tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  */
  16
  17 #include "util/record.h"
  18 #include <api/fs/tracing_path.h>
  19 #ifdef HAVE_LIBBPF_SUPPORT
  20 #include <bpf/bpf.h>
  21 #include <bpf/libbpf.h>
  22 #ifdef HAVE_BPF_SKEL
  23 #include "bpf_skel/augmented_raw_syscalls.skel.h"
  24 #endif
  25 #endif
  26 #include "util/bpf_map.h"
  27 #include "util/rlimit.h"
  28 #include "builtin.h"
  29 #include "util/cgroup.h"
  30 #include "util/color.h"
  31 #include "util/config.h"
  32 #include "util/debug.h"
  33 #include "util/dso.h"
  34 #include "util/env.h"
  35 #include "util/event.h"
  36 #include "util/evsel.h"
  37 #include "util/evsel_fprintf.h"
  38 #include "util/synthetic-events.h"
  39 #include "util/evlist.h"
  40 #include "util/evswitch.h"
  41 #include "util/mmap.h"
  42 #include <subcmd/pager.h>
  43 #include <subcmd/exec-cmd.h>
  44 #include "util/machine.h"
  45 #include "util/map.h"
  46 #include "util/symbol.h"
  47 #include "util/path.h"
  48 #include "util/session.h"
  49 #include "util/thread.h"
  50 #include <subcmd/parse-options.h>
  51 #include "util/strlist.h"
  52 #include "util/intlist.h"
  53 #include "util/thread_map.h"
  54 #include "util/stat.h"
  55 #include "util/tool.h"
  56 #include "util/util.h"
  57 #include "trace/beauty/beauty.h"
  58 #include "trace-event.h"
  59 #include "util/parse-events.h"
  60 #include "util/tracepoint.h"
  61 #include "callchain.h"
  62 #include "print_binary.h"
  63 #include "string2.h"
  64 #include "syscalltbl.h"
  65 #include "rb_resort.h"
  66 #include "../perf.h"
  67
  68 #include <errno.h>
  69 #include <inttypes.h>
  70 #include <poll.h>
  71 #include <signal.h>
  72 #include <stdlib.h>
  73 #include <string.h>
  74 #include <linux/err.h>
  75 #include <linux/filter.h>
  76 #include <linux/kernel.h>
  77 #include <linux/list_sort.h>
  78 #include <linux/random.h>
  79 #include <linux/stringify.h>
  80 #include <linux/time64.h>
  81 #include <linux/zalloc.h>
  82 #include <fcntl.h>
  83 #include <sys/sysmacros.h>
  84
  85 #include <linux/ctype.h>
  86 #include <perf/mmap.h>
  87
  88 #ifdef HAVE_LIBTRACEEVENT
  89 #include <traceevent/event-parse.h>
  90 #endif
  91
  92 #ifndef O_CLOEXEC
  93 # define O_CLOEXEC              02000000
  94 #endif
  95
  96 #ifndef F_LINUX_SPECIFIC_BASE
  97 # define F_LINUX_SPECIFIC_BASE  1024
  98 #endif
  99
 100 #define RAW_SYSCALL_ARGS_NUM    6
 101
 102 /*
 103  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
 104  */
 105 struct syscall_arg_fmt {
 106         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 107         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
 108         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 109         void       *parm;
 110         const char *name;
 111         u16        nr_entries; // for arrays
 112         bool       show_zero;
 113 };
 114
 115 struct syscall_fmt {
 116         const char *name;
 117         const char *alias;
 118         struct {
 119                 const char *sys_enter,
 120                            *sys_exit;
 121         }          bpf_prog_name;
 122         struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
 123         u8         nr_args;
 124         bool       errpid;
 125         bool       timeout;
 126         bool       hexret;
 127 };
 128
 129 struct trace {
 130         struct perf_tool        tool;
 131         struct syscalltbl       *sctbl;
 132         struct {
 133                 struct syscall  *table;
 134                 struct {
 135                         struct evsel *sys_enter,
 136                                 *sys_exit,
 137                                 *bpf_output;
 138                 }               events;
 139         } syscalls;
 140 #ifdef HAVE_BPF_SKEL
 141         struct augmented_raw_syscalls_bpf *skel;
 142 #endif
 143         struct record_opts      opts;
 144         struct evlist   *evlist;
 145         struct machine          *host;
 146         struct thread           *current;
 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
 417                 return 0;
 418         }
 419
 420         return -ENOMEM;
 421 }
 422
 423 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 424 {
 425         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 426
 427         if (sc != NULL) {
 428                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 429                 if (syscall_id == NULL)
 430                         syscall_id = evsel__field(tp, "__syscall_nr");
 431                 if (syscall_id == NULL ||
 432                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 433                         return -EINVAL;
 434
 435                 return 0;
 436         }
 437
 438         return -ENOMEM;
 439 }
 440
 441 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 442 {
 443         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 444
 445         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 446 }
 447
 448 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 449 {
 450         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 451
 452         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 453 }
 454
 455 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 456 {
 457         if (evsel__syscall_tp(evsel) != NULL) {
 458                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 459                         return -ENOENT;
 460
 461                 evsel->handler = handler;
 462                 return 0;
 463         }
 464
 465         return -ENOMEM;
 466 }
 467
 468 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 469 {
 470         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 471
 472         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 473         if (IS_ERR(evsel))
 474                 evsel = evsel__newtp("syscalls", direction);
 475
 476         if (IS_ERR(evsel))
 477                 return NULL;
 478
 479         if (evsel__init_raw_syscall_tp(evsel, handler))
 480                 goto out_delete;
 481
 482         return evsel;
 483
 484 out_delete:
 485         evsel__delete_priv(evsel);
 486         return NULL;
 487 }
 488
 489 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 490         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 491            fields->name.integer(&fields->name, sample); })
 492
 493 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 494         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 495            fields->name.pointer(&fields->name, sample); })
 496
 497 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 498 {
 499         int idx = val - sa->offset;
 500
 501         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 502                 size_t printed = scnprintf(bf, size, intfmt, val);
 503                 if (show_suffix)
 504                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 505                 return printed;
 506         }
 507
 508         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 509 }
 510
 511 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 512 {
 513         int idx = val - sa->offset;
 514
 515         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 516                 size_t printed = scnprintf(bf, size, intfmt, val);
 517                 if (show_prefix)
 518                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 519                 return printed;
 520         }
 521
 522         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 523 }
 524
 525 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 526                                                 const char *intfmt,
 527                                                 struct syscall_arg *arg)
 528 {
 529         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 530 }
 531
 532 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 533                                               struct syscall_arg *arg)
 534 {
 535         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 536 }
 537
 538 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 539
 540 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 541 {
 542         return strarray__strtoul(arg->parm, bf, size, ret);
 543 }
 544
 545 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 546 {
 547         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 548 }
 549
 550 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 551 {
 552         return strarrays__strtoul(arg->parm, bf, size, ret);
 553 }
 554
 555 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 556 {
 557         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 558 }
 559
 560 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 561 {
 562         size_t printed;
 563         int i;
 564
 565         for (i = 0; i < sas->nr_entries; ++i) {
 566                 struct strarray *sa = sas->entries[i];
 567                 int idx = val - sa->offset;
 568
 569                 if (idx >= 0 && idx < sa->nr_entries) {
 570                         if (sa->entries[idx] == NULL)
 571                                 break;
 572                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 573                 }
 574         }
 575
 576         printed = scnprintf(bf, size, intfmt, val);
 577         if (show_prefix)
 578                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 579         return printed;
 580 }
 581
 582 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 583 {
 584         int i;
 585
 586         for (i = 0; i < sa->nr_entries; ++i) {
 587                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 588                         *ret = sa->offset + i;
 589                         return true;
 590                 }
 591         }
 592
 593         return false;
 594 }
 595
 596 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 597 {
 598         u64 val = 0;
 599         char *tok = bf, *sep, *end;
 600
 601         *ret = 0;
 602
 603         while (size != 0) {
 604                 int toklen = size;
 605
 606                 sep = memchr(tok, '|', size);
 607                 if (sep != NULL) {
 608                         size -= sep - tok + 1;
 609
 610                         end = sep - 1;
 611                         while (end > tok && isspace(*end))
 612                                 --end;
 613
 614                         toklen = end - tok + 1;
 615                 }
 616
 617                 while (isspace(*tok))
 618                         ++tok;
 619
 620                 if (isalpha(*tok) || *tok == '_') {
 621                         if (!strarray__strtoul(sa, tok, toklen, &val))
 622                                 return false;
 623                 } else
 624                         val = strtoul(tok, NULL, 0);
 625
 626                 *ret |= (1 << (val - 1));
 627
 628                 if (sep == NULL)
 629                         break;
 630                 tok = sep + 1;
 631         }
 632
 633         return true;
 634 }
 635
 636 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 637 {
 638         int i;
 639
 640         for (i = 0; i < sas->nr_entries; ++i) {
 641                 struct strarray *sa = sas->entries[i];
 642
 643                 if (strarray__strtoul(sa, bf, size, ret))
 644                         return true;
 645         }
 646
 647         return false;
 648 }
 649
 650 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 651                                         struct syscall_arg *arg)
 652 {
 653         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 654 }
 655
 656 #ifndef AT_FDCWD
 657 #define AT_FDCWD        -100
 658 #endif
 659
 660 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 661                                            struct syscall_arg *arg)
 662 {
 663         int fd = arg->val;
 664         const char *prefix = "AT_FD";
 665
 666         if (fd == AT_FDCWD)
 667                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 668
 669         return syscall_arg__scnprintf_fd(bf, size, arg);
 670 }
 671
 672 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 673
 674 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 675                                               struct syscall_arg *arg);
 676
 677 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 678
 679 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 680 {
 681         return scnprintf(bf, size, "%#lx", arg->val);
 682 }
 683
 684 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 685 {
 686         if (arg->val == 0)
 687                 return scnprintf(bf, size, "NULL");
 688         return syscall_arg__scnprintf_hex(bf, size, arg);
 689 }
 690
 691 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 692 {
 693         return scnprintf(bf, size, "%d", arg->val);
 694 }
 695
 696 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 697 {
 698         return scnprintf(bf, size, "%ld", arg->val);
 699 }
 700
 701 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 702 {
 703         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 704         //     fill missing comms using thread__set_comm()...
 705         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 706         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 707 }
 708
 709 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 710
 711 static const char *bpf_cmd[] = {
 712         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 713         "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 714         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 715         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 716         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 717         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 718         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 719         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 720         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 721         "LINK_DETACH", "PROG_BIND_MAP",
 722 };
 723 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 724
 725 static const char *fsmount_flags[] = {
 726         [1] = "CLOEXEC",
 727 };
 728 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 729
 730 #include "trace/beauty/generated/fsconfig_arrays.c"
 731
 732 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 733
 734 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 735 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 736
 737 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 738 static DEFINE_STRARRAY(itimers, "ITIMER_");
 739
 740 static const char *keyctl_options[] = {
 741         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 742         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 743         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 744         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 745         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 746 };
 747 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 748
 749 static const char *whences[] = { "SET", "CUR", "END",
 750 #ifdef SEEK_DATA
 751 "DATA",
 752 #endif
 753 #ifdef SEEK_HOLE
 754 "HOLE",
 755 #endif
 756 };
 757 static DEFINE_STRARRAY(whences, "SEEK_");
 758
 759 static const char *fcntl_cmds[] = {
 760         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 761         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 762         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 763         "GETOWNER_UIDS",
 764 };
 765 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 766
 767 static const char *fcntl_linux_specific_cmds[] = {
 768         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 769         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 770         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 771 };
 772
 773 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 774
 775 static struct strarray *fcntl_cmds_arrays[] = {
 776         &strarray__fcntl_cmds,
 777         &strarray__fcntl_linux_specific_cmds,
 778 };
 779
 780 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 781
 782 static const char *rlimit_resources[] = {
 783         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 784         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 785         "RTTIME",
 786 };
 787 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 788
 789 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 790 static DEFINE_STRARRAY(sighow, "SIG_");
 791
 792 static const char *clockid[] = {
 793         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 794         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 795         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 796 };
 797 static DEFINE_STRARRAY(clockid, "CLOCK_");
 798
 799 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 800                                                  struct syscall_arg *arg)
 801 {
 802         bool show_prefix = arg->show_string_prefix;
 803         const char *suffix = "_OK";
 804         size_t printed = 0;
 805         int mode = arg->val;
 806
 807         if (mode == F_OK) /* 0 */
 808                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 809 #define P_MODE(n) \
 810         if (mode & n##_OK) { \
 811                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 812                 mode &= ~n##_OK; \
 813         }
 814
 815         P_MODE(R);
 816         P_MODE(W);
 817         P_MODE(X);
 818 #undef P_MODE
 819
 820         if (mode)
 821                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 822
 823         return printed;
 824 }
 825
 826 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 827
 828 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 829                                               struct syscall_arg *arg);
 830
 831 #define SCA_FILENAME syscall_arg__scnprintf_filename
 832
 833 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 834                                                 struct syscall_arg *arg)
 835 {
 836         bool show_prefix = arg->show_string_prefix;
 837         const char *prefix = "O_";
 838         int printed = 0, flags = arg->val;
 839
 840 #define P_FLAG(n) \
 841         if (flags & O_##n) { \
 842                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 843                 flags &= ~O_##n; \
 844         }
 845
 846         P_FLAG(CLOEXEC);
 847         P_FLAG(NONBLOCK);
 848 #undef P_FLAG
 849
 850         if (flags)
 851                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 852
 853         return printed;
 854 }
 855
 856 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 857
 858 #ifndef GRND_NONBLOCK
 859 #define GRND_NONBLOCK   0x0001
 860 #endif
 861 #ifndef GRND_RANDOM
 862 #define GRND_RANDOM     0x0002
 863 #endif
 864
 865 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 866                                                    struct syscall_arg *arg)
 867 {
 868         bool show_prefix = arg->show_string_prefix;
 869         const char *prefix = "GRND_";
 870         int printed = 0, flags = arg->val;
 871
 872 #define P_FLAG(n) \
 873         if (flags & GRND_##n) { \
 874                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 875                 flags &= ~GRND_##n; \
 876         }
 877
 878         P_FLAG(RANDOM);
 879         P_FLAG(NONBLOCK);
 880 #undef P_FLAG
 881
 882         if (flags)
 883                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 884
 885         return printed;
 886 }
 887
 888 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 889
 890 #define STRARRAY(name, array) \
 891           { .scnprintf  = SCA_STRARRAY, \
 892             .strtoul    = STUL_STRARRAY, \
 893             .parm       = &strarray__##array, }
 894
 895 #define STRARRAY_FLAGS(name, array) \
 896           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 897             .strtoul    = STUL_STRARRAY_FLAGS, \
 898             .parm       = &strarray__##array, }
 899
 900 #include "trace/beauty/arch_errno_names.c"
 901 #include "trace/beauty/eventfd.c"
 902 #include "trace/beauty/futex_op.c"
 903 #include "trace/beauty/futex_val3.c"
 904 #include "trace/beauty/mmap.c"
 905 #include "trace/beauty/mode_t.c"
 906 #include "trace/beauty/msg_flags.c"
 907 #include "trace/beauty/open_flags.c"
 908 #include "trace/beauty/perf_event_open.c"
 909 #include "trace/beauty/pid.c"
 910 #include "trace/beauty/sched_policy.c"
 911 #include "trace/beauty/seccomp.c"
 912 #include "trace/beauty/signum.c"
 913 #include "trace/beauty/socket_type.c"
 914 #include "trace/beauty/waitid_options.c"
 915
 916 static const struct syscall_fmt syscall_fmts[] = {
 917         { .name     = "access",
 918           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 919         { .name     = "arch_prctl",
 920           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 921                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 922         { .name     = "bind",
 923           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 924                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 925                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 926         { .name     = "bpf",
 927           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 928         { .name     = "brk",        .hexret = true,
 929           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 930         { .name     = "clock_gettime",
 931           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 932         { .name     = "clock_nanosleep",
 933           .arg = { [2] = { .scnprintf = SCA_TIMESPEC,  /* rqtp */ }, }, },
 934         { .name     = "clone",      .errpid = true, .nr_args = 5,
 935           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 936                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 937                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 938                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 939                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 940         { .name     = "close",
 941           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 942         { .name     = "connect",
 943           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 944                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 945                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 946         { .name     = "epoll_ctl",
 947           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 948         { .name     = "eventfd2",
 949           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 950         { .name     = "fchmodat",
 951           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 952         { .name     = "fchownat",
 953           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 954         { .name     = "fcntl",
 955           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 956                            .strtoul   = STUL_STRARRAYS,
 957                            .parm      = &strarrays__fcntl_cmds_arrays,
 958                            .show_zero = true, },
 959                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 960         { .name     = "flock",
 961           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 962         { .name     = "fsconfig",
 963           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 964         { .name     = "fsmount",
 965           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 966                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 967         { .name     = "fspick",
 968           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 969                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 970                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 971         { .name     = "fstat", .alias = "newfstat", },
 972         { .name     = "fstatat", .alias = "newfstatat", },
 973         { .name     = "futex",
 974           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 975                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 976         { .name     = "futimesat",
 977           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 978         { .name     = "getitimer",
 979           .arg = { [0] = STRARRAY(which, itimers), }, },
 980         { .name     = "getpid",     .errpid = true, },
 981         { .name     = "getpgid",    .errpid = true, },
 982         { .name     = "getppid",    .errpid = true, },
 983         { .name     = "getrandom",
 984           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 985         { .name     = "getrlimit",
 986           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 987         { .name     = "getsockopt",
 988           .arg = { [1] = STRARRAY(level, socket_level), }, },
 989         { .name     = "gettid",     .errpid = true, },
 990         { .name     = "ioctl",
 991           .arg = {
 992 #if defined(__i386__) || defined(__x86_64__)
 993 /*
 994  * FIXME: Make this available to all arches.
 995  */
 996                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 997                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 998 #else
 999                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
1000 #endif
1001         { .name     = "kcmp",       .nr_args = 5,
1002           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
1003                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
1004                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
1005                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
1006                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1007         { .name     = "keyctl",
1008           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1009         { .name     = "kill",
1010           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1011         { .name     = "linkat",
1012           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1013         { .name     = "lseek",
1014           .arg = { [2] = STRARRAY(whence, whences), }, },
1015         { .name     = "lstat", .alias = "newlstat", },
1016         { .name     = "madvise",
1017           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1018                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1019         { .name     = "mkdirat",
1020           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1021         { .name     = "mknodat",
1022           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1023         { .name     = "mmap",       .hexret = true,
1024 /* The standard mmap maps to old_mmap on s390x */
1025 #if defined(__s390x__)
1026         .alias = "old_mmap",
1027 #endif
1028           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1029                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1030                            .strtoul   = STUL_STRARRAY_FLAGS,
1031                            .parm      = &strarray__mmap_flags, },
1032                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1033         { .name     = "mount",
1034           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1035                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1036                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1037         { .name     = "move_mount",
1038           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1039                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1040                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1041                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1042                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1043         { .name     = "mprotect",
1044           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1045                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1046         { .name     = "mq_unlink",
1047           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1048         { .name     = "mremap",     .hexret = true,
1049           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1050         { .name     = "name_to_handle_at",
1051           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1052         { .name     = "newfstatat",
1053           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1054         { .name     = "open",
1055           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1056         { .name     = "open_by_handle_at",
1057           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1058                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1059         { .name     = "openat",
1060           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1061                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1062         { .name     = "perf_event_open",
1063           .arg = { [0] = { .scnprintf = SCA_PERF_ATTR,  /* attr */ },
1064                    [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1065                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1066                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1067         { .name     = "pipe2",
1068           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1069         { .name     = "pkey_alloc",
1070           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1071         { .name     = "pkey_free",
1072           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1073         { .name     = "pkey_mprotect",
1074           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1075                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1076                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1077         { .name     = "poll", .timeout = true, },
1078         { .name     = "ppoll", .timeout = true, },
1079         { .name     = "prctl",
1080           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1081                            .strtoul   = STUL_STRARRAY,
1082                            .parm      = &strarray__prctl_options, },
1083                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1084                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1085         { .name     = "pread", .alias = "pread64", },
1086         { .name     = "preadv", .alias = "pread", },
1087         { .name     = "prlimit64",
1088           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1089         { .name     = "pwrite", .alias = "pwrite64", },
1090         { .name     = "readlinkat",
1091           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1092         { .name     = "recvfrom",
1093           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1094         { .name     = "recvmmsg",
1095           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1096         { .name     = "recvmsg",
1097           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1098         { .name     = "renameat",
1099           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1100                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1101         { .name     = "renameat2",
1102           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1103                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1104                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1105         { .name     = "rt_sigaction",
1106           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1107         { .name     = "rt_sigprocmask",
1108           .arg = { [0] = STRARRAY(how, sighow), }, },
1109         { .name     = "rt_sigqueueinfo",
1110           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1111         { .name     = "rt_tgsigqueueinfo",
1112           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1113         { .name     = "sched_setscheduler",
1114           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1115         { .name     = "seccomp",
1116           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1117                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1118         { .name     = "select", .timeout = true, },
1119         { .name     = "sendfile", .alias = "sendfile64", },
1120         { .name     = "sendmmsg",
1121           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1122         { .name     = "sendmsg",
1123           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1124         { .name     = "sendto",
1125           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1126                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1127         { .name     = "set_tid_address", .errpid = true, },
1128         { .name     = "setitimer",
1129           .arg = { [0] = STRARRAY(which, itimers), }, },
1130         { .name     = "setrlimit",
1131           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1132         { .name     = "setsockopt",
1133           .arg = { [1] = STRARRAY(level, socket_level), }, },
1134         { .name     = "socket",
1135           .arg = { [0] = STRARRAY(family, socket_families),
1136                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1137                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1138         { .name     = "socketpair",
1139           .arg = { [0] = STRARRAY(family, socket_families),
1140                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1141                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1142         { .name     = "stat", .alias = "newstat", },
1143         { .name     = "statx",
1144           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1145                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1146                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1147         { .name     = "swapoff",
1148           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1149         { .name     = "swapon",
1150           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1151         { .name     = "symlinkat",
1152           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1153         { .name     = "sync_file_range",
1154           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1155         { .name     = "tgkill",
1156           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1157         { .name     = "tkill",
1158           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1159         { .name     = "umount2", .alias = "umount",
1160           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1161         { .name     = "uname", .alias = "newuname", },
1162         { .name     = "unlinkat",
1163           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1164         { .name     = "utimensat",
1165           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1166         { .name     = "wait4",      .errpid = true,
1167           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1168         { .name     = "waitid",     .errpid = true,
1169           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1170 };
1171
1172 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1173 {
1174         const struct syscall_fmt *fmt = fmtp;
1175         return strcmp(name, fmt->name);
1176 }
1177
1178 static const struct syscall_fmt *__syscall_fmt__find(const struct syscall_fmt *fmts,
1179                                                      const int nmemb,
1180                                                      const char *name)
1181 {
1182         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1183 }
1184
1185 static const struct syscall_fmt *syscall_fmt__find(const char *name)
1186 {
1187         const int nmemb = ARRAY_SIZE(syscall_fmts);
1188         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1189 }
1190
1191 static const struct syscall_fmt *__syscall_fmt__find_by_alias(const struct syscall_fmt *fmts,
1192                                                               const int nmemb, const char *alias)
1193 {
1194         int i;
1195
1196         for (i = 0; i < nmemb; ++i) {
1197                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1198                         return &fmts[i];
1199         }
1200
1201         return NULL;
1202 }
1203
1204 static const struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1205 {
1206         const int nmemb = ARRAY_SIZE(syscall_fmts);
1207         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1208 }
1209
1210 /*
1211  * is_exit: is this "exit" or "exit_group"?
1212  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1213  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1214  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1215  */
1216 struct syscall {
1217         struct tep_event    *tp_format;
1218         int                 nr_args;
1219         int                 args_size;
1220         struct {
1221                 struct bpf_program *sys_enter,
1222                                    *sys_exit;
1223         }                   bpf_prog;
1224         bool                is_exit;
1225         bool                is_open;
1226         bool                nonexistent;
1227         struct tep_format_field *args;
1228         const char          *name;
1229         const struct syscall_fmt  *fmt;
1230         struct syscall_arg_fmt *arg_fmt;
1231 };
1232
1233 /*
1234  * We need to have this 'calculated' boolean because in some cases we really
1235  * don't know what is the duration of a syscall, for instance, when we start
1236  * a session and some threads are waiting for a syscall to finish, say 'poll',
1237  * in which case all we can do is to print "( ? ) for duration and for the
1238  * start timestamp.
1239  */
1240 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1241 {
1242         double duration = (double)t / NSEC_PER_MSEC;
1243         size_t printed = fprintf(fp, "(");
1244
1245         if (!calculated)
1246                 printed += fprintf(fp, "         ");
1247         else if (duration >= 1.0)
1248                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1249         else if (duration >= 0.01)
1250                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1251         else
1252                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1253         return printed + fprintf(fp, "): ");
1254 }
1255
1256 /**
1257  * filename.ptr: The filename char pointer that will be vfs_getname'd
1258  * filename.entry_str_pos: Where to insert the string translated from
1259  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1260  * ret_scnprintf: syscall args may set this to a different syscall return
1261  *                formatter, for instance, fcntl may return fds, file flags, etc.
1262  */
1263 struct thread_trace {
1264         u64               entry_time;
1265         bool              entry_pending;
1266         unsigned long     nr_events;
1267         unsigned long     pfmaj, pfmin;
1268         char              *entry_str;
1269         double            runtime_ms;
1270         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1271         struct {
1272                 unsigned long ptr;
1273                 short int     entry_str_pos;
1274                 bool          pending_open;
1275                 unsigned int  namelen;
1276                 char          *name;
1277         } filename;
1278         struct {
1279                 int           max;
1280                 struct file   *table;
1281         } files;
1282
1283         struct intlist *syscall_stats;
1284 };
1285
1286 static struct thread_trace *thread_trace__new(void)
1287 {
1288         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1289
1290         if (ttrace) {
1291                 ttrace->files.max = -1;
1292                 ttrace->syscall_stats = intlist__new(NULL);
1293         }
1294
1295         return ttrace;
1296 }
1297
1298 static void thread_trace__free_files(struct thread_trace *ttrace);
1299
1300 static void thread_trace__delete(void *pttrace)
1301 {
1302         struct thread_trace *ttrace = pttrace;
1303
1304         if (!ttrace)
1305                 return;
1306
1307         intlist__delete(ttrace->syscall_stats);
1308         ttrace->syscall_stats = NULL;
1309         thread_trace__free_files(ttrace);
1310         zfree(&ttrace->entry_str);
1311         free(ttrace);
1312 }
1313
1314 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1315 {
1316         struct thread_trace *ttrace;
1317
1318         if (thread == NULL)
1319                 goto fail;
1320
1321         if (thread__priv(thread) == NULL)
1322                 thread__set_priv(thread, thread_trace__new());
1323
1324         if (thread__priv(thread) == NULL)
1325                 goto fail;
1326
1327         ttrace = thread__priv(thread);
1328         ++ttrace->nr_events;
1329
1330         return ttrace;
1331 fail:
1332         color_fprintf(fp, PERF_COLOR_RED,
1333                       "WARNING: not enough memory, dropping samples!\n");
1334         return NULL;
1335 }
1336
1337
1338 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1339                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1340 {
1341         struct thread_trace *ttrace = thread__priv(arg->thread);
1342
1343         ttrace->ret_scnprintf = ret_scnprintf;
1344 }
1345
1346 #define TRACE_PFMAJ             (1 << 0)
1347 #define TRACE_PFMIN             (1 << 1)
1348
1349 static const size_t trace__entry_str_size = 2048;
1350
1351 static void thread_trace__free_files(struct thread_trace *ttrace)
1352 {
1353         for (int i = 0; i < ttrace->files.max; ++i) {
1354                 struct file *file = ttrace->files.table + i;
1355                 zfree(&file->pathname);
1356         }
1357
1358         zfree(&ttrace->files.table);
1359         ttrace->files.max  = -1;
1360 }
1361
1362 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1363 {
1364         if (fd < 0)
1365                 return NULL;
1366
1367         if (fd > ttrace->files.max) {
1368                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1369
1370                 if (nfiles == NULL)
1371                         return NULL;
1372
1373                 if (ttrace->files.max != -1) {
1374                         memset(nfiles + ttrace->files.max + 1, 0,
1375                                (fd - ttrace->files.max) * sizeof(struct file));
1376                 } else {
1377                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1378                 }
1379
1380                 ttrace->files.table = nfiles;
1381                 ttrace->files.max   = fd;
1382         }
1383
1384         return ttrace->files.table + fd;
1385 }
1386
1387 struct file *thread__files_entry(struct thread *thread, int fd)
1388 {
1389         return thread_trace__files_entry(thread__priv(thread), fd);
1390 }
1391
1392 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1393 {
1394         struct thread_trace *ttrace = thread__priv(thread);
1395         struct file *file = thread_trace__files_entry(ttrace, fd);
1396
1397         if (file != NULL) {
1398                 struct stat st;
1399                 if (stat(pathname, &st) == 0)
1400                         file->dev_maj = major(st.st_rdev);
1401                 file->pathname = strdup(pathname);
1402                 if (file->pathname)
1403                         return 0;
1404         }
1405
1406         return -1;
1407 }
1408
1409 static int thread__read_fd_path(struct thread *thread, int fd)
1410 {
1411         char linkname[PATH_MAX], pathname[PATH_MAX];
1412         struct stat st;
1413         int ret;
1414
1415         if (thread__pid(thread) == thread__tid(thread)) {
1416                 scnprintf(linkname, sizeof(linkname),
1417                           "/proc/%d/fd/%d", thread__pid(thread), fd);
1418         } else {
1419                 scnprintf(linkname, sizeof(linkname),
1420                           "/proc/%d/task/%d/fd/%d",
1421                           thread__pid(thread), thread__tid(thread), fd);
1422         }
1423
1424         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1425                 return -1;
1426
1427         ret = readlink(linkname, pathname, sizeof(pathname));
1428
1429         if (ret < 0 || ret > st.st_size)
1430                 return -1;
1431
1432         pathname[ret] = '\0';
1433         return trace__set_fd_pathname(thread, fd, pathname);
1434 }
1435
1436 static const char *thread__fd_path(struct thread *thread, int fd,
1437                                    struct trace *trace)
1438 {
1439         struct thread_trace *ttrace = thread__priv(thread);
1440
1441         if (ttrace == NULL || trace->fd_path_disabled)
1442                 return NULL;
1443
1444         if (fd < 0)
1445                 return NULL;
1446
1447         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1448                 if (!trace->live)
1449                         return NULL;
1450                 ++trace->stats.proc_getname;
1451                 if (thread__read_fd_path(thread, fd))
1452                         return NULL;
1453         }
1454
1455         return ttrace->files.table[fd].pathname;
1456 }
1457
1458 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1459 {
1460         int fd = arg->val;
1461         size_t printed = scnprintf(bf, size, "%d", fd);
1462         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1463
1464         if (path)
1465                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1466
1467         return printed;
1468 }
1469
1470 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1471 {
1472         size_t printed = scnprintf(bf, size, "%d", fd);
1473         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1474
1475         if (thread) {
1476                 const char *path = thread__fd_path(thread, fd, trace);
1477
1478                 if (path)
1479                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1480
1481                 thread__put(thread);
1482         }
1483
1484         return printed;
1485 }
1486
1487 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1488                                               struct syscall_arg *arg)
1489 {
1490         int fd = arg->val;
1491         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1492         struct thread_trace *ttrace = thread__priv(arg->thread);
1493
1494         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1495                 zfree(&ttrace->files.table[fd].pathname);
1496
1497         return printed;
1498 }
1499
1500 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1501                                      unsigned long ptr)
1502 {
1503         struct thread_trace *ttrace = thread__priv(thread);
1504
1505         ttrace->filename.ptr = ptr;
1506         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1507 }
1508
1509 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1510 {
1511         struct augmented_arg *augmented_arg = arg->augmented.args;
1512         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1513         /*
1514          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1515          * we would have two strings, each prefixed by its size.
1516          */
1517         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1518
1519         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1520         arg->augmented.size -= consumed;
1521
1522         return printed;
1523 }
1524
1525 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1526                                               struct syscall_arg *arg)
1527 {
1528         unsigned long ptr = arg->val;
1529
1530         if (arg->augmented.args)
1531                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1532
1533         if (!arg->trace->vfs_getname)
1534                 return scnprintf(bf, size, "%#x", ptr);
1535
1536         thread__set_filename_pos(arg->thread, bf, ptr);
1537         return 0;
1538 }
1539
1540 static bool trace__filter_duration(struct trace *trace, double t)
1541 {
1542         return t < (trace->duration_filter * NSEC_PER_MSEC);
1543 }
1544
1545 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1546 {
1547         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1548
1549         return fprintf(fp, "%10.3f ", ts);
1550 }
1551
1552 /*
1553  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1554  * using ttrace->entry_time for a thread that receives a sys_exit without
1555  * first having received a sys_enter ("poll" issued before tracing session
1556  * starts, lost sys_enter exit due to ring buffer overflow).
1557  */
1558 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1559 {
1560         if (tstamp > 0)
1561                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1562
1563         return fprintf(fp, "         ? ");
1564 }
1565
1566 static pid_t workload_pid = -1;
1567 static volatile sig_atomic_t done = false;
1568 static volatile sig_atomic_t interrupted = false;
1569
1570 static void sighandler_interrupt(int sig __maybe_unused)
1571 {
1572         done = interrupted = true;
1573 }
1574
1575 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1576                             void *context __maybe_unused)
1577 {
1578         if (info->si_pid == workload_pid)
1579                 done = true;
1580 }
1581
1582 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1583 {
1584         size_t printed = 0;
1585
1586         if (trace->multiple_threads) {
1587                 if (trace->show_comm)
1588                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1589                 printed += fprintf(fp, "%d ", thread__tid(thread));
1590         }
1591
1592         return printed;
1593 }
1594
1595 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1596                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1597 {
1598         size_t printed = 0;
1599
1600         if (trace->show_tstamp)
1601                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1602         if (trace->show_duration)
1603                 printed += fprintf_duration(duration, duration_calculated, fp);
1604         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1605 }
1606
1607 static int trace__process_event(struct trace *trace, struct machine *machine,
1608                                 union perf_event *event, struct perf_sample *sample)
1609 {
1610         int ret = 0;
1611
1612         switch (event->header.type) {
1613         case PERF_RECORD_LOST:
1614                 color_fprintf(trace->output, PERF_COLOR_RED,
1615                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1616                 ret = machine__process_lost_event(machine, event, sample);
1617                 break;
1618         default:
1619                 ret = machine__process_event(machine, event, sample);
1620                 break;
1621         }
1622
1623         return ret;
1624 }
1625
1626 static int trace__tool_process(struct perf_tool *tool,
1627                                union perf_event *event,
1628                                struct perf_sample *sample,
1629                                struct machine *machine)
1630 {
1631         struct trace *trace = container_of(tool, struct trace, tool);
1632         return trace__process_event(trace, machine, event, sample);
1633 }
1634
1635 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1636 {
1637         struct machine *machine = vmachine;
1638
1639         if (machine->kptr_restrict_warned)
1640                 return NULL;
1641
1642         if (symbol_conf.kptr_restrict) {
1643                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1644                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1645                            "Kernel samples will not be resolved.\n");
1646                 machine->kptr_restrict_warned = true;
1647                 return NULL;
1648         }
1649
1650         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1651 }
1652
1653 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1654 {
1655         int err = symbol__init(NULL);
1656
1657         if (err)
1658                 return err;
1659
1660         trace->host = machine__new_host();
1661         if (trace->host == NULL)
1662                 return -ENOMEM;
1663
1664         thread__set_priv_destructor(thread_trace__delete);
1665
1666         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1667         if (err < 0)
1668                 goto out;
1669
1670         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1671                                             evlist->core.threads, trace__tool_process,
1672                                             true, false, 1);
1673 out:
1674         if (err)
1675                 symbol__exit();
1676
1677         return err;
1678 }
1679
1680 static void trace__symbols__exit(struct trace *trace)
1681 {
1682         machine__exit(trace->host);
1683         trace->host = NULL;
1684
1685         symbol__exit();
1686 }
1687
1688 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1689 {
1690         int idx;
1691
1692         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1693                 nr_args = sc->fmt->nr_args;
1694
1695         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1696         if (sc->arg_fmt == NULL)
1697                 return -1;
1698
1699         for (idx = 0; idx < nr_args; ++idx) {
1700                 if (sc->fmt)
1701                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1702         }
1703
1704         sc->nr_args = nr_args;
1705         return 0;
1706 }
1707
1708 static const struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1709         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1710         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1711 };
1712
1713 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1714 {
1715        const struct syscall_arg_fmt *fmt = fmtp;
1716        return strcmp(name, fmt->name);
1717 }
1718
1719 static const struct syscall_arg_fmt *
1720 __syscall_arg_fmt__find_by_name(const struct syscall_arg_fmt *fmts, const int nmemb,
1721                                 const char *name)
1722 {
1723        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1724 }
1725
1726 static const struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1727 {
1728        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1729        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1730 }
1731
1732 static struct tep_format_field *
1733 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1734 {
1735         struct tep_format_field *last_field = NULL;
1736         int len;
1737
1738         for (; field; field = field->next, ++arg) {
1739                 last_field = field;
1740
1741                 if (arg->scnprintf)
1742                         continue;
1743
1744                 len = strlen(field->name);
1745
1746                 if (strcmp(field->type, "const char *") == 0 &&
1747                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1748                      strstr(field->name, "path") != NULL))
1749                         arg->scnprintf = SCA_FILENAME;
1750                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1751                         arg->scnprintf = SCA_PTR;
1752                 else if (strcmp(field->type, "pid_t") == 0)
1753                         arg->scnprintf = SCA_PID;
1754                 else if (strcmp(field->type, "umode_t") == 0)
1755                         arg->scnprintf = SCA_MODE_T;
1756                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1757                         arg->scnprintf = SCA_CHAR_ARRAY;
1758                         arg->nr_entries = field->arraylen;
1759                 } else if ((strcmp(field->type, "int") == 0 ||
1760                           strcmp(field->type, "unsigned int") == 0 ||
1761                           strcmp(field->type, "long") == 0) &&
1762                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1763                         /*
1764                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1765                          * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1766                          * 65 int
1767                          * 23 unsigned int
1768                          * 7 unsigned long
1769                          */
1770                         arg->scnprintf = SCA_FD;
1771                 } else {
1772                         const struct syscall_arg_fmt *fmt =
1773                                 syscall_arg_fmt__find_by_name(field->name);
1774
1775                         if (fmt) {
1776                                 arg->scnprintf = fmt->scnprintf;
1777                                 arg->strtoul   = fmt->strtoul;
1778                         }
1779                 }
1780         }
1781
1782         return last_field;
1783 }
1784
1785 static int syscall__set_arg_fmts(struct syscall *sc)
1786 {
1787         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1788
1789         if (last_field)
1790                 sc->args_size = last_field->offset + last_field->size;
1791
1792         return 0;
1793 }
1794
1795 static int trace__read_syscall_info(struct trace *trace, int id)
1796 {
1797         char tp_name[128];
1798         struct syscall *sc;
1799         const char *name = syscalltbl__name(trace->sctbl, id);
1800
1801 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1802         if (trace->syscalls.table == NULL) {
1803                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1804                 if (trace->syscalls.table == NULL)
1805                         return -ENOMEM;
1806         }
1807 #else
1808         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1809                 // When using libaudit we don't know beforehand what is the max syscall id
1810                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1811
1812                 if (table == NULL)
1813                         return -ENOMEM;
1814
1815                 // Need to memset from offset 0 and +1 members if brand new
1816                 if (trace->syscalls.table == NULL)
1817                         memset(table, 0, (id + 1) * sizeof(*sc));
1818                 else
1819                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1820
1821                 trace->syscalls.table         = table;
1822                 trace->sctbl->syscalls.max_id = id;
1823         }
1824 #endif
1825         sc = trace->syscalls.table + id;
1826         if (sc->nonexistent)
1827                 return -EEXIST;
1828
1829         if (name == NULL) {
1830                 sc->nonexistent = true;
1831                 return -EEXIST;
1832         }
1833
1834         sc->name = name;
1835         sc->fmt  = syscall_fmt__find(sc->name);
1836
1837         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1838         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1839
1840         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1841                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1842                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1843         }
1844
1845         /*
1846          * Fails to read trace point format via sysfs node, so the trace point
1847          * doesn't exist.  Set the 'nonexistent' flag as true.
1848          */
1849         if (IS_ERR(sc->tp_format)) {
1850                 sc->nonexistent = true;
1851                 return PTR_ERR(sc->tp_format);
1852         }
1853
1854         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1855                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1856                 return -ENOMEM;
1857
1858         sc->args = sc->tp_format->format.fields;
1859         /*
1860          * We need to check and discard the first variable '__syscall_nr'
1861          * or 'nr' that mean the syscall number. It is needless here.
1862          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1863          */
1864         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1865                 sc->args = sc->args->next;
1866                 --sc->nr_args;
1867         }
1868
1869         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1870         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1871
1872         return syscall__set_arg_fmts(sc);
1873 }
1874
1875 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1876 {
1877         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1878
1879         if (fmt != NULL) {
1880                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1881                 return 0;
1882         }
1883
1884         return -ENOMEM;
1885 }
1886
1887 static int intcmp(const void *a, const void *b)
1888 {
1889         const int *one = a, *another = b;
1890
1891         return *one - *another;
1892 }
1893
1894 static int trace__validate_ev_qualifier(struct trace *trace)
1895 {
1896         int err = 0;
1897         bool printed_invalid_prefix = false;
1898         struct str_node *pos;
1899         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1900
1901         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1902                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1903
1904         if (trace->ev_qualifier_ids.entries == NULL) {
1905                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1906                        trace->output);
1907                 err = -EINVAL;
1908                 goto out;
1909         }
1910
1911         strlist__for_each_entry(pos, trace->ev_qualifier) {
1912                 const char *sc = pos->s;
1913                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1914
1915                 if (id < 0) {
1916                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1917                         if (id >= 0)
1918                                 goto matches;
1919
1920                         if (!printed_invalid_prefix) {
1921                                 pr_debug("Skipping unknown syscalls: ");
1922                                 printed_invalid_prefix = true;
1923                         } else {
1924                                 pr_debug(", ");
1925                         }
1926
1927                         pr_debug("%s", sc);
1928                         continue;
1929                 }
1930 matches:
1931                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1932                 if (match_next == -1)
1933                         continue;
1934
1935                 while (1) {
1936                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1937                         if (id < 0)
1938                                 break;
1939                         if (nr_allocated == nr_used) {
1940                                 void *entries;
1941
1942                                 nr_allocated += 8;
1943                                 entries = realloc(trace->ev_qualifier_ids.entries,
1944                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1945                                 if (entries == NULL) {
1946                                         err = -ENOMEM;
1947                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1948                                         goto out_free;
1949                                 }
1950                                 trace->ev_qualifier_ids.entries = entries;
1951                         }
1952                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1953                 }
1954         }
1955
1956         trace->ev_qualifier_ids.nr = nr_used;
1957         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1958 out:
1959         if (printed_invalid_prefix)
1960                 pr_debug("\n");
1961         return err;
1962 out_free:
1963         zfree(&trace->ev_qualifier_ids.entries);
1964         trace->ev_qualifier_ids.nr = 0;
1965         goto out;
1966 }
1967
1968 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1969 {
1970         bool in_ev_qualifier;
1971
1972         if (trace->ev_qualifier_ids.nr == 0)
1973                 return true;
1974
1975         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1976                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1977
1978         if (in_ev_qualifier)
1979                return !trace->not_ev_qualifier;
1980
1981         return trace->not_ev_qualifier;
1982 }
1983
1984 /*
1985  * args is to be interpreted as a series of longs but we need to handle
1986  * 8-byte unaligned accesses. args points to raw_data within the event
1987  * and raw_data is guaranteed to be 8-byte unaligned because it is
1988  * preceded by raw_size which is a u32. So we need to copy args to a temp
1989  * variable to read it. Most notably this avoids extended load instructions
1990  * on unaligned addresses
1991  */
1992 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1993 {
1994         unsigned long val;
1995         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1996
1997         memcpy(&val, p, sizeof(val));
1998         return val;
1999 }
2000
2001 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
2002                                       struct syscall_arg *arg)
2003 {
2004         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
2005                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
2006
2007         return scnprintf(bf, size, "arg%d: ", arg->idx);
2008 }
2009
2010 /*
2011  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
2012  * as mount 'flags' argument that needs ignoring some magic flag, see comment
2013  * in tools/perf/trace/beauty/mount_flags.c
2014  */
2015 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
2016 {
2017         if (fmt && fmt->mask_val)
2018                 return fmt->mask_val(arg, val);
2019
2020         return val;
2021 }
2022
2023 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
2024                                              struct syscall_arg *arg, unsigned long val)
2025 {
2026         if (fmt && fmt->scnprintf) {
2027                 arg->val = val;
2028                 if (fmt->parm)
2029                         arg->parm = fmt->parm;
2030                 return fmt->scnprintf(bf, size, arg);
2031         }
2032         return scnprintf(bf, size, "%ld", val);
2033 }
2034
2035 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2036                                       unsigned char *args, void *augmented_args, int augmented_args_size,
2037                                       struct trace *trace, struct thread *thread)
2038 {
2039         size_t printed = 0;
2040         unsigned long val;
2041         u8 bit = 1;
2042         struct syscall_arg arg = {
2043                 .args   = args,
2044                 .augmented = {
2045                         .size = augmented_args_size,
2046                         .args = augmented_args,
2047                 },
2048                 .idx    = 0,
2049                 .mask   = 0,
2050                 .trace  = trace,
2051                 .thread = thread,
2052                 .show_string_prefix = trace->show_string_prefix,
2053         };
2054         struct thread_trace *ttrace = thread__priv(thread);
2055
2056         /*
2057          * Things like fcntl will set this in its 'cmd' formatter to pick the
2058          * right formatter for the return value (an fd? file flags?), which is
2059          * not needed for syscalls that always return a given type, say an fd.
2060          */
2061         ttrace->ret_scnprintf = NULL;
2062
2063         if (sc->args != NULL) {
2064                 struct tep_format_field *field;
2065
2066                 for (field = sc->args; field;
2067                      field = field->next, ++arg.idx, bit <<= 1) {
2068                         if (arg.mask & bit)
2069                                 continue;
2070
2071                         arg.fmt = &sc->arg_fmt[arg.idx];
2072                         val = syscall_arg__val(&arg, arg.idx);
2073                         /*
2074                          * Some syscall args need some mask, most don't and
2075                          * return val untouched.
2076                          */
2077                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2078
2079                         /*
2080                          * Suppress this argument if its value is zero and
2081                          * and we don't have a string associated in an
2082                          * strarray for it.
2083                          */
2084                         if (val == 0 &&
2085                             !trace->show_zeros &&
2086                             !(sc->arg_fmt &&
2087                               (sc->arg_fmt[arg.idx].show_zero ||
2088                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2089                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2090                               sc->arg_fmt[arg.idx].parm))
2091                                 continue;
2092
2093                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2094
2095                         if (trace->show_arg_names)
2096                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2097
2098                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2099                                                                   bf + printed, size - printed, &arg, val);
2100                 }
2101         } else if (IS_ERR(sc->tp_format)) {
2102                 /*
2103                  * If we managed to read the tracepoint /format file, then we
2104                  * may end up not having any args, like with gettid(), so only
2105                  * print the raw args when we didn't manage to read it.
2106                  */
2107                 while (arg.idx < sc->nr_args) {
2108                         if (arg.mask & bit)
2109                                 goto next_arg;
2110                         val = syscall_arg__val(&arg, arg.idx);
2111                         if (printed)
2112                                 printed += scnprintf(bf + printed, size - printed, ", ");
2113                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2114                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2115 next_arg:
2116                         ++arg.idx;
2117                         bit <<= 1;
2118                 }
2119         }
2120
2121         return printed;
2122 }
2123
2124 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2125                                   union perf_event *event,
2126                                   struct perf_sample *sample);
2127
2128 static struct syscall *trace__syscall_info(struct trace *trace,
2129                                            struct evsel *evsel, int id)
2130 {
2131         int err = 0;
2132
2133         if (id < 0) {
2134
2135                 /*
2136                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2137                  * before that, leaving at a higher verbosity level till that is
2138                  * explained. Reproduced with plain ftrace with:
2139                  *
2140                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2141                  * grep "NR -1 " /t/trace_pipe
2142                  *
2143                  * After generating some load on the machine.
2144                  */
2145                 if (verbose > 1) {
2146                         static u64 n;
2147                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2148                                 id, evsel__name(evsel), ++n);
2149                 }
2150                 return NULL;
2151         }
2152
2153         err = -EINVAL;
2154
2155 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2156         if (id > trace->sctbl->syscalls.max_id) {
2157 #else
2158         if (id >= trace->sctbl->syscalls.max_id) {
2159                 /*
2160                  * With libaudit we don't know beforehand what is the max_id,
2161                  * so we let trace__read_syscall_info() figure that out as we
2162                  * go on reading syscalls.
2163                  */
2164                 err = trace__read_syscall_info(trace, id);
2165                 if (err)
2166 #endif
2167                 goto out_cant_read;
2168         }
2169
2170         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2171             (err = trace__read_syscall_info(trace, id)) != 0)
2172                 goto out_cant_read;
2173
2174         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2175                 goto out_cant_read;
2176
2177         return &trace->syscalls.table[id];
2178
2179 out_cant_read:
2180         if (verbose > 0) {
2181                 char sbuf[STRERR_BUFSIZE];
2182                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2183                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2184                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2185                 fputs(" information\n", trace->output);
2186         }
2187         return NULL;
2188 }
2189
2190 struct syscall_stats {
2191         struct stats stats;
2192         u64          nr_failures;
2193         int          max_errno;
2194         u32          *errnos;
2195 };
2196
2197 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2198                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2199 {
2200         struct int_node *inode;
2201         struct syscall_stats *stats;
2202         u64 duration = 0;
2203
2204         inode = intlist__findnew(ttrace->syscall_stats, id);
2205         if (inode == NULL)
2206                 return;
2207
2208         stats = inode->priv;
2209         if (stats == NULL) {
2210                 stats = zalloc(sizeof(*stats));
2211                 if (stats == NULL)
2212                         return;
2213
2214                 init_stats(&stats->stats);
2215                 inode->priv = stats;
2216         }
2217
2218         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2219                 duration = sample->time - ttrace->entry_time;
2220
2221         update_stats(&stats->stats, duration);
2222
2223         if (err < 0) {
2224                 ++stats->nr_failures;
2225
2226                 if (!errno_summary)
2227                         return;
2228
2229                 err = -err;
2230                 if (err > stats->max_errno) {
2231                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2232
2233                         if (new_errnos) {
2234                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2235                         } else {
2236                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2237                                          thread__comm_str(thread), thread__pid(thread),
2238                                          thread__tid(thread));
2239                                 return;
2240                         }
2241
2242                         stats->errnos = new_errnos;
2243                         stats->max_errno = err;
2244                 }
2245
2246                 ++stats->errnos[err - 1];
2247         }
2248 }
2249
2250 static int trace__printf_interrupted_entry(struct trace *trace)
2251 {
2252         struct thread_trace *ttrace;
2253         size_t printed;
2254         int len;
2255
2256         if (trace->failure_only || trace->current == NULL)
2257                 return 0;
2258
2259         ttrace = thread__priv(trace->current);
2260
2261         if (!ttrace->entry_pending)
2262                 return 0;
2263
2264         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2265         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2266
2267         if (len < trace->args_alignment - 4)
2268                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2269
2270         printed += fprintf(trace->output, " ...\n");
2271
2272         ttrace->entry_pending = false;
2273         ++trace->nr_events_printed;
2274
2275         return printed;
2276 }
2277
2278 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2279                                  struct perf_sample *sample, struct thread *thread)
2280 {
2281         int printed = 0;
2282
2283         if (trace->print_sample) {
2284                 double ts = (double)sample->time / NSEC_PER_MSEC;
2285
2286                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2287                                    evsel__name(evsel), ts,
2288                                    thread__comm_str(thread),
2289                                    sample->pid, sample->tid, sample->cpu);
2290         }
2291
2292         return printed;
2293 }
2294
2295 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2296 {
2297         void *augmented_args = NULL;
2298         /*
2299          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2300          * and there we get all 6 syscall args plus the tracepoint common fields
2301          * that gets calculated at the start and the syscall_nr (another long).
2302          * So we check if that is the case and if so don't look after the
2303          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2304          * which is fixed.
2305          *
2306          * We'll revisit this later to pass s->args_size to the BPF augmenter
2307          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2308          * copies only what we need for each syscall, like what happens when we
2309          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2310          * traffic to just what is needed for each syscall.
2311          */
2312         int args_size = raw_augmented_args_size ?: sc->args_size;
2313
2314         *augmented_args_size = sample->raw_size - args_size;
2315         if (*augmented_args_size > 0)
2316                 augmented_args = sample->raw_data + args_size;
2317
2318         return augmented_args;
2319 }
2320
2321 static void syscall__exit(struct syscall *sc)
2322 {
2323         if (!sc)
2324                 return;
2325
2326         zfree(&sc->arg_fmt);
2327 }
2328
2329 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2330                             union perf_event *event __maybe_unused,
2331                             struct perf_sample *sample)
2332 {
2333         char *msg;
2334         void *args;
2335         int printed = 0;
2336         struct thread *thread;
2337         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2338         int augmented_args_size = 0;
2339         void *augmented_args = NULL;
2340         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2341         struct thread_trace *ttrace;
2342
2343         if (sc == NULL)
2344                 return -1;
2345
2346         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2347         ttrace = thread__trace(thread, trace->output);
2348         if (ttrace == NULL)
2349                 goto out_put;
2350
2351         trace__fprintf_sample(trace, evsel, sample, thread);
2352
2353         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2354
2355         if (ttrace->entry_str == NULL) {
2356                 ttrace->entry_str = malloc(trace__entry_str_size);
2357                 if (!ttrace->entry_str)
2358                         goto out_put;
2359         }
2360
2361         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2362                 trace__printf_interrupted_entry(trace);
2363         /*
2364          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2365          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2366          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2367          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2368          * so when handling, say the openat syscall, we end up getting 6 args for the
2369          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2370          * thinking that the extra 2 u64 args are the augmented filename, so just check
2371          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2372          */
2373         if (evsel != trace->syscalls.events.sys_enter)
2374                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2375         ttrace->entry_time = sample->time;
2376         msg = ttrace->entry_str;
2377         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2378
2379         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2380                                            args, augmented_args, augmented_args_size, trace, thread);
2381
2382         if (sc->is_exit) {
2383                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2384                         int alignment = 0;
2385
2386                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2387                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2388                         if (trace->args_alignment > printed)
2389                                 alignment = trace->args_alignment - printed;
2390                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2391                 }
2392         } else {
2393                 ttrace->entry_pending = true;
2394                 /* See trace__vfs_getname & trace__sys_exit */
2395                 ttrace->filename.pending_open = false;
2396         }
2397
2398         if (trace->current != thread) {
2399                 thread__put(trace->current);
2400                 trace->current = thread__get(thread);
2401         }
2402         err = 0;
2403 out_put:
2404         thread__put(thread);
2405         return err;
2406 }
2407
2408 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2409                                     struct perf_sample *sample)
2410 {
2411         struct thread_trace *ttrace;
2412         struct thread *thread;
2413         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2414         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2415         char msg[1024];
2416         void *args, *augmented_args = NULL;
2417         int augmented_args_size;
2418
2419         if (sc == NULL)
2420                 return -1;
2421
2422         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2423         ttrace = thread__trace(thread, trace->output);
2424         /*
2425          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2426          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2427          */
2428         if (ttrace == NULL)
2429                 goto out_put;
2430
2431         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2432         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2433         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2434         fprintf(trace->output, "%s", msg);
2435         err = 0;
2436 out_put:
2437         thread__put(thread);
2438         return err;
2439 }
2440
2441 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2442                                     struct perf_sample *sample,
2443                                     struct callchain_cursor *cursor)
2444 {
2445         struct addr_location al;
2446         int max_stack = evsel->core.attr.sample_max_stack ?
2447                         evsel->core.attr.sample_max_stack :
2448                         trace->max_stack;
2449         int err = -1;
2450
2451         addr_location__init(&al);
2452         if (machine__resolve(trace->host, &al, sample) < 0)
2453                 goto out;
2454
2455         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2456 out:
2457         addr_location__exit(&al);
2458         return err;
2459 }
2460
2461 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2462 {
2463         /* TODO: user-configurable print_opts */
2464         const unsigned int print_opts = EVSEL__PRINT_SYM |
2465                                         EVSEL__PRINT_DSO |
2466                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2467
2468         return sample__fprintf_callchain(sample, 38, print_opts, get_tls_callchain_cursor(), symbol_conf.bt_stop_list, trace->output);
2469 }
2470
2471 static const char *errno_to_name(struct evsel *evsel, int err)
2472 {
2473         struct perf_env *env = evsel__env(evsel);
2474
2475         return perf_env__arch_strerrno(env, err);
2476 }
2477
2478 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2479                            union perf_event *event __maybe_unused,
2480                            struct perf_sample *sample)
2481 {
2482         long ret;
2483         u64 duration = 0;
2484         bool duration_calculated = false;
2485         struct thread *thread;
2486         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2487         int alignment = trace->args_alignment;
2488         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2489         struct thread_trace *ttrace;
2490
2491         if (sc == NULL)
2492                 return -1;
2493
2494         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2495         ttrace = thread__trace(thread, trace->output);
2496         if (ttrace == NULL)
2497                 goto out_put;
2498
2499         trace__fprintf_sample(trace, evsel, sample, thread);
2500
2501         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2502
2503         if (trace->summary)
2504                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2505
2506         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2507                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2508                 ttrace->filename.pending_open = false;
2509                 ++trace->stats.vfs_getname;
2510         }
2511
2512         if (ttrace->entry_time) {
2513                 duration = sample->time - ttrace->entry_time;
2514                 if (trace__filter_duration(trace, duration))
2515                         goto out;
2516                 duration_calculated = true;
2517         } else if (trace->duration_filter)
2518                 goto out;
2519
2520         if (sample->callchain) {
2521                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2522
2523                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2524                 if (callchain_ret == 0) {
2525                         if (cursor->nr < trace->min_stack)
2526                                 goto out;
2527                         callchain_ret = 1;
2528                 }
2529         }
2530
2531         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2532                 goto out;
2533
2534         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2535
2536         if (ttrace->entry_pending) {
2537                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2538         } else {
2539                 printed += fprintf(trace->output, " ... [");
2540                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2541                 printed += 9;
2542                 printed += fprintf(trace->output, "]: %s()", sc->name);
2543         }
2544
2545         printed++; /* the closing ')' */
2546
2547         if (alignment > printed)
2548                 alignment -= printed;
2549         else
2550                 alignment = 0;
2551
2552         fprintf(trace->output, ")%*s= ", alignment, " ");
2553
2554         if (sc->fmt == NULL) {
2555                 if (ret < 0)
2556                         goto errno_print;
2557 signed_print:
2558                 fprintf(trace->output, "%ld", ret);
2559         } else if (ret < 0) {
2560 errno_print: {
2561                 char bf[STRERR_BUFSIZE];
2562                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2563                            *e = errno_to_name(evsel, -ret);
2564
2565                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2566         }
2567         } else if (ret == 0 && sc->fmt->timeout)
2568                 fprintf(trace->output, "0 (Timeout)");
2569         else if (ttrace->ret_scnprintf) {
2570                 char bf[1024];
2571                 struct syscall_arg arg = {
2572                         .val    = ret,
2573                         .thread = thread,
2574                         .trace  = trace,
2575                 };
2576                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2577                 ttrace->ret_scnprintf = NULL;
2578                 fprintf(trace->output, "%s", bf);
2579         } else if (sc->fmt->hexret)
2580                 fprintf(trace->output, "%#lx", ret);
2581         else if (sc->fmt->errpid) {
2582                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2583
2584                 if (child != NULL) {
2585                         fprintf(trace->output, "%ld", ret);
2586                         if (thread__comm_set(child))
2587                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2588                         thread__put(child);
2589                 }
2590         } else
2591                 goto signed_print;
2592
2593         fputc('\n', trace->output);
2594
2595         /*
2596          * We only consider an 'event' for the sake of --max-events a non-filtered
2597          * sys_enter + sys_exit and other tracepoint events.
2598          */
2599         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2600                 interrupted = true;
2601
2602         if (callchain_ret > 0)
2603                 trace__fprintf_callchain(trace, sample);
2604         else if (callchain_ret < 0)
2605                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2606 out:
2607         ttrace->entry_pending = false;
2608         err = 0;
2609 out_put:
2610         thread__put(thread);
2611         return err;
2612 }
2613
2614 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2615                               union perf_event *event __maybe_unused,
2616                               struct perf_sample *sample)
2617 {
2618         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2619         struct thread_trace *ttrace;
2620         size_t filename_len, entry_str_len, to_move;
2621         ssize_t remaining_space;
2622         char *pos;
2623         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2624
2625         if (!thread)
2626                 goto out;
2627
2628         ttrace = thread__priv(thread);
2629         if (!ttrace)
2630                 goto out_put;
2631
2632         filename_len = strlen(filename);
2633         if (filename_len == 0)
2634                 goto out_put;
2635
2636         if (ttrace->filename.namelen < filename_len) {
2637                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2638
2639                 if (f == NULL)
2640                         goto out_put;
2641
2642                 ttrace->filename.namelen = filename_len;
2643                 ttrace->filename.name = f;
2644         }
2645
2646         strcpy(ttrace->filename.name, filename);
2647         ttrace->filename.pending_open = true;
2648
2649         if (!ttrace->filename.ptr)
2650                 goto out_put;
2651
2652         entry_str_len = strlen(ttrace->entry_str);
2653         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2654         if (remaining_space <= 0)
2655                 goto out_put;
2656
2657         if (filename_len > (size_t)remaining_space) {
2658                 filename += filename_len - remaining_space;
2659                 filename_len = remaining_space;
2660         }
2661
2662         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2663         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2664         memmove(pos + filename_len, pos, to_move);
2665         memcpy(pos, filename, filename_len);
2666
2667         ttrace->filename.ptr = 0;
2668         ttrace->filename.entry_str_pos = 0;
2669 out_put:
2670         thread__put(thread);
2671 out:
2672         return 0;
2673 }
2674
2675 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2676                                      union perf_event *event __maybe_unused,
2677                                      struct perf_sample *sample)
2678 {
2679         u64 runtime = evsel__intval(evsel, sample, "runtime");
2680         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2681         struct thread *thread = machine__findnew_thread(trace->host,
2682                                                         sample->pid,
2683                                                         sample->tid);
2684         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2685
2686         if (ttrace == NULL)
2687                 goto out_dump;
2688
2689         ttrace->runtime_ms += runtime_ms;
2690         trace->runtime_ms += runtime_ms;
2691 out_put:
2692         thread__put(thread);
2693         return 0;
2694
2695 out_dump:
2696         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2697                evsel->name,
2698                evsel__strval(evsel, sample, "comm"),
2699                (pid_t)evsel__intval(evsel, sample, "pid"),
2700                runtime,
2701                evsel__intval(evsel, sample, "vruntime"));
2702         goto out_put;
2703 }
2704
2705 static int bpf_output__printer(enum binary_printer_ops op,
2706                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2707 {
2708         unsigned char ch = (unsigned char)val;
2709
2710         switch (op) {
2711         case BINARY_PRINT_CHAR_DATA:
2712                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2713         case BINARY_PRINT_DATA_BEGIN:
2714         case BINARY_PRINT_LINE_BEGIN:
2715         case BINARY_PRINT_ADDR:
2716         case BINARY_PRINT_NUM_DATA:
2717         case BINARY_PRINT_NUM_PAD:
2718         case BINARY_PRINT_SEP:
2719         case BINARY_PRINT_CHAR_PAD:
2720         case BINARY_PRINT_LINE_END:
2721         case BINARY_PRINT_DATA_END:
2722         default:
2723                 break;
2724         }
2725
2726         return 0;
2727 }
2728
2729 static void bpf_output__fprintf(struct trace *trace,
2730                                 struct perf_sample *sample)
2731 {
2732         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2733                         bpf_output__printer, NULL, trace->output);
2734         ++trace->nr_events_printed;
2735 }
2736
2737 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2738                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2739 {
2740         char bf[2048];
2741         size_t size = sizeof(bf);
2742         struct tep_format_field *field = evsel->tp_format->format.fields;
2743         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2744         size_t printed = 0;
2745         unsigned long val;
2746         u8 bit = 1;
2747         struct syscall_arg syscall_arg = {
2748                 .augmented = {
2749                         .size = augmented_args_size,
2750                         .args = augmented_args,
2751                 },
2752                 .idx    = 0,
2753                 .mask   = 0,
2754                 .trace  = trace,
2755                 .thread = thread,
2756                 .show_string_prefix = trace->show_string_prefix,
2757         };
2758
2759         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2760                 if (syscall_arg.mask & bit)
2761                         continue;
2762
2763                 syscall_arg.len = 0;
2764                 syscall_arg.fmt = arg;
2765                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2766                         int offset = field->offset;
2767
2768                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2769                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2770                                 syscall_arg.len = offset >> 16;
2771                                 offset &= 0xffff;
2772                                 if (tep_field_is_relative(field->flags))
2773                                         offset += field->offset + field->size;
2774                         }
2775
2776                         val = (uintptr_t)(sample->raw_data + offset);
2777                 } else
2778                         val = format_field__intval(field, sample, evsel->needs_swap);
2779                 /*
2780                  * Some syscall args need some mask, most don't and
2781                  * return val untouched.
2782                  */
2783                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2784
2785                 /*
2786                  * Suppress this argument if its value is zero and
2787                  * we don't have a string associated in an
2788                  * strarray for it.
2789                  */
2790                 if (val == 0 &&
2791                     !trace->show_zeros &&
2792                     !((arg->show_zero ||
2793                        arg->scnprintf == SCA_STRARRAY ||
2794                        arg->scnprintf == SCA_STRARRAYS) &&
2795                       arg->parm))
2796                         continue;
2797
2798                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2799
2800                 if (trace->show_arg_names)
2801                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2802
2803                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2804         }
2805
2806         return printed + fprintf(trace->output, "%s", bf);
2807 }
2808
2809 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2810                                 union perf_event *event __maybe_unused,
2811                                 struct perf_sample *sample)
2812 {
2813         struct thread *thread;
2814         int callchain_ret = 0;
2815         /*
2816          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2817          * this event's max_events having been hit and this is an entry coming
2818          * from the ring buffer that we should discard, since the max events
2819          * have already been considered/printed.
2820          */
2821         if (evsel->disabled)
2822                 return 0;
2823
2824         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2825
2826         if (sample->callchain) {
2827                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2828
2829                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2830                 if (callchain_ret == 0) {
2831                         if (cursor->nr < trace->min_stack)
2832                                 goto out;
2833                         callchain_ret = 1;
2834                 }
2835         }
2836
2837         trace__printf_interrupted_entry(trace);
2838         trace__fprintf_tstamp(trace, sample->time, trace->output);
2839
2840         if (trace->trace_syscalls && trace->show_duration)
2841                 fprintf(trace->output, "(         ): ");
2842
2843         if (thread)
2844                 trace__fprintf_comm_tid(trace, thread, trace->output);
2845
2846         if (evsel == trace->syscalls.events.bpf_output) {
2847                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2848                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2849
2850                 if (sc) {
2851                         fprintf(trace->output, "%s(", sc->name);
2852                         trace__fprintf_sys_enter(trace, evsel, sample);
2853                         fputc(')', trace->output);
2854                         goto newline;
2855                 }
2856
2857                 /*
2858                  * XXX: Not having the associated syscall info or not finding/adding
2859                  *      the thread should never happen, but if it does...
2860                  *      fall thru and print it as a bpf_output event.
2861                  */
2862         }
2863
2864         fprintf(trace->output, "%s(", evsel->name);
2865
2866         if (evsel__is_bpf_output(evsel)) {
2867                 bpf_output__fprintf(trace, sample);
2868         } else if (evsel->tp_format) {
2869                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2870                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2871                         if (trace->libtraceevent_print) {
2872                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2873                                                       sample->raw_data, sample->raw_size,
2874                                                       trace->output);
2875                         } else {
2876                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2877                         }
2878                 }
2879         }
2880
2881 newline:
2882         fprintf(trace->output, ")\n");
2883
2884         if (callchain_ret > 0)
2885                 trace__fprintf_callchain(trace, sample);
2886         else if (callchain_ret < 0)
2887                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2888
2889         ++trace->nr_events_printed;
2890
2891         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2892                 evsel__disable(evsel);
2893                 evsel__close(evsel);
2894         }
2895 out:
2896         thread__put(thread);
2897         return 0;
2898 }
2899
2900 static void print_location(FILE *f, struct perf_sample *sample,
2901                            struct addr_location *al,
2902                            bool print_dso, bool print_sym)
2903 {
2904
2905         if ((verbose > 0 || print_dso) && al->map)
2906                 fprintf(f, "%s@", map__dso(al->map)->long_name);
2907
2908         if ((verbose > 0 || print_sym) && al->sym)
2909                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2910                         al->addr - al->sym->start);
2911         else if (al->map)
2912                 fprintf(f, "0x%" PRIx64, al->addr);
2913         else
2914                 fprintf(f, "0x%" PRIx64, sample->addr);
2915 }
2916
2917 static int trace__pgfault(struct trace *trace,
2918                           struct evsel *evsel,
2919                           union perf_event *event __maybe_unused,
2920                           struct perf_sample *sample)
2921 {
2922         struct thread *thread;
2923         struct addr_location al;
2924         char map_type = 'd';
2925         struct thread_trace *ttrace;
2926         int err = -1;
2927         int callchain_ret = 0;
2928
2929         addr_location__init(&al);
2930         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2931
2932         if (sample->callchain) {
2933                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2934
2935                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2936                 if (callchain_ret == 0) {
2937                         if (cursor->nr < trace->min_stack)
2938                                 goto out_put;
2939                         callchain_ret = 1;
2940                 }
2941         }
2942
2943         ttrace = thread__trace(thread, trace->output);
2944         if (ttrace == NULL)
2945                 goto out_put;
2946
2947         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2948                 ttrace->pfmaj++;
2949         else
2950                 ttrace->pfmin++;
2951
2952         if (trace->summary_only)
2953                 goto out;
2954
2955         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2956
2957         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2958
2959         fprintf(trace->output, "%sfault [",
2960                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2961                 "maj" : "min");
2962
2963         print_location(trace->output, sample, &al, false, true);
2964
2965         fprintf(trace->output, "] => ");
2966
2967         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2968
2969         if (!al.map) {
2970                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2971
2972                 if (al.map)
2973                         map_type = 'x';
2974                 else
2975                         map_type = '?';
2976         }
2977
2978         print_location(trace->output, sample, &al, true, false);
2979
2980         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2981
2982         if (callchain_ret > 0)
2983                 trace__fprintf_callchain(trace, sample);
2984         else if (callchain_ret < 0)
2985                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2986
2987         ++trace->nr_events_printed;
2988 out:
2989         err = 0;
2990 out_put:
2991         thread__put(thread);
2992         addr_location__exit(&al);
2993         return err;
2994 }
2995
2996 static void trace__set_base_time(struct trace *trace,
2997                                  struct evsel *evsel,
2998                                  struct perf_sample *sample)
2999 {
3000         /*
3001          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
3002          * and don't use sample->time unconditionally, we may end up having
3003          * some other event in the future without PERF_SAMPLE_TIME for good
3004          * reason, i.e. we may not be interested in its timestamps, just in
3005          * it taking place, picking some piece of information when it
3006          * appears in our event stream (vfs_getname comes to mind).
3007          */
3008         if (trace->base_time == 0 && !trace->full_time &&
3009             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
3010                 trace->base_time = sample->time;
3011 }
3012
3013 static int trace__process_sample(struct perf_tool *tool,
3014                                  union perf_event *event,
3015                                  struct perf_sample *sample,
3016                                  struct evsel *evsel,
3017                                  struct machine *machine __maybe_unused)
3018 {
3019         struct trace *trace = container_of(tool, struct trace, tool);
3020         struct thread *thread;
3021         int err = 0;
3022
3023         tracepoint_handler handler = evsel->handler;
3024
3025         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
3026         if (thread && thread__is_filtered(thread))
3027                 goto out;
3028
3029         trace__set_base_time(trace, evsel, sample);
3030
3031         if (handler) {
3032                 ++trace->nr_events;
3033                 handler(trace, evsel, event, sample);
3034         }
3035 out:
3036         thread__put(thread);
3037         return err;
3038 }
3039
3040 static int trace__record(struct trace *trace, int argc, const char **argv)
3041 {
3042         unsigned int rec_argc, i, j;
3043         const char **rec_argv;
3044         const char * const record_args[] = {
3045                 "record",
3046                 "-R",
3047                 "-m", "1024",
3048                 "-c", "1",
3049         };
3050         pid_t pid = getpid();
3051         char *filter = asprintf__tp_filter_pids(1, &pid);
3052         const char * const sc_args[] = { "-e", };
3053         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3054         const char * const majpf_args[] = { "-e", "major-faults" };
3055         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3056         const char * const minpf_args[] = { "-e", "minor-faults" };
3057         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3058         int err = -1;
3059
3060         /* +3 is for the event string below and the pid filter */
3061         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3062                 majpf_args_nr + minpf_args_nr + argc;
3063         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3064
3065         if (rec_argv == NULL || filter == NULL)
3066                 goto out_free;
3067
3068         j = 0;
3069         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3070                 rec_argv[j++] = record_args[i];
3071
3072         if (trace->trace_syscalls) {
3073                 for (i = 0; i < sc_args_nr; i++)
3074                         rec_argv[j++] = sc_args[i];
3075
3076                 /* event string may be different for older kernels - e.g., RHEL6 */
3077                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3078                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3079                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3080                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3081                 else {
3082                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3083                         goto out_free;
3084                 }
3085         }
3086
3087         rec_argv[j++] = "--filter";
3088         rec_argv[j++] = filter;
3089
3090         if (trace->trace_pgfaults & TRACE_PFMAJ)
3091                 for (i = 0; i < majpf_args_nr; i++)
3092                         rec_argv[j++] = majpf_args[i];
3093
3094         if (trace->trace_pgfaults & TRACE_PFMIN)
3095                 for (i = 0; i < minpf_args_nr; i++)
3096                         rec_argv[j++] = minpf_args[i];
3097
3098         for (i = 0; i < (unsigned int)argc; i++)
3099                 rec_argv[j++] = argv[i];
3100
3101         err = cmd_record(j, rec_argv);
3102 out_free:
3103         free(filter);
3104         free(rec_argv);
3105         return err;
3106 }
3107
3108 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3109
3110 static bool evlist__add_vfs_getname(struct evlist *evlist)
3111 {
3112         bool found = false;
3113         struct evsel *evsel, *tmp;
3114         struct parse_events_error err;
3115         int ret;
3116
3117         parse_events_error__init(&err);
3118         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3119         parse_events_error__exit(&err);
3120         if (ret)
3121                 return false;
3122
3123         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3124                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3125                         continue;
3126
3127                 if (evsel__field(evsel, "pathname")) {
3128                         evsel->handler = trace__vfs_getname;
3129                         found = true;
3130                         continue;
3131                 }
3132
3133                 list_del_init(&evsel->core.node);
3134                 evsel->evlist = NULL;
3135                 evsel__delete(evsel);
3136         }
3137
3138         return found;
3139 }
3140
3141 static struct evsel *evsel__new_pgfault(u64 config)
3142 {
3143         struct evsel *evsel;
3144         struct perf_event_attr attr = {
3145                 .type = PERF_TYPE_SOFTWARE,
3146                 .mmap_data = 1,
3147         };
3148
3149         attr.config = config;
3150         attr.sample_period = 1;
3151
3152         event_attr_init(&attr);
3153
3154         evsel = evsel__new(&attr);
3155         if (evsel)
3156                 evsel->handler = trace__pgfault;
3157
3158         return evsel;
3159 }
3160
3161 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3162 {
3163         struct evsel *evsel;
3164
3165         evlist__for_each_entry(evlist, evsel) {
3166                 evsel_trace__delete(evsel->priv);
3167                 evsel->priv = NULL;
3168         }
3169 }
3170
3171 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3172 {
3173         const u32 type = event->header.type;
3174         struct evsel *evsel;
3175
3176         if (type != PERF_RECORD_SAMPLE) {
3177                 trace__process_event(trace, trace->host, event, sample);
3178                 return;
3179         }
3180
3181         evsel = evlist__id2evsel(trace->evlist, sample->id);
3182         if (evsel == NULL) {
3183                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3184                 return;
3185         }
3186
3187         if (evswitch__discard(&trace->evswitch, evsel))
3188                 return;
3189
3190         trace__set_base_time(trace, evsel, sample);
3191
3192         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3193             sample->raw_data == NULL) {
3194                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3195                        evsel__name(evsel), sample->tid,
3196                        sample->cpu, sample->raw_size);
3197         } else {
3198                 tracepoint_handler handler = evsel->handler;
3199                 handler(trace, evsel, event, sample);
3200         }
3201
3202         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3203                 interrupted = true;
3204 }
3205
3206 static int trace__add_syscall_newtp(struct trace *trace)
3207 {
3208         int ret = -1;
3209         struct evlist *evlist = trace->evlist;
3210         struct evsel *sys_enter, *sys_exit;
3211
3212         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3213         if (sys_enter == NULL)
3214                 goto out;
3215
3216         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3217                 goto out_delete_sys_enter;
3218
3219         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3220         if (sys_exit == NULL)
3221                 goto out_delete_sys_enter;
3222
3223         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3224                 goto out_delete_sys_exit;
3225
3226         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3227         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3228
3229         evlist__add(evlist, sys_enter);
3230         evlist__add(evlist, sys_exit);
3231
3232         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3233                 /*
3234                  * We're interested only in the user space callchain
3235                  * leading to the syscall, allow overriding that for
3236                  * debugging reasons using --kernel_syscall_callchains
3237                  */
3238                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3239         }
3240
3241         trace->syscalls.events.sys_enter = sys_enter;
3242         trace->syscalls.events.sys_exit  = sys_exit;
3243
3244         ret = 0;
3245 out:
3246         return ret;
3247
3248 out_delete_sys_exit:
3249         evsel__delete_priv(sys_exit);
3250 out_delete_sys_enter:
3251         evsel__delete_priv(sys_enter);
3252         goto out;
3253 }
3254
3255 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3256 {
3257         int err = -1;
3258         struct evsel *sys_exit;
3259         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3260                                                 trace->ev_qualifier_ids.nr,
3261                                                 trace->ev_qualifier_ids.entries);
3262
3263         if (filter == NULL)
3264                 goto out_enomem;
3265
3266         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3267                 sys_exit = trace->syscalls.events.sys_exit;
3268                 err = evsel__append_tp_filter(sys_exit, filter);
3269         }
3270
3271         free(filter);
3272 out:
3273         return err;
3274 out_enomem:
3275         errno = ENOMEM;
3276         goto out;
3277 }
3278
3279 #ifdef HAVE_BPF_SKEL
3280 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3281 {
3282         struct bpf_program *pos, *prog = NULL;
3283         const char *sec_name;
3284
3285         if (trace->skel->obj == NULL)
3286                 return NULL;
3287
3288         bpf_object__for_each_program(pos, trace->skel->obj) {
3289                 sec_name = bpf_program__section_name(pos);
3290                 if (sec_name && !strcmp(sec_name, name)) {
3291                         prog = pos;
3292                         break;
3293                 }
3294         }
3295
3296         return prog;
3297 }
3298
3299 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3300                                                         const char *prog_name, const char *type)
3301 {
3302         struct bpf_program *prog;
3303
3304         if (prog_name == NULL) {
3305                 char default_prog_name[256];
3306                 scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->name);
3307                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3308                 if (prog != NULL)
3309                         goto out_found;
3310                 if (sc->fmt && sc->fmt->alias) {
3311                         scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->fmt->alias);
3312                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3313                         if (prog != NULL)
3314                                 goto out_found;
3315                 }
3316                 goto out_unaugmented;
3317         }
3318
3319         prog = trace__find_bpf_program_by_title(trace, prog_name);
3320
3321         if (prog != NULL) {
3322 out_found:
3323                 return prog;
3324         }
3325
3326         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3327                  prog_name, type, sc->name);
3328 out_unaugmented:
3329         return trace->skel->progs.syscall_unaugmented;
3330 }
3331
3332 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3333 {
3334         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3335
3336         if (sc == NULL)
3337                 return;
3338
3339         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3340         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3341 }
3342
3343 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3344 {
3345         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3346         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3347 }
3348
3349 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3350 {
3351         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3352         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3353 }
3354
3355 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3356 {
3357         struct tep_format_field *field, *candidate_field;
3358         int id;
3359
3360         /*
3361          * We're only interested in syscalls that have a pointer:
3362          */
3363         for (field = sc->args; field; field = field->next) {
3364                 if (field->flags & TEP_FIELD_IS_POINTER)
3365                         goto try_to_find_pair;
3366         }
3367
3368         return NULL;
3369
3370 try_to_find_pair:
3371         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3372                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3373                 struct bpf_program *pair_prog;
3374                 bool is_candidate = false;
3375
3376                 if (pair == NULL || pair == sc ||
3377                     pair->bpf_prog.sys_enter == trace->skel->progs.syscall_unaugmented)
3378                         continue;
3379
3380                 for (field = sc->args, candidate_field = pair->args;
3381                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3382                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3383                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3384
3385                         if (is_pointer) {
3386                                if (!candidate_is_pointer) {
3387                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3388                                         continue;
3389                                }
3390                         } else {
3391                                 if (candidate_is_pointer) {
3392                                         // The candidate might copy a pointer we don't have, skip it.
3393                                         goto next_candidate;
3394                                 }
3395                                 continue;
3396                         }
3397
3398                         if (strcmp(field->type, candidate_field->type))
3399                                 goto next_candidate;
3400
3401                         /*
3402                          * This is limited in the BPF program but sys_write
3403                          * uses "const char *" for its "buf" arg so we need to
3404                          * use some heuristic that is kinda future proof...
3405                          */
3406                         if (strcmp(field->type, "const char *") == 0 &&
3407                             !(strstr(field->name, "name") ||
3408                               strstr(field->name, "path") ||
3409                               strstr(field->name, "file") ||
3410                               strstr(field->name, "root") ||
3411                               strstr(field->name, "description")))
3412                                 goto next_candidate;
3413
3414                         is_candidate = true;
3415                 }
3416
3417                 if (!is_candidate)
3418                         goto next_candidate;
3419
3420                 /*
3421                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3422                  * then it may be collecting that and we then can't use it, as it would collect
3423                  * more than what is common to the two syscalls.
3424                  */
3425                 if (candidate_field) {
3426                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3427                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3428                                         goto next_candidate;
3429                 }
3430
3431                 pair_prog = pair->bpf_prog.sys_enter;
3432                 /*
3433                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3434                  * have been searched for, so search it here and if it returns the
3435                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3436                  * program for a filtered syscall on a non-filtered one.
3437                  *
3438                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3439                  * useful for "renameat2".
3440                  */
3441                 if (pair_prog == NULL) {
3442                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3443                         if (pair_prog == trace->skel->progs.syscall_unaugmented)
3444                                 goto next_candidate;
3445                 }
3446
3447                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3448                 return pair_prog;
3449         next_candidate:
3450                 continue;
3451         }
3452
3453         return NULL;
3454 }
3455
3456 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3457 {
3458         int map_enter_fd = bpf_map__fd(trace->skel->maps.syscalls_sys_enter);
3459         int map_exit_fd  = bpf_map__fd(trace->skel->maps.syscalls_sys_exit);
3460         int err = 0, key;
3461
3462         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3463                 int prog_fd;
3464
3465                 if (!trace__syscall_enabled(trace, key))
3466                         continue;
3467
3468                 trace__init_syscall_bpf_progs(trace, key);
3469
3470                 // It'll get at least the "!raw_syscalls:unaugmented"
3471                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3472                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3473                 if (err)
3474                         break;
3475                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3476                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3477                 if (err)
3478                         break;
3479         }
3480
3481         /*
3482          * Now lets do a second pass looking for enabled syscalls without
3483          * an augmenter that have a signature that is a superset of another
3484          * syscall with an augmenter so that we can auto-reuse it.
3485          *
3486          * I.e. if we have an augmenter for the "open" syscall that has
3487          * this signature:
3488          *
3489          *   int open(const char *pathname, int flags, mode_t mode);
3490          *
3491          * I.e. that will collect just the first string argument, then we
3492          * can reuse it for the 'creat' syscall, that has this signature:
3493          *
3494          *   int creat(const char *pathname, mode_t mode);
3495          *
3496          * and for:
3497          *
3498          *   int stat(const char *pathname, struct stat *statbuf);
3499          *   int lstat(const char *pathname, struct stat *statbuf);
3500          *
3501          * Because the 'open' augmenter will collect the first arg as a string,
3502          * and leave alone all the other args, which already helps with
3503          * beautifying 'stat' and 'lstat''s pathname arg.
3504          *
3505          * Then, in time, when 'stat' gets an augmenter that collects both
3506          * first and second arg (this one on the raw_syscalls:sys_exit prog
3507          * array tail call, then that one will be used.
3508          */
3509         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3510                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3511                 struct bpf_program *pair_prog;
3512                 int prog_fd;
3513
3514                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3515                         continue;
3516
3517                 /*
3518                  * For now we're just reusing the sys_enter prog, and if it
3519                  * already has an augmenter, we don't need to find one.
3520                  */
3521                 if (sc->bpf_prog.sys_enter != trace->skel->progs.syscall_unaugmented)
3522                         continue;
3523
3524                 /*
3525                  * Look at all the other syscalls for one that has a signature
3526                  * that is close enough that we can share:
3527                  */
3528                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3529                 if (pair_prog == NULL)
3530                         continue;
3531
3532                 sc->bpf_prog.sys_enter = pair_prog;
3533
3534                 /*
3535                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3536                  * with the fd for the program we're reusing:
3537                  */
3538                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3539                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3540                 if (err)
3541                         break;
3542         }
3543
3544         return err;
3545 }
3546 #endif // HAVE_BPF_SKEL
3547
3548 static int trace__set_ev_qualifier_filter(struct trace *trace)
3549 {
3550         if (trace->syscalls.events.sys_enter)
3551                 return trace__set_ev_qualifier_tp_filter(trace);
3552         return 0;
3553 }
3554
3555 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3556                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3557 {
3558         int err = 0;
3559 #ifdef HAVE_LIBBPF_SUPPORT
3560         bool value = true;
3561         int map_fd = bpf_map__fd(map);
3562         size_t i;
3563
3564         for (i = 0; i < npids; ++i) {
3565                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3566                 if (err)
3567                         break;
3568         }
3569 #endif
3570         return err;
3571 }
3572
3573 static int trace__set_filter_loop_pids(struct trace *trace)
3574 {
3575         unsigned int nr = 1, err;
3576         pid_t pids[32] = {
3577                 getpid(),
3578         };
3579         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3580
3581         while (thread && nr < ARRAY_SIZE(pids)) {
3582                 struct thread *parent = machine__find_thread(trace->host,
3583                                                              thread__ppid(thread),
3584                                                              thread__ppid(thread));
3585
3586                 if (parent == NULL)
3587                         break;
3588
3589                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3590                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3591                         pids[nr++] = thread__tid(parent);
3592                         break;
3593                 }
3594                 thread = parent;
3595         }
3596
3597         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3598         if (!err && trace->filter_pids.map)
3599                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3600
3601         return err;
3602 }
3603
3604 static int trace__set_filter_pids(struct trace *trace)
3605 {
3606         int err = 0;
3607         /*
3608          * Better not use !target__has_task() here because we need to cover the
3609          * case where no threads were specified in the command line, but a
3610          * workload was, and in that case we will fill in the thread_map when
3611          * we fork the workload in evlist__prepare_workload.
3612          */
3613         if (trace->filter_pids.nr > 0) {
3614                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3615                                                     trace->filter_pids.entries);
3616                 if (!err && trace->filter_pids.map) {
3617                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3618                                                        trace->filter_pids.entries);
3619                 }
3620         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3621                 err = trace__set_filter_loop_pids(trace);
3622         }
3623
3624         return err;
3625 }
3626
3627 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3628 {
3629         struct evlist *evlist = trace->evlist;
3630         struct perf_sample sample;
3631         int err = evlist__parse_sample(evlist, event, &sample);
3632
3633         if (err)
3634                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3635         else
3636                 trace__handle_event(trace, event, &sample);
3637
3638         return 0;
3639 }
3640
3641 static int __trace__flush_events(struct trace *trace)
3642 {
3643         u64 first = ordered_events__first_time(&trace->oe.data);
3644         u64 flush = trace->oe.last - NSEC_PER_SEC;
3645
3646         /* Is there some thing to flush.. */
3647         if (first && first < flush)
3648                 return ordered_events__flush_time(&trace->oe.data, flush);
3649
3650         return 0;
3651 }
3652
3653 static int trace__flush_events(struct trace *trace)
3654 {
3655         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3656 }
3657
3658 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3659 {
3660         int err;
3661
3662         if (!trace->sort_events)
3663                 return __trace__deliver_event(trace, event);
3664
3665         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3666         if (err && err != -1)
3667                 return err;
3668
3669         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3670         if (err)
3671                 return err;
3672
3673         return trace__flush_events(trace);
3674 }
3675
3676 static int ordered_events__deliver_event(struct ordered_events *oe,
3677                                          struct ordered_event *event)
3678 {
3679         struct trace *trace = container_of(oe, struct trace, oe.data);
3680
3681         return __trace__deliver_event(trace, event->event);
3682 }
3683
3684 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3685 {
3686         struct tep_format_field *field;
3687         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3688
3689         if (evsel->tp_format == NULL || fmt == NULL)
3690                 return NULL;
3691
3692         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3693                 if (strcmp(field->name, arg) == 0)
3694                         return fmt;
3695
3696         return NULL;
3697 }
3698
3699 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3700 {
3701         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3702
3703         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3704                 char *right = tok + 1, *right_end;
3705
3706                 if (*right == '=')
3707                         ++right;
3708
3709                 while (isspace(*right))
3710                         ++right;
3711
3712                 if (*right == '\0')
3713                         break;
3714
3715                 while (!isalpha(*left))
3716                         if (++left == tok) {
3717                                 /*
3718                                  * Bail out, can't find the name of the argument that is being
3719                                  * used in the filter, let it try to set this filter, will fail later.
3720                                  */
3721                                 return 0;
3722                         }
3723
3724                 right_end = right + 1;
3725                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3726                         ++right_end;
3727
3728                 if (isalpha(*right)) {
3729                         struct syscall_arg_fmt *fmt;
3730                         int left_size = tok - left,
3731                             right_size = right_end - right;
3732                         char arg[128];
3733
3734                         while (isspace(left[left_size - 1]))
3735                                 --left_size;
3736
3737                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3738
3739                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3740                         if (fmt == NULL) {
3741                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3742                                        arg, evsel->name, evsel->filter);
3743                                 return -1;
3744                         }
3745
3746                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3747                                  arg, (int)(right - tok), tok, right_size, right);
3748
3749                         if (fmt->strtoul) {
3750                                 u64 val;
3751                                 struct syscall_arg syscall_arg = {
3752                                         .parm = fmt->parm,
3753                                 };
3754
3755                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3756                                         char *n, expansion[19];
3757                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3758                                         int expansion_offset = right - new_filter;
3759
3760                                         pr_debug("%s", expansion);
3761
3762                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3763                                                 pr_debug(" out of memory!\n");
3764                                                 free(new_filter);
3765                                                 return -1;
3766                                         }
3767                                         if (new_filter != evsel->filter)
3768                                                 free(new_filter);
3769                                         left = n + expansion_offset + expansion_lenght;
3770                                         new_filter = n;
3771                                 } else {
3772                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3773                                                right_size, right, arg, evsel->name, evsel->filter);
3774                                         return -1;
3775                                 }
3776                         } else {
3777                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3778                                        arg, evsel->name, evsel->filter);
3779                                 return -1;
3780                         }
3781
3782                         pr_debug("\n");
3783                 } else {
3784                         left = right_end;
3785                 }
3786         }
3787
3788         if (new_filter != evsel->filter) {
3789                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3790                 evsel__set_filter(evsel, new_filter);
3791                 free(new_filter);
3792         }
3793
3794         return 0;
3795 }
3796
3797 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3798 {
3799         struct evlist *evlist = trace->evlist;
3800         struct evsel *evsel;
3801
3802         evlist__for_each_entry(evlist, evsel) {
3803                 if (evsel->filter == NULL)
3804                         continue;
3805
3806                 if (trace__expand_filter(trace, evsel)) {
3807                         *err_evsel = evsel;
3808                         return -1;
3809                 }
3810         }
3811
3812         return 0;
3813 }
3814
3815 static int trace__run(struct trace *trace, int argc, const char **argv)
3816 {
3817         struct evlist *evlist = trace->evlist;
3818         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3819         int err = -1, i;
3820         unsigned long before;
3821         const bool forks = argc > 0;
3822         bool draining = false;
3823
3824         trace->live = true;
3825
3826         if (!trace->raw_augmented_syscalls) {
3827                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3828                         goto out_error_raw_syscalls;
3829
3830                 if (trace->trace_syscalls)
3831                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3832         }
3833
3834         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3835                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3836                 if (pgfault_maj == NULL)
3837                         goto out_error_mem;
3838                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3839                 evlist__add(evlist, pgfault_maj);
3840         }
3841
3842         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3843                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3844                 if (pgfault_min == NULL)
3845                         goto out_error_mem;
3846                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3847                 evlist__add(evlist, pgfault_min);
3848         }
3849
3850         /* Enable ignoring missing threads when -u/-p option is defined. */
3851         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3852
3853         if (trace->sched &&
3854             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3855                 goto out_error_sched_stat_runtime;
3856         /*
3857          * If a global cgroup was set, apply it to all the events without an
3858          * explicit cgroup. I.e.:
3859          *
3860          *      trace -G A -e sched:*switch
3861          *
3862          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3863          * _and_ sched:sched_switch to the 'A' cgroup, while:
3864          *
3865          * trace -e sched:*switch -G A
3866          *
3867          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3868          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3869          * a cgroup (on the root cgroup, sys wide, etc).
3870          *
3871          * Multiple cgroups:
3872          *
3873          * trace -G A -e sched:*switch -G B
3874          *
3875          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3876          * to the 'B' cgroup.
3877          *
3878          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3879          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3880          */
3881         if (trace->cgroup)
3882                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3883
3884         err = evlist__create_maps(evlist, &trace->opts.target);
3885         if (err < 0) {
3886                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3887                 goto out_delete_evlist;
3888         }
3889
3890         err = trace__symbols_init(trace, evlist);
3891         if (err < 0) {
3892                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3893                 goto out_delete_evlist;
3894         }
3895
3896         evlist__config(evlist, &trace->opts, &callchain_param);
3897
3898         if (forks) {
3899                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3900                 if (err < 0) {
3901                         fprintf(trace->output, "Couldn't run the workload!\n");
3902                         goto out_delete_evlist;
3903                 }
3904                 workload_pid = evlist->workload.pid;
3905         }
3906
3907         err = evlist__open(evlist);
3908         if (err < 0)
3909                 goto out_error_open;
3910 #ifdef HAVE_BPF_SKEL
3911         if (trace->syscalls.events.bpf_output) {
3912                 struct perf_cpu cpu;
3913
3914                 /*
3915                  * Set up the __augmented_syscalls__ BPF map to hold for each
3916                  * CPU the bpf-output event's file descriptor.
3917                  */
3918                 perf_cpu_map__for_each_cpu(cpu, i, trace->syscalls.events.bpf_output->core.cpus) {
3919                         bpf_map__update_elem(trace->skel->maps.__augmented_syscalls__,
3920                                         &cpu.cpu, sizeof(int),
3921                                         xyarray__entry(trace->syscalls.events.bpf_output->core.fd,
3922                                                        cpu.cpu, 0),
3923                                         sizeof(__u32), BPF_ANY);
3924                 }
3925         }
3926 #endif
3927         err = trace__set_filter_pids(trace);
3928         if (err < 0)
3929                 goto out_error_mem;
3930
3931 #ifdef HAVE_BPF_SKEL
3932         if (trace->skel && trace->skel->progs.sys_enter)
3933                 trace__init_syscalls_bpf_prog_array_maps(trace);
3934 #endif
3935
3936         if (trace->ev_qualifier_ids.nr > 0) {
3937                 err = trace__set_ev_qualifier_filter(trace);
3938                 if (err < 0)
3939                         goto out_errno;
3940
3941                 if (trace->syscalls.events.sys_exit) {
3942                         pr_debug("event qualifier tracepoint filter: %s\n",
3943                                  trace->syscalls.events.sys_exit->filter);
3944                 }
3945         }
3946
3947         /*
3948          * If the "close" syscall is not traced, then we will not have the
3949          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3950          * fd->pathname table and were ending up showing the last value set by
3951          * syscalls opening a pathname and associating it with a descriptor or
3952          * reading it from /proc/pid/fd/ in cases where that doesn't make
3953          * sense.
3954          *
3955          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3956          *  not in use.
3957          */
3958         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3959
3960         err = trace__expand_filters(trace, &evsel);
3961         if (err)
3962                 goto out_delete_evlist;
3963         err = evlist__apply_filters(evlist, &evsel);
3964         if (err < 0)
3965                 goto out_error_apply_filters;
3966
3967         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3968         if (err < 0)
3969                 goto out_error_mmap;
3970
3971         if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
3972                 evlist__enable(evlist);
3973
3974         if (forks)
3975                 evlist__start_workload(evlist);
3976
3977         if (trace->opts.target.initial_delay) {
3978                 usleep(trace->opts.target.initial_delay * 1000);
3979                 evlist__enable(evlist);
3980         }
3981
3982         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3983                 perf_thread_map__nr(evlist->core.threads) > 1 ||
3984                 evlist__first(evlist)->core.attr.inherit;
3985
3986         /*
3987          * Now that we already used evsel->core.attr to ask the kernel to setup the
3988          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3989          * trace__resolve_callchain(), allowing per-event max-stack settings
3990          * to override an explicitly set --max-stack global setting.
3991          */
3992         evlist__for_each_entry(evlist, evsel) {
3993                 if (evsel__has_callchain(evsel) &&
3994                     evsel->core.attr.sample_max_stack == 0)
3995                         evsel->core.attr.sample_max_stack = trace->max_stack;
3996         }
3997 again:
3998         before = trace->nr_events;
3999
4000         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4001                 union perf_event *event;
4002                 struct mmap *md;
4003
4004                 md = &evlist->mmap[i];
4005                 if (perf_mmap__read_init(&md->core) < 0)
4006                         continue;
4007
4008                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4009                         ++trace->nr_events;
4010
4011                         err = trace__deliver_event(trace, event);
4012                         if (err)
4013                                 goto out_disable;
4014
4015                         perf_mmap__consume(&md->core);
4016
4017                         if (interrupted)
4018                                 goto out_disable;
4019
4020                         if (done && !draining) {
4021                                 evlist__disable(evlist);
4022                                 draining = true;
4023                         }
4024                 }
4025                 perf_mmap__read_done(&md->core);
4026         }
4027
4028         if (trace->nr_events == before) {
4029                 int timeout = done ? 100 : -1;
4030
4031                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4032                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4033                                 draining = true;
4034
4035                         goto again;
4036                 } else {
4037                         if (trace__flush_events(trace))
4038                                 goto out_disable;
4039                 }
4040         } else {
4041                 goto again;
4042         }
4043
4044 out_disable:
4045         thread__zput(trace->current);
4046
4047         evlist__disable(evlist);
4048
4049         if (trace->sort_events)
4050                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4051
4052         if (!err) {
4053                 if (trace->summary)
4054                         trace__fprintf_thread_summary(trace, trace->output);
4055
4056                 if (trace->show_tool_stats) {
4057                         fprintf(trace->output, "Stats:\n "
4058                                                " vfs_getname : %" PRIu64 "\n"
4059                                                " proc_getname: %" PRIu64 "\n",
4060                                 trace->stats.vfs_getname,
4061                                 trace->stats.proc_getname);
4062                 }
4063         }
4064
4065 out_delete_evlist:
4066         trace__symbols__exit(trace);
4067         evlist__free_syscall_tp_fields(evlist);
4068         evlist__delete(evlist);
4069         cgroup__put(trace->cgroup);
4070         trace->evlist = NULL;
4071         trace->live = false;
4072         return err;
4073 {
4074         char errbuf[BUFSIZ];
4075
4076 out_error_sched_stat_runtime:
4077         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4078         goto out_error;
4079
4080 out_error_raw_syscalls:
4081         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4082         goto out_error;
4083
4084 out_error_mmap:
4085         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4086         goto out_error;
4087
4088 out_error_open:
4089         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4090
4091 out_error:
4092         fprintf(trace->output, "%s\n", errbuf);
4093         goto out_delete_evlist;
4094
4095 out_error_apply_filters:
4096         fprintf(trace->output,
4097                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4098                 evsel->filter, evsel__name(evsel), errno,
4099                 str_error_r(errno, errbuf, sizeof(errbuf)));
4100         goto out_delete_evlist;
4101 }
4102 out_error_mem:
4103         fprintf(trace->output, "Not enough memory to run!\n");
4104         goto out_delete_evlist;
4105
4106 out_errno:
4107         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4108         goto out_delete_evlist;
4109 }
4110
4111 static int trace__replay(struct trace *trace)
4112 {
4113         const struct evsel_str_handler handlers[] = {
4114                 { "probe:vfs_getname",       trace__vfs_getname, },
4115         };
4116         struct perf_data data = {
4117                 .path  = input_name,
4118                 .mode  = PERF_DATA_MODE_READ,
4119                 .force = trace->force,
4120         };
4121         struct perf_session *session;
4122         struct evsel *evsel;
4123         int err = -1;
4124
4125         trace->tool.sample        = trace__process_sample;
4126         trace->tool.mmap          = perf_event__process_mmap;
4127         trace->tool.mmap2         = perf_event__process_mmap2;
4128         trace->tool.comm          = perf_event__process_comm;
4129         trace->tool.exit          = perf_event__process_exit;
4130         trace->tool.fork          = perf_event__process_fork;
4131         trace->tool.attr          = perf_event__process_attr;
4132         trace->tool.tracing_data  = perf_event__process_tracing_data;
4133         trace->tool.build_id      = perf_event__process_build_id;
4134         trace->tool.namespaces    = perf_event__process_namespaces;
4135
4136         trace->tool.ordered_events = true;
4137         trace->tool.ordering_requires_timestamps = true;
4138
4139         /* add tid to output */
4140         trace->multiple_threads = true;
4141
4142         session = perf_session__new(&data, &trace->tool);
4143         if (IS_ERR(session))
4144                 return PTR_ERR(session);
4145
4146         if (trace->opts.target.pid)
4147                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4148
4149         if (trace->opts.target.tid)
4150                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4151
4152         if (symbol__init(&session->header.env) < 0)
4153                 goto out;
4154
4155         trace->host = &session->machines.host;
4156
4157         err = perf_session__set_tracepoints_handlers(session, handlers);
4158         if (err)
4159                 goto out;
4160
4161         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4162         trace->syscalls.events.sys_enter = evsel;
4163         /* older kernels have syscalls tp versus raw_syscalls */
4164         if (evsel == NULL)
4165                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4166
4167         if (evsel &&
4168             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4169             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4170                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4171                 goto out;
4172         }
4173
4174         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4175         trace->syscalls.events.sys_exit = evsel;
4176         if (evsel == NULL)
4177                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4178         if (evsel &&
4179             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4180             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4181                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4182                 goto out;
4183         }
4184
4185         evlist__for_each_entry(session->evlist, evsel) {
4186                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4187                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4188                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4189                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4190                         evsel->handler = trace__pgfault;
4191         }
4192
4193         setup_pager();
4194
4195         err = perf_session__process_events(session);
4196         if (err)
4197                 pr_err("Failed to process events, error %d", err);
4198
4199         else if (trace->summary)
4200                 trace__fprintf_thread_summary(trace, trace->output);
4201
4202 out:
4203         perf_session__delete(session);
4204
4205         return err;
4206 }
4207
4208 static size_t trace__fprintf_threads_header(FILE *fp)
4209 {
4210         size_t printed;
4211
4212         printed  = fprintf(fp, "\n Summary of events:\n\n");
4213
4214         return printed;
4215 }
4216
4217 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4218         struct syscall_stats *stats;
4219         double               msecs;
4220         int                  syscall;
4221 )
4222 {
4223         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4224         struct syscall_stats *stats = source->priv;
4225
4226         entry->syscall = source->i;
4227         entry->stats   = stats;
4228         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4229 }
4230
4231 static size_t thread__dump_stats(struct thread_trace *ttrace,
4232                                  struct trace *trace, FILE *fp)
4233 {
4234         size_t printed = 0;
4235         struct syscall *sc;
4236         struct rb_node *nd;
4237         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4238
4239         if (syscall_stats == NULL)
4240                 return 0;
4241
4242         printed += fprintf(fp, "\n");
4243
4244         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4245         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4246         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4247
4248         resort_rb__for_each_entry(nd, syscall_stats) {
4249                 struct syscall_stats *stats = syscall_stats_entry->stats;
4250                 if (stats) {
4251                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4252                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4253                         double avg = avg_stats(&stats->stats);
4254                         double pct;
4255                         u64 n = (u64)stats->stats.n;
4256
4257                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4258                         avg /= NSEC_PER_MSEC;
4259
4260                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4261                         printed += fprintf(fp, "   %-15s", sc->name);
4262                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4263                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4264                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4265
4266                         if (trace->errno_summary && stats->nr_failures) {
4267                                 int e;
4268
4269                                 for (e = 0; e < stats->max_errno; ++e) {
4270                                         if (stats->errnos[e] != 0)
4271                                                 fprintf(fp, "\t\t\t\t%s: %d\n", perf_env__arch_strerrno(trace->host->env, e + 1), stats->errnos[e]);
4272                                 }
4273                         }
4274                 }
4275         }
4276
4277         resort_rb__delete(syscall_stats);
4278         printed += fprintf(fp, "\n\n");
4279
4280         return printed;
4281 }
4282
4283 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4284 {
4285         size_t printed = 0;
4286         struct thread_trace *ttrace = thread__priv(thread);
4287         double ratio;
4288
4289         if (ttrace == NULL)
4290                 return 0;
4291
4292         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4293
4294         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread__tid(thread));
4295         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4296         printed += fprintf(fp, "%.1f%%", ratio);
4297         if (ttrace->pfmaj)
4298                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4299         if (ttrace->pfmin)
4300                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4301         if (trace->sched)
4302                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4303         else if (fputc('\n', fp) != EOF)
4304                 ++printed;
4305
4306         printed += thread__dump_stats(ttrace, trace, fp);
4307
4308         return printed;
4309 }
4310
4311 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4312 {
4313         return ttrace ? ttrace->nr_events : 0;
4314 }
4315
4316 static int trace_nr_events_cmp(void *priv __maybe_unused,
4317                                const struct list_head *la,
4318                                const struct list_head *lb)
4319 {
4320         struct thread_list *a = list_entry(la, struct thread_list, list);
4321         struct thread_list *b = list_entry(lb, struct thread_list, list);
4322         unsigned long a_nr_events = thread__nr_events(thread__priv(a->thread));
4323         unsigned long b_nr_events = thread__nr_events(thread__priv(b->thread));
4324
4325         if (a_nr_events != b_nr_events)
4326                 return a_nr_events < b_nr_events ? -1 : 1;
4327
4328         /* Identical number of threads, place smaller tids first. */
4329         return thread__tid(a->thread) < thread__tid(b->thread)
4330                 ? -1
4331                 : (thread__tid(a->thread) > thread__tid(b->thread) ? 1 : 0);
4332 }
4333
4334 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4335 {
4336         size_t printed = trace__fprintf_threads_header(fp);
4337         LIST_HEAD(threads);
4338
4339         if (machine__thread_list(trace->host, &threads) == 0) {
4340                 struct thread_list *pos;
4341
4342                 list_sort(NULL, &threads, trace_nr_events_cmp);
4343
4344                 list_for_each_entry(pos, &threads, list)
4345                         printed += trace__fprintf_thread(fp, pos->thread, trace);
4346         }
4347         thread_list__delete(&threads);
4348         return printed;
4349 }
4350
4351 static int trace__set_duration(const struct option *opt, const char *str,
4352                                int unset __maybe_unused)
4353 {
4354         struct trace *trace = opt->value;
4355
4356         trace->duration_filter = atof(str);
4357         return 0;
4358 }
4359
4360 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4361                                               int unset __maybe_unused)
4362 {
4363         int ret = -1;
4364         size_t i;
4365         struct trace *trace = opt->value;
4366         /*
4367          * FIXME: introduce a intarray class, plain parse csv and create a
4368          * { int nr, int entries[] } struct...
4369          */
4370         struct intlist *list = intlist__new(str);
4371
4372         if (list == NULL)
4373                 return -1;
4374
4375         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4376         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4377
4378         if (trace->filter_pids.entries == NULL)
4379                 goto out;
4380
4381         trace->filter_pids.entries[0] = getpid();
4382
4383         for (i = 1; i < trace->filter_pids.nr; ++i)
4384                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4385
4386         intlist__delete(list);
4387         ret = 0;
4388 out:
4389         return ret;
4390 }
4391
4392 static int trace__open_output(struct trace *trace, const char *filename)
4393 {
4394         struct stat st;
4395
4396         if (!stat(filename, &st) && st.st_size) {
4397                 char oldname[PATH_MAX];
4398
4399                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4400                 unlink(oldname);
4401                 rename(filename, oldname);
4402         }
4403
4404         trace->output = fopen(filename, "w");
4405
4406         return trace->output == NULL ? -errno : 0;
4407 }
4408
4409 static int parse_pagefaults(const struct option *opt, const char *str,
4410                             int unset __maybe_unused)
4411 {
4412         int *trace_pgfaults = opt->value;
4413
4414         if (strcmp(str, "all") == 0)
4415                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4416         else if (strcmp(str, "maj") == 0)
4417                 *trace_pgfaults |= TRACE_PFMAJ;
4418         else if (strcmp(str, "min") == 0)
4419                 *trace_pgfaults |= TRACE_PFMIN;
4420         else
4421                 return -1;
4422
4423         return 0;
4424 }
4425
4426 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4427 {
4428         struct evsel *evsel;
4429
4430         evlist__for_each_entry(evlist, evsel) {
4431                 if (evsel->handler == NULL)
4432                         evsel->handler = handler;
4433         }
4434 }
4435
4436 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4437 {
4438         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4439
4440         if (fmt) {
4441                 const struct syscall_fmt *scfmt = syscall_fmt__find(name);
4442
4443                 if (scfmt) {
4444                         int skip = 0;
4445
4446                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4447                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4448                                 ++skip;
4449
4450                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4451                 }
4452         }
4453 }
4454
4455 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4456 {
4457         struct evsel *evsel;
4458
4459         evlist__for_each_entry(evlist, evsel) {
4460                 if (evsel->priv || !evsel->tp_format)
4461                         continue;
4462
4463                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4464                         evsel__init_tp_arg_scnprintf(evsel);
4465                         continue;
4466                 }
4467
4468                 if (evsel__init_syscall_tp(evsel))
4469                         return -1;
4470
4471                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4472                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4473
4474                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4475                                 return -1;
4476
4477                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4478                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4479                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4480
4481                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4482                                 return -1;
4483
4484                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4485                 }
4486         }
4487
4488         return 0;
4489 }
4490
4491 /*
4492  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4493  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4494  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4495  *
4496  * It'd be better to introduce a parse_options() variant that would return a
4497  * list with the terms it didn't match to an event...
4498  */
4499 static int trace__parse_events_option(const struct option *opt, const char *str,
4500                                       int unset __maybe_unused)
4501 {
4502         struct trace *trace = (struct trace *)opt->value;
4503         const char *s = str;
4504         char *sep = NULL, *lists[2] = { NULL, NULL, };
4505         int len = strlen(str) + 1, err = -1, list, idx;
4506         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4507         char group_name[PATH_MAX];
4508         const struct syscall_fmt *fmt;
4509
4510         if (strace_groups_dir == NULL)
4511                 return -1;
4512
4513         if (*s == '!') {
4514                 ++s;
4515                 trace->not_ev_qualifier = true;
4516         }
4517
4518         while (1) {
4519                 if ((sep = strchr(s, ',')) != NULL)
4520                         *sep = '\0';
4521
4522                 list = 0;
4523                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4524                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4525                         list = 1;
4526                         goto do_concat;
4527                 }
4528
4529                 fmt = syscall_fmt__find_by_alias(s);
4530                 if (fmt != NULL) {
4531                         list = 1;
4532                         s = fmt->name;
4533                 } else {
4534                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4535                         if (access(group_name, R_OK) == 0)
4536                                 list = 1;
4537                 }
4538 do_concat:
4539                 if (lists[list]) {
4540                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4541                 } else {
4542                         lists[list] = malloc(len);
4543                         if (lists[list] == NULL)
4544                                 goto out;
4545                         strcpy(lists[list], s);
4546                 }
4547
4548                 if (!sep)
4549                         break;
4550
4551                 *sep = ',';
4552                 s = sep + 1;
4553         }
4554
4555         if (lists[1] != NULL) {
4556                 struct strlist_config slist_config = {
4557                         .dirname = strace_groups_dir,
4558                 };
4559
4560                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4561                 if (trace->ev_qualifier == NULL) {
4562                         fputs("Not enough memory to parse event qualifier", trace->output);
4563                         goto out;
4564                 }
4565
4566                 if (trace__validate_ev_qualifier(trace))
4567                         goto out;
4568                 trace->trace_syscalls = true;
4569         }
4570
4571         err = 0;
4572
4573         if (lists[0]) {
4574                 struct parse_events_option_args parse_events_option_args = {
4575                         .evlistp = &trace->evlist,
4576                 };
4577                 struct option o = {
4578                         .value = &parse_events_option_args,
4579                 };
4580                 err = parse_events_option(&o, lists[0], 0);
4581         }
4582 out:
4583         free(strace_groups_dir);
4584         free(lists[0]);
4585         free(lists[1]);
4586         if (sep)
4587                 *sep = ',';
4588
4589         return err;
4590 }
4591
4592 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4593 {
4594         struct trace *trace = opt->value;
4595
4596         if (!list_empty(&trace->evlist->core.entries)) {
4597                 struct option o = {
4598                         .value = &trace->evlist,
4599                 };
4600                 return parse_cgroups(&o, str, unset);
4601         }
4602         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4603
4604         return 0;
4605 }
4606
4607 static int trace__config(const char *var, const char *value, void *arg)
4608 {
4609         struct trace *trace = arg;
4610         int err = 0;
4611
4612         if (!strcmp(var, "trace.add_events")) {
4613                 trace->perfconfig_events = strdup(value);
4614                 if (trace->perfconfig_events == NULL) {
4615                         pr_err("Not enough memory for %s\n", "trace.add_events");
4616                         return -1;
4617                 }
4618         } else if (!strcmp(var, "trace.show_timestamp")) {
4619                 trace->show_tstamp = perf_config_bool(var, value);
4620         } else if (!strcmp(var, "trace.show_duration")) {
4621                 trace->show_duration = perf_config_bool(var, value);
4622         } else if (!strcmp(var, "trace.show_arg_names")) {
4623                 trace->show_arg_names = perf_config_bool(var, value);
4624                 if (!trace->show_arg_names)
4625                         trace->show_zeros = true;
4626         } else if (!strcmp(var, "trace.show_zeros")) {
4627                 bool new_show_zeros = perf_config_bool(var, value);
4628                 if (!trace->show_arg_names && !new_show_zeros) {
4629                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4630                         goto out;
4631                 }
4632                 trace->show_zeros = new_show_zeros;
4633         } else if (!strcmp(var, "trace.show_prefix")) {
4634                 trace->show_string_prefix = perf_config_bool(var, value);
4635         } else if (!strcmp(var, "trace.no_inherit")) {
4636                 trace->opts.no_inherit = perf_config_bool(var, value);
4637         } else if (!strcmp(var, "trace.args_alignment")) {
4638                 int args_alignment = 0;
4639                 if (perf_config_int(&args_alignment, var, value) == 0)
4640                         trace->args_alignment = args_alignment;
4641         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4642                 if (strcasecmp(value, "libtraceevent") == 0)
4643                         trace->libtraceevent_print = true;
4644                 else if (strcasecmp(value, "libbeauty") == 0)
4645                         trace->libtraceevent_print = false;
4646         }
4647 out:
4648         return err;
4649 }
4650
4651 static void trace__exit(struct trace *trace)
4652 {
4653         int i;
4654
4655         strlist__delete(trace->ev_qualifier);
4656         zfree(&trace->ev_qualifier_ids.entries);
4657         if (trace->syscalls.table) {
4658                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4659                         syscall__exit(&trace->syscalls.table[i]);
4660                 zfree(&trace->syscalls.table);
4661         }
4662         syscalltbl__delete(trace->sctbl);
4663         zfree(&trace->perfconfig_events);
4664 }
4665
4666 #ifdef HAVE_BPF_SKEL
4667 static int bpf__setup_bpf_output(struct evlist *evlist)
4668 {
4669         int err = parse_event(evlist, "bpf-output/no-inherit=1,name=__augmented_syscalls__/");
4670
4671         if (err)
4672                 pr_debug("ERROR: failed to create the \"__augmented_syscalls__\" bpf-output event\n");
4673
4674         return err;
4675 }
4676 #endif
4677
4678 int cmd_trace(int argc, const char **argv)
4679 {
4680         const char *trace_usage[] = {
4681                 "perf trace [<options>] [<command>]",
4682                 "perf trace [<options>] -- <command> [<options>]",
4683                 "perf trace record [<options>] [<command>]",
4684                 "perf trace record [<options>] -- <command> [<options>]",
4685                 NULL
4686         };
4687         struct trace trace = {
4688                 .opts = {
4689                         .target = {
4690                                 .uid       = UINT_MAX,
4691                                 .uses_mmap = true,
4692                         },
4693                         .user_freq     = UINT_MAX,
4694                         .user_interval = ULLONG_MAX,
4695                         .no_buffering  = true,
4696                         .mmap_pages    = UINT_MAX,
4697                 },
4698                 .output = stderr,
4699                 .show_comm = true,
4700                 .show_tstamp = true,
4701                 .show_duration = true,
4702                 .show_arg_names = true,
4703                 .args_alignment = 70,
4704                 .trace_syscalls = false,
4705                 .kernel_syscallchains = false,
4706                 .max_stack = UINT_MAX,
4707                 .max_events = ULONG_MAX,
4708         };
4709         const char *output_name = NULL;
4710         const struct option trace_options[] = {
4711         OPT_CALLBACK('e', "event", &trace, "event",
4712                      "event/syscall selector. use 'perf list' to list available events",
4713                      trace__parse_events_option),
4714         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4715                      "event filter", parse_filter),
4716         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4717                     "show the thread COMM next to its id"),
4718         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4719         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4720                      trace__parse_events_option),
4721         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4722         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4723         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4724                     "trace events on existing process id"),
4725         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4726                     "trace events on existing thread id"),
4727         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4728                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4729         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4730                     "system-wide collection from all CPUs"),
4731         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4732                     "list of cpus to monitor"),
4733         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4734                     "child tasks do not inherit counters"),
4735         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4736                      "number of mmap data pages", evlist__parse_mmap_pages),
4737         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4738                    "user to profile"),
4739         OPT_CALLBACK(0, "duration", &trace, "float",
4740                      "show only events with duration > N.M ms",
4741                      trace__set_duration),
4742         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4743         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4744         OPT_BOOLEAN('T', "time", &trace.full_time,
4745                     "Show full timestamp, not time relative to first start"),
4746         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4747                     "Show only syscalls that failed"),
4748         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4749                     "Show only syscall summary with statistics"),
4750         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4751                     "Show all syscalls and summary with statistics"),
4752         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4753                     "Show errno stats per syscall, use with -s or -S"),
4754         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4755                      "Trace pagefaults", parse_pagefaults, "maj"),
4756         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4757         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4758         OPT_CALLBACK(0, "call-graph", &trace.opts,
4759                      "record_mode[,record_size]", record_callchain_help,
4760                      &record_parse_callchain_opt),
4761         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4762                     "Use libtraceevent to print the tracepoint arguments."),
4763         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4764                     "Show the kernel callchains on the syscall exit path"),
4765         OPT_ULONG(0, "max-events", &trace.max_events,
4766                 "Set the maximum number of events to print, exit after that is reached. "),
4767         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4768                      "Set the minimum stack depth when parsing the callchain, "
4769                      "anything below the specified depth will be ignored."),
4770         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4771                      "Set the maximum stack depth when parsing the callchain, "
4772                      "anything beyond the specified depth will be ignored. "
4773                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4774         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4775                         "Sort batch of events before processing, use if getting out of order events"),
4776         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4777                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4778         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4779                         "per thread proc mmap processing timeout in ms"),
4780         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4781                      trace__parse_cgroups),
4782         OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4783                      "ms to wait before starting measurement after program "
4784                      "start"),
4785         OPTS_EVSWITCH(&trace.evswitch),
4786         OPT_END()
4787         };
4788         bool __maybe_unused max_stack_user_set = true;
4789         bool mmap_pages_user_set = true;
4790         struct evsel *evsel;
4791         const char * const trace_subcommands[] = { "record", NULL };
4792         int err = -1;
4793         char bf[BUFSIZ];
4794         struct sigaction sigchld_act;
4795
4796         signal(SIGSEGV, sighandler_dump_stack);
4797         signal(SIGFPE, sighandler_dump_stack);
4798         signal(SIGINT, sighandler_interrupt);
4799
4800         memset(&sigchld_act, 0, sizeof(sigchld_act));
4801         sigchld_act.sa_flags = SA_SIGINFO;
4802         sigchld_act.sa_sigaction = sighandler_chld;
4803         sigaction(SIGCHLD, &sigchld_act, NULL);
4804
4805         trace.evlist = evlist__new();
4806         trace.sctbl = syscalltbl__new();
4807
4808         if (trace.evlist == NULL || trace.sctbl == NULL) {
4809                 pr_err("Not enough memory to run!\n");
4810                 err = -ENOMEM;
4811                 goto out;
4812         }
4813
4814         /*
4815          * Parsing .perfconfig may entail creating a BPF event, that may need
4816          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4817          * is too small. This affects just this process, not touching the
4818          * global setting. If it fails we'll get something in 'perf trace -v'
4819          * to help diagnose the problem.
4820          */
4821         rlimit__bump_memlock();
4822
4823         err = perf_config(trace__config, &trace);
4824         if (err)
4825                 goto out;
4826
4827         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4828                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4829
4830         /*
4831          * Here we already passed thru trace__parse_events_option() and it has
4832          * already figured out if -e syscall_name, if not but if --event
4833          * foo:bar was used, the user is interested _just_ in those, say,
4834          * tracepoint events, not in the strace-like syscall-name-based mode.
4835          *
4836          * This is important because we need to check if strace-like mode is
4837          * needed to decided if we should filter out the eBPF
4838          * __augmented_syscalls__ code, if it is in the mix, say, via
4839          * .perfconfig trace.add_events, and filter those out.
4840          */
4841         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4842             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4843                 trace.trace_syscalls = true;
4844         }
4845         /*
4846          * Now that we have --verbose figured out, lets see if we need to parse
4847          * events from .perfconfig, so that if those events fail parsing, say some
4848          * BPF program fails, then we'll be able to use --verbose to see what went
4849          * wrong in more detail.
4850          */
4851         if (trace.perfconfig_events != NULL) {
4852                 struct parse_events_error parse_err;
4853
4854                 parse_events_error__init(&parse_err);
4855                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4856                 if (err)
4857                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4858                 parse_events_error__exit(&parse_err);
4859                 if (err)
4860                         goto out;
4861         }
4862
4863         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4864                 usage_with_options_msg(trace_usage, trace_options,
4865                                        "cgroup monitoring only available in system-wide mode");
4866         }
4867
4868 #ifdef HAVE_BPF_SKEL
4869         if (!trace.trace_syscalls)
4870                 goto skip_augmentation;
4871
4872         trace.skel = augmented_raw_syscalls_bpf__open();
4873         if (!trace.skel) {
4874                 pr_debug("Failed to open augmented syscalls BPF skeleton");
4875         } else {
4876                 /*
4877                  * Disable attaching the BPF programs except for sys_enter and
4878                  * sys_exit that tail call into this as necessary.
4879                  */
4880                 struct bpf_program *prog;
4881
4882                 bpf_object__for_each_program(prog, trace.skel->obj) {
4883                         if (prog != trace.skel->progs.sys_enter && prog != trace.skel->progs.sys_exit)
4884                                 bpf_program__set_autoattach(prog, /*autoattach=*/false);
4885                 }
4886
4887                 err = augmented_raw_syscalls_bpf__load(trace.skel);
4888
4889                 if (err < 0) {
4890                         libbpf_strerror(err, bf, sizeof(bf));
4891                         pr_debug("Failed to load augmented syscalls BPF skeleton: %s\n", bf);
4892                 } else {
4893                         augmented_raw_syscalls_bpf__attach(trace.skel);
4894                         trace__add_syscall_newtp(&trace);
4895                 }
4896         }
4897
4898         err = bpf__setup_bpf_output(trace.evlist);
4899         if (err) {
4900                 libbpf_strerror(err, bf, sizeof(bf));
4901                 pr_err("ERROR: Setup BPF output event failed: %s\n", bf);
4902                 goto out;
4903         }
4904         trace.syscalls.events.bpf_output = evlist__last(trace.evlist);
4905         assert(!strcmp(evsel__name(trace.syscalls.events.bpf_output), "__augmented_syscalls__"));
4906 skip_augmentation:
4907 #endif
4908         err = -1;
4909
4910         if (trace.trace_pgfaults) {
4911                 trace.opts.sample_address = true;
4912                 trace.opts.sample_time = true;
4913         }
4914
4915         if (trace.opts.mmap_pages == UINT_MAX)
4916                 mmap_pages_user_set = false;
4917
4918         if (trace.max_stack == UINT_MAX) {
4919                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4920                 max_stack_user_set = false;
4921         }
4922
4923 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4924         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4925                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4926         }
4927 #endif
4928
4929         if (callchain_param.enabled) {
4930                 if (!mmap_pages_user_set && geteuid() == 0)
4931                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4932
4933                 symbol_conf.use_callchain = true;
4934         }
4935
4936         if (trace.evlist->core.nr_entries > 0) {
4937                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4938                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4939                         perror("failed to set syscalls:* tracepoint fields");
4940                         goto out;
4941                 }
4942         }
4943
4944         if (trace.sort_events) {
4945                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4946                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4947         }
4948
4949         /*
4950          * If we are augmenting syscalls, then combine what we put in the
4951          * __augmented_syscalls__ BPF map with what is in the
4952          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4953          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4954          *
4955          * We'll switch to look at two BPF maps, one for sys_enter and the
4956          * other for sys_exit when we start augmenting the sys_exit paths with
4957          * buffers that are being copied from kernel to userspace, think 'read'
4958          * syscall.
4959          */
4960         if (trace.syscalls.events.bpf_output) {
4961                 evlist__for_each_entry(trace.evlist, evsel) {
4962                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4963
4964                         if (raw_syscalls_sys_exit) {
4965                                 trace.raw_augmented_syscalls = true;
4966                                 goto init_augmented_syscall_tp;
4967                         }
4968
4969                         if (trace.syscalls.events.bpf_output->priv == NULL &&
4970                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
4971                                 struct evsel *augmented = trace.syscalls.events.bpf_output;
4972                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
4973                                     evsel__init_augmented_syscall_tp_args(augmented))
4974                                         goto out;
4975                                 /*
4976                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4977                                  * Above we made sure we can get from the payload the tp fields
4978                                  * that we get from syscalls:sys_enter tracefs format file.
4979                                  */
4980                                 augmented->handler = trace__sys_enter;
4981                                 /*
4982                                  * Now we do the same for the *syscalls:sys_enter event so that
4983                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4984                                  * as not to filter it, then we'll handle it just like we would
4985                                  * for the BPF_OUTPUT one:
4986                                  */
4987                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
4988                                     evsel__init_augmented_syscall_tp_args(evsel))
4989                                         goto out;
4990                                 evsel->handler = trace__sys_enter;
4991                         }
4992
4993                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
4994                                 struct syscall_tp *sc;
4995 init_augmented_syscall_tp:
4996                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
4997                                         goto out;
4998                                 sc = __evsel__syscall_tp(evsel);
4999                                 /*
5000                                  * For now with BPF raw_augmented we hook into
5001                                  * raw_syscalls:sys_enter and there we get all
5002                                  * 6 syscall args plus the tracepoint common
5003                                  * fields and the syscall_nr (another long).
5004                                  * So we check if that is the case and if so
5005                                  * don't look after the sc->args_size but
5006                                  * always after the full raw_syscalls:sys_enter
5007                                  * payload, which is fixed.
5008                                  *
5009                                  * We'll revisit this later to pass
5010                                  * s->args_size to the BPF augmenter (now
5011                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5012                                  * so that it copies only what we need for each
5013                                  * syscall, like what happens when we use
5014                                  * syscalls:sys_enter_NAME, so that we reduce
5015                                  * the kernel/userspace traffic to just what is
5016                                  * needed for each syscall.
5017                                  */
5018                                 if (trace.raw_augmented_syscalls)
5019                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5020                                 evsel__init_augmented_syscall_tp_ret(evsel);
5021                                 evsel->handler = trace__sys_exit;
5022                         }
5023                 }
5024         }
5025
5026         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5027                 return trace__record(&trace, argc-1, &argv[1]);
5028
5029         /* Using just --errno-summary will trigger --summary */
5030         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5031                 trace.summary_only = true;
5032
5033         /* summary_only implies summary option, but don't overwrite summary if set */
5034         if (trace.summary_only)
5035                 trace.summary = trace.summary_only;
5036
5037         if (output_name != NULL) {
5038                 err = trace__open_output(&trace, output_name);
5039                 if (err < 0) {
5040                         perror("failed to create output file");
5041                         goto out;
5042                 }
5043         }
5044
5045         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5046         if (err)
5047                 goto out_close;
5048
5049         err = target__validate(&trace.opts.target);
5050         if (err) {
5051                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5052                 fprintf(trace.output, "%s", bf);
5053                 goto out_close;
5054         }
5055
5056         err = target__parse_uid(&trace.opts.target);
5057         if (err) {
5058                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5059                 fprintf(trace.output, "%s", bf);
5060                 goto out_close;
5061         }
5062
5063         if (!argc && target__none(&trace.opts.target))
5064                 trace.opts.target.system_wide = true;
5065
5066         if (input_name)
5067                 err = trace__replay(&trace);
5068         else
5069                 err = trace__run(&trace, argc, argv);
5070
5071 out_close:
5072         if (output_name != NULL)
5073                 fclose(trace.output);
5074 out:
5075         trace__exit(&trace);
5076 #ifdef HAVE_BPF_SKEL
5077         augmented_raw_syscalls_bpf__destroy(trace.skel);
5078 #endif
5079         return err;
5080 }