tools/perf/builtin-trace.c

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