tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  */
  16
  17 #include "util/record.h"
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "util/rlimit.h"
  23 #include "builtin.h"
  24 #include "util/cgroup.h"
  25 #include "util/color.h"
  26 #include "util/config.h"
  27 #include "util/debug.h"
  28 #include "util/dso.h"
  29 #include "util/env.h"
  30 #include "util/event.h"
  31 #include "util/evsel.h"
  32 #include "util/evsel_fprintf.h"
  33 #include "util/synthetic-events.h"
  34 #include "util/evlist.h"
  35 #include "util/evswitch.h"
  36 #include "util/mmap.h"
  37 #include <subcmd/pager.h>
  38 #include <subcmd/exec-cmd.h>
  39 #include "util/machine.h"
  40 #include "util/map.h"
  41 #include "util/symbol.h"
  42 #include "util/path.h"
  43 #include "util/session.h"
  44 #include "util/thread.h"
  45 #include <subcmd/parse-options.h>
  46 #include "util/strlist.h"
  47 #include "util/intlist.h"
  48 #include "util/thread_map.h"
  49 #include "util/stat.h"
  50 #include "util/tool.h"
  51 #include "util/util.h"
  52 #include "trace/beauty/beauty.h"
  53 #include "trace-event.h"
  54 #include "util/parse-events.h"
  55 #include "util/bpf-loader.h"
  56 #include "util/tracepoint.h"
  57 #include "callchain.h"
  58 #include "print_binary.h"
  59 #include "string2.h"
  60 #include "syscalltbl.h"
  61 #include "rb_resort.h"
  62 #include "../perf.h"
  63
  64 #include <errno.h>
  65 #include <inttypes.h>
  66 #include <poll.h>
  67 #include <signal.h>
  68 #include <stdlib.h>
  69 #include <string.h>
  70 #include <linux/err.h>
  71 #include <linux/filter.h>
  72 #include <linux/kernel.h>
  73 #include <linux/random.h>
  74 #include <linux/stringify.h>
  75 #include <linux/time64.h>
  76 #include <linux/zalloc.h>
  77 #include <fcntl.h>
  78 #include <sys/sysmacros.h>
  79
  80 #include <linux/ctype.h>
  81 #include <perf/mmap.h>
  82
  83 #ifndef O_CLOEXEC
  84 # define O_CLOEXEC              02000000
  85 #endif
  86
  87 #ifndef F_LINUX_SPECIFIC_BASE
  88 # define F_LINUX_SPECIFIC_BASE  1024
  89 #endif
  90
  91 /*
  92  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  93  */
  94 struct syscall_arg_fmt {
  95         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
  96         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
  97         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
  98         void       *parm;
  99         const char *name;
 100         u16        nr_entries; // for arrays
 101         bool       show_zero;
 102 };
 103
 104 struct syscall_fmt {
 105         const char *name;
 106         const char *alias;
 107         struct {
 108                 const char *sys_enter,
 109                            *sys_exit;
 110         }          bpf_prog_name;
 111         struct syscall_arg_fmt arg[6];
 112         u8         nr_args;
 113         bool       errpid;
 114         bool       timeout;
 115         bool       hexret;
 116 };
 117
 118 struct trace {
 119         struct perf_tool        tool;
 120         struct syscalltbl       *sctbl;
 121         struct {
 122                 struct syscall  *table;
 123                 struct bpf_map  *map;
 124                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 125                         struct bpf_map  *sys_enter,
 126                                         *sys_exit;
 127                 }               prog_array;
 128                 struct {
 129                         struct evsel *sys_enter,
 130                                           *sys_exit,
 131                                           *augmented;
 132                 }               events;
 133                 struct bpf_program *unaugmented_prog;
 134         } syscalls;
 135         struct {
 136                 struct bpf_map *map;
 137         } dump;
 138         struct record_opts      opts;
 139         struct evlist   *evlist;
 140         struct machine          *host;
 141         struct thread           *current;
 142         struct bpf_object       *bpf_obj;
 143         struct cgroup           *cgroup;
 144         u64                     base_time;
 145         FILE                    *output;
 146         unsigned long           nr_events;
 147         unsigned long           nr_events_printed;
 148         unsigned long           max_events;
 149         struct evswitch         evswitch;
 150         struct strlist          *ev_qualifier;
 151         struct {
 152                 size_t          nr;
 153                 int             *entries;
 154         }                       ev_qualifier_ids;
 155         struct {
 156                 size_t          nr;
 157                 pid_t           *entries;
 158                 struct bpf_map  *map;
 159         }                       filter_pids;
 160         double                  duration_filter;
 161         double                  runtime_ms;
 162         struct {
 163                 u64             vfs_getname,
 164                                 proc_getname;
 165         } stats;
 166         unsigned int            max_stack;
 167         unsigned int            min_stack;
 168         int                     raw_augmented_syscalls_args_size;
 169         bool                    raw_augmented_syscalls;
 170         bool                    fd_path_disabled;
 171         bool                    sort_events;
 172         bool                    not_ev_qualifier;
 173         bool                    live;
 174         bool                    full_time;
 175         bool                    sched;
 176         bool                    multiple_threads;
 177         bool                    summary;
 178         bool                    summary_only;
 179         bool                    errno_summary;
 180         bool                    failure_only;
 181         bool                    show_comm;
 182         bool                    print_sample;
 183         bool                    show_tool_stats;
 184         bool                    trace_syscalls;
 185         bool                    libtraceevent_print;
 186         bool                    kernel_syscallchains;
 187         s16                     args_alignment;
 188         bool                    show_tstamp;
 189         bool                    show_duration;
 190         bool                    show_zeros;
 191         bool                    show_arg_names;
 192         bool                    show_string_prefix;
 193         bool                    force;
 194         bool                    vfs_getname;
 195         int                     trace_pgfaults;
 196         char                    *perfconfig_events;
 197         struct {
 198                 struct ordered_events   data;
 199                 u64                     last;
 200         } oe;
 201 };
 202
 203 struct tp_field {
 204         int offset;
 205         union {
 206                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 207                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 208         };
 209 };
 210
 211 #define TP_UINT_FIELD(bits) \
 212 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 213 { \
 214         u##bits value; \
 215         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 216         return value;  \
 217 }
 218
 219 TP_UINT_FIELD(8);
 220 TP_UINT_FIELD(16);
 221 TP_UINT_FIELD(32);
 222 TP_UINT_FIELD(64);
 223
 224 #define TP_UINT_FIELD__SWAPPED(bits) \
 225 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 226 { \
 227         u##bits value; \
 228         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 229         return bswap_##bits(value);\
 230 }
 231
 232 TP_UINT_FIELD__SWAPPED(16);
 233 TP_UINT_FIELD__SWAPPED(32);
 234 TP_UINT_FIELD__SWAPPED(64);
 235
 236 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 237 {
 238         field->offset = offset;
 239
 240         switch (size) {
 241         case 1:
 242                 field->integer = tp_field__u8;
 243                 break;
 244         case 2:
 245                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 246                 break;
 247         case 4:
 248                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 249                 break;
 250         case 8:
 251                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 252                 break;
 253         default:
 254                 return -1;
 255         }
 256
 257         return 0;
 258 }
 259
 260 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 261 {
 262         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 263 }
 264
 265 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 266 {
 267         return sample->raw_data + field->offset;
 268 }
 269
 270 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 271 {
 272         field->offset = offset;
 273         field->pointer = tp_field__ptr;
 274         return 0;
 275 }
 276
 277 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 278 {
 279         return __tp_field__init_ptr(field, format_field->offset);
 280 }
 281
 282 struct syscall_tp {
 283         struct tp_field id;
 284         union {
 285                 struct tp_field args, ret;
 286         };
 287 };
 288
 289 /*
 290  * The evsel->priv as used by 'perf trace'
 291  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 292  * fmt: for all the other tracepoints
 293  */
 294 struct evsel_trace {
 295         struct syscall_tp       sc;
 296         struct syscall_arg_fmt  *fmt;
 297 };
 298
 299 static struct evsel_trace *evsel_trace__new(void)
 300 {
 301         return zalloc(sizeof(struct evsel_trace));
 302 }
 303
 304 static void evsel_trace__delete(struct evsel_trace *et)
 305 {
 306         if (et == NULL)
 307                 return;
 308
 309         zfree(&et->fmt);
 310         free(et);
 311 }
 312
 313 /*
 314  * Used with raw_syscalls:sys_{enter,exit} and with the
 315  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 316  */
 317 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 318 {
 319         struct evsel_trace *et = evsel->priv;
 320
 321         return &et->sc;
 322 }
 323
 324 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 325 {
 326         if (evsel->priv == NULL) {
 327                 evsel->priv = evsel_trace__new();
 328                 if (evsel->priv == NULL)
 329                         return NULL;
 330         }
 331
 332         return __evsel__syscall_tp(evsel);
 333 }
 334
 335 /*
 336  * Used with all the other tracepoints.
 337  */
 338 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 339 {
 340         struct evsel_trace *et = evsel->priv;
 341
 342         return et->fmt;
 343 }
 344
 345 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 346 {
 347         struct evsel_trace *et = evsel->priv;
 348
 349         if (evsel->priv == NULL) {
 350                 et = evsel->priv = evsel_trace__new();
 351
 352                 if (et == NULL)
 353                         return NULL;
 354         }
 355
 356         if (et->fmt == NULL) {
 357                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 358                 if (et->fmt == NULL)
 359                         goto out_delete;
 360         }
 361
 362         return __evsel__syscall_arg_fmt(evsel);
 363
 364 out_delete:
 365         evsel_trace__delete(evsel->priv);
 366         evsel->priv = NULL;
 367         return NULL;
 368 }
 369
 370 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 371 {
 372         struct tep_format_field *format_field = evsel__field(evsel, name);
 373
 374         if (format_field == NULL)
 375                 return -1;
 376
 377         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 378 }
 379
 380 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 381         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 382            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 383
 384 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 385 {
 386         struct tep_format_field *format_field = evsel__field(evsel, name);
 387
 388         if (format_field == NULL)
 389                 return -1;
 390
 391         return tp_field__init_ptr(field, format_field);
 392 }
 393
 394 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 395         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 396            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 397
 398 static void evsel__delete_priv(struct evsel *evsel)
 399 {
 400         zfree(&evsel->priv);
 401         evsel__delete(evsel);
 402 }
 403
 404 static int evsel__init_syscall_tp(struct evsel *evsel)
 405 {
 406         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 407
 408         if (sc != NULL) {
 409                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 410                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 411                         return -ENOENT;
 412                 return 0;
 413         }
 414
 415         return -ENOMEM;
 416 }
 417
 418 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 419 {
 420         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 421
 422         if (sc != NULL) {
 423                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 424                 if (syscall_id == NULL)
 425                         syscall_id = evsel__field(tp, "__syscall_nr");
 426                 if (syscall_id == NULL ||
 427                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 428                         return -EINVAL;
 429
 430                 return 0;
 431         }
 432
 433         return -ENOMEM;
 434 }
 435
 436 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 437 {
 438         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 439
 440         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 441 }
 442
 443 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 444 {
 445         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 446
 447         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 448 }
 449
 450 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 451 {
 452         if (evsel__syscall_tp(evsel) != NULL) {
 453                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 454                         return -ENOENT;
 455
 456                 evsel->handler = handler;
 457                 return 0;
 458         }
 459
 460         return -ENOMEM;
 461 }
 462
 463 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 464 {
 465         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 466
 467         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 468         if (IS_ERR(evsel))
 469                 evsel = evsel__newtp("syscalls", direction);
 470
 471         if (IS_ERR(evsel))
 472                 return NULL;
 473
 474         if (evsel__init_raw_syscall_tp(evsel, handler))
 475                 goto out_delete;
 476
 477         return evsel;
 478
 479 out_delete:
 480         evsel__delete_priv(evsel);
 481         return NULL;
 482 }
 483
 484 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 485         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 486            fields->name.integer(&fields->name, sample); })
 487
 488 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 489         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 490            fields->name.pointer(&fields->name, sample); })
 491
 492 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 493 {
 494         int idx = val - sa->offset;
 495
 496         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 497                 size_t printed = scnprintf(bf, size, intfmt, val);
 498                 if (show_suffix)
 499                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 500                 return printed;
 501         }
 502
 503         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 504 }
 505
 506 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 507 {
 508         int idx = val - sa->offset;
 509
 510         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 511                 size_t printed = scnprintf(bf, size, intfmt, val);
 512                 if (show_prefix)
 513                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 514                 return printed;
 515         }
 516
 517         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 518 }
 519
 520 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 521                                                 const char *intfmt,
 522                                                 struct syscall_arg *arg)
 523 {
 524         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 525 }
 526
 527 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 528                                               struct syscall_arg *arg)
 529 {
 530         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 531 }
 532
 533 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 534
 535 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 536 {
 537         return strarray__strtoul(arg->parm, bf, size, ret);
 538 }
 539
 540 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 541 {
 542         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 543 }
 544
 545 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 546 {
 547         return strarrays__strtoul(arg->parm, bf, size, ret);
 548 }
 549
 550 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 551 {
 552         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 553 }
 554
 555 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 556 {
 557         size_t printed;
 558         int i;
 559
 560         for (i = 0; i < sas->nr_entries; ++i) {
 561                 struct strarray *sa = sas->entries[i];
 562                 int idx = val - sa->offset;
 563
 564                 if (idx >= 0 && idx < sa->nr_entries) {
 565                         if (sa->entries[idx] == NULL)
 566                                 break;
 567                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 568                 }
 569         }
 570
 571         printed = scnprintf(bf, size, intfmt, val);
 572         if (show_prefix)
 573                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 574         return printed;
 575 }
 576
 577 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 578 {
 579         int i;
 580
 581         for (i = 0; i < sa->nr_entries; ++i) {
 582                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 583                         *ret = sa->offset + i;
 584                         return true;
 585                 }
 586         }
 587
 588         return false;
 589 }
 590
 591 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 592 {
 593         u64 val = 0;
 594         char *tok = bf, *sep, *end;
 595
 596         *ret = 0;
 597
 598         while (size != 0) {
 599                 int toklen = size;
 600
 601                 sep = memchr(tok, '|', size);
 602                 if (sep != NULL) {
 603                         size -= sep - tok + 1;
 604
 605                         end = sep - 1;
 606                         while (end > tok && isspace(*end))
 607                                 --end;
 608
 609                         toklen = end - tok + 1;
 610                 }
 611
 612                 while (isspace(*tok))
 613                         ++tok;
 614
 615                 if (isalpha(*tok) || *tok == '_') {
 616                         if (!strarray__strtoul(sa, tok, toklen, &val))
 617                                 return false;
 618                 } else
 619                         val = strtoul(tok, NULL, 0);
 620
 621                 *ret |= (1 << (val - 1));
 622
 623                 if (sep == NULL)
 624                         break;
 625                 tok = sep + 1;
 626         }
 627
 628         return true;
 629 }
 630
 631 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 632 {
 633         int i;
 634
 635         for (i = 0; i < sas->nr_entries; ++i) {
 636                 struct strarray *sa = sas->entries[i];
 637
 638                 if (strarray__strtoul(sa, bf, size, ret))
 639                         return true;
 640         }
 641
 642         return false;
 643 }
 644
 645 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 646                                         struct syscall_arg *arg)
 647 {
 648         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 649 }
 650
 651 #ifndef AT_FDCWD
 652 #define AT_FDCWD        -100
 653 #endif
 654
 655 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 656                                            struct syscall_arg *arg)
 657 {
 658         int fd = arg->val;
 659         const char *prefix = "AT_FD";
 660
 661         if (fd == AT_FDCWD)
 662                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 663
 664         return syscall_arg__scnprintf_fd(bf, size, arg);
 665 }
 666
 667 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 668
 669 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 670                                               struct syscall_arg *arg);
 671
 672 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 673
 674 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 675 {
 676         return scnprintf(bf, size, "%#lx", arg->val);
 677 }
 678
 679 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 680 {
 681         if (arg->val == 0)
 682                 return scnprintf(bf, size, "NULL");
 683         return syscall_arg__scnprintf_hex(bf, size, arg);
 684 }
 685
 686 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 687 {
 688         return scnprintf(bf, size, "%d", arg->val);
 689 }
 690
 691 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 692 {
 693         return scnprintf(bf, size, "%ld", arg->val);
 694 }
 695
 696 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 697 {
 698         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 699         //     fill missing comms using thread__set_comm()...
 700         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 701         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 702 }
 703
 704 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 705
 706 static const char *bpf_cmd[] = {
 707         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 708         "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 709         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 710         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 711         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 712         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 713         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 714         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 715         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 716         "LINK_DETACH", "PROG_BIND_MAP",
 717 };
 718 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 719
 720 static const char *fsmount_flags[] = {
 721         [1] = "CLOEXEC",
 722 };
 723 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 724
 725 #include "trace/beauty/generated/fsconfig_arrays.c"
 726
 727 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 728
 729 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 730 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 731
 732 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 733 static DEFINE_STRARRAY(itimers, "ITIMER_");
 734
 735 static const char *keyctl_options[] = {
 736         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 737         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 738         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 739         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 740         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 741 };
 742 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 743
 744 static const char *whences[] = { "SET", "CUR", "END",
 745 #ifdef SEEK_DATA
 746 "DATA",
 747 #endif
 748 #ifdef SEEK_HOLE
 749 "HOLE",
 750 #endif
 751 };
 752 static DEFINE_STRARRAY(whences, "SEEK_");
 753
 754 static const char *fcntl_cmds[] = {
 755         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 756         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 757         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 758         "GETOWNER_UIDS",
 759 };
 760 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 761
 762 static const char *fcntl_linux_specific_cmds[] = {
 763         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 764         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 765         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 766 };
 767
 768 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 769
 770 static struct strarray *fcntl_cmds_arrays[] = {
 771         &strarray__fcntl_cmds,
 772         &strarray__fcntl_linux_specific_cmds,
 773 };
 774
 775 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 776
 777 static const char *rlimit_resources[] = {
 778         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 779         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 780         "RTTIME",
 781 };
 782 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 783
 784 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 785 static DEFINE_STRARRAY(sighow, "SIG_");
 786
 787 static const char *clockid[] = {
 788         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 789         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 790         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 791 };
 792 static DEFINE_STRARRAY(clockid, "CLOCK_");
 793
 794 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 795                                                  struct syscall_arg *arg)
 796 {
 797         bool show_prefix = arg->show_string_prefix;
 798         const char *suffix = "_OK";
 799         size_t printed = 0;
 800         int mode = arg->val;
 801
 802         if (mode == F_OK) /* 0 */
 803                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 804 #define P_MODE(n) \
 805         if (mode & n##_OK) { \
 806                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 807                 mode &= ~n##_OK; \
 808         }
 809
 810         P_MODE(R);
 811         P_MODE(W);
 812         P_MODE(X);
 813 #undef P_MODE
 814
 815         if (mode)
 816                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 817
 818         return printed;
 819 }
 820
 821 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 822
 823 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 824                                               struct syscall_arg *arg);
 825
 826 #define SCA_FILENAME syscall_arg__scnprintf_filename
 827
 828 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 829                                                 struct syscall_arg *arg)
 830 {
 831         bool show_prefix = arg->show_string_prefix;
 832         const char *prefix = "O_";
 833         int printed = 0, flags = arg->val;
 834
 835 #define P_FLAG(n) \
 836         if (flags & O_##n) { \
 837                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 838                 flags &= ~O_##n; \
 839         }
 840
 841         P_FLAG(CLOEXEC);
 842         P_FLAG(NONBLOCK);
 843 #undef P_FLAG
 844
 845         if (flags)
 846                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 847
 848         return printed;
 849 }
 850
 851 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 852
 853 #ifndef GRND_NONBLOCK
 854 #define GRND_NONBLOCK   0x0001
 855 #endif
 856 #ifndef GRND_RANDOM
 857 #define GRND_RANDOM     0x0002
 858 #endif
 859
 860 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 861                                                    struct syscall_arg *arg)
 862 {
 863         bool show_prefix = arg->show_string_prefix;
 864         const char *prefix = "GRND_";
 865         int printed = 0, flags = arg->val;
 866
 867 #define P_FLAG(n) \
 868         if (flags & GRND_##n) { \
 869                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 870                 flags &= ~GRND_##n; \
 871         }
 872
 873         P_FLAG(RANDOM);
 874         P_FLAG(NONBLOCK);
 875 #undef P_FLAG
 876
 877         if (flags)
 878                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 879
 880         return printed;
 881 }
 882
 883 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 884
 885 #define STRARRAY(name, array) \
 886           { .scnprintf  = SCA_STRARRAY, \
 887             .strtoul    = STUL_STRARRAY, \
 888             .parm       = &strarray__##array, }
 889
 890 #define STRARRAY_FLAGS(name, array) \
 891           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 892             .strtoul    = STUL_STRARRAY_FLAGS, \
 893             .parm       = &strarray__##array, }
 894
 895 #include "trace/beauty/arch_errno_names.c"
 896 #include "trace/beauty/eventfd.c"
 897 #include "trace/beauty/futex_op.c"
 898 #include "trace/beauty/futex_val3.c"
 899 #include "trace/beauty/mmap.c"
 900 #include "trace/beauty/mode_t.c"
 901 #include "trace/beauty/msg_flags.c"
 902 #include "trace/beauty/open_flags.c"
 903 #include "trace/beauty/perf_event_open.c"
 904 #include "trace/beauty/pid.c"
 905 #include "trace/beauty/sched_policy.c"
 906 #include "trace/beauty/seccomp.c"
 907 #include "trace/beauty/signum.c"
 908 #include "trace/beauty/socket_type.c"
 909 #include "trace/beauty/waitid_options.c"
 910
 911 static struct syscall_fmt syscall_fmts[] = {
 912         { .name     = "access",
 913           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 914         { .name     = "arch_prctl",
 915           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 916                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 917         { .name     = "bind",
 918           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 919                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 920                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 921         { .name     = "bpf",
 922           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 923         { .name     = "brk",        .hexret = true,
 924           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 925         { .name     = "clock_gettime",
 926           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 927         { .name     = "clone",      .errpid = true, .nr_args = 5,
 928           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 929                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 930                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 931                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 932                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 933         { .name     = "close",
 934           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 935         { .name     = "connect",
 936           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 937                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 938                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 939         { .name     = "epoll_ctl",
 940           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 941         { .name     = "eventfd2",
 942           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 943         { .name     = "fchmodat",
 944           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 945         { .name     = "fchownat",
 946           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 947         { .name     = "fcntl",
 948           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 949                            .strtoul   = STUL_STRARRAYS,
 950                            .parm      = &strarrays__fcntl_cmds_arrays,
 951                            .show_zero = true, },
 952                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 953         { .name     = "flock",
 954           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 955         { .name     = "fsconfig",
 956           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 957         { .name     = "fsmount",
 958           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 959                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 960         { .name     = "fspick",
 961           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 962                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 963                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 964         { .name     = "fstat", .alias = "newfstat", },
 965         { .name     = "fstatat", .alias = "newfstatat", },
 966         { .name     = "futex",
 967           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 968                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 969         { .name     = "futimesat",
 970           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 971         { .name     = "getitimer",
 972           .arg = { [0] = STRARRAY(which, itimers), }, },
 973         { .name     = "getpid",     .errpid = true, },
 974         { .name     = "getpgid",    .errpid = true, },
 975         { .name     = "getppid",    .errpid = true, },
 976         { .name     = "getrandom",
 977           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 978         { .name     = "getrlimit",
 979           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 980         { .name     = "getsockopt",
 981           .arg = { [1] = STRARRAY(level, socket_level), }, },
 982         { .name     = "gettid",     .errpid = true, },
 983         { .name     = "ioctl",
 984           .arg = {
 985 #if defined(__i386__) || defined(__x86_64__)
 986 /*
 987  * FIXME: Make this available to all arches.
 988  */
 989                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 990                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 991 #else
 992                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 993 #endif
 994         { .name     = "kcmp",       .nr_args = 5,
 995           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 996                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 997                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 998                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 999                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1000         { .name     = "keyctl",
1001           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1002         { .name     = "kill",
1003           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1004         { .name     = "linkat",
1005           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1006         { .name     = "lseek",
1007           .arg = { [2] = STRARRAY(whence, whences), }, },
1008         { .name     = "lstat", .alias = "newlstat", },
1009         { .name     = "madvise",
1010           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1011                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1012         { .name     = "mkdirat",
1013           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1014         { .name     = "mknodat",
1015           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1016         { .name     = "mmap",       .hexret = true,
1017 /* The standard mmap maps to old_mmap on s390x */
1018 #if defined(__s390x__)
1019         .alias = "old_mmap",
1020 #endif
1021           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1022                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1023                            .strtoul   = STUL_STRARRAY_FLAGS,
1024                            .parm      = &strarray__mmap_flags, },
1025                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1026         { .name     = "mount",
1027           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1028                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1029                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1030         { .name     = "move_mount",
1031           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1032                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1033                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1034                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1035                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1036         { .name     = "mprotect",
1037           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1038                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1039         { .name     = "mq_unlink",
1040           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1041         { .name     = "mremap",     .hexret = true,
1042           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1043         { .name     = "name_to_handle_at",
1044           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1045         { .name     = "newfstatat",
1046           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1047         { .name     = "open",
1048           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1049         { .name     = "open_by_handle_at",
1050           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1051                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1052         { .name     = "openat",
1053           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1054                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1055         { .name     = "perf_event_open",
1056           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1057                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1058                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1059         { .name     = "pipe2",
1060           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1061         { .name     = "pkey_alloc",
1062           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1063         { .name     = "pkey_free",
1064           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1065         { .name     = "pkey_mprotect",
1066           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1067                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1068                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1069         { .name     = "poll", .timeout = true, },
1070         { .name     = "ppoll", .timeout = true, },
1071         { .name     = "prctl",
1072           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1073                            .strtoul   = STUL_STRARRAY,
1074                            .parm      = &strarray__prctl_options, },
1075                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1076                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1077         { .name     = "pread", .alias = "pread64", },
1078         { .name     = "preadv", .alias = "pread", },
1079         { .name     = "prlimit64",
1080           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1081         { .name     = "pwrite", .alias = "pwrite64", },
1082         { .name     = "readlinkat",
1083           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1084         { .name     = "recvfrom",
1085           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1086         { .name     = "recvmmsg",
1087           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1088         { .name     = "recvmsg",
1089           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1090         { .name     = "renameat",
1091           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1092                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1093         { .name     = "renameat2",
1094           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1095                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1096                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1097         { .name     = "rt_sigaction",
1098           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1099         { .name     = "rt_sigprocmask",
1100           .arg = { [0] = STRARRAY(how, sighow), }, },
1101         { .name     = "rt_sigqueueinfo",
1102           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1103         { .name     = "rt_tgsigqueueinfo",
1104           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1105         { .name     = "sched_setscheduler",
1106           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1107         { .name     = "seccomp",
1108           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1109                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1110         { .name     = "select", .timeout = true, },
1111         { .name     = "sendfile", .alias = "sendfile64", },
1112         { .name     = "sendmmsg",
1113           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1114         { .name     = "sendmsg",
1115           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1116         { .name     = "sendto",
1117           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1118                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1119         { .name     = "set_tid_address", .errpid = true, },
1120         { .name     = "setitimer",
1121           .arg = { [0] = STRARRAY(which, itimers), }, },
1122         { .name     = "setrlimit",
1123           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1124         { .name     = "setsockopt",
1125           .arg = { [1] = STRARRAY(level, socket_level), }, },
1126         { .name     = "socket",
1127           .arg = { [0] = STRARRAY(family, socket_families),
1128                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1129                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1130         { .name     = "socketpair",
1131           .arg = { [0] = STRARRAY(family, socket_families),
1132                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1133                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1134         { .name     = "stat", .alias = "newstat", },
1135         { .name     = "statx",
1136           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1137                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1138                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1139         { .name     = "swapoff",
1140           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1141         { .name     = "swapon",
1142           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1143         { .name     = "symlinkat",
1144           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1145         { .name     = "sync_file_range",
1146           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1147         { .name     = "tgkill",
1148           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1149         { .name     = "tkill",
1150           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1151         { .name     = "umount2", .alias = "umount",
1152           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1153         { .name     = "uname", .alias = "newuname", },
1154         { .name     = "unlinkat",
1155           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1156         { .name     = "utimensat",
1157           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1158         { .name     = "wait4",      .errpid = true,
1159           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1160         { .name     = "waitid",     .errpid = true,
1161           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1162 };
1163
1164 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1165 {
1166         const struct syscall_fmt *fmt = fmtp;
1167         return strcmp(name, fmt->name);
1168 }
1169
1170 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1171 {
1172         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1173 }
1174
1175 static struct syscall_fmt *syscall_fmt__find(const char *name)
1176 {
1177         const int nmemb = ARRAY_SIZE(syscall_fmts);
1178         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1179 }
1180
1181 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1182 {
1183         int i;
1184
1185         for (i = 0; i < nmemb; ++i) {
1186                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1187                         return &fmts[i];
1188         }
1189
1190         return NULL;
1191 }
1192
1193 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1194 {
1195         const int nmemb = ARRAY_SIZE(syscall_fmts);
1196         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1197 }
1198
1199 /*
1200  * is_exit: is this "exit" or "exit_group"?
1201  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1202  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1203  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1204  */
1205 struct syscall {
1206         struct tep_event    *tp_format;
1207         int                 nr_args;
1208         int                 args_size;
1209         struct {
1210                 struct bpf_program *sys_enter,
1211                                    *sys_exit;
1212         }                   bpf_prog;
1213         bool                is_exit;
1214         bool                is_open;
1215         bool                nonexistent;
1216         struct tep_format_field *args;
1217         const char          *name;
1218         struct syscall_fmt  *fmt;
1219         struct syscall_arg_fmt *arg_fmt;
1220 };
1221
1222 /*
1223  * Must match what is in the BPF program:
1224  *
1225  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1226  */
1227 struct bpf_map_syscall_entry {
1228         bool    enabled;
1229         u16     string_args_len[6];
1230 };
1231
1232 /*
1233  * We need to have this 'calculated' boolean because in some cases we really
1234  * don't know what is the duration of a syscall, for instance, when we start
1235  * a session and some threads are waiting for a syscall to finish, say 'poll',
1236  * in which case all we can do is to print "( ? ) for duration and for the
1237  * start timestamp.
1238  */
1239 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1240 {
1241         double duration = (double)t / NSEC_PER_MSEC;
1242         size_t printed = fprintf(fp, "(");
1243
1244         if (!calculated)
1245                 printed += fprintf(fp, "         ");
1246         else if (duration >= 1.0)
1247                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1248         else if (duration >= 0.01)
1249                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1250         else
1251                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1252         return printed + fprintf(fp, "): ");
1253 }
1254
1255 /**
1256  * filename.ptr: The filename char pointer that will be vfs_getname'd
1257  * filename.entry_str_pos: Where to insert the string translated from
1258  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1259  * ret_scnprintf: syscall args may set this to a different syscall return
1260  *                formatter, for instance, fcntl may return fds, file flags, etc.
1261  */
1262 struct thread_trace {
1263         u64               entry_time;
1264         bool              entry_pending;
1265         unsigned long     nr_events;
1266         unsigned long     pfmaj, pfmin;
1267         char              *entry_str;
1268         double            runtime_ms;
1269         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1270         struct {
1271                 unsigned long ptr;
1272                 short int     entry_str_pos;
1273                 bool          pending_open;
1274                 unsigned int  namelen;
1275                 char          *name;
1276         } filename;
1277         struct {
1278                 int           max;
1279                 struct file   *table;
1280         } files;
1281
1282         struct intlist *syscall_stats;
1283 };
1284
1285 static struct thread_trace *thread_trace__new(void)
1286 {
1287         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1288
1289         if (ttrace) {
1290                 ttrace->files.max = -1;
1291                 ttrace->syscall_stats = intlist__new(NULL);
1292         }
1293
1294         return ttrace;
1295 }
1296
1297 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1298 {
1299         struct thread_trace *ttrace;
1300
1301         if (thread == NULL)
1302                 goto fail;
1303
1304         if (thread__priv(thread) == NULL)
1305                 thread__set_priv(thread, thread_trace__new());
1306
1307         if (thread__priv(thread) == NULL)
1308                 goto fail;
1309
1310         ttrace = thread__priv(thread);
1311         ++ttrace->nr_events;
1312
1313         return ttrace;
1314 fail:
1315         color_fprintf(fp, PERF_COLOR_RED,
1316                       "WARNING: not enough memory, dropping samples!\n");
1317         return NULL;
1318 }
1319
1320
1321 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1322                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1323 {
1324         struct thread_trace *ttrace = thread__priv(arg->thread);
1325
1326         ttrace->ret_scnprintf = ret_scnprintf;
1327 }
1328
1329 #define TRACE_PFMAJ             (1 << 0)
1330 #define TRACE_PFMIN             (1 << 1)
1331
1332 static const size_t trace__entry_str_size = 2048;
1333
1334 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1335 {
1336         if (fd < 0)
1337                 return NULL;
1338
1339         if (fd > ttrace->files.max) {
1340                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1341
1342                 if (nfiles == NULL)
1343                         return NULL;
1344
1345                 if (ttrace->files.max != -1) {
1346                         memset(nfiles + ttrace->files.max + 1, 0,
1347                                (fd - ttrace->files.max) * sizeof(struct file));
1348                 } else {
1349                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1350                 }
1351
1352                 ttrace->files.table = nfiles;
1353                 ttrace->files.max   = fd;
1354         }
1355
1356         return ttrace->files.table + fd;
1357 }
1358
1359 struct file *thread__files_entry(struct thread *thread, int fd)
1360 {
1361         return thread_trace__files_entry(thread__priv(thread), fd);
1362 }
1363
1364 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1365 {
1366         struct thread_trace *ttrace = thread__priv(thread);
1367         struct file *file = thread_trace__files_entry(ttrace, fd);
1368
1369         if (file != NULL) {
1370                 struct stat st;
1371                 if (stat(pathname, &st) == 0)
1372                         file->dev_maj = major(st.st_rdev);
1373                 file->pathname = strdup(pathname);
1374                 if (file->pathname)
1375                         return 0;
1376         }
1377
1378         return -1;
1379 }
1380
1381 static int thread__read_fd_path(struct thread *thread, int fd)
1382 {
1383         char linkname[PATH_MAX], pathname[PATH_MAX];
1384         struct stat st;
1385         int ret;
1386
1387         if (thread->pid_ == thread->tid) {
1388                 scnprintf(linkname, sizeof(linkname),
1389                           "/proc/%d/fd/%d", thread->pid_, fd);
1390         } else {
1391                 scnprintf(linkname, sizeof(linkname),
1392                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1393         }
1394
1395         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1396                 return -1;
1397
1398         ret = readlink(linkname, pathname, sizeof(pathname));
1399
1400         if (ret < 0 || ret > st.st_size)
1401                 return -1;
1402
1403         pathname[ret] = '\0';
1404         return trace__set_fd_pathname(thread, fd, pathname);
1405 }
1406
1407 static const char *thread__fd_path(struct thread *thread, int fd,
1408                                    struct trace *trace)
1409 {
1410         struct thread_trace *ttrace = thread__priv(thread);
1411
1412         if (ttrace == NULL || trace->fd_path_disabled)
1413                 return NULL;
1414
1415         if (fd < 0)
1416                 return NULL;
1417
1418         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1419                 if (!trace->live)
1420                         return NULL;
1421                 ++trace->stats.proc_getname;
1422                 if (thread__read_fd_path(thread, fd))
1423                         return NULL;
1424         }
1425
1426         return ttrace->files.table[fd].pathname;
1427 }
1428
1429 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1430 {
1431         int fd = arg->val;
1432         size_t printed = scnprintf(bf, size, "%d", fd);
1433         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1434
1435         if (path)
1436                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1437
1438         return printed;
1439 }
1440
1441 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1442 {
1443         size_t printed = scnprintf(bf, size, "%d", fd);
1444         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1445
1446         if (thread) {
1447                 const char *path = thread__fd_path(thread, fd, trace);
1448
1449                 if (path)
1450                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1451
1452                 thread__put(thread);
1453         }
1454
1455         return printed;
1456 }
1457
1458 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1459                                               struct syscall_arg *arg)
1460 {
1461         int fd = arg->val;
1462         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1463         struct thread_trace *ttrace = thread__priv(arg->thread);
1464
1465         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1466                 zfree(&ttrace->files.table[fd].pathname);
1467
1468         return printed;
1469 }
1470
1471 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1472                                      unsigned long ptr)
1473 {
1474         struct thread_trace *ttrace = thread__priv(thread);
1475
1476         ttrace->filename.ptr = ptr;
1477         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1478 }
1479
1480 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1481 {
1482         struct augmented_arg *augmented_arg = arg->augmented.args;
1483         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1484         /*
1485          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1486          * we would have two strings, each prefixed by its size.
1487          */
1488         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1489
1490         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1491         arg->augmented.size -= consumed;
1492
1493         return printed;
1494 }
1495
1496 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1497                                               struct syscall_arg *arg)
1498 {
1499         unsigned long ptr = arg->val;
1500
1501         if (arg->augmented.args)
1502                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1503
1504         if (!arg->trace->vfs_getname)
1505                 return scnprintf(bf, size, "%#x", ptr);
1506
1507         thread__set_filename_pos(arg->thread, bf, ptr);
1508         return 0;
1509 }
1510
1511 static bool trace__filter_duration(struct trace *trace, double t)
1512 {
1513         return t < (trace->duration_filter * NSEC_PER_MSEC);
1514 }
1515
1516 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1517 {
1518         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1519
1520         return fprintf(fp, "%10.3f ", ts);
1521 }
1522
1523 /*
1524  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1525  * using ttrace->entry_time for a thread that receives a sys_exit without
1526  * first having received a sys_enter ("poll" issued before tracing session
1527  * starts, lost sys_enter exit due to ring buffer overflow).
1528  */
1529 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1530 {
1531         if (tstamp > 0)
1532                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1533
1534         return fprintf(fp, "         ? ");
1535 }
1536
1537 static pid_t workload_pid = -1;
1538 static bool done = false;
1539 static bool interrupted = false;
1540
1541 static void sighandler_interrupt(int sig __maybe_unused)
1542 {
1543         done = interrupted = true;
1544 }
1545
1546 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1547                             void *context __maybe_unused)
1548 {
1549         if (info->si_pid == workload_pid)
1550                 done = true;
1551 }
1552
1553 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1554 {
1555         size_t printed = 0;
1556
1557         if (trace->multiple_threads) {
1558                 if (trace->show_comm)
1559                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1560                 printed += fprintf(fp, "%d ", thread->tid);
1561         }
1562
1563         return printed;
1564 }
1565
1566 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1567                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1568 {
1569         size_t printed = 0;
1570
1571         if (trace->show_tstamp)
1572                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1573         if (trace->show_duration)
1574                 printed += fprintf_duration(duration, duration_calculated, fp);
1575         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1576 }
1577
1578 static int trace__process_event(struct trace *trace, struct machine *machine,
1579                                 union perf_event *event, struct perf_sample *sample)
1580 {
1581         int ret = 0;
1582
1583         switch (event->header.type) {
1584         case PERF_RECORD_LOST:
1585                 color_fprintf(trace->output, PERF_COLOR_RED,
1586                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1587                 ret = machine__process_lost_event(machine, event, sample);
1588                 break;
1589         default:
1590                 ret = machine__process_event(machine, event, sample);
1591                 break;
1592         }
1593
1594         return ret;
1595 }
1596
1597 static int trace__tool_process(struct perf_tool *tool,
1598                                union perf_event *event,
1599                                struct perf_sample *sample,
1600                                struct machine *machine)
1601 {
1602         struct trace *trace = container_of(tool, struct trace, tool);
1603         return trace__process_event(trace, machine, event, sample);
1604 }
1605
1606 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1607 {
1608         struct machine *machine = vmachine;
1609
1610         if (machine->kptr_restrict_warned)
1611                 return NULL;
1612
1613         if (symbol_conf.kptr_restrict) {
1614                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1615                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1616                            "Kernel samples will not be resolved.\n");
1617                 machine->kptr_restrict_warned = true;
1618                 return NULL;
1619         }
1620
1621         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1622 }
1623
1624 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1625 {
1626         int err = symbol__init(NULL);
1627
1628         if (err)
1629                 return err;
1630
1631         trace->host = machine__new_host();
1632         if (trace->host == NULL)
1633                 return -ENOMEM;
1634
1635         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1636         if (err < 0)
1637                 goto out;
1638
1639         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1640                                             evlist->core.threads, trace__tool_process,
1641                                             true, false, 1);
1642 out:
1643         if (err)
1644                 symbol__exit();
1645
1646         return err;
1647 }
1648
1649 static void trace__symbols__exit(struct trace *trace)
1650 {
1651         machine__exit(trace->host);
1652         trace->host = NULL;
1653
1654         symbol__exit();
1655 }
1656
1657 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1658 {
1659         int idx;
1660
1661         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1662                 nr_args = sc->fmt->nr_args;
1663
1664         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1665         if (sc->arg_fmt == NULL)
1666                 return -1;
1667
1668         for (idx = 0; idx < nr_args; ++idx) {
1669                 if (sc->fmt)
1670                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1671         }
1672
1673         sc->nr_args = nr_args;
1674         return 0;
1675 }
1676
1677 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1678         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1679         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1680 };
1681
1682 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1683 {
1684        const struct syscall_arg_fmt *fmt = fmtp;
1685        return strcmp(name, fmt->name);
1686 }
1687
1688 static struct syscall_arg_fmt *
1689 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1690 {
1691        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1692 }
1693
1694 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1695 {
1696        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1697        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1698 }
1699
1700 static struct tep_format_field *
1701 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1702 {
1703         struct tep_format_field *last_field = NULL;
1704         int len;
1705
1706         for (; field; field = field->next, ++arg) {
1707                 last_field = field;
1708
1709                 if (arg->scnprintf)
1710                         continue;
1711
1712                 len = strlen(field->name);
1713
1714                 if (strcmp(field->type, "const char *") == 0 &&
1715                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1716                      strstr(field->name, "path") != NULL))
1717                         arg->scnprintf = SCA_FILENAME;
1718                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1719                         arg->scnprintf = SCA_PTR;
1720                 else if (strcmp(field->type, "pid_t") == 0)
1721                         arg->scnprintf = SCA_PID;
1722                 else if (strcmp(field->type, "umode_t") == 0)
1723                         arg->scnprintf = SCA_MODE_T;
1724                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1725                         arg->scnprintf = SCA_CHAR_ARRAY;
1726                         arg->nr_entries = field->arraylen;
1727                 } else if ((strcmp(field->type, "int") == 0 ||
1728                           strcmp(field->type, "unsigned int") == 0 ||
1729                           strcmp(field->type, "long") == 0) &&
1730                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1731                         /*
1732                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1733                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1734                          * 65 int
1735                          * 23 unsigned int
1736                          * 7 unsigned long
1737                          */
1738                         arg->scnprintf = SCA_FD;
1739                } else {
1740                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1741
1742                         if (fmt) {
1743                                 arg->scnprintf = fmt->scnprintf;
1744                                 arg->strtoul   = fmt->strtoul;
1745                         }
1746                 }
1747         }
1748
1749         return last_field;
1750 }
1751
1752 static int syscall__set_arg_fmts(struct syscall *sc)
1753 {
1754         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1755
1756         if (last_field)
1757                 sc->args_size = last_field->offset + last_field->size;
1758
1759         return 0;
1760 }
1761
1762 static int trace__read_syscall_info(struct trace *trace, int id)
1763 {
1764         char tp_name[128];
1765         struct syscall *sc;
1766         const char *name = syscalltbl__name(trace->sctbl, id);
1767
1768 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1769         if (trace->syscalls.table == NULL) {
1770                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1771                 if (trace->syscalls.table == NULL)
1772                         return -ENOMEM;
1773         }
1774 #else
1775         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1776                 // When using libaudit we don't know beforehand what is the max syscall id
1777                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1778
1779                 if (table == NULL)
1780                         return -ENOMEM;
1781
1782                 // Need to memset from offset 0 and +1 members if brand new
1783                 if (trace->syscalls.table == NULL)
1784                         memset(table, 0, (id + 1) * sizeof(*sc));
1785                 else
1786                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1787
1788                 trace->syscalls.table         = table;
1789                 trace->sctbl->syscalls.max_id = id;
1790         }
1791 #endif
1792         sc = trace->syscalls.table + id;
1793         if (sc->nonexistent)
1794                 return 0;
1795
1796         if (name == NULL) {
1797                 sc->nonexistent = true;
1798                 return 0;
1799         }
1800
1801         sc->name = name;
1802         sc->fmt  = syscall_fmt__find(sc->name);
1803
1804         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1805         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1806
1807         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1808                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1809                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1810         }
1811
1812         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1813                 return -ENOMEM;
1814
1815         if (IS_ERR(sc->tp_format))
1816                 return PTR_ERR(sc->tp_format);
1817
1818         sc->args = sc->tp_format->format.fields;
1819         /*
1820          * We need to check and discard the first variable '__syscall_nr'
1821          * or 'nr' that mean the syscall number. It is needless here.
1822          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1823          */
1824         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1825                 sc->args = sc->args->next;
1826                 --sc->nr_args;
1827         }
1828
1829         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1830         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1831
1832         return syscall__set_arg_fmts(sc);
1833 }
1834
1835 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1836 {
1837         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1838
1839         if (fmt != NULL) {
1840                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1841                 return 0;
1842         }
1843
1844         return -ENOMEM;
1845 }
1846
1847 static int intcmp(const void *a, const void *b)
1848 {
1849         const int *one = a, *another = b;
1850
1851         return *one - *another;
1852 }
1853
1854 static int trace__validate_ev_qualifier(struct trace *trace)
1855 {
1856         int err = 0;
1857         bool printed_invalid_prefix = false;
1858         struct str_node *pos;
1859         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1860
1861         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1862                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1863
1864         if (trace->ev_qualifier_ids.entries == NULL) {
1865                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1866                        trace->output);
1867                 err = -EINVAL;
1868                 goto out;
1869         }
1870
1871         strlist__for_each_entry(pos, trace->ev_qualifier) {
1872                 const char *sc = pos->s;
1873                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1874
1875                 if (id < 0) {
1876                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1877                         if (id >= 0)
1878                                 goto matches;
1879
1880                         if (!printed_invalid_prefix) {
1881                                 pr_debug("Skipping unknown syscalls: ");
1882                                 printed_invalid_prefix = true;
1883                         } else {
1884                                 pr_debug(", ");
1885                         }
1886
1887                         pr_debug("%s", sc);
1888                         continue;
1889                 }
1890 matches:
1891                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1892                 if (match_next == -1)
1893                         continue;
1894
1895                 while (1) {
1896                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1897                         if (id < 0)
1898                                 break;
1899                         if (nr_allocated == nr_used) {
1900                                 void *entries;
1901
1902                                 nr_allocated += 8;
1903                                 entries = realloc(trace->ev_qualifier_ids.entries,
1904                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1905                                 if (entries == NULL) {
1906                                         err = -ENOMEM;
1907                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1908                                         goto out_free;
1909                                 }
1910                                 trace->ev_qualifier_ids.entries = entries;
1911                         }
1912                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1913                 }
1914         }
1915
1916         trace->ev_qualifier_ids.nr = nr_used;
1917         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1918 out:
1919         if (printed_invalid_prefix)
1920                 pr_debug("\n");
1921         return err;
1922 out_free:
1923         zfree(&trace->ev_qualifier_ids.entries);
1924         trace->ev_qualifier_ids.nr = 0;
1925         goto out;
1926 }
1927
1928 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1929 {
1930         bool in_ev_qualifier;
1931
1932         if (trace->ev_qualifier_ids.nr == 0)
1933                 return true;
1934
1935         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1936                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1937
1938         if (in_ev_qualifier)
1939                return !trace->not_ev_qualifier;
1940
1941         return trace->not_ev_qualifier;
1942 }
1943
1944 /*
1945  * args is to be interpreted as a series of longs but we need to handle
1946  * 8-byte unaligned accesses. args points to raw_data within the event
1947  * and raw_data is guaranteed to be 8-byte unaligned because it is
1948  * preceded by raw_size which is a u32. So we need to copy args to a temp
1949  * variable to read it. Most notably this avoids extended load instructions
1950  * on unaligned addresses
1951  */
1952 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1953 {
1954         unsigned long val;
1955         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1956
1957         memcpy(&val, p, sizeof(val));
1958         return val;
1959 }
1960
1961 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1962                                       struct syscall_arg *arg)
1963 {
1964         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1965                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1966
1967         return scnprintf(bf, size, "arg%d: ", arg->idx);
1968 }
1969
1970 /*
1971  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1972  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1973  * in tools/perf/trace/beauty/mount_flags.c
1974  */
1975 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1976 {
1977         if (fmt && fmt->mask_val)
1978                 return fmt->mask_val(arg, val);
1979
1980         return val;
1981 }
1982
1983 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1984                                              struct syscall_arg *arg, unsigned long val)
1985 {
1986         if (fmt && fmt->scnprintf) {
1987                 arg->val = val;
1988                 if (fmt->parm)
1989                         arg->parm = fmt->parm;
1990                 return fmt->scnprintf(bf, size, arg);
1991         }
1992         return scnprintf(bf, size, "%ld", val);
1993 }
1994
1995 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1996                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1997                                       struct trace *trace, struct thread *thread)
1998 {
1999         size_t printed = 0;
2000         unsigned long val;
2001         u8 bit = 1;
2002         struct syscall_arg arg = {
2003                 .args   = args,
2004                 .augmented = {
2005                         .size = augmented_args_size,
2006                         .args = augmented_args,
2007                 },
2008                 .idx    = 0,
2009                 .mask   = 0,
2010                 .trace  = trace,
2011                 .thread = thread,
2012                 .show_string_prefix = trace->show_string_prefix,
2013         };
2014         struct thread_trace *ttrace = thread__priv(thread);
2015
2016         /*
2017          * Things like fcntl will set this in its 'cmd' formatter to pick the
2018          * right formatter for the return value (an fd? file flags?), which is
2019          * not needed for syscalls that always return a given type, say an fd.
2020          */
2021         ttrace->ret_scnprintf = NULL;
2022
2023         if (sc->args != NULL) {
2024                 struct tep_format_field *field;
2025
2026                 for (field = sc->args; field;
2027                      field = field->next, ++arg.idx, bit <<= 1) {
2028                         if (arg.mask & bit)
2029                                 continue;
2030
2031                         arg.fmt = &sc->arg_fmt[arg.idx];
2032                         val = syscall_arg__val(&arg, arg.idx);
2033                         /*
2034                          * Some syscall args need some mask, most don't and
2035                          * return val untouched.
2036                          */
2037                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2038
2039                         /*
2040                          * Suppress this argument if its value is zero and
2041                          * and we don't have a string associated in an
2042                          * strarray for it.
2043                          */
2044                         if (val == 0 &&
2045                             !trace->show_zeros &&
2046                             !(sc->arg_fmt &&
2047                               (sc->arg_fmt[arg.idx].show_zero ||
2048                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2049                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2050                               sc->arg_fmt[arg.idx].parm))
2051                                 continue;
2052
2053                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2054
2055                         if (trace->show_arg_names)
2056                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2057
2058                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2059                                                                   bf + printed, size - printed, &arg, val);
2060                 }
2061         } else if (IS_ERR(sc->tp_format)) {
2062                 /*
2063                  * If we managed to read the tracepoint /format file, then we
2064                  * may end up not having any args, like with gettid(), so only
2065                  * print the raw args when we didn't manage to read it.
2066                  */
2067                 while (arg.idx < sc->nr_args) {
2068                         if (arg.mask & bit)
2069                                 goto next_arg;
2070                         val = syscall_arg__val(&arg, arg.idx);
2071                         if (printed)
2072                                 printed += scnprintf(bf + printed, size - printed, ", ");
2073                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2074                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2075 next_arg:
2076                         ++arg.idx;
2077                         bit <<= 1;
2078                 }
2079         }
2080
2081         return printed;
2082 }
2083
2084 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2085                                   union perf_event *event,
2086                                   struct perf_sample *sample);
2087
2088 static struct syscall *trace__syscall_info(struct trace *trace,
2089                                            struct evsel *evsel, int id)
2090 {
2091         int err = 0;
2092
2093         if (id < 0) {
2094
2095                 /*
2096                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2097                  * before that, leaving at a higher verbosity level till that is
2098                  * explained. Reproduced with plain ftrace with:
2099                  *
2100                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2101                  * grep "NR -1 " /t/trace_pipe
2102                  *
2103                  * After generating some load on the machine.
2104                  */
2105                 if (verbose > 1) {
2106                         static u64 n;
2107                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2108                                 id, evsel__name(evsel), ++n);
2109                 }
2110                 return NULL;
2111         }
2112
2113         err = -EINVAL;
2114
2115 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2116         if (id > trace->sctbl->syscalls.max_id) {
2117 #else
2118         if (id >= trace->sctbl->syscalls.max_id) {
2119                 /*
2120                  * With libaudit we don't know beforehand what is the max_id,
2121                  * so we let trace__read_syscall_info() figure that out as we
2122                  * go on reading syscalls.
2123                  */
2124                 err = trace__read_syscall_info(trace, id);
2125                 if (err)
2126 #endif
2127                 goto out_cant_read;
2128         }
2129
2130         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2131             (err = trace__read_syscall_info(trace, id)) != 0)
2132                 goto out_cant_read;
2133
2134         if (trace->syscalls.table[id].name == NULL) {
2135                 if (trace->syscalls.table[id].nonexistent)
2136                         return NULL;
2137                 goto out_cant_read;
2138         }
2139
2140         return &trace->syscalls.table[id];
2141
2142 out_cant_read:
2143         if (verbose > 0) {
2144                 char sbuf[STRERR_BUFSIZE];
2145                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2146                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2147                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2148                 fputs(" information\n", trace->output);
2149         }
2150         return NULL;
2151 }
2152
2153 struct syscall_stats {
2154         struct stats stats;
2155         u64          nr_failures;
2156         int          max_errno;
2157         u32          *errnos;
2158 };
2159
2160 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2161                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2162 {
2163         struct int_node *inode;
2164         struct syscall_stats *stats;
2165         u64 duration = 0;
2166
2167         inode = intlist__findnew(ttrace->syscall_stats, id);
2168         if (inode == NULL)
2169                 return;
2170
2171         stats = inode->priv;
2172         if (stats == NULL) {
2173                 stats = zalloc(sizeof(*stats));
2174                 if (stats == NULL)
2175                         return;
2176
2177                 init_stats(&stats->stats);
2178                 inode->priv = stats;
2179         }
2180
2181         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2182                 duration = sample->time - ttrace->entry_time;
2183
2184         update_stats(&stats->stats, duration);
2185
2186         if (err < 0) {
2187                 ++stats->nr_failures;
2188
2189                 if (!errno_summary)
2190                         return;
2191
2192                 err = -err;
2193                 if (err > stats->max_errno) {
2194                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2195
2196                         if (new_errnos) {
2197                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2198                         } else {
2199                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2200                                          thread__comm_str(thread), thread->pid_, thread->tid);
2201                                 return;
2202                         }
2203
2204                         stats->errnos = new_errnos;
2205                         stats->max_errno = err;
2206                 }
2207
2208                 ++stats->errnos[err - 1];
2209         }
2210 }
2211
2212 static int trace__printf_interrupted_entry(struct trace *trace)
2213 {
2214         struct thread_trace *ttrace;
2215         size_t printed;
2216         int len;
2217
2218         if (trace->failure_only || trace->current == NULL)
2219                 return 0;
2220
2221         ttrace = thread__priv(trace->current);
2222
2223         if (!ttrace->entry_pending)
2224                 return 0;
2225
2226         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2227         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2228
2229         if (len < trace->args_alignment - 4)
2230                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2231
2232         printed += fprintf(trace->output, " ...\n");
2233
2234         ttrace->entry_pending = false;
2235         ++trace->nr_events_printed;
2236
2237         return printed;
2238 }
2239
2240 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2241                                  struct perf_sample *sample, struct thread *thread)
2242 {
2243         int printed = 0;
2244
2245         if (trace->print_sample) {
2246                 double ts = (double)sample->time / NSEC_PER_MSEC;
2247
2248                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2249                                    evsel__name(evsel), ts,
2250                                    thread__comm_str(thread),
2251                                    sample->pid, sample->tid, sample->cpu);
2252         }
2253
2254         return printed;
2255 }
2256
2257 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2258 {
2259         void *augmented_args = NULL;
2260         /*
2261          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2262          * and there we get all 6 syscall args plus the tracepoint common fields
2263          * that gets calculated at the start and the syscall_nr (another long).
2264          * So we check if that is the case and if so don't look after the
2265          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2266          * which is fixed.
2267          *
2268          * We'll revisit this later to pass s->args_size to the BPF augmenter
2269          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2270          * copies only what we need for each syscall, like what happens when we
2271          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2272          * traffic to just what is needed for each syscall.
2273          */
2274         int args_size = raw_augmented_args_size ?: sc->args_size;
2275
2276         *augmented_args_size = sample->raw_size - args_size;
2277         if (*augmented_args_size > 0)
2278                 augmented_args = sample->raw_data + args_size;
2279
2280         return augmented_args;
2281 }
2282
2283 static void syscall__exit(struct syscall *sc)
2284 {
2285         if (!sc)
2286                 return;
2287
2288         free(sc->arg_fmt);
2289 }
2290
2291 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2292                             union perf_event *event __maybe_unused,
2293                             struct perf_sample *sample)
2294 {
2295         char *msg;
2296         void *args;
2297         int printed = 0;
2298         struct thread *thread;
2299         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2300         int augmented_args_size = 0;
2301         void *augmented_args = NULL;
2302         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2303         struct thread_trace *ttrace;
2304
2305         if (sc == NULL)
2306                 return -1;
2307
2308         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2309         ttrace = thread__trace(thread, trace->output);
2310         if (ttrace == NULL)
2311                 goto out_put;
2312
2313         trace__fprintf_sample(trace, evsel, sample, thread);
2314
2315         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2316
2317         if (ttrace->entry_str == NULL) {
2318                 ttrace->entry_str = malloc(trace__entry_str_size);
2319                 if (!ttrace->entry_str)
2320                         goto out_put;
2321         }
2322
2323         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2324                 trace__printf_interrupted_entry(trace);
2325         /*
2326          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2327          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2328          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2329          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2330          * so when handling, say the openat syscall, we end up getting 6 args for the
2331          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2332          * thinking that the extra 2 u64 args are the augmented filename, so just check
2333          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2334          */
2335         if (evsel != trace->syscalls.events.sys_enter)
2336                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2337         ttrace->entry_time = sample->time;
2338         msg = ttrace->entry_str;
2339         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2340
2341         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2342                                            args, augmented_args, augmented_args_size, trace, thread);
2343
2344         if (sc->is_exit) {
2345                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2346                         int alignment = 0;
2347
2348                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2349                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2350                         if (trace->args_alignment > printed)
2351                                 alignment = trace->args_alignment - printed;
2352                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2353                 }
2354         } else {
2355                 ttrace->entry_pending = true;
2356                 /* See trace__vfs_getname & trace__sys_exit */
2357                 ttrace->filename.pending_open = false;
2358         }
2359
2360         if (trace->current != thread) {
2361                 thread__put(trace->current);
2362                 trace->current = thread__get(thread);
2363         }
2364         err = 0;
2365 out_put:
2366         thread__put(thread);
2367         return err;
2368 }
2369
2370 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2371                                     struct perf_sample *sample)
2372 {
2373         struct thread_trace *ttrace;
2374         struct thread *thread;
2375         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2376         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2377         char msg[1024];
2378         void *args, *augmented_args = NULL;
2379         int augmented_args_size;
2380
2381         if (sc == NULL)
2382                 return -1;
2383
2384         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2385         ttrace = thread__trace(thread, trace->output);
2386         /*
2387          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2388          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2389          */
2390         if (ttrace == NULL)
2391                 goto out_put;
2392
2393         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2394         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2395         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2396         fprintf(trace->output, "%s", msg);
2397         err = 0;
2398 out_put:
2399         thread__put(thread);
2400         return err;
2401 }
2402
2403 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2404                                     struct perf_sample *sample,
2405                                     struct callchain_cursor *cursor)
2406 {
2407         struct addr_location al;
2408         int max_stack = evsel->core.attr.sample_max_stack ?
2409                         evsel->core.attr.sample_max_stack :
2410                         trace->max_stack;
2411         int err;
2412
2413         if (machine__resolve(trace->host, &al, sample) < 0)
2414                 return -1;
2415
2416         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2417         addr_location__put(&al);
2418         return err;
2419 }
2420
2421 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2422 {
2423         /* TODO: user-configurable print_opts */
2424         const unsigned int print_opts = EVSEL__PRINT_SYM |
2425                                         EVSEL__PRINT_DSO |
2426                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2427
2428         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2429 }
2430
2431 static const char *errno_to_name(struct evsel *evsel, int err)
2432 {
2433         struct perf_env *env = evsel__env(evsel);
2434         const char *arch_name = perf_env__arch(env);
2435
2436         return arch_syscalls__strerrno(arch_name, err);
2437 }
2438
2439 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2440                            union perf_event *event __maybe_unused,
2441                            struct perf_sample *sample)
2442 {
2443         long ret;
2444         u64 duration = 0;
2445         bool duration_calculated = false;
2446         struct thread *thread;
2447         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2448         int alignment = trace->args_alignment;
2449         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2450         struct thread_trace *ttrace;
2451
2452         if (sc == NULL)
2453                 return -1;
2454
2455         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2456         ttrace = thread__trace(thread, trace->output);
2457         if (ttrace == NULL)
2458                 goto out_put;
2459
2460         trace__fprintf_sample(trace, evsel, sample, thread);
2461
2462         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2463
2464         if (trace->summary)
2465                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2466
2467         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2468                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2469                 ttrace->filename.pending_open = false;
2470                 ++trace->stats.vfs_getname;
2471         }
2472
2473         if (ttrace->entry_time) {
2474                 duration = sample->time - ttrace->entry_time;
2475                 if (trace__filter_duration(trace, duration))
2476                         goto out;
2477                 duration_calculated = true;
2478         } else if (trace->duration_filter)
2479                 goto out;
2480
2481         if (sample->callchain) {
2482                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2483                 if (callchain_ret == 0) {
2484                         if (callchain_cursor.nr < trace->min_stack)
2485                                 goto out;
2486                         callchain_ret = 1;
2487                 }
2488         }
2489
2490         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2491                 goto out;
2492
2493         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2494
2495         if (ttrace->entry_pending) {
2496                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2497         } else {
2498                 printed += fprintf(trace->output, " ... [");
2499                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2500                 printed += 9;
2501                 printed += fprintf(trace->output, "]: %s()", sc->name);
2502         }
2503
2504         printed++; /* the closing ')' */
2505
2506         if (alignment > printed)
2507                 alignment -= printed;
2508         else
2509                 alignment = 0;
2510
2511         fprintf(trace->output, ")%*s= ", alignment, " ");
2512
2513         if (sc->fmt == NULL) {
2514                 if (ret < 0)
2515                         goto errno_print;
2516 signed_print:
2517                 fprintf(trace->output, "%ld", ret);
2518         } else if (ret < 0) {
2519 errno_print: {
2520                 char bf[STRERR_BUFSIZE];
2521                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2522                            *e = errno_to_name(evsel, -ret);
2523
2524                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2525         }
2526         } else if (ret == 0 && sc->fmt->timeout)
2527                 fprintf(trace->output, "0 (Timeout)");
2528         else if (ttrace->ret_scnprintf) {
2529                 char bf[1024];
2530                 struct syscall_arg arg = {
2531                         .val    = ret,
2532                         .thread = thread,
2533                         .trace  = trace,
2534                 };
2535                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2536                 ttrace->ret_scnprintf = NULL;
2537                 fprintf(trace->output, "%s", bf);
2538         } else if (sc->fmt->hexret)
2539                 fprintf(trace->output, "%#lx", ret);
2540         else if (sc->fmt->errpid) {
2541                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2542
2543                 if (child != NULL) {
2544                         fprintf(trace->output, "%ld", ret);
2545                         if (child->comm_set)
2546                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2547                         thread__put(child);
2548                 }
2549         } else
2550                 goto signed_print;
2551
2552         fputc('\n', trace->output);
2553
2554         /*
2555          * We only consider an 'event' for the sake of --max-events a non-filtered
2556          * sys_enter + sys_exit and other tracepoint events.
2557          */
2558         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2559                 interrupted = true;
2560
2561         if (callchain_ret > 0)
2562                 trace__fprintf_callchain(trace, sample);
2563         else if (callchain_ret < 0)
2564                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2565 out:
2566         ttrace->entry_pending = false;
2567         err = 0;
2568 out_put:
2569         thread__put(thread);
2570         return err;
2571 }
2572
2573 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2574                               union perf_event *event __maybe_unused,
2575                               struct perf_sample *sample)
2576 {
2577         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2578         struct thread_trace *ttrace;
2579         size_t filename_len, entry_str_len, to_move;
2580         ssize_t remaining_space;
2581         char *pos;
2582         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2583
2584         if (!thread)
2585                 goto out;
2586
2587         ttrace = thread__priv(thread);
2588         if (!ttrace)
2589                 goto out_put;
2590
2591         filename_len = strlen(filename);
2592         if (filename_len == 0)
2593                 goto out_put;
2594
2595         if (ttrace->filename.namelen < filename_len) {
2596                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2597
2598                 if (f == NULL)
2599                         goto out_put;
2600
2601                 ttrace->filename.namelen = filename_len;
2602                 ttrace->filename.name = f;
2603         }
2604
2605         strcpy(ttrace->filename.name, filename);
2606         ttrace->filename.pending_open = true;
2607
2608         if (!ttrace->filename.ptr)
2609                 goto out_put;
2610
2611         entry_str_len = strlen(ttrace->entry_str);
2612         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2613         if (remaining_space <= 0)
2614                 goto out_put;
2615
2616         if (filename_len > (size_t)remaining_space) {
2617                 filename += filename_len - remaining_space;
2618                 filename_len = remaining_space;
2619         }
2620
2621         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2622         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2623         memmove(pos + filename_len, pos, to_move);
2624         memcpy(pos, filename, filename_len);
2625
2626         ttrace->filename.ptr = 0;
2627         ttrace->filename.entry_str_pos = 0;
2628 out_put:
2629         thread__put(thread);
2630 out:
2631         return 0;
2632 }
2633
2634 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2635                                      union perf_event *event __maybe_unused,
2636                                      struct perf_sample *sample)
2637 {
2638         u64 runtime = evsel__intval(evsel, sample, "runtime");
2639         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2640         struct thread *thread = machine__findnew_thread(trace->host,
2641                                                         sample->pid,
2642                                                         sample->tid);
2643         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2644
2645         if (ttrace == NULL)
2646                 goto out_dump;
2647
2648         ttrace->runtime_ms += runtime_ms;
2649         trace->runtime_ms += runtime_ms;
2650 out_put:
2651         thread__put(thread);
2652         return 0;
2653
2654 out_dump:
2655         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2656                evsel->name,
2657                evsel__strval(evsel, sample, "comm"),
2658                (pid_t)evsel__intval(evsel, sample, "pid"),
2659                runtime,
2660                evsel__intval(evsel, sample, "vruntime"));
2661         goto out_put;
2662 }
2663
2664 static int bpf_output__printer(enum binary_printer_ops op,
2665                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2666 {
2667         unsigned char ch = (unsigned char)val;
2668
2669         switch (op) {
2670         case BINARY_PRINT_CHAR_DATA:
2671                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2672         case BINARY_PRINT_DATA_BEGIN:
2673         case BINARY_PRINT_LINE_BEGIN:
2674         case BINARY_PRINT_ADDR:
2675         case BINARY_PRINT_NUM_DATA:
2676         case BINARY_PRINT_NUM_PAD:
2677         case BINARY_PRINT_SEP:
2678         case BINARY_PRINT_CHAR_PAD:
2679         case BINARY_PRINT_LINE_END:
2680         case BINARY_PRINT_DATA_END:
2681         default:
2682                 break;
2683         }
2684
2685         return 0;
2686 }
2687
2688 static void bpf_output__fprintf(struct trace *trace,
2689                                 struct perf_sample *sample)
2690 {
2691         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2692                         bpf_output__printer, NULL, trace->output);
2693         ++trace->nr_events_printed;
2694 }
2695
2696 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2697                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2698 {
2699         char bf[2048];
2700         size_t size = sizeof(bf);
2701         struct tep_format_field *field = evsel->tp_format->format.fields;
2702         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2703         size_t printed = 0;
2704         unsigned long val;
2705         u8 bit = 1;
2706         struct syscall_arg syscall_arg = {
2707                 .augmented = {
2708                         .size = augmented_args_size,
2709                         .args = augmented_args,
2710                 },
2711                 .idx    = 0,
2712                 .mask   = 0,
2713                 .trace  = trace,
2714                 .thread = thread,
2715                 .show_string_prefix = trace->show_string_prefix,
2716         };
2717
2718         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2719                 if (syscall_arg.mask & bit)
2720                         continue;
2721
2722                 syscall_arg.len = 0;
2723                 syscall_arg.fmt = arg;
2724                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2725                         int offset = field->offset;
2726
2727                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2728                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2729                                 syscall_arg.len = offset >> 16;
2730                                 offset &= 0xffff;
2731                                 if (field->flags & TEP_FIELD_IS_RELATIVE)
2732                                         offset += field->offset + field->size;
2733                         }
2734
2735                         val = (uintptr_t)(sample->raw_data + offset);
2736                 } else
2737                         val = format_field__intval(field, sample, evsel->needs_swap);
2738                 /*
2739                  * Some syscall args need some mask, most don't and
2740                  * return val untouched.
2741                  */
2742                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2743
2744                 /*
2745                  * Suppress this argument if its value is zero and
2746                  * we don't have a string associated in an
2747                  * strarray for it.
2748                  */
2749                 if (val == 0 &&
2750                     !trace->show_zeros &&
2751                     !((arg->show_zero ||
2752                        arg->scnprintf == SCA_STRARRAY ||
2753                        arg->scnprintf == SCA_STRARRAYS) &&
2754                       arg->parm))
2755                         continue;
2756
2757                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2758
2759                 if (trace->show_arg_names)
2760                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2761
2762                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2763         }
2764
2765         return printed + fprintf(trace->output, "%s", bf);
2766 }
2767
2768 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2769                                 union perf_event *event __maybe_unused,
2770                                 struct perf_sample *sample)
2771 {
2772         struct thread *thread;
2773         int callchain_ret = 0;
2774         /*
2775          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2776          * this event's max_events having been hit and this is an entry coming
2777          * from the ring buffer that we should discard, since the max events
2778          * have already been considered/printed.
2779          */
2780         if (evsel->disabled)
2781                 return 0;
2782
2783         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2784
2785         if (sample->callchain) {
2786                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2787                 if (callchain_ret == 0) {
2788                         if (callchain_cursor.nr < trace->min_stack)
2789                                 goto out;
2790                         callchain_ret = 1;
2791                 }
2792         }
2793
2794         trace__printf_interrupted_entry(trace);
2795         trace__fprintf_tstamp(trace, sample->time, trace->output);
2796
2797         if (trace->trace_syscalls && trace->show_duration)
2798                 fprintf(trace->output, "(         ): ");
2799
2800         if (thread)
2801                 trace__fprintf_comm_tid(trace, thread, trace->output);
2802
2803         if (evsel == trace->syscalls.events.augmented) {
2804                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2805                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2806
2807                 if (sc) {
2808                         fprintf(trace->output, "%s(", sc->name);
2809                         trace__fprintf_sys_enter(trace, evsel, sample);
2810                         fputc(')', trace->output);
2811                         goto newline;
2812                 }
2813
2814                 /*
2815                  * XXX: Not having the associated syscall info or not finding/adding
2816                  *      the thread should never happen, but if it does...
2817                  *      fall thru and print it as a bpf_output event.
2818                  */
2819         }
2820
2821         fprintf(trace->output, "%s(", evsel->name);
2822
2823         if (evsel__is_bpf_output(evsel)) {
2824                 bpf_output__fprintf(trace, sample);
2825         } else if (evsel->tp_format) {
2826                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2827                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2828                         if (trace->libtraceevent_print) {
2829                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2830                                                       sample->raw_data, sample->raw_size,
2831                                                       trace->output);
2832                         } else {
2833                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2834                         }
2835                 }
2836         }
2837
2838 newline:
2839         fprintf(trace->output, ")\n");
2840
2841         if (callchain_ret > 0)
2842                 trace__fprintf_callchain(trace, sample);
2843         else if (callchain_ret < 0)
2844                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2845
2846         ++trace->nr_events_printed;
2847
2848         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2849                 evsel__disable(evsel);
2850                 evsel__close(evsel);
2851         }
2852 out:
2853         thread__put(thread);
2854         return 0;
2855 }
2856
2857 static void print_location(FILE *f, struct perf_sample *sample,
2858                            struct addr_location *al,
2859                            bool print_dso, bool print_sym)
2860 {
2861
2862         if ((verbose > 0 || print_dso) && al->map)
2863                 fprintf(f, "%s@", al->map->dso->long_name);
2864
2865         if ((verbose > 0 || print_sym) && al->sym)
2866                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2867                         al->addr - al->sym->start);
2868         else if (al->map)
2869                 fprintf(f, "0x%" PRIx64, al->addr);
2870         else
2871                 fprintf(f, "0x%" PRIx64, sample->addr);
2872 }
2873
2874 static int trace__pgfault(struct trace *trace,
2875                           struct evsel *evsel,
2876                           union perf_event *event __maybe_unused,
2877                           struct perf_sample *sample)
2878 {
2879         struct thread *thread;
2880         struct addr_location al;
2881         char map_type = 'd';
2882         struct thread_trace *ttrace;
2883         int err = -1;
2884         int callchain_ret = 0;
2885
2886         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2887
2888         if (sample->callchain) {
2889                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2890                 if (callchain_ret == 0) {
2891                         if (callchain_cursor.nr < trace->min_stack)
2892                                 goto out_put;
2893                         callchain_ret = 1;
2894                 }
2895         }
2896
2897         ttrace = thread__trace(thread, trace->output);
2898         if (ttrace == NULL)
2899                 goto out_put;
2900
2901         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2902                 ttrace->pfmaj++;
2903         else
2904                 ttrace->pfmin++;
2905
2906         if (trace->summary_only)
2907                 goto out;
2908
2909         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2910
2911         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2912
2913         fprintf(trace->output, "%sfault [",
2914                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2915                 "maj" : "min");
2916
2917         print_location(trace->output, sample, &al, false, true);
2918
2919         fprintf(trace->output, "] => ");
2920
2921         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2922
2923         if (!al.map) {
2924                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2925
2926                 if (al.map)
2927                         map_type = 'x';
2928                 else
2929                         map_type = '?';
2930         }
2931
2932         print_location(trace->output, sample, &al, true, false);
2933
2934         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2935
2936         if (callchain_ret > 0)
2937                 trace__fprintf_callchain(trace, sample);
2938         else if (callchain_ret < 0)
2939                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2940
2941         ++trace->nr_events_printed;
2942 out:
2943         err = 0;
2944 out_put:
2945         thread__put(thread);
2946         return err;
2947 }
2948
2949 static void trace__set_base_time(struct trace *trace,
2950                                  struct evsel *evsel,
2951                                  struct perf_sample *sample)
2952 {
2953         /*
2954          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2955          * and don't use sample->time unconditionally, we may end up having
2956          * some other event in the future without PERF_SAMPLE_TIME for good
2957          * reason, i.e. we may not be interested in its timestamps, just in
2958          * it taking place, picking some piece of information when it
2959          * appears in our event stream (vfs_getname comes to mind).
2960          */
2961         if (trace->base_time == 0 && !trace->full_time &&
2962             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2963                 trace->base_time = sample->time;
2964 }
2965
2966 static int trace__process_sample(struct perf_tool *tool,
2967                                  union perf_event *event,
2968                                  struct perf_sample *sample,
2969                                  struct evsel *evsel,
2970                                  struct machine *machine __maybe_unused)
2971 {
2972         struct trace *trace = container_of(tool, struct trace, tool);
2973         struct thread *thread;
2974         int err = 0;
2975
2976         tracepoint_handler handler = evsel->handler;
2977
2978         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2979         if (thread && thread__is_filtered(thread))
2980                 goto out;
2981
2982         trace__set_base_time(trace, evsel, sample);
2983
2984         if (handler) {
2985                 ++trace->nr_events;
2986                 handler(trace, evsel, event, sample);
2987         }
2988 out:
2989         thread__put(thread);
2990         return err;
2991 }
2992
2993 static int trace__record(struct trace *trace, int argc, const char **argv)
2994 {
2995         unsigned int rec_argc, i, j;
2996         const char **rec_argv;
2997         const char * const record_args[] = {
2998                 "record",
2999                 "-R",
3000                 "-m", "1024",
3001                 "-c", "1",
3002         };
3003         pid_t pid = getpid();
3004         char *filter = asprintf__tp_filter_pids(1, &pid);
3005         const char * const sc_args[] = { "-e", };
3006         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3007         const char * const majpf_args[] = { "-e", "major-faults" };
3008         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3009         const char * const minpf_args[] = { "-e", "minor-faults" };
3010         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3011         int err = -1;
3012
3013         /* +3 is for the event string below and the pid filter */
3014         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3015                 majpf_args_nr + minpf_args_nr + argc;
3016         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3017
3018         if (rec_argv == NULL || filter == NULL)
3019                 goto out_free;
3020
3021         j = 0;
3022         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3023                 rec_argv[j++] = record_args[i];
3024
3025         if (trace->trace_syscalls) {
3026                 for (i = 0; i < sc_args_nr; i++)
3027                         rec_argv[j++] = sc_args[i];
3028
3029                 /* event string may be different for older kernels - e.g., RHEL6 */
3030                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3031                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3032                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3033                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3034                 else {
3035                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3036                         goto out_free;
3037                 }
3038         }
3039
3040         rec_argv[j++] = "--filter";
3041         rec_argv[j++] = filter;
3042
3043         if (trace->trace_pgfaults & TRACE_PFMAJ)
3044                 for (i = 0; i < majpf_args_nr; i++)
3045                         rec_argv[j++] = majpf_args[i];
3046
3047         if (trace->trace_pgfaults & TRACE_PFMIN)
3048                 for (i = 0; i < minpf_args_nr; i++)
3049                         rec_argv[j++] = minpf_args[i];
3050
3051         for (i = 0; i < (unsigned int)argc; i++)
3052                 rec_argv[j++] = argv[i];
3053
3054         err = cmd_record(j, rec_argv);
3055 out_free:
3056         free(filter);
3057         free(rec_argv);
3058         return err;
3059 }
3060
3061 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3062
3063 static bool evlist__add_vfs_getname(struct evlist *evlist)
3064 {
3065         bool found = false;
3066         struct evsel *evsel, *tmp;
3067         struct parse_events_error err;
3068         int ret;
3069
3070         parse_events_error__init(&err);
3071         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3072         parse_events_error__exit(&err);
3073         if (ret)
3074                 return false;
3075
3076         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3077                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3078                         continue;
3079
3080                 if (evsel__field(evsel, "pathname")) {
3081                         evsel->handler = trace__vfs_getname;
3082                         found = true;
3083                         continue;
3084                 }
3085
3086                 list_del_init(&evsel->core.node);
3087                 evsel->evlist = NULL;
3088                 evsel__delete(evsel);
3089         }
3090
3091         return found;
3092 }
3093
3094 static struct evsel *evsel__new_pgfault(u64 config)
3095 {
3096         struct evsel *evsel;
3097         struct perf_event_attr attr = {
3098                 .type = PERF_TYPE_SOFTWARE,
3099                 .mmap_data = 1,
3100         };
3101
3102         attr.config = config;
3103         attr.sample_period = 1;
3104
3105         event_attr_init(&attr);
3106
3107         evsel = evsel__new(&attr);
3108         if (evsel)
3109                 evsel->handler = trace__pgfault;
3110
3111         return evsel;
3112 }
3113
3114 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3115 {
3116         struct evsel *evsel;
3117
3118         evlist__for_each_entry(evlist, evsel) {
3119                 struct evsel_trace *et = evsel->priv;
3120
3121                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3122                         continue;
3123
3124                 free(et->fmt);
3125                 free(et);
3126         }
3127 }
3128
3129 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3130 {
3131         const u32 type = event->header.type;
3132         struct evsel *evsel;
3133
3134         if (type != PERF_RECORD_SAMPLE) {
3135                 trace__process_event(trace, trace->host, event, sample);
3136                 return;
3137         }
3138
3139         evsel = evlist__id2evsel(trace->evlist, sample->id);
3140         if (evsel == NULL) {
3141                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3142                 return;
3143         }
3144
3145         if (evswitch__discard(&trace->evswitch, evsel))
3146                 return;
3147
3148         trace__set_base_time(trace, evsel, sample);
3149
3150         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3151             sample->raw_data == NULL) {
3152                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3153                        evsel__name(evsel), sample->tid,
3154                        sample->cpu, sample->raw_size);
3155         } else {
3156                 tracepoint_handler handler = evsel->handler;
3157                 handler(trace, evsel, event, sample);
3158         }
3159
3160         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3161                 interrupted = true;
3162 }
3163
3164 static int trace__add_syscall_newtp(struct trace *trace)
3165 {
3166         int ret = -1;
3167         struct evlist *evlist = trace->evlist;
3168         struct evsel *sys_enter, *sys_exit;
3169
3170         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3171         if (sys_enter == NULL)
3172                 goto out;
3173
3174         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3175                 goto out_delete_sys_enter;
3176
3177         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3178         if (sys_exit == NULL)
3179                 goto out_delete_sys_enter;
3180
3181         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3182                 goto out_delete_sys_exit;
3183
3184         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3185         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3186
3187         evlist__add(evlist, sys_enter);
3188         evlist__add(evlist, sys_exit);
3189
3190         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3191                 /*
3192                  * We're interested only in the user space callchain
3193                  * leading to the syscall, allow overriding that for
3194                  * debugging reasons using --kernel_syscall_callchains
3195                  */
3196                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3197         }
3198
3199         trace->syscalls.events.sys_enter = sys_enter;
3200         trace->syscalls.events.sys_exit  = sys_exit;
3201
3202         ret = 0;
3203 out:
3204         return ret;
3205
3206 out_delete_sys_exit:
3207         evsel__delete_priv(sys_exit);
3208 out_delete_sys_enter:
3209         evsel__delete_priv(sys_enter);
3210         goto out;
3211 }
3212
3213 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3214 {
3215         int err = -1;
3216         struct evsel *sys_exit;
3217         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3218                                                 trace->ev_qualifier_ids.nr,
3219                                                 trace->ev_qualifier_ids.entries);
3220
3221         if (filter == NULL)
3222                 goto out_enomem;
3223
3224         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3225                 sys_exit = trace->syscalls.events.sys_exit;
3226                 err = evsel__append_tp_filter(sys_exit, filter);
3227         }
3228
3229         free(filter);
3230 out:
3231         return err;
3232 out_enomem:
3233         errno = ENOMEM;
3234         goto out;
3235 }
3236
3237 #ifdef HAVE_LIBBPF_SUPPORT
3238 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3239 {
3240         if (trace->bpf_obj == NULL)
3241                 return NULL;
3242
3243         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3244 }
3245
3246 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3247 {
3248         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3249 }
3250
3251 static void trace__set_bpf_map_syscalls(struct trace *trace)
3252 {
3253         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3254         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3255         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3256 }
3257
3258 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3259 {
3260         struct bpf_program *pos, *prog = NULL;
3261         const char *sec_name;
3262
3263         if (trace->bpf_obj == NULL)
3264                 return NULL;
3265
3266         bpf_object__for_each_program(pos, trace->bpf_obj) {
3267                 sec_name = bpf_program__section_name(pos);
3268                 if (sec_name && !strcmp(sec_name, name)) {
3269                         prog = pos;
3270                         break;
3271                 }
3272         }
3273
3274         return prog;
3275 }
3276
3277 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3278                                                         const char *prog_name, const char *type)
3279 {
3280         struct bpf_program *prog;
3281
3282         if (prog_name == NULL) {
3283                 char default_prog_name[256];
3284                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3285                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3286                 if (prog != NULL)
3287                         goto out_found;
3288                 if (sc->fmt && sc->fmt->alias) {
3289                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3290                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3291                         if (prog != NULL)
3292                                 goto out_found;
3293                 }
3294                 goto out_unaugmented;
3295         }
3296
3297         prog = trace__find_bpf_program_by_title(trace, prog_name);
3298
3299         if (prog != NULL) {
3300 out_found:
3301                 return prog;
3302         }
3303
3304         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3305                  prog_name, type, sc->name);
3306 out_unaugmented:
3307         return trace->syscalls.unaugmented_prog;
3308 }
3309
3310 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3311 {
3312         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3313
3314         if (sc == NULL)
3315                 return;
3316
3317         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3318         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3319 }
3320
3321 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3322 {
3323         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3324         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3325 }
3326
3327 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3328 {
3329         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3330         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3331 }
3332
3333 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3334 {
3335         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3336         int arg = 0;
3337
3338         if (sc == NULL)
3339                 goto out;
3340
3341         for (; arg < sc->nr_args; ++arg) {
3342                 entry->string_args_len[arg] = 0;
3343                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3344                         /* Should be set like strace -s strsize */
3345                         entry->string_args_len[arg] = PATH_MAX;
3346                 }
3347         }
3348 out:
3349         for (; arg < 6; ++arg)
3350                 entry->string_args_len[arg] = 0;
3351 }
3352 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3353 {
3354         int fd = bpf_map__fd(trace->syscalls.map);
3355         struct bpf_map_syscall_entry value = {
3356                 .enabled = !trace->not_ev_qualifier,
3357         };
3358         int err = 0;
3359         size_t i;
3360
3361         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3362                 int key = trace->ev_qualifier_ids.entries[i];
3363
3364                 if (value.enabled) {
3365                         trace__init_bpf_map_syscall_args(trace, key, &value);
3366                         trace__init_syscall_bpf_progs(trace, key);
3367                 }
3368
3369                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3370                 if (err)
3371                         break;
3372         }
3373
3374         return err;
3375 }
3376
3377 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3378 {
3379         int fd = bpf_map__fd(trace->syscalls.map);
3380         struct bpf_map_syscall_entry value = {
3381                 .enabled = enabled,
3382         };
3383         int err = 0, key;
3384
3385         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3386                 if (enabled)
3387                         trace__init_bpf_map_syscall_args(trace, key, &value);
3388
3389                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3390                 if (err)
3391                         break;
3392         }
3393
3394         return err;
3395 }
3396
3397 static int trace__init_syscalls_bpf_map(struct trace *trace)
3398 {
3399         bool enabled = true;
3400
3401         if (trace->ev_qualifier_ids.nr)
3402                 enabled = trace->not_ev_qualifier;
3403
3404         return __trace__init_syscalls_bpf_map(trace, enabled);
3405 }
3406
3407 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3408 {
3409         struct tep_format_field *field, *candidate_field;
3410         int id;
3411
3412         /*
3413          * We're only interested in syscalls that have a pointer:
3414          */
3415         for (field = sc->args; field; field = field->next) {
3416                 if (field->flags & TEP_FIELD_IS_POINTER)
3417                         goto try_to_find_pair;
3418         }
3419
3420         return NULL;
3421
3422 try_to_find_pair:
3423         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3424                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3425                 struct bpf_program *pair_prog;
3426                 bool is_candidate = false;
3427
3428                 if (pair == NULL || pair == sc ||
3429                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3430                         continue;
3431
3432                 for (field = sc->args, candidate_field = pair->args;
3433                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3434                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3435                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3436
3437                         if (is_pointer) {
3438                                if (!candidate_is_pointer) {
3439                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3440                                         continue;
3441                                }
3442                         } else {
3443                                 if (candidate_is_pointer) {
3444                                         // The candidate might copy a pointer we don't have, skip it.
3445                                         goto next_candidate;
3446                                 }
3447                                 continue;
3448                         }
3449
3450                         if (strcmp(field->type, candidate_field->type))
3451                                 goto next_candidate;
3452
3453                         is_candidate = true;
3454                 }
3455
3456                 if (!is_candidate)
3457                         goto next_candidate;
3458
3459                 /*
3460                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3461                  * then it may be collecting that and we then can't use it, as it would collect
3462                  * more than what is common to the two syscalls.
3463                  */
3464                 if (candidate_field) {
3465                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3466                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3467                                         goto next_candidate;
3468                 }
3469
3470                 pair_prog = pair->bpf_prog.sys_enter;
3471                 /*
3472                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3473                  * have been searched for, so search it here and if it returns the
3474                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3475                  * program for a filtered syscall on a non-filtered one.
3476                  *
3477                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3478                  * useful for "renameat2".
3479                  */
3480                 if (pair_prog == NULL) {
3481                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3482                         if (pair_prog == trace->syscalls.unaugmented_prog)
3483                                 goto next_candidate;
3484                 }
3485
3486                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3487                 return pair_prog;
3488         next_candidate:
3489                 continue;
3490         }
3491
3492         return NULL;
3493 }
3494
3495 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3496 {
3497         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3498             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3499         int err = 0, key;
3500
3501         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3502                 int prog_fd;
3503
3504                 if (!trace__syscall_enabled(trace, key))
3505                         continue;
3506
3507                 trace__init_syscall_bpf_progs(trace, key);
3508
3509                 // It'll get at least the "!raw_syscalls:unaugmented"
3510                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3511                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3512                 if (err)
3513                         break;
3514                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3515                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3516                 if (err)
3517                         break;
3518         }
3519
3520         /*
3521          * Now lets do a second pass looking for enabled syscalls without
3522          * an augmenter that have a signature that is a superset of another
3523          * syscall with an augmenter so that we can auto-reuse it.
3524          *
3525          * I.e. if we have an augmenter for the "open" syscall that has
3526          * this signature:
3527          *
3528          *   int open(const char *pathname, int flags, mode_t mode);
3529          *
3530          * I.e. that will collect just the first string argument, then we
3531          * can reuse it for the 'creat' syscall, that has this signature:
3532          *
3533          *   int creat(const char *pathname, mode_t mode);
3534          *
3535          * and for:
3536          *
3537          *   int stat(const char *pathname, struct stat *statbuf);
3538          *   int lstat(const char *pathname, struct stat *statbuf);
3539          *
3540          * Because the 'open' augmenter will collect the first arg as a string,
3541          * and leave alone all the other args, which already helps with
3542          * beautifying 'stat' and 'lstat''s pathname arg.
3543          *
3544          * Then, in time, when 'stat' gets an augmenter that collects both
3545          * first and second arg (this one on the raw_syscalls:sys_exit prog
3546          * array tail call, then that one will be used.
3547          */
3548         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3549                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3550                 struct bpf_program *pair_prog;
3551                 int prog_fd;
3552
3553                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3554                         continue;
3555
3556                 /*
3557                  * For now we're just reusing the sys_enter prog, and if it
3558                  * already has an augmenter, we don't need to find one.
3559                  */
3560                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3561                         continue;
3562
3563                 /*
3564                  * Look at all the other syscalls for one that has a signature
3565                  * that is close enough that we can share:
3566                  */
3567                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3568                 if (pair_prog == NULL)
3569                         continue;
3570
3571                 sc->bpf_prog.sys_enter = pair_prog;
3572
3573                 /*
3574                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3575                  * with the fd for the program we're reusing:
3576                  */
3577                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3578                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3579                 if (err)
3580                         break;
3581         }
3582
3583
3584         return err;
3585 }
3586
3587 static void trace__delete_augmented_syscalls(struct trace *trace)
3588 {
3589         struct evsel *evsel, *tmp;
3590
3591         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3592         evsel__delete(trace->syscalls.events.augmented);
3593         trace->syscalls.events.augmented = NULL;
3594
3595         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3596                 if (evsel->bpf_obj == trace->bpf_obj) {
3597                         evlist__remove(trace->evlist, evsel);
3598                         evsel__delete(evsel);
3599                 }
3600
3601         }
3602
3603         bpf_object__close(trace->bpf_obj);
3604         trace->bpf_obj = NULL;
3605 }
3606 #else // HAVE_LIBBPF_SUPPORT
3607 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3608                                                    const char *name __maybe_unused)
3609 {
3610         return NULL;
3611 }
3612
3613 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3614 {
3615 }
3616
3617 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3618 {
3619 }
3620
3621 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3622 {
3623         return 0;
3624 }
3625
3626 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3627 {
3628         return 0;
3629 }
3630
3631 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3632                                                             const char *name __maybe_unused)
3633 {
3634         return NULL;
3635 }
3636
3637 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3638 {
3639         return 0;
3640 }
3641
3642 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3643 {
3644 }
3645 #endif // HAVE_LIBBPF_SUPPORT
3646
3647 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3648 {
3649         struct evsel *evsel;
3650
3651         evlist__for_each_entry(trace->evlist, evsel) {
3652                 if (evsel == trace->syscalls.events.augmented ||
3653                     evsel->bpf_obj == trace->bpf_obj)
3654                         continue;
3655
3656                 return false;
3657         }
3658
3659         return true;
3660 }
3661
3662 static int trace__set_ev_qualifier_filter(struct trace *trace)
3663 {
3664         if (trace->syscalls.map)
3665                 return trace__set_ev_qualifier_bpf_filter(trace);
3666         if (trace->syscalls.events.sys_enter)
3667                 return trace__set_ev_qualifier_tp_filter(trace);
3668         return 0;
3669 }
3670
3671 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3672                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3673 {
3674         int err = 0;
3675 #ifdef HAVE_LIBBPF_SUPPORT
3676         bool value = true;
3677         int map_fd = bpf_map__fd(map);
3678         size_t i;
3679
3680         for (i = 0; i < npids; ++i) {
3681                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3682                 if (err)
3683                         break;
3684         }
3685 #endif
3686         return err;
3687 }
3688
3689 static int trace__set_filter_loop_pids(struct trace *trace)
3690 {
3691         unsigned int nr = 1, err;
3692         pid_t pids[32] = {
3693                 getpid(),
3694         };
3695         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3696
3697         while (thread && nr < ARRAY_SIZE(pids)) {
3698                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3699
3700                 if (parent == NULL)
3701                         break;
3702
3703                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3704                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3705                         pids[nr++] = parent->tid;
3706                         break;
3707                 }
3708                 thread = parent;
3709         }
3710
3711         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3712         if (!err && trace->filter_pids.map)
3713                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3714
3715         return err;
3716 }
3717
3718 static int trace__set_filter_pids(struct trace *trace)
3719 {
3720         int err = 0;
3721         /*
3722          * Better not use !target__has_task() here because we need to cover the
3723          * case where no threads were specified in the command line, but a
3724          * workload was, and in that case we will fill in the thread_map when
3725          * we fork the workload in evlist__prepare_workload.
3726          */
3727         if (trace->filter_pids.nr > 0) {
3728                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3729                                                     trace->filter_pids.entries);
3730                 if (!err && trace->filter_pids.map) {
3731                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3732                                                        trace->filter_pids.entries);
3733                 }
3734         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3735                 err = trace__set_filter_loop_pids(trace);
3736         }
3737
3738         return err;
3739 }
3740
3741 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3742 {
3743         struct evlist *evlist = trace->evlist;
3744         struct perf_sample sample;
3745         int err = evlist__parse_sample(evlist, event, &sample);
3746
3747         if (err)
3748                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3749         else
3750                 trace__handle_event(trace, event, &sample);
3751
3752         return 0;
3753 }
3754
3755 static int __trace__flush_events(struct trace *trace)
3756 {
3757         u64 first = ordered_events__first_time(&trace->oe.data);
3758         u64 flush = trace->oe.last - NSEC_PER_SEC;
3759
3760         /* Is there some thing to flush.. */
3761         if (first && first < flush)
3762                 return ordered_events__flush_time(&trace->oe.data, flush);
3763
3764         return 0;
3765 }
3766
3767 static int trace__flush_events(struct trace *trace)
3768 {
3769         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3770 }
3771
3772 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3773 {
3774         int err;
3775
3776         if (!trace->sort_events)
3777                 return __trace__deliver_event(trace, event);
3778
3779         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3780         if (err && err != -1)
3781                 return err;
3782
3783         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3784         if (err)
3785                 return err;
3786
3787         return trace__flush_events(trace);
3788 }
3789
3790 static int ordered_events__deliver_event(struct ordered_events *oe,
3791                                          struct ordered_event *event)
3792 {
3793         struct trace *trace = container_of(oe, struct trace, oe.data);
3794
3795         return __trace__deliver_event(trace, event->event);
3796 }
3797
3798 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3799 {
3800         struct tep_format_field *field;
3801         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3802
3803         if (evsel->tp_format == NULL || fmt == NULL)
3804                 return NULL;
3805
3806         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3807                 if (strcmp(field->name, arg) == 0)
3808                         return fmt;
3809
3810         return NULL;
3811 }
3812
3813 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3814 {
3815         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3816
3817         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3818                 char *right = tok + 1, *right_end;
3819
3820                 if (*right == '=')
3821                         ++right;
3822
3823                 while (isspace(*right))
3824                         ++right;
3825
3826                 if (*right == '\0')
3827                         break;
3828
3829                 while (!isalpha(*left))
3830                         if (++left == tok) {
3831                                 /*
3832                                  * Bail out, can't find the name of the argument that is being
3833                                  * used in the filter, let it try to set this filter, will fail later.
3834                                  */
3835                                 return 0;
3836                         }
3837
3838                 right_end = right + 1;
3839                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3840                         ++right_end;
3841
3842                 if (isalpha(*right)) {
3843                         struct syscall_arg_fmt *fmt;
3844                         int left_size = tok - left,
3845                             right_size = right_end - right;
3846                         char arg[128];
3847
3848                         while (isspace(left[left_size - 1]))
3849                                 --left_size;
3850
3851                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3852
3853                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3854                         if (fmt == NULL) {
3855                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3856                                        arg, evsel->name, evsel->filter);
3857                                 return -1;
3858                         }
3859
3860                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3861                                  arg, (int)(right - tok), tok, right_size, right);
3862
3863                         if (fmt->strtoul) {
3864                                 u64 val;
3865                                 struct syscall_arg syscall_arg = {
3866                                         .parm = fmt->parm,
3867                                 };
3868
3869                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3870                                         char *n, expansion[19];
3871                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3872                                         int expansion_offset = right - new_filter;
3873
3874                                         pr_debug("%s", expansion);
3875
3876                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3877                                                 pr_debug(" out of memory!\n");
3878                                                 free(new_filter);
3879                                                 return -1;
3880                                         }
3881                                         if (new_filter != evsel->filter)
3882                                                 free(new_filter);
3883                                         left = n + expansion_offset + expansion_lenght;
3884                                         new_filter = n;
3885                                 } else {
3886                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3887                                                right_size, right, arg, evsel->name, evsel->filter);
3888                                         return -1;
3889                                 }
3890                         } else {
3891                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3892                                        arg, evsel->name, evsel->filter);
3893                                 return -1;
3894                         }
3895
3896                         pr_debug("\n");
3897                 } else {
3898                         left = right_end;
3899                 }
3900         }
3901
3902         if (new_filter != evsel->filter) {
3903                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3904                 evsel__set_filter(evsel, new_filter);
3905                 free(new_filter);
3906         }
3907
3908         return 0;
3909 }
3910
3911 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3912 {
3913         struct evlist *evlist = trace->evlist;
3914         struct evsel *evsel;
3915
3916         evlist__for_each_entry(evlist, evsel) {
3917                 if (evsel->filter == NULL)
3918                         continue;
3919
3920                 if (trace__expand_filter(trace, evsel)) {
3921                         *err_evsel = evsel;
3922                         return -1;
3923                 }
3924         }
3925
3926         return 0;
3927 }
3928
3929 static int trace__run(struct trace *trace, int argc, const char **argv)
3930 {
3931         struct evlist *evlist = trace->evlist;
3932         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3933         int err = -1, i;
3934         unsigned long before;
3935         const bool forks = argc > 0;
3936         bool draining = false;
3937
3938         trace->live = true;
3939
3940         if (!trace->raw_augmented_syscalls) {
3941                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3942                         goto out_error_raw_syscalls;
3943
3944                 if (trace->trace_syscalls)
3945                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3946         }
3947
3948         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3949                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3950                 if (pgfault_maj == NULL)
3951                         goto out_error_mem;
3952                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3953                 evlist__add(evlist, pgfault_maj);
3954         }
3955
3956         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3957                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3958                 if (pgfault_min == NULL)
3959                         goto out_error_mem;
3960                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3961                 evlist__add(evlist, pgfault_min);
3962         }
3963
3964         /* Enable ignoring missing threads when -u/-p option is defined. */
3965         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3966
3967         if (trace->sched &&
3968             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3969                 goto out_error_sched_stat_runtime;
3970         /*
3971          * If a global cgroup was set, apply it to all the events without an
3972          * explicit cgroup. I.e.:
3973          *
3974          *      trace -G A -e sched:*switch
3975          *
3976          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3977          * _and_ sched:sched_switch to the 'A' cgroup, while:
3978          *
3979          * trace -e sched:*switch -G A
3980          *
3981          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3982          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3983          * a cgroup (on the root cgroup, sys wide, etc).
3984          *
3985          * Multiple cgroups:
3986          *
3987          * trace -G A -e sched:*switch -G B
3988          *
3989          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3990          * to the 'B' cgroup.
3991          *
3992          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3993          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3994          */
3995         if (trace->cgroup)
3996                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3997
3998         err = evlist__create_maps(evlist, &trace->opts.target);
3999         if (err < 0) {
4000                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
4001                 goto out_delete_evlist;
4002         }
4003
4004         err = trace__symbols_init(trace, evlist);
4005         if (err < 0) {
4006                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
4007                 goto out_delete_evlist;
4008         }
4009
4010         evlist__config(evlist, &trace->opts, &callchain_param);
4011
4012         if (forks) {
4013                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
4014                 if (err < 0) {
4015                         fprintf(trace->output, "Couldn't run the workload!\n");
4016                         goto out_delete_evlist;
4017                 }
4018                 workload_pid = evlist->workload.pid;
4019         }
4020
4021         err = evlist__open(evlist);
4022         if (err < 0)
4023                 goto out_error_open;
4024
4025         err = bpf__apply_obj_config();
4026         if (err) {
4027                 char errbuf[BUFSIZ];
4028
4029                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
4030                 pr_err("ERROR: Apply config to BPF failed: %s\n",
4031                          errbuf);
4032                 goto out_error_open;
4033         }
4034
4035         err = trace__set_filter_pids(trace);
4036         if (err < 0)
4037                 goto out_error_mem;
4038
4039         if (trace->syscalls.map)
4040                 trace__init_syscalls_bpf_map(trace);
4041
4042         if (trace->syscalls.prog_array.sys_enter)
4043                 trace__init_syscalls_bpf_prog_array_maps(trace);
4044
4045         if (trace->ev_qualifier_ids.nr > 0) {
4046                 err = trace__set_ev_qualifier_filter(trace);
4047                 if (err < 0)
4048                         goto out_errno;
4049
4050                 if (trace->syscalls.events.sys_exit) {
4051                         pr_debug("event qualifier tracepoint filter: %s\n",
4052                                  trace->syscalls.events.sys_exit->filter);
4053                 }
4054         }
4055
4056         /*
4057          * If the "close" syscall is not traced, then we will not have the
4058          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4059          * fd->pathname table and were ending up showing the last value set by
4060          * syscalls opening a pathname and associating it with a descriptor or
4061          * reading it from /proc/pid/fd/ in cases where that doesn't make
4062          * sense.
4063          *
4064          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4065          *  not in use.
4066          */
4067         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4068
4069         err = trace__expand_filters(trace, &evsel);
4070         if (err)
4071                 goto out_delete_evlist;
4072         err = evlist__apply_filters(evlist, &evsel);
4073         if (err < 0)
4074                 goto out_error_apply_filters;
4075
4076         if (trace->dump.map)
4077                 bpf_map__fprintf(trace->dump.map, trace->output);
4078
4079         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4080         if (err < 0)
4081                 goto out_error_mmap;
4082
4083         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4084                 evlist__enable(evlist);
4085
4086         if (forks)
4087                 evlist__start_workload(evlist);
4088
4089         if (trace->opts.initial_delay) {
4090                 usleep(trace->opts.initial_delay * 1000);
4091                 evlist__enable(evlist);
4092         }
4093
4094         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4095                                   evlist->core.threads->nr > 1 ||
4096                                   evlist__first(evlist)->core.attr.inherit;
4097
4098         /*
4099          * Now that we already used evsel->core.attr to ask the kernel to setup the
4100          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4101          * trace__resolve_callchain(), allowing per-event max-stack settings
4102          * to override an explicitly set --max-stack global setting.
4103          */
4104         evlist__for_each_entry(evlist, evsel) {
4105                 if (evsel__has_callchain(evsel) &&
4106                     evsel->core.attr.sample_max_stack == 0)
4107                         evsel->core.attr.sample_max_stack = trace->max_stack;
4108         }
4109 again:
4110         before = trace->nr_events;
4111
4112         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4113                 union perf_event *event;
4114                 struct mmap *md;
4115
4116                 md = &evlist->mmap[i];
4117                 if (perf_mmap__read_init(&md->core) < 0)
4118                         continue;
4119
4120                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4121                         ++trace->nr_events;
4122
4123                         err = trace__deliver_event(trace, event);
4124                         if (err)
4125                                 goto out_disable;
4126
4127                         perf_mmap__consume(&md->core);
4128
4129                         if (interrupted)
4130                                 goto out_disable;
4131
4132                         if (done && !draining) {
4133                                 evlist__disable(evlist);
4134                                 draining = true;
4135                         }
4136                 }
4137                 perf_mmap__read_done(&md->core);
4138         }
4139
4140         if (trace->nr_events == before) {
4141                 int timeout = done ? 100 : -1;
4142
4143                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4144                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4145                                 draining = true;
4146
4147                         goto again;
4148                 } else {
4149                         if (trace__flush_events(trace))
4150                                 goto out_disable;
4151                 }
4152         } else {
4153                 goto again;
4154         }
4155
4156 out_disable:
4157         thread__zput(trace->current);
4158
4159         evlist__disable(evlist);
4160
4161         if (trace->sort_events)
4162                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4163
4164         if (!err) {
4165                 if (trace->summary)
4166                         trace__fprintf_thread_summary(trace, trace->output);
4167
4168                 if (trace->show_tool_stats) {
4169                         fprintf(trace->output, "Stats:\n "
4170                                                " vfs_getname : %" PRIu64 "\n"
4171                                                " proc_getname: %" PRIu64 "\n",
4172                                 trace->stats.vfs_getname,
4173                                 trace->stats.proc_getname);
4174                 }
4175         }
4176
4177 out_delete_evlist:
4178         trace__symbols__exit(trace);
4179         evlist__free_syscall_tp_fields(evlist);
4180         evlist__delete(evlist);
4181         cgroup__put(trace->cgroup);
4182         trace->evlist = NULL;
4183         trace->live = false;
4184         return err;
4185 {
4186         char errbuf[BUFSIZ];
4187
4188 out_error_sched_stat_runtime:
4189         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4190         goto out_error;
4191
4192 out_error_raw_syscalls:
4193         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4194         goto out_error;
4195
4196 out_error_mmap:
4197         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4198         goto out_error;
4199
4200 out_error_open:
4201         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4202
4203 out_error:
4204         fprintf(trace->output, "%s\n", errbuf);
4205         goto out_delete_evlist;
4206
4207 out_error_apply_filters:
4208         fprintf(trace->output,
4209                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4210                 evsel->filter, evsel__name(evsel), errno,
4211                 str_error_r(errno, errbuf, sizeof(errbuf)));
4212         goto out_delete_evlist;
4213 }
4214 out_error_mem:
4215         fprintf(trace->output, "Not enough memory to run!\n");
4216         goto out_delete_evlist;
4217
4218 out_errno:
4219         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4220         goto out_delete_evlist;
4221 }
4222
4223 static int trace__replay(struct trace *trace)
4224 {
4225         const struct evsel_str_handler handlers[] = {
4226                 { "probe:vfs_getname",       trace__vfs_getname, },
4227         };
4228         struct perf_data data = {
4229                 .path  = input_name,
4230                 .mode  = PERF_DATA_MODE_READ,
4231                 .force = trace->force,
4232         };
4233         struct perf_session *session;
4234         struct evsel *evsel;
4235         int err = -1;
4236
4237         trace->tool.sample        = trace__process_sample;
4238         trace->tool.mmap          = perf_event__process_mmap;
4239         trace->tool.mmap2         = perf_event__process_mmap2;
4240         trace->tool.comm          = perf_event__process_comm;
4241         trace->tool.exit          = perf_event__process_exit;
4242         trace->tool.fork          = perf_event__process_fork;
4243         trace->tool.attr          = perf_event__process_attr;
4244         trace->tool.tracing_data  = perf_event__process_tracing_data;
4245         trace->tool.build_id      = perf_event__process_build_id;
4246         trace->tool.namespaces    = perf_event__process_namespaces;
4247
4248         trace->tool.ordered_events = true;
4249         trace->tool.ordering_requires_timestamps = true;
4250
4251         /* add tid to output */
4252         trace->multiple_threads = true;
4253
4254         session = perf_session__new(&data, &trace->tool);
4255         if (IS_ERR(session))
4256                 return PTR_ERR(session);
4257
4258         if (trace->opts.target.pid)
4259                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4260
4261         if (trace->opts.target.tid)
4262                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4263
4264         if (symbol__init(&session->header.env) < 0)
4265                 goto out;
4266
4267         trace->host = &session->machines.host;
4268
4269         err = perf_session__set_tracepoints_handlers(session, handlers);
4270         if (err)
4271                 goto out;
4272
4273         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4274         trace->syscalls.events.sys_enter = evsel;
4275         /* older kernels have syscalls tp versus raw_syscalls */
4276         if (evsel == NULL)
4277                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4278
4279         if (evsel &&
4280             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4281             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4282                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4283                 goto out;
4284         }
4285
4286         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4287         trace->syscalls.events.sys_exit = evsel;
4288         if (evsel == NULL)
4289                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4290         if (evsel &&
4291             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4292             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4293                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4294                 goto out;
4295         }
4296
4297         evlist__for_each_entry(session->evlist, evsel) {
4298                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4299                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4300                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4301                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4302                         evsel->handler = trace__pgfault;
4303         }
4304
4305         setup_pager();
4306
4307         err = perf_session__process_events(session);
4308         if (err)
4309                 pr_err("Failed to process events, error %d", err);
4310
4311         else if (trace->summary)
4312                 trace__fprintf_thread_summary(trace, trace->output);
4313
4314 out:
4315         perf_session__delete(session);
4316
4317         return err;
4318 }
4319
4320 static size_t trace__fprintf_threads_header(FILE *fp)
4321 {
4322         size_t printed;
4323
4324         printed  = fprintf(fp, "\n Summary of events:\n\n");
4325
4326         return printed;
4327 }
4328
4329 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4330         struct syscall_stats *stats;
4331         double               msecs;
4332         int                  syscall;
4333 )
4334 {
4335         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4336         struct syscall_stats *stats = source->priv;
4337
4338         entry->syscall = source->i;
4339         entry->stats   = stats;
4340         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4341 }
4342
4343 static size_t thread__dump_stats(struct thread_trace *ttrace,
4344                                  struct trace *trace, FILE *fp)
4345 {
4346         size_t printed = 0;
4347         struct syscall *sc;
4348         struct rb_node *nd;
4349         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4350
4351         if (syscall_stats == NULL)
4352                 return 0;
4353
4354         printed += fprintf(fp, "\n");
4355
4356         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4357         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4358         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4359
4360         resort_rb__for_each_entry(nd, syscall_stats) {
4361                 struct syscall_stats *stats = syscall_stats_entry->stats;
4362                 if (stats) {
4363                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4364                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4365                         double avg = avg_stats(&stats->stats);
4366                         double pct;
4367                         u64 n = (u64)stats->stats.n;
4368
4369                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4370                         avg /= NSEC_PER_MSEC;
4371
4372                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4373                         printed += fprintf(fp, "   %-15s", sc->name);
4374                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4375                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4376                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4377
4378                         if (trace->errno_summary && stats->nr_failures) {
4379                                 const char *arch_name = perf_env__arch(trace->host->env);
4380                                 int e;
4381
4382                                 for (e = 0; e < stats->max_errno; ++e) {
4383                                         if (stats->errnos[e] != 0)
4384                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4385                                 }
4386                         }
4387                 }
4388         }
4389
4390         resort_rb__delete(syscall_stats);
4391         printed += fprintf(fp, "\n\n");
4392
4393         return printed;
4394 }
4395
4396 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4397 {
4398         size_t printed = 0;
4399         struct thread_trace *ttrace = thread__priv(thread);
4400         double ratio;
4401
4402         if (ttrace == NULL)
4403                 return 0;
4404
4405         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4406
4407         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4408         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4409         printed += fprintf(fp, "%.1f%%", ratio);
4410         if (ttrace->pfmaj)
4411                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4412         if (ttrace->pfmin)
4413                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4414         if (trace->sched)
4415                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4416         else if (fputc('\n', fp) != EOF)
4417                 ++printed;
4418
4419         printed += thread__dump_stats(ttrace, trace, fp);
4420
4421         return printed;
4422 }
4423
4424 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4425 {
4426         return ttrace ? ttrace->nr_events : 0;
4427 }
4428
4429 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4430         struct thread *thread;
4431 )
4432 {
4433         entry->thread = rb_entry(nd, struct thread, rb_node);
4434 }
4435
4436 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4437 {
4438         size_t printed = trace__fprintf_threads_header(fp);
4439         struct rb_node *nd;
4440         int i;
4441
4442         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4443                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4444
4445                 if (threads == NULL) {
4446                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4447                         return 0;
4448                 }
4449
4450                 resort_rb__for_each_entry(nd, threads)
4451                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4452
4453                 resort_rb__delete(threads);
4454         }
4455         return printed;
4456 }
4457
4458 static int trace__set_duration(const struct option *opt, const char *str,
4459                                int unset __maybe_unused)
4460 {
4461         struct trace *trace = opt->value;
4462
4463         trace->duration_filter = atof(str);
4464         return 0;
4465 }
4466
4467 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4468                                               int unset __maybe_unused)
4469 {
4470         int ret = -1;
4471         size_t i;
4472         struct trace *trace = opt->value;
4473         /*
4474          * FIXME: introduce a intarray class, plain parse csv and create a
4475          * { int nr, int entries[] } struct...
4476          */
4477         struct intlist *list = intlist__new(str);
4478
4479         if (list == NULL)
4480                 return -1;
4481
4482         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4483         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4484
4485         if (trace->filter_pids.entries == NULL)
4486                 goto out;
4487
4488         trace->filter_pids.entries[0] = getpid();
4489
4490         for (i = 1; i < trace->filter_pids.nr; ++i)
4491                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4492
4493         intlist__delete(list);
4494         ret = 0;
4495 out:
4496         return ret;
4497 }
4498
4499 static int trace__open_output(struct trace *trace, const char *filename)
4500 {
4501         struct stat st;
4502
4503         if (!stat(filename, &st) && st.st_size) {
4504                 char oldname[PATH_MAX];
4505
4506                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4507                 unlink(oldname);
4508                 rename(filename, oldname);
4509         }
4510
4511         trace->output = fopen(filename, "w");
4512
4513         return trace->output == NULL ? -errno : 0;
4514 }
4515
4516 static int parse_pagefaults(const struct option *opt, const char *str,
4517                             int unset __maybe_unused)
4518 {
4519         int *trace_pgfaults = opt->value;
4520
4521         if (strcmp(str, "all") == 0)
4522                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4523         else if (strcmp(str, "maj") == 0)
4524                 *trace_pgfaults |= TRACE_PFMAJ;
4525         else if (strcmp(str, "min") == 0)
4526                 *trace_pgfaults |= TRACE_PFMIN;
4527         else
4528                 return -1;
4529
4530         return 0;
4531 }
4532
4533 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4534 {
4535         struct evsel *evsel;
4536
4537         evlist__for_each_entry(evlist, evsel) {
4538                 if (evsel->handler == NULL)
4539                         evsel->handler = handler;
4540         }
4541 }
4542
4543 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4544 {
4545         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4546
4547         if (fmt) {
4548                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4549
4550                 if (scfmt) {
4551                         int skip = 0;
4552
4553                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4554                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4555                                 ++skip;
4556
4557                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4558                 }
4559         }
4560 }
4561
4562 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4563 {
4564         struct evsel *evsel;
4565
4566         evlist__for_each_entry(evlist, evsel) {
4567                 if (evsel->priv || !evsel->tp_format)
4568                         continue;
4569
4570                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4571                         evsel__init_tp_arg_scnprintf(evsel);
4572                         continue;
4573                 }
4574
4575                 if (evsel__init_syscall_tp(evsel))
4576                         return -1;
4577
4578                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4579                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4580
4581                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4582                                 return -1;
4583
4584                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4585                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4586                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4587
4588                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4589                                 return -1;
4590
4591                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4592                 }
4593         }
4594
4595         return 0;
4596 }
4597
4598 /*
4599  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4600  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4601  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4602  *
4603  * It'd be better to introduce a parse_options() variant that would return a
4604  * list with the terms it didn't match to an event...
4605  */
4606 static int trace__parse_events_option(const struct option *opt, const char *str,
4607                                       int unset __maybe_unused)
4608 {
4609         struct trace *trace = (struct trace *)opt->value;
4610         const char *s = str;
4611         char *sep = NULL, *lists[2] = { NULL, NULL, };
4612         int len = strlen(str) + 1, err = -1, list, idx;
4613         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4614         char group_name[PATH_MAX];
4615         struct syscall_fmt *fmt;
4616
4617         if (strace_groups_dir == NULL)
4618                 return -1;
4619
4620         if (*s == '!') {
4621                 ++s;
4622                 trace->not_ev_qualifier = true;
4623         }
4624
4625         while (1) {
4626                 if ((sep = strchr(s, ',')) != NULL)
4627                         *sep = '\0';
4628
4629                 list = 0;
4630                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4631                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4632                         list = 1;
4633                         goto do_concat;
4634                 }
4635
4636                 fmt = syscall_fmt__find_by_alias(s);
4637                 if (fmt != NULL) {
4638                         list = 1;
4639                         s = fmt->name;
4640                 } else {
4641                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4642                         if (access(group_name, R_OK) == 0)
4643                                 list = 1;
4644                 }
4645 do_concat:
4646                 if (lists[list]) {
4647                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4648                 } else {
4649                         lists[list] = malloc(len);
4650                         if (lists[list] == NULL)
4651                                 goto out;
4652                         strcpy(lists[list], s);
4653                 }
4654
4655                 if (!sep)
4656                         break;
4657
4658                 *sep = ',';
4659                 s = sep + 1;
4660         }
4661
4662         if (lists[1] != NULL) {
4663                 struct strlist_config slist_config = {
4664                         .dirname = strace_groups_dir,
4665                 };
4666
4667                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4668                 if (trace->ev_qualifier == NULL) {
4669                         fputs("Not enough memory to parse event qualifier", trace->output);
4670                         goto out;
4671                 }
4672
4673                 if (trace__validate_ev_qualifier(trace))
4674                         goto out;
4675                 trace->trace_syscalls = true;
4676         }
4677
4678         err = 0;
4679
4680         if (lists[0]) {
4681                 struct option o = {
4682                         .value = &trace->evlist,
4683                 };
4684                 err = parse_events_option(&o, lists[0], 0);
4685         }
4686 out:
4687         free(strace_groups_dir);
4688         free(lists[0]);
4689         free(lists[1]);
4690         if (sep)
4691                 *sep = ',';
4692
4693         return err;
4694 }
4695
4696 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4697 {
4698         struct trace *trace = opt->value;
4699
4700         if (!list_empty(&trace->evlist->core.entries)) {
4701                 struct option o = {
4702                         .value = &trace->evlist,
4703                 };
4704                 return parse_cgroups(&o, str, unset);
4705         }
4706         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4707
4708         return 0;
4709 }
4710
4711 static int trace__config(const char *var, const char *value, void *arg)
4712 {
4713         struct trace *trace = arg;
4714         int err = 0;
4715
4716         if (!strcmp(var, "trace.add_events")) {
4717                 trace->perfconfig_events = strdup(value);
4718                 if (trace->perfconfig_events == NULL) {
4719                         pr_err("Not enough memory for %s\n", "trace.add_events");
4720                         return -1;
4721                 }
4722         } else if (!strcmp(var, "trace.show_timestamp")) {
4723                 trace->show_tstamp = perf_config_bool(var, value);
4724         } else if (!strcmp(var, "trace.show_duration")) {
4725                 trace->show_duration = perf_config_bool(var, value);
4726         } else if (!strcmp(var, "trace.show_arg_names")) {
4727                 trace->show_arg_names = perf_config_bool(var, value);
4728                 if (!trace->show_arg_names)
4729                         trace->show_zeros = true;
4730         } else if (!strcmp(var, "trace.show_zeros")) {
4731                 bool new_show_zeros = perf_config_bool(var, value);
4732                 if (!trace->show_arg_names && !new_show_zeros) {
4733                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4734                         goto out;
4735                 }
4736                 trace->show_zeros = new_show_zeros;
4737         } else if (!strcmp(var, "trace.show_prefix")) {
4738                 trace->show_string_prefix = perf_config_bool(var, value);
4739         } else if (!strcmp(var, "trace.no_inherit")) {
4740                 trace->opts.no_inherit = perf_config_bool(var, value);
4741         } else if (!strcmp(var, "trace.args_alignment")) {
4742                 int args_alignment = 0;
4743                 if (perf_config_int(&args_alignment, var, value) == 0)
4744                         trace->args_alignment = args_alignment;
4745         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4746                 if (strcasecmp(value, "libtraceevent") == 0)
4747                         trace->libtraceevent_print = true;
4748                 else if (strcasecmp(value, "libbeauty") == 0)
4749                         trace->libtraceevent_print = false;
4750         }
4751 out:
4752         return err;
4753 }
4754
4755 static void trace__exit(struct trace *trace)
4756 {
4757         int i;
4758
4759         strlist__delete(trace->ev_qualifier);
4760         free(trace->ev_qualifier_ids.entries);
4761         if (trace->syscalls.table) {
4762                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4763                         syscall__exit(&trace->syscalls.table[i]);
4764                 free(trace->syscalls.table);
4765         }
4766         syscalltbl__delete(trace->sctbl);
4767         zfree(&trace->perfconfig_events);
4768 }
4769
4770 int cmd_trace(int argc, const char **argv)
4771 {
4772         const char *trace_usage[] = {
4773                 "perf trace [<options>] [<command>]",
4774                 "perf trace [<options>] -- <command> [<options>]",
4775                 "perf trace record [<options>] [<command>]",
4776                 "perf trace record [<options>] -- <command> [<options>]",
4777                 NULL
4778         };
4779         struct trace trace = {
4780                 .opts = {
4781                         .target = {
4782                                 .uid       = UINT_MAX,
4783                                 .uses_mmap = true,
4784                         },
4785                         .user_freq     = UINT_MAX,
4786                         .user_interval = ULLONG_MAX,
4787                         .no_buffering  = true,
4788                         .mmap_pages    = UINT_MAX,
4789                 },
4790                 .output = stderr,
4791                 .show_comm = true,
4792                 .show_tstamp = true,
4793                 .show_duration = true,
4794                 .show_arg_names = true,
4795                 .args_alignment = 70,
4796                 .trace_syscalls = false,
4797                 .kernel_syscallchains = false,
4798                 .max_stack = UINT_MAX,
4799                 .max_events = ULONG_MAX,
4800         };
4801         const char *map_dump_str = NULL;
4802         const char *output_name = NULL;
4803         const struct option trace_options[] = {
4804         OPT_CALLBACK('e', "event", &trace, "event",
4805                      "event/syscall selector. use 'perf list' to list available events",
4806                      trace__parse_events_option),
4807         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4808                      "event filter", parse_filter),
4809         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4810                     "show the thread COMM next to its id"),
4811         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4812         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4813                      trace__parse_events_option),
4814         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4815         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4816         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4817                     "trace events on existing process id"),
4818         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4819                     "trace events on existing thread id"),
4820         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4821                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4822         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4823                     "system-wide collection from all CPUs"),
4824         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4825                     "list of cpus to monitor"),
4826         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4827                     "child tasks do not inherit counters"),
4828         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4829                      "number of mmap data pages", evlist__parse_mmap_pages),
4830         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4831                    "user to profile"),
4832         OPT_CALLBACK(0, "duration", &trace, "float",
4833                      "show only events with duration > N.M ms",
4834                      trace__set_duration),
4835 #ifdef HAVE_LIBBPF_SUPPORT
4836         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4837 #endif
4838         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4839         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4840         OPT_BOOLEAN('T', "time", &trace.full_time,
4841                     "Show full timestamp, not time relative to first start"),
4842         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4843                     "Show only syscalls that failed"),
4844         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4845                     "Show only syscall summary with statistics"),
4846         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4847                     "Show all syscalls and summary with statistics"),
4848         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4849                     "Show errno stats per syscall, use with -s or -S"),
4850         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4851                      "Trace pagefaults", parse_pagefaults, "maj"),
4852         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4853         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4854         OPT_CALLBACK(0, "call-graph", &trace.opts,
4855                      "record_mode[,record_size]", record_callchain_help,
4856                      &record_parse_callchain_opt),
4857         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4858                     "Use libtraceevent to print the tracepoint arguments."),
4859         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4860                     "Show the kernel callchains on the syscall exit path"),
4861         OPT_ULONG(0, "max-events", &trace.max_events,
4862                 "Set the maximum number of events to print, exit after that is reached. "),
4863         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4864                      "Set the minimum stack depth when parsing the callchain, "
4865                      "anything below the specified depth will be ignored."),
4866         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4867                      "Set the maximum stack depth when parsing the callchain, "
4868                      "anything beyond the specified depth will be ignored. "
4869                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4870         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4871                         "Sort batch of events before processing, use if getting out of order events"),
4872         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4873                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4874         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4875                         "per thread proc mmap processing timeout in ms"),
4876         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4877                      trace__parse_cgroups),
4878         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4879                      "ms to wait before starting measurement after program "
4880                      "start"),
4881         OPTS_EVSWITCH(&trace.evswitch),
4882         OPT_END()
4883         };
4884         bool __maybe_unused max_stack_user_set = true;
4885         bool mmap_pages_user_set = true;
4886         struct evsel *evsel;
4887         const char * const trace_subcommands[] = { "record", NULL };
4888         int err = -1;
4889         char bf[BUFSIZ];
4890         struct sigaction sigchld_act;
4891
4892         signal(SIGSEGV, sighandler_dump_stack);
4893         signal(SIGFPE, sighandler_dump_stack);
4894         signal(SIGINT, sighandler_interrupt);
4895
4896         memset(&sigchld_act, 0, sizeof(sigchld_act));
4897         sigchld_act.sa_flags = SA_SIGINFO;
4898         sigchld_act.sa_sigaction = sighandler_chld;
4899         sigaction(SIGCHLD, &sigchld_act, NULL);
4900
4901         trace.evlist = evlist__new();
4902         trace.sctbl = syscalltbl__new();
4903
4904         if (trace.evlist == NULL || trace.sctbl == NULL) {
4905                 pr_err("Not enough memory to run!\n");
4906                 err = -ENOMEM;
4907                 goto out;
4908         }
4909
4910         /*
4911          * Parsing .perfconfig may entail creating a BPF event, that may need
4912          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4913          * is too small. This affects just this process, not touching the
4914          * global setting. If it fails we'll get something in 'perf trace -v'
4915          * to help diagnose the problem.
4916          */
4917         rlimit__bump_memlock();
4918
4919         err = perf_config(trace__config, &trace);
4920         if (err)
4921                 goto out;
4922
4923         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4924                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4925
4926         /*
4927          * Here we already passed thru trace__parse_events_option() and it has
4928          * already figured out if -e syscall_name, if not but if --event
4929          * foo:bar was used, the user is interested _just_ in those, say,
4930          * tracepoint events, not in the strace-like syscall-name-based mode.
4931          *
4932          * This is important because we need to check if strace-like mode is
4933          * needed to decided if we should filter out the eBPF
4934          * __augmented_syscalls__ code, if it is in the mix, say, via
4935          * .perfconfig trace.add_events, and filter those out.
4936          */
4937         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4938             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4939                 trace.trace_syscalls = true;
4940         }
4941         /*
4942          * Now that we have --verbose figured out, lets see if we need to parse
4943          * events from .perfconfig, so that if those events fail parsing, say some
4944          * BPF program fails, then we'll be able to use --verbose to see what went
4945          * wrong in more detail.
4946          */
4947         if (trace.perfconfig_events != NULL) {
4948                 struct parse_events_error parse_err;
4949
4950                 parse_events_error__init(&parse_err);
4951                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4952                 if (err)
4953                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4954                 parse_events_error__exit(&parse_err);
4955                 if (err)
4956                         goto out;
4957         }
4958
4959         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4960                 usage_with_options_msg(trace_usage, trace_options,
4961                                        "cgroup monitoring only available in system-wide mode");
4962         }
4963
4964         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4965         if (IS_ERR(evsel)) {
4966                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4967                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4968                 goto out;
4969         }
4970
4971         if (evsel) {
4972                 trace.syscalls.events.augmented = evsel;
4973
4974                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4975                 if (evsel == NULL) {
4976                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4977                         goto out;
4978                 }
4979
4980                 if (evsel->bpf_obj == NULL) {
4981                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4982                         goto out;
4983                 }
4984
4985                 trace.bpf_obj = evsel->bpf_obj;
4986
4987                 /*
4988                  * If we have _just_ the augmenter event but don't have a
4989                  * explicit --syscalls, then assume we want all strace-like
4990                  * syscalls:
4991                  */
4992                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4993                         trace.trace_syscalls = true;
4994                 /*
4995                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4996                  * strace-like syscall tracing is not set, then we need to trow
4997                  * away the augmenter, i.e. all the events that were created
4998                  * from that BPF object file.
4999                  *
5000                  * This is more to fix the current .perfconfig trace.add_events
5001                  * style of setting up the strace-like eBPF based syscall point
5002                  * payload augmenter.
5003                  *
5004                  * All this complexity will be avoided by adding an alternative
5005                  * to trace.add_events in the form of
5006                  * trace.bpf_augmented_syscalls, that will be only parsed if we
5007                  * need it.
5008                  *
5009                  * .perfconfig trace.add_events is still useful if we want, for
5010                  * instance, have msr_write.msr in some .perfconfig profile based
5011                  * 'perf trace --config determinism.profile' mode, where for some
5012                  * particular goal/workload type we want a set of events and
5013                  * output mode (with timings, etc) instead of having to add
5014                  * all via the command line.
5015                  *
5016                  * Also --config to specify an alternate .perfconfig file needs
5017                  * to be implemented.
5018                  */
5019                 if (!trace.trace_syscalls) {
5020                         trace__delete_augmented_syscalls(&trace);
5021                 } else {
5022                         trace__set_bpf_map_filtered_pids(&trace);
5023                         trace__set_bpf_map_syscalls(&trace);
5024                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
5025                 }
5026         }
5027
5028         err = bpf__setup_stdout(trace.evlist);
5029         if (err) {
5030                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
5031                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
5032                 goto out;
5033         }
5034
5035         err = -1;
5036
5037         if (map_dump_str) {
5038                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
5039                 if (trace.dump.map == NULL) {
5040                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
5041                         goto out;
5042                 }
5043         }
5044
5045         if (trace.trace_pgfaults) {
5046                 trace.opts.sample_address = true;
5047                 trace.opts.sample_time = true;
5048         }
5049
5050         if (trace.opts.mmap_pages == UINT_MAX)
5051                 mmap_pages_user_set = false;
5052
5053         if (trace.max_stack == UINT_MAX) {
5054                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
5055                 max_stack_user_set = false;
5056         }
5057
5058 #ifdef HAVE_DWARF_UNWIND_SUPPORT
5059         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
5060                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
5061         }
5062 #endif
5063
5064         if (callchain_param.enabled) {
5065                 if (!mmap_pages_user_set && geteuid() == 0)
5066                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5067
5068                 symbol_conf.use_callchain = true;
5069         }
5070
5071         if (trace.evlist->core.nr_entries > 0) {
5072                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5073                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5074                         perror("failed to set syscalls:* tracepoint fields");
5075                         goto out;
5076                 }
5077         }
5078
5079         if (trace.sort_events) {
5080                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5081                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5082         }
5083
5084         /*
5085          * If we are augmenting syscalls, then combine what we put in the
5086          * __augmented_syscalls__ BPF map with what is in the
5087          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5088          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5089          *
5090          * We'll switch to look at two BPF maps, one for sys_enter and the
5091          * other for sys_exit when we start augmenting the sys_exit paths with
5092          * buffers that are being copied from kernel to userspace, think 'read'
5093          * syscall.
5094          */
5095         if (trace.syscalls.events.augmented) {
5096                 evlist__for_each_entry(trace.evlist, evsel) {
5097                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5098
5099                         if (raw_syscalls_sys_exit) {
5100                                 trace.raw_augmented_syscalls = true;
5101                                 goto init_augmented_syscall_tp;
5102                         }
5103
5104                         if (trace.syscalls.events.augmented->priv == NULL &&
5105                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5106                                 struct evsel *augmented = trace.syscalls.events.augmented;
5107                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5108                                     evsel__init_augmented_syscall_tp_args(augmented))
5109                                         goto out;
5110                                 /*
5111                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5112                                  * Above we made sure we can get from the payload the tp fields
5113                                  * that we get from syscalls:sys_enter tracefs format file.
5114                                  */
5115                                 augmented->handler = trace__sys_enter;
5116                                 /*
5117                                  * Now we do the same for the *syscalls:sys_enter event so that
5118                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5119                                  * as not to filter it, then we'll handle it just like we would
5120                                  * for the BPF_OUTPUT one:
5121                                  */
5122                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5123                                     evsel__init_augmented_syscall_tp_args(evsel))
5124                                         goto out;
5125                                 evsel->handler = trace__sys_enter;
5126                         }
5127
5128                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5129                                 struct syscall_tp *sc;
5130 init_augmented_syscall_tp:
5131                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5132                                         goto out;
5133                                 sc = __evsel__syscall_tp(evsel);
5134                                 /*
5135                                  * For now with BPF raw_augmented we hook into
5136                                  * raw_syscalls:sys_enter and there we get all
5137                                  * 6 syscall args plus the tracepoint common
5138                                  * fields and the syscall_nr (another long).
5139                                  * So we check if that is the case and if so
5140                                  * don't look after the sc->args_size but
5141                                  * always after the full raw_syscalls:sys_enter
5142                                  * payload, which is fixed.
5143                                  *
5144                                  * We'll revisit this later to pass
5145                                  * s->args_size to the BPF augmenter (now
5146                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5147                                  * so that it copies only what we need for each
5148                                  * syscall, like what happens when we use
5149                                  * syscalls:sys_enter_NAME, so that we reduce
5150                                  * the kernel/userspace traffic to just what is
5151                                  * needed for each syscall.
5152                                  */
5153                                 if (trace.raw_augmented_syscalls)
5154                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5155                                 evsel__init_augmented_syscall_tp_ret(evsel);
5156                                 evsel->handler = trace__sys_exit;
5157                         }
5158                 }
5159         }
5160
5161         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5162                 return trace__record(&trace, argc-1, &argv[1]);
5163
5164         /* Using just --errno-summary will trigger --summary */
5165         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5166                 trace.summary_only = true;
5167
5168         /* summary_only implies summary option, but don't overwrite summary if set */
5169         if (trace.summary_only)
5170                 trace.summary = trace.summary_only;
5171
5172         if (output_name != NULL) {
5173                 err = trace__open_output(&trace, output_name);
5174                 if (err < 0) {
5175                         perror("failed to create output file");
5176                         goto out;
5177                 }
5178         }
5179
5180         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5181         if (err)
5182                 goto out_close;
5183
5184         err = target__validate(&trace.opts.target);
5185         if (err) {
5186                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5187                 fprintf(trace.output, "%s", bf);
5188                 goto out_close;
5189         }
5190
5191         err = target__parse_uid(&trace.opts.target);
5192         if (err) {
5193                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5194                 fprintf(trace.output, "%s", bf);
5195                 goto out_close;
5196         }
5197
5198         if (!argc && target__none(&trace.opts.target))
5199                 trace.opts.target.system_wide = true;
5200
5201         if (input_name)
5202                 err = trace__replay(&trace);
5203         else
5204                 err = trace__run(&trace, argc, argv);
5205
5206 out_close:
5207         if (output_name != NULL)
5208                 fclose(trace.output);
5209 out:
5210         trace__exit(&trace);
5211         return err;
5212 }