##n) { \
 835                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 836                 flags &= ~O_##n; \
 837         }
 838
 839         P_FLAG(CLOEXEC);
 840         P_FLAG(NONBLOCK);
 841 #undef P_FLAG
 842
 843         if (flags)
 844                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 845
 846         return printed;
 847 }
 848
 849 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 850
 851 #ifndef GRND_NONBLOCK
 852 #define GRND_NONBLOCK   0x0001
 853 #endif
 854 #ifndef GRND_RANDOM
 855 #define GRND_RANDOM     0x0002
 856 #endif
 857
 858 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 859                                                    struct syscall_arg *arg)
 860 {
 861         bool show_prefix = arg->show_string_prefix;
 862         const char *prefix = "GRND_";
 863         int printed = 0, flags = arg->val;
 864
 865 #define P_FLAG(n) \
 866         if (flags & GRND_##n) { \
 867                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 868                 flags &= ~GRND_##n; \
 869         }
 870
 871         P_FLAG(RANDOM);
 872         P_FLAG(NONBLOCK);
 873 #undef P_FLAG
 874
 875         if (flags)
 876                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 877
 878         return printed;
 879 }
 880
 881 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 882
 883 #define STRARRAY(name, array) \
 884           { .scnprintf  = SCA_STRARRAY, \
 885             .strtoul    = STUL_STRARRAY, \
 886             .parm       = &strarray__##array, }
 887
 888 #define STRARRAY_FLAGS(name, array) \
 889           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 890             .strtoul    = STUL_STRARRAY_FLAGS, \
 891             .parm       = &strarray__##array, }
 892
 893 #include "trace/beauty/arch_errno_names.c"
 894 #include "trace/beauty/eventfd.c"
 895 #include "trace/beauty/futex_op.c"
 896 #include "trace/beauty/futex_val3.c"
 897 #include "trace/beauty/mmap.c"
 898 #include "trace/beauty/mode_t.c"
 899 #include "trace/beauty/msg_flags.c"
 900 #include "trace/beauty/open_flags.c"
 901 #include "trace/beauty/perf_event_open.c"
 902 #include "trace/beauty/pid.c"
 903 #include "trace/beauty/sched_policy.c"
 904 #include "trace/beauty/seccomp.c"
 905 #include "trace/beauty/signum.c"
 906 #include "trace/beauty/socket_type.c"
 907 #include "trace/beauty/waitid_options.c"
 908
 909 static struct syscall_fmt syscall_fmts[] = {
 910         { .name     = "access",
 911           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 912         { .name     = "arch_prctl",
 913           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 914                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 915         { .name     = "bind",
 916           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 917                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 918                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 919         { .name     = "bpf",
 920           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 921         { .name     = "brk",        .hexret = true,
 922           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 923         { .name     = "clock_gettime",
 924           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 925         { .name     = "clone",      .errpid = true, .nr_args = 5,
 926           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 927                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 928                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 929                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 930                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 931         { .name     = "close",
 932           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 933         { .name     = "connect",
 934           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 935                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 936                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 937         { .name     = "epoll_ctl",
 938           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 939         { .name     = "eventfd2",
 940           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 941         { .name     = "fchmodat",
 942           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 943         { .name     = "fchownat",
 944           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 945         { .name     = "fcntl",
 946           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 947                            .strtoul   = STUL_STRARRAYS,
 948                            .parm      = &strarrays__fcntl_cmds_arrays,
 949                            .show_zero = true, },
 950                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 951         { .name     = "flock",
 952           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 953         { .name     = "fsconfig",
 954           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 955         { .name     = "fsmount",
 956           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 957                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 958         { .name     = "fspick",
 959           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 960                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 961                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 962         { .name     = "fstat", .alias = "newfstat", },
 963         { .name     = "fstatat", .alias = "newfstatat", },
 964         { .name     = "futex",
 965           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 966                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 967         { .name     = "futimesat",
 968           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 969         { .name     = "getitimer",
 970           .arg = { [0] = STRARRAY(which, itimers), }, },
 971         { .name     = "getpid",     .errpid = true, },
 972         { .name     = "getpgid",    .errpid = true, },
 973         { .name     = "getppid",    .errpid = true, },
 974         { .name     = "getrandom",
 975           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 976         { .name     = "getrlimit",
 977           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 978         { .name     = "gettid",     .errpid = true, },
 979         { .name     = "ioctl",
 980           .arg = {
 981 #if defined(__i386__) || defined(__x86_64__)
 982 /*
 983  * FIXME: Make this available to all arches.
 984  */
 985                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 986                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 987 #else
 988                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 989 #endif
 990         { .name     = "kcmp",       .nr_args = 5,
 991           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 992                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 993                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 994                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 995                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 996         { .name     = "keyctl",
 997           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 998         { .name     = "kill",
 999           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1000         { .name     = "linkat",
1001           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1002         { .name     = "lseek",
1003           .arg = { [2] = STRARRAY(whence, whences), }, },
1004         { .name     = "lstat", .alias = "newlstat", },
1005         { .name     = "madvise",
1006           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1007                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1008         { .name     = "mkdirat",
1009           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1010         { .name     = "mknodat",
1011           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1012         { .name     = "mmap",       .hexret = true,
1013 /* The standard mmap maps to old_mmap on s390x */
1014 #if defined(__s390x__)
1015         .alias = "old_mmap",
1016 #endif
1017           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1018                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1019                            .strtoul   = STUL_STRARRAY_FLAGS,
1020                            .parm      = &strarray__mmap_flags, },
1021                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1022         { .name     = "mount",
1023           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1024                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1025                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1026         { .name     = "move_mount",
1027           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1028                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1029                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1030                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1031                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1032         { .name     = "mprotect",
1033           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1034                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1035         { .name     = "mq_unlink",
1036           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1037         { .name     = "mremap",     .hexret = true,
1038           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1039         { .name     = "name_to_handle_at",
1040           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1041         { .name     = "newfstatat",
1042           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1043         { .name     = "open",
1044           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1045         { .name     = "open_by_handle_at",
1046           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1047                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1048         { .name     = "openat",
1049           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1050                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1051         { .name     = "perf_event_open",
1052           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1053                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1054                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1055         { .name     = "pipe2",
1056           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1057         { .name     = "pkey_alloc",
1058           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1059         { .name     = "pkey_free",
1060           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1061         { .name     = "pkey_mprotect",
1062           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1063                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1064                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1065         { .name     = "poll", .timeout = true, },
1066         { .name     = "ppoll", .timeout = true, },
1067         { .name     = "prctl",
1068           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
1069                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1070                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1071         { .name     = "pread", .alias = "pread64", },
1072         { .name     = "preadv", .alias = "pread", },
1073         { .name     = "prlimit64",
1074           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1075         { .name     = "pwrite", .alias = "pwrite64", },
1076         { .name     = "readlinkat",
1077           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1078         { .name     = "recvfrom",
1079           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1080         { .name     = "recvmmsg",
1081           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1082         { .name     = "recvmsg",
1083           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1084         { .name     = "renameat",
1085           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1086                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1087         { .name     = "renameat2",
1088           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1089                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1090                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1091         { .name     = "rt_sigaction",
1092           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1093         { .name     = "rt_sigprocmask",
1094           .arg = { [0] = STRARRAY(how, sighow), }, },
1095         { .name     = "rt_sigqueueinfo",
1096           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1097         { .name     = "rt_tgsigqueueinfo",
1098           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1099         { .name     = "sched_setscheduler",
1100           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1101         { .name     = "seccomp",
1102           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1103                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1104         { .name     = "select", .timeout = true, },
1105         { .name     = "sendfile", .alias = "sendfile64", },
1106         { .name     = "sendmmsg",
1107           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1108         { .name     = "sendmsg",
1109           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1110         { .name     = "sendto",
1111           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1112                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1113         { .name     = "set_tid_address", .errpid = true, },
1114         { .name     = "setitimer",
1115           .arg = { [0] = STRARRAY(which, itimers), }, },
1116         { .name     = "setrlimit",
1117           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1118         { .name     = "socket",
1119           .arg = { [0] = STRARRAY(family, socket_families),
1120                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1121                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1122         { .name     = "socketpair",
1123           .arg = { [0] = STRARRAY(family, socket_families),
1124                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1125                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1126         { .name     = "stat", .alias = "newstat", },
1127         { .name     = "statx",
1128           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1129                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1130                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1131         { .name     = "swapoff",
1132           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1133         { .name     = "swapon",
1134           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1135         { .name     = "symlinkat",
1136           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1137         { .name     = "sync_file_range",
1138           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1139         { .name     = "tgkill",
1140           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1141         { .name     = "tkill",
1142           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1143         { .name     = "umount2", .alias = "umount",
1144           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1145         { .name     = "uname", .alias = "newuname", },
1146         { .name     = "unlinkat",
1147           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1148         { .name     = "utimensat",
1149           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1150         { .name     = "wait4",      .errpid = true,
1151           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1152         { .name     = "waitid",     .errpid = true,
1153           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1154 };
1155
1156 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1157 {
1158         const struct syscall_fmt *fmt = fmtp;
1159         return strcmp(name, fmt->name);
1160 }
1161
1162 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1163 {
1164         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1165 }
1166
1167 static struct syscall_fmt *syscall_fmt__find(const char *name)
1168 {
1169         const int nmemb = ARRAY_SIZE(syscall_fmts);
1170         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1171 }
1172
1173 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1174 {
1175         int i;
1176
1177         for (i = 0; i < nmemb; ++i) {
1178                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1179                         return &fmts[i];
1180         }
1181
1182         return NULL;
1183 }
1184
1185 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1186 {
1187         const int nmemb = ARRAY_SIZE(syscall_fmts);
1188         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1189 }
1190
1191 /*
1192  * is_exit: is this "exit" or "exit_group"?
1193  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1194  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1195  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1196  */
1197 struct syscall {
1198         struct tep_event    *tp_format;
1199         int                 nr_args;
1200         int                 args_size;
1201         struct {
1202                 struct bpf_program *sys_enter,
1203                                    *sys_exit;
1204         }                   bpf_prog;
1205         bool                is_exit;
1206         bool                is_open;
1207         bool                nonexistent;
1208         struct tep_format_field *args;
1209         const char          *name;
1210         struct syscall_fmt  *fmt;
1211         struct syscall_arg_fmt *arg_fmt;
1212 };
1213
1214 /*
1215  * Must match what is in the BPF program:
1216  *
1217  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1218  */
1219 struct bpf_map_syscall_entry {
1220         bool    enabled;
1221         u16     string_args_len[6];
1222 };
1223
1224 /*
1225  * We need to have this 'calculated' boolean because in some cases we really
1226  * don't know what is the duration of a syscall, for instance, when we start
1227  * a session and some threads are waiting for a syscall to finish, say 'poll',
1228  * in which case all we can do is to print "( ? ) for duration and for the
1229  * start timestamp.
1230  */
1231 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1232 {
1233         double duration = (double)t / NSEC_PER_MSEC;
1234         size_t printed = fprintf(fp, "(");
1235
1236         if (!calculated)
1237                 printed += fprintf(fp, "         ");
1238         else if (duration >= 1.0)
1239                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1240         else if (duration >= 0.01)
1241                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1242         else
1243                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1244         return printed + fprintf(fp, "): ");
1245 }
1246
1247 /**
1248  * filename.ptr: The filename char pointer that will be vfs_getname'd
1249  * filename.entry_str_pos: Where to insert the string translated from
1250  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1251  * ret_scnprintf: syscall args may set this to a different syscall return
1252  *                formatter, for instance, fcntl may return fds, file flags, etc.
1253  */
1254 struct thread_trace {
1255         u64               entry_time;
1256         bool              entry_pending;
1257         unsigned long     nr_events;
1258         unsigned long     pfmaj, pfmin;
1259         char              *entry_str;
1260         double            runtime_ms;
1261         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1262         struct {
1263                 unsigned long ptr;
1264                 short int     entry_str_pos;
1265                 bool          pending_open;
1266                 unsigned int  namelen;
1267                 char          *name;
1268         } filename;
1269         struct {
1270                 int           max;
1271                 struct file   *table;
1272         } files;
1273
1274         struct intlist *syscall_stats;
1275 };
1276
1277 static struct thread_trace *thread_trace__new(void)
1278 {
1279         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1280
1281         if (ttrace) {
1282                 ttrace->files.max = -1;
1283                 ttrace->syscall_stats = intlist__new(NULL);
1284         }
1285
1286         return ttrace;
1287 }
1288
1289 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1290 {
1291         struct thread_trace *ttrace;
1292
1293         if (thread == NULL)
1294                 goto fail;
1295
1296         if (thread__priv(thread) == NULL)
1297                 thread__set_priv(thread, thread_trace__new());
1298
1299         if (thread__priv(thread) == NULL)
1300                 goto fail;
1301
1302         ttrace = thread__priv(thread);
1303         ++ttrace->nr_events;
1304
1305         return ttrace;
1306 fail:
1307         color_fprintf(fp, PERF_COLOR_RED,
1308                       "WARNING: not enough memory, dropping samples!\n");
1309         return NULL;
1310 }
1311
1312
1313 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1314                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1315 {
1316         struct thread_trace *ttrace = thread__priv(arg->thread);
1317
1318         ttrace->ret_scnprintf = ret_scnprintf;
1319 }
1320
1321 #define TRACE_PFMAJ             (1 << 0)
1322 #define TRACE_PFMIN             (1 << 1)
1323
1324 static const size_t trace__entry_str_size = 2048;
1325
1326 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1327 {
1328         if (fd < 0)
1329                 return NULL;
1330
1331         if (fd > ttrace->files.max) {
1332                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1333
1334                 if (nfiles == NULL)
1335                         return NULL;
1336
1337                 if (ttrace->files.max != -1) {
1338                         memset(nfiles + ttrace->files.max + 1, 0,
1339                                (fd - ttrace->files.max) * sizeof(struct file));
1340                 } else {
1341                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1342                 }
1343
1344                 ttrace->files.table = nfiles;
1345                 ttrace->files.max   = fd;
1346         }
1347
1348         return ttrace->files.table + fd;
1349 }
1350
1351 struct file *thread__files_entry(struct thread *thread, int fd)
1352 {
1353         return thread_trace__files_entry(thread__priv(thread), fd);
1354 }
1355
1356 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1357 {
1358         struct thread_trace *ttrace = thread__priv(thread);
1359         struct file *file = thread_trace__files_entry(ttrace, fd);
1360
1361         if (file != NULL) {
1362                 struct stat st;
1363                 if (stat(pathname, &st) == 0)
1364                         file->dev_maj = major(st.st_rdev);
1365                 file->pathname = strdup(pathname);
1366                 if (file->pathname)
1367                         return 0;
1368         }
1369
1370         return -1;
1371 }
1372
1373 static int thread__read_fd_path(struct thread *thread, int fd)
1374 {
1375         char linkname[PATH_MAX], pathname[PATH_MAX];
1376         struct stat st;
1377         int ret;
1378
1379         if (thread->pid_ == thread->tid) {
1380                 scnprintf(linkname, sizeof(linkname),
1381                           "/proc/%d/fd/%d", thread->pid_, fd);
1382         } else {
1383                 scnprintf(linkname, sizeof(linkname),
1384                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1385         }
1386
1387         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1388                 return -1;
1389
1390         ret = readlink(linkname, pathname, sizeof(pathname));
1391
1392         if (ret < 0 || ret > st.st_size)
1393                 return -1;
1394
1395         pathname[ret] = '\0';
1396         return trace__set_fd_pathname(thread, fd, pathname);
1397 }
1398
1399 static const char *thread__fd_path(struct thread *thread, int fd,
1400                                    struct trace *trace)
1401 {
1402         struct thread_trace *ttrace = thread__priv(thread);
1403
1404         if (ttrace == NULL || trace->fd_path_disabled)
1405                 return NULL;
1406
1407         if (fd < 0)
1408                 return NULL;
1409
1410         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1411                 if (!trace->live)
1412                         return NULL;
1413                 ++trace->stats.proc_getname;
1414                 if (thread__read_fd_path(thread, fd))
1415                         return NULL;
1416         }
1417
1418         return ttrace->files.table[fd].pathname;
1419 }
1420
1421 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1422 {
1423         int fd = arg->val;
1424         size_t printed = scnprintf(bf, size, "%d", fd);
1425         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1426
1427         if (path)
1428                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1429
1430         return printed;
1431 }
1432
1433 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1434 {
1435         size_t printed = scnprintf(bf, size, "%d", fd);
1436         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1437
1438         if (thread) {
1439                 const char *path = thread__fd_path(thread, fd, trace);
1440
1441                 if (path)
1442                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1443
1444                 thread__put(thread);
1445         }
1446
1447         return printed;
1448 }
1449
1450 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1451                                               struct syscall_arg *arg)
1452 {
1453         int fd = arg->val;
1454         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1455         struct thread_trace *ttrace = thread__priv(arg->thread);
1456
1457         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1458                 zfree(&ttrace->files.table[fd].pathname);
1459
1460         return printed;
1461 }
1462
1463 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1464                                      unsigned long ptr)
1465 {
1466         struct thread_trace *ttrace = thread__priv(thread);
1467
1468         ttrace->filename.ptr = ptr;
1469         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1470 }
1471
1472 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1473 {
1474         struct augmented_arg *augmented_arg = arg->augmented.args;
1475         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1476         /*
1477          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1478          * we would have two strings, each prefixed by its size.
1479          */
1480         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1481
1482         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1483         arg->augmented.size -= consumed;
1484
1485         return printed;
1486 }
1487
1488 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1489                                               struct syscall_arg *arg)
1490 {
1491         unsigned long ptr = arg->val;
1492
1493         if (arg->augmented.args)
1494                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1495
1496         if (!arg->trace->vfs_getname)
1497                 return scnprintf(bf, size, "%#x", ptr);
1498
1499         thread__set_filename_pos(arg->thread, bf, ptr);
1500         return 0;
1501 }
1502
1503 static bool trace__filter_duration(struct trace *trace, double t)
1504 {
1505         return t < (trace->duration_filter * NSEC_PER_MSEC);
1506 }
1507
1508 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1509 {
1510         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1511
1512         return fprintf(fp, "%10.3f ", ts);
1513 }
1514
1515 /*
1516  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1517  * using ttrace->entry_time for a thread that receives a sys_exit without
1518  * first having received a sys_enter ("poll" issued before tracing session
1519  * starts, lost sys_enter exit due to ring buffer overflow).
1520  */
1521 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1522 {
1523         if (tstamp > 0)
1524                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1525
1526         return fprintf(fp, "         ? ");
1527 }
1528
1529 static bool done = false;
1530 static bool interrupted = false;
1531
1532 static void sig_handler(int sig)
1533 {
1534         done = true;
1535         interrupted = sig == SIGINT;
1536 }
1537
1538 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1539 {
1540         size_t printed = 0;
1541
1542         if (trace->multiple_threads) {
1543                 if (trace->show_comm)
1544                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1545                 printed += fprintf(fp, "%d ", thread->tid);
1546         }
1547
1548         return printed;
1549 }
1550
1551 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1552                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1553 {
1554         size_t printed = 0;
1555
1556         if (trace->show_tstamp)
1557                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1558         if (trace->show_duration)
1559                 printed += fprintf_duration(duration, duration_calculated, fp);
1560         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1561 }
1562
1563 static int trace__process_event(struct trace *trace, struct machine *machine,
1564                                 union perf_event *event, struct perf_sample *sample)
1565 {
1566         int ret = 0;
1567
1568         switch (event->header.type) {
1569         case PERF_RECORD_LOST:
1570                 color_fprintf(trace->output, PERF_COLOR_RED,
1571                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1572                 ret = machine__process_lost_event(machine, event, sample);
1573                 break;
1574         default:
1575                 ret = machine__process_event(machine, event, sample);
1576                 break;
1577         }
1578
1579         return ret;
1580 }
1581
1582 static int trace__tool_process(struct perf_tool *tool,
1583                                union perf_event *event,
1584                                struct perf_sample *sample,
1585                                struct machine *machine)
1586 {
1587         struct trace *trace = container_of(tool, struct trace, tool);
1588         return trace__process_event(trace, machine, event, sample);
1589 }
1590
1591 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1592 {
1593         struct machine *machine = vmachine;
1594
1595         if (machine->kptr_restrict_warned)
1596                 return NULL;
1597
1598         if (symbol_conf.kptr_restrict) {
1599                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1600                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1601                            "Kernel samples will not be resolved.\n");
1602                 machine->kptr_restrict_warned = true;
1603                 return NULL;
1604         }
1605
1606         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1607 }
1608
1609 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1610 {
1611         int err = symbol__init(NULL);
1612
1613         if (err)
1614                 return err;
1615
1616         trace->host = machine__new_host();
1617         if (trace->host == NULL)
1618                 return -ENOMEM;
1619
1620         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1621         if (err < 0)
1622                 goto out;
1623
1624         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1625                                             evlist->core.threads, trace__tool_process, false,
1626                                             1);
1627 out:
1628         if (err)
1629                 symbol__exit();
1630
1631         return err;
1632 }
1633
1634 static void trace__symbols__exit(struct trace *trace)
1635 {
1636         machine__exit(trace->host);
1637         trace->host = NULL;
1638
1639         symbol__exit();
1640 }
1641
1642 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1643 {
1644         int idx;
1645
1646         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1647                 nr_args = sc->fmt->nr_args;
1648
1649         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1650         if (sc->arg_fmt == NULL)
1651                 return -1;
1652
1653         for (idx = 0; idx < nr_args; ++idx) {
1654                 if (sc->fmt)
1655                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1656         }
1657
1658         sc->nr_args = nr_args;
1659         return 0;
1660 }
1661
1662 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1663         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1664         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1665 };
1666
1667 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1668 {
1669        const struct syscall_arg_fmt *fmt = fmtp;
1670        return strcmp(name, fmt->name);
1671 }
1672
1673 static struct syscall_arg_fmt *
1674 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1675 {
1676        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1677 }
1678
1679 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1680 {
1681        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1682        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1683 }
1684
1685 static struct tep_format_field *
1686 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1687 {
1688         struct tep_format_field *last_field = NULL;
1689         int len;
1690
1691         for (; field; field = field->next, ++arg) {
1692                 last_field = field;
1693
1694                 if (arg->scnprintf)
1695                         continue;
1696
1697                 len = strlen(field->name);
1698
1699                 if (strcmp(field->type, "const char *") == 0 &&
1700                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1701                      strstr(field->name, "path") != NULL))
1702                         arg->scnprintf = SCA_FILENAME;
1703                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1704                         arg->scnprintf = SCA_PTR;
1705                 else if (strcmp(field->type, "pid_t") == 0)
1706                         arg->scnprintf = SCA_PID;
1707                 else if (strcmp(field->type, "umode_t") == 0)
1708                         arg->scnprintf = SCA_MODE_T;
1709                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1710                         arg->scnprintf = SCA_CHAR_ARRAY;
1711                         arg->nr_entries = field->arraylen;
1712                 } else if ((strcmp(field->type, "int") == 0 ||
1713                           strcmp(field->type, "unsigned int") == 0 ||
1714                           strcmp(field->type, "long") == 0) &&
1715                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1716                         /*
1717                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1718                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1719                          * 65 int
1720                          * 23 unsigned int
1721                          * 7 unsigned long
1722                          */
1723                         arg->scnprintf = SCA_FD;
1724                } else {
1725                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1726
1727                         if (fmt) {
1728                                 arg->scnprintf = fmt->scnprintf;
1729                                 arg->strtoul   = fmt->strtoul;
1730                         }
1731                 }
1732         }
1733
1734         return last_field;
1735 }
1736
1737 static int syscall__set_arg_fmts(struct syscall *sc)
1738 {
1739         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1740
1741         if (last_field)
1742                 sc->args_size = last_field->offset + last_field->size;
1743
1744         return 0;
1745 }
1746
1747 static int trace__read_syscall_info(struct trace *trace, int id)
1748 {
1749         char tp_name[128];
1750         struct syscall *sc;
1751         const char *name = syscalltbl__name(trace->sctbl, id);
1752
1753         if (trace->syscalls.table == NULL) {
1754                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1755                 if (trace->syscalls.table == NULL)
1756                         return -ENOMEM;
1757         }
1758
1759         sc = trace->syscalls.table + id;
1760         if (sc->nonexistent)
1761                 return 0;
1762
1763         if (name == NULL) {
1764                 sc->nonexistent = true;
1765                 return 0;
1766         }
1767
1768         sc->name = name;
1769         sc->fmt  = syscall_fmt__find(sc->name);
1770
1771         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1772         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1773
1774         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1775                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1776                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1777         }
1778
1779         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1780                 return -ENOMEM;
1781
1782         if (IS_ERR(sc->tp_format))
1783                 return PTR_ERR(sc->tp_format);
1784
1785         sc->args = sc->tp_format->format.fields;
1786         /*
1787          * We need to check and discard the first variable '__syscall_nr'
1788          * or 'nr' that mean the syscall number. It is needless here.
1789          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1790          */
1791         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1792                 sc->args = sc->args->next;
1793                 --sc->nr_args;
1794         }
1795
1796         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1797         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1798
1799         return syscall__set_arg_fmts(sc);
1800 }
1801
1802 static int perf_evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1803 {
1804         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1805
1806         if (fmt != NULL) {
1807                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1808                 return 0;
1809         }
1810
1811         return -ENOMEM;
1812 }
1813
1814 static int intcmp(const void *a, const void *b)
1815 {
1816         const int *one = a, *another = b;
1817
1818         return *one - *another;
1819 }
1820
1821 static int trace__validate_ev_qualifier(struct trace *trace)
1822 {
1823         int err = 0;
1824         bool printed_invalid_prefix = false;
1825         struct str_node *pos;
1826         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1827
1828         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1829                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1830
1831         if (trace->ev_qualifier_ids.entries == NULL) {
1832                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1833                        trace->output);
1834                 err = -EINVAL;
1835                 goto out;
1836         }
1837
1838         strlist__for_each_entry(pos, trace->ev_qualifier) {
1839                 const char *sc = pos->s;
1840                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1841
1842                 if (id < 0) {
1843                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1844                         if (id >= 0)
1845                                 goto matches;
1846
1847                         if (!printed_invalid_prefix) {
1848                                 pr_debug("Skipping unknown syscalls: ");
1849                                 printed_invalid_prefix = true;
1850                         } else {
1851                                 pr_debug(", ");
1852                         }
1853
1854                         pr_debug("%s", sc);
1855                         continue;
1856                 }
1857 matches:
1858                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1859                 if (match_next == -1)
1860                         continue;
1861
1862                 while (1) {
1863                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1864                         if (id < 0)
1865                                 break;
1866                         if (nr_allocated == nr_used) {
1867                                 void *entries;
1868
1869                                 nr_allocated += 8;
1870                                 entries = realloc(trace->ev_qualifier_ids.entries,
1871                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1872                                 if (entries == NULL) {
1873                                         err = -ENOMEM;
1874                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1875                                         goto out_free;
1876                                 }
1877                                 trace->ev_qualifier_ids.entries = entries;
1878                         }
1879                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1880                 }
1881         }
1882
1883         trace->ev_qualifier_ids.nr = nr_used;
1884         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1885 out:
1886         if (printed_invalid_prefix)
1887                 pr_debug("\n");
1888         return err;
1889 out_free:
1890         zfree(&trace->ev_qualifier_ids.entries);
1891         trace->ev_qualifier_ids.nr = 0;
1892         goto out;
1893 }
1894
1895 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1896 {
1897         bool in_ev_qualifier;
1898
1899         if (trace->ev_qualifier_ids.nr == 0)
1900                 return true;
1901
1902         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1903                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1904
1905         if (in_ev_qualifier)
1906                return !trace->not_ev_qualifier;
1907
1908         return trace->not_ev_qualifier;
1909 }
1910
1911 /*
1912  * args is to be interpreted as a series of longs but we need to handle
1913  * 8-byte unaligned accesses. args points to raw_data within the event
1914  * and raw_data is guaranteed to be 8-byte unaligned because it is
1915  * preceded by raw_size which is a u32. So we need to copy args to a temp
1916  * variable to read it. Most notably this avoids extended load instructions
1917  * on unaligned addresses
1918  */
1919 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1920 {
1921         unsigned long val;
1922         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1923
1924         memcpy(&val, p, sizeof(val));
1925         return val;
1926 }
1927
1928 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1929                                       struct syscall_arg *arg)
1930 {
1931         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1932                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1933
1934         return scnprintf(bf, size, "arg%d: ", arg->idx);
1935 }
1936
1937 /*
1938  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1939  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1940  * in tools/perf/trace/beauty/mount_flags.c
1941  */
1942 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1943 {
1944         if (fmt && fmt->mask_val)
1945                 return fmt->mask_val(arg, val);
1946
1947         return val;
1948 }
1949
1950 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1951                                              struct syscall_arg *arg, unsigned long val)
1952 {
1953         if (fmt && fmt->scnprintf) {
1954                 arg->val = val;
1955                 if (fmt->parm)
1956                         arg->parm = fmt->parm;
1957                 return fmt->scnprintf(bf, size, arg);
1958         }
1959         return scnprintf(bf, size, "%ld", val);
1960 }
1961
1962 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1963                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1964                                       struct trace *trace, struct thread *thread)
1965 {
1966         size_t printed = 0;
1967         unsigned long val;
1968         u8 bit = 1;
1969         struct syscall_arg arg = {
1970                 .args   = args,
1971                 .augmented = {
1972                         .size = augmented_args_size,
1973                         .args = augmented_args,
1974                 },
1975                 .idx    = 0,
1976                 .mask   = 0,
1977                 .trace  = trace,
1978                 .thread = thread,
1979                 .show_string_prefix = trace->show_string_prefix,
1980         };
1981         struct thread_trace *ttrace = thread__priv(thread);
1982
1983         /*
1984          * Things like fcntl will set this in its 'cmd' formatter to pick the
1985          * right formatter for the return value (an fd? file flags?), which is
1986          * not needed for syscalls that always return a given type, say an fd.
1987          */
1988         ttrace->ret_scnprintf = NULL;
1989
1990         if (sc->args != NULL) {
1991                 struct tep_format_field *field;
1992
1993                 for (field = sc->args; field;
1994                      field = field->next, ++arg.idx, bit <<= 1) {
1995                         if (arg.mask & bit)
1996                                 continue;
1997
1998                         arg.fmt = &sc->arg_fmt[arg.idx];
1999                         val = syscall_arg__val(&arg, arg.idx);
2000                         /*
2001                          * Some syscall args need some mask, most don't and
2002                          * return val untouched.
2003                          */
2004                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2005
2006                         /*
2007                          * Suppress this argument if its value is zero and
2008                          * and we don't have a string associated in an
2009                          * strarray for it.
2010                          */
2011                         if (val == 0 &&
2012                             !trace->show_zeros &&
2013                             !(sc->arg_fmt &&
2014                               (sc->arg_fmt[arg.idx].show_zero ||
2015                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2016                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2017                               sc->arg_fmt[arg.idx].parm))
2018                                 continue;
2019
2020                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2021
2022                         if (trace->show_arg_names)
2023                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2024
2025                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2026                                                                   bf + printed, size - printed, &arg, val);
2027                 }
2028         } else if (IS_ERR(sc->tp_format)) {
2029                 /*
2030                  * If we managed to read the tracepoint /format file, then we
2031                  * may end up not having any args, like with gettid(), so only
2032                  * print the raw args when we didn't manage to read it.
2033                  */
2034                 while (arg.idx < sc->nr_args) {
2035                         if (arg.mask & bit)
2036                                 goto next_arg;
2037                         val = syscall_arg__val(&arg, arg.idx);
2038                         if (printed)
2039                                 printed += scnprintf(bf + printed, size - printed, ", ");
2040                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2041                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2042 next_arg:
2043                         ++arg.idx;
2044                         bit <<= 1;
2045                 }
2046         }
2047
2048         return printed;
2049 }
2050
2051 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2052                                   union perf_event *event,
2053                                   struct perf_sample *sample);
2054
2055 static struct syscall *trace__syscall_info(struct trace *trace,
2056                                            struct evsel *evsel, int id)
2057 {
2058         int err = 0;
2059
2060         if (id < 0) {
2061
2062                 /*
2063                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2064                  * before that, leaving at a higher verbosity level till that is
2065                  * explained. Reproduced with plain ftrace with:
2066                  *
2067                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2068                  * grep "NR -1 " /t/trace_pipe
2069                  *
2070                  * After generating some load on the machine.
2071                  */
2072                 if (verbose > 1) {
2073                         static u64 n;
2074                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2075                                 id, perf_evsel__name(evsel), ++n);
2076                 }
2077                 return NULL;
2078         }
2079
2080         err = -EINVAL;
2081
2082         if (id > trace->sctbl->syscalls.max_id)
2083                 goto out_cant_read;
2084
2085         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2086             (err = trace__read_syscall_info(trace, id)) != 0)
2087                 goto out_cant_read;
2088
2089         if (trace->syscalls.table[id].name == NULL) {
2090                 if (trace->syscalls.table[id].nonexistent)
2091                         return NULL;
2092                 goto out_cant_read;
2093         }
2094
2095         return &trace->syscalls.table[id];
2096
2097 out_cant_read:
2098         if (verbose > 0) {
2099                 char sbuf[STRERR_BUFSIZE];
2100                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2101                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2102                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2103                 fputs(" information\n", trace->output);
2104         }
2105         return NULL;
2106 }
2107
2108 struct syscall_stats {
2109         struct stats stats;
2110         u64          nr_failures;
2111         int          max_errno;
2112         u32          *errnos;
2113 };
2114
2115 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2116                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2117 {
2118         struct int_node *inode;
2119         struct syscall_stats *stats;
2120         u64 duration = 0;
2121
2122         inode = intlist__findnew(ttrace->syscall_stats, id);
2123         if (inode == NULL)
2124                 return;
2125
2126         stats = inode->priv;
2127         if (stats == NULL) {
2128                 stats = malloc(sizeof(*stats));
2129                 if (stats == NULL)
2130                         return;
2131
2132                 stats->nr_failures = 0;
2133                 stats->max_errno   = 0;
2134                 stats->errnos      = NULL;
2135                 init_stats(&stats->stats);
2136                 inode->priv = stats;
2137         }
2138
2139         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2140                 duration = sample->time - ttrace->entry_time;
2141
2142         update_stats(&stats->stats, duration);
2143
2144         if (err < 0) {
2145                 ++stats->nr_failures;
2146
2147                 if (!errno_summary)
2148                         return;
2149
2150                 err = -err;
2151                 if (err > stats->max_errno) {
2152                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2153
2154                         if (new_errnos) {
2155                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2156                         } else {
2157                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2158                                          thread__comm_str(thread), thread->pid_, thread->tid);
2159                                 return;
2160                         }
2161
2162                         stats->errnos = new_errnos;
2163                         stats->max_errno = err;
2164                 }
2165
2166                 ++stats->errnos[err - 1];
2167         }
2168 }
2169
2170 static int trace__printf_interrupted_entry(struct trace *trace)
2171 {
2172         struct thread_trace *ttrace;
2173         size_t printed;
2174         int len;
2175
2176         if (trace->failure_only || trace->current == NULL)
2177                 return 0;
2178
2179         ttrace = thread__priv(trace->current);
2180
2181         if (!ttrace->entry_pending)
2182                 return 0;
2183
2184         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2185         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2186
2187         if (len < trace->args_alignment - 4)
2188                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2189
2190         printed += fprintf(trace->output, " ...\n");
2191
2192         ttrace->entry_pending = false;
2193         ++trace->nr_events_printed;
2194
2195         return printed;
2196 }
2197
2198 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2199                                  struct perf_sample *sample, struct thread *thread)
2200 {
2201         int printed = 0;
2202
2203         if (trace->print_sample) {
2204                 double ts = (double)sample->time / NSEC_PER_MSEC;
2205
2206                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2207                                    perf_evsel__name(evsel), ts,
2208                                    thread__comm_str(thread),
2209                                    sample->pid, sample->tid, sample->cpu);
2210         }
2211
2212         return printed;
2213 }
2214
2215 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2216 {
2217         void *augmented_args = NULL;
2218         /*
2219          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2220          * and there we get all 6 syscall args plus the tracepoint common fields
2221          * that gets calculated at the start and the syscall_nr (another long).
2222          * So we check if that is the case and if so don't look after the
2223          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2224          * which is fixed.
2225          *
2226          * We'll revisit this later to pass s->args_size to the BPF augmenter
2227          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2228          * copies only what we need for each syscall, like what happens when we
2229          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2230          * traffic to just what is needed for each syscall.
2231          */
2232         int args_size = raw_augmented_args_size ?: sc->args_size;
2233
2234         *augmented_args_size = sample->raw_size - args_size;
2235         if (*augmented_args_size > 0)
2236                 augmented_args = sample->raw_data + args_size;
2237
2238         return augmented_args;
2239 }
2240
2241 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2242                             union perf_event *event __maybe_unused,
2243                             struct perf_sample *sample)
2244 {
2245         char *msg;
2246         void *args;
2247         int printed = 0;
2248         struct thread *thread;
2249         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2250         int augmented_args_size = 0;
2251         void *augmented_args = NULL;
2252         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2253         struct thread_trace *ttrace;
2254
2255         if (sc == NULL)
2256                 return -1;
2257
2258         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2259         ttrace = thread__trace(thread, trace->output);
2260         if (ttrace == NULL)
2261                 goto out_put;
2262
2263         trace__fprintf_sample(trace, evsel, sample, thread);
2264
2265         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2266
2267         if (ttrace->entry_str == NULL) {
2268                 ttrace->entry_str = malloc(trace__entry_str_size);
2269                 if (!ttrace->entry_str)
2270                         goto out_put;
2271         }
2272
2273         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2274                 trace__printf_interrupted_entry(trace);
2275         /*
2276          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2277          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2278          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2279          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2280          * so when handling, say the openat syscall, we end up getting 6 args for the
2281          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2282          * thinking that the extra 2 u64 args are the augmented filename, so just check
2283          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2284          */
2285         if (evsel != trace->syscalls.events.sys_enter)
2286                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2287         ttrace->entry_time = sample->time;
2288         msg = ttrace->entry_str;
2289         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2290
2291         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2292                                            args, augmented_args, augmented_args_size, trace, thread);
2293
2294         if (sc->is_exit) {
2295                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2296                         int alignment = 0;
2297
2298                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2299                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2300                         if (trace->args_alignment > printed)
2301                                 alignment = trace->args_alignment - printed;
2302                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2303                 }
2304         } else {
2305                 ttrace->entry_pending = true;
2306                 /* See trace__vfs_getname & trace__sys_exit */
2307                 ttrace->filename.pending_open = false;
2308         }
2309
2310         if (trace->current != thread) {
2311                 thread__put(trace->current);
2312                 trace->current = thread__get(thread);
2313         }
2314         err = 0;
2315 out_put:
2316         thread__put(thread);
2317         return err;
2318 }
2319
2320 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2321                                     struct perf_sample *sample)
2322 {
2323         struct thread_trace *ttrace;
2324         struct thread *thread;
2325         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2326         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2327         char msg[1024];
2328         void *args, *augmented_args = NULL;
2329         int augmented_args_size;
2330
2331         if (sc == NULL)
2332                 return -1;
2333
2334         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2335         ttrace = thread__trace(thread, trace->output);
2336         /*
2337          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2338          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2339          */
2340         if (ttrace == NULL)
2341                 goto out_put;
2342
2343         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2344         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2345         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2346         fprintf(trace->output, "%s", msg);
2347         err = 0;
2348 out_put:
2349         thread__put(thread);
2350         return err;
2351 }
2352
2353 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2354                                     struct perf_sample *sample,
2355                                     struct callchain_cursor *cursor)
2356 {
2357         struct addr_location al;
2358         int max_stack = evsel->core.attr.sample_max_stack ?
2359                         evsel->core.attr.sample_max_stack :
2360                         trace->max_stack;
2361         int err;
2362
2363         if (machine__resolve(trace->host, &al, sample) < 0)
2364                 return -1;
2365
2366         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2367         addr_location__put(&al);
2368         return err;
2369 }
2370
2371 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2372 {
2373         /* TODO: user-configurable print_opts */
2374         const unsigned int print_opts = EVSEL__PRINT_SYM |
2375                                         EVSEL__PRINT_DSO |
2376                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2377
2378         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2379 }
2380
2381 static const char *errno_to_name(struct evsel *evsel, int err)
2382 {
2383         struct perf_env *env = perf_evsel__env(evsel);
2384         const char *arch_name = perf_env__arch(env);
2385
2386         return arch_syscalls__strerrno(arch_name, err);
2387 }
2388
2389 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2390                            union perf_event *event __maybe_unused,
2391                            struct perf_sample *sample)
2392 {
2393         long ret;
2394         u64 duration = 0;
2395         bool duration_calculated = false;
2396         struct thread *thread;
2397         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2398         int alignment = trace->args_alignment;
2399         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2400         struct thread_trace *ttrace;
2401
2402         if (sc == NULL)
2403                 return -1;
2404
2405         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2406         ttrace = thread__trace(thread, trace->output);
2407         if (ttrace == NULL)
2408                 goto out_put;
2409
2410         trace__fprintf_sample(trace, evsel, sample, thread);
2411
2412         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2413
2414         if (trace->summary)
2415                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2416
2417         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2418                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2419                 ttrace->filename.pending_open = false;
2420                 ++trace->stats.vfs_getname;
2421         }
2422
2423         if (ttrace->entry_time) {
2424                 duration = sample->time - ttrace->entry_time;
2425                 if (trace__filter_duration(trace, duration))
2426                         goto out;
2427                 duration_calculated = true;
2428         } else if (trace->duration_filter)
2429                 goto out;
2430
2431         if (sample->callchain) {
2432                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2433                 if (callchain_ret == 0) {
2434                         if (callchain_cursor.nr < trace->min_stack)
2435                                 goto out;
2436                         callchain_ret = 1;
2437                 }
2438         }
2439
2440         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2441                 goto out;
2442
2443         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2444
2445         if (ttrace->entry_pending) {
2446                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2447         } else {
2448                 printed += fprintf(trace->output, " ... [");
2449                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2450                 printed += 9;
2451                 printed += fprintf(trace->output, "]: %s()", sc->name);
2452         }
2453
2454         printed++; /* the closing ')' */
2455
2456         if (alignment > printed)
2457                 alignment -= printed;
2458         else
2459                 alignment = 0;
2460
2461         fprintf(trace->output, ")%*s= ", alignment, " ");
2462
2463         if (sc->fmt == NULL) {
2464                 if (ret < 0)
2465                         goto errno_print;
2466 signed_print:
2467                 fprintf(trace->output, "%ld", ret);
2468         } else if (ret < 0) {
2469 errno_print: {
2470                 char bf[STRERR_BUFSIZE];
2471                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2472                            *e = errno_to_name(evsel, -ret);
2473
2474                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2475         }
2476         } else if (ret == 0 && sc->fmt->timeout)
2477                 fprintf(trace->output, "0 (Timeout)");
2478         else if (ttrace->ret_scnprintf) {
2479                 char bf[1024];
2480                 struct syscall_arg arg = {
2481                         .val    = ret,
2482                         .thread = thread,
2483                         .trace  = trace,
2484                 };
2485                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2486                 ttrace->ret_scnprintf = NULL;
2487                 fprintf(trace->output, "%s", bf);
2488         } else if (sc->fmt->hexret)
2489                 fprintf(trace->output, "%#lx", ret);
2490         else if (sc->fmt->errpid) {
2491                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2492
2493                 if (child != NULL) {
2494                         fprintf(trace->output, "%ld", ret);
2495                         if (child->comm_set)
2496                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2497                         thread__put(child);
2498                 }
2499         } else
2500                 goto signed_print;
2501
2502         fputc('\n', trace->output);
2503
2504         /*
2505          * We only consider an 'event' for the sake of --max-events a non-filtered
2506          * sys_enter + sys_exit and other tracepoint events.
2507          */
2508         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2509                 interrupted = true;
2510
2511         if (callchain_ret > 0)
2512                 trace__fprintf_callchain(trace, sample);
2513         else if (callchain_ret < 0)
2514                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2515 out:
2516         ttrace->entry_pending = false;
2517         err = 0;
2518 out_put:
2519         thread__put(thread);
2520         return err;
2521 }
2522
2523 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2524                               union perf_event *event __maybe_unused,
2525                               struct perf_sample *sample)
2526 {
2527         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2528         struct thread_trace *ttrace;
2529         size_t filename_len, entry_str_len, to_move;
2530         ssize_t remaining_space;
2531         char *pos;
2532         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2533
2534         if (!thread)
2535                 goto out;
2536
2537         ttrace = thread__priv(thread);
2538         if (!ttrace)
2539                 goto out_put;
2540
2541         filename_len = strlen(filename);
2542         if (filename_len == 0)
2543                 goto out_put;
2544
2545         if (ttrace->filename.namelen < filename_len) {
2546                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2547
2548                 if (f == NULL)
2549                         goto out_put;
2550
2551                 ttrace->filename.namelen = filename_len;
2552                 ttrace->filename.name = f;
2553         }
2554
2555         strcpy(ttrace->filename.name, filename);
2556         ttrace->filename.pending_open = true;
2557
2558         if (!ttrace->filename.ptr)
2559                 goto out_put;
2560
2561         entry_str_len = strlen(ttrace->entry_str);
2562         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2563         if (remaining_space <= 0)
2564                 goto out_put;
2565
2566         if (filename_len > (size_t)remaining_space) {
2567                 filename += filename_len - remaining_space;
2568                 filename_len = remaining_space;
2569         }
2570
2571         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2572         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2573         memmove(pos + filename_len, pos, to_move);
2574         memcpy(pos, filename, filename_len);
2575
2576         ttrace->filename.ptr = 0;
2577         ttrace->filename.entry_str_pos = 0;
2578 out_put:
2579         thread__put(thread);
2580 out:
2581         return 0;
2582 }
2583
2584 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2585                                      union perf_event *event __maybe_unused,
2586                                      struct perf_sample *sample)
2587 {
2588         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2589         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2590         struct thread *thread = machine__findnew_thread(trace->host,
2591                                                         sample->pid,
2592                                                         sample->tid);
2593         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2594
2595         if (ttrace == NULL)
2596                 goto out_dump;
2597
2598         ttrace->runtime_ms += runtime_ms;
2599         trace->runtime_ms += runtime_ms;
2600 out_put:
2601         thread__put(thread);
2602         return 0;
2603
2604 out_dump:
2605         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2606                evsel->name,
2607                perf_evsel__strval(evsel, sample, "comm"),
2608                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2609                runtime,
2610                perf_evsel__intval(evsel, sample, "vruntime"));
2611         goto out_put;
2612 }
2613
2614 static int bpf_output__printer(enum binary_printer_ops op,
2615                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2616 {
2617         unsigned char ch = (unsigned char)val;
2618
2619         switch (op) {
2620         case BINARY_PRINT_CHAR_DATA:
2621                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2622         case BINARY_PRINT_DATA_BEGIN:
2623         case BINARY_PRINT_LINE_BEGIN:
2624         case BINARY_PRINT_ADDR:
2625         case BINARY_PRINT_NUM_DATA:
2626         case BINARY_PRINT_NUM_PAD:
2627         case BINARY_PRINT_SEP:
2628         case BINARY_PRINT_CHAR_PAD:
2629         case BINARY_PRINT_LINE_END:
2630         case BINARY_PRINT_DATA_END:
2631         default:
2632                 break;
2633         }
2634
2635         return 0;
2636 }
2637
2638 static void bpf_output__fprintf(struct trace *trace,
2639                                 struct perf_sample *sample)
2640 {
2641         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2642                         bpf_output__printer, NULL, trace->output);
2643         ++trace->nr_events_printed;
2644 }
2645
2646 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2647                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2648 {
2649         char bf[2048];
2650         size_t size = sizeof(bf);
2651         struct tep_format_field *field = evsel->tp_format->format.fields;
2652         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2653         size_t printed = 0;
2654         unsigned long val;
2655         u8 bit = 1;
2656         struct syscall_arg syscall_arg = {
2657                 .augmented = {
2658                         .size = augmented_args_size,
2659                         .args = augmented_args,
2660                 },
2661                 .idx    = 0,
2662                 .mask   = 0,
2663                 .trace  = trace,
2664                 .thread = thread,
2665                 .show_string_prefix = trace->show_string_prefix,
2666         };
2667
2668         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2669                 if (syscall_arg.mask & bit)
2670                         continue;
2671
2672                 syscall_arg.len = 0;
2673                 syscall_arg.fmt = arg;
2674                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2675                         int offset = field->offset;
2676
2677                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2678                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2679                                 syscall_arg.len = offset >> 16;
2680                                 offset &= 0xffff;
2681                         }
2682
2683                         val = (uintptr_t)(sample->raw_data + offset);
2684                 } else
2685                         val = format_field__intval(field, sample, evsel->needs_swap);
2686                 /*
2687                  * Some syscall args need some mask, most don't and
2688                  * return val untouched.
2689                  */
2690                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2691
2692                 /*
2693                  * Suppress this argument if its value is zero and
2694                  * and we don't have a string associated in an
2695                  * strarray for it.
2696                  */
2697                 if (val == 0 &&
2698                     !trace->show_zeros &&
2699                     !((arg->show_zero ||
2700                        arg->scnprintf == SCA_STRARRAY ||
2701                        arg->scnprintf == SCA_STRARRAYS) &&
2702                       arg->parm))
2703                         continue;
2704
2705                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2706
2707                 /*
2708                  * XXX Perhaps we should have a show_tp_arg_names,
2709                  * leaving show_arg_names just for syscalls?
2710                  */
2711                 if (1 || trace->show_arg_names)
2712                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2713
2714                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2715         }
2716
2717         return printed + fprintf(trace->output, "%s", bf);
2718 }
2719
2720 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2721                                 union perf_event *event __maybe_unused,
2722                                 struct perf_sample *sample)
2723 {
2724         struct thread *thread;
2725         int callchain_ret = 0;
2726         /*
2727          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2728          * this event's max_events having been hit and this is an entry coming
2729          * from the ring buffer that we should discard, since the max events
2730          * have already been considered/printed.
2731          */
2732         if (evsel->disabled)
2733                 return 0;
2734
2735         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2736
2737         if (sample->callchain) {
2738                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2739                 if (callchain_ret == 0) {
2740                         if (callchain_cursor.nr < trace->min_stack)
2741                                 goto out;
2742                         callchain_ret = 1;
2743                 }
2744         }
2745
2746         trace__printf_interrupted_entry(trace);
2747         trace__fprintf_tstamp(trace, sample->time, trace->output);
2748
2749         if (trace->trace_syscalls && trace->show_duration)
2750                 fprintf(trace->output, "(         ): ");
2751
2752         if (thread)
2753                 trace__fprintf_comm_tid(trace, thread, trace->output);
2754
2755         if (evsel == trace->syscalls.events.augmented) {
2756                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2757                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2758
2759                 if (sc) {
2760                         fprintf(trace->output, "%s(", sc->name);
2761                         trace__fprintf_sys_enter(trace, evsel, sample);
2762                         fputc(')', trace->output);
2763                         goto newline;
2764                 }
2765
2766                 /*
2767                  * XXX: Not having the associated syscall info or not finding/adding
2768                  *      the thread should never happen, but if it does...
2769                  *      fall thru and print it as a bpf_output event.
2770                  */
2771         }
2772
2773         fprintf(trace->output, "%s(", evsel->name);
2774
2775         if (perf_evsel__is_bpf_output(evsel)) {
2776                 bpf_output__fprintf(trace, sample);
2777         } else if (evsel->tp_format) {
2778                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2779                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2780                         if (trace->libtraceevent_print) {
2781                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2782                                                       sample->raw_data, sample->raw_size,
2783                                                       trace->output);
2784                         } else {
2785                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2786                         }
2787                 }
2788         }
2789
2790 newline:
2791         fprintf(trace->output, ")\n");
2792
2793         if (callchain_ret > 0)
2794                 trace__fprintf_callchain(trace, sample);
2795         else if (callchain_ret < 0)
2796                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2797
2798         ++trace->nr_events_printed;
2799
2800         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2801                 evsel__disable(evsel);
2802                 evsel__close(evsel);
2803         }
2804 out:
2805         thread__put(thread);
2806         return 0;
2807 }
2808
2809 static void print_location(FILE *f, struct perf_sample *sample,
2810                            struct addr_location *al,
2811                            bool print_dso, bool print_sym)
2812 {
2813
2814         if ((verbose > 0 || print_dso) && al->map)
2815                 fprintf(f, "%s@", al->map->dso->long_name);
2816
2817         if ((verbose > 0 || print_sym) && al->sym)
2818                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2819                         al->addr - al->sym->start);
2820         else if (al->map)
2821                 fprintf(f, "0x%" PRIx64, al->addr);
2822         else
2823                 fprintf(f, "0x%" PRIx64, sample->addr);
2824 }
2825
2826 static int trace__pgfault(struct trace *trace,
2827                           struct evsel *evsel,
2828                           union perf_event *event __maybe_unused,
2829                           struct perf_sample *sample)
2830 {
2831         struct thread *thread;
2832         struct addr_location al;
2833         char map_type = 'd';
2834         struct thread_trace *ttrace;
2835         int err = -1;
2836         int callchain_ret = 0;
2837
2838         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2839
2840         if (sample->callchain) {
2841                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2842                 if (callchain_ret == 0) {
2843                         if (callchain_cursor.nr < trace->min_stack)
2844                                 goto out_put;
2845                         callchain_ret = 1;
2846                 }
2847         }
2848
2849         ttrace = thread__trace(thread, trace->output);
2850         if (ttrace == NULL)
2851                 goto out_put;
2852
2853         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2854                 ttrace->pfmaj++;
2855         else
2856                 ttrace->pfmin++;
2857
2858         if (trace->summary_only)
2859                 goto out;
2860
2861         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2862
2863         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2864
2865         fprintf(trace->output, "%sfault [",
2866                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2867                 "maj" : "min");
2868
2869         print_location(trace->output, sample, &al, false, true);
2870
2871         fprintf(trace->output, "] => ");
2872
2873         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2874
2875         if (!al.map) {
2876                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2877
2878                 if (al.map)
2879                         map_type = 'x';
2880                 else
2881                         map_type = '?';
2882         }
2883
2884         print_location(trace->output, sample, &al, true, false);
2885
2886         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2887
2888         if (callchain_ret > 0)
2889                 trace__fprintf_callchain(trace, sample);
2890         else if (callchain_ret < 0)
2891                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2892
2893         ++trace->nr_events_printed;
2894 out:
2895         err = 0;
2896 out_put:
2897         thread__put(thread);
2898         return err;
2899 }
2900
2901 static void trace__set_base_time(struct trace *trace,
2902                                  struct evsel *evsel,
2903                                  struct perf_sample *sample)
2904 {
2905         /*
2906          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2907          * and don't use sample->time unconditionally, we may end up having
2908          * some other event in the future without PERF_SAMPLE_TIME for good
2909          * reason, i.e. we may not be interested in its timestamps, just in
2910          * it taking place, picking some piece of information when it
2911          * appears in our event stream (vfs_getname comes to mind).
2912          */
2913         if (trace->base_time == 0 && !trace->full_time &&
2914             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2915                 trace->base_time = sample->time;
2916 }
2917
2918 static int trace__process_sample(struct perf_tool *tool,
2919                                  union perf_event *event,
2920                                  struct perf_sample *sample,
2921                                  struct evsel *evsel,
2922                                  struct machine *machine __maybe_unused)
2923 {
2924         struct trace *trace = container_of(tool, struct trace, tool);
2925         struct thread *thread;
2926         int err = 0;
2927
2928         tracepoint_handler handler = evsel->handler;
2929
2930         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2931         if (thread && thread__is_filtered(thread))
2932                 goto out;
2933
2934         trace__set_base_time(trace, evsel, sample);
2935
2936         if (handler) {
2937                 ++trace->nr_events;
2938                 handler(trace, evsel, event, sample);
2939         }
2940 out:
2941         thread__put(thread);
2942         return err;
2943 }
2944
2945 static int trace__record(struct trace *trace, int argc, const char **argv)
2946 {
2947         unsigned int rec_argc, i, j;
2948         const char **rec_argv;
2949         const char * const record_args[] = {
2950                 "record",
2951                 "-R",
2952                 "-m", "1024",
2953                 "-c", "1",
2954         };
2955         pid_t pid = getpid();
2956         char *filter = asprintf__tp_filter_pids(1, &pid);
2957         const char * const sc_args[] = { "-e", };
2958         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2959         const char * const majpf_args[] = { "-e", "major-faults" };
2960         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2961         const char * const minpf_args[] = { "-e", "minor-faults" };
2962         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2963         int err = -1;
2964
2965         /* +3 is for the event string below and the pid filter */
2966         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
2967                 majpf_args_nr + minpf_args_nr + argc;
2968         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2969
2970         if (rec_argv == NULL || filter == NULL)
2971                 goto out_free;
2972
2973         j = 0;
2974         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2975                 rec_argv[j++] = record_args[i];
2976
2977         if (trace->trace_syscalls) {
2978                 for (i = 0; i < sc_args_nr; i++)
2979                         rec_argv[j++] = sc_args[i];
2980
2981                 /* event string may be different for older kernels - e.g., RHEL6 */
2982                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2983                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2984                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2985                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2986                 else {
2987                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2988                         goto out_free;
2989                 }
2990         }
2991
2992         rec_argv[j++] = "--filter";
2993         rec_argv[j++] = filter;
2994
2995         if (trace->trace_pgfaults & TRACE_PFMAJ)
2996                 for (i = 0; i < majpf_args_nr; i++)
2997                         rec_argv[j++] = majpf_args[i];
2998
2999         if (trace->trace_pgfaults & TRACE_PFMIN)
3000                 for (i = 0; i < minpf_args_nr; i++)
3001                         rec_argv[j++] = minpf_args[i];
3002
3003         for (i = 0; i < (unsigned int)argc; i++)
3004                 rec_argv[j++] = argv[i];
3005
3006         err = cmd_record(j, rec_argv);
3007 out_free:
3008         free(filter);
3009         free(rec_argv);
3010         return err;
3011 }
3012
3013 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3014
3015 static bool evlist__add_vfs_getname(struct evlist *evlist)
3016 {
3017         bool found = false;
3018         struct evsel *evsel, *tmp;
3019         struct parse_events_error err = { .idx = 0, };
3020         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
3021
3022         if (ret)
3023                 return false;
3024
3025         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3026                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
3027                         continue;
3028
3029                 if (perf_evsel__field(evsel, "pathname")) {
3030                         evsel->handler = trace__vfs_getname;
3031                         found = true;
3032                         continue;
3033                 }
3034
3035                 list_del_init(&evsel->core.node);
3036                 evsel->evlist = NULL;
3037                 evsel__delete(evsel);
3038         }
3039
3040         return found;
3041 }
3042
3043 static struct evsel *perf_evsel__new_pgfault(u64 config)
3044 {
3045         struct evsel *evsel;
3046         struct perf_event_attr attr = {
3047                 .type = PERF_TYPE_SOFTWARE,
3048                 .mmap_data = 1,
3049         };
3050
3051         attr.config = config;
3052         attr.sample_period = 1;
3053
3054         event_attr_init(&attr);
3055
3056         evsel = evsel__new(&attr);
3057         if (evsel)
3058                 evsel->handler = trace__pgfault;
3059
3060         return evsel;
3061 }
3062
3063 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3064 {
3065         const u32 type = event->header.type;
3066         struct evsel *evsel;
3067
3068         if (type != PERF_RECORD_SAMPLE) {
3069                 trace__process_event(trace, trace->host, event, sample);
3070                 return;
3071         }
3072
3073         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
3074         if (evsel == NULL) {
3075                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3076                 return;
3077         }
3078
3079         if (evswitch__discard(&trace->evswitch, evsel))
3080                 return;
3081
3082         trace__set_base_time(trace, evsel, sample);
3083
3084         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3085             sample->raw_data == NULL) {
3086                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3087                        perf_evsel__name(evsel), sample->tid,
3088                        sample->cpu, sample->raw_size);
3089         } else {
3090                 tracepoint_handler handler = evsel->handler;
3091                 handler(trace, evsel, event, sample);
3092         }
3093
3094         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3095                 interrupted = true;
3096 }
3097
3098 static int trace__add_syscall_newtp(struct trace *trace)
3099 {
3100         int ret = -1;
3101         struct evlist *evlist = trace->evlist;
3102         struct evsel *sys_enter, *sys_exit;
3103
3104         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3105         if (sys_enter == NULL)
3106                 goto out;
3107
3108         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3109                 goto out_delete_sys_enter;
3110
3111         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3112         if (sys_exit == NULL)
3113                 goto out_delete_sys_enter;
3114
3115         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3116                 goto out_delete_sys_exit;
3117
3118         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3119         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3120
3121         evlist__add(evlist, sys_enter);
3122         evlist__add(evlist, sys_exit);
3123
3124         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3125                 /*
3126                  * We're interested only in the user space callchain
3127                  * leading to the syscall, allow overriding that for
3128                  * debugging reasons using --kernel_syscall_callchains
3129                  */
3130                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3131         }
3132
3133         trace->syscalls.events.sys_enter = sys_enter;
3134         trace->syscalls.events.sys_exit  = sys_exit;
3135
3136         ret = 0;
3137 out:
3138         return ret;
3139
3140 out_delete_sys_exit:
3141         evsel__delete_priv(sys_exit);
3142 out_delete_sys_enter:
3143         evsel__delete_priv(sys_enter);
3144         goto out;
3145 }
3146
3147 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3148 {
3149         int err = -1;
3150         struct evsel *sys_exit;
3151         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3152                                                 trace->ev_qualifier_ids.nr,
3153                                                 trace->ev_qualifier_ids.entries);
3154
3155         if (filter == NULL)
3156                 goto out_enomem;
3157
3158         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
3159                                           filter)) {
3160                 sys_exit = trace->syscalls.events.sys_exit;
3161                 err = perf_evsel__append_tp_filter(sys_exit, filter);
3162         }
3163
3164         free(filter);
3165 out:
3166         return err;
3167 out_enomem:
3168         errno = ENOMEM;
3169         goto out;
3170 }
3171
3172 #ifdef HAVE_LIBBPF_SUPPORT
3173 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3174 {
3175         if (trace->bpf_obj == NULL)
3176                 return NULL;
3177
3178         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3179 }
3180
3181 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3182                                                         const char *prog_name, const char *type)
3183 {
3184         struct bpf_program *prog;
3185
3186         if (prog_name == NULL) {
3187                 char default_prog_name[256];
3188                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3189                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3190                 if (prog != NULL)
3191                         goto out_found;
3192                 if (sc->fmt && sc->fmt->alias) {
3193                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3194                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3195                         if (prog != NULL)
3196                                 goto out_found;
3197                 }
3198                 goto out_unaugmented;
3199         }
3200
3201         prog = trace__find_bpf_program_by_title(trace, prog_name);
3202
3203         if (prog != NULL) {
3204 out_found:
3205                 return prog;
3206         }
3207
3208         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3209                  prog_name, type, sc->name);
3210 out_unaugmented:
3211         return trace->syscalls.unaugmented_prog;
3212 }
3213
3214 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3215 {
3216         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3217
3218         if (sc == NULL)
3219                 return;
3220
3221         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3222         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3223 }
3224
3225 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3226 {
3227         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3228         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3229 }
3230
3231 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3232 {
3233         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3234         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3235 }
3236
3237 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3238 {
3239         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3240         int arg = 0;
3241
3242         if (sc == NULL)
3243                 goto out;
3244
3245         for (; arg < sc->nr_args; ++arg) {
3246                 entry->string_args_len[arg] = 0;
3247                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3248                         /* Should be set like strace -s strsize */
3249                         entry->string_args_len[arg] = PATH_MAX;
3250                 }
3251         }
3252 out:
3253         for (; arg < 6; ++arg)
3254                 entry->string_args_len[arg] = 0;
3255 }
3256 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3257 {
3258         int fd = bpf_map__fd(trace->syscalls.map);
3259         struct bpf_map_syscall_entry value = {
3260                 .enabled = !trace->not_ev_qualifier,
3261         };
3262         int err = 0;
3263         size_t i;
3264
3265         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3266                 int key = trace->ev_qualifier_ids.entries[i];
3267
3268                 if (value.enabled) {
3269                         trace__init_bpf_map_syscall_args(trace, key, &value);
3270                         trace__init_syscall_bpf_progs(trace, key);
3271                 }
3272
3273                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3274                 if (err)
3275                         break;
3276         }
3277
3278         return err;
3279 }
3280
3281 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3282 {
3283         int fd = bpf_map__fd(trace->syscalls.map);
3284         struct bpf_map_syscall_entry value = {
3285                 .enabled = enabled,
3286         };
3287         int err = 0, key;
3288
3289         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3290                 if (enabled)
3291                         trace__init_bpf_map_syscall_args(trace, key, &value);
3292
3293                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3294                 if (err)
3295                         break;
3296         }
3297
3298         return err;
3299 }
3300
3301 static int trace__init_syscalls_bpf_map(struct trace *trace)
3302 {
3303         bool enabled = true;
3304
3305         if (trace->ev_qualifier_ids.nr)
3306                 enabled = trace->not_ev_qualifier;
3307
3308         return __trace__init_syscalls_bpf_map(trace, enabled);
3309 }
3310
3311 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3312 {
3313         struct tep_format_field *field, *candidate_field;
3314         int id;
3315
3316         /*
3317          * We're only interested in syscalls that have a pointer:
3318          */
3319         for (field = sc->args; field; field = field->next) {
3320                 if (field->flags & TEP_FIELD_IS_POINTER)
3321                         goto try_to_find_pair;
3322         }
3323
3324         return NULL;
3325
3326 try_to_find_pair:
3327         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3328                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3329                 struct bpf_program *pair_prog;
3330                 bool is_candidate = false;
3331
3332                 if (pair == NULL || pair == sc ||
3333                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3334                         continue;
3335
3336                 for (field = sc->args, candidate_field = pair->args;
3337                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3338                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3339                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3340
3341                         if (is_pointer) {
3342                                if (!candidate_is_pointer) {
3343                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3344                                         continue;
3345                                }
3346                         } else {
3347                                 if (candidate_is_pointer) {
3348                                         // The candidate might copy a pointer we don't have, skip it.
3349                                         goto next_candidate;
3350                                 }
3351                                 continue;
3352                         }
3353
3354                         if (strcmp(field->type, candidate_field->type))
3355                                 goto next_candidate;
3356
3357                         is_candidate = true;
3358                 }
3359
3360                 if (!is_candidate)
3361                         goto next_candidate;
3362
3363                 /*
3364                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3365                  * then it may be collecting that and we then can't use it, as it would collect
3366                  * more than what is common to the two syscalls.
3367                  */
3368                 if (candidate_field) {
3369                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3370                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3371                                         goto next_candidate;
3372                 }
3373
3374                 pair_prog = pair->bpf_prog.sys_enter;
3375                 /*
3376                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3377                  * have been searched for, so search it here and if it returns the
3378                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3379                  * program for a filtered syscall on a non-filtered one.
3380                  *
3381                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3382                  * useful for "renameat2".
3383                  */
3384                 if (pair_prog == NULL) {
3385                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3386                         if (pair_prog == trace->syscalls.unaugmented_prog)
3387                                 goto next_candidate;
3388                 }
3389
3390                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3391                 return pair_prog;
3392         next_candidate:
3393                 continue;
3394         }
3395
3396         return NULL;
3397 }
3398
3399 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3400 {
3401         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3402             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3403         int err = 0, key;
3404
3405         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3406                 int prog_fd;
3407
3408                 if (!trace__syscall_enabled(trace, key))
3409                         continue;
3410
3411                 trace__init_syscall_bpf_progs(trace, key);
3412
3413                 // It'll get at least the "!raw_syscalls:unaugmented"
3414                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3415                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3416                 if (err)
3417                         break;
3418                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3419                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3420                 if (err)
3421                         break;
3422         }
3423
3424         /*
3425          * Now lets do a second pass looking for enabled syscalls without
3426          * an augmenter that have a signature that is a superset of another
3427          * syscall with an augmenter so that we can auto-reuse it.
3428          *
3429          * I.e. if we have an augmenter for the "open" syscall that has
3430          * this signature:
3431          *
3432          *   int open(const char *pathname, int flags, mode_t mode);
3433          *
3434          * I.e. that will collect just the first string argument, then we
3435          * can reuse it for the 'creat' syscall, that has this signature:
3436          *
3437          *   int creat(const char *pathname, mode_t mode);
3438          *
3439          * and for:
3440          *
3441          *   int stat(const char *pathname, struct stat *statbuf);
3442          *   int lstat(const char *pathname, struct stat *statbuf);
3443          *
3444          * Because the 'open' augmenter will collect the first arg as a string,
3445          * and leave alone all the other args, which already helps with
3446          * beautifying 'stat' and 'lstat''s pathname arg.
3447          *
3448          * Then, in time, when 'stat' gets an augmenter that collects both
3449          * first and second arg (this one on the raw_syscalls:sys_exit prog
3450          * array tail call, then that one will be used.
3451          */
3452         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3453                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3454                 struct bpf_program *pair_prog;
3455                 int prog_fd;
3456
3457                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3458                         continue;
3459
3460                 /*
3461                  * For now we're just reusing the sys_enter prog, and if it
3462                  * already has an augmenter, we don't need to find one.
3463                  */
3464                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3465                         continue;
3466
3467                 /*
3468                  * Look at all the other syscalls for one that has a signature
3469                  * that is close enough that we can share:
3470                  */
3471                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3472                 if (pair_prog == NULL)
3473                         continue;
3474
3475                 sc->bpf_prog.sys_enter = pair_prog;
3476
3477                 /*
3478                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3479                  * with the fd for the program we're reusing:
3480                  */
3481                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3482                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3483                 if (err)
3484                         break;
3485         }
3486
3487
3488         return err;
3489 }
3490
3491 static void trace__delete_augmented_syscalls(struct trace *trace)
3492 {
3493         struct evsel *evsel, *tmp;
3494
3495         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3496         evsel__delete(trace->syscalls.events.augmented);
3497         trace->syscalls.events.augmented = NULL;
3498
3499         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3500                 if (evsel->bpf_obj == trace->bpf_obj) {
3501                         evlist__remove(trace->evlist, evsel);
3502                         evsel__delete(evsel);
3503                 }
3504
3505         }
3506
3507         bpf_object__close(trace->bpf_obj);
3508         trace->bpf_obj = NULL;
3509 }
3510 #else // HAVE_LIBBPF_SUPPORT
3511 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3512 {
3513         return 0;
3514 }
3515
3516 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3517 {
3518         return 0;
3519 }
3520
3521 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3522                                                             const char *name __maybe_unused)
3523 {
3524         return NULL;
3525 }
3526
3527 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3528 {
3529         return 0;
3530 }
3531
3532 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3533 {
3534 }
3535 #endif // HAVE_LIBBPF_SUPPORT
3536
3537 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3538 {
3539         struct evsel *evsel;
3540
3541         evlist__for_each_entry(trace->evlist, evsel) {
3542                 if (evsel == trace->syscalls.events.augmented ||
3543                     evsel->bpf_obj == trace->bpf_obj)
3544                         continue;
3545
3546                 return false;
3547         }
3548
3549         return true;
3550 }
3551
3552 static int trace__set_ev_qualifier_filter(struct trace *trace)
3553 {
3554         if (trace->syscalls.map)
3555                 return trace__set_ev_qualifier_bpf_filter(trace);
3556         if (trace->syscalls.events.sys_enter)
3557                 return trace__set_ev_qualifier_tp_filter(trace);
3558         return 0;
3559 }
3560
3561 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3562                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3563 {
3564         int err = 0;
3565 #ifdef HAVE_LIBBPF_SUPPORT
3566         bool value = true;
3567         int map_fd = bpf_map__fd(map);
3568         size_t i;
3569
3570         for (i = 0; i < npids; ++i) {
3571                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3572                 if (err)
3573                         break;
3574         }
3575 #endif
3576         return err;
3577 }
3578
3579 static int trace__set_filter_loop_pids(struct trace *trace)
3580 {
3581         unsigned int nr = 1, err;
3582         pid_t pids[32] = {
3583                 getpid(),
3584         };
3585         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3586
3587         while (thread && nr < ARRAY_SIZE(pids)) {
3588                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3589
3590                 if (parent == NULL)
3591                         break;
3592
3593                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3594                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3595                         pids[nr++] = parent->tid;
3596                         break;
3597                 }
3598                 thread = parent;
3599         }
3600
3601         err = perf_evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3602         if (!err && trace->filter_pids.map)
3603                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3604
3605         return err;
3606 }
3607
3608 static int trace__set_filter_pids(struct trace *trace)
3609 {
3610         int err = 0;
3611         /*
3612          * Better not use !target__has_task() here because we need to cover the
3613          * case where no threads were specified in the command line, but a
3614          * workload was, and in that case we will fill in the thread_map when
3615          * we fork the workload in perf_evlist__prepare_workload.
3616          */
3617         if (trace->filter_pids.nr > 0) {
3618                 err = perf_evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3619                                                          trace->filter_pids.entries);
3620                 if (!err && trace->filter_pids.map) {
3621                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3622                                                        trace->filter_pids.entries);
3623                 }
3624         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3625                 err = trace__set_filter_loop_pids(trace);
3626         }
3627
3628         return err;
3629 }
3630
3631 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3632 {
3633         struct evlist *evlist = trace->evlist;
3634         struct perf_sample sample;
3635         int err;
3636
3637         err = perf_evlist__parse_sample(evlist, event, &sample);
3638         if (err)
3639                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3640         else
3641                 trace__handle_event(trace, event, &sample);
3642
3643         return 0;
3644 }
3645
3646 static int __trace__flush_events(struct trace *trace)
3647 {
3648         u64 first = ordered_events__first_time(&trace->oe.data);
3649         u64 flush = trace->oe.last - NSEC_PER_SEC;
3650
3651         /* Is there some thing to flush.. */
3652         if (first && first < flush)
3653                 return ordered_events__flush_time(&trace->oe.data, flush);
3654
3655         return 0;
3656 }
3657
3658 static int trace__flush_events(struct trace *trace)
3659 {
3660         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3661 }
3662
3663 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3664 {
3665         int err;
3666
3667         if (!trace->sort_events)
3668                 return __trace__deliver_event(trace, event);
3669
3670         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3671         if (err && err != -1)
3672                 return err;
3673
3674         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3675         if (err)
3676                 return err;
3677
3678         return trace__flush_events(trace);
3679 }
3680
3681 static int ordered_events__deliver_event(struct ordered_events *oe,
3682                                          struct ordered_event *event)
3683 {
3684         struct trace *trace = container_of(oe, struct trace, oe.data);
3685
3686         return __trace__deliver_event(trace, event->event);
3687 }
3688
3689 static struct syscall_arg_fmt *perf_evsel__syscall_arg_fmt(struct evsel *evsel, char *arg)
3690 {
3691         struct tep_format_field *field;
3692         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3693
3694         if (evsel->tp_format == NULL || fmt == NULL)
3695                 return NULL;
3696
3697         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3698                 if (strcmp(field->name, arg) == 0)
3699                         return fmt;
3700
3701         return NULL;
3702 }
3703
3704 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3705 {
3706         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3707
3708         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3709                 char *right = tok + 1, *right_end;
3710
3711                 if (*right == '=')
3712                         ++right;
3713
3714                 while (isspace(*right))
3715                         ++right;
3716
3717                 if (*right == '\0')
3718                         break;
3719
3720                 while (!isalpha(*left))
3721                         if (++left == tok) {
3722                                 /*
3723                                  * Bail out, can't find the name of the argument that is being
3724                                  * used in the filter, let it try to set this filter, will fail later.
3725                                  */
3726                                 return 0;
3727                         }
3728
3729                 right_end = right + 1;
3730                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3731                         ++right_end;
3732
3733                 if (isalpha(*right)) {
3734                         struct syscall_arg_fmt *fmt;
3735                         int left_size = tok - left,
3736                             right_size = right_end - right;
3737                         char arg[128];
3738
3739                         while (isspace(left[left_size - 1]))
3740                                 --left_size;
3741
3742                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3743
3744                         fmt = perf_evsel__syscall_arg_fmt(evsel, arg);
3745                         if (fmt == NULL) {
3746                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3747                                        arg, evsel->name, evsel->filter);
3748                                 return -1;
3749                         }
3750
3751                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3752                                  arg, (int)(right - tok), tok, right_size, right);
3753
3754                         if (fmt->strtoul) {
3755                                 u64 val;
3756                                 struct syscall_arg syscall_arg = {
3757                                         .parm = fmt->parm,
3758                                 };
3759
3760                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3761                                         char *n, expansion[19];
3762                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3763                                         int expansion_offset = right - new_filter;
3764
3765                                         pr_debug("%s", expansion);
3766
3767                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3768                                                 pr_debug(" out of memory!\n");
3769                                                 free(new_filter);
3770                                                 return -1;
3771                                         }
3772                                         if (new_filter != evsel->filter)
3773                                                 free(new_filter);
3774                                         left = n + expansion_offset + expansion_lenght;
3775                                         new_filter = n;
3776                                 } else {
3777                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3778                                                right_size, right, arg, evsel->name, evsel->filter);
3779                                         return -1;
3780                                 }
3781                         } else {
3782                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3783                                        arg, evsel->name, evsel->filter);
3784                                 return -1;
3785                         }
3786
3787                         pr_debug("\n");
3788                 } else {
3789                         left = right_end;
3790                 }
3791         }
3792
3793         if (new_filter != evsel->filter) {
3794                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3795                 perf_evsel__set_filter(evsel, new_filter);
3796                 free(new_filter);
3797         }
3798
3799         return 0;
3800 }
3801
3802 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3803 {
3804         struct evlist *evlist = trace->evlist;
3805         struct evsel *evsel;
3806
3807         evlist__for_each_entry(evlist, evsel) {
3808                 if (evsel->filter == NULL)
3809                         continue;
3810
3811                 if (trace__expand_filter(trace, evsel)) {
3812                         *err_evsel = evsel;
3813                         return -1;
3814                 }
3815         }
3816
3817         return 0;
3818 }
3819
3820 static int trace__run(struct trace *trace, int argc, const char **argv)
3821 {
3822         struct evlist *evlist = trace->evlist;
3823         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3824         int err = -1, i;
3825         unsigned long before;
3826         const bool forks = argc > 0;
3827         bool draining = false;
3828
3829         trace->live = true;
3830
3831         if (!trace->raw_augmented_syscalls) {
3832                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3833                         goto out_error_raw_syscalls;
3834
3835                 if (trace->trace_syscalls)
3836                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3837         }
3838
3839         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3840                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3841                 if (pgfault_maj == NULL)
3842                         goto out_error_mem;
3843                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3844                 evlist__add(evlist, pgfault_maj);
3845         }
3846
3847         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3848                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3849                 if (pgfault_min == NULL)
3850                         goto out_error_mem;
3851                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3852                 evlist__add(evlist, pgfault_min);
3853         }
3854
3855         if (trace->sched &&
3856             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3857                                    trace__sched_stat_runtime))
3858                 goto out_error_sched_stat_runtime;
3859         /*
3860          * If a global cgroup was set, apply it to all the events without an
3861          * explicit cgroup. I.e.:
3862          *
3863          *      trace -G A -e sched:*switch
3864          *
3865          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3866          * _and_ sched:sched_switch to the 'A' cgroup, while:
3867          *
3868          * trace -e sched:*switch -G A
3869          *
3870          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3871          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3872          * a cgroup (on the root cgroup, sys wide, etc).
3873          *
3874          * Multiple cgroups:
3875          *
3876          * trace -G A -e sched:*switch -G B
3877          *
3878          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3879          * to the 'B' cgroup.
3880          *
3881          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3882          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3883          */
3884         if (trace->cgroup)
3885                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3886
3887         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3888         if (err < 0) {
3889                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3890                 goto out_delete_evlist;
3891         }
3892
3893         err = trace__symbols_init(trace, evlist);
3894         if (err < 0) {
3895                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3896                 goto out_delete_evlist;
3897         }
3898
3899         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3900
3901         signal(SIGCHLD, sig_handler);
3902         signal(SIGINT, sig_handler);
3903
3904         if (forks) {
3905                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3906                                                     argv, false, NULL);
3907                 if (err < 0) {
3908                         fprintf(trace->output, "Couldn't run the workload!\n");
3909                         goto out_delete_evlist;
3910                 }
3911         }
3912
3913         err = evlist__open(evlist);
3914         if (err < 0)
3915                 goto out_error_open;
3916
3917         err = bpf__apply_obj_config();
3918         if (err) {
3919                 char errbuf[BUFSIZ];
3920
3921                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3922                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3923                          errbuf);
3924                 goto out_error_open;
3925         }
3926
3927         err = trace__set_filter_pids(trace);
3928         if (err < 0)
3929                 goto out_error_mem;
3930
3931         if (trace->syscalls.map)
3932                 trace__init_syscalls_bpf_map(trace);
3933
3934         if (trace->syscalls.prog_array.sys_enter)
3935                 trace__init_syscalls_bpf_prog_array_maps(trace);
3936
3937         if (trace->ev_qualifier_ids.nr > 0) {
3938                 err = trace__set_ev_qualifier_filter(trace);
3939                 if (err < 0)
3940                         goto out_errno;
3941
3942                 if (trace->syscalls.events.sys_exit) {
3943                         pr_debug("event qualifier tracepoint filter: %s\n",
3944                                  trace->syscalls.events.sys_exit->filter);
3945                 }
3946         }
3947
3948         /*
3949          * If the "close" syscall is not traced, then we will not have the
3950          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3951          * fd->pathname table and were ending up showing the last value set by
3952          * syscalls opening a pathname and associating it with a descriptor or
3953          * reading it from /proc/pid/fd/ in cases where that doesn't make
3954          * sense.
3955          *
3956          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3957          *  not in use.
3958          */
3959         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3960
3961         err = trace__expand_filters(trace, &evsel);
3962         if (err)
3963                 goto out_delete_evlist;
3964         err = perf_evlist__apply_filters(evlist, &evsel);
3965         if (err < 0)
3966                 goto out_error_apply_filters;
3967
3968         if (trace->dump.map)
3969                 bpf_map__fprintf(trace->dump.map, trace->output);
3970
3971         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3972         if (err < 0)
3973                 goto out_error_mmap;
3974
3975         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3976                 evlist__enable(evlist);
3977
3978         if (forks)
3979                 perf_evlist__start_workload(evlist);
3980
3981         if (trace->opts.initial_delay) {
3982                 usleep(trace->opts.initial_delay * 1000);
3983                 evlist__enable(evlist);
3984         }
3985
3986         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3987                                   evlist->core.threads->nr > 1 ||
3988                                   evlist__first(evlist)->core.attr.inherit;
3989
3990         /*
3991          * Now that we already used evsel->core.attr to ask the kernel to setup the
3992          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3993          * trace__resolve_callchain(), allowing per-event max-stack settings
3994          * to override an explicitly set --max-stack global setting.
3995          */
3996         evlist__for_each_entry(evlist, evsel) {
3997                 if (evsel__has_callchain(evsel) &&
3998                     evsel->core.attr.sample_max_stack == 0)
3999                         evsel->core.attr.sample_max_stack = trace->max_stack;
4000         }
4001 again:
4002         before = trace->nr_events;
4003
4004         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4005                 union perf_event *event;
4006                 struct mmap *md;
4007
4008                 md = &evlist->mmap[i];
4009                 if (perf_mmap__read_init(&md->core) < 0)
4010                         continue;
4011
4012                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4013                         ++trace->nr_events;
4014
4015                         err = trace__deliver_event(trace, event);
4016                         if (err)
4017                                 goto out_disable;
4018
4019                         perf_mmap__consume(&md->core);
4020
4021                         if (interrupted)
4022                                 goto out_disable;
4023
4024                         if (done && !draining) {
4025                                 evlist__disable(evlist);
4026                                 draining = true;
4027                         }
4028                 }
4029                 perf_mmap__read_done(&md->core);
4030         }
4031
4032         if (trace->nr_events == before) {
4033                 int timeout = done ? 100 : -1;
4034
4035                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4036                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4037                                 draining = true;
4038
4039                         goto again;
4040                 } else {
4041                         if (trace__flush_events(trace))
4042                                 goto out_disable;
4043                 }
4044         } else {
4045                 goto again;
4046         }
4047
4048 out_disable:
4049         thread__zput(trace->current);
4050
4051         evlist__disable(evlist);
4052
4053         if (trace->sort_events)
4054                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4055
4056         if (!err) {
4057                 if (trace->summary)
4058                         trace__fprintf_thread_summary(trace, trace->output);
4059
4060                 if (trace->show_tool_stats) {
4061                         fprintf(trace->output, "Stats:\n "
4062                                                " vfs_getname : %" PRIu64 "\n"
4063                                                " proc_getname: %" PRIu64 "\n",
4064                                 trace->stats.vfs_getname,
4065                                 trace->stats.proc_getname);
4066                 }
4067         }
4068
4069 out_delete_evlist:
4070         trace__symbols__exit(trace);
4071
4072         evlist__delete(evlist);
4073         cgroup__put(trace->cgroup);
4074         trace->evlist = NULL;
4075         trace->live = false;
4076         return err;
4077 {
4078         char errbuf[BUFSIZ];
4079
4080 out_error_sched_stat_runtime:
4081         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4082         goto out_error;
4083
4084 out_error_raw_syscalls:
4085         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4086         goto out_error;
4087
4088 out_error_mmap:
4089         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4090         goto out_error;
4091
4092 out_error_open:
4093         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4094
4095 out_error:
4096         fprintf(trace->output, "%s\n", errbuf);
4097         goto out_delete_evlist;
4098
4099 out_error_apply_filters:
4100         fprintf(trace->output,
4101                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4102                 evsel->filter, perf_evsel__name(evsel), errno,
4103                 str_error_r(errno, errbuf, sizeof(errbuf)));
4104         goto out_delete_evlist;
4105 }
4106 out_error_mem:
4107         fprintf(trace->output, "Not enough memory to run!\n");
4108         goto out_delete_evlist;
4109
4110 out_errno:
4111         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4112         goto out_delete_evlist;
4113 }
4114
4115 static int trace__replay(struct trace *trace)
4116 {
4117         const struct evsel_str_handler handlers[] = {
4118                 { "probe:vfs_getname",       trace__vfs_getname, },
4119         };
4120         struct perf_data data = {
4121                 .path  = input_name,
4122                 .mode  = PERF_DATA_MODE_READ,
4123                 .force = trace->force,
4124         };
4125         struct perf_session *session;
4126         struct evsel *evsel;
4127         int err = -1;
4128
4129         trace->tool.sample        = trace__process_sample;
4130         trace->tool.mmap          = perf_event__process_mmap;
4131         trace->tool.mmap2         = perf_event__process_mmap2;
4132         trace->tool.comm          = perf_event__process_comm;
4133         trace->tool.exit          = perf_event__process_exit;
4134         trace->tool.fork          = perf_event__process_fork;
4135         trace->tool.attr          = perf_event__process_attr;
4136         trace->tool.tracing_data  = perf_event__process_tracing_data;
4137         trace->tool.build_id      = perf_event__process_build_id;
4138         trace->tool.namespaces    = perf_event__process_namespaces;
4139
4140         trace->tool.ordered_events = true;
4141         trace->tool.ordering_requires_timestamps = true;
4142
4143         /* add tid to output */
4144         trace->multiple_threads = true;
4145
4146         session = perf_session__new(&data, false, &trace->tool);
4147         if (IS_ERR(session))
4148                 return PTR_ERR(session);
4149
4150         if (trace->opts.target.pid)
4151                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4152
4153         if (trace->opts.target.tid)
4154                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4155
4156         if (symbol__init(&session->header.env) < 0)
4157                 goto out;
4158
4159         trace->host = &session->machines.host;
4160
4161         err = perf_session__set_tracepoints_handlers(session, handlers);
4162         if (err)
4163                 goto out;
4164
4165         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4166                                                      "raw_syscalls:sys_enter");
4167         /* older kernels have syscalls tp versus raw_syscalls */
4168         if (evsel == NULL)
4169                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4170                                                              "syscalls:sys_enter");
4171
4172         if (evsel &&
4173             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4174             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4175                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4176                 goto out;
4177         }
4178
4179         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4180                                                      "raw_syscalls:sys_exit");
4181         if (evsel == NULL)
4182                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4183                                                              "syscalls:sys_exit");
4184         if (evsel &&
4185             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4186             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4187                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4188                 goto out;
4189         }
4190
4191         evlist__for_each_entry(session->evlist, evsel) {
4192                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4193                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4194                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4195                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4196                         evsel->handler = trace__pgfault;
4197         }
4198
4199         setup_pager();
4200
4201         err = perf_session__process_events(session);
4202         if (err)
4203                 pr_err("Failed to process events, error %d", err);
4204
4205         else if (trace->summary)
4206                 trace__fprintf_thread_summary(trace, trace->output);
4207
4208 out:
4209         perf_session__delete(session);
4210
4211         return err;
4212 }
4213
4214 static size_t trace__fprintf_threads_header(FILE *fp)
4215 {
4216         size_t printed;
4217
4218         printed  = fprintf(fp, "\n Summary of events:\n\n");
4219
4220         return printed;
4221 }
4222
4223 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4224         struct syscall_stats *stats;
4225         double               msecs;
4226         int                  syscall;
4227 )
4228 {
4229         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4230         struct syscall_stats *stats = source->priv;
4231
4232         entry->syscall = source->i;
4233         entry->stats   = stats;
4234         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4235 }
4236
4237 static size_t thread__dump_stats(struct thread_trace *ttrace,
4238                                  struct trace *trace, FILE *fp)
4239 {
4240         size_t printed = 0;
4241         struct syscall *sc;
4242         struct rb_node *nd;
4243         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4244
4245         if (syscall_stats == NULL)
4246                 return 0;
4247
4248         printed += fprintf(fp, "\n");
4249
4250         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4251         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4252         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4253
4254         resort_rb__for_each_entry(nd, syscall_stats) {
4255                 struct syscall_stats *stats = syscall_stats_entry->stats;
4256                 if (stats) {
4257                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4258                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4259                         double avg = avg_stats(&stats->stats);
4260                         double pct;
4261                         u64 n = (u64)stats->stats.n;
4262
4263                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4264                         avg /= NSEC_PER_MSEC;
4265
4266                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4267                         printed += fprintf(fp, "   %-15s", sc->name);
4268                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4269                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4270                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4271
4272                         if (trace->errno_summary && stats->nr_failures) {
4273                                 const char *arch_name = perf_env__arch(trace->host->env);
4274                                 int e;
4275
4276                                 for (e = 0; e < stats->max_errno; ++e) {
4277                                         if (stats->errnos[e] != 0)
4278                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4279                                 }
4280                         }
4281                 }
4282         }
4283
4284         resort_rb__delete(syscall_stats);
4285         printed += fprintf(fp, "\n\n");
4286
4287         return printed;
4288 }
4289
4290 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4291 {
4292         size_t printed = 0;
4293         struct thread_trace *ttrace = thread__priv(thread);
4294         double ratio;
4295
4296         if (ttrace == NULL)
4297                 return 0;
4298
4299         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4300
4301         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4302         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4303         printed += fprintf(fp, "%.1f%%", ratio);
4304         if (ttrace->pfmaj)
4305                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4306         if (ttrace->pfmin)
4307                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4308         if (trace->sched)
4309                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4310         else if (fputc('\n', fp) != EOF)
4311                 ++printed;
4312
4313         printed += thread__dump_stats(ttrace, trace, fp);
4314
4315         return printed;
4316 }
4317
4318 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4319 {
4320         return ttrace ? ttrace->nr_events : 0;
4321 }
4322
4323 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4324         struct thread *thread;
4325 )
4326 {
4327         entry->thread = rb_entry(nd, struct thread, rb_node);
4328 }
4329
4330 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4331 {
4332         size_t printed = trace__fprintf_threads_header(fp);
4333         struct rb_node *nd;
4334         int i;
4335
4336         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4337                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4338
4339                 if (threads == NULL) {
4340                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4341                         return 0;
4342                 }
4343
4344                 resort_rb__for_each_entry(nd, threads)
4345                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4346
4347                 resort_rb__delete(threads);
4348         }
4349         return printed;
4350 }
4351
4352 static int trace__set_duration(const struct option *opt, const char *str,
4353                                int unset __maybe_unused)
4354 {
4355         struct trace *trace = opt->value;
4356
4357         trace->duration_filter = atof(str);
4358         return 0;
4359 }
4360
4361 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4362                                               int unset __maybe_unused)
4363 {
4364         int ret = -1;
4365         size_t i;
4366         struct trace *trace = opt->value;
4367         /*
4368          * FIXME: introduce a intarray class, plain parse csv and create a
4369          * { int nr, int entries[] } struct...
4370          */
4371         struct intlist *list = intlist__new(str);
4372
4373         if (list == NULL)
4374                 return -1;
4375
4376         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4377         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4378
4379         if (trace->filter_pids.entries == NULL)
4380                 goto out;
4381
4382         trace->filter_pids.entries[0] = getpid();
4383
4384         for (i = 1; i < trace->filter_pids.nr; ++i)
4385                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4386
4387         intlist__delete(list);
4388         ret = 0;
4389 out:
4390         return ret;
4391 }
4392
4393 static int trace__open_output(struct trace *trace, const char *filename)
4394 {
4395         struct stat st;
4396
4397         if (!stat(filename, &st) && st.st_size) {
4398                 char oldname[PATH_MAX];
4399
4400                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4401                 unlink(oldname);
4402                 rename(filename, oldname);
4403         }
4404
4405         trace->output = fopen(filename, "w");
4406
4407         return trace->output == NULL ? -errno : 0;
4408 }
4409
4410 static int parse_pagefaults(const struct option *opt, const char *str,
4411                             int unset __maybe_unused)
4412 {
4413         int *trace_pgfaults = opt->value;
4414
4415         if (strcmp(str, "all") == 0)
4416                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4417         else if (strcmp(str, "maj") == 0)
4418                 *trace_pgfaults |= TRACE_PFMAJ;
4419         else if (strcmp(str, "min") == 0)
4420                 *trace_pgfaults |= TRACE_PFMIN;
4421         else
4422                 return -1;
4423
4424         return 0;
4425 }
4426
4427 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4428 {
4429         struct evsel *evsel;
4430
4431         evlist__for_each_entry(evlist, evsel) {
4432                 if (evsel->handler == NULL)
4433                         evsel->handler = handler;
4434         }
4435 }
4436
4437 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4438 {
4439         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4440
4441         if (fmt) {
4442                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4443
4444                 if (scfmt) {
4445                         int skip = 0;
4446
4447                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4448                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4449                                 ++skip;
4450
4451                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4452                 }
4453         }
4454 }
4455
4456 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4457 {
4458         struct evsel *evsel;
4459
4460         evlist__for_each_entry(evlist, evsel) {
4461                 if (evsel->priv || !evsel->tp_format)
4462                         continue;
4463
4464                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4465                         perf_evsel__init_tp_arg_scnprintf(evsel);
4466                         continue;
4467                 }
4468
4469                 if (perf_evsel__init_syscall_tp(evsel))
4470                         return -1;
4471
4472                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4473                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4474
4475                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4476                                 return -1;
4477
4478                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4479                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4480                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4481
4482                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4483                                 return -1;
4484
4485                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4486                 }
4487         }
4488
4489         return 0;
4490 }
4491
4492 /*
4493  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4494  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4495  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4496  *
4497  * It'd be better to introduce a parse_options() variant that would return a
4498  * list with the terms it didn't match to an event...
4499  */
4500 static int trace__parse_events_option(const struct option *opt, const char *str,
4501                                       int unset __maybe_unused)
4502 {
4503         struct trace *trace = (struct trace *)opt->value;
4504         const char *s = str;
4505         char *sep = NULL, *lists[2] = { NULL, NULL, };
4506         int len = strlen(str) + 1, err = -1, list, idx;
4507         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4508         char group_name[PATH_MAX];
4509         struct syscall_fmt *fmt;
4510
4511         if (strace_groups_dir == NULL)
4512                 return -1;
4513
4514         if (*s == '!') {
4515                 ++s;
4516                 trace->not_ev_qualifier = true;
4517         }
4518
4519         while (1) {
4520                 if ((sep = strchr(s, ',')) != NULL)
4521                         *sep = '\0';
4522
4523                 list = 0;
4524                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4525                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4526                         list = 1;
4527                         goto do_concat;
4528                 }
4529
4530                 fmt = syscall_fmt__find_by_alias(s);
4531                 if (fmt != NULL) {
4532                         list = 1;
4533                         s = fmt->name;
4534                 } else {
4535                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4536                         if (access(group_name, R_OK) == 0)
4537                                 list = 1;
4538                 }
4539 do_concat:
4540                 if (lists[list]) {
4541                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4542                 } else {
4543                         lists[list] = malloc(len);
4544                         if (lists[list] == NULL)
4545                                 goto out;
4546                         strcpy(lists[list], s);
4547                 }
4548
4549                 if (!sep)
4550                         break;
4551
4552                 *sep = ',';
4553                 s = sep + 1;
4554         }
4555
4556         if (lists[1] != NULL) {
4557                 struct strlist_config slist_config = {
4558                         .dirname = strace_groups_dir,
4559                 };
4560
4561                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4562                 if (trace->ev_qualifier == NULL) {
4563                         fputs("Not enough memory to parse event qualifier", trace->output);
4564                         goto out;
4565                 }
4566
4567                 if (trace__validate_ev_qualifier(trace))
4568                         goto out;
4569                 trace->trace_syscalls = true;
4570         }
4571
4572         err = 0;
4573
4574         if (lists[0]) {
4575                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4576                                                "event selector. use 'perf list' to list available events",
4577                                                parse_events_option);
4578                 err = parse_events_option(&o, lists[0], 0);
4579         }
4580 out:
4581         if (sep)
4582                 *sep = ',';
4583
4584         return err;
4585 }
4586
4587 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4588 {
4589         struct trace *trace = opt->value;
4590
4591         if (!list_empty(&trace->evlist->core.entries))
4592                 return parse_cgroups(opt, str, unset);
4593
4594         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4595
4596         return 0;
4597 }
4598
4599 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4600 {
4601         if (trace->bpf_obj == NULL)
4602                 return NULL;
4603
4604         return bpf_object__find_map_by_name(trace->bpf_obj, name);
4605 }
4606
4607 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4608 {
4609         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4610 }
4611
4612 static void trace__set_bpf_map_syscalls(struct trace *trace)
4613 {
4614         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4615         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4616         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4617 }
4618
4619 static int trace__config(const char *var, const char *value, void *arg)
4620 {
4621         struct trace *trace = arg;
4622         int err = 0;
4623
4624         if (!strcmp(var, "trace.add_events")) {
4625                 trace->perfconfig_events = strdup(value);
4626                 if (trace->perfconfig_events == NULL) {
4627                         pr_err("Not enough memory for %s\n", "trace.add_events");
4628                         return -1;
4629                 }
4630         } else if (!strcmp(var, "trace.show_timestamp")) {
4631                 trace->show_tstamp = perf_config_bool(var, value);
4632         } else if (!strcmp(var, "trace.show_duration")) {
4633                 trace->show_duration = perf_config_bool(var, value);
4634         } else if (!strcmp(var, "trace.show_arg_names")) {
4635                 trace->show_arg_names = perf_config_bool(var, value);
4636                 if (!trace->show_arg_names)
4637                         trace->show_zeros = true;
4638         } else if (!strcmp(var, "trace.show_zeros")) {
4639                 bool new_show_zeros = perf_config_bool(var, value);
4640                 if (!trace->show_arg_names && !new_show_zeros) {
4641                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4642                         goto out;
4643                 }
4644                 trace->show_zeros = new_show_zeros;
4645         } else if (!strcmp(var, "trace.show_prefix")) {
4646                 trace->show_string_prefix = perf_config_bool(var, value);
4647         } else if (!strcmp(var, "trace.no_inherit")) {
4648                 trace->opts.no_inherit = perf_config_bool(var, value);
4649         } else if (!strcmp(var, "trace.args_alignment")) {
4650                 int args_alignment = 0;
4651                 if (perf_config_int(&args_alignment, var, value) == 0)
4652                         trace->args_alignment = args_alignment;
4653         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4654                 if (strcasecmp(value, "libtraceevent") == 0)
4655                         trace->libtraceevent_print = true;
4656                 else if (strcasecmp(value, "libbeauty") == 0)
4657                         trace->libtraceevent_print = false;
4658         }
4659 out:
4660         return err;
4661 }
4662
4663 int cmd_trace(int argc, const char **argv)
4664 {
4665         const char *trace_usage[] = {
4666                 "perf trace [<options>] [<command>]",
4667                 "perf trace [<options>] -- <command> [<options>]",
4668                 "perf trace record [<options>] [<command>]",
4669                 "perf trace record [<options>] -- <command> [<options>]",
4670                 NULL
4671         };
4672         struct trace trace = {
4673                 .opts = {
4674                         .target = {
4675                                 .uid       = UINT_MAX,
4676                                 .uses_mmap = true,
4677                         },
4678                         .user_freq     = UINT_MAX,
4679                         .user_interval = ULLONG_MAX,
4680                         .no_buffering  = true,
4681                         .mmap_pages    = UINT_MAX,
4682                 },
4683                 .output = stderr,
4684                 .show_comm = true,
4685                 .show_tstamp = true,
4686                 .show_duration = true,
4687                 .show_arg_names = true,
4688                 .args_alignment = 70,
4689                 .trace_syscalls = false,
4690                 .kernel_syscallchains = false,
4691                 .max_stack = UINT_MAX,
4692                 .max_events = ULONG_MAX,
4693         };
4694         const char *map_dump_str = NULL;
4695         const char *output_name = NULL;
4696         const struct option trace_options[] = {
4697         OPT_CALLBACK('e', "event", &trace, "event",
4698                      "event/syscall selector. use 'perf list' to list available events",
4699                      trace__parse_events_option),
4700         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4701                      "event filter", parse_filter),
4702         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4703                     "show the thread COMM next to its id"),
4704         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4705         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4706                      trace__parse_events_option),
4707         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4708         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4709         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4710                     "trace events on existing process id"),
4711         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4712                     "trace events on existing thread id"),
4713         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4714                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4715         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4716                     "system-wide collection from all CPUs"),
4717         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4718                     "list of cpus to monitor"),
4719         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4720                     "child tasks do not inherit counters"),
4721         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4722                      "number of mmap data pages",
4723                      perf_evlist__parse_mmap_pages),
4724         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4725                    "user to profile"),
4726         OPT_CALLBACK(0, "duration", &trace, "float",
4727                      "show only events with duration > N.M ms",
4728                      trace__set_duration),
4729 #ifdef HAVE_LIBBPF_SUPPORT
4730         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4731 #endif
4732         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4733         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4734         OPT_BOOLEAN('T', "time", &trace.full_time,
4735                     "Show full timestamp, not time relative to first start"),
4736         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4737                     "Show only syscalls that failed"),
4738         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4739                     "Show only syscall summary with statistics"),
4740         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4741                     "Show all syscalls and summary with statistics"),
4742         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4743                     "Show errno stats per syscall, use with -s or -S"),
4744         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4745                      "Trace pagefaults", parse_pagefaults, "maj"),
4746         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4747         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4748         OPT_CALLBACK(0, "call-graph", &trace.opts,
4749                      "record_mode[,record_size]", record_callchain_help,
4750                      &record_parse_callchain_opt),
4751         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4752                     "Use libtraceevent to print the tracepoint arguments."),
4753         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4754                     "Show the kernel callchains on the syscall exit path"),
4755         OPT_ULONG(0, "max-events", &trace.max_events,
4756                 "Set the maximum number of events to print, exit after that is reached. "),
4757         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4758                      "Set the minimum stack depth when parsing the callchain, "
4759                      "anything below the specified depth will be ignored."),
4760         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4761                      "Set the maximum stack depth when parsing the callchain, "
4762                      "anything beyond the specified depth will be ignored. "
4763                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4764         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4765                         "Sort batch of events before processing, use if getting out of order events"),
4766         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4767                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4768         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4769                         "per thread proc mmap processing timeout in ms"),
4770         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4771                      trace__parse_cgroups),
4772         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4773                      "ms to wait before starting measurement after program "
4774                      "start"),
4775         OPTS_EVSWITCH(&trace.evswitch),
4776         OPT_END()
4777         };
4778         bool __maybe_unused max_stack_user_set = true;
4779         bool mmap_pages_user_set = true;
4780         struct evsel *evsel;
4781         const char * const trace_subcommands[] = { "record", NULL };
4782         int err = -1;
4783         char bf[BUFSIZ];
4784
4785         signal(SIGSEGV, sighandler_dump_stack);
4786         signal(SIGFPE, sighandler_dump_stack);
4787
4788         trace.evlist = evlist__new();
4789         trace.sctbl = syscalltbl__new();
4790
4791         if (trace.evlist == NULL || trace.sctbl == NULL) {
4792                 pr_err("Not enough memory to run!\n");
4793                 err = -ENOMEM;
4794                 goto out;
4795         }
4796
4797         /*
4798          * Parsing .perfconfig may entail creating a BPF event, that may need
4799          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4800          * is too small. This affects just this process, not touching the
4801          * global setting. If it fails we'll get something in 'perf trace -v'
4802          * to help diagnose the problem.
4803          */
4804         rlimit__bump_memlock();
4805
4806         err = perf_config(trace__config, &trace);
4807         if (err)
4808                 goto out;
4809
4810         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4811                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4812
4813         /*
4814          * Here we already passed thru trace__parse_events_option() and it has
4815          * already figured out if -e syscall_name, if not but if --event
4816          * foo:bar was used, the user is interested _just_ in those, say,
4817          * tracepoint events, not in the strace-like syscall-name-based mode.
4818          *
4819          * This is important because we need to check if strace-like mode is
4820          * needed to decided if we should filter out the eBPF
4821          * __augmented_syscalls__ code, if it is in the mix, say, via
4822          * .perfconfig trace.add_events, and filter those out.
4823          */
4824         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4825             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4826                 trace.trace_syscalls = true;
4827         }
4828         /*
4829          * Now that we have --verbose figured out, lets see if we need to parse
4830          * events from .perfconfig, so that if those events fail parsing, say some
4831          * BPF program fails, then we'll be able to use --verbose to see what went
4832          * wrong in more detail.
4833          */
4834         if (trace.perfconfig_events != NULL) {
4835                 struct parse_events_error parse_err = { .idx = 0, };
4836
4837                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4838                 if (err) {
4839                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4840                         goto out;
4841                 }
4842         }
4843
4844         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4845                 usage_with_options_msg(trace_usage, trace_options,
4846                                        "cgroup monitoring only available in system-wide mode");
4847         }
4848
4849         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4850         if (IS_ERR(evsel)) {
4851                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4852                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4853                 goto out;
4854         }
4855
4856         if (evsel) {
4857                 trace.syscalls.events.augmented = evsel;
4858
4859                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4860                 if (evsel == NULL) {
4861                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4862                         goto out;
4863                 }
4864
4865                 if (evsel->bpf_obj == NULL) {
4866                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4867                         goto out;
4868                 }
4869
4870                 trace.bpf_obj = evsel->bpf_obj;
4871
4872                 /*
4873                  * If we have _just_ the augmenter event but don't have a
4874                  * explicit --syscalls, then assume we want all strace-like
4875                  * syscalls:
4876                  */
4877                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4878                         trace.trace_syscalls = true;
4879                 /*
4880                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4881                  * strace-like syscall tracing is not set, then we need to trow
4882                  * away the augmenter, i.e. all the events that were created
4883                  * from that BPF object file.
4884                  *
4885                  * This is more to fix the current .perfconfig trace.add_events
4886                  * style of setting up the strace-like eBPF based syscall point
4887                  * payload augmenter.
4888                  *
4889                  * All this complexity will be avoided by adding an alternative
4890                  * to trace.add_events in the form of
4891                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4892                  * need it.
4893                  *
4894                  * .perfconfig trace.add_events is still useful if we want, for
4895                  * instance, have msr_write.msr in some .perfconfig profile based
4896                  * 'perf trace --config determinism.profile' mode, where for some
4897                  * particular goal/workload type we want a set of events and
4898                  * output mode (with timings, etc) instead of having to add
4899                  * all via the command line.
4900                  *
4901                  * Also --config to specify an alternate .perfconfig file needs
4902                  * to be implemented.
4903                  */
4904                 if (!trace.trace_syscalls) {
4905                         trace__delete_augmented_syscalls(&trace);
4906                 } else {
4907                         trace__set_bpf_map_filtered_pids(&trace);
4908                         trace__set_bpf_map_syscalls(&trace);
4909                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4910                 }
4911         }
4912
4913         err = bpf__setup_stdout(trace.evlist);
4914         if (err) {
4915                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4916                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4917                 goto out;
4918         }
4919
4920         err = -1;
4921
4922         if (map_dump_str) {
4923                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4924                 if (trace.dump.map == NULL) {
4925                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4926                         goto out;
4927                 }
4928         }
4929
4930         if (trace.trace_pgfaults) {
4931                 trace.opts.sample_address = true;
4932                 trace.opts.sample_time = true;
4933         }
4934
4935         if (trace.opts.mmap_pages == UINT_MAX)
4936                 mmap_pages_user_set = false;
4937
4938         if (trace.max_stack == UINT_MAX) {
4939                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4940                 max_stack_user_set = false;
4941         }
4942
4943 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4944         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4945                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4946         }
4947 #endif
4948
4949         if (callchain_param.enabled) {
4950                 if (!mmap_pages_user_set && geteuid() == 0)
4951                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4952
4953                 symbol_conf.use_callchain = true;
4954         }
4955
4956         if (trace.evlist->core.nr_entries > 0) {
4957                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4958                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4959                         perror("failed to set syscalls:* tracepoint fields");
4960                         goto out;
4961                 }
4962         }
4963
4964         if (trace.sort_events) {
4965                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4966                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4967         }
4968
4969         /*
4970          * If we are augmenting syscalls, then combine what we put in the
4971          * __augmented_syscalls__ BPF map with what is in the
4972          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4973          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4974          *
4975          * We'll switch to look at two BPF maps, one for sys_enter and the
4976          * other for sys_exit when we start augmenting the sys_exit paths with
4977          * buffers that are being copied from kernel to userspace, think 'read'
4978          * syscall.
4979          */
4980         if (trace.syscalls.events.augmented) {
4981                 evlist__for_each_entry(trace.evlist, evsel) {
4982                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4983
4984                         if (raw_syscalls_sys_exit) {
4985                                 trace.raw_augmented_syscalls = true;
4986                                 goto init_augmented_syscall_tp;
4987                         }
4988
4989                         if (trace.syscalls.events.augmented->priv == NULL &&
4990                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
4991                                 struct evsel *augmented = trace.syscalls.events.augmented;
4992                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
4993                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
4994                                         goto out;
4995                                 /*
4996                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4997                                  * Above we made sure we can get from the payload the tp fields
4998                                  * that we get from syscalls:sys_enter tracefs format file.
4999                                  */
5000                                 augmented->handler = trace__sys_enter;
5001                                 /*
5002                                  * Now we do the same for the *syscalls:sys_enter event so that
5003                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5004                                  * as not to filter it, then we'll handle it just like we would
5005                                  * for the BPF_OUTPUT one:
5006                                  */
5007                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
5008                                     perf_evsel__init_augmented_syscall_tp_args(evsel))
5009                                         goto out;
5010                                 evsel->handler = trace__sys_enter;
5011                         }
5012
5013                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
5014                                 struct syscall_tp *sc;
5015 init_augmented_syscall_tp:
5016                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
5017                                         goto out;
5018                                 sc = __evsel__syscall_tp(evsel);
5019                                 /*
5020                                  * For now with BPF raw_augmented we hook into
5021                                  * raw_syscalls:sys_enter and there we get all
5022                                  * 6 syscall args plus the tracepoint common
5023                                  * fields and the syscall_nr (another long).
5024                                  * So we check if that is the case and if so
5025                                  * don't look after the sc->args_size but
5026                                  * always after the full raw_syscalls:sys_enter
5027                                  * payload, which is fixed.
5028                                  *
5029                                  * We'll revisit this later to pass
5030                                  * s->args_size to the BPF augmenter (now
5031                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5032                                  * so that it copies only what we need for each
5033                                  * syscall, like what happens when we use
5034                                  * syscalls:sys_enter_NAME, so that we reduce
5035                                  * the kernel/userspace traffic to just what is
5036                                  * needed for each syscall.
5037                                  */
5038                                 if (trace.raw_augmented_syscalls)
5039                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5040                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
5041                                 evsel->handler = trace__sys_exit;
5042                         }
5043                 }
5044         }
5045
5046         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5047                 return trace__record(&trace, argc-1, &argv[1]);
5048
5049         /* Using just --errno-summary will trigger --summary */
5050         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5051                 trace.summary_only = true;
5052
5053         /* summary_only implies summary option, but don't overwrite summary if set */
5054         if (trace.summary_only)
5055                 trace.summary = trace.summary_only;
5056
5057         if (output_name != NULL) {
5058                 err = trace__open_output(&trace, output_name);
5059                 if (err < 0) {
5060                         perror("failed to create output file");
5061                         goto out;
5062                 }
5063         }
5064
5065         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5066         if (err)
5067                 goto out_close;
5068
5069         err = target__validate(&trace.opts.target);
5070         if (err) {
5071                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5072                 fprintf(trace.output, "%s", bf);
5073                 goto out_close;
5074         }
5075
5076         err = target__parse_uid(&trace.opts.target);
5077         if (err) {
5078                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5079                 fprintf(trace.output, "%s", bf);
5080                 goto out_close;
5081         }
5082
5083         if (!argc && target__none(&trace.opts.target))
5084                 trace.opts.target.system_wide = true;
5085
5086         if (input_name)
5087                 err = trace__replay(&trace);
5088         else
5089                 err = trace__run(&trace, argc, argv);
5090
5091 out_close:
5092         if (output_name != NULL)
5093                 fclose(trace.output);
5094 out:
5095         zfree(&trace.perfconfig_events);
5096         return err;
5097 }