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 <[email protected]>
  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 <api/fs/tracing_path.h>
  19 #include <bpf/bpf.h>
  20 #include "util/bpf_map.h"
  21 #include "util/rlimit.h"
  22 #include "builtin.h"
  23 #include "util/cgroup.h"
  24 #include "util/color.h"
  25 #include "util/config.h"
  26 #include "util/debug.h"
  27 #include "util/dso.h"
  28 #include "util/env.h"
  29 #include "util/event.h"
  30 #include "util/evsel.h"
  31 #include "util/evsel_fprintf.h"
  32 #include "util/synthetic-events.h"
  33 #include "util/evlist.h"
  34 #include "util/evswitch.h"
  35 #include "util/mmap.h"
  36 #include <subcmd/pager.h>
  37 #include <subcmd/exec-cmd.h>
  38 #include "util/machine.h"
  39 #include "util/map.h"
  40 #include "util/symbol.h"
  41 #include "util/path.h"
  42 #include "util/session.h"
  43 #include "util/thread.h"
  44 #include <subcmd/parse-options.h>
  45 #include "util/strlist.h"
  46 #include "util/intlist.h"
  47 #include "util/thread_map.h"
  48 #include "util/stat.h"
  49 #include "util/tool.h"
  50 #include "util/util.h"
  51 #include "trace/beauty/beauty.h"
  52 #include "trace-event.h"
  53 #include "util/parse-events.h"
  54 #include "util/bpf-loader.h"
  55 #include "util/tracepoint.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 #ifdef HAVE_LIBTRACEEVENT
  83 #include <traceevent/event-parse.h>
  84 #endif
  85
  86 #ifndef O_CLOEXEC
  87 # define O_CLOEXEC              02000000
  88 #endif
  89
  90 #ifndef F_LINUX_SPECIFIC_BASE
  91 # define F_LINUX_SPECIFIC_BASE  1024
  92 #endif
  93
  94 #define RAW_SYSCALL_ARGS_NUM    6
  95
  96 /*
  97  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  98  */
  99 struct syscall_arg_fmt {
 100         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 101         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
 102         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 103         void       *parm;
 104         const char *name;
 105         u16        nr_entries; // for arrays
 106         bool       show_zero;
 107 };
 108
 109 struct syscall_fmt {
 110         const char *name;
 111         const char *alias;
 112         struct {
 113                 const char *sys_enter,
 114                            *sys_exit;
 115         }          bpf_prog_name;
 116         struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
 117         u8         nr_args;
 118         bool       errpid;
 119         bool       timeout;
 120         bool       hexret;
 121 };
 122
 123 struct trace {
 124         struct perf_tool        tool;
 125         struct syscalltbl       *sctbl;
 126         struct {
 127                 struct syscall  *table;
 128                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 129                         struct bpf_map  *sys_enter,
 130                                         *sys_exit;
 131                 }               prog_array;
 132                 struct {
 133                         struct evsel *sys_enter,
 134                                           *sys_exit,
 135                                           *augmented;
 136                 }               events;
 137                 struct bpf_program *unaugmented_prog;
 138         } syscalls;
 139         struct {
 140                 struct bpf_map *map;
 141         } dump;
 142         struct record_opts      opts;
 143         struct evlist   *evlist;
 144         struct machine          *host;
 145         struct thread           *current;
 146         struct bpf_object       *bpf_obj;
 147         struct cgroup           *cgroup;
 148         u64                     base_time;
 149         FILE                    *output;
 150         unsigned long           nr_events;
 151         unsigned long           nr_events_printed;
 152         unsigned long           max_events;
 153         struct evswitch         evswitch;
 154         struct strlist          *ev_qualifier;
 155         struct {
 156                 size_t          nr;
 157                 int             *entries;
 158         }                       ev_qualifier_ids;
 159         struct {
 160                 size_t          nr;
 161                 pid_t           *entries;
 162                 struct bpf_map  *map;
 163         }                       filter_pids;
 164         double                  duration_filter;
 165         double                  runtime_ms;
 166         struct {
 167                 u64             vfs_getname,
 168                                 proc_getname;
 169         } stats;
 170         unsigned int            max_stack;
 171         unsigned int            min_stack;
 172         int                     raw_augmented_syscalls_args_size;
 173         bool                    raw_augmented_syscalls;
 174         bool                    fd_path_disabled;
 175         bool                    sort_events;
 176         bool                    not_ev_qualifier;
 177         bool                    live;
 178         bool                    full_time;
 179         bool                    sched;
 180         bool                    multiple_threads;
 181         bool                    summary;
 182         bool                    summary_only;
 183         bool                    errno_summary;
 184         bool                    failure_only;
 185         bool                    show_comm;
 186         bool                    print_sample;
 187         bool                    show_tool_stats;
 188         bool                    trace_syscalls;
 189         bool                    libtraceevent_print;
 190         bool                    kernel_syscallchains;
 191         s16                     args_alignment;
 192         bool                    show_tstamp;
 193         bool                    show_duration;
 194         bool                    show_zeros;
 195         bool                    show_arg_names;
 196         bool                    show_string_prefix;
 197         bool                    force;
 198         bool                    vfs_getname;
 199         int                     trace_pgfaults;
 200         char                    *perfconfig_events;
 201         struct {
 202                 struct ordered_events   data;
 203                 u64                     last;
 204         } oe;
 205 };
 206
 207 struct tp_field {
 208         int offset;
 209         union {
 210                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 211                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 212         };
 213 };
 214
 215 #define TP_UINT_FIELD(bits) \
 216 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 217 { \
 218         u##bits value; \
 219         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 220         return value;  \
 221 }
 222
 223 TP_UINT_FIELD(8);
 224 TP_UINT_FIELD(16);
 225 TP_UINT_FIELD(32);
 226 TP_UINT_FIELD(64);
 227
 228 #define TP_UINT_FIELD__SWAPPED(bits) \
 229 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 230 { \
 231         u##bits value; \
 232         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 233         return bswap_##bits(value);\
 234 }
 235
 236 TP_UINT_FIELD__SWAPPED(16);
 237 TP_UINT_FIELD__SWAPPED(32);
 238 TP_UINT_FIELD__SWAPPED(64);
 239
 240 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 241 {
 242         field->offset = offset;
 243
 244         switch (size) {
 245         case 1:
 246                 field->integer = tp_field__u8;
 247                 break;
 248         case 2:
 249                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 250                 break;
 251         case 4:
 252                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 253                 break;
 254         case 8:
 255                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 256                 break;
 257         default:
 258                 return -1;
 259         }
 260
 261         return 0;
 262 }
 263
 264 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 265 {
 266         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 267 }
 268
 269 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 270 {
 271         return sample->raw_data + field->offset;
 272 }
 273
 274 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 275 {
 276         field->offset = offset;
 277         field->pointer = tp_field__ptr;
 278         return 0;
 279 }
 280
 281 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 282 {
 283         return __tp_field__init_ptr(field, format_field->offset);
 284 }
 285
 286 struct syscall_tp {
 287         struct tp_field id;
 288         union {
 289                 struct tp_field args, ret;
 290         };
 291 };
 292
 293 /*
 294  * The evsel->priv as used by 'perf trace'
 295  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 296  * fmt: for all the other tracepoints
 297  */
 298 struct evsel_trace {
 299         struct syscall_tp       sc;
 300         struct syscall_arg_fmt  *fmt;
 301 };
 302
 303 static struct evsel_trace *evsel_trace__new(void)
 304 {
 305         return zalloc(sizeof(struct evsel_trace));
 306 }
 307
 308 static void evsel_trace__delete(struct evsel_trace *et)
 309 {
 310         if (et == NULL)
 311                 return;
 312
 313         zfree(&et->fmt);
 314         free(et);
 315 }
 316
 317 /*
 318  * Used with raw_syscalls:sys_{enter,exit} and with the
 319  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 320  */
 321 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 322 {
 323         struct evsel_trace *et = evsel->priv;
 324
 325         return &et->sc;
 326 }
 327
 328 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 329 {
 330         if (evsel->priv == NULL) {
 331                 evsel->priv = evsel_trace__new();
 332                 if (evsel->priv == NULL)
 333                         return NULL;
 334         }
 335
 336         return __evsel__syscall_tp(evsel);
 337 }
 338
 339 /*
 340  * Used with all the other tracepoints.
 341  */
 342 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 343 {
 344         struct evsel_trace *et = evsel->priv;
 345
 346         return et->fmt;
 347 }
 348
 349 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 350 {
 351         struct evsel_trace *et = evsel->priv;
 352
 353         if (evsel->priv == NULL) {
 354                 et = evsel->priv = evsel_trace__new();
 355
 356                 if (et == NULL)
 357                         return NULL;
 358         }
 359
 360         if (et->fmt == NULL) {
 361                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 362                 if (et->fmt == NULL)
 363                         goto out_delete;
 364         }
 365
 366         return __evsel__syscall_arg_fmt(evsel);
 367
 368 out_delete:
 369         evsel_trace__delete(evsel->priv);
 370         evsel->priv = NULL;
 371         return NULL;
 372 }
 373
 374 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 375 {
 376         struct tep_format_field *format_field = evsel__field(evsel, name);
 377
 378         if (format_field == NULL)
 379                 return -1;
 380
 381         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 382 }
 383
 384 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 385         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 386            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 387
 388 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 389 {
 390         struct tep_format_field *format_field = evsel__field(evsel, name);
 391
 392         if (format_field == NULL)
 393                 return -1;
 394
 395         return tp_field__init_ptr(field, format_field);
 396 }
 397
 398 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 399         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 400            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 401
 402 static void evsel__delete_priv(struct evsel *evsel)
 403 {
 404         zfree(&evsel->priv);
 405         evsel__delete(evsel);
 406 }
 407
 408 static int evsel__init_syscall_tp(struct evsel *evsel)
 409 {
 410         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 411
 412         if (sc != NULL) {
 413                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 414                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 415                         return -ENOENT;
 416                 return 0;
 417         }
 418
 419         return -ENOMEM;
 420 }
 421
 422 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 423 {
 424         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 425
 426         if (sc != NULL) {
 427                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 428                 if (syscall_id == NULL)
 429                         syscall_id = evsel__field(tp, "__syscall_nr");
 430                 if (syscall_id == NULL ||
 431                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 432                         return -EINVAL;
 433
 434                 return 0;
 435         }
 436
 437         return -ENOMEM;
 438 }
 439
 440 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 441 {
 442         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 443
 444         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 445 }
 446
 447 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 448 {
 449         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 450
 451         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 452 }
 453
 454 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 455 {
 456         if (evsel__syscall_tp(evsel) != NULL) {
 457                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 458                         return -ENOENT;
 459
 460                 evsel->handler = handler;
 461                 return 0;
 462         }
 463
 464         return -ENOMEM;
 465 }
 466
 467 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 468 {
 469         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 470
 471         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 472         if (IS_ERR(evsel))
 473                 evsel = evsel__newtp("syscalls", direction);
 474
 475         if (IS_ERR(evsel))
 476                 return NULL;
 477
 478         if (evsel__init_raw_syscall_tp(evsel, handler))
 479                 goto out_delete;
 480
 481         return evsel;
 482
 483 out_delete:
 484         evsel__delete_priv(evsel);
 485         return NULL;
 486 }
 487
 488 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 489         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 490            fields->name.integer(&fields->name, sample); })
 491
 492 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 493         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 494            fields->name.pointer(&fields->name, sample); })
 495
 496 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 497 {
 498         int idx = val - sa->offset;
 499
 500         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 501                 size_t printed = scnprintf(bf, size, intfmt, val);
 502                 if (show_suffix)
 503                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 504                 return printed;
 505         }
 506
 507         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 508 }
 509
 510 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 511 {
 512         int idx = val - sa->offset;
 513
 514         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 515                 size_t printed = scnprintf(bf, size, intfmt, val);
 516                 if (show_prefix)
 517                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 518                 return printed;
 519         }
 520
 521         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 522 }
 523
 524 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 525                                                 const char *intfmt,
 526                                                 struct syscall_arg *arg)
 527 {
 528         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 529 }
 530
 531 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 532                                               struct syscall_arg *arg)
 533 {
 534         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 535 }
 536
 537 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 538
 539 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 540 {
 541         return strarray__strtoul(arg->parm, bf, size, ret);
 542 }
 543
 544 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 545 {
 546         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 547 }
 548
 549 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 550 {
 551         return strarrays__strtoul(arg->parm, bf, size, ret);
 552 }
 553
 554 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 555 {
 556         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 557 }
 558
 559 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 560 {
 561         size_t printed;
 562         int i;
 563
 564         for (i = 0; i < sas->nr_entries; ++i) {
 565                 struct strarray *sa = sas->entries[i];
 566                 int idx = val - sa->offset;
 567
 568                 if (idx >= 0 && idx < sa->nr_entries) {
 569                         if (sa->entries[idx] == NULL)
 570                                 break;
 571                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 572                 }
 573         }
 574
 575         printed = scnprintf(bf, size, intfmt, val);
 576         if (show_prefix)
 577                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 578         return printed;
 579 }
 580
 581 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 582 {
 583         int i;
 584
 585         for (i = 0; i < sa->nr_entries; ++i) {
 586                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 587                         *ret = sa->offset + i;
 588                         return true;
 589                 }
 590         }
 591
 592         return false;
 593 }
 594
 595 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 596 {
 597         u64 val = 0;
 598         char *tok = bf, *sep, *end;
 599
 600         *ret = 0;
 601
 602         while (size != 0) {
 603                 int toklen = size;
 604
 605                 sep = memchr(tok, '|', size);
 606                 if (sep != NULL) {
 607                         size -= sep - tok + 1;
 608
 609                         end = sep - 1;
 610                         while (end > tok && isspace(*end))
 611                                 --end;
 612
 613                         toklen = end - tok + 1;
 614                 }
 615
 616                 while (isspace(*tok))
 617                         ++tok;
 618
 619                 if (isalpha(*tok) || *tok == '_') {
 620                         if (!strarray__strtoul(sa, tok, toklen, &val))
 621                                 return false;
 622                 } else
 623                         val = strtoul(tok, NULL, 0);
 624
 625                 *ret |= (1 << (val - 1));
 626
 627                 if (sep == NULL)
 628                         break;
 629                 tok = sep + 1;
 630         }
 631
 632         return true;
 633 }
 634
 635 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 636 {
 637         int i;
 638
 639         for (i = 0; i < sas->nr_entries; ++i) {
 640                 struct strarray *sa = sas->entries[i];
 641
 642                 if (strarray__strtoul(sa, bf, size, ret))
 643                         return true;
 644         }
 645
 646         return false;
 647 }
 648
 649 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 650                                         struct syscall_arg *arg)
 651 {
 652         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 653 }
 654
 655 #ifndef AT_FDCWD
 656 #define AT_FDCWD        -100
 657 #endif
 658
 659 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 660                                            struct syscall_arg *arg)
 661 {
 662         int fd = arg->val;
 663         const char *prefix = "AT_FD";
 664
 665         if (fd == AT_FDCWD)
 666                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 667
 668         return syscall_arg__scnprintf_fd(bf, size, arg);
 669 }
 670
 671 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 672
 673 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 674                                               struct syscall_arg *arg);
 675
 676 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 677
 678 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 679 {
 680         return scnprintf(bf, size, "%#lx", arg->val);
 681 }
 682
 683 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 684 {
 685         if (arg->val == 0)
 686                 return scnprintf(bf, size, "NULL");
 687         return syscall_arg__scnprintf_hex(bf, size, arg);
 688 }
 689
 690 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 691 {
 692         return scnprintf(bf, size, "%d", arg->val);
 693 }
 694
 695 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 696 {
 697         return scnprintf(bf, size, "%ld", arg->val);
 698 }
 699
 700 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 701 {
 702         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 703         //     fill missing comms using thread__set_comm()...
 704         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 705         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 706 }
 707
 708 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 709
 710 static const char *bpf_cmd[] = {
 711         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 712         "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 713         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 714         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 715         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 716         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 717         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 718         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 719         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 720         "LINK_DETACH", "PROG_BIND_MAP",
 721 };
 722 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 723
 724 static const char *fsmount_flags[] = {
 725         [1] = "CLOEXEC",
 726 };
 727 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 728
 729 #include "trace/beauty/generated/fsconfig_arrays.c"
 730
 731 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 732
 733 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 734 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 735
 736 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 737 static DEFINE_STRARRAY(itimers, "ITIMER_");
 738
 739 static const char *keyctl_options[] = {
 740         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 741         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 742         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 743         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 744         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 745 };
 746 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 747
 748 static const char *whences[] = { "SET", "CUR", "END",
 749 #ifdef SEEK_DATA
 750 "DATA",
 751 #endif
 752 #ifdef SEEK_HOLE
 753 "HOLE",
 754 #endif
 755 };
 756 static DEFINE_STRARRAY(whences, "SEEK_");
 757
 758 static const char *fcntl_cmds[] = {
 759         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 760         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 761         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 762         "GETOWNER_UIDS",
 763 };
 764 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 765
 766 static const char *fcntl_linux_specific_cmds[] = {
 767         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 768         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 769         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 770 };
 771
 772 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 773
 774 static struct strarray *fcntl_cmds_arrays[] = {
 775         &strarray__fcntl_cmds,
 776         &strarray__fcntl_linux_specific_cmds,
 777 };
 778
 779 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 780
 781 static const char *rlimit_resources[] = {
 782         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 783         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 784         "RTTIME",
 785 };
 786 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 787
 788 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 789 static DEFINE_STRARRAY(sighow, "SIG_");
 790
 791 static const char *clockid[] = {
 792         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 793         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 794         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 795 };
 796 static DEFINE_STRARRAY(clockid, "CLOCK_");
 797
 798 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 799                                                  struct syscall_arg *arg)
 800 {
 801         bool show_prefix = arg->show_string_prefix;
 802         const char *suffix = "_OK";
 803         size_t printed = 0;
 804         int mode = arg->val;
 805
 806         if (mode == F_OK) /* 0 */
 807                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 808 #define P_MODE(n) \
 809         if (mode & n##_OK) { \
 810                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 811                 mode &= ~n##_OK; \
 812         }
 813
 814         P_MODE(R);
 815         P_MODE(W);
 816         P_MODE(X);
 817 #undef P_MODE
 818
 819         if (mode)
 820                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 821
 822         return printed;
 823 }
 824
 825 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 826
 827 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 828                                               struct syscall_arg *arg);
 829
 830 #define SCA_FILENAME syscall_arg__scnprintf_filename
 831
 832 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 833                                                 struct syscall_arg *arg)
 834 {
 835         bool show_prefix = arg->show_string_prefix;
 836         const char *prefix = "O_";
 837         int printed = 0, flags = arg->val;
 838
 839 #define P_FLAG(n) \
 840         if (flags & O_##n) { \
 841                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 842                 flags &= ~O_##n; \
 843         }
 844
 845         P_FLAG(CLOEXEC);
 846         P_FLAG(NONBLOCK);
 847 #undef P_FLAG
 848
 849         if (flags)
 850                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 851
 852         return printed;
 853 }
 854
 855 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 856
 857 #ifndef GRND_NONBLOCK
 858 #define GRND_NONBLOCK   0x0001
 859 #endif
 860 #ifndef GRND_RANDOM
 861 #define GRND_RANDOM     0x0002
 862 #endif
 863
 864 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 865                                                    struct syscall_arg *arg)
 866 {
 867         bool show_prefix = arg->show_string_prefix;
 868         const char *prefix = "GRND_";
 869         int printed = 0, flags = arg->val;
 870
 871 #define P_FLAG(n) \
 872         if (flags & GRND_##n) { \
 873                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 874                 flags &= ~GRND_##n; \
 875         }
 876
 877         P_FLAG(RANDOM);
 878         P_FLAG(NONBLOCK);
 879 #undef P_FLAG
 880
 881         if (flags)
 882                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 883
 884         return printed;
 885 }
 886
 887 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 888
 889 #define STRARRAY(name, array) \
 890           { .scnprintf  = SCA_STRARRAY, \
 891             .strtoul    = STUL_STRARRAY, \
 892             .parm       = &strarray__##array, }
 893
 894 #define STRARRAY_FLAGS(name, array) \
 895           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 896             .strtoul    = STUL_STRARRAY_FLAGS, \
 897             .parm       = &strarray__##array, }
 898
 899 #include "trace/beauty/arch_errno_names.c"
 900 #include "trace/beauty/eventfd.c"
 901 #include "trace/beauty/futex_op.c"
 902 #include "trace/beauty/futex_val3.c"
 903 #include "trace/beauty/mmap.c"
 904 #include "trace/beauty/mode_t.c"
 905 #include "trace/beauty/msg_flags.c"
 906 #include "trace/beauty/open_flags.c"
 907 #include "trace/beauty/perf_event_open.c"
 908 #include "trace/beauty/pid.c"
 909 #include "trace/beauty/sched_policy.c"
 910 #include "trace/beauty/seccomp.c"
 911 #include "trace/beauty/signum.c"
 912 #include "trace/beauty/socket_type.c"
 913 #include "trace/beauty/waitid_options.c"
 914
 915 static struct syscall_fmt syscall_fmts[] = {
 916         { .name     = "access",
 917           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 918         { .name     = "arch_prctl",
 919           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 920                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 921         { .name     = "bind",
 922           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 923                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 924                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 925         { .name     = "bpf",
 926           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 927         { .name     = "brk",        .hexret = true,
 928           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 929         { .name     = "clock_gettime",
 930           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 931         { .name     = "clock_nanosleep",
 932           .arg = { [2] = { .scnprintf = SCA_TIMESPEC,  /* rqtp */ }, }, },
 933         { .name     = "clone",      .errpid = true, .nr_args = 5,
 934           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 935                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 936                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 937                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 938                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 939         { .name     = "close",
 940           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 941         { .name     = "connect",
 942           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 943                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 944                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 945         { .name     = "epoll_ctl",
 946           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 947         { .name     = "eventfd2",
 948           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 949         { .name     = "fchmodat",
 950           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 951         { .name     = "fchownat",
 952           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 953         { .name     = "fcntl",
 954           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 955                            .strtoul   = STUL_STRARRAYS,
 956                            .parm      = &strarrays__fcntl_cmds_arrays,
 957                            .show_zero = true, },
 958                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 959         { .name     = "flock",
 960           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 961         { .name     = "fsconfig",
 962           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 963         { .name     = "fsmount",
 964           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 965                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 966         { .name     = "fspick",
 967           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 968                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 969                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 970         { .name     = "fstat", .alias = "newfstat", },
 971         { .name     = "fstatat", .alias = "newfstatat", },
 972         { .name     = "futex",
 973           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 974                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 975         { .name     = "futimesat",
 976           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 977         { .name     = "getitimer",
 978           .arg = { [0] = STRARRAY(which, itimers), }, },
 979         { .name     = "getpid",     .errpid = true, },
 980         { .name     = "getpgid",    .errpid = true, },
 981         { .name     = "getppid",    .errpid = true, },
 982         { .name     = "getrandom",
 983           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 984         { .name     = "getrlimit",
 985           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 986         { .name     = "getsockopt",
 987           .arg = { [1] = STRARRAY(level, socket_level), }, },
 988         { .name     = "gettid",     .errpid = true, },
 989         { .name     = "ioctl",
 990           .arg = {
 991 #if defined(__i386__) || defined(__x86_64__)
 992 /*
 993  * FIXME: Make this available to all arches.
 994  */
 995                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 996                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 997 #else
 998                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 999 #endif
1000         { .name     = "kcmp",       .nr_args = 5,
1001           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
1002                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
1003                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
1004                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
1005                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1006         { .name     = "keyctl",
1007           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1008         { .name     = "kill",
1009           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1010         { .name     = "linkat",
1011           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1012         { .name     = "lseek",
1013           .arg = { [2] = STRARRAY(whence, whences), }, },
1014         { .name     = "lstat", .alias = "newlstat", },
1015         { .name     = "madvise",
1016           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1017                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1018         { .name     = "mkdirat",
1019           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1020         { .name     = "mknodat",
1021           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1022         { .name     = "mmap",       .hexret = true,
1023 /* The standard mmap maps to old_mmap on s390x */
1024 #if defined(__s390x__)
1025         .alias = "old_mmap",
1026 #endif
1027           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1028                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1029                            .strtoul   = STUL_STRARRAY_FLAGS,
1030                            .parm      = &strarray__mmap_flags, },
1031                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1032         { .name     = "mount",
1033           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1034                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1035                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1036         { .name     = "move_mount",
1037           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1038                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1039                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1040                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1041                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1042         { .name     = "mprotect",
1043           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1044                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1045         { .name     = "mq_unlink",
1046           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1047         { .name     = "mremap",     .hexret = true,
1048           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1049         { .name     = "name_to_handle_at",
1050           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1051         { .name     = "newfstatat",
1052           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1053         { .name     = "open",
1054           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1055         { .name     = "open_by_handle_at",
1056           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1057                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1058         { .name     = "openat",
1059           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1060                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1061         { .name     = "perf_event_open",
1062           .arg = { [0] = { .scnprintf = SCA_PERF_ATTR,  /* attr */ },
1063                    [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1064                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1065                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1066         { .name     = "pipe2",
1067           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1068         { .name     = "pkey_alloc",
1069           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1070         { .name     = "pkey_free",
1071           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1072         { .name     = "pkey_mprotect",
1073           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1074                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1075                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1076         { .name     = "poll", .timeout = true, },
1077         { .name     = "ppoll", .timeout = true, },
1078         { .name     = "prctl",
1079           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1080                            .strtoul   = STUL_STRARRAY,
1081                            .parm      = &strarray__prctl_options, },
1082                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1083                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1084         { .name     = "pread", .alias = "pread64", },
1085         { .name     = "preadv", .alias = "pread", },
1086         { .name     = "prlimit64",
1087           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1088         { .name     = "pwrite", .alias = "pwrite64", },
1089         { .name     = "readlinkat",
1090           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1091         { .name     = "recvfrom",
1092           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1093         { .name     = "recvmmsg",
1094           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1095         { .name     = "recvmsg",
1096           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1097         { .name     = "renameat",
1098           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1099                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1100         { .name     = "renameat2",
1101           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1102                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1103                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1104         { .name     = "rt_sigaction",
1105           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1106         { .name     = "rt_sigprocmask",
1107           .arg = { [0] = STRARRAY(how, sighow), }, },
1108         { .name     = "rt_sigqueueinfo",
1109           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1110         { .name     = "rt_tgsigqueueinfo",
1111           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1112         { .name     = "sched_setscheduler",
1113           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1114         { .name     = "seccomp",
1115           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1116                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1117         { .name     = "select", .timeout = true, },
1118         { .name     = "sendfile", .alias = "sendfile64", },
1119         { .name     = "sendmmsg",
1120           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1121         { .name     = "sendmsg",
1122           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1123         { .name     = "sendto",
1124           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1125                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1126         { .name     = "set_tid_address", .errpid = true, },
1127         { .name     = "setitimer",
1128           .arg = { [0] = STRARRAY(which, itimers), }, },
1129         { .name     = "setrlimit",
1130           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1131         { .name     = "setsockopt",
1132           .arg = { [1] = STRARRAY(level, socket_level), }, },
1133         { .name     = "socket",
1134           .arg = { [0] = STRARRAY(family, socket_families),
1135                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1136                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1137         { .name     = "socketpair",
1138           .arg = { [0] = STRARRAY(family, socket_families),
1139                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1140                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1141         { .name     = "stat", .alias = "newstat", },
1142         { .name     = "statx",
1143           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1144                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1145                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1146         { .name     = "swapoff",
1147           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1148         { .name     = "swapon",
1149           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1150         { .name     = "symlinkat",
1151           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1152         { .name     = "sync_file_range",
1153           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1154         { .name     = "tgkill",
1155           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1156         { .name     = "tkill",
1157           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1158         { .name     = "umount2", .alias = "umount",
1159           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1160         { .name     = "uname", .alias = "newuname", },
1161         { .name     = "unlinkat",
1162           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1163         { .name     = "utimensat",
1164           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1165         { .name     = "wait4",      .errpid = true,
1166           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1167         { .name     = "waitid",     .errpid = true,
1168           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1169 };
1170
1171 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1172 {
1173         const struct syscall_fmt *fmt = fmtp;
1174         return strcmp(name, fmt->name);
1175 }
1176
1177 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1178 {
1179         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1180 }
1181
1182 static struct syscall_fmt *syscall_fmt__find(const char *name)
1183 {
1184         const int nmemb = ARRAY_SIZE(syscall_fmts);
1185         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1186 }
1187
1188 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1189 {
1190         int i;
1191
1192         for (i = 0; i < nmemb; ++i) {
1193                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1194                         return &fmts[i];
1195         }
1196
1197         return NULL;
1198 }
1199
1200 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1201 {
1202         const int nmemb = ARRAY_SIZE(syscall_fmts);
1203         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1204 }
1205
1206 /*
1207  * is_exit: is this "exit" or "exit_group"?
1208  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1209  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1210  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1211  */
1212 struct syscall {
1213         struct tep_event    *tp_format;
1214         int                 nr_args;
1215         int                 args_size;
1216         struct {
1217                 struct bpf_program *sys_enter,
1218                                    *sys_exit;
1219         }                   bpf_prog;
1220         bool                is_exit;
1221         bool                is_open;
1222         bool                nonexistent;
1223         struct tep_format_field *args;
1224         const char          *name;
1225         struct syscall_fmt  *fmt;
1226         struct syscall_arg_fmt *arg_fmt;
1227 };
1228
1229 /*
1230  * We need to have this 'calculated' boolean because in some cases we really
1231  * don't know what is the duration of a syscall, for instance, when we start
1232  * a session and some threads are waiting for a syscall to finish, say 'poll',
1233  * in which case all we can do is to print "( ? ) for duration and for the
1234  * start timestamp.
1235  */
1236 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1237 {
1238         double duration = (double)t / NSEC_PER_MSEC;
1239         size_t printed = fprintf(fp, "(");
1240
1241         if (!calculated)
1242                 printed += fprintf(fp, "         ");
1243         else if (duration >= 1.0)
1244                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1245         else if (duration >= 0.01)
1246                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1247         else
1248                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1249         return printed + fprintf(fp, "): ");
1250 }
1251
1252 /**
1253  * filename.ptr: The filename char pointer that will be vfs_getname'd
1254  * filename.entry_str_pos: Where to insert the string translated from
1255  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1256  * ret_scnprintf: syscall args may set this to a different syscall return
1257  *                formatter, for instance, fcntl may return fds, file flags, etc.
1258  */
1259 struct thread_trace {
1260         u64               entry_time;
1261         bool              entry_pending;
1262         unsigned long     nr_events;
1263         unsigned long     pfmaj, pfmin;
1264         char              *entry_str;
1265         double            runtime_ms;
1266         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1267         struct {
1268                 unsigned long ptr;
1269                 short int     entry_str_pos;
1270                 bool          pending_open;
1271                 unsigned int  namelen;
1272                 char          *name;
1273         } filename;
1274         struct {
1275                 int           max;
1276                 struct file   *table;
1277         } files;
1278
1279         struct intlist *syscall_stats;
1280 };
1281
1282 static struct thread_trace *thread_trace__new(void)
1283 {
1284         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1285
1286         if (ttrace) {
1287                 ttrace->files.max = -1;
1288                 ttrace->syscall_stats = intlist__new(NULL);
1289         }
1290
1291         return ttrace;
1292 }
1293
1294 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1295 {
1296         struct thread_trace *ttrace;
1297
1298         if (thread == NULL)
1299                 goto fail;
1300
1301         if (thread__priv(thread) == NULL)
1302                 thread__set_priv(thread, thread_trace__new());
1303
1304         if (thread__priv(thread) == NULL)
1305                 goto fail;
1306
1307         ttrace = thread__priv(thread);
1308         ++ttrace->nr_events;
1309
1310         return ttrace;
1311 fail:
1312         color_fprintf(fp, PERF_COLOR_RED,
1313                       "WARNING: not enough memory, dropping samples!\n");
1314         return NULL;
1315 }
1316
1317
1318 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1319                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1320 {
1321         struct thread_trace *ttrace = thread__priv(arg->thread);
1322
1323         ttrace->ret_scnprintf = ret_scnprintf;
1324 }
1325
1326 #define TRACE_PFMAJ             (1 << 0)
1327 #define TRACE_PFMIN             (1 << 1)
1328
1329 static const size_t trace__entry_str_size = 2048;
1330
1331 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1332 {
1333         if (fd < 0)
1334                 return NULL;
1335
1336         if (fd > ttrace->files.max) {
1337                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1338
1339                 if (nfiles == NULL)
1340                         return NULL;
1341
1342                 if (ttrace->files.max != -1) {
1343                         memset(nfiles + ttrace->files.max + 1, 0,
1344                                (fd - ttrace->files.max) * sizeof(struct file));
1345                 } else {
1346                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1347                 }
1348
1349                 ttrace->files.table = nfiles;
1350                 ttrace->files.max   = fd;
1351         }
1352
1353         return ttrace->files.table + fd;
1354 }
1355
1356 struct file *thread__files_entry(struct thread *thread, int fd)
1357 {
1358         return thread_trace__files_entry(thread__priv(thread), fd);
1359 }
1360
1361 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1362 {
1363         struct thread_trace *ttrace = thread__priv(thread);
1364         struct file *file = thread_trace__files_entry(ttrace, fd);
1365
1366         if (file != NULL) {
1367                 struct stat st;
1368                 if (stat(pathname, &st) == 0)
1369                         file->dev_maj = major(st.st_rdev);
1370                 file->pathname = strdup(pathname);
1371                 if (file->pathname)
1372                         return 0;
1373         }
1374
1375         return -1;
1376 }
1377
1378 static int thread__read_fd_path(struct thread *thread, int fd)
1379 {
1380         char linkname[PATH_MAX], pathname[PATH_MAX];
1381         struct stat st;
1382         int ret;
1383
1384         if (thread->pid_ == thread->tid) {
1385                 scnprintf(linkname, sizeof(linkname),
1386                           "/proc/%d/fd/%d", thread->pid_, fd);
1387         } else {
1388                 scnprintf(linkname, sizeof(linkname),
1389                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1390         }
1391
1392         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1393                 return -1;
1394
1395         ret = readlink(linkname, pathname, sizeof(pathname));
1396
1397         if (ret < 0 || ret > st.st_size)
1398                 return -1;
1399
1400         pathname[ret] = '\0';
1401         return trace__set_fd_pathname(thread, fd, pathname);
1402 }
1403
1404 static const char *thread__fd_path(struct thread *thread, int fd,
1405                                    struct trace *trace)
1406 {
1407         struct thread_trace *ttrace = thread__priv(thread);
1408
1409         if (ttrace == NULL || trace->fd_path_disabled)
1410                 return NULL;
1411
1412         if (fd < 0)
1413                 return NULL;
1414
1415         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1416                 if (!trace->live)
1417                         return NULL;
1418                 ++trace->stats.proc_getname;
1419                 if (thread__read_fd_path(thread, fd))
1420                         return NULL;
1421         }
1422
1423         return ttrace->files.table[fd].pathname;
1424 }
1425
1426 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1427 {
1428         int fd = arg->val;
1429         size_t printed = scnprintf(bf, size, "%d", fd);
1430         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1431
1432         if (path)
1433                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1434
1435         return printed;
1436 }
1437
1438 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1439 {
1440         size_t printed = scnprintf(bf, size, "%d", fd);
1441         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1442
1443         if (thread) {
1444                 const char *path = thread__fd_path(thread, fd, trace);
1445
1446                 if (path)
1447                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1448
1449                 thread__put(thread);
1450         }
1451
1452         return printed;
1453 }
1454
1455 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1456                                               struct syscall_arg *arg)
1457 {
1458         int fd = arg->val;
1459         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1460         struct thread_trace *ttrace = thread__priv(arg->thread);
1461
1462         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1463                 zfree(&ttrace->files.table[fd].pathname);
1464
1465         return printed;
1466 }
1467
1468 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1469                                      unsigned long ptr)
1470 {
1471         struct thread_trace *ttrace = thread__priv(thread);
1472
1473         ttrace->filename.ptr = ptr;
1474         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1475 }
1476
1477 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1478 {
1479         struct augmented_arg *augmented_arg = arg->augmented.args;
1480         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1481         /*
1482          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1483          * we would have two strings, each prefixed by its size.
1484          */
1485         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1486
1487         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1488         arg->augmented.size -= consumed;
1489
1490         return printed;
1491 }
1492
1493 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1494                                               struct syscall_arg *arg)
1495 {
1496         unsigned long ptr = arg->val;
1497
1498         if (arg->augmented.args)
1499                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1500
1501         if (!arg->trace->vfs_getname)
1502                 return scnprintf(bf, size, "%#x", ptr);
1503
1504         thread__set_filename_pos(arg->thread, bf, ptr);
1505         return 0;
1506 }
1507
1508 static bool trace__filter_duration(struct trace *trace, double t)
1509 {
1510         return t < (trace->duration_filter * NSEC_PER_MSEC);
1511 }
1512
1513 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1514 {
1515         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1516
1517         return fprintf(fp, "%10.3f ", ts);
1518 }
1519
1520 /*
1521  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1522  * using ttrace->entry_time for a thread that receives a sys_exit without
1523  * first having received a sys_enter ("poll" issued before tracing session
1524  * starts, lost sys_enter exit due to ring buffer overflow).
1525  */
1526 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1527 {
1528         if (tstamp > 0)
1529                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1530
1531         return fprintf(fp, "         ? ");
1532 }
1533
1534 static pid_t workload_pid = -1;
1535 static volatile sig_atomic_t done = false;
1536 static volatile sig_atomic_t interrupted = false;
1537
1538 static void sighandler_interrupt(int sig __maybe_unused)
1539 {
1540         done = interrupted = true;
1541 }
1542
1543 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1544                             void *context __maybe_unused)
1545 {
1546         if (info->si_pid == workload_pid)
1547                 done = true;
1548 }
1549
1550 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1551 {
1552         size_t printed = 0;
1553
1554         if (trace->multiple_threads) {
1555                 if (trace->show_comm)
1556                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1557                 printed += fprintf(fp, "%d ", thread->tid);
1558         }
1559
1560         return printed;
1561 }
1562
1563 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1564                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1565 {
1566         size_t printed = 0;
1567
1568         if (trace->show_tstamp)
1569                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1570         if (trace->show_duration)
1571                 printed += fprintf_duration(duration, duration_calculated, fp);
1572         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1573 }
1574
1575 static int trace__process_event(struct trace *trace, struct machine *machine,
1576                                 union perf_event *event, struct perf_sample *sample)
1577 {
1578         int ret = 0;
1579
1580         switch (event->header.type) {
1581         case PERF_RECORD_LOST:
1582                 color_fprintf(trace->output, PERF_COLOR_RED,
1583                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1584                 ret = machine__process_lost_event(machine, event, sample);
1585                 break;
1586         default:
1587                 ret = machine__process_event(machine, event, sample);
1588                 break;
1589         }
1590
1591         return ret;
1592 }
1593
1594 static int trace__tool_process(struct perf_tool *tool,
1595                                union perf_event *event,
1596                                struct perf_sample *sample,
1597                                struct machine *machine)
1598 {
1599         struct trace *trace = container_of(tool, struct trace, tool);
1600         return trace__process_event(trace, machine, event, sample);
1601 }
1602
1603 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1604 {
1605         struct machine *machine = vmachine;
1606
1607         if (machine->kptr_restrict_warned)
1608                 return NULL;
1609
1610         if (symbol_conf.kptr_restrict) {
1611                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1612                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1613                            "Kernel samples will not be resolved.\n");
1614                 machine->kptr_restrict_warned = true;
1615                 return NULL;
1616         }
1617
1618         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1619 }
1620
1621 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1622 {
1623         int err = symbol__init(NULL);
1624
1625         if (err)
1626                 return err;
1627
1628         trace->host = machine__new_host();
1629         if (trace->host == NULL)
1630                 return -ENOMEM;
1631
1632         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1633         if (err < 0)
1634                 goto out;
1635
1636         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1637                                             evlist->core.threads, trace__tool_process,
1638                                             true, false, 1);
1639 out:
1640         if (err)
1641                 symbol__exit();
1642
1643         return err;
1644 }
1645
1646 static void trace__symbols__exit(struct trace *trace)
1647 {
1648         machine__exit(trace->host);
1649         trace->host = NULL;
1650
1651         symbol__exit();
1652 }
1653
1654 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1655 {
1656         int idx;
1657
1658         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1659                 nr_args = sc->fmt->nr_args;
1660
1661         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1662         if (sc->arg_fmt == NULL)
1663                 return -1;
1664
1665         for (idx = 0; idx < nr_args; ++idx) {
1666                 if (sc->fmt)
1667                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1668         }
1669
1670         sc->nr_args = nr_args;
1671         return 0;
1672 }
1673
1674 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1675         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1676         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1677 };
1678
1679 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1680 {
1681        const struct syscall_arg_fmt *fmt = fmtp;
1682        return strcmp(name, fmt->name);
1683 }
1684
1685 static struct syscall_arg_fmt *
1686 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1687 {
1688        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1689 }
1690
1691 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1692 {
1693        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1694        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1695 }
1696
1697 static struct tep_format_field *
1698 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1699 {
1700         struct tep_format_field *last_field = NULL;
1701         int len;
1702
1703         for (; field; field = field->next, ++arg) {
1704                 last_field = field;
1705
1706                 if (arg->scnprintf)
1707                         continue;
1708
1709                 len = strlen(field->name);
1710
1711                 if (strcmp(field->type, "const char *") == 0 &&
1712                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1713                      strstr(field->name, "path") != NULL))
1714                         arg->scnprintf = SCA_FILENAME;
1715                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1716                         arg->scnprintf = SCA_PTR;
1717                 else if (strcmp(field->type, "pid_t") == 0)
1718                         arg->scnprintf = SCA_PID;
1719                 else if (strcmp(field->type, "umode_t") == 0)
1720                         arg->scnprintf = SCA_MODE_T;
1721                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1722                         arg->scnprintf = SCA_CHAR_ARRAY;
1723                         arg->nr_entries = field->arraylen;
1724                 } else if ((strcmp(field->type, "int") == 0 ||
1725                           strcmp(field->type, "unsigned int") == 0 ||
1726                           strcmp(field->type, "long") == 0) &&
1727                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1728                         /*
1729                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1730                          * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1731                          * 65 int
1732                          * 23 unsigned int
1733                          * 7 unsigned long
1734                          */
1735                         arg->scnprintf = SCA_FD;
1736                } else {
1737                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1738
1739                         if (fmt) {
1740                                 arg->scnprintf = fmt->scnprintf;
1741                                 arg->strtoul   = fmt->strtoul;
1742                         }
1743                 }
1744         }
1745
1746         return last_field;
1747 }
1748
1749 static int syscall__set_arg_fmts(struct syscall *sc)
1750 {
1751         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1752
1753         if (last_field)
1754                 sc->args_size = last_field->offset + last_field->size;
1755
1756         return 0;
1757 }
1758
1759 static int trace__read_syscall_info(struct trace *trace, int id)
1760 {
1761         char tp_name[128];
1762         struct syscall *sc;
1763         const char *name = syscalltbl__name(trace->sctbl, id);
1764
1765 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1766         if (trace->syscalls.table == NULL) {
1767                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1768                 if (trace->syscalls.table == NULL)
1769                         return -ENOMEM;
1770         }
1771 #else
1772         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1773                 // When using libaudit we don't know beforehand what is the max syscall id
1774                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1775
1776                 if (table == NULL)
1777                         return -ENOMEM;
1778
1779                 // Need to memset from offset 0 and +1 members if brand new
1780                 if (trace->syscalls.table == NULL)
1781                         memset(table, 0, (id + 1) * sizeof(*sc));
1782                 else
1783                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1784
1785                 trace->syscalls.table         = table;
1786                 trace->sctbl->syscalls.max_id = id;
1787         }
1788 #endif
1789         sc = trace->syscalls.table + id;
1790         if (sc->nonexistent)
1791                 return -EEXIST;
1792
1793         if (name == NULL) {
1794                 sc->nonexistent = true;
1795                 return -EEXIST;
1796         }
1797
1798         sc->name = name;
1799         sc->fmt  = syscall_fmt__find(sc->name);
1800
1801         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1802         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1803
1804         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1805                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1806                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1807         }
1808
1809         /*
1810          * Fails to read trace point format via sysfs node, so the trace point
1811          * doesn't exist.  Set the 'nonexistent' flag as true.
1812          */
1813         if (IS_ERR(sc->tp_format)) {
1814                 sc->nonexistent = true;
1815                 return PTR_ERR(sc->tp_format);
1816         }
1817
1818         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1819                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1820                 return -ENOMEM;
1821
1822         sc->args = sc->tp_format->format.fields;
1823         /*
1824          * We need to check and discard the first variable '__syscall_nr'
1825          * or 'nr' that mean the syscall number. It is needless here.
1826          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1827          */
1828         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1829                 sc->args = sc->args->next;
1830                 --sc->nr_args;
1831         }
1832
1833         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1834         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1835
1836         return syscall__set_arg_fmts(sc);
1837 }
1838
1839 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1840 {
1841         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1842
1843         if (fmt != NULL) {
1844                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1845                 return 0;
1846         }
1847
1848         return -ENOMEM;
1849 }
1850
1851 static int intcmp(const void *a, const void *b)
1852 {
1853         const int *one = a, *another = b;
1854
1855         return *one - *another;
1856 }
1857
1858 static int trace__validate_ev_qualifier(struct trace *trace)
1859 {
1860         int err = 0;
1861         bool printed_invalid_prefix = false;
1862         struct str_node *pos;
1863         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1864
1865         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1866                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1867
1868         if (trace->ev_qualifier_ids.entries == NULL) {
1869                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1870                        trace->output);
1871                 err = -EINVAL;
1872                 goto out;
1873         }
1874
1875         strlist__for_each_entry(pos, trace->ev_qualifier) {
1876                 const char *sc = pos->s;
1877                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1878
1879                 if (id < 0) {
1880                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1881                         if (id >= 0)
1882                                 goto matches;
1883
1884                         if (!printed_invalid_prefix) {
1885                                 pr_debug("Skipping unknown syscalls: ");
1886                                 printed_invalid_prefix = true;
1887                         } else {
1888                                 pr_debug(", ");
1889                         }
1890
1891                         pr_debug("%s", sc);
1892                         continue;
1893                 }
1894 matches:
1895                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1896                 if (match_next == -1)
1897                         continue;
1898
1899                 while (1) {
1900                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1901                         if (id < 0)
1902                                 break;
1903                         if (nr_allocated == nr_used) {
1904                                 void *entries;
1905
1906                                 nr_allocated += 8;
1907                                 entries = realloc(trace->ev_qualifier_ids.entries,
1908                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1909                                 if (entries == NULL) {
1910                                         err = -ENOMEM;
1911                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1912                                         goto out_free;
1913                                 }
1914                                 trace->ev_qualifier_ids.entries = entries;
1915                         }
1916                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1917                 }
1918         }
1919
1920         trace->ev_qualifier_ids.nr = nr_used;
1921         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1922 out:
1923         if (printed_invalid_prefix)
1924                 pr_debug("\n");
1925         return err;
1926 out_free:
1927         zfree(&trace->ev_qualifier_ids.entries);
1928         trace->ev_qualifier_ids.nr = 0;
1929         goto out;
1930 }
1931
1932 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1933 {
1934         bool in_ev_qualifier;
1935
1936         if (trace->ev_qualifier_ids.nr == 0)
1937                 return true;
1938
1939         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1940                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1941
1942         if (in_ev_qualifier)
1943                return !trace->not_ev_qualifier;
1944
1945         return trace->not_ev_qualifier;
1946 }
1947
1948 /*
1949  * args is to be interpreted as a series of longs but we need to handle
1950  * 8-byte unaligned accesses. args points to raw_data within the event
1951  * and raw_data is guaranteed to be 8-byte unaligned because it is
1952  * preceded by raw_size which is a u32. So we need to copy args to a temp
1953  * variable to read it. Most notably this avoids extended load instructions
1954  * on unaligned addresses
1955  */
1956 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1957 {
1958         unsigned long val;
1959         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1960
1961         memcpy(&val, p, sizeof(val));
1962         return val;
1963 }
1964
1965 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1966                                       struct syscall_arg *arg)
1967 {
1968         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1969                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1970
1971         return scnprintf(bf, size, "arg%d: ", arg->idx);
1972 }
1973
1974 /*
1975  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1976  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1977  * in tools/perf/trace/beauty/mount_flags.c
1978  */
1979 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1980 {
1981         if (fmt && fmt->mask_val)
1982                 return fmt->mask_val(arg, val);
1983
1984         return val;
1985 }
1986
1987 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1988                                              struct syscall_arg *arg, unsigned long val)
1989 {
1990         if (fmt && fmt->scnprintf) {
1991                 arg->val = val;
1992                 if (fmt->parm)
1993                         arg->parm = fmt->parm;
1994                 return fmt->scnprintf(bf, size, arg);
1995         }
1996         return scnprintf(bf, size, "%ld", val);
1997 }
1998
1999 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2000                                       unsigned char *args, void *augmented_args, int augmented_args_size,
2001                                       struct trace *trace, struct thread *thread)
2002 {
2003         size_t printed = 0;
2004         unsigned long val;
2005         u8 bit = 1;
2006         struct syscall_arg arg = {
2007                 .args   = args,
2008                 .augmented = {
2009                         .size = augmented_args_size,
2010                         .args = augmented_args,
2011                 },
2012                 .idx    = 0,
2013                 .mask   = 0,
2014                 .trace  = trace,
2015                 .thread = thread,
2016                 .show_string_prefix = trace->show_string_prefix,
2017         };
2018         struct thread_trace *ttrace = thread__priv(thread);
2019
2020         /*
2021          * Things like fcntl will set this in its 'cmd' formatter to pick the
2022          * right formatter for the return value (an fd? file flags?), which is
2023          * not needed for syscalls that always return a given type, say an fd.
2024          */
2025         ttrace->ret_scnprintf = NULL;
2026
2027         if (sc->args != NULL) {
2028                 struct tep_format_field *field;
2029
2030                 for (field = sc->args; field;
2031                      field = field->next, ++arg.idx, bit <<= 1) {
2032                         if (arg.mask & bit)
2033                                 continue;
2034
2035                         arg.fmt = &sc->arg_fmt[arg.idx];
2036                         val = syscall_arg__val(&arg, arg.idx);
2037                         /*
2038                          * Some syscall args need some mask, most don't and
2039                          * return val untouched.
2040                          */
2041                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2042
2043                         /*
2044                          * Suppress this argument if its value is zero and
2045                          * and we don't have a string associated in an
2046                          * strarray for it.
2047                          */
2048                         if (val == 0 &&
2049                             !trace->show_zeros &&
2050                             !(sc->arg_fmt &&
2051                               (sc->arg_fmt[arg.idx].show_zero ||
2052                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2053                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2054                               sc->arg_fmt[arg.idx].parm))
2055                                 continue;
2056
2057                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2058
2059                         if (trace->show_arg_names)
2060                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2061
2062                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2063                                                                   bf + printed, size - printed, &arg, val);
2064                 }
2065         } else if (IS_ERR(sc->tp_format)) {
2066                 /*
2067                  * If we managed to read the tracepoint /format file, then we
2068                  * may end up not having any args, like with gettid(), so only
2069                  * print the raw args when we didn't manage to read it.
2070                  */
2071                 while (arg.idx < sc->nr_args) {
2072                         if (arg.mask & bit)
2073                                 goto next_arg;
2074                         val = syscall_arg__val(&arg, arg.idx);
2075                         if (printed)
2076                                 printed += scnprintf(bf + printed, size - printed, ", ");
2077                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2078                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2079 next_arg:
2080                         ++arg.idx;
2081                         bit <<= 1;
2082                 }
2083         }
2084
2085         return printed;
2086 }
2087
2088 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2089                                   union perf_event *event,
2090                                   struct perf_sample *sample);
2091
2092 static struct syscall *trace__syscall_info(struct trace *trace,
2093                                            struct evsel *evsel, int id)
2094 {
2095         int err = 0;
2096
2097         if (id < 0) {
2098
2099                 /*
2100                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2101                  * before that, leaving at a higher verbosity level till that is
2102                  * explained. Reproduced with plain ftrace with:
2103                  *
2104                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2105                  * grep "NR -1 " /t/trace_pipe
2106                  *
2107                  * After generating some load on the machine.
2108                  */
2109                 if (verbose > 1) {
2110                         static u64 n;
2111                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2112                                 id, evsel__name(evsel), ++n);
2113                 }
2114                 return NULL;
2115         }
2116
2117         err = -EINVAL;
2118
2119 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2120         if (id > trace->sctbl->syscalls.max_id) {
2121 #else
2122         if (id >= trace->sctbl->syscalls.max_id) {
2123                 /*
2124                  * With libaudit we don't know beforehand what is the max_id,
2125                  * so we let trace__read_syscall_info() figure that out as we
2126                  * go on reading syscalls.
2127                  */
2128                 err = trace__read_syscall_info(trace, id);
2129                 if (err)
2130 #endif
2131                 goto out_cant_read;
2132         }
2133
2134         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2135             (err = trace__read_syscall_info(trace, id)) != 0)
2136                 goto out_cant_read;
2137
2138         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2139                 goto out_cant_read;
2140
2141         return &trace->syscalls.table[id];
2142
2143 out_cant_read:
2144         if (verbose > 0) {
2145                 char sbuf[STRERR_BUFSIZE];
2146                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2147                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2148                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2149                 fputs(" information\n", trace->output);
2150         }
2151         return NULL;
2152 }
2153
2154 struct syscall_stats {
2155         struct stats stats;
2156         u64          nr_failures;
2157         int          max_errno;
2158         u32          *errnos;
2159 };
2160
2161 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2162                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2163 {
2164         struct int_node *inode;
2165         struct syscall_stats *stats;
2166         u64 duration = 0;
2167
2168         inode = intlist__findnew(ttrace->syscall_stats, id);
2169         if (inode == NULL)
2170                 return;
2171
2172         stats = inode->priv;
2173         if (stats == NULL) {
2174                 stats = zalloc(sizeof(*stats));
2175                 if (stats == NULL)
2176                         return;
2177
2178                 init_stats(&stats->stats);
2179                 inode->priv = stats;
2180         }
2181
2182         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2183                 duration = sample->time - ttrace->entry_time;
2184
2185         update_stats(&stats->stats, duration);
2186
2187         if (err < 0) {
2188                 ++stats->nr_failures;
2189
2190                 if (!errno_summary)
2191                         return;
2192
2193                 err = -err;
2194                 if (err > stats->max_errno) {
2195                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2196
2197                         if (new_errnos) {
2198                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2199                         } else {
2200                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2201                                          thread__comm_str(thread), thread->pid_, thread->tid);
2202                                 return;
2203                         }
2204
2205                         stats->errnos = new_errnos;
2206                         stats->max_errno = err;
2207                 }
2208
2209                 ++stats->errnos[err - 1];
2210         }
2211 }
2212
2213 static int trace__printf_interrupted_entry(struct trace *trace)
2214 {
2215         struct thread_trace *ttrace;
2216         size_t printed;
2217         int len;
2218
2219         if (trace->failure_only || trace->current == NULL)
2220                 return 0;
2221
2222         ttrace = thread__priv(trace->current);
2223
2224         if (!ttrace->entry_pending)
2225                 return 0;
2226
2227         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2228         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2229
2230         if (len < trace->args_alignment - 4)
2231                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2232
2233         printed += fprintf(trace->output, " ...\n");
2234
2235         ttrace->entry_pending = false;
2236         ++trace->nr_events_printed;
2237
2238         return printed;
2239 }
2240
2241 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2242                                  struct perf_sample *sample, struct thread *thread)
2243 {
2244         int printed = 0;
2245
2246         if (trace->print_sample) {
2247                 double ts = (double)sample->time / NSEC_PER_MSEC;
2248
2249                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2250                                    evsel__name(evsel), ts,
2251                                    thread__comm_str(thread),
2252                                    sample->pid, sample->tid, sample->cpu);
2253         }
2254
2255         return printed;
2256 }
2257
2258 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2259 {
2260         void *augmented_args = NULL;
2261         /*
2262          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2263          * and there we get all 6 syscall args plus the tracepoint common fields
2264          * that gets calculated at the start and the syscall_nr (another long).
2265          * So we check if that is the case and if so don't look after the
2266          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2267          * which is fixed.
2268          *
2269          * We'll revisit this later to pass s->args_size to the BPF augmenter
2270          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2271          * copies only what we need for each syscall, like what happens when we
2272          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2273          * traffic to just what is needed for each syscall.
2274          */
2275         int args_size = raw_augmented_args_size ?: sc->args_size;
2276
2277         *augmented_args_size = sample->raw_size - args_size;
2278         if (*augmented_args_size > 0)
2279                 augmented_args = sample->raw_data + args_size;
2280
2281         return augmented_args;
2282 }
2283
2284 static void syscall__exit(struct syscall *sc)
2285 {
2286         if (!sc)
2287                 return;
2288
2289         free(sc->arg_fmt);
2290 }
2291
2292 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2293                             union perf_event *event __maybe_unused,
2294                             struct perf_sample *sample)
2295 {
2296         char *msg;
2297         void *args;
2298         int printed = 0;
2299         struct thread *thread;
2300         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2301         int augmented_args_size = 0;
2302         void *augmented_args = NULL;
2303         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2304         struct thread_trace *ttrace;
2305
2306         if (sc == NULL)
2307                 return -1;
2308
2309         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2310         ttrace = thread__trace(thread, trace->output);
2311         if (ttrace == NULL)
2312                 goto out_put;
2313
2314         trace__fprintf_sample(trace, evsel, sample, thread);
2315
2316         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2317
2318         if (ttrace->entry_str == NULL) {
2319                 ttrace->entry_str = malloc(trace__entry_str_size);
2320                 if (!ttrace->entry_str)
2321                         goto out_put;
2322         }
2323
2324         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2325                 trace__printf_interrupted_entry(trace);
2326         /*
2327          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2328          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2329          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2330          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2331          * so when handling, say the openat syscall, we end up getting 6 args for the
2332          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2333          * thinking that the extra 2 u64 args are the augmented filename, so just check
2334          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2335          */
2336         if (evsel != trace->syscalls.events.sys_enter)
2337                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2338         ttrace->entry_time = sample->time;
2339         msg = ttrace->entry_str;
2340         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2341
2342         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2343                                            args, augmented_args, augmented_args_size, trace, thread);
2344
2345         if (sc->is_exit) {
2346                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2347                         int alignment = 0;
2348
2349                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2350                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2351                         if (trace->args_alignment > printed)
2352                                 alignment = trace->args_alignment - printed;
2353                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2354                 }
2355         } else {
2356                 ttrace->entry_pending = true;
2357                 /* See trace__vfs_getname & trace__sys_exit */
2358                 ttrace->filename.pending_open = false;
2359         }
2360
2361         if (trace->current != thread) {
2362                 thread__put(trace->current);
2363                 trace->current = thread__get(thread);
2364         }
2365         err = 0;
2366 out_put:
2367         thread__put(thread);
2368         return err;
2369 }
2370
2371 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2372                                     struct perf_sample *sample)
2373 {
2374         struct thread_trace *ttrace;
2375         struct thread *thread;
2376         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2377         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2378         char msg[1024];
2379         void *args, *augmented_args = NULL;
2380         int augmented_args_size;
2381
2382         if (sc == NULL)
2383                 return -1;
2384
2385         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2386         ttrace = thread__trace(thread, trace->output);
2387         /*
2388          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2389          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2390          */
2391         if (ttrace == NULL)
2392                 goto out_put;
2393
2394         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2395         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2396         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2397         fprintf(trace->output, "%s", msg);
2398         err = 0;
2399 out_put:
2400         thread__put(thread);
2401         return err;
2402 }
2403
2404 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2405                                     struct perf_sample *sample,
2406                                     struct callchain_cursor *cursor)
2407 {
2408         struct addr_location al;
2409         int max_stack = evsel->core.attr.sample_max_stack ?
2410                         evsel->core.attr.sample_max_stack :
2411                         trace->max_stack;
2412         int err;
2413
2414         if (machine__resolve(trace->host, &al, sample) < 0)
2415                 return -1;
2416
2417         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2418         addr_location__put(&al);
2419         return err;
2420 }
2421
2422 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2423 {
2424         /* TODO: user-configurable print_opts */
2425         const unsigned int print_opts = EVSEL__PRINT_SYM |
2426                                         EVSEL__PRINT_DSO |
2427                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2428
2429         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2430 }
2431
2432 static const char *errno_to_name(struct evsel *evsel, int err)
2433 {
2434         struct perf_env *env = evsel__env(evsel);
2435         const char *arch_name = perf_env__arch(env);
2436
2437         return arch_syscalls__strerrno(arch_name, err);
2438 }
2439
2440 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2441                            union perf_event *event __maybe_unused,
2442                            struct perf_sample *sample)
2443 {
2444         long ret;
2445         u64 duration = 0;
2446         bool duration_calculated = false;
2447         struct thread *thread;
2448         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2449         int alignment = trace->args_alignment;
2450         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2451         struct thread_trace *ttrace;
2452
2453         if (sc == NULL)
2454                 return -1;
2455
2456         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2457         ttrace = thread__trace(thread, trace->output);
2458         if (ttrace == NULL)
2459                 goto out_put;
2460
2461         trace__fprintf_sample(trace, evsel, sample, thread);
2462
2463         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2464
2465         if (trace->summary)
2466                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2467
2468         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2469                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2470                 ttrace->filename.pending_open = false;
2471                 ++trace->stats.vfs_getname;
2472         }
2473
2474         if (ttrace->entry_time) {
2475                 duration = sample->time - ttrace->entry_time;
2476                 if (trace__filter_duration(trace, duration))
2477                         goto out;
2478                 duration_calculated = true;
2479         } else if (trace->duration_filter)
2480                 goto out;
2481
2482         if (sample->callchain) {
2483                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2484                 if (callchain_ret == 0) {
2485                         if (callchain_cursor.nr < trace->min_stack)
2486                                 goto out;
2487                         callchain_ret = 1;
2488                 }
2489         }
2490
2491         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2492                 goto out;
2493
2494         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2495
2496         if (ttrace->entry_pending) {
2497                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2498         } else {
2499                 printed += fprintf(trace->output, " ... [");
2500                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2501                 printed += 9;
2502                 printed += fprintf(trace->output, "]: %s()", sc->name);
2503         }
2504
2505         printed++; /* the closing ')' */
2506
2507         if (alignment > printed)
2508                 alignment -= printed;
2509         else
2510                 alignment = 0;
2511
2512         fprintf(trace->output, ")%*s= ", alignment, " ");
2513
2514         if (sc->fmt == NULL) {
2515                 if (ret < 0)
2516                         goto errno_print;
2517 signed_print:
2518                 fprintf(trace->output, "%ld", ret);
2519         } else if (ret < 0) {
2520 errno_print: {
2521                 char bf[STRERR_BUFSIZE];
2522                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2523                            *e = errno_to_name(evsel, -ret);
2524
2525                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2526         }
2527         } else if (ret == 0 && sc->fmt->timeout)
2528                 fprintf(trace->output, "0 (Timeout)");
2529         else if (ttrace->ret_scnprintf) {
2530                 char bf[1024];
2531                 struct syscall_arg arg = {
2532                         .val    = ret,
2533                         .thread = thread,
2534                         .trace  = trace,
2535                 };
2536                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2537                 ttrace->ret_scnprintf = NULL;
2538                 fprintf(trace->output, "%s", bf);
2539         } else if (sc->fmt->hexret)
2540                 fprintf(trace->output, "%#lx", ret);
2541         else if (sc->fmt->errpid) {
2542                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2543
2544                 if (child != NULL) {
2545                         fprintf(trace->output, "%ld", ret);
2546                         if (child->comm_set)
2547                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2548                         thread__put(child);
2549                 }
2550         } else
2551                 goto signed_print;
2552
2553         fputc('\n', trace->output);
2554
2555         /*
2556          * We only consider an 'event' for the sake of --max-events a non-filtered
2557          * sys_enter + sys_exit and other tracepoint events.
2558          */
2559         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2560                 interrupted = true;
2561
2562         if (callchain_ret > 0)
2563                 trace__fprintf_callchain(trace, sample);
2564         else if (callchain_ret < 0)
2565                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2566 out:
2567         ttrace->entry_pending = false;
2568         err = 0;
2569 out_put:
2570         thread__put(thread);
2571         return err;
2572 }
2573
2574 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2575                               union perf_event *event __maybe_unused,
2576                               struct perf_sample *sample)
2577 {
2578         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2579         struct thread_trace *ttrace;
2580         size_t filename_len, entry_str_len, to_move;
2581         ssize_t remaining_space;
2582         char *pos;
2583         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2584
2585         if (!thread)
2586                 goto out;
2587
2588         ttrace = thread__priv(thread);
2589         if (!ttrace)
2590                 goto out_put;
2591
2592         filename_len = strlen(filename);
2593         if (filename_len == 0)
2594                 goto out_put;
2595
2596         if (ttrace->filename.namelen < filename_len) {
2597                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2598
2599                 if (f == NULL)
2600                         goto out_put;
2601
2602                 ttrace->filename.namelen = filename_len;
2603                 ttrace->filename.name = f;
2604         }
2605
2606         strcpy(ttrace->filename.name, filename);
2607         ttrace->filename.pending_open = true;
2608
2609         if (!ttrace->filename.ptr)
2610                 goto out_put;
2611
2612         entry_str_len = strlen(ttrace->entry_str);
2613         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2614         if (remaining_space <= 0)
2615                 goto out_put;
2616
2617         if (filename_len > (size_t)remaining_space) {
2618                 filename += filename_len - remaining_space;
2619                 filename_len = remaining_space;
2620         }
2621
2622         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2623         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2624         memmove(pos + filename_len, pos, to_move);
2625         memcpy(pos, filename, filename_len);
2626
2627         ttrace->filename.ptr = 0;
2628         ttrace->filename.entry_str_pos = 0;
2629 out_put:
2630         thread__put(thread);
2631 out:
2632         return 0;
2633 }
2634
2635 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2636                                      union perf_event *event __maybe_unused,
2637                                      struct perf_sample *sample)
2638 {
2639         u64 runtime = evsel__intval(evsel, sample, "runtime");
2640         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2641         struct thread *thread = machine__findnew_thread(trace->host,
2642                                                         sample->pid,
2643                                                         sample->tid);
2644         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2645
2646         if (ttrace == NULL)
2647                 goto out_dump;
2648
2649         ttrace->runtime_ms += runtime_ms;
2650         trace->runtime_ms += runtime_ms;
2651 out_put:
2652         thread__put(thread);
2653         return 0;
2654
2655 out_dump:
2656         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2657                evsel->name,
2658                evsel__strval(evsel, sample, "comm"),
2659                (pid_t)evsel__intval(evsel, sample, "pid"),
2660                runtime,
2661                evsel__intval(evsel, sample, "vruntime"));
2662         goto out_put;
2663 }
2664
2665 static int bpf_output__printer(enum binary_printer_ops op,
2666                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2667 {
2668         unsigned char ch = (unsigned char)val;
2669
2670         switch (op) {
2671         case BINARY_PRINT_CHAR_DATA:
2672                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2673         case BINARY_PRINT_DATA_BEGIN:
2674         case BINARY_PRINT_LINE_BEGIN:
2675         case BINARY_PRINT_ADDR:
2676         case BINARY_PRINT_NUM_DATA:
2677         case BINARY_PRINT_NUM_PAD:
2678         case BINARY_PRINT_SEP:
2679         case BINARY_PRINT_CHAR_PAD:
2680         case BINARY_PRINT_LINE_END:
2681         case BINARY_PRINT_DATA_END:
2682         default:
2683                 break;
2684         }
2685
2686         return 0;
2687 }
2688
2689 static void bpf_output__fprintf(struct trace *trace,
2690                                 struct perf_sample *sample)
2691 {
2692         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2693                         bpf_output__printer, NULL, trace->output);
2694         ++trace->nr_events_printed;
2695 }
2696
2697 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2698                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2699 {
2700         char bf[2048];
2701         size_t size = sizeof(bf);
2702         struct tep_format_field *field = evsel->tp_format->format.fields;
2703         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2704         size_t printed = 0;
2705         unsigned long val;
2706         u8 bit = 1;
2707         struct syscall_arg syscall_arg = {
2708                 .augmented = {
2709                         .size = augmented_args_size,
2710                         .args = augmented_args,
2711                 },
2712                 .idx    = 0,
2713                 .mask   = 0,
2714                 .trace  = trace,
2715                 .thread = thread,
2716                 .show_string_prefix = trace->show_string_prefix,
2717         };
2718
2719         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2720                 if (syscall_arg.mask & bit)
2721                         continue;
2722
2723                 syscall_arg.len = 0;
2724                 syscall_arg.fmt = arg;
2725                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2726                         int offset = field->offset;
2727
2728                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2729                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2730                                 syscall_arg.len = offset >> 16;
2731                                 offset &= 0xffff;
2732 #ifdef HAVE_LIBTRACEEVENT_TEP_FIELD_IS_RELATIVE
2733                                 if (field->flags & TEP_FIELD_IS_RELATIVE)
2734                                         offset += field->offset + field->size;
2735 #endif
2736                         }
2737
2738                         val = (uintptr_t)(sample->raw_data + offset);
2739                 } else
2740                         val = format_field__intval(field, sample, evsel->needs_swap);
2741                 /*
2742                  * Some syscall args need some mask, most don't and
2743                  * return val untouched.
2744                  */
2745                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2746
2747                 /*
2748                  * Suppress this argument if its value is zero and
2749                  * we don't have a string associated in an
2750                  * strarray for it.
2751                  */
2752                 if (val == 0 &&
2753                     !trace->show_zeros &&
2754                     !((arg->show_zero ||
2755                        arg->scnprintf == SCA_STRARRAY ||
2756                        arg->scnprintf == SCA_STRARRAYS) &&
2757                       arg->parm))
2758                         continue;
2759
2760                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2761
2762                 if (trace->show_arg_names)
2763                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2764
2765                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2766         }
2767
2768         return printed + fprintf(trace->output, "%s", bf);
2769 }
2770
2771 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2772                                 union perf_event *event __maybe_unused,
2773                                 struct perf_sample *sample)
2774 {
2775         struct thread *thread;
2776         int callchain_ret = 0;
2777         /*
2778          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2779          * this event's max_events having been hit and this is an entry coming
2780          * from the ring buffer that we should discard, since the max events
2781          * have already been considered/printed.
2782          */
2783         if (evsel->disabled)
2784                 return 0;
2785
2786         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2787
2788         if (sample->callchain) {
2789                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2790                 if (callchain_ret == 0) {
2791                         if (callchain_cursor.nr < trace->min_stack)
2792                                 goto out;
2793                         callchain_ret = 1;
2794                 }
2795         }
2796
2797         trace__printf_interrupted_entry(trace);
2798         trace__fprintf_tstamp(trace, sample->time, trace->output);
2799
2800         if (trace->trace_syscalls && trace->show_duration)
2801                 fprintf(trace->output, "(         ): ");
2802
2803         if (thread)
2804                 trace__fprintf_comm_tid(trace, thread, trace->output);
2805
2806         if (evsel == trace->syscalls.events.augmented) {
2807                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2808                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2809
2810                 if (sc) {
2811                         fprintf(trace->output, "%s(", sc->name);
2812                         trace__fprintf_sys_enter(trace, evsel, sample);
2813                         fputc(')', trace->output);
2814                         goto newline;
2815                 }
2816
2817                 /*
2818                  * XXX: Not having the associated syscall info or not finding/adding
2819                  *      the thread should never happen, but if it does...
2820                  *      fall thru and print it as a bpf_output event.
2821                  */
2822         }
2823
2824         fprintf(trace->output, "%s(", evsel->name);
2825
2826         if (evsel__is_bpf_output(evsel)) {
2827                 bpf_output__fprintf(trace, sample);
2828         } else if (evsel->tp_format) {
2829                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2830                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2831                         if (trace->libtraceevent_print) {
2832                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2833                                                       sample->raw_data, sample->raw_size,
2834                                                       trace->output);
2835                         } else {
2836                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2837                         }
2838                 }
2839         }
2840
2841 newline:
2842         fprintf(trace->output, ")\n");
2843
2844         if (callchain_ret > 0)
2845                 trace__fprintf_callchain(trace, sample);
2846         else if (callchain_ret < 0)
2847                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2848
2849         ++trace->nr_events_printed;
2850
2851         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2852                 evsel__disable(evsel);
2853                 evsel__close(evsel);
2854         }
2855 out:
2856         thread__put(thread);
2857         return 0;
2858 }
2859
2860 static void print_location(FILE *f, struct perf_sample *sample,
2861                            struct addr_location *al,
2862                            bool print_dso, bool print_sym)
2863 {
2864
2865         if ((verbose > 0 || print_dso) && al->map)
2866                 fprintf(f, "%s@", al->map->dso->long_name);
2867
2868         if ((verbose > 0 || print_sym) && al->sym)
2869                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2870                         al->addr - al->sym->start);
2871         else if (al->map)
2872                 fprintf(f, "0x%" PRIx64, al->addr);
2873         else
2874                 fprintf(f, "0x%" PRIx64, sample->addr);
2875 }
2876
2877 static int trace__pgfault(struct trace *trace,
2878                           struct evsel *evsel,
2879                           union perf_event *event __maybe_unused,
2880                           struct perf_sample *sample)
2881 {
2882         struct thread *thread;
2883         struct addr_location al;
2884         char map_type = 'd';
2885         struct thread_trace *ttrace;
2886         int err = -1;
2887         int callchain_ret = 0;
2888
2889         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2890
2891         if (sample->callchain) {
2892                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2893                 if (callchain_ret == 0) {
2894                         if (callchain_cursor.nr < trace->min_stack)
2895                                 goto out_put;
2896                         callchain_ret = 1;
2897                 }
2898         }
2899
2900         ttrace = thread__trace(thread, trace->output);
2901         if (ttrace == NULL)
2902                 goto out_put;
2903
2904         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2905                 ttrace->pfmaj++;
2906         else
2907                 ttrace->pfmin++;
2908
2909         if (trace->summary_only)
2910                 goto out;
2911
2912         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2913
2914         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2915
2916         fprintf(trace->output, "%sfault [",
2917                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2918                 "maj" : "min");
2919
2920         print_location(trace->output, sample, &al, false, true);
2921
2922         fprintf(trace->output, "] => ");
2923
2924         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2925
2926         if (!al.map) {
2927                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2928
2929                 if (al.map)
2930                         map_type = 'x';
2931                 else
2932                         map_type = '?';
2933         }
2934
2935         print_location(trace->output, sample, &al, true, false);
2936
2937         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2938
2939         if (callchain_ret > 0)
2940                 trace__fprintf_callchain(trace, sample);
2941         else if (callchain_ret < 0)
2942                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2943
2944         ++trace->nr_events_printed;
2945 out:
2946         err = 0;
2947 out_put:
2948         thread__put(thread);
2949         return err;
2950 }
2951
2952 static void trace__set_base_time(struct trace *trace,
2953                                  struct evsel *evsel,
2954                                  struct perf_sample *sample)
2955 {
2956         /*
2957          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2958          * and don't use sample->time unconditionally, we may end up having
2959          * some other event in the future without PERF_SAMPLE_TIME for good
2960          * reason, i.e. we may not be interested in its timestamps, just in
2961          * it taking place, picking some piece of information when it
2962          * appears in our event stream (vfs_getname comes to mind).
2963          */
2964         if (trace->base_time == 0 && !trace->full_time &&
2965             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2966                 trace->base_time = sample->time;
2967 }
2968
2969 static int trace__process_sample(struct perf_tool *tool,
2970                                  union perf_event *event,
2971                                  struct perf_sample *sample,
2972                                  struct evsel *evsel,
2973                                  struct machine *machine __maybe_unused)
2974 {
2975         struct trace *trace = container_of(tool, struct trace, tool);
2976         struct thread *thread;
2977         int err = 0;
2978
2979         tracepoint_handler handler = evsel->handler;
2980
2981         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2982         if (thread && thread__is_filtered(thread))
2983                 goto out;
2984
2985         trace__set_base_time(trace, evsel, sample);
2986
2987         if (handler) {
2988                 ++trace->nr_events;
2989                 handler(trace, evsel, event, sample);
2990         }
2991 out:
2992         thread__put(thread);
2993         return err;
2994 }
2995
2996 static int trace__record(struct trace *trace, int argc, const char **argv)
2997 {
2998         unsigned int rec_argc, i, j;
2999         const char **rec_argv;
3000         const char * const record_args[] = {
3001                 "record",
3002                 "-R",
3003                 "-m", "1024",
3004                 "-c", "1",
3005         };
3006         pid_t pid = getpid();
3007         char *filter = asprintf__tp_filter_pids(1, &pid);
3008         const char * const sc_args[] = { "-e", };
3009         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3010         const char * const majpf_args[] = { "-e", "major-faults" };
3011         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3012         const char * const minpf_args[] = { "-e", "minor-faults" };
3013         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3014         int err = -1;
3015
3016         /* +3 is for the event string below and the pid filter */
3017         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3018                 majpf_args_nr + minpf_args_nr + argc;
3019         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3020
3021         if (rec_argv == NULL || filter == NULL)
3022                 goto out_free;
3023
3024         j = 0;
3025         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3026                 rec_argv[j++] = record_args[i];
3027
3028         if (trace->trace_syscalls) {
3029                 for (i = 0; i < sc_args_nr; i++)
3030                         rec_argv[j++] = sc_args[i];
3031
3032                 /* event string may be different for older kernels - e.g., RHEL6 */
3033                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3034                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3035                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3036                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3037                 else {
3038                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3039                         goto out_free;
3040                 }
3041         }
3042
3043         rec_argv[j++] = "--filter";
3044         rec_argv[j++] = filter;
3045
3046         if (trace->trace_pgfaults & TRACE_PFMAJ)
3047                 for (i = 0; i < majpf_args_nr; i++)
3048                         rec_argv[j++] = majpf_args[i];
3049
3050         if (trace->trace_pgfaults & TRACE_PFMIN)
3051                 for (i = 0; i < minpf_args_nr; i++)
3052                         rec_argv[j++] = minpf_args[i];
3053
3054         for (i = 0; i < (unsigned int)argc; i++)
3055                 rec_argv[j++] = argv[i];
3056
3057         err = cmd_record(j, rec_argv);
3058 out_free:
3059         free(filter);
3060         free(rec_argv);
3061         return err;
3062 }
3063
3064 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3065
3066 static bool evlist__add_vfs_getname(struct evlist *evlist)
3067 {
3068         bool found = false;
3069         struct evsel *evsel, *tmp;
3070         struct parse_events_error err;
3071         int ret;
3072
3073         parse_events_error__init(&err);
3074         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3075         parse_events_error__exit(&err);
3076         if (ret)
3077                 return false;
3078
3079         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3080                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3081                         continue;
3082
3083                 if (evsel__field(evsel, "pathname")) {
3084                         evsel->handler = trace__vfs_getname;
3085                         found = true;
3086                         continue;
3087                 }
3088
3089                 list_del_init(&evsel->core.node);
3090                 evsel->evlist = NULL;
3091                 evsel__delete(evsel);
3092         }
3093
3094         return found;
3095 }
3096
3097 static struct evsel *evsel__new_pgfault(u64 config)
3098 {
3099         struct evsel *evsel;
3100         struct perf_event_attr attr = {
3101                 .type = PERF_TYPE_SOFTWARE,
3102                 .mmap_data = 1,
3103         };
3104
3105         attr.config = config;
3106         attr.sample_period = 1;
3107
3108         event_attr_init(&attr);
3109
3110         evsel = evsel__new(&attr);
3111         if (evsel)
3112                 evsel->handler = trace__pgfault;
3113
3114         return evsel;
3115 }
3116
3117 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3118 {
3119         struct evsel *evsel;
3120
3121         evlist__for_each_entry(evlist, evsel) {
3122                 struct evsel_trace *et = evsel->priv;
3123
3124                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3125                         continue;
3126
3127                 free(et->fmt);
3128                 free(et);
3129         }
3130 }
3131
3132 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3133 {
3134         const u32 type = event->header.type;
3135         struct evsel *evsel;
3136
3137         if (type != PERF_RECORD_SAMPLE) {
3138                 trace__process_event(trace, trace->host, event, sample);
3139                 return;
3140         }
3141
3142         evsel = evlist__id2evsel(trace->evlist, sample->id);
3143         if (evsel == NULL) {
3144                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3145                 return;
3146         }
3147
3148         if (evswitch__discard(&trace->evswitch, evsel))
3149                 return;
3150
3151         trace__set_base_time(trace, evsel, sample);
3152
3153         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3154             sample->raw_data == NULL) {
3155                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3156                        evsel__name(evsel), sample->tid,
3157                        sample->cpu, sample->raw_size);
3158         } else {
3159                 tracepoint_handler handler = evsel->handler;
3160                 handler(trace, evsel, event, sample);
3161         }
3162
3163         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3164                 interrupted = true;
3165 }
3166
3167 static int trace__add_syscall_newtp(struct trace *trace)
3168 {
3169         int ret = -1;
3170         struct evlist *evlist = trace->evlist;
3171         struct evsel *sys_enter, *sys_exit;
3172
3173         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3174         if (sys_enter == NULL)
3175                 goto out;
3176
3177         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3178                 goto out_delete_sys_enter;
3179
3180         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3181         if (sys_exit == NULL)
3182                 goto out_delete_sys_enter;
3183
3184         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3185                 goto out_delete_sys_exit;
3186
3187         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3188         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3189
3190         evlist__add(evlist, sys_enter);
3191         evlist__add(evlist, sys_exit);
3192
3193         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3194                 /*
3195                  * We're interested only in the user space callchain
3196                  * leading to the syscall, allow overriding that for
3197                  * debugging reasons using --kernel_syscall_callchains
3198                  */
3199                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3200         }
3201
3202         trace->syscalls.events.sys_enter = sys_enter;
3203         trace->syscalls.events.sys_exit  = sys_exit;
3204
3205         ret = 0;
3206 out:
3207         return ret;
3208
3209 out_delete_sys_exit:
3210         evsel__delete_priv(sys_exit);
3211 out_delete_sys_enter:
3212         evsel__delete_priv(sys_enter);
3213         goto out;
3214 }
3215
3216 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3217 {
3218         int err = -1;
3219         struct evsel *sys_exit;
3220         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3221                                                 trace->ev_qualifier_ids.nr,
3222                                                 trace->ev_qualifier_ids.entries);
3223
3224         if (filter == NULL)
3225                 goto out_enomem;
3226
3227         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3228                 sys_exit = trace->syscalls.events.sys_exit;
3229                 err = evsel__append_tp_filter(sys_exit, filter);
3230         }
3231
3232         free(filter);
3233 out:
3234         return err;
3235 out_enomem:
3236         errno = ENOMEM;
3237         goto out;
3238 }
3239
3240 #ifdef HAVE_LIBBPF_SUPPORT
3241 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3242 {
3243         if (trace->bpf_obj == NULL)
3244                 return NULL;
3245
3246         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3247 }
3248
3249 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3250 {
3251         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3252 }
3253
3254 static void trace__set_bpf_map_syscalls(struct trace *trace)
3255 {
3256         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3257         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3258 }
3259
3260 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3261 {
3262         struct bpf_program *pos, *prog = NULL;
3263         const char *sec_name;
3264
3265         if (trace->bpf_obj == NULL)
3266                 return NULL;
3267
3268         bpf_object__for_each_program(pos, trace->bpf_obj) {
3269                 sec_name = bpf_program__section_name(pos);
3270                 if (sec_name && !strcmp(sec_name, name)) {
3271                         prog = pos;
3272                         break;
3273                 }
3274         }
3275
3276         return prog;
3277 }
3278
3279 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3280                                                         const char *prog_name, const char *type)
3281 {
3282         struct bpf_program *prog;
3283
3284         if (prog_name == NULL) {
3285                 char default_prog_name[256];
3286                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3287                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3288                 if (prog != NULL)
3289                         goto out_found;
3290                 if (sc->fmt && sc->fmt->alias) {
3291                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3292                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3293                         if (prog != NULL)
3294                                 goto out_found;
3295                 }
3296                 goto out_unaugmented;
3297         }
3298
3299         prog = trace__find_bpf_program_by_title(trace, prog_name);
3300
3301         if (prog != NULL) {
3302 out_found:
3303                 return prog;
3304         }
3305
3306         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3307                  prog_name, type, sc->name);
3308 out_unaugmented:
3309         return trace->syscalls.unaugmented_prog;
3310 }
3311
3312 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3313 {
3314         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3315
3316         if (sc == NULL)
3317                 return;
3318
3319         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3320         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3321 }
3322
3323 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3324 {
3325         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3326         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3327 }
3328
3329 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3330 {
3331         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3332         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3333 }
3334
3335 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3336 {
3337         struct tep_format_field *field, *candidate_field;
3338         int id;
3339
3340         /*
3341          * We're only interested in syscalls that have a pointer:
3342          */
3343         for (field = sc->args; field; field = field->next) {
3344                 if (field->flags & TEP_FIELD_IS_POINTER)
3345                         goto try_to_find_pair;
3346         }
3347
3348         return NULL;
3349
3350 try_to_find_pair:
3351         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3352                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3353                 struct bpf_program *pair_prog;
3354                 bool is_candidate = false;
3355
3356                 if (pair == NULL || pair == sc ||
3357                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3358                         continue;
3359
3360                 for (field = sc->args, candidate_field = pair->args;
3361                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3362                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3363                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3364
3365                         if (is_pointer) {
3366                                if (!candidate_is_pointer) {
3367                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3368                                         continue;
3369                                }
3370                         } else {
3371                                 if (candidate_is_pointer) {
3372                                         // The candidate might copy a pointer we don't have, skip it.
3373                                         goto next_candidate;
3374                                 }
3375                                 continue;
3376                         }
3377
3378                         if (strcmp(field->type, candidate_field->type))
3379                                 goto next_candidate;
3380
3381                         is_candidate = true;
3382                 }
3383
3384                 if (!is_candidate)
3385                         goto next_candidate;
3386
3387                 /*
3388                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3389                  * then it may be collecting that and we then can't use it, as it would collect
3390                  * more than what is common to the two syscalls.
3391                  */
3392                 if (candidate_field) {
3393                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3394                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3395                                         goto next_candidate;
3396                 }
3397
3398                 pair_prog = pair->bpf_prog.sys_enter;
3399                 /*
3400                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3401                  * have been searched for, so search it here and if it returns the
3402                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3403                  * program for a filtered syscall on a non-filtered one.
3404                  *
3405                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3406                  * useful for "renameat2".
3407                  */
3408                 if (pair_prog == NULL) {
3409                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3410                         if (pair_prog == trace->syscalls.unaugmented_prog)
3411                                 goto next_candidate;
3412                 }
3413
3414                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3415                 return pair_prog;
3416         next_candidate:
3417                 continue;
3418         }
3419
3420         return NULL;
3421 }
3422
3423 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3424 {
3425         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3426             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3427         int err = 0, key;
3428
3429         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3430                 int prog_fd;
3431
3432                 if (!trace__syscall_enabled(trace, key))
3433                         continue;
3434
3435                 trace__init_syscall_bpf_progs(trace, key);
3436
3437                 // It'll get at least the "!raw_syscalls:unaugmented"
3438                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3439                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3440                 if (err)
3441                         break;
3442                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3443                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3444                 if (err)
3445                         break;
3446         }
3447
3448         /*
3449          * Now lets do a second pass looking for enabled syscalls without
3450          * an augmenter that have a signature that is a superset of another
3451          * syscall with an augmenter so that we can auto-reuse it.
3452          *
3453          * I.e. if we have an augmenter for the "open" syscall that has
3454          * this signature:
3455          *
3456          *   int open(const char *pathname, int flags, mode_t mode);
3457          *
3458          * I.e. that will collect just the first string argument, then we
3459          * can reuse it for the 'creat' syscall, that has this signature:
3460          *
3461          *   int creat(const char *pathname, mode_t mode);
3462          *
3463          * and for:
3464          *
3465          *   int stat(const char *pathname, struct stat *statbuf);
3466          *   int lstat(const char *pathname, struct stat *statbuf);
3467          *
3468          * Because the 'open' augmenter will collect the first arg as a string,
3469          * and leave alone all the other args, which already helps with
3470          * beautifying 'stat' and 'lstat''s pathname arg.
3471          *
3472          * Then, in time, when 'stat' gets an augmenter that collects both
3473          * first and second arg (this one on the raw_syscalls:sys_exit prog
3474          * array tail call, then that one will be used.
3475          */
3476         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3477                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3478                 struct bpf_program *pair_prog;
3479                 int prog_fd;
3480
3481                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3482                         continue;
3483
3484                 /*
3485                  * For now we're just reusing the sys_enter prog, and if it
3486                  * already has an augmenter, we don't need to find one.
3487                  */
3488                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3489                         continue;
3490
3491                 /*
3492                  * Look at all the other syscalls for one that has a signature
3493                  * that is close enough that we can share:
3494                  */
3495                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3496                 if (pair_prog == NULL)
3497                         continue;
3498
3499                 sc->bpf_prog.sys_enter = pair_prog;
3500
3501                 /*
3502                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3503                  * with the fd for the program we're reusing:
3504                  */
3505                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3506                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3507                 if (err)
3508                         break;
3509         }
3510
3511
3512         return err;
3513 }
3514
3515 static void trace__delete_augmented_syscalls(struct trace *trace)
3516 {
3517         struct evsel *evsel, *tmp;
3518
3519         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3520         evsel__delete(trace->syscalls.events.augmented);
3521         trace->syscalls.events.augmented = NULL;
3522
3523         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3524                 if (evsel->bpf_obj == trace->bpf_obj) {
3525                         evlist__remove(trace->evlist, evsel);
3526                         evsel__delete(evsel);
3527                 }
3528
3529         }
3530
3531         bpf_object__close(trace->bpf_obj);
3532         trace->bpf_obj = NULL;
3533 }
3534 #else // HAVE_LIBBPF_SUPPORT
3535 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3536                                                    const char *name __maybe_unused)
3537 {
3538         return NULL;
3539 }
3540
3541 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3542 {
3543 }
3544
3545 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3546 {
3547 }
3548
3549 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3550                                                             const char *name __maybe_unused)
3551 {
3552         return NULL;
3553 }
3554
3555 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3556 {
3557         return 0;
3558 }
3559
3560 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3561 {
3562 }
3563 #endif // HAVE_LIBBPF_SUPPORT
3564
3565 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3566 {
3567         struct evsel *evsel;
3568
3569         evlist__for_each_entry(trace->evlist, evsel) {
3570                 if (evsel == trace->syscalls.events.augmented ||
3571                     evsel->bpf_obj == trace->bpf_obj)
3572                         continue;
3573
3574                 return false;
3575         }
3576
3577         return true;
3578 }
3579
3580 static int trace__set_ev_qualifier_filter(struct trace *trace)
3581 {
3582         if (trace->syscalls.events.sys_enter)
3583                 return trace__set_ev_qualifier_tp_filter(trace);
3584         return 0;
3585 }
3586
3587 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3588                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3589 {
3590         int err = 0;
3591 #ifdef HAVE_LIBBPF_SUPPORT
3592         bool value = true;
3593         int map_fd = bpf_map__fd(map);
3594         size_t i;
3595
3596         for (i = 0; i < npids; ++i) {
3597                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3598                 if (err)
3599                         break;
3600         }
3601 #endif
3602         return err;
3603 }
3604
3605 static int trace__set_filter_loop_pids(struct trace *trace)
3606 {
3607         unsigned int nr = 1, err;
3608         pid_t pids[32] = {
3609                 getpid(),
3610         };
3611         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3612
3613         while (thread && nr < ARRAY_SIZE(pids)) {
3614                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3615
3616                 if (parent == NULL)
3617                         break;
3618
3619                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3620                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3621                         pids[nr++] = parent->tid;
3622                         break;
3623                 }
3624                 thread = parent;
3625         }
3626
3627         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3628         if (!err && trace->filter_pids.map)
3629                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3630
3631         return err;
3632 }
3633
3634 static int trace__set_filter_pids(struct trace *trace)
3635 {
3636         int err = 0;
3637         /*
3638          * Better not use !target__has_task() here because we need to cover the
3639          * case where no threads were specified in the command line, but a
3640          * workload was, and in that case we will fill in the thread_map when
3641          * we fork the workload in evlist__prepare_workload.
3642          */
3643         if (trace->filter_pids.nr > 0) {
3644                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3645                                                     trace->filter_pids.entries);
3646                 if (!err && trace->filter_pids.map) {
3647                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3648                                                        trace->filter_pids.entries);
3649                 }
3650         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3651                 err = trace__set_filter_loop_pids(trace);
3652         }
3653
3654         return err;
3655 }
3656
3657 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3658 {
3659         struct evlist *evlist = trace->evlist;
3660         struct perf_sample sample;
3661         int err = evlist__parse_sample(evlist, event, &sample);
3662
3663         if (err)
3664                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3665         else
3666                 trace__handle_event(trace, event, &sample);
3667
3668         return 0;
3669 }
3670
3671 static int __trace__flush_events(struct trace *trace)
3672 {
3673         u64 first = ordered_events__first_time(&trace->oe.data);
3674         u64 flush = trace->oe.last - NSEC_PER_SEC;
3675
3676         /* Is there some thing to flush.. */
3677         if (first && first < flush)
3678                 return ordered_events__flush_time(&trace->oe.data, flush);
3679
3680         return 0;
3681 }
3682
3683 static int trace__flush_events(struct trace *trace)
3684 {
3685         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3686 }
3687
3688 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3689 {
3690         int err;
3691
3692         if (!trace->sort_events)
3693                 return __trace__deliver_event(trace, event);
3694
3695         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3696         if (err && err != -1)
3697                 return err;
3698
3699         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3700         if (err)
3701                 return err;
3702
3703         return trace__flush_events(trace);
3704 }
3705
3706 static int ordered_events__deliver_event(struct ordered_events *oe,
3707                                          struct ordered_event *event)
3708 {
3709         struct trace *trace = container_of(oe, struct trace, oe.data);
3710
3711         return __trace__deliver_event(trace, event->event);
3712 }
3713
3714 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3715 {
3716         struct tep_format_field *field;
3717         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3718
3719         if (evsel->tp_format == NULL || fmt == NULL)
3720                 return NULL;
3721
3722         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3723                 if (strcmp(field->name, arg) == 0)
3724                         return fmt;
3725
3726         return NULL;
3727 }
3728
3729 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3730 {
3731         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3732
3733         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3734                 char *right = tok + 1, *right_end;
3735
3736                 if (*right == '=')
3737                         ++right;
3738
3739                 while (isspace(*right))
3740                         ++right;
3741
3742                 if (*right == '\0')
3743                         break;
3744
3745                 while (!isalpha(*left))
3746                         if (++left == tok) {
3747                                 /*
3748                                  * Bail out, can't find the name of the argument that is being
3749                                  * used in the filter, let it try to set this filter, will fail later.
3750                                  */
3751                                 return 0;
3752                         }
3753
3754                 right_end = right + 1;
3755                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3756                         ++right_end;
3757
3758                 if (isalpha(*right)) {
3759                         struct syscall_arg_fmt *fmt;
3760                         int left_size = tok - left,
3761                             right_size = right_end - right;
3762                         char arg[128];
3763
3764                         while (isspace(left[left_size - 1]))
3765                                 --left_size;
3766
3767                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3768
3769                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3770                         if (fmt == NULL) {
3771                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3772                                        arg, evsel->name, evsel->filter);
3773                                 return -1;
3774                         }
3775
3776                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3777                                  arg, (int)(right - tok), tok, right_size, right);
3778
3779                         if (fmt->strtoul) {
3780                                 u64 val;
3781                                 struct syscall_arg syscall_arg = {
3782                                         .parm = fmt->parm,
3783                                 };
3784
3785                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3786                                         char *n, expansion[19];
3787                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3788                                         int expansion_offset = right - new_filter;
3789
3790                                         pr_debug("%s", expansion);
3791
3792                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3793                                                 pr_debug(" out of memory!\n");
3794                                                 free(new_filter);
3795                                                 return -1;
3796                                         }
3797                                         if (new_filter != evsel->filter)
3798                                                 free(new_filter);
3799                                         left = n + expansion_offset + expansion_lenght;
3800                                         new_filter = n;
3801                                 } else {
3802                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3803                                                right_size, right, arg, evsel->name, evsel->filter);
3804                                         return -1;
3805                                 }
3806                         } else {
3807                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3808                                        arg, evsel->name, evsel->filter);
3809                                 return -1;
3810                         }
3811
3812                         pr_debug("\n");
3813                 } else {
3814                         left = right_end;
3815                 }
3816         }
3817
3818         if (new_filter != evsel->filter) {
3819                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3820                 evsel__set_filter(evsel, new_filter);
3821                 free(new_filter);
3822         }
3823
3824         return 0;
3825 }
3826
3827 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3828 {
3829         struct evlist *evlist = trace->evlist;
3830         struct evsel *evsel;
3831
3832         evlist__for_each_entry(evlist, evsel) {
3833                 if (evsel->filter == NULL)
3834                         continue;
3835
3836                 if (trace__expand_filter(trace, evsel)) {
3837                         *err_evsel = evsel;
3838                         return -1;
3839                 }
3840         }
3841
3842         return 0;
3843 }
3844
3845 static int trace__run(struct trace *trace, int argc, const char **argv)
3846 {
3847         struct evlist *evlist = trace->evlist;
3848         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3849         int err = -1, i;
3850         unsigned long before;
3851         const bool forks = argc > 0;
3852         bool draining = false;
3853
3854         trace->live = true;
3855
3856         if (!trace->raw_augmented_syscalls) {
3857                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3858                         goto out_error_raw_syscalls;
3859
3860                 if (trace->trace_syscalls)
3861                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3862         }
3863
3864         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3865                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3866                 if (pgfault_maj == NULL)
3867                         goto out_error_mem;
3868                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3869                 evlist__add(evlist, pgfault_maj);
3870         }
3871
3872         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3873                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3874                 if (pgfault_min == NULL)
3875                         goto out_error_mem;
3876                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3877                 evlist__add(evlist, pgfault_min);
3878         }
3879
3880         /* Enable ignoring missing threads when -u/-p option is defined. */
3881         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3882
3883         if (trace->sched &&
3884             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3885                 goto out_error_sched_stat_runtime;
3886         /*
3887          * If a global cgroup was set, apply it to all the events without an
3888          * explicit cgroup. I.e.:
3889          *
3890          *      trace -G A -e sched:*switch
3891          *
3892          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3893          * _and_ sched:sched_switch to the 'A' cgroup, while:
3894          *
3895          * trace -e sched:*switch -G A
3896          *
3897          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3898          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3899          * a cgroup (on the root cgroup, sys wide, etc).
3900          *
3901          * Multiple cgroups:
3902          *
3903          * trace -G A -e sched:*switch -G B
3904          *
3905          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3906          * to the 'B' cgroup.
3907          *
3908          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3909          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3910          */
3911         if (trace->cgroup)
3912                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3913
3914         err = evlist__create_maps(evlist, &trace->opts.target);
3915         if (err < 0) {
3916                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3917                 goto out_delete_evlist;
3918         }
3919
3920         err = trace__symbols_init(trace, evlist);
3921         if (err < 0) {
3922                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3923                 goto out_delete_evlist;
3924         }
3925
3926         evlist__config(evlist, &trace->opts, &callchain_param);
3927
3928         if (forks) {
3929                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3930                 if (err < 0) {
3931                         fprintf(trace->output, "Couldn't run the workload!\n");
3932                         goto out_delete_evlist;
3933                 }
3934                 workload_pid = evlist->workload.pid;
3935         }
3936
3937         err = evlist__open(evlist);
3938         if (err < 0)
3939                 goto out_error_open;
3940
3941         err = bpf__apply_obj_config();
3942         if (err) {
3943                 char errbuf[BUFSIZ];
3944
3945                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3946                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3947                          errbuf);
3948                 goto out_error_open;
3949         }
3950
3951         err = trace__set_filter_pids(trace);
3952         if (err < 0)
3953                 goto out_error_mem;
3954
3955         if (trace->syscalls.prog_array.sys_enter)
3956                 trace__init_syscalls_bpf_prog_array_maps(trace);
3957
3958         if (trace->ev_qualifier_ids.nr > 0) {
3959                 err = trace__set_ev_qualifier_filter(trace);
3960                 if (err < 0)
3961                         goto out_errno;
3962
3963                 if (trace->syscalls.events.sys_exit) {
3964                         pr_debug("event qualifier tracepoint filter: %s\n",
3965                                  trace->syscalls.events.sys_exit->filter);
3966                 }
3967         }
3968
3969         /*
3970          * If the "close" syscall is not traced, then we will not have the
3971          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3972          * fd->pathname table and were ending up showing the last value set by
3973          * syscalls opening a pathname and associating it with a descriptor or
3974          * reading it from /proc/pid/fd/ in cases where that doesn't make
3975          * sense.
3976          *
3977          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3978          *  not in use.
3979          */
3980         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3981
3982         err = trace__expand_filters(trace, &evsel);
3983         if (err)
3984                 goto out_delete_evlist;
3985         err = evlist__apply_filters(evlist, &evsel);
3986         if (err < 0)
3987                 goto out_error_apply_filters;
3988
3989         if (trace->dump.map)
3990                 bpf_map__fprintf(trace->dump.map, trace->output);
3991
3992         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3993         if (err < 0)
3994                 goto out_error_mmap;
3995
3996         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3997                 evlist__enable(evlist);
3998
3999         if (forks)
4000                 evlist__start_workload(evlist);
4001
4002         if (trace->opts.initial_delay) {
4003                 usleep(trace->opts.initial_delay * 1000);
4004                 evlist__enable(evlist);
4005         }
4006
4007         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4008                 perf_thread_map__nr(evlist->core.threads) > 1 ||
4009                 evlist__first(evlist)->core.attr.inherit;
4010
4011         /*
4012          * Now that we already used evsel->core.attr to ask the kernel to setup the
4013          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4014          * trace__resolve_callchain(), allowing per-event max-stack settings
4015          * to override an explicitly set --max-stack global setting.
4016          */
4017         evlist__for_each_entry(evlist, evsel) {
4018                 if (evsel__has_callchain(evsel) &&
4019                     evsel->core.attr.sample_max_stack == 0)
4020                         evsel->core.attr.sample_max_stack = trace->max_stack;
4021         }
4022 again:
4023         before = trace->nr_events;
4024
4025         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4026                 union perf_event *event;
4027                 struct mmap *md;
4028
4029                 md = &evlist->mmap[i];
4030                 if (perf_mmap__read_init(&md->core) < 0)
4031                         continue;
4032
4033                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4034                         ++trace->nr_events;
4035
4036                         err = trace__deliver_event(trace, event);
4037                         if (err)
4038                                 goto out_disable;
4039
4040                         perf_mmap__consume(&md->core);
4041
4042                         if (interrupted)
4043                                 goto out_disable;
4044
4045                         if (done && !draining) {
4046                                 evlist__disable(evlist);
4047                                 draining = true;
4048                         }
4049                 }
4050                 perf_mmap__read_done(&md->core);
4051         }
4052
4053         if (trace->nr_events == before) {
4054                 int timeout = done ? 100 : -1;
4055
4056                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4057                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4058                                 draining = true;
4059
4060                         goto again;
4061                 } else {
4062                         if (trace__flush_events(trace))
4063                                 goto out_disable;
4064                 }
4065         } else {
4066                 goto again;
4067         }
4068
4069 out_disable:
4070         thread__zput(trace->current);
4071
4072         evlist__disable(evlist);
4073
4074         if (trace->sort_events)
4075                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4076
4077         if (!err) {
4078                 if (trace->summary)
4079                         trace__fprintf_thread_summary(trace, trace->output);
4080
4081                 if (trace->show_tool_stats) {
4082                         fprintf(trace->output, "Stats:\n "
4083                                                " vfs_getname : %" PRIu64 "\n"
4084                                                " proc_getname: %" PRIu64 "\n",
4085                                 trace->stats.vfs_getname,
4086                                 trace->stats.proc_getname);
4087                 }
4088         }
4089
4090 out_delete_evlist:
4091         trace__symbols__exit(trace);
4092         evlist__free_syscall_tp_fields(evlist);
4093         evlist__delete(evlist);
4094         cgroup__put(trace->cgroup);
4095         trace->evlist = NULL;
4096         trace->live = false;
4097         return err;
4098 {
4099         char errbuf[BUFSIZ];
4100
4101 out_error_sched_stat_runtime:
4102         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4103         goto out_error;
4104
4105 out_error_raw_syscalls:
4106         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4107         goto out_error;
4108
4109 out_error_mmap:
4110         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4111         goto out_error;
4112
4113 out_error_open:
4114         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4115
4116 out_error:
4117         fprintf(trace->output, "%s\n", errbuf);
4118         goto out_delete_evlist;
4119
4120 out_error_apply_filters:
4121         fprintf(trace->output,
4122                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4123                 evsel->filter, evsel__name(evsel), errno,
4124                 str_error_r(errno, errbuf, sizeof(errbuf)));
4125         goto out_delete_evlist;
4126 }
4127 out_error_mem:
4128         fprintf(trace->output, "Not enough memory to run!\n");
4129         goto out_delete_evlist;
4130
4131 out_errno:
4132         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4133         goto out_delete_evlist;
4134 }
4135
4136 static int trace__replay(struct trace *trace)
4137 {
4138         const struct evsel_str_handler handlers[] = {
4139                 { "probe:vfs_getname",       trace__vfs_getname, },
4140         };
4141         struct perf_data data = {
4142                 .path  = input_name,
4143                 .mode  = PERF_DATA_MODE_READ,
4144                 .force = trace->force,
4145         };
4146         struct perf_session *session;
4147         struct evsel *evsel;
4148         int err = -1;
4149
4150         trace->tool.sample        = trace__process_sample;
4151         trace->tool.mmap          = perf_event__process_mmap;
4152         trace->tool.mmap2         = perf_event__process_mmap2;
4153         trace->tool.comm          = perf_event__process_comm;
4154         trace->tool.exit          = perf_event__process_exit;
4155         trace->tool.fork          = perf_event__process_fork;
4156         trace->tool.attr          = perf_event__process_attr;
4157         trace->tool.tracing_data  = perf_event__process_tracing_data;
4158         trace->tool.build_id      = perf_event__process_build_id;
4159         trace->tool.namespaces    = perf_event__process_namespaces;
4160
4161         trace->tool.ordered_events = true;
4162         trace->tool.ordering_requires_timestamps = true;
4163
4164         /* add tid to output */
4165         trace->multiple_threads = true;
4166
4167         session = perf_session__new(&data, &trace->tool);
4168         if (IS_ERR(session))
4169                 return PTR_ERR(session);
4170
4171         if (trace->opts.target.pid)
4172                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4173
4174         if (trace->opts.target.tid)
4175                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4176
4177         if (symbol__init(&session->header.env) < 0)
4178                 goto out;
4179
4180         trace->host = &session->machines.host;
4181
4182         err = perf_session__set_tracepoints_handlers(session, handlers);
4183         if (err)
4184                 goto out;
4185
4186         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4187         trace->syscalls.events.sys_enter = evsel;
4188         /* older kernels have syscalls tp versus raw_syscalls */
4189         if (evsel == NULL)
4190                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4191
4192         if (evsel &&
4193             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4194             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4195                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4196                 goto out;
4197         }
4198
4199         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4200         trace->syscalls.events.sys_exit = evsel;
4201         if (evsel == NULL)
4202                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4203         if (evsel &&
4204             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4205             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4206                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4207                 goto out;
4208         }
4209
4210         evlist__for_each_entry(session->evlist, evsel) {
4211                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4212                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4213                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4214                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4215                         evsel->handler = trace__pgfault;
4216         }
4217
4218         setup_pager();
4219
4220         err = perf_session__process_events(session);
4221         if (err)
4222                 pr_err("Failed to process events, error %d", err);
4223
4224         else if (trace->summary)
4225                 trace__fprintf_thread_summary(trace, trace->output);
4226
4227 out:
4228         perf_session__delete(session);
4229
4230         return err;
4231 }
4232
4233 static size_t trace__fprintf_threads_header(FILE *fp)
4234 {
4235         size_t printed;
4236
4237         printed  = fprintf(fp, "\n Summary of events:\n\n");
4238
4239         return printed;
4240 }
4241
4242 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4243         struct syscall_stats *stats;
4244         double               msecs;
4245         int                  syscall;
4246 )
4247 {
4248         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4249         struct syscall_stats *stats = source->priv;
4250
4251         entry->syscall = source->i;
4252         entry->stats   = stats;
4253         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4254 }
4255
4256 static size_t thread__dump_stats(struct thread_trace *ttrace,
4257                                  struct trace *trace, FILE *fp)
4258 {
4259         size_t printed = 0;
4260         struct syscall *sc;
4261         struct rb_node *nd;
4262         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4263
4264         if (syscall_stats == NULL)
4265                 return 0;
4266
4267         printed += fprintf(fp, "\n");
4268
4269         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4270         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4271         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4272
4273         resort_rb__for_each_entry(nd, syscall_stats) {
4274                 struct syscall_stats *stats = syscall_stats_entry->stats;
4275                 if (stats) {
4276                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4277                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4278                         double avg = avg_stats(&stats->stats);
4279                         double pct;
4280                         u64 n = (u64)stats->stats.n;
4281
4282                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4283                         avg /= NSEC_PER_MSEC;
4284
4285                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4286                         printed += fprintf(fp, "   %-15s", sc->name);
4287                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4288                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4289                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4290
4291                         if (trace->errno_summary && stats->nr_failures) {
4292                                 const char *arch_name = perf_env__arch(trace->host->env);
4293                                 int e;
4294
4295                                 for (e = 0; e < stats->max_errno; ++e) {
4296                                         if (stats->errnos[e] != 0)
4297                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4298                                 }
4299                         }
4300                 }
4301         }
4302
4303         resort_rb__delete(syscall_stats);
4304         printed += fprintf(fp, "\n\n");
4305
4306         return printed;
4307 }
4308
4309 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4310 {
4311         size_t printed = 0;
4312         struct thread_trace *ttrace = thread__priv(thread);
4313         double ratio;
4314
4315         if (ttrace == NULL)
4316                 return 0;
4317
4318         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4319
4320         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4321         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4322         printed += fprintf(fp, "%.1f%%", ratio);
4323         if (ttrace->pfmaj)
4324                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4325         if (ttrace->pfmin)
4326                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4327         if (trace->sched)
4328                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4329         else if (fputc('\n', fp) != EOF)
4330                 ++printed;
4331
4332         printed += thread__dump_stats(ttrace, trace, fp);
4333
4334         return printed;
4335 }
4336
4337 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4338 {
4339         return ttrace ? ttrace->nr_events : 0;
4340 }
4341
4342 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4343         struct thread *thread;
4344 )
4345 {
4346         entry->thread = rb_entry(nd, struct thread, rb_node);
4347 }
4348
4349 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4350 {
4351         size_t printed = trace__fprintf_threads_header(fp);
4352         struct rb_node *nd;
4353         int i;
4354
4355         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4356                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4357
4358                 if (threads == NULL) {
4359                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4360                         return 0;
4361                 }
4362
4363                 resort_rb__for_each_entry(nd, threads)
4364                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4365
4366                 resort_rb__delete(threads);
4367         }
4368         return printed;
4369 }
4370
4371 static int trace__set_duration(const struct option *opt, const char *str,
4372                                int unset __maybe_unused)
4373 {
4374         struct trace *trace = opt->value;
4375
4376         trace->duration_filter = atof(str);
4377         return 0;
4378 }
4379
4380 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4381                                               int unset __maybe_unused)
4382 {
4383         int ret = -1;
4384         size_t i;
4385         struct trace *trace = opt->value;
4386         /*
4387          * FIXME: introduce a intarray class, plain parse csv and create a
4388          * { int nr, int entries[] } struct...
4389          */
4390         struct intlist *list = intlist__new(str);
4391
4392         if (list == NULL)
4393                 return -1;
4394
4395         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4396         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4397
4398         if (trace->filter_pids.entries == NULL)
4399                 goto out;
4400
4401         trace->filter_pids.entries[0] = getpid();
4402
4403         for (i = 1; i < trace->filter_pids.nr; ++i)
4404                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4405
4406         intlist__delete(list);
4407         ret = 0;
4408 out:
4409         return ret;
4410 }
4411
4412 static int trace__open_output(struct trace *trace, const char *filename)
4413 {
4414         struct stat st;
4415
4416         if (!stat(filename, &st) && st.st_size) {
4417                 char oldname[PATH_MAX];
4418
4419                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4420                 unlink(oldname);
4421                 rename(filename, oldname);
4422         }
4423
4424         trace->output = fopen(filename, "w");
4425
4426         return trace->output == NULL ? -errno : 0;
4427 }
4428
4429 static int parse_pagefaults(const struct option *opt, const char *str,
4430                             int unset __maybe_unused)
4431 {
4432         int *trace_pgfaults = opt->value;
4433
4434         if (strcmp(str, "all") == 0)
4435                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4436         else if (strcmp(str, "maj") == 0)
4437                 *trace_pgfaults |= TRACE_PFMAJ;
4438         else if (strcmp(str, "min") == 0)
4439                 *trace_pgfaults |= TRACE_PFMIN;
4440         else
4441                 return -1;
4442
4443         return 0;
4444 }
4445
4446 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4447 {
4448         struct evsel *evsel;
4449
4450         evlist__for_each_entry(evlist, evsel) {
4451                 if (evsel->handler == NULL)
4452                         evsel->handler = handler;
4453         }
4454 }
4455
4456 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4457 {
4458         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4459
4460         if (fmt) {
4461                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4462
4463                 if (scfmt) {
4464                         int skip = 0;
4465
4466                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4467                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4468                                 ++skip;
4469
4470                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4471                 }
4472         }
4473 }
4474
4475 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4476 {
4477         struct evsel *evsel;
4478
4479         evlist__for_each_entry(evlist, evsel) {
4480                 if (evsel->priv || !evsel->tp_format)
4481                         continue;
4482
4483                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4484                         evsel__init_tp_arg_scnprintf(evsel);
4485                         continue;
4486                 }
4487
4488                 if (evsel__init_syscall_tp(evsel))
4489                         return -1;
4490
4491                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4492                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4493
4494                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4495                                 return -1;
4496
4497                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4498                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4499                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4500
4501                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4502                                 return -1;
4503
4504                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4505                 }
4506         }
4507
4508         return 0;
4509 }
4510
4511 /*
4512  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4513  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4514  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4515  *
4516  * It'd be better to introduce a parse_options() variant that would return a
4517  * list with the terms it didn't match to an event...
4518  */
4519 static int trace__parse_events_option(const struct option *opt, const char *str,
4520                                       int unset __maybe_unused)
4521 {
4522         struct trace *trace = (struct trace *)opt->value;
4523         const char *s = str;
4524         char *sep = NULL, *lists[2] = { NULL, NULL, };
4525         int len = strlen(str) + 1, err = -1, list, idx;
4526         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4527         char group_name[PATH_MAX];
4528         struct syscall_fmt *fmt;
4529
4530         if (strace_groups_dir == NULL)
4531                 return -1;
4532
4533         if (*s == '!') {
4534                 ++s;
4535                 trace->not_ev_qualifier = true;
4536         }
4537
4538         while (1) {
4539                 if ((sep = strchr(s, ',')) != NULL)
4540                         *sep = '\0';
4541
4542                 list = 0;
4543                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4544                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4545                         list = 1;
4546                         goto do_concat;
4547                 }
4548
4549                 fmt = syscall_fmt__find_by_alias(s);
4550                 if (fmt != NULL) {
4551                         list = 1;
4552                         s = fmt->name;
4553                 } else {
4554                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4555                         if (access(group_name, R_OK) == 0)
4556                                 list = 1;
4557                 }
4558 do_concat:
4559                 if (lists[list]) {
4560                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4561                 } else {
4562                         lists[list] = malloc(len);
4563                         if (lists[list] == NULL)
4564                                 goto out;
4565                         strcpy(lists[list], s);
4566                 }
4567
4568                 if (!sep)
4569                         break;
4570
4571                 *sep = ',';
4572                 s = sep + 1;
4573         }
4574
4575         if (lists[1] != NULL) {
4576                 struct strlist_config slist_config = {
4577                         .dirname = strace_groups_dir,
4578                 };
4579
4580                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4581                 if (trace->ev_qualifier == NULL) {
4582                         fputs("Not enough memory to parse event qualifier", trace->output);
4583                         goto out;
4584                 }
4585
4586                 if (trace__validate_ev_qualifier(trace))
4587                         goto out;
4588                 trace->trace_syscalls = true;
4589         }
4590
4591         err = 0;
4592
4593         if (lists[0]) {
4594                 struct option o = {
4595                         .value = &trace->evlist,
4596                 };
4597                 err = parse_events_option(&o, lists[0], 0);
4598         }
4599 out:
4600         free(strace_groups_dir);
4601         free(lists[0]);
4602         free(lists[1]);
4603         if (sep)
4604                 *sep = ',';
4605
4606         return err;
4607 }
4608
4609 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4610 {
4611         struct trace *trace = opt->value;
4612
4613         if (!list_empty(&trace->evlist->core.entries)) {
4614                 struct option o = {
4615                         .value = &trace->evlist,
4616                 };
4617                 return parse_cgroups(&o, str, unset);
4618         }
4619         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4620
4621         return 0;
4622 }
4623
4624 static int trace__config(const char *var, const char *value, void *arg)
4625 {
4626         struct trace *trace = arg;
4627         int err = 0;
4628
4629         if (!strcmp(var, "trace.add_events")) {
4630                 trace->perfconfig_events = strdup(value);
4631                 if (trace->perfconfig_events == NULL) {
4632                         pr_err("Not enough memory for %s\n", "trace.add_events");
4633                         return -1;
4634                 }
4635         } else if (!strcmp(var, "trace.show_timestamp")) {
4636                 trace->show_tstamp = perf_config_bool(var, value);
4637         } else if (!strcmp(var, "trace.show_duration")) {
4638                 trace->show_duration = perf_config_bool(var, value);
4639         } else if (!strcmp(var, "trace.show_arg_names")) {
4640                 trace->show_arg_names = perf_config_bool(var, value);
4641                 if (!trace->show_arg_names)
4642                         trace->show_zeros = true;
4643         } else if (!strcmp(var, "trace.show_zeros")) {
4644                 bool new_show_zeros = perf_config_bool(var, value);
4645                 if (!trace->show_arg_names && !new_show_zeros) {
4646                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4647                         goto out;
4648                 }
4649                 trace->show_zeros = new_show_zeros;
4650         } else if (!strcmp(var, "trace.show_prefix")) {
4651                 trace->show_string_prefix = perf_config_bool(var, value);
4652         } else if (!strcmp(var, "trace.no_inherit")) {
4653                 trace->opts.no_inherit = perf_config_bool(var, value);
4654         } else if (!strcmp(var, "trace.args_alignment")) {
4655                 int args_alignment = 0;
4656                 if (perf_config_int(&args_alignment, var, value) == 0)
4657                         trace->args_alignment = args_alignment;
4658         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4659                 if (strcasecmp(value, "libtraceevent") == 0)
4660                         trace->libtraceevent_print = true;
4661                 else if (strcasecmp(value, "libbeauty") == 0)
4662                         trace->libtraceevent_print = false;
4663         }
4664 out:
4665         return err;
4666 }
4667
4668 static void trace__exit(struct trace *trace)
4669 {
4670         int i;
4671
4672         strlist__delete(trace->ev_qualifier);
4673         free(trace->ev_qualifier_ids.entries);
4674         if (trace->syscalls.table) {
4675                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4676                         syscall__exit(&trace->syscalls.table[i]);
4677                 free(trace->syscalls.table);
4678         }
4679         syscalltbl__delete(trace->sctbl);
4680         zfree(&trace->perfconfig_events);
4681 }
4682
4683 int cmd_trace(int argc, const char **argv)
4684 {
4685         const char *trace_usage[] = {
4686                 "perf trace [<options>] [<command>]",
4687                 "perf trace [<options>] -- <command> [<options>]",
4688                 "perf trace record [<options>] [<command>]",
4689                 "perf trace record [<options>] -- <command> [<options>]",
4690                 NULL
4691         };
4692         struct trace trace = {
4693                 .opts = {
4694                         .target = {
4695                                 .uid       = UINT_MAX,
4696                                 .uses_mmap = true,
4697                         },
4698                         .user_freq     = UINT_MAX,
4699                         .user_interval = ULLONG_MAX,
4700                         .no_buffering  = true,
4701                         .mmap_pages    = UINT_MAX,
4702                 },
4703                 .output = stderr,
4704                 .show_comm = true,
4705                 .show_tstamp = true,
4706                 .show_duration = true,
4707                 .show_arg_names = true,
4708                 .args_alignment = 70,
4709                 .trace_syscalls = false,
4710                 .kernel_syscallchains = false,
4711                 .max_stack = UINT_MAX,
4712                 .max_events = ULONG_MAX,
4713         };
4714         const char *map_dump_str = NULL;
4715         const char *output_name = NULL;
4716         const struct option trace_options[] = {
4717         OPT_CALLBACK('e', "event", &trace, "event",
4718                      "event/syscall selector. use 'perf list' to list available events",
4719                      trace__parse_events_option),
4720         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4721                      "event filter", parse_filter),
4722         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4723                     "show the thread COMM next to its id"),
4724         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4725         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4726                      trace__parse_events_option),
4727         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4728         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4729         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4730                     "trace events on existing process id"),
4731         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4732                     "trace events on existing thread id"),
4733         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4734                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4735         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4736                     "system-wide collection from all CPUs"),
4737         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4738                     "list of cpus to monitor"),
4739         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4740                     "child tasks do not inherit counters"),
4741         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4742                      "number of mmap data pages", evlist__parse_mmap_pages),
4743         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4744                    "user to profile"),
4745         OPT_CALLBACK(0, "duration", &trace, "float",
4746                      "show only events with duration > N.M ms",
4747                      trace__set_duration),
4748 #ifdef HAVE_LIBBPF_SUPPORT
4749         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4750 #endif
4751         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4752         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4753         OPT_BOOLEAN('T', "time", &trace.full_time,
4754                     "Show full timestamp, not time relative to first start"),
4755         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4756                     "Show only syscalls that failed"),
4757         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4758                     "Show only syscall summary with statistics"),
4759         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4760                     "Show all syscalls and summary with statistics"),
4761         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4762                     "Show errno stats per syscall, use with -s or -S"),
4763         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4764                      "Trace pagefaults", parse_pagefaults, "maj"),
4765         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4766         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4767         OPT_CALLBACK(0, "call-graph", &trace.opts,
4768                      "record_mode[,record_size]", record_callchain_help,
4769                      &record_parse_callchain_opt),
4770         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4771                     "Use libtraceevent to print the tracepoint arguments."),
4772         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4773                     "Show the kernel callchains on the syscall exit path"),
4774         OPT_ULONG(0, "max-events", &trace.max_events,
4775                 "Set the maximum number of events to print, exit after that is reached. "),
4776         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4777                      "Set the minimum stack depth when parsing the callchain, "
4778                      "anything below the specified depth will be ignored."),
4779         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4780                      "Set the maximum stack depth when parsing the callchain, "
4781                      "anything beyond the specified depth will be ignored. "
4782                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4783         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4784                         "Sort batch of events before processing, use if getting out of order events"),
4785         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4786                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4787         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4788                         "per thread proc mmap processing timeout in ms"),
4789         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4790                      trace__parse_cgroups),
4791         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4792                      "ms to wait before starting measurement after program "
4793                      "start"),
4794         OPTS_EVSWITCH(&trace.evswitch),
4795         OPT_END()
4796         };
4797         bool __maybe_unused max_stack_user_set = true;
4798         bool mmap_pages_user_set = true;
4799         struct evsel *evsel;
4800         const char * const trace_subcommands[] = { "record", NULL };
4801         int err = -1;
4802         char bf[BUFSIZ];
4803         struct sigaction sigchld_act;
4804
4805         signal(SIGSEGV, sighandler_dump_stack);
4806         signal(SIGFPE, sighandler_dump_stack);
4807         signal(SIGINT, sighandler_interrupt);
4808
4809         memset(&sigchld_act, 0, sizeof(sigchld_act));
4810         sigchld_act.sa_flags = SA_SIGINFO;
4811         sigchld_act.sa_sigaction = sighandler_chld;
4812         sigaction(SIGCHLD, &sigchld_act, NULL);
4813
4814         trace.evlist = evlist__new();
4815         trace.sctbl = syscalltbl__new();
4816
4817         if (trace.evlist == NULL || trace.sctbl == NULL) {
4818                 pr_err("Not enough memory to run!\n");
4819                 err = -ENOMEM;
4820                 goto out;
4821         }
4822
4823         /*
4824          * Parsing .perfconfig may entail creating a BPF event, that may need
4825          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4826          * is too small. This affects just this process, not touching the
4827          * global setting. If it fails we'll get something in 'perf trace -v'
4828          * to help diagnose the problem.
4829          */
4830         rlimit__bump_memlock();
4831
4832         err = perf_config(trace__config, &trace);
4833         if (err)
4834                 goto out;
4835
4836         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4837                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4838
4839         /*
4840          * Here we already passed thru trace__parse_events_option() and it has
4841          * already figured out if -e syscall_name, if not but if --event
4842          * foo:bar was used, the user is interested _just_ in those, say,
4843          * tracepoint events, not in the strace-like syscall-name-based mode.
4844          *
4845          * This is important because we need to check if strace-like mode is
4846          * needed to decided if we should filter out the eBPF
4847          * __augmented_syscalls__ code, if it is in the mix, say, via
4848          * .perfconfig trace.add_events, and filter those out.
4849          */
4850         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4851             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4852                 trace.trace_syscalls = true;
4853         }
4854         /*
4855          * Now that we have --verbose figured out, lets see if we need to parse
4856          * events from .perfconfig, so that if those events fail parsing, say some
4857          * BPF program fails, then we'll be able to use --verbose to see what went
4858          * wrong in more detail.
4859          */
4860         if (trace.perfconfig_events != NULL) {
4861                 struct parse_events_error parse_err;
4862
4863                 parse_events_error__init(&parse_err);
4864                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4865                 if (err)
4866                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4867                 parse_events_error__exit(&parse_err);
4868                 if (err)
4869                         goto out;
4870         }
4871
4872         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4873                 usage_with_options_msg(trace_usage, trace_options,
4874                                        "cgroup monitoring only available in system-wide mode");
4875         }
4876
4877         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4878         if (IS_ERR(evsel)) {
4879                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4880                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4881                 goto out;
4882         }
4883
4884         if (evsel) {
4885                 trace.syscalls.events.augmented = evsel;
4886
4887                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4888                 if (evsel == NULL) {
4889                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4890                         goto out;
4891                 }
4892
4893                 if (evsel->bpf_obj == NULL) {
4894                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4895                         goto out;
4896                 }
4897
4898                 trace.bpf_obj = evsel->bpf_obj;
4899
4900                 /*
4901                  * If we have _just_ the augmenter event but don't have a
4902                  * explicit --syscalls, then assume we want all strace-like
4903                  * syscalls:
4904                  */
4905                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4906                         trace.trace_syscalls = true;
4907                 /*
4908                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4909                  * strace-like syscall tracing is not set, then we need to trow
4910                  * away the augmenter, i.e. all the events that were created
4911                  * from that BPF object file.
4912                  *
4913                  * This is more to fix the current .perfconfig trace.add_events
4914                  * style of setting up the strace-like eBPF based syscall point
4915                  * payload augmenter.
4916                  *
4917                  * All this complexity will be avoided by adding an alternative
4918                  * to trace.add_events in the form of
4919                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4920                  * need it.
4921                  *
4922                  * .perfconfig trace.add_events is still useful if we want, for
4923                  * instance, have msr_write.msr in some .perfconfig profile based
4924                  * 'perf trace --config determinism.profile' mode, where for some
4925                  * particular goal/workload type we want a set of events and
4926                  * output mode (with timings, etc) instead of having to add
4927                  * all via the command line.
4928                  *
4929                  * Also --config to specify an alternate .perfconfig file needs
4930                  * to be implemented.
4931                  */
4932                 if (!trace.trace_syscalls) {
4933                         trace__delete_augmented_syscalls(&trace);
4934                 } else {
4935                         trace__set_bpf_map_filtered_pids(&trace);
4936                         trace__set_bpf_map_syscalls(&trace);
4937                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4938                 }
4939         }
4940
4941         err = bpf__setup_stdout(trace.evlist);
4942         if (err) {
4943                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4944                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4945                 goto out;
4946         }
4947
4948         err = -1;
4949
4950         if (map_dump_str) {
4951                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4952                 if (trace.dump.map == NULL) {
4953                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4954                         goto out;
4955                 }
4956         }
4957
4958         if (trace.trace_pgfaults) {
4959                 trace.opts.sample_address = true;
4960                 trace.opts.sample_time = true;
4961         }
4962
4963         if (trace.opts.mmap_pages == UINT_MAX)
4964                 mmap_pages_user_set = false;
4965
4966         if (trace.max_stack == UINT_MAX) {
4967                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4968                 max_stack_user_set = false;
4969         }
4970
4971 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4972         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4973                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4974         }
4975 #endif
4976
4977         if (callchain_param.enabled) {
4978                 if (!mmap_pages_user_set && geteuid() == 0)
4979                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4980
4981                 symbol_conf.use_callchain = true;
4982         }
4983
4984         if (trace.evlist->core.nr_entries > 0) {
4985                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4986                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4987                         perror("failed to set syscalls:* tracepoint fields");
4988                         goto out;
4989                 }
4990         }
4991
4992         if (trace.sort_events) {
4993                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4994                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4995         }
4996
4997         /*
4998          * If we are augmenting syscalls, then combine what we put in the
4999          * __augmented_syscalls__ BPF map with what is in the
5000          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5001          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5002          *
5003          * We'll switch to look at two BPF maps, one for sys_enter and the
5004          * other for sys_exit when we start augmenting the sys_exit paths with
5005          * buffers that are being copied from kernel to userspace, think 'read'
5006          * syscall.
5007          */
5008         if (trace.syscalls.events.augmented) {
5009                 evlist__for_each_entry(trace.evlist, evsel) {
5010                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5011
5012                         if (raw_syscalls_sys_exit) {
5013                                 trace.raw_augmented_syscalls = true;
5014                                 goto init_augmented_syscall_tp;
5015                         }
5016
5017                         if (trace.syscalls.events.augmented->priv == NULL &&
5018                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5019                                 struct evsel *augmented = trace.syscalls.events.augmented;
5020                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5021                                     evsel__init_augmented_syscall_tp_args(augmented))
5022                                         goto out;
5023                                 /*
5024                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5025                                  * Above we made sure we can get from the payload the tp fields
5026                                  * that we get from syscalls:sys_enter tracefs format file.
5027                                  */
5028                                 augmented->handler = trace__sys_enter;
5029                                 /*
5030                                  * Now we do the same for the *syscalls:sys_enter event so that
5031                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5032                                  * as not to filter it, then we'll handle it just like we would
5033                                  * for the BPF_OUTPUT one:
5034                                  */
5035                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5036                                     evsel__init_augmented_syscall_tp_args(evsel))
5037                                         goto out;
5038                                 evsel->handler = trace__sys_enter;
5039                         }
5040
5041                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5042                                 struct syscall_tp *sc;
5043 init_augmented_syscall_tp:
5044                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5045                                         goto out;
5046                                 sc = __evsel__syscall_tp(evsel);
5047                                 /*
5048                                  * For now with BPF raw_augmented we hook into
5049                                  * raw_syscalls:sys_enter and there we get all
5050                                  * 6 syscall args plus the tracepoint common
5051                                  * fields and the syscall_nr (another long).
5052                                  * So we check if that is the case and if so
5053                                  * don't look after the sc->args_size but
5054                                  * always after the full raw_syscalls:sys_enter
5055                                  * payload, which is fixed.
5056                                  *
5057                                  * We'll revisit this later to pass
5058                                  * s->args_size to the BPF augmenter (now
5059                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5060                                  * so that it copies only what we need for each
5061                                  * syscall, like what happens when we use
5062                                  * syscalls:sys_enter_NAME, so that we reduce
5063                                  * the kernel/userspace traffic to just what is
5064                                  * needed for each syscall.
5065                                  */
5066                                 if (trace.raw_augmented_syscalls)
5067                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5068                                 evsel__init_augmented_syscall_tp_ret(evsel);
5069                                 evsel->handler = trace__sys_exit;
5070                         }
5071                 }
5072         }
5073
5074         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5075                 return trace__record(&trace, argc-1, &argv[1]);
5076
5077         /* Using just --errno-summary will trigger --summary */
5078         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5079                 trace.summary_only = true;
5080
5081         /* summary_only implies summary option, but don't overwrite summary if set */
5082         if (trace.summary_only)
5083                 trace.summary = trace.summary_only;
5084
5085         if (output_name != NULL) {
5086                 err = trace__open_output(&trace, output_name);
5087                 if (err < 0) {
5088                         perror("failed to create output file");
5089                         goto out;
5090                 }
5091         }
5092
5093         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5094         if (err)
5095                 goto out_close;
5096
5097         err = target__validate(&trace.opts.target);
5098         if (err) {
5099                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5100                 fprintf(trace.output, "%s", bf);
5101                 goto out_close;
5102         }
5103
5104         err = target__parse_uid(&trace.opts.target);
5105         if (err) {
5106                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5107                 fprintf(trace.output, "%s", bf);
5108                 goto out_close;
5109         }
5110
5111         if (!argc && target__none(&trace.opts.target))
5112                 trace.opts.target.system_wide = true;
5113
5114         if (input_name)
5115                 err = trace__replay(&trace);
5116         else
5117                 err = trace__run(&trace, argc, argv);
5118
5119 out_close:
5120         if (output_name != NULL)
5121                 fclose(trace.output);
5122 out:
5123         trace__exit(&trace);
5124         return err;
5125 }