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                            .strtoul   = STUL_STRARRAY,
1070                            .parm      = &strarray__prctl_options, },
1071                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1072                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1073         { .name     = "pread", .alias = "pread64", },
1074         { .name     = "preadv", .alias = "pread", },
1075         { .name     = "prlimit64",
1076           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1077         { .name     = "pwrite", .alias = "pwrite64", },
1078         { .name     = "readlinkat",
1079           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1080         { .name     = "recvfrom",
1081           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1082         { .name     = "recvmmsg",
1083           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1084         { .name     = "recvmsg",
1085           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1086         { .name     = "renameat",
1087           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1088                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1089         { .name     = "renameat2",
1090           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1091                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1092                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1093         { .name     = "rt_sigaction",
1094           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1095         { .name     = "rt_sigprocmask",
1096           .arg = { [0] = STRARRAY(how, sighow), }, },
1097         { .name     = "rt_sigqueueinfo",
1098           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1099         { .name     = "rt_tgsigqueueinfo",
1100           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1101         { .name     = "sched_setscheduler",
1102           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1103         { .name     = "seccomp",
1104           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1105                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1106         { .name     = "select", .timeout = true, },
1107         { .name     = "sendfile", .alias = "sendfile64", },
1108         { .name     = "sendmmsg",
1109           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1110         { .name     = "sendmsg",
1111           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1112         { .name     = "sendto",
1113           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1114                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1115         { .name     = "set_tid_address", .errpid = true, },
1116         { .name     = "setitimer",
1117           .arg = { [0] = STRARRAY(which, itimers), }, },
1118         { .name     = "setrlimit",
1119           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1120         { .name     = "socket",
1121           .arg = { [0] = STRARRAY(family, socket_families),
1122                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1123                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1124         { .name     = "socketpair",
1125           .arg = { [0] = STRARRAY(family, socket_families),
1126                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1127                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1128         { .name     = "stat", .alias = "newstat", },
1129         { .name     = "statx",
1130           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1131                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1132                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1133         { .name     = "swapoff",
1134           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1135         { .name     = "swapon",
1136           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1137         { .name     = "symlinkat",
1138           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1139         { .name     = "sync_file_range",
1140           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1141         { .name     = "tgkill",
1142           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1143         { .name     = "tkill",
1144           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1145         { .name     = "umount2", .alias = "umount",
1146           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1147         { .name     = "uname", .alias = "newuname", },
1148         { .name     = "unlinkat",
1149           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1150         { .name     = "utimensat",
1151           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1152         { .name     = "wait4",      .errpid = true,
1153           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1154         { .name     = "waitid",     .errpid = true,
1155           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1156 };
1157
1158 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1159 {
1160         const struct syscall_fmt *fmt = fmtp;
1161         return strcmp(name, fmt->name);
1162 }
1163
1164 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1165 {
1166         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1167 }
1168
1169 static struct syscall_fmt *syscall_fmt__find(const char *name)
1170 {
1171         const int nmemb = ARRAY_SIZE(syscall_fmts);
1172         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1173 }
1174
1175 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1176 {
1177         int i;
1178
1179         for (i = 0; i < nmemb; ++i) {
1180                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1181                         return &fmts[i];
1182         }
1183
1184         return NULL;
1185 }
1186
1187 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1188 {
1189         const int nmemb = ARRAY_SIZE(syscall_fmts);
1190         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1191 }
1192
1193 /*
1194  * is_exit: is this "exit" or "exit_group"?
1195  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1196  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1197  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1198  */
1199 struct syscall {
1200         struct tep_event    *tp_format;
1201         int                 nr_args;
1202         int                 args_size;
1203         struct {
1204                 struct bpf_program *sys_enter,
1205                                    *sys_exit;
1206         }                   bpf_prog;
1207         bool                is_exit;
1208         bool                is_open;
1209         bool                nonexistent;
1210         struct tep_format_field *args;
1211         const char          *name;
1212         struct syscall_fmt  *fmt;
1213         struct syscall_arg_fmt *arg_fmt;
1214 };
1215
1216 /*
1217  * Must match what is in the BPF program:
1218  *
1219  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1220  */
1221 struct bpf_map_syscall_entry {
1222         bool    enabled;
1223         u16     string_args_len[6];
1224 };
1225
1226 /*
1227  * We need to have this 'calculated' boolean because in some cases we really
1228  * don't know what is the duration of a syscall, for instance, when we start
1229  * a session and some threads are waiting for a syscall to finish, say 'poll',
1230  * in which case all we can do is to print "( ? ) for duration and for the
1231  * start timestamp.
1232  */
1233 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1234 {
1235         double duration = (double)t / NSEC_PER_MSEC;
1236         size_t printed = fprintf(fp, "(");
1237
1238         if (!calculated)
1239                 printed += fprintf(fp, "         ");
1240         else if (duration >= 1.0)
1241                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1242         else if (duration >= 0.01)
1243                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1244         else
1245                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1246         return printed + fprintf(fp, "): ");
1247 }
1248
1249 /**
1250  * filename.ptr: The filename char pointer that will be vfs_getname'd
1251  * filename.entry_str_pos: Where to insert the string translated from
1252  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1253  * ret_scnprintf: syscall args may set this to a different syscall return
1254  *                formatter, for instance, fcntl may return fds, file flags, etc.
1255  */
1256 struct thread_trace {
1257         u64               entry_time;
1258         bool              entry_pending;
1259         unsigned long     nr_events;
1260         unsigned long     pfmaj, pfmin;
1261         char              *entry_str;
1262         double            runtime_ms;
1263         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1264         struct {
1265                 unsigned long ptr;
1266                 short int     entry_str_pos;
1267                 bool          pending_open;
1268                 unsigned int  namelen;
1269                 char          *name;
1270         } filename;
1271         struct {
1272                 int           max;
1273                 struct file   *table;
1274         } files;
1275
1276         struct intlist *syscall_stats;
1277 };
1278
1279 static struct thread_trace *thread_trace__new(void)
1280 {
1281         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1282
1283         if (ttrace) {
1284                 ttrace->files.max = -1;
1285                 ttrace->syscall_stats = intlist__new(NULL);
1286         }
1287
1288         return ttrace;
1289 }
1290
1291 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1292 {
1293         struct thread_trace *ttrace;
1294
1295         if (thread == NULL)
1296                 goto fail;
1297
1298         if (thread__priv(thread) == NULL)
1299                 thread__set_priv(thread, thread_trace__new());
1300
1301         if (thread__priv(thread) == NULL)
1302                 goto fail;
1303
1304         ttrace = thread__priv(thread);
1305         ++ttrace->nr_events;
1306
1307         return ttrace;
1308 fail:
1309         color_fprintf(fp, PERF_COLOR_RED,
1310                       "WARNING: not enough memory, dropping samples!\n");
1311         return NULL;
1312 }
1313
1314
1315 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1316                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1317 {
1318         struct thread_trace *ttrace = thread__priv(arg->thread);
1319
1320         ttrace->ret_scnprintf = ret_scnprintf;
1321 }
1322
1323 #define TRACE_PFMAJ             (1 << 0)
1324 #define TRACE_PFMIN             (1 << 1)
1325
1326 static const size_t trace__entry_str_size = 2048;
1327
1328 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1329 {
1330         if (fd < 0)
1331                 return NULL;
1332
1333         if (fd > ttrace->files.max) {
1334                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1335
1336                 if (nfiles == NULL)
1337                         return NULL;
1338
1339                 if (ttrace->files.max != -1) {
1340                         memset(nfiles + ttrace->files.max + 1, 0,
1341                                (fd - ttrace->files.max) * sizeof(struct file));
1342                 } else {
1343                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1344                 }
1345
1346                 ttrace->files.table = nfiles;
1347                 ttrace->files.max   = fd;
1348         }
1349
1350         return ttrace->files.table + fd;
1351 }
1352
1353 struct file *thread__files_entry(struct thread *thread, int fd)
1354 {
1355         return thread_trace__files_entry(thread__priv(thread), fd);
1356 }
1357
1358 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1359 {
1360         struct thread_trace *ttrace = thread__priv(thread);
1361         struct file *file = thread_trace__files_entry(ttrace, fd);
1362
1363         if (file != NULL) {
1364                 struct stat st;
1365                 if (stat(pathname, &st) == 0)
1366                         file->dev_maj = major(st.st_rdev);
1367                 file->pathname = strdup(pathname);
1368                 if (file->pathname)
1369                         return 0;
1370         }
1371
1372         return -1;
1373 }
1374
1375 static int thread__read_fd_path(struct thread *thread, int fd)
1376 {
1377         char linkname[PATH_MAX], pathname[PATH_MAX];
1378         struct stat st;
1379         int ret;
1380
1381         if (thread->pid_ == thread->tid) {
1382                 scnprintf(linkname, sizeof(linkname),
1383                           "/proc/%d/fd/%d", thread->pid_, fd);
1384         } else {
1385                 scnprintf(linkname, sizeof(linkname),
1386                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1387         }
1388
1389         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1390                 return -1;
1391
1392         ret = readlink(linkname, pathname, sizeof(pathname));
1393
1394         if (ret < 0 || ret > st.st_size)
1395                 return -1;
1396
1397         pathname[ret] = '\0';
1398         return trace__set_fd_pathname(thread, fd, pathname);
1399 }
1400
1401 static const char *thread__fd_path(struct thread *thread, int fd,
1402                                    struct trace *trace)
1403 {
1404         struct thread_trace *ttrace = thread__priv(thread);
1405
1406         if (ttrace == NULL || trace->fd_path_disabled)
1407                 return NULL;
1408
1409         if (fd < 0)
1410                 return NULL;
1411
1412         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1413                 if (!trace->live)
1414                         return NULL;
1415                 ++trace->stats.proc_getname;
1416                 if (thread__read_fd_path(thread, fd))
1417                         return NULL;
1418         }
1419
1420         return ttrace->files.table[fd].pathname;
1421 }
1422
1423 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1424 {
1425         int fd = arg->val;
1426         size_t printed = scnprintf(bf, size, "%d", fd);
1427         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1428
1429         if (path)
1430                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1431
1432         return printed;
1433 }
1434
1435 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1436 {
1437         size_t printed = scnprintf(bf, size, "%d", fd);
1438         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1439
1440         if (thread) {
1441                 const char *path = thread__fd_path(thread, fd, trace);
1442
1443                 if (path)
1444                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1445
1446                 thread__put(thread);
1447         }
1448
1449         return printed;
1450 }
1451
1452 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1453                                               struct syscall_arg *arg)
1454 {
1455         int fd = arg->val;
1456         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1457         struct thread_trace *ttrace = thread__priv(arg->thread);
1458
1459         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1460                 zfree(&ttrace->files.table[fd].pathname);
1461
1462         return printed;
1463 }
1464
1465 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1466                                      unsigned long ptr)
1467 {
1468         struct thread_trace *ttrace = thread__priv(thread);
1469
1470         ttrace->filename.ptr = ptr;
1471         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1472 }
1473
1474 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1475 {
1476         struct augmented_arg *augmented_arg = arg->augmented.args;
1477         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1478         /*
1479          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1480          * we would have two strings, each prefixed by its size.
1481          */
1482         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1483
1484         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1485         arg->augmented.size -= consumed;
1486
1487         return printed;
1488 }
1489
1490 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1491                                               struct syscall_arg *arg)
1492 {
1493         unsigned long ptr = arg->val;
1494
1495         if (arg->augmented.args)
1496                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1497
1498         if (!arg->trace->vfs_getname)
1499                 return scnprintf(bf, size, "%#x", ptr);
1500
1501         thread__set_filename_pos(arg->thread, bf, ptr);
1502         return 0;
1503 }
1504
1505 static bool trace__filter_duration(struct trace *trace, double t)
1506 {
1507         return t < (trace->duration_filter * NSEC_PER_MSEC);
1508 }
1509
1510 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1511 {
1512         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1513
1514         return fprintf(fp, "%10.3f ", ts);
1515 }
1516
1517 /*
1518  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1519  * using ttrace->entry_time for a thread that receives a sys_exit without
1520  * first having received a sys_enter ("poll" issued before tracing session
1521  * starts, lost sys_enter exit due to ring buffer overflow).
1522  */
1523 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1524 {
1525         if (tstamp > 0)
1526                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1527
1528         return fprintf(fp, "         ? ");
1529 }
1530
1531 static bool done = false;
1532 static bool interrupted = false;
1533
1534 static void sig_handler(int sig)
1535 {
1536         done = true;
1537         interrupted = sig == SIGINT;
1538 }
1539
1540 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1541 {
1542         size_t printed = 0;
1543
1544         if (trace->multiple_threads) {
1545                 if (trace->show_comm)
1546                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1547                 printed += fprintf(fp, "%d ", thread->tid);
1548         }
1549
1550         return printed;
1551 }
1552
1553 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1554                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1555 {
1556         size_t printed = 0;
1557
1558         if (trace->show_tstamp)
1559                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1560         if (trace->show_duration)
1561                 printed += fprintf_duration(duration, duration_calculated, fp);
1562         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1563 }
1564
1565 static int trace__process_event(struct trace *trace, struct machine *machine,
1566                                 union perf_event *event, struct perf_sample *sample)
1567 {
1568         int ret = 0;
1569
1570         switch (event->header.type) {
1571         case PERF_RECORD_LOST:
1572                 color_fprintf(trace->output, PERF_COLOR_RED,
1573                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1574                 ret = machine__process_lost_event(machine, event, sample);
1575                 break;
1576         default:
1577                 ret = machine__process_event(machine, event, sample);
1578                 break;
1579         }
1580
1581         return ret;
1582 }
1583
1584 static int trace__tool_process(struct perf_tool *tool,
1585                                union perf_event *event,
1586                                struct perf_sample *sample,
1587                                struct machine *machine)
1588 {
1589         struct trace *trace = container_of(tool, struct trace, tool);
1590         return trace__process_event(trace, machine, event, sample);
1591 }
1592
1593 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1594 {
1595         struct machine *machine = vmachine;
1596
1597         if (machine->kptr_restrict_warned)
1598                 return NULL;
1599
1600         if (symbol_conf.kptr_restrict) {
1601                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1602                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1603                            "Kernel samples will not be resolved.\n");
1604                 machine->kptr_restrict_warned = true;
1605                 return NULL;
1606         }
1607
1608         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1609 }
1610
1611 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1612 {
1613         int err = symbol__init(NULL);
1614
1615         if (err)
1616                 return err;
1617
1618         trace->host = machine__new_host();
1619         if (trace->host == NULL)
1620                 return -ENOMEM;
1621
1622         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1623         if (err < 0)
1624                 goto out;
1625
1626         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1627                                             evlist->core.threads, trace__tool_process, false,
1628                                             1);
1629 out:
1630         if (err)
1631                 symbol__exit();
1632
1633         return err;
1634 }
1635
1636 static void trace__symbols__exit(struct trace *trace)
1637 {
1638         machine__exit(trace->host);
1639         trace->host = NULL;
1640
1641         symbol__exit();
1642 }
1643
1644 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1645 {
1646         int idx;
1647
1648         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1649                 nr_args = sc->fmt->nr_args;
1650
1651         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1652         if (sc->arg_fmt == NULL)
1653                 return -1;
1654
1655         for (idx = 0; idx < nr_args; ++idx) {
1656                 if (sc->fmt)
1657                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1658         }
1659
1660         sc->nr_args = nr_args;
1661         return 0;
1662 }
1663
1664 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1665         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1666         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1667 };
1668
1669 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1670 {
1671        const struct syscall_arg_fmt *fmt = fmtp;
1672        return strcmp(name, fmt->name);
1673 }
1674
1675 static struct syscall_arg_fmt *
1676 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1677 {
1678        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1679 }
1680
1681 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1682 {
1683        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1684        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1685 }
1686
1687 static struct tep_format_field *
1688 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1689 {
1690         struct tep_format_field *last_field = NULL;
1691         int len;
1692
1693         for (; field; field = field->next, ++arg) {
1694                 last_field = field;
1695
1696                 if (arg->scnprintf)
1697                         continue;
1698
1699                 len = strlen(field->name);
1700
1701                 if (strcmp(field->type, "const char *") == 0 &&
1702                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1703                      strstr(field->name, "path") != NULL))
1704                         arg->scnprintf = SCA_FILENAME;
1705                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1706                         arg->scnprintf = SCA_PTR;
1707                 else if (strcmp(field->type, "pid_t") == 0)
1708                         arg->scnprintf = SCA_PID;
1709                 else if (strcmp(field->type, "umode_t") == 0)
1710                         arg->scnprintf = SCA_MODE_T;
1711                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1712                         arg->scnprintf = SCA_CHAR_ARRAY;
1713                         arg->nr_entries = field->arraylen;
1714                 } else if ((strcmp(field->type, "int") == 0 ||
1715                           strcmp(field->type, "unsigned int") == 0 ||
1716                           strcmp(field->type, "long") == 0) &&
1717                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1718                         /*
1719                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1720                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1721                          * 65 int
1722                          * 23 unsigned int
1723                          * 7 unsigned long
1724                          */
1725                         arg->scnprintf = SCA_FD;
1726                } else {
1727                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1728
1729                         if (fmt) {
1730                                 arg->scnprintf = fmt->scnprintf;
1731                                 arg->strtoul   = fmt->strtoul;
1732                         }
1733                 }
1734         }
1735
1736         return last_field;
1737 }
1738
1739 static int syscall__set_arg_fmts(struct syscall *sc)
1740 {
1741         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1742
1743         if (last_field)
1744                 sc->args_size = last_field->offset + last_field->size;
1745
1746         return 0;
1747 }
1748
1749 static int trace__read_syscall_info(struct trace *trace, int id)
1750 {
1751         char tp_name[128];
1752         struct syscall *sc;
1753         const char *name = syscalltbl__name(trace->sctbl, id);
1754
1755         if (trace->syscalls.table == NULL) {
1756                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1757                 if (trace->syscalls.table == NULL)
1758                         return -ENOMEM;
1759         }
1760
1761         sc = trace->syscalls.table + id;
1762         if (sc->nonexistent)
1763                 return 0;
1764
1765         if (name == NULL) {
1766                 sc->nonexistent = true;
1767                 return 0;
1768         }
1769
1770         sc->name = name;
1771         sc->fmt  = syscall_fmt__find(sc->name);
1772
1773         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1774         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1775
1776         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1777                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1778                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1779         }
1780
1781         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1782                 return -ENOMEM;
1783
1784         if (IS_ERR(sc->tp_format))
1785                 return PTR_ERR(sc->tp_format);
1786
1787         sc->args = sc->tp_format->format.fields;
1788         /*
1789          * We need to check and discard the first variable '__syscall_nr'
1790          * or 'nr' that mean the syscall number. It is needless here.
1791          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1792          */
1793         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1794                 sc->args = sc->args->next;
1795                 --sc->nr_args;
1796         }
1797
1798         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1799         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1800
1801         return syscall__set_arg_fmts(sc);
1802 }
1803
1804 static int perf_evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1805 {
1806         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1807
1808         if (fmt != NULL) {
1809                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1810                 return 0;
1811         }
1812
1813         return -ENOMEM;
1814 }
1815
1816 static int intcmp(const void *a, const void *b)
1817 {
1818         const int *one = a, *another = b;
1819
1820         return *one - *another;
1821 }
1822
1823 static int trace__validate_ev_qualifier(struct trace *trace)
1824 {
1825         int err = 0;
1826         bool printed_invalid_prefix = false;
1827         struct str_node *pos;
1828         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1829
1830         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1831                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1832
1833         if (trace->ev_qualifier_ids.entries == NULL) {
1834                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1835                        trace->output);
1836                 err = -EINVAL;
1837                 goto out;
1838         }
1839
1840         strlist__for_each_entry(pos, trace->ev_qualifier) {
1841                 const char *sc = pos->s;
1842                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1843
1844                 if (id < 0) {
1845                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1846                         if (id >= 0)
1847                                 goto matches;
1848
1849                         if (!printed_invalid_prefix) {
1850                                 pr_debug("Skipping unknown syscalls: ");
1851                                 printed_invalid_prefix = true;
1852                         } else {
1853                                 pr_debug(", ");
1854                         }
1855
1856                         pr_debug("%s", sc);
1857                         continue;
1858                 }
1859 matches:
1860                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1861                 if (match_next == -1)
1862                         continue;
1863
1864                 while (1) {
1865                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1866                         if (id < 0)
1867                                 break;
1868                         if (nr_allocated == nr_used) {
1869                                 void *entries;
1870
1871                                 nr_allocated += 8;
1872                                 entries = realloc(trace->ev_qualifier_ids.entries,
1873                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1874                                 if (entries == NULL) {
1875                                         err = -ENOMEM;
1876                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1877                                         goto out_free;
1878                                 }
1879                                 trace->ev_qualifier_ids.entries = entries;
1880                         }
1881                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1882                 }
1883         }
1884
1885         trace->ev_qualifier_ids.nr = nr_used;
1886         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1887 out:
1888         if (printed_invalid_prefix)
1889                 pr_debug("\n");
1890         return err;
1891 out_free:
1892         zfree(&trace->ev_qualifier_ids.entries);
1893         trace->ev_qualifier_ids.nr = 0;
1894         goto out;
1895 }
1896
1897 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1898 {
1899         bool in_ev_qualifier;
1900
1901         if (trace->ev_qualifier_ids.nr == 0)
1902                 return true;
1903
1904         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1905                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1906
1907         if (in_ev_qualifier)
1908                return !trace->not_ev_qualifier;
1909
1910         return trace->not_ev_qualifier;
1911 }
1912
1913 /*
1914  * args is to be interpreted as a series of longs but we need to handle
1915  * 8-byte unaligned accesses. args points to raw_data within the event
1916  * and raw_data is guaranteed to be 8-byte unaligned because it is
1917  * preceded by raw_size which is a u32. So we need to copy args to a temp
1918  * variable to read it. Most notably this avoids extended load instructions
1919  * on unaligned addresses
1920  */
1921 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1922 {
1923         unsigned long val;
1924         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1925
1926         memcpy(&val, p, sizeof(val));
1927         return val;
1928 }
1929
1930 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1931                                       struct syscall_arg *arg)
1932 {
1933         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1934                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1935
1936         return scnprintf(bf, size, "arg%d: ", arg->idx);
1937 }
1938
1939 /*
1940  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1941  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1942  * in tools/perf/trace/beauty/mount_flags.c
1943  */
1944 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1945 {
1946         if (fmt && fmt->mask_val)
1947                 return fmt->mask_val(arg, val);
1948
1949         return val;
1950 }
1951
1952 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1953                                              struct syscall_arg *arg, unsigned long val)
1954 {
1955         if (fmt && fmt->scnprintf) {
1956                 arg->val = val;
1957                 if (fmt->parm)
1958                         arg->parm = fmt->parm;
1959                 return fmt->scnprintf(bf, size, arg);
1960         }
1961         return scnprintf(bf, size, "%ld", val);
1962 }
1963
1964 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1965                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1966                                       struct trace *trace, struct thread *thread)
1967 {
1968         size_t printed = 0;
1969         unsigned long val;
1970         u8 bit = 1;
1971         struct syscall_arg arg = {
1972                 .args   = args,
1973                 .augmented = {
1974                         .size = augmented_args_size,
1975                         .args = augmented_args,
1976                 },
1977                 .idx    = 0,
1978                 .mask   = 0,
1979                 .trace  = trace,
1980                 .thread = thread,
1981                 .show_string_prefix = trace->show_string_prefix,
1982         };
1983         struct thread_trace *ttrace = thread__priv(thread);
1984
1985         /*
1986          * Things like fcntl will set this in its 'cmd' formatter to pick the
1987          * right formatter for the return value (an fd? file flags?), which is
1988          * not needed for syscalls that always return a given type, say an fd.
1989          */
1990         ttrace->ret_scnprintf = NULL;
1991
1992         if (sc->args != NULL) {
1993                 struct tep_format_field *field;
1994
1995                 for (field = sc->args; field;
1996                      field = field->next, ++arg.idx, bit <<= 1) {
1997                         if (arg.mask & bit)
1998                                 continue;
1999
2000                         arg.fmt = &sc->arg_fmt[arg.idx];
2001                         val = syscall_arg__val(&arg, arg.idx);
2002                         /*
2003                          * Some syscall args need some mask, most don't and
2004                          * return val untouched.
2005                          */
2006                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2007
2008                         /*
2009                          * Suppress this argument if its value is zero and
2010                          * and we don't have a string associated in an
2011                          * strarray for it.
2012                          */
2013                         if (val == 0 &&
2014                             !trace->show_zeros &&
2015                             !(sc->arg_fmt &&
2016                               (sc->arg_fmt[arg.idx].show_zero ||
2017                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2018                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2019                               sc->arg_fmt[arg.idx].parm))
2020                                 continue;
2021
2022                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2023
2024                         if (trace->show_arg_names)
2025                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2026
2027                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2028                                                                   bf + printed, size - printed, &arg, val);
2029                 }
2030         } else if (IS_ERR(sc->tp_format)) {
2031                 /*
2032                  * If we managed to read the tracepoint /format file, then we
2033                  * may end up not having any args, like with gettid(), so only
2034                  * print the raw args when we didn't manage to read it.
2035                  */
2036                 while (arg.idx < sc->nr_args) {
2037                         if (arg.mask & bit)
2038                                 goto next_arg;
2039                         val = syscall_arg__val(&arg, arg.idx);
2040                         if (printed)
2041                                 printed += scnprintf(bf + printed, size - printed, ", ");
2042                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2043                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2044 next_arg:
2045                         ++arg.idx;
2046                         bit <<= 1;
2047                 }
2048         }
2049
2050         return printed;
2051 }
2052
2053 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2054                                   union perf_event *event,
2055                                   struct perf_sample *sample);
2056
2057 static struct syscall *trace__syscall_info(struct trace *trace,
2058                                            struct evsel *evsel, int id)
2059 {
2060         int err = 0;
2061
2062         if (id < 0) {
2063
2064                 /*
2065                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2066                  * before that, leaving at a higher verbosity level till that is
2067                  * explained. Reproduced with plain ftrace with:
2068                  *
2069                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2070                  * grep "NR -1 " /t/trace_pipe
2071                  *
2072                  * After generating some load on the machine.
2073                  */
2074                 if (verbose > 1) {
2075                         static u64 n;
2076                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2077                                 id, perf_evsel__name(evsel), ++n);
2078                 }
2079                 return NULL;
2080         }
2081
2082         err = -EINVAL;
2083
2084         if (id > trace->sctbl->syscalls.max_id)
2085                 goto out_cant_read;
2086
2087         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2088             (err = trace__read_syscall_info(trace, id)) != 0)
2089                 goto out_cant_read;
2090
2091         if (trace->syscalls.table[id].name == NULL) {
2092                 if (trace->syscalls.table[id].nonexistent)
2093                         return NULL;
2094                 goto out_cant_read;
2095         }
2096
2097         return &trace->syscalls.table[id];
2098
2099 out_cant_read:
2100         if (verbose > 0) {
2101                 char sbuf[STRERR_BUFSIZE];
2102                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2103                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2104                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2105                 fputs(" information\n", trace->output);
2106         }
2107         return NULL;
2108 }
2109
2110 struct syscall_stats {
2111         struct stats stats;
2112         u64          nr_failures;
2113         int          max_errno;
2114         u32          *errnos;
2115 };
2116
2117 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2118                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2119 {
2120         struct int_node *inode;
2121         struct syscall_stats *stats;
2122         u64 duration = 0;
2123
2124         inode = intlist__findnew(ttrace->syscall_stats, id);
2125         if (inode == NULL)
2126                 return;
2127
2128         stats = inode->priv;
2129         if (stats == NULL) {
2130                 stats = malloc(sizeof(*stats));
2131                 if (stats == NULL)
2132                         return;
2133
2134                 stats->nr_failures = 0;
2135                 stats->max_errno   = 0;
2136                 stats->errnos      = NULL;
2137                 init_stats(&stats->stats);
2138                 inode->priv = stats;
2139         }
2140
2141         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2142                 duration = sample->time - ttrace->entry_time;
2143
2144         update_stats(&stats->stats, duration);
2145
2146         if (err < 0) {
2147                 ++stats->nr_failures;
2148
2149                 if (!errno_summary)
2150                         return;
2151
2152                 err = -err;
2153                 if (err > stats->max_errno) {
2154                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2155
2156                         if (new_errnos) {
2157                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2158                         } else {
2159                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2160                                          thread__comm_str(thread), thread->pid_, thread->tid);
2161                                 return;
2162                         }
2163
2164                         stats->errnos = new_errnos;
2165                         stats->max_errno = err;
2166                 }
2167
2168                 ++stats->errnos[err - 1];
2169         }
2170 }
2171
2172 static int trace__printf_interrupted_entry(struct trace *trace)
2173 {
2174         struct thread_trace *ttrace;
2175         size_t printed;
2176         int len;
2177
2178         if (trace->failure_only || trace->current == NULL)
2179                 return 0;
2180
2181         ttrace = thread__priv(trace->current);
2182
2183         if (!ttrace->entry_pending)
2184                 return 0;
2185
2186         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2187         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2188
2189         if (len < trace->args_alignment - 4)
2190                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2191
2192         printed += fprintf(trace->output, " ...\n");
2193
2194         ttrace->entry_pending = false;
2195         ++trace->nr_events_printed;
2196
2197         return printed;
2198 }
2199
2200 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2201                                  struct perf_sample *sample, struct thread *thread)
2202 {
2203         int printed = 0;
2204
2205         if (trace->print_sample) {
2206                 double ts = (double)sample->time / NSEC_PER_MSEC;
2207
2208                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2209                                    perf_evsel__name(evsel), ts,
2210                                    thread__comm_str(thread),
2211                                    sample->pid, sample->tid, sample->cpu);
2212         }
2213
2214         return printed;
2215 }
2216
2217 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2218 {
2219         void *augmented_args = NULL;
2220         /*
2221          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2222          * and there we get all 6 syscall args plus the tracepoint common fields
2223          * that gets calculated at the start and the syscall_nr (another long).
2224          * So we check if that is the case and if so don't look after the
2225          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2226          * which is fixed.
2227          *
2228          * We'll revisit this later to pass s->args_size to the BPF augmenter
2229          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2230          * copies only what we need for each syscall, like what happens when we
2231          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2232          * traffic to just what is needed for each syscall.
2233          */
2234         int args_size = raw_augmented_args_size ?: sc->args_size;
2235
2236         *augmented_args_size = sample->raw_size - args_size;
2237         if (*augmented_args_size > 0)
2238                 augmented_args = sample->raw_data + args_size;
2239
2240         return augmented_args;
2241 }
2242
2243 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2244                             union perf_event *event __maybe_unused,
2245                             struct perf_sample *sample)
2246 {
2247         char *msg;
2248         void *args;
2249         int printed = 0;
2250         struct thread *thread;
2251         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2252         int augmented_args_size = 0;
2253         void *augmented_args = NULL;
2254         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2255         struct thread_trace *ttrace;
2256
2257         if (sc == NULL)
2258                 return -1;
2259
2260         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2261         ttrace = thread__trace(thread, trace->output);
2262         if (ttrace == NULL)
2263                 goto out_put;
2264
2265         trace__fprintf_sample(trace, evsel, sample, thread);
2266
2267         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2268
2269         if (ttrace->entry_str == NULL) {
2270                 ttrace->entry_str = malloc(trace__entry_str_size);
2271                 if (!ttrace->entry_str)
2272                         goto out_put;
2273         }
2274
2275         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2276                 trace__printf_interrupted_entry(trace);
2277         /*
2278          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2279          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2280          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2281          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2282          * so when handling, say the openat syscall, we end up getting 6 args for the
2283          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2284          * thinking that the extra 2 u64 args are the augmented filename, so just check
2285          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2286          */
2287         if (evsel != trace->syscalls.events.sys_enter)
2288                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2289         ttrace->entry_time = sample->time;
2290         msg = ttrace->entry_str;
2291         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2292
2293         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2294                                            args, augmented_args, augmented_args_size, trace, thread);
2295
2296         if (sc->is_exit) {
2297                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2298                         int alignment = 0;
2299
2300                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2301                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2302                         if (trace->args_alignment > printed)
2303                                 alignment = trace->args_alignment - printed;
2304                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2305                 }
2306         } else {
2307                 ttrace->entry_pending = true;
2308                 /* See trace__vfs_getname & trace__sys_exit */
2309                 ttrace->filename.pending_open = false;
2310         }
2311
2312         if (trace->current != thread) {
2313                 thread__put(trace->current);
2314                 trace->current = thread__get(thread);
2315         }
2316         err = 0;
2317 out_put:
2318         thread__put(thread);
2319         return err;
2320 }
2321
2322 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2323                                     struct perf_sample *sample)
2324 {
2325         struct thread_trace *ttrace;
2326         struct thread *thread;
2327         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2328         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2329         char msg[1024];
2330         void *args, *augmented_args = NULL;
2331         int augmented_args_size;
2332
2333         if (sc == NULL)
2334                 return -1;
2335
2336         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2337         ttrace = thread__trace(thread, trace->output);
2338         /*
2339          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2340          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2341          */
2342         if (ttrace == NULL)
2343                 goto out_put;
2344
2345         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2346         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2347         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2348         fprintf(trace->output, "%s", msg);
2349         err = 0;
2350 out_put:
2351         thread__put(thread);
2352         return err;
2353 }
2354
2355 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2356                                     struct perf_sample *sample,
2357                                     struct callchain_cursor *cursor)
2358 {
2359         struct addr_location al;
2360         int max_stack = evsel->core.attr.sample_max_stack ?
2361                         evsel->core.attr.sample_max_stack :
2362                         trace->max_stack;
2363         int err;
2364
2365         if (machine__resolve(trace->host, &al, sample) < 0)
2366                 return -1;
2367
2368         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2369         addr_location__put(&al);
2370         return err;
2371 }
2372
2373 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2374 {
2375         /* TODO: user-configurable print_opts */
2376         const unsigned int print_opts = EVSEL__PRINT_SYM |
2377                                         EVSEL__PRINT_DSO |
2378                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2379
2380         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2381 }
2382
2383 static const char *errno_to_name(struct evsel *evsel, int err)
2384 {
2385         struct perf_env *env = perf_evsel__env(evsel);
2386         const char *arch_name = perf_env__arch(env);
2387
2388         return arch_syscalls__strerrno(arch_name, err);
2389 }
2390
2391 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2392                            union perf_event *event __maybe_unused,
2393                            struct perf_sample *sample)
2394 {
2395         long ret;
2396         u64 duration = 0;
2397         bool duration_calculated = false;
2398         struct thread *thread;
2399         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2400         int alignment = trace->args_alignment;
2401         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2402         struct thread_trace *ttrace;
2403
2404         if (sc == NULL)
2405                 return -1;
2406
2407         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2408         ttrace = thread__trace(thread, trace->output);
2409         if (ttrace == NULL)
2410                 goto out_put;
2411
2412         trace__fprintf_sample(trace, evsel, sample, thread);
2413
2414         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2415
2416         if (trace->summary)
2417                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2418
2419         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2420                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2421                 ttrace->filename.pending_open = false;
2422                 ++trace->stats.vfs_getname;
2423         }
2424
2425         if (ttrace->entry_time) {
2426                 duration = sample->time - ttrace->entry_time;
2427                 if (trace__filter_duration(trace, duration))
2428                         goto out;
2429                 duration_calculated = true;
2430         } else if (trace->duration_filter)
2431                 goto out;
2432
2433         if (sample->callchain) {
2434                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2435                 if (callchain_ret == 0) {
2436                         if (callchain_cursor.nr < trace->min_stack)
2437                                 goto out;
2438                         callchain_ret = 1;
2439                 }
2440         }
2441
2442         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2443                 goto out;
2444
2445         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2446
2447         if (ttrace->entry_pending) {
2448                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2449         } else {
2450                 printed += fprintf(trace->output, " ... [");
2451                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2452                 printed += 9;
2453                 printed += fprintf(trace->output, "]: %s()", sc->name);
2454         }
2455
2456         printed++; /* the closing ')' */
2457
2458         if (alignment > printed)
2459                 alignment -= printed;
2460         else
2461                 alignment = 0;
2462
2463         fprintf(trace->output, ")%*s= ", alignment, " ");
2464
2465         if (sc->fmt == NULL) {
2466                 if (ret < 0)
2467                         goto errno_print;
2468 signed_print:
2469                 fprintf(trace->output, "%ld", ret);
2470         } else if (ret < 0) {
2471 errno_print: {
2472                 char bf[STRERR_BUFSIZE];
2473                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2474                            *e = errno_to_name(evsel, -ret);
2475
2476                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2477         }
2478         } else if (ret == 0 && sc->fmt->timeout)
2479                 fprintf(trace->output, "0 (Timeout)");
2480         else if (ttrace->ret_scnprintf) {
2481                 char bf[1024];
2482                 struct syscall_arg arg = {
2483                         .val    = ret,
2484                         .thread = thread,
2485                         .trace  = trace,
2486                 };
2487                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2488                 ttrace->ret_scnprintf = NULL;
2489                 fprintf(trace->output, "%s", bf);
2490         } else if (sc->fmt->hexret)
2491                 fprintf(trace->output, "%#lx", ret);
2492         else if (sc->fmt->errpid) {
2493                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2494
2495                 if (child != NULL) {
2496                         fprintf(trace->output, "%ld", ret);
2497                         if (child->comm_set)
2498                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2499                         thread__put(child);
2500                 }
2501         } else
2502                 goto signed_print;
2503
2504         fputc('\n', trace->output);
2505
2506         /*
2507          * We only consider an 'event' for the sake of --max-events a non-filtered
2508          * sys_enter + sys_exit and other tracepoint events.
2509          */
2510         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2511                 interrupted = true;
2512
2513         if (callchain_ret > 0)
2514                 trace__fprintf_callchain(trace, sample);
2515         else if (callchain_ret < 0)
2516                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2517 out:
2518         ttrace->entry_pending = false;
2519         err = 0;
2520 out_put:
2521         thread__put(thread);
2522         return err;
2523 }
2524
2525 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2526                               union perf_event *event __maybe_unused,
2527                               struct perf_sample *sample)
2528 {
2529         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2530         struct thread_trace *ttrace;
2531         size_t filename_len, entry_str_len, to_move;
2532         ssize_t remaining_space;
2533         char *pos;
2534         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2535
2536         if (!thread)
2537                 goto out;
2538
2539         ttrace = thread__priv(thread);
2540         if (!ttrace)
2541                 goto out_put;
2542
2543         filename_len = strlen(filename);
2544         if (filename_len == 0)
2545                 goto out_put;
2546
2547         if (ttrace->filename.namelen < filename_len) {
2548                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2549
2550                 if (f == NULL)
2551                         goto out_put;
2552
2553                 ttrace->filename.namelen = filename_len;
2554                 ttrace->filename.name = f;
2555         }
2556
2557         strcpy(ttrace->filename.name, filename);
2558         ttrace->filename.pending_open = true;
2559
2560         if (!ttrace->filename.ptr)
2561                 goto out_put;
2562
2563         entry_str_len = strlen(ttrace->entry_str);
2564         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2565         if (remaining_space <= 0)
2566                 goto out_put;
2567
2568         if (filename_len > (size_t)remaining_space) {
2569                 filename += filename_len - remaining_space;
2570                 filename_len = remaining_space;
2571         }
2572
2573         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2574         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2575         memmove(pos + filename_len, pos, to_move);
2576         memcpy(pos, filename, filename_len);
2577
2578         ttrace->filename.ptr = 0;
2579         ttrace->filename.entry_str_pos = 0;
2580 out_put:
2581         thread__put(thread);
2582 out:
2583         return 0;
2584 }
2585
2586 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2587                                      union perf_event *event __maybe_unused,
2588                                      struct perf_sample *sample)
2589 {
2590         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2591         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2592         struct thread *thread = machine__findnew_thread(trace->host,
2593                                                         sample->pid,
2594                                                         sample->tid);
2595         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2596
2597         if (ttrace == NULL)
2598                 goto out_dump;
2599
2600         ttrace->runtime_ms += runtime_ms;
2601         trace->runtime_ms += runtime_ms;
2602 out_put:
2603         thread__put(thread);
2604         return 0;
2605
2606 out_dump:
2607         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2608                evsel->name,
2609                perf_evsel__strval(evsel, sample, "comm"),
2610                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2611                runtime,
2612                perf_evsel__intval(evsel, sample, "vruntime"));
2613         goto out_put;
2614 }
2615
2616 static int bpf_output__printer(enum binary_printer_ops op,
2617                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2618 {
2619         unsigned char ch = (unsigned char)val;
2620
2621         switch (op) {
2622         case BINARY_PRINT_CHAR_DATA:
2623                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2624         case BINARY_PRINT_DATA_BEGIN:
2625         case BINARY_PRINT_LINE_BEGIN:
2626         case BINARY_PRINT_ADDR:
2627         case BINARY_PRINT_NUM_DATA:
2628         case BINARY_PRINT_NUM_PAD:
2629         case BINARY_PRINT_SEP:
2630         case BINARY_PRINT_CHAR_PAD:
2631         case BINARY_PRINT_LINE_END:
2632         case BINARY_PRINT_DATA_END:
2633         default:
2634                 break;
2635         }
2636
2637         return 0;
2638 }
2639
2640 static void bpf_output__fprintf(struct trace *trace,
2641                                 struct perf_sample *sample)
2642 {
2643         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2644                         bpf_output__printer, NULL, trace->output);
2645         ++trace->nr_events_printed;
2646 }
2647
2648 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2649                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2650 {
2651         char bf[2048];
2652         size_t size = sizeof(bf);
2653         struct tep_format_field *field = evsel->tp_format->format.fields;
2654         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2655         size_t printed = 0;
2656         unsigned long val;
2657         u8 bit = 1;
2658         struct syscall_arg syscall_arg = {
2659                 .augmented = {
2660                         .size = augmented_args_size,
2661                         .args = augmented_args,
2662                 },
2663                 .idx    = 0,
2664                 .mask   = 0,
2665                 .trace  = trace,
2666                 .thread = thread,
2667                 .show_string_prefix = trace->show_string_prefix,
2668         };
2669
2670         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2671                 if (syscall_arg.mask & bit)
2672                         continue;
2673
2674                 syscall_arg.len = 0;
2675                 syscall_arg.fmt = arg;
2676                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2677                         int offset = field->offset;
2678
2679                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2680                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2681                                 syscall_arg.len = offset >> 16;
2682                                 offset &= 0xffff;
2683                         }
2684
2685                         val = (uintptr_t)(sample->raw_data + offset);
2686                 } else
2687                         val = format_field__intval(field, sample, evsel->needs_swap);
2688                 /*
2689                  * Some syscall args need some mask, most don't and
2690                  * return val untouched.
2691                  */
2692                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2693
2694                 /*
2695                  * Suppress this argument if its value is zero and
2696                  * and we don't have a string associated in an
2697                  * strarray for it.
2698                  */
2699                 if (val == 0 &&
2700                     !trace->show_zeros &&
2701                     !((arg->show_zero ||
2702                        arg->scnprintf == SCA_STRARRAY ||
2703                        arg->scnprintf == SCA_STRARRAYS) &&
2704                       arg->parm))
2705                         continue;
2706
2707                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2708
2709                 /*
2710                  * XXX Perhaps we should have a show_tp_arg_names,
2711                  * leaving show_arg_names just for syscalls?
2712                  */
2713                 if (1 || trace->show_arg_names)
2714                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2715
2716                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2717         }
2718
2719         return printed + fprintf(trace->output, "%s", bf);
2720 }
2721
2722 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2723                                 union perf_event *event __maybe_unused,
2724                                 struct perf_sample *sample)
2725 {
2726         struct thread *thread;
2727         int callchain_ret = 0;
2728         /*
2729          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2730          * this event's max_events having been hit and this is an entry coming
2731          * from the ring buffer that we should discard, since the max events
2732          * have already been considered/printed.
2733          */
2734         if (evsel->disabled)
2735                 return 0;
2736
2737         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2738
2739         if (sample->callchain) {
2740                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2741                 if (callchain_ret == 0) {
2742                         if (callchain_cursor.nr < trace->min_stack)
2743                                 goto out;
2744                         callchain_ret = 1;
2745                 }
2746         }
2747
2748         trace__printf_interrupted_entry(trace);
2749         trace__fprintf_tstamp(trace, sample->time, trace->output);
2750
2751         if (trace->trace_syscalls && trace->show_duration)
2752                 fprintf(trace->output, "(         ): ");
2753
2754         if (thread)
2755                 trace__fprintf_comm_tid(trace, thread, trace->output);
2756
2757         if (evsel == trace->syscalls.events.augmented) {
2758                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2759                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2760
2761                 if (sc) {
2762                         fprintf(trace->output, "%s(", sc->name);
2763                         trace__fprintf_sys_enter(trace, evsel, sample);
2764                         fputc(')', trace->output);
2765                         goto newline;
2766                 }
2767
2768                 /*
2769                  * XXX: Not having the associated syscall info or not finding/adding
2770                  *      the thread should never happen, but if it does...
2771                  *      fall thru and print it as a bpf_output event.
2772                  */
2773         }
2774
2775         fprintf(trace->output, "%s(", evsel->name);
2776
2777         if (perf_evsel__is_bpf_output(evsel)) {
2778                 bpf_output__fprintf(trace, sample);
2779         } else if (evsel->tp_format) {
2780                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2781                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2782                         if (trace->libtraceevent_print) {
2783                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2784                                                       sample->raw_data, sample->raw_size,
2785                                                       trace->output);
2786                         } else {
2787                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2788                         }
2789                 }
2790         }
2791
2792 newline:
2793         fprintf(trace->output, ")\n");
2794
2795         if (callchain_ret > 0)
2796                 trace__fprintf_callchain(trace, sample);
2797         else if (callchain_ret < 0)
2798                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2799
2800         ++trace->nr_events_printed;
2801
2802         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2803                 evsel__disable(evsel);
2804                 evsel__close(evsel);
2805         }
2806 out:
2807         thread__put(thread);
2808         return 0;
2809 }
2810
2811 static void print_location(FILE *f, struct perf_sample *sample,
2812                            struct addr_location *al,
2813                            bool print_dso, bool print_sym)
2814 {
2815
2816         if ((verbose > 0 || print_dso) && al->map)
2817                 fprintf(f, "%s@", al->map->dso->long_name);
2818
2819         if ((verbose > 0 || print_sym) && al->sym)
2820                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2821                         al->addr - al->sym->start);
2822         else if (al->map)
2823                 fprintf(f, "0x%" PRIx64, al->addr);
2824         else
2825                 fprintf(f, "0x%" PRIx64, sample->addr);
2826 }
2827
2828 static int trace__pgfault(struct trace *trace,
2829                           struct evsel *evsel,
2830                           union perf_event *event __maybe_unused,
2831                           struct perf_sample *sample)
2832 {
2833         struct thread *thread;
2834         struct addr_location al;
2835         char map_type = 'd';
2836         struct thread_trace *ttrace;
2837         int err = -1;
2838         int callchain_ret = 0;
2839
2840         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2841
2842         if (sample->callchain) {
2843                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2844                 if (callchain_ret == 0) {
2845                         if (callchain_cursor.nr < trace->min_stack)
2846                                 goto out_put;
2847                         callchain_ret = 1;
2848                 }
2849         }
2850
2851         ttrace = thread__trace(thread, trace->output);
2852         if (ttrace == NULL)
2853                 goto out_put;
2854
2855         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2856                 ttrace->pfmaj++;
2857         else
2858                 ttrace->pfmin++;
2859
2860         if (trace->summary_only)
2861                 goto out;
2862
2863         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2864
2865         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2866
2867         fprintf(trace->output, "%sfault [",
2868                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2869                 "maj" : "min");
2870
2871         print_location(trace->output, sample, &al, false, true);
2872
2873         fprintf(trace->output, "] => ");
2874
2875         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2876
2877         if (!al.map) {
2878                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2879
2880                 if (al.map)
2881                         map_type = 'x';
2882                 else
2883                         map_type = '?';
2884         }
2885
2886         print_location(trace->output, sample, &al, true, false);
2887
2888         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2889
2890         if (callchain_ret > 0)
2891                 trace__fprintf_callchain(trace, sample);
2892         else if (callchain_ret < 0)
2893                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2894
2895         ++trace->nr_events_printed;
2896 out:
2897         err = 0;
2898 out_put:
2899         thread__put(thread);
2900         return err;
2901 }
2902
2903 static void trace__set_base_time(struct trace *trace,
2904                                  struct evsel *evsel,
2905                                  struct perf_sample *sample)
2906 {
2907         /*
2908          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2909          * and don't use sample->time unconditionally, we may end up having
2910          * some other event in the future without PERF_SAMPLE_TIME for good
2911          * reason, i.e. we may not be interested in its timestamps, just in
2912          * it taking place, picking some piece of information when it
2913          * appears in our event stream (vfs_getname comes to mind).
2914          */
2915         if (trace->base_time == 0 && !trace->full_time &&
2916             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2917                 trace->base_time = sample->time;
2918 }
2919
2920 static int trace__process_sample(struct perf_tool *tool,
2921                                  union perf_event *event,
2922                                  struct perf_sample *sample,
2923                                  struct evsel *evsel,
2924                                  struct machine *machine __maybe_unused)
2925 {
2926         struct trace *trace = container_of(tool, struct trace, tool);
2927         struct thread *thread;
2928         int err = 0;
2929
2930         tracepoint_handler handler = evsel->handler;
2931
2932         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2933         if (thread && thread__is_filtered(thread))
2934                 goto out;
2935
2936         trace__set_base_time(trace, evsel, sample);
2937
2938         if (handler) {
2939                 ++trace->nr_events;
2940                 handler(trace, evsel, event, sample);
2941         }
2942 out:
2943         thread__put(thread);
2944         return err;
2945 }
2946
2947 static int trace__record(struct trace *trace, int argc, const char **argv)
2948 {
2949         unsigned int rec_argc, i, j;
2950         const char **rec_argv;
2951         const char * const record_args[] = {
2952                 "record",
2953                 "-R",
2954                 "-m", "1024",
2955                 "-c", "1",
2956         };
2957         pid_t pid = getpid();
2958         char *filter = asprintf__tp_filter_pids(1, &pid);
2959         const char * const sc_args[] = { "-e", };
2960         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2961         const char * const majpf_args[] = { "-e", "major-faults" };
2962         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2963         const char * const minpf_args[] = { "-e", "minor-faults" };
2964         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2965         int err = -1;
2966
2967         /* +3 is for the event string below and the pid filter */
2968         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
2969                 majpf_args_nr + minpf_args_nr + argc;
2970         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2971
2972         if (rec_argv == NULL || filter == NULL)
2973                 goto out_free;
2974
2975         j = 0;
2976         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2977                 rec_argv[j++] = record_args[i];
2978
2979         if (trace->trace_syscalls) {
2980                 for (i = 0; i < sc_args_nr; i++)
2981                         rec_argv[j++] = sc_args[i];
2982
2983                 /* event string may be different for older kernels - e.g., RHEL6 */
2984                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2985                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2986                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2987                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2988                 else {
2989                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2990                         goto out_free;
2991                 }
2992         }
2993
2994         rec_argv[j++] = "--filter";
2995         rec_argv[j++] = filter;
2996
2997         if (trace->trace_pgfaults & TRACE_PFMAJ)
2998                 for (i = 0; i < majpf_args_nr; i++)
2999                         rec_argv[j++] = majpf_args[i];
3000
3001         if (trace->trace_pgfaults & TRACE_PFMIN)
3002                 for (i = 0; i < minpf_args_nr; i++)
3003                         rec_argv[j++] = minpf_args[i];
3004
3005         for (i = 0; i < (unsigned int)argc; i++)
3006                 rec_argv[j++] = argv[i];
3007
3008         err = cmd_record(j, rec_argv);
3009 out_free:
3010         free(filter);
3011         free(rec_argv);
3012         return err;
3013 }
3014
3015 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3016
3017 static bool evlist__add_vfs_getname(struct evlist *evlist)
3018 {
3019         bool found = false;
3020         struct evsel *evsel, *tmp;
3021         struct parse_events_error err;
3022         int ret;
3023
3024         bzero(&err, sizeof(err));
3025         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3026         if (ret) {
3027                 free(err.str);
3028                 free(err.help);
3029                 free(err.first_str);
3030                 free(err.first_help);
3031                 return false;
3032         }
3033
3034         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3035                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
3036                         continue;
3037
3038                 if (perf_evsel__field(evsel, "pathname")) {
3039                         evsel->handler = trace__vfs_getname;
3040                         found = true;
3041                         continue;
3042                 }
3043
3044                 list_del_init(&evsel->core.node);
3045                 evsel->evlist = NULL;
3046                 evsel__delete(evsel);
3047         }
3048
3049         return found;
3050 }
3051
3052 static struct evsel *perf_evsel__new_pgfault(u64 config)
3053 {
3054         struct evsel *evsel;
3055         struct perf_event_attr attr = {
3056                 .type = PERF_TYPE_SOFTWARE,
3057                 .mmap_data = 1,
3058         };
3059
3060         attr.config = config;
3061         attr.sample_period = 1;
3062
3063         event_attr_init(&attr);
3064
3065         evsel = evsel__new(&attr);
3066         if (evsel)
3067                 evsel->handler = trace__pgfault;
3068
3069         return evsel;
3070 }
3071
3072 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3073 {
3074         const u32 type = event->header.type;
3075         struct evsel *evsel;
3076
3077         if (type != PERF_RECORD_SAMPLE) {
3078                 trace__process_event(trace, trace->host, event, sample);
3079                 return;
3080         }
3081
3082         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
3083         if (evsel == NULL) {
3084                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3085                 return;
3086         }
3087
3088         if (evswitch__discard(&trace->evswitch, evsel))
3089                 return;
3090
3091         trace__set_base_time(trace, evsel, sample);
3092
3093         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3094             sample->raw_data == NULL) {
3095                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3096                        perf_evsel__name(evsel), sample->tid,
3097                        sample->cpu, sample->raw_size);
3098         } else {
3099                 tracepoint_handler handler = evsel->handler;
3100                 handler(trace, evsel, event, sample);
3101         }
3102
3103         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3104                 interrupted = true;
3105 }
3106
3107 static int trace__add_syscall_newtp(struct trace *trace)
3108 {
3109         int ret = -1;
3110         struct evlist *evlist = trace->evlist;
3111         struct evsel *sys_enter, *sys_exit;
3112
3113         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3114         if (sys_enter == NULL)
3115                 goto out;
3116
3117         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3118                 goto out_delete_sys_enter;
3119
3120         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3121         if (sys_exit == NULL)
3122                 goto out_delete_sys_enter;
3123
3124         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3125                 goto out_delete_sys_exit;
3126
3127         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3128         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3129
3130         evlist__add(evlist, sys_enter);
3131         evlist__add(evlist, sys_exit);
3132
3133         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3134                 /*
3135                  * We're interested only in the user space callchain
3136                  * leading to the syscall, allow overriding that for
3137                  * debugging reasons using --kernel_syscall_callchains
3138                  */
3139                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3140         }
3141
3142         trace->syscalls.events.sys_enter = sys_enter;
3143         trace->syscalls.events.sys_exit  = sys_exit;
3144
3145         ret = 0;
3146 out:
3147         return ret;
3148
3149 out_delete_sys_exit:
3150         evsel__delete_priv(sys_exit);
3151 out_delete_sys_enter:
3152         evsel__delete_priv(sys_enter);
3153         goto out;
3154 }
3155
3156 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3157 {
3158         int err = -1;
3159         struct evsel *sys_exit;
3160         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3161                                                 trace->ev_qualifier_ids.nr,
3162                                                 trace->ev_qualifier_ids.entries);
3163
3164         if (filter == NULL)
3165                 goto out_enomem;
3166
3167         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
3168                                           filter)) {
3169                 sys_exit = trace->syscalls.events.sys_exit;
3170                 err = perf_evsel__append_tp_filter(sys_exit, filter);
3171         }
3172
3173         free(filter);
3174 out:
3175         return err;
3176 out_enomem:
3177         errno = ENOMEM;
3178         goto out;
3179 }
3180
3181 #ifdef HAVE_LIBBPF_SUPPORT
3182 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3183 {
3184         if (trace->bpf_obj == NULL)
3185                 return NULL;
3186
3187         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3188 }
3189
3190 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3191                                                         const char *prog_name, const char *type)
3192 {
3193         struct bpf_program *prog;
3194
3195         if (prog_name == NULL) {
3196                 char default_prog_name[256];
3197                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3198                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3199                 if (prog != NULL)
3200                         goto out_found;
3201                 if (sc->fmt && sc->fmt->alias) {
3202                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3203                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3204                         if (prog != NULL)
3205                                 goto out_found;
3206                 }
3207                 goto out_unaugmented;
3208         }
3209
3210         prog = trace__find_bpf_program_by_title(trace, prog_name);
3211
3212         if (prog != NULL) {
3213 out_found:
3214                 return prog;
3215         }
3216
3217         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3218                  prog_name, type, sc->name);
3219 out_unaugmented:
3220         return trace->syscalls.unaugmented_prog;
3221 }
3222
3223 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3224 {
3225         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3226
3227         if (sc == NULL)
3228                 return;
3229
3230         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3231         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3232 }
3233
3234 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3235 {
3236         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3237         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3238 }
3239
3240 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3241 {
3242         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3243         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3244 }
3245
3246 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3247 {
3248         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3249         int arg = 0;
3250
3251         if (sc == NULL)
3252                 goto out;
3253
3254         for (; arg < sc->nr_args; ++arg) {
3255                 entry->string_args_len[arg] = 0;
3256                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3257                         /* Should be set like strace -s strsize */
3258                         entry->string_args_len[arg] = PATH_MAX;
3259                 }
3260         }
3261 out:
3262         for (; arg < 6; ++arg)
3263                 entry->string_args_len[arg] = 0;
3264 }
3265 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3266 {
3267         int fd = bpf_map__fd(trace->syscalls.map);
3268         struct bpf_map_syscall_entry value = {
3269                 .enabled = !trace->not_ev_qualifier,
3270         };
3271         int err = 0;
3272         size_t i;
3273
3274         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3275                 int key = trace->ev_qualifier_ids.entries[i];
3276
3277                 if (value.enabled) {
3278                         trace__init_bpf_map_syscall_args(trace, key, &value);
3279                         trace__init_syscall_bpf_progs(trace, key);
3280                 }
3281
3282                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3283                 if (err)
3284                         break;
3285         }
3286
3287         return err;
3288 }
3289
3290 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3291 {
3292         int fd = bpf_map__fd(trace->syscalls.map);
3293         struct bpf_map_syscall_entry value = {
3294                 .enabled = enabled,
3295         };
3296         int err = 0, key;
3297
3298         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3299                 if (enabled)
3300                         trace__init_bpf_map_syscall_args(trace, key, &value);
3301
3302                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3303                 if (err)
3304                         break;
3305         }
3306
3307         return err;
3308 }
3309
3310 static int trace__init_syscalls_bpf_map(struct trace *trace)
3311 {
3312         bool enabled = true;
3313
3314         if (trace->ev_qualifier_ids.nr)
3315                 enabled = trace->not_ev_qualifier;
3316
3317         return __trace__init_syscalls_bpf_map(trace, enabled);
3318 }
3319
3320 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3321 {
3322         struct tep_format_field *field, *candidate_field;
3323         int id;
3324
3325         /*
3326          * We're only interested in syscalls that have a pointer:
3327          */
3328         for (field = sc->args; field; field = field->next) {
3329                 if (field->flags & TEP_FIELD_IS_POINTER)
3330                         goto try_to_find_pair;
3331         }
3332
3333         return NULL;
3334
3335 try_to_find_pair:
3336         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3337                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3338                 struct bpf_program *pair_prog;
3339                 bool is_candidate = false;
3340
3341                 if (pair == NULL || pair == sc ||
3342                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3343                         continue;
3344
3345                 for (field = sc->args, candidate_field = pair->args;
3346                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3347                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3348                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3349
3350                         if (is_pointer) {
3351                                if (!candidate_is_pointer) {
3352                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3353                                         continue;
3354                                }
3355                         } else {
3356                                 if (candidate_is_pointer) {
3357                                         // The candidate might copy a pointer we don't have, skip it.
3358                                         goto next_candidate;
3359                                 }
3360                                 continue;
3361                         }
3362
3363                         if (strcmp(field->type, candidate_field->type))
3364                                 goto next_candidate;
3365
3366                         is_candidate = true;
3367                 }
3368
3369                 if (!is_candidate)
3370                         goto next_candidate;
3371
3372                 /*
3373                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3374                  * then it may be collecting that and we then can't use it, as it would collect
3375                  * more than what is common to the two syscalls.
3376                  */
3377                 if (candidate_field) {
3378                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3379                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3380                                         goto next_candidate;
3381                 }
3382
3383                 pair_prog = pair->bpf_prog.sys_enter;
3384                 /*
3385                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3386                  * have been searched for, so search it here and if it returns the
3387                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3388                  * program for a filtered syscall on a non-filtered one.
3389                  *
3390                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3391                  * useful for "renameat2".
3392                  */
3393                 if (pair_prog == NULL) {
3394                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3395                         if (pair_prog == trace->syscalls.unaugmented_prog)
3396                                 goto next_candidate;
3397                 }
3398
3399                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3400                 return pair_prog;
3401         next_candidate:
3402                 continue;
3403         }
3404
3405         return NULL;
3406 }
3407
3408 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3409 {
3410         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3411             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3412         int err = 0, key;
3413
3414         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3415                 int prog_fd;
3416
3417                 if (!trace__syscall_enabled(trace, key))
3418                         continue;
3419
3420                 trace__init_syscall_bpf_progs(trace, key);
3421
3422                 // It'll get at least the "!raw_syscalls:unaugmented"
3423                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3424                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3425                 if (err)
3426                         break;
3427                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3428                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3429                 if (err)
3430                         break;
3431         }
3432
3433         /*
3434          * Now lets do a second pass looking for enabled syscalls without
3435          * an augmenter that have a signature that is a superset of another
3436          * syscall with an augmenter so that we can auto-reuse it.
3437          *
3438          * I.e. if we have an augmenter for the "open" syscall that has
3439          * this signature:
3440          *
3441          *   int open(const char *pathname, int flags, mode_t mode);
3442          *
3443          * I.e. that will collect just the first string argument, then we
3444          * can reuse it for the 'creat' syscall, that has this signature:
3445          *
3446          *   int creat(const char *pathname, mode_t mode);
3447          *
3448          * and for:
3449          *
3450          *   int stat(const char *pathname, struct stat *statbuf);
3451          *   int lstat(const char *pathname, struct stat *statbuf);
3452          *
3453          * Because the 'open' augmenter will collect the first arg as a string,
3454          * and leave alone all the other args, which already helps with
3455          * beautifying 'stat' and 'lstat''s pathname arg.
3456          *
3457          * Then, in time, when 'stat' gets an augmenter that collects both
3458          * first and second arg (this one on the raw_syscalls:sys_exit prog
3459          * array tail call, then that one will be used.
3460          */
3461         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3462                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3463                 struct bpf_program *pair_prog;
3464                 int prog_fd;
3465
3466                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3467                         continue;
3468
3469                 /*
3470                  * For now we're just reusing the sys_enter prog, and if it
3471                  * already has an augmenter, we don't need to find one.
3472                  */
3473                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3474                         continue;
3475
3476                 /*
3477                  * Look at all the other syscalls for one that has a signature
3478                  * that is close enough that we can share:
3479                  */
3480                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3481                 if (pair_prog == NULL)
3482                         continue;
3483
3484                 sc->bpf_prog.sys_enter = pair_prog;
3485
3486                 /*
3487                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3488                  * with the fd for the program we're reusing:
3489                  */
3490                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3491                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3492                 if (err)
3493                         break;
3494         }
3495
3496
3497         return err;
3498 }
3499
3500 static void trace__delete_augmented_syscalls(struct trace *trace)
3501 {
3502         struct evsel *evsel, *tmp;
3503
3504         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3505         evsel__delete(trace->syscalls.events.augmented);
3506         trace->syscalls.events.augmented = NULL;
3507
3508         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3509                 if (evsel->bpf_obj == trace->bpf_obj) {
3510                         evlist__remove(trace->evlist, evsel);
3511                         evsel__delete(evsel);
3512                 }
3513
3514         }
3515
3516         bpf_object__close(trace->bpf_obj);
3517         trace->bpf_obj = NULL;
3518 }
3519 #else // HAVE_LIBBPF_SUPPORT
3520 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3521 {
3522         return 0;
3523 }
3524
3525 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3526 {
3527         return 0;
3528 }
3529
3530 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3531                                                             const char *name __maybe_unused)
3532 {
3533         return NULL;
3534 }
3535
3536 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3537 {
3538         return 0;
3539 }
3540
3541 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3542 {
3543 }
3544 #endif // HAVE_LIBBPF_SUPPORT
3545
3546 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3547 {
3548         struct evsel *evsel;
3549
3550         evlist__for_each_entry(trace->evlist, evsel) {
3551                 if (evsel == trace->syscalls.events.augmented ||
3552                     evsel->bpf_obj == trace->bpf_obj)
3553                         continue;
3554
3555                 return false;
3556         }
3557
3558         return true;
3559 }
3560
3561 static int trace__set_ev_qualifier_filter(struct trace *trace)
3562 {
3563         if (trace->syscalls.map)
3564                 return trace__set_ev_qualifier_bpf_filter(trace);
3565         if (trace->syscalls.events.sys_enter)
3566                 return trace__set_ev_qualifier_tp_filter(trace);
3567         return 0;
3568 }
3569
3570 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3571                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3572 {
3573         int err = 0;
3574 #ifdef HAVE_LIBBPF_SUPPORT
3575         bool value = true;
3576         int map_fd = bpf_map__fd(map);
3577         size_t i;
3578
3579         for (i = 0; i < npids; ++i) {
3580                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3581                 if (err)
3582                         break;
3583         }
3584 #endif
3585         return err;
3586 }
3587
3588 static int trace__set_filter_loop_pids(struct trace *trace)
3589 {
3590         unsigned int nr = 1, err;
3591         pid_t pids[32] = {
3592                 getpid(),
3593         };
3594         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3595
3596         while (thread && nr < ARRAY_SIZE(pids)) {
3597                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3598
3599                 if (parent == NULL)
3600                         break;
3601
3602                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3603                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3604                         pids[nr++] = parent->tid;
3605                         break;
3606                 }
3607                 thread = parent;
3608         }
3609
3610         err = perf_evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3611         if (!err && trace->filter_pids.map)
3612                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3613
3614         return err;
3615 }
3616
3617 static int trace__set_filter_pids(struct trace *trace)
3618 {
3619         int err = 0;
3620         /*
3621          * Better not use !target__has_task() here because we need to cover the
3622          * case where no threads were specified in the command line, but a
3623          * workload was, and in that case we will fill in the thread_map when
3624          * we fork the workload in perf_evlist__prepare_workload.
3625          */
3626         if (trace->filter_pids.nr > 0) {
3627                 err = perf_evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3628                                                          trace->filter_pids.entries);
3629                 if (!err && trace->filter_pids.map) {
3630                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3631                                                        trace->filter_pids.entries);
3632                 }
3633         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3634                 err = trace__set_filter_loop_pids(trace);
3635         }
3636
3637         return err;
3638 }
3639
3640 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3641 {
3642         struct evlist *evlist = trace->evlist;
3643         struct perf_sample sample;
3644         int err;
3645
3646         err = perf_evlist__parse_sample(evlist, event, &sample);
3647         if (err)
3648                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3649         else
3650                 trace__handle_event(trace, event, &sample);
3651
3652         return 0;
3653 }
3654
3655 static int __trace__flush_events(struct trace *trace)
3656 {
3657         u64 first = ordered_events__first_time(&trace->oe.data);
3658         u64 flush = trace->oe.last - NSEC_PER_SEC;
3659
3660         /* Is there some thing to flush.. */
3661         if (first && first < flush)
3662                 return ordered_events__flush_time(&trace->oe.data, flush);
3663
3664         return 0;
3665 }
3666
3667 static int trace__flush_events(struct trace *trace)
3668 {
3669         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3670 }
3671
3672 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3673 {
3674         int err;
3675
3676         if (!trace->sort_events)
3677                 return __trace__deliver_event(trace, event);
3678
3679         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3680         if (err && err != -1)
3681                 return err;
3682
3683         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3684         if (err)
3685                 return err;
3686
3687         return trace__flush_events(trace);
3688 }
3689
3690 static int ordered_events__deliver_event(struct ordered_events *oe,
3691                                          struct ordered_event *event)
3692 {
3693         struct trace *trace = container_of(oe, struct trace, oe.data);
3694
3695         return __trace__deliver_event(trace, event->event);
3696 }
3697
3698 static struct syscall_arg_fmt *perf_evsel__syscall_arg_fmt(struct evsel *evsel, char *arg)
3699 {
3700         struct tep_format_field *field;
3701         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3702
3703         if (evsel->tp_format == NULL || fmt == NULL)
3704                 return NULL;
3705
3706         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3707                 if (strcmp(field->name, arg) == 0)
3708                         return fmt;
3709
3710         return NULL;
3711 }
3712
3713 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3714 {
3715         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3716
3717         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3718                 char *right = tok + 1, *right_end;
3719
3720                 if (*right == '=')
3721                         ++right;
3722
3723                 while (isspace(*right))
3724                         ++right;
3725
3726                 if (*right == '\0')
3727                         break;
3728
3729                 while (!isalpha(*left))
3730                         if (++left == tok) {
3731                                 /*
3732                                  * Bail out, can't find the name of the argument that is being
3733                                  * used in the filter, let it try to set this filter, will fail later.
3734                                  */
3735                                 return 0;
3736                         }
3737
3738                 right_end = right + 1;
3739                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3740                         ++right_end;
3741
3742                 if (isalpha(*right)) {
3743                         struct syscall_arg_fmt *fmt;
3744                         int left_size = tok - left,
3745                             right_size = right_end - right;
3746                         char arg[128];
3747
3748                         while (isspace(left[left_size - 1]))
3749                                 --left_size;
3750
3751                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3752
3753                         fmt = perf_evsel__syscall_arg_fmt(evsel, arg);
3754                         if (fmt == NULL) {
3755                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3756                                        arg, evsel->name, evsel->filter);
3757                                 return -1;
3758                         }
3759
3760                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3761                                  arg, (int)(right - tok), tok, right_size, right);
3762
3763                         if (fmt->strtoul) {
3764                                 u64 val;
3765                                 struct syscall_arg syscall_arg = {
3766                                         .parm = fmt->parm,
3767                                 };
3768
3769                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3770                                         char *n, expansion[19];
3771                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3772                                         int expansion_offset = right - new_filter;
3773
3774                                         pr_debug("%s", expansion);
3775
3776                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3777                                                 pr_debug(" out of memory!\n");
3778                                                 free(new_filter);
3779                                                 return -1;
3780                                         }
3781                                         if (new_filter != evsel->filter)
3782                                                 free(new_filter);
3783                                         left = n + expansion_offset + expansion_lenght;
3784                                         new_filter = n;
3785                                 } else {
3786                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3787                                                right_size, right, arg, evsel->name, evsel->filter);
3788                                         return -1;
3789                                 }
3790                         } else {
3791                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3792                                        arg, evsel->name, evsel->filter);
3793                                 return -1;
3794                         }
3795
3796                         pr_debug("\n");
3797                 } else {
3798                         left = right_end;
3799                 }
3800         }
3801
3802         if (new_filter != evsel->filter) {
3803                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3804                 perf_evsel__set_filter(evsel, new_filter);
3805                 free(new_filter);
3806         }
3807
3808         return 0;
3809 }
3810
3811 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3812 {
3813         struct evlist *evlist = trace->evlist;
3814         struct evsel *evsel;
3815
3816         evlist__for_each_entry(evlist, evsel) {
3817                 if (evsel->filter == NULL)
3818                         continue;
3819
3820                 if (trace__expand_filter(trace, evsel)) {
3821                         *err_evsel = evsel;
3822                         return -1;
3823                 }
3824         }
3825
3826         return 0;
3827 }
3828
3829 static int trace__run(struct trace *trace, int argc, const char **argv)
3830 {
3831         struct evlist *evlist = trace->evlist;
3832         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3833         int err = -1, i;
3834         unsigned long before;
3835         const bool forks = argc > 0;
3836         bool draining = false;
3837
3838         trace->live = true;
3839
3840         if (!trace->raw_augmented_syscalls) {
3841                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3842                         goto out_error_raw_syscalls;
3843
3844                 if (trace->trace_syscalls)
3845                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3846         }
3847
3848         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3849                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3850                 if (pgfault_maj == NULL)
3851                         goto out_error_mem;
3852                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3853                 evlist__add(evlist, pgfault_maj);
3854         }
3855
3856         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3857                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3858                 if (pgfault_min == NULL)
3859                         goto out_error_mem;
3860                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3861                 evlist__add(evlist, pgfault_min);
3862         }
3863
3864         if (trace->sched &&
3865             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3866                                    trace__sched_stat_runtime))
3867                 goto out_error_sched_stat_runtime;
3868         /*
3869          * If a global cgroup was set, apply it to all the events without an
3870          * explicit cgroup. I.e.:
3871          *
3872          *      trace -G A -e sched:*switch
3873          *
3874          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3875          * _and_ sched:sched_switch to the 'A' cgroup, while:
3876          *
3877          * trace -e sched:*switch -G A
3878          *
3879          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3880          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3881          * a cgroup (on the root cgroup, sys wide, etc).
3882          *
3883          * Multiple cgroups:
3884          *
3885          * trace -G A -e sched:*switch -G B
3886          *
3887          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3888          * to the 'B' cgroup.
3889          *
3890          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3891          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3892          */
3893         if (trace->cgroup)
3894                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3895
3896         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3897         if (err < 0) {
3898                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3899                 goto out_delete_evlist;
3900         }
3901
3902         err = trace__symbols_init(trace, evlist);
3903         if (err < 0) {
3904                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3905                 goto out_delete_evlist;
3906         }
3907
3908         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3909
3910         signal(SIGCHLD, sig_handler);
3911         signal(SIGINT, sig_handler);
3912
3913         if (forks) {
3914                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3915                                                     argv, false, NULL);
3916                 if (err < 0) {
3917                         fprintf(trace->output, "Couldn't run the workload!\n");
3918                         goto out_delete_evlist;
3919                 }
3920         }
3921
3922         err = evlist__open(evlist);
3923         if (err < 0)
3924                 goto out_error_open;
3925
3926         err = bpf__apply_obj_config();
3927         if (err) {
3928                 char errbuf[BUFSIZ];
3929
3930                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3931                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3932                          errbuf);
3933                 goto out_error_open;
3934         }
3935
3936         err = trace__set_filter_pids(trace);
3937         if (err < 0)
3938                 goto out_error_mem;
3939
3940         if (trace->syscalls.map)
3941                 trace__init_syscalls_bpf_map(trace);
3942
3943         if (trace->syscalls.prog_array.sys_enter)
3944                 trace__init_syscalls_bpf_prog_array_maps(trace);
3945
3946         if (trace->ev_qualifier_ids.nr > 0) {
3947                 err = trace__set_ev_qualifier_filter(trace);
3948                 if (err < 0)
3949                         goto out_errno;
3950
3951                 if (trace->syscalls.events.sys_exit) {
3952                         pr_debug("event qualifier tracepoint filter: %s\n",
3953                                  trace->syscalls.events.sys_exit->filter);
3954                 }
3955         }
3956
3957         /*
3958          * If the "close" syscall is not traced, then we will not have the
3959          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3960          * fd->pathname table and were ending up showing the last value set by
3961          * syscalls opening a pathname and associating it with a descriptor or
3962          * reading it from /proc/pid/fd/ in cases where that doesn't make
3963          * sense.
3964          *
3965          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3966          *  not in use.
3967          */
3968         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3969
3970         err = trace__expand_filters(trace, &evsel);
3971         if (err)
3972                 goto out_delete_evlist;
3973         err = perf_evlist__apply_filters(evlist, &evsel);
3974         if (err < 0)
3975                 goto out_error_apply_filters;
3976
3977         if (trace->dump.map)
3978                 bpf_map__fprintf(trace->dump.map, trace->output);
3979
3980         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3981         if (err < 0)
3982                 goto out_error_mmap;
3983
3984         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3985                 evlist__enable(evlist);
3986
3987         if (forks)
3988                 perf_evlist__start_workload(evlist);
3989
3990         if (trace->opts.initial_delay) {
3991                 usleep(trace->opts.initial_delay * 1000);
3992                 evlist__enable(evlist);
3993         }
3994
3995         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3996                                   evlist->core.threads->nr > 1 ||
3997                                   evlist__first(evlist)->core.attr.inherit;
3998
3999         /*
4000          * Now that we already used evsel->core.attr to ask the kernel to setup the
4001          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4002          * trace__resolve_callchain(), allowing per-event max-stack settings
4003          * to override an explicitly set --max-stack global setting.
4004          */
4005         evlist__for_each_entry(evlist, evsel) {
4006                 if (evsel__has_callchain(evsel) &&
4007                     evsel->core.attr.sample_max_stack == 0)
4008                         evsel->core.attr.sample_max_stack = trace->max_stack;
4009         }
4010 again:
4011         before = trace->nr_events;
4012
4013         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4014                 union perf_event *event;
4015                 struct mmap *md;
4016
4017                 md = &evlist->mmap[i];
4018                 if (perf_mmap__read_init(&md->core) < 0)
4019                         continue;
4020
4021                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4022                         ++trace->nr_events;
4023
4024                         err = trace__deliver_event(trace, event);
4025                         if (err)
4026                                 goto out_disable;
4027
4028                         perf_mmap__consume(&md->core);
4029
4030                         if (interrupted)
4031                                 goto out_disable;
4032
4033                         if (done && !draining) {
4034                                 evlist__disable(evlist);
4035                                 draining = true;
4036                         }
4037                 }
4038                 perf_mmap__read_done(&md->core);
4039         }
4040
4041         if (trace->nr_events == before) {
4042                 int timeout = done ? 100 : -1;
4043
4044                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4045                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4046                                 draining = true;
4047
4048                         goto again;
4049                 } else {
4050                         if (trace__flush_events(trace))
4051                                 goto out_disable;
4052                 }
4053         } else {
4054                 goto again;
4055         }
4056
4057 out_disable:
4058         thread__zput(trace->current);
4059
4060         evlist__disable(evlist);
4061
4062         if (trace->sort_events)
4063                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4064
4065         if (!err) {
4066                 if (trace->summary)
4067                         trace__fprintf_thread_summary(trace, trace->output);
4068
4069                 if (trace->show_tool_stats) {
4070                         fprintf(trace->output, "Stats:\n "
4071                                                " vfs_getname : %" PRIu64 "\n"
4072                                                " proc_getname: %" PRIu64 "\n",
4073                                 trace->stats.vfs_getname,
4074                                 trace->stats.proc_getname);
4075                 }
4076         }
4077
4078 out_delete_evlist:
4079         trace__symbols__exit(trace);
4080
4081         evlist__delete(evlist);
4082         cgroup__put(trace->cgroup);
4083         trace->evlist = NULL;
4084         trace->live = false;
4085         return err;
4086 {
4087         char errbuf[BUFSIZ];
4088
4089 out_error_sched_stat_runtime:
4090         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4091         goto out_error;
4092
4093 out_error_raw_syscalls:
4094         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4095         goto out_error;
4096
4097 out_error_mmap:
4098         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4099         goto out_error;
4100
4101 out_error_open:
4102         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4103
4104 out_error:
4105         fprintf(trace->output, "%s\n", errbuf);
4106         goto out_delete_evlist;
4107
4108 out_error_apply_filters:
4109         fprintf(trace->output,
4110                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4111                 evsel->filter, perf_evsel__name(evsel), errno,
4112                 str_error_r(errno, errbuf, sizeof(errbuf)));
4113         goto out_delete_evlist;
4114 }
4115 out_error_mem:
4116         fprintf(trace->output, "Not enough memory to run!\n");
4117         goto out_delete_evlist;
4118
4119 out_errno:
4120         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4121         goto out_delete_evlist;
4122 }
4123
4124 static int trace__replay(struct trace *trace)
4125 {
4126         const struct evsel_str_handler handlers[] = {
4127                 { "probe:vfs_getname",       trace__vfs_getname, },
4128         };
4129         struct perf_data data = {
4130                 .path  = input_name,
4131                 .mode  = PERF_DATA_MODE_READ,
4132                 .force = trace->force,
4133         };
4134         struct perf_session *session;
4135         struct evsel *evsel;
4136         int err = -1;
4137
4138         trace->tool.sample        = trace__process_sample;
4139         trace->tool.mmap          = perf_event__process_mmap;
4140         trace->tool.mmap2         = perf_event__process_mmap2;
4141         trace->tool.comm          = perf_event__process_comm;
4142         trace->tool.exit          = perf_event__process_exit;
4143         trace->tool.fork          = perf_event__process_fork;
4144         trace->tool.attr          = perf_event__process_attr;
4145         trace->tool.tracing_data  = perf_event__process_tracing_data;
4146         trace->tool.build_id      = perf_event__process_build_id;
4147         trace->tool.namespaces    = perf_event__process_namespaces;
4148
4149         trace->tool.ordered_events = true;
4150         trace->tool.ordering_requires_timestamps = true;
4151
4152         /* add tid to output */
4153         trace->multiple_threads = true;
4154
4155         session = perf_session__new(&data, false, &trace->tool);
4156         if (IS_ERR(session))
4157                 return PTR_ERR(session);
4158
4159         if (trace->opts.target.pid)
4160                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4161
4162         if (trace->opts.target.tid)
4163                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4164
4165         if (symbol__init(&session->header.env) < 0)
4166                 goto out;
4167
4168         trace->host = &session->machines.host;
4169
4170         err = perf_session__set_tracepoints_handlers(session, handlers);
4171         if (err)
4172                 goto out;
4173
4174         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4175                                                      "raw_syscalls:sys_enter");
4176         /* older kernels have syscalls tp versus raw_syscalls */
4177         if (evsel == NULL)
4178                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4179                                                              "syscalls:sys_enter");
4180
4181         if (evsel &&
4182             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4183             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4184                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4185                 goto out;
4186         }
4187
4188         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4189                                                      "raw_syscalls:sys_exit");
4190         if (evsel == NULL)
4191                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4192                                                              "syscalls:sys_exit");
4193         if (evsel &&
4194             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4195             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4196                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4197                 goto out;
4198         }
4199
4200         evlist__for_each_entry(session->evlist, evsel) {
4201                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4202                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4203                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4204                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4205                         evsel->handler = trace__pgfault;
4206         }
4207
4208         setup_pager();
4209
4210         err = perf_session__process_events(session);
4211         if (err)
4212                 pr_err("Failed to process events, error %d", err);
4213
4214         else if (trace->summary)
4215                 trace__fprintf_thread_summary(trace, trace->output);
4216
4217 out:
4218         perf_session__delete(session);
4219
4220         return err;
4221 }
4222
4223 static size_t trace__fprintf_threads_header(FILE *fp)
4224 {
4225         size_t printed;
4226
4227         printed  = fprintf(fp, "\n Summary of events:\n\n");
4228
4229         return printed;
4230 }
4231
4232 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4233         struct syscall_stats *stats;
4234         double               msecs;
4235         int                  syscall;
4236 )
4237 {
4238         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4239         struct syscall_stats *stats = source->priv;
4240
4241         entry->syscall = source->i;
4242         entry->stats   = stats;
4243         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4244 }
4245
4246 static size_t thread__dump_stats(struct thread_trace *ttrace,
4247                                  struct trace *trace, FILE *fp)
4248 {
4249         size_t printed = 0;
4250         struct syscall *sc;
4251         struct rb_node *nd;
4252         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4253
4254         if (syscall_stats == NULL)
4255                 return 0;
4256
4257         printed += fprintf(fp, "\n");
4258
4259         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4260         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4261         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4262
4263         resort_rb__for_each_entry(nd, syscall_stats) {
4264                 struct syscall_stats *stats = syscall_stats_entry->stats;
4265                 if (stats) {
4266                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4267                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4268                         double avg = avg_stats(&stats->stats);
4269                         double pct;
4270                         u64 n = (u64)stats->stats.n;
4271
4272                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4273                         avg /= NSEC_PER_MSEC;
4274
4275                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4276                         printed += fprintf(fp, "   %-15s", sc->name);
4277                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4278                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4279                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4280
4281                         if (trace->errno_summary && stats->nr_failures) {
4282                                 const char *arch_name = perf_env__arch(trace->host->env);
4283                                 int e;
4284
4285                                 for (e = 0; e < stats->max_errno; ++e) {
4286                                         if (stats->errnos[e] != 0)
4287                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4288                                 }
4289                         }
4290                 }
4291         }
4292
4293         resort_rb__delete(syscall_stats);
4294         printed += fprintf(fp, "\n\n");
4295
4296         return printed;
4297 }
4298
4299 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4300 {
4301         size_t printed = 0;
4302         struct thread_trace *ttrace = thread__priv(thread);
4303         double ratio;
4304
4305         if (ttrace == NULL)
4306                 return 0;
4307
4308         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4309
4310         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4311         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4312         printed += fprintf(fp, "%.1f%%", ratio);
4313         if (ttrace->pfmaj)
4314                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4315         if (ttrace->pfmin)
4316                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4317         if (trace->sched)
4318                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4319         else if (fputc('\n', fp) != EOF)
4320                 ++printed;
4321
4322         printed += thread__dump_stats(ttrace, trace, fp);
4323
4324         return printed;
4325 }
4326
4327 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4328 {
4329         return ttrace ? ttrace->nr_events : 0;
4330 }
4331
4332 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4333         struct thread *thread;
4334 )
4335 {
4336         entry->thread = rb_entry(nd, struct thread, rb_node);
4337 }
4338
4339 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4340 {
4341         size_t printed = trace__fprintf_threads_header(fp);
4342         struct rb_node *nd;
4343         int i;
4344
4345         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4346                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4347
4348                 if (threads == NULL) {
4349                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4350                         return 0;
4351                 }
4352
4353                 resort_rb__for_each_entry(nd, threads)
4354                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4355
4356                 resort_rb__delete(threads);
4357         }
4358         return printed;
4359 }
4360
4361 static int trace__set_duration(const struct option *opt, const char *str,
4362                                int unset __maybe_unused)
4363 {
4364         struct trace *trace = opt->value;
4365
4366         trace->duration_filter = atof(str);
4367         return 0;
4368 }
4369
4370 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4371                                               int unset __maybe_unused)
4372 {
4373         int ret = -1;
4374         size_t i;
4375         struct trace *trace = opt->value;
4376         /*
4377          * FIXME: introduce a intarray class, plain parse csv and create a
4378          * { int nr, int entries[] } struct...
4379          */
4380         struct intlist *list = intlist__new(str);
4381
4382         if (list == NULL)
4383                 return -1;
4384
4385         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4386         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4387
4388         if (trace->filter_pids.entries == NULL)
4389                 goto out;
4390
4391         trace->filter_pids.entries[0] = getpid();
4392
4393         for (i = 1; i < trace->filter_pids.nr; ++i)
4394                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4395
4396         intlist__delete(list);
4397         ret = 0;
4398 out:
4399         return ret;
4400 }
4401
4402 static int trace__open_output(struct trace *trace, const char *filename)
4403 {
4404         struct stat st;
4405
4406         if (!stat(filename, &st) && st.st_size) {
4407                 char oldname[PATH_MAX];
4408
4409                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4410                 unlink(oldname);
4411                 rename(filename, oldname);
4412         }
4413
4414         trace->output = fopen(filename, "w");
4415
4416         return trace->output == NULL ? -errno : 0;
4417 }
4418
4419 static int parse_pagefaults(const struct option *opt, const char *str,
4420                             int unset __maybe_unused)
4421 {
4422         int *trace_pgfaults = opt->value;
4423
4424         if (strcmp(str, "all") == 0)
4425                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4426         else if (strcmp(str, "maj") == 0)
4427                 *trace_pgfaults |= TRACE_PFMAJ;
4428         else if (strcmp(str, "min") == 0)
4429                 *trace_pgfaults |= TRACE_PFMIN;
4430         else
4431                 return -1;
4432
4433         return 0;
4434 }
4435
4436 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4437 {
4438         struct evsel *evsel;
4439
4440         evlist__for_each_entry(evlist, evsel) {
4441                 if (evsel->handler == NULL)
4442                         evsel->handler = handler;
4443         }
4444 }
4445
4446 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4447 {
4448         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4449
4450         if (fmt) {
4451                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4452
4453                 if (scfmt) {
4454                         int skip = 0;
4455
4456                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4457                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4458                                 ++skip;
4459
4460                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4461                 }
4462         }
4463 }
4464
4465 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4466 {
4467         struct evsel *evsel;
4468
4469         evlist__for_each_entry(evlist, evsel) {
4470                 if (evsel->priv || !evsel->tp_format)
4471                         continue;
4472
4473                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4474                         perf_evsel__init_tp_arg_scnprintf(evsel);
4475                         continue;
4476                 }
4477
4478                 if (perf_evsel__init_syscall_tp(evsel))
4479                         return -1;
4480
4481                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4482                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4483
4484                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4485                                 return -1;
4486
4487                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4488                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4489                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4490
4491                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4492                                 return -1;
4493
4494                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4495                 }
4496         }
4497
4498         return 0;
4499 }
4500
4501 /*
4502  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4503  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4504  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4505  *
4506  * It'd be better to introduce a parse_options() variant that would return a
4507  * list with the terms it didn't match to an event...
4508  */
4509 static int trace__parse_events_option(const struct option *opt, const char *str,
4510                                       int unset __maybe_unused)
4511 {
4512         struct trace *trace = (struct trace *)opt->value;
4513         const char *s = str;
4514         char *sep = NULL, *lists[2] = { NULL, NULL, };
4515         int len = strlen(str) + 1, err = -1, list, idx;
4516         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4517         char group_name[PATH_MAX];
4518         struct syscall_fmt *fmt;
4519
4520         if (strace_groups_dir == NULL)
4521                 return -1;
4522
4523         if (*s == '!') {
4524                 ++s;
4525                 trace->not_ev_qualifier = true;
4526         }
4527
4528         while (1) {
4529                 if ((sep = strchr(s, ',')) != NULL)
4530                         *sep = '\0';
4531
4532                 list = 0;
4533                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4534                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4535                         list = 1;
4536                         goto do_concat;
4537                 }
4538
4539                 fmt = syscall_fmt__find_by_alias(s);
4540                 if (fmt != NULL) {
4541                         list = 1;
4542                         s = fmt->name;
4543                 } else {
4544                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4545                         if (access(group_name, R_OK) == 0)
4546                                 list = 1;
4547                 }
4548 do_concat:
4549                 if (lists[list]) {
4550                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4551                 } else {
4552                         lists[list] = malloc(len);
4553                         if (lists[list] == NULL)
4554                                 goto out;
4555                         strcpy(lists[list], s);
4556                 }
4557
4558                 if (!sep)
4559                         break;
4560
4561                 *sep = ',';
4562                 s = sep + 1;
4563         }
4564
4565         if (lists[1] != NULL) {
4566                 struct strlist_config slist_config = {
4567                         .dirname = strace_groups_dir,
4568                 };
4569
4570                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4571                 if (trace->ev_qualifier == NULL) {
4572                         fputs("Not enough memory to parse event qualifier", trace->output);
4573                         goto out;
4574                 }
4575
4576                 if (trace__validate_ev_qualifier(trace))
4577                         goto out;
4578                 trace->trace_syscalls = true;
4579         }
4580
4581         err = 0;
4582
4583         if (lists[0]) {
4584                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4585                                                "event selector. use 'perf list' to list available events",
4586                                                parse_events_option);
4587                 err = parse_events_option(&o, lists[0], 0);
4588         }
4589 out:
4590         if (sep)
4591                 *sep = ',';
4592
4593         return err;
4594 }
4595
4596 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4597 {
4598         struct trace *trace = opt->value;
4599
4600         if (!list_empty(&trace->evlist->core.entries))
4601                 return parse_cgroups(opt, str, unset);
4602
4603         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4604
4605         return 0;
4606 }
4607
4608 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4609 {
4610         if (trace->bpf_obj == NULL)
4611                 return NULL;
4612
4613         return bpf_object__find_map_by_name(trace->bpf_obj, name);
4614 }
4615
4616 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4617 {
4618         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4619 }
4620
4621 static void trace__set_bpf_map_syscalls(struct trace *trace)
4622 {
4623         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4624         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4625         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4626 }
4627
4628 static int trace__config(const char *var, const char *value, void *arg)
4629 {
4630         struct trace *trace = arg;
4631         int err = 0;
4632
4633         if (!strcmp(var, "trace.add_events")) {
4634                 trace->perfconfig_events = strdup(value);
4635                 if (trace->perfconfig_events == NULL) {
4636                         pr_err("Not enough memory for %s\n", "trace.add_events");
4637                         return -1;
4638                 }
4639         } else if (!strcmp(var, "trace.show_timestamp")) {
4640                 trace->show_tstamp = perf_config_bool(var, value);
4641         } else if (!strcmp(var, "trace.show_duration")) {
4642                 trace->show_duration = perf_config_bool(var, value);
4643         } else if (!strcmp(var, "trace.show_arg_names")) {
4644                 trace->show_arg_names = perf_config_bool(var, value);
4645                 if (!trace->show_arg_names)
4646                         trace->show_zeros = true;
4647         } else if (!strcmp(var, "trace.show_zeros")) {
4648                 bool new_show_zeros = perf_config_bool(var, value);
4649                 if (!trace->show_arg_names && !new_show_zeros) {
4650                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4651                         goto out;
4652                 }
4653                 trace->show_zeros = new_show_zeros;
4654         } else if (!strcmp(var, "trace.show_prefix")) {
4655                 trace->show_string_prefix = perf_config_bool(var, value);
4656         } else if (!strcmp(var, "trace.no_inherit")) {
4657                 trace->opts.no_inherit = perf_config_bool(var, value);
4658         } else if (!strcmp(var, "trace.args_alignment")) {
4659                 int args_alignment = 0;
4660                 if (perf_config_int(&args_alignment, var, value) == 0)
4661                         trace->args_alignment = args_alignment;
4662         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4663                 if (strcasecmp(value, "libtraceevent") == 0)
4664                         trace->libtraceevent_print = true;
4665                 else if (strcasecmp(value, "libbeauty") == 0)
4666                         trace->libtraceevent_print = false;
4667         }
4668 out:
4669         return err;
4670 }
4671
4672 int cmd_trace(int argc, const char **argv)
4673 {
4674         const char *trace_usage[] = {
4675                 "perf trace [<options>] [<command>]",
4676                 "perf trace [<options>] -- <command> [<options>]",
4677                 "perf trace record [<options>] [<command>]",
4678                 "perf trace record [<options>] -- <command> [<options>]",
4679                 NULL
4680         };
4681         struct trace trace = {
4682                 .opts = {
4683                         .target = {
4684                                 .uid       = UINT_MAX,
4685                                 .uses_mmap = true,
4686                         },
4687                         .user_freq     = UINT_MAX,
4688                         .user_interval = ULLONG_MAX,
4689                         .no_buffering  = true,
4690                         .mmap_pages    = UINT_MAX,
4691                 },
4692                 .output = stderr,
4693                 .show_comm = true,
4694                 .show_tstamp = true,
4695                 .show_duration = true,
4696                 .show_arg_names = true,
4697                 .args_alignment = 70,
4698                 .trace_syscalls = false,
4699                 .kernel_syscallchains = false,
4700                 .max_stack = UINT_MAX,
4701                 .max_events = ULONG_MAX,
4702         };
4703         const char *map_dump_str = NULL;
4704         const char *output_name = NULL;
4705         const struct option trace_options[] = {
4706         OPT_CALLBACK('e', "event", &trace, "event",
4707                      "event/syscall selector. use 'perf list' to list available events",
4708                      trace__parse_events_option),
4709         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4710                      "event filter", parse_filter),
4711         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4712                     "show the thread COMM next to its id"),
4713         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4714         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4715                      trace__parse_events_option),
4716         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4717         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4718         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4719                     "trace events on existing process id"),
4720         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4721                     "trace events on existing thread id"),
4722         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4723                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4724         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4725                     "system-wide collection from all CPUs"),
4726         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4727                     "list of cpus to monitor"),
4728         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4729                     "child tasks do not inherit counters"),
4730         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4731                      "number of mmap data pages",
4732                      perf_evlist__parse_mmap_pages),
4733         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4734                    "user to profile"),
4735         OPT_CALLBACK(0, "duration", &trace, "float",
4736                      "show only events with duration > N.M ms",
4737                      trace__set_duration),
4738 #ifdef HAVE_LIBBPF_SUPPORT
4739         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4740 #endif
4741         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4742         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4743         OPT_BOOLEAN('T', "time", &trace.full_time,
4744                     "Show full timestamp, not time relative to first start"),
4745         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4746                     "Show only syscalls that failed"),
4747         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4748                     "Show only syscall summary with statistics"),
4749         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4750                     "Show all syscalls and summary with statistics"),
4751         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4752                     "Show errno stats per syscall, use with -s or -S"),
4753         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4754                      "Trace pagefaults", parse_pagefaults, "maj"),
4755         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4756         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4757         OPT_CALLBACK(0, "call-graph", &trace.opts,
4758                      "record_mode[,record_size]", record_callchain_help,
4759                      &record_parse_callchain_opt),
4760         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4761                     "Use libtraceevent to print the tracepoint arguments."),
4762         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4763                     "Show the kernel callchains on the syscall exit path"),
4764         OPT_ULONG(0, "max-events", &trace.max_events,
4765                 "Set the maximum number of events to print, exit after that is reached. "),
4766         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4767                      "Set the minimum stack depth when parsing the callchain, "
4768                      "anything below the specified depth will be ignored."),
4769         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4770                      "Set the maximum stack depth when parsing the callchain, "
4771                      "anything beyond the specified depth will be ignored. "
4772                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4773         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4774                         "Sort batch of events before processing, use if getting out of order events"),
4775         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4776                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4777         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4778                         "per thread proc mmap processing timeout in ms"),
4779         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4780                      trace__parse_cgroups),
4781         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4782                      "ms to wait before starting measurement after program "
4783                      "start"),
4784         OPTS_EVSWITCH(&trace.evswitch),
4785         OPT_END()
4786         };
4787         bool __maybe_unused max_stack_user_set = true;
4788         bool mmap_pages_user_set = true;
4789         struct evsel *evsel;
4790         const char * const trace_subcommands[] = { "record", NULL };
4791         int err = -1;
4792         char bf[BUFSIZ];
4793
4794         signal(SIGSEGV, sighandler_dump_stack);
4795         signal(SIGFPE, sighandler_dump_stack);
4796
4797         trace.evlist = evlist__new();
4798         trace.sctbl = syscalltbl__new();
4799
4800         if (trace.evlist == NULL || trace.sctbl == NULL) {
4801                 pr_err("Not enough memory to run!\n");
4802                 err = -ENOMEM;
4803                 goto out;
4804         }
4805
4806         /*
4807          * Parsing .perfconfig may entail creating a BPF event, that may need
4808          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4809          * is too small. This affects just this process, not touching the
4810          * global setting. If it fails we'll get something in 'perf trace -v'
4811          * to help diagnose the problem.
4812          */
4813         rlimit__bump_memlock();
4814
4815         err = perf_config(trace__config, &trace);
4816         if (err)
4817                 goto out;
4818
4819         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4820                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4821
4822         /*
4823          * Here we already passed thru trace__parse_events_option() and it has
4824          * already figured out if -e syscall_name, if not but if --event
4825          * foo:bar was used, the user is interested _just_ in those, say,
4826          * tracepoint events, not in the strace-like syscall-name-based mode.
4827          *
4828          * This is important because we need to check if strace-like mode is
4829          * needed to decided if we should filter out the eBPF
4830          * __augmented_syscalls__ code, if it is in the mix, say, via
4831          * .perfconfig trace.add_events, and filter those out.
4832          */
4833         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4834             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4835                 trace.trace_syscalls = true;
4836         }
4837         /*
4838          * Now that we have --verbose figured out, lets see if we need to parse
4839          * events from .perfconfig, so that if those events fail parsing, say some
4840          * BPF program fails, then we'll be able to use --verbose to see what went
4841          * wrong in more detail.
4842          */
4843         if (trace.perfconfig_events != NULL) {
4844                 struct parse_events_error parse_err;
4845
4846                 bzero(&parse_err, sizeof(parse_err));
4847                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4848                 if (err) {
4849                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4850                         goto out;
4851                 }
4852         }
4853
4854         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4855                 usage_with_options_msg(trace_usage, trace_options,
4856                                        "cgroup monitoring only available in system-wide mode");
4857         }
4858
4859         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4860         if (IS_ERR(evsel)) {
4861                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4862                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4863                 goto out;
4864         }
4865
4866         if (evsel) {
4867                 trace.syscalls.events.augmented = evsel;
4868
4869                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4870                 if (evsel == NULL) {
4871                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4872                         goto out;
4873                 }
4874
4875                 if (evsel->bpf_obj == NULL) {
4876                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4877                         goto out;
4878                 }
4879
4880                 trace.bpf_obj = evsel->bpf_obj;
4881
4882                 /*
4883                  * If we have _just_ the augmenter event but don't have a
4884                  * explicit --syscalls, then assume we want all strace-like
4885                  * syscalls:
4886                  */
4887                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4888                         trace.trace_syscalls = true;
4889                 /*
4890                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4891                  * strace-like syscall tracing is not set, then we need to trow
4892                  * away the augmenter, i.e. all the events that were created
4893                  * from that BPF object file.
4894                  *
4895                  * This is more to fix the current .perfconfig trace.add_events
4896                  * style of setting up the strace-like eBPF based syscall point
4897                  * payload augmenter.
4898                  *
4899                  * All this complexity will be avoided by adding an alternative
4900                  * to trace.add_events in the form of
4901                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4902                  * need it.
4903                  *
4904                  * .perfconfig trace.add_events is still useful if we want, for
4905                  * instance, have msr_write.msr in some .perfconfig profile based
4906                  * 'perf trace --config determinism.profile' mode, where for some
4907                  * particular goal/workload type we want a set of events and
4908                  * output mode (with timings, etc) instead of having to add
4909                  * all via the command line.
4910                  *
4911                  * Also --config to specify an alternate .perfconfig file needs
4912                  * to be implemented.
4913                  */
4914                 if (!trace.trace_syscalls) {
4915                         trace__delete_augmented_syscalls(&trace);
4916                 } else {
4917                         trace__set_bpf_map_filtered_pids(&trace);
4918                         trace__set_bpf_map_syscalls(&trace);
4919                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4920                 }
4921         }
4922
4923         err = bpf__setup_stdout(trace.evlist);
4924         if (err) {
4925                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4926                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4927                 goto out;
4928         }
4929
4930         err = -1;
4931
4932         if (map_dump_str) {
4933                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4934                 if (trace.dump.map == NULL) {
4935                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4936                         goto out;
4937                 }
4938         }
4939
4940         if (trace.trace_pgfaults) {
4941                 trace.opts.sample_address = true;
4942                 trace.opts.sample_time = true;
4943         }
4944
4945         if (trace.opts.mmap_pages == UINT_MAX)
4946                 mmap_pages_user_set = false;
4947
4948         if (trace.max_stack == UINT_MAX) {
4949                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4950                 max_stack_user_set = false;
4951         }
4952
4953 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4954         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4955                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4956         }
4957 #endif
4958
4959         if (callchain_param.enabled) {
4960                 if (!mmap_pages_user_set && geteuid() == 0)
4961                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4962
4963                 symbol_conf.use_callchain = true;
4964         }
4965
4966         if (trace.evlist->core.nr_entries > 0) {
4967                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4968                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4969                         perror("failed to set syscalls:* tracepoint fields");
4970                         goto out;
4971                 }
4972         }
4973
4974         if (trace.sort_events) {
4975                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4976                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4977         }
4978
4979         /*
4980          * If we are augmenting syscalls, then combine what we put in the
4981          * __augmented_syscalls__ BPF map with what is in the
4982          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4983          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4984          *
4985          * We'll switch to look at two BPF maps, one for sys_enter and the
4986          * other for sys_exit when we start augmenting the sys_exit paths with
4987          * buffers that are being copied from kernel to userspace, think 'read'
4988          * syscall.
4989          */
4990         if (trace.syscalls.events.augmented) {
4991                 evlist__for_each_entry(trace.evlist, evsel) {
4992                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4993
4994                         if (raw_syscalls_sys_exit) {
4995                                 trace.raw_augmented_syscalls = true;
4996                                 goto init_augmented_syscall_tp;
4997                         }
4998
4999                         if (trace.syscalls.events.augmented->priv == NULL &&
5000                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
5001                                 struct evsel *augmented = trace.syscalls.events.augmented;
5002                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
5003                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
5004                                         goto out;
5005                                 /*
5006                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5007                                  * Above we made sure we can get from the payload the tp fields
5008                                  * that we get from syscalls:sys_enter tracefs format file.
5009                                  */
5010                                 augmented->handler = trace__sys_enter;
5011                                 /*
5012                                  * Now we do the same for the *syscalls:sys_enter event so that
5013                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5014                                  * as not to filter it, then we'll handle it just like we would
5015                                  * for the BPF_OUTPUT one:
5016                                  */
5017                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
5018                                     perf_evsel__init_augmented_syscall_tp_args(evsel))
5019                                         goto out;
5020                                 evsel->handler = trace__sys_enter;
5021                         }
5022
5023                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
5024                                 struct syscall_tp *sc;
5025 init_augmented_syscall_tp:
5026                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
5027                                         goto out;
5028                                 sc = __evsel__syscall_tp(evsel);
5029                                 /*
5030                                  * For now with BPF raw_augmented we hook into
5031                                  * raw_syscalls:sys_enter and there we get all
5032                                  * 6 syscall args plus the tracepoint common
5033                                  * fields and the syscall_nr (another long).
5034                                  * So we check if that is the case and if so
5035                                  * don't look after the sc->args_size but
5036                                  * always after the full raw_syscalls:sys_enter
5037                                  * payload, which is fixed.
5038                                  *
5039                                  * We'll revisit this later to pass
5040                                  * s->args_size to the BPF augmenter (now
5041                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5042                                  * so that it copies only what we need for each
5043                                  * syscall, like what happens when we use
5044                                  * syscalls:sys_enter_NAME, so that we reduce
5045                                  * the kernel/userspace traffic to just what is
5046                                  * needed for each syscall.
5047                                  */
5048                                 if (trace.raw_augmented_syscalls)
5049                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5050                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
5051                                 evsel->handler = trace__sys_exit;
5052                         }
5053                 }
5054         }
5055
5056         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5057                 return trace__record(&trace, argc-1, &argv[1]);
5058
5059         /* Using just --errno-summary will trigger --summary */
5060         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5061                 trace.summary_only = true;
5062
5063         /* summary_only implies summary option, but don't overwrite summary if set */
5064         if (trace.summary_only)
5065                 trace.summary = trace.summary_only;
5066
5067         if (output_name != NULL) {
5068                 err = trace__open_output(&trace, output_name);
5069                 if (err < 0) {
5070                         perror("failed to create output file");
5071                         goto out;
5072                 }
5073         }
5074
5075         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5076         if (err)
5077                 goto out_close;
5078
5079         err = target__validate(&trace.opts.target);
5080         if (err) {
5081                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5082                 fprintf(trace.output, "%s", bf);
5083                 goto out_close;
5084         }
5085
5086         err = target__parse_uid(&trace.opts.target);
5087         if (err) {
5088                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5089                 fprintf(trace.output, "%s", bf);
5090                 goto out_close;
5091         }
5092
5093         if (!argc && target__none(&trace.opts.target))
5094                 trace.opts.target.system_wide = true;
5095
5096         if (input_name)
5097                 err = trace__replay(&trace);
5098         else
5099                 err = trace__run(&trace, argc, argv);
5100
5101 out_close:
5102         if (output_name != NULL)
5103                 fclose(trace.output);
5104 out:
5105         zfree(&trace.perfconfig_events);
5106         return err;
5107 }