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 #endif
  22 #include "util/bpf_map.h"
  23 #include "util/rlimit.h"
  24 #include "builtin.h"
  25 #include "util/cgroup.h"
  26 #include "util/color.h"
  27 #include "util/config.h"
  28 #include "util/debug.h"
  29 #include "util/dso.h"
  30 #include "util/env.h"
  31 #include "util/event.h"
  32 #include "util/evsel.h"
  33 #include "util/evsel_fprintf.h"
  34 #include "util/synthetic-events.h"
  35 #include "util/evlist.h"
  36 #include "util/evswitch.h"
  37 #include "util/mmap.h"
  38 #include <subcmd/pager.h>
  39 #include <subcmd/exec-cmd.h>
  40 #include "util/machine.h"
  41 #include "util/map.h"
  42 #include "util/symbol.h"
  43 #include "util/path.h"
  44 #include "util/session.h"
  45 #include "util/thread.h"
  46 #include <subcmd/parse-options.h>
  47 #include "util/strlist.h"
  48 #include "util/intlist.h"
  49 #include "util/thread_map.h"
  50 #include "util/stat.h"
  51 #include "util/tool.h"
  52 #include "util/util.h"
  53 #include "trace/beauty/beauty.h"
  54 #include "trace-event.h"
  55 #include "util/parse-events.h"
  56 #include "util/bpf-loader.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 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 struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1180 {
1181         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1182 }
1183
1184 static struct syscall_fmt *syscall_fmt__find(const char *name)
1185 {
1186         const int nmemb = ARRAY_SIZE(syscall_fmts);
1187         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1188 }
1189
1190 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1191 {
1192         int i;
1193
1194         for (i = 0; i < nmemb; ++i) {
1195                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1196                         return &fmts[i];
1197         }
1198
1199         return NULL;
1200 }
1201
1202 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1203 {
1204         const int nmemb = ARRAY_SIZE(syscall_fmts);
1205         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1206 }
1207
1208 /*
1209  * is_exit: is this "exit" or "exit_group"?
1210  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1211  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1212  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1213  */
1214 struct syscall {
1215         struct tep_event    *tp_format;
1216         int                 nr_args;
1217         int                 args_size;
1218         struct {
1219                 struct bpf_program *sys_enter,
1220                                    *sys_exit;
1221         }                   bpf_prog;
1222         bool                is_exit;
1223         bool                is_open;
1224         bool                nonexistent;
1225         struct tep_format_field *args;
1226         const char          *name;
1227         struct syscall_fmt  *fmt;
1228         struct syscall_arg_fmt *arg_fmt;
1229 };
1230
1231 /*
1232  * We need to have this 'calculated' boolean because in some cases we really
1233  * don't know what is the duration of a syscall, for instance, when we start
1234  * a session and some threads are waiting for a syscall to finish, say 'poll',
1235  * in which case all we can do is to print "( ? ) for duration and for the
1236  * start timestamp.
1237  */
1238 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1239 {
1240         double duration = (double)t / NSEC_PER_MSEC;
1241         size_t printed = fprintf(fp, "(");
1242
1243         if (!calculated)
1244                 printed += fprintf(fp, "         ");
1245         else if (duration >= 1.0)
1246                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1247         else if (duration >= 0.01)
1248                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1249         else
1250                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1251         return printed + fprintf(fp, "): ");
1252 }
1253
1254 /**
1255  * filename.ptr: The filename char pointer that will be vfs_getname'd
1256  * filename.entry_str_pos: Where to insert the string translated from
1257  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1258  * ret_scnprintf: syscall args may set this to a different syscall return
1259  *                formatter, for instance, fcntl may return fds, file flags, etc.
1260  */
1261 struct thread_trace {
1262         u64               entry_time;
1263         bool              entry_pending;
1264         unsigned long     nr_events;
1265         unsigned long     pfmaj, pfmin;
1266         char              *entry_str;
1267         double            runtime_ms;
1268         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1269         struct {
1270                 unsigned long ptr;
1271                 short int     entry_str_pos;
1272                 bool          pending_open;
1273                 unsigned int  namelen;
1274                 char          *name;
1275         } filename;
1276         struct {
1277                 int           max;
1278                 struct file   *table;
1279         } files;
1280
1281         struct intlist *syscall_stats;
1282 };
1283
1284 static struct thread_trace *thread_trace__new(void)
1285 {
1286         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1287
1288         if (ttrace) {
1289                 ttrace->files.max = -1;
1290                 ttrace->syscall_stats = intlist__new(NULL);
1291         }
1292
1293         return ttrace;
1294 }
1295
1296 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1297 {
1298         struct thread_trace *ttrace;
1299
1300         if (thread == NULL)
1301                 goto fail;
1302
1303         if (thread__priv(thread) == NULL)
1304                 thread__set_priv(thread, thread_trace__new());
1305
1306         if (thread__priv(thread) == NULL)
1307                 goto fail;
1308
1309         ttrace = thread__priv(thread);
1310         ++ttrace->nr_events;
1311
1312         return ttrace;
1313 fail:
1314         color_fprintf(fp, PERF_COLOR_RED,
1315                       "WARNING: not enough memory, dropping samples!\n");
1316         return NULL;
1317 }
1318
1319
1320 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1321                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1322 {
1323         struct thread_trace *ttrace = thread__priv(arg->thread);
1324
1325         ttrace->ret_scnprintf = ret_scnprintf;
1326 }
1327
1328 #define TRACE_PFMAJ             (1 << 0)
1329 #define TRACE_PFMIN             (1 << 1)
1330
1331 static const size_t trace__entry_str_size = 2048;
1332
1333 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1334 {
1335         if (fd < 0)
1336                 return NULL;
1337
1338         if (fd > ttrace->files.max) {
1339                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1340
1341                 if (nfiles == NULL)
1342                         return NULL;
1343
1344                 if (ttrace->files.max != -1) {
1345                         memset(nfiles + ttrace->files.max + 1, 0,
1346                                (fd - ttrace->files.max) * sizeof(struct file));
1347                 } else {
1348                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1349                 }
1350
1351                 ttrace->files.table = nfiles;
1352                 ttrace->files.max   = fd;
1353         }
1354
1355         return ttrace->files.table + fd;
1356 }
1357
1358 struct file *thread__files_entry(struct thread *thread, int fd)
1359 {
1360         return thread_trace__files_entry(thread__priv(thread), fd);
1361 }
1362
1363 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1364 {
1365         struct thread_trace *ttrace = thread__priv(thread);
1366         struct file *file = thread_trace__files_entry(ttrace, fd);
1367
1368         if (file != NULL) {
1369                 struct stat st;
1370                 if (stat(pathname, &st) == 0)
1371                         file->dev_maj = major(st.st_rdev);
1372                 file->pathname = strdup(pathname);
1373                 if (file->pathname)
1374                         return 0;
1375         }
1376
1377         return -1;
1378 }
1379
1380 static int thread__read_fd_path(struct thread *thread, int fd)
1381 {
1382         char linkname[PATH_MAX], pathname[PATH_MAX];
1383         struct stat st;
1384         int ret;
1385
1386         if (thread->pid_ == thread->tid) {
1387                 scnprintf(linkname, sizeof(linkname),
1388                           "/proc/%d/fd/%d", thread->pid_, fd);
1389         } else {
1390                 scnprintf(linkname, sizeof(linkname),
1391                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1392         }
1393
1394         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1395                 return -1;
1396
1397         ret = readlink(linkname, pathname, sizeof(pathname));
1398
1399         if (ret < 0 || ret > st.st_size)
1400                 return -1;
1401
1402         pathname[ret] = '\0';
1403         return trace__set_fd_pathname(thread, fd, pathname);
1404 }
1405
1406 static const char *thread__fd_path(struct thread *thread, int fd,
1407                                    struct trace *trace)
1408 {
1409         struct thread_trace *ttrace = thread__priv(thread);
1410
1411         if (ttrace == NULL || trace->fd_path_disabled)
1412                 return NULL;
1413
1414         if (fd < 0)
1415                 return NULL;
1416
1417         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1418                 if (!trace->live)
1419                         return NULL;
1420                 ++trace->stats.proc_getname;
1421                 if (thread__read_fd_path(thread, fd))
1422                         return NULL;
1423         }
1424
1425         return ttrace->files.table[fd].pathname;
1426 }
1427
1428 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1429 {
1430         int fd = arg->val;
1431         size_t printed = scnprintf(bf, size, "%d", fd);
1432         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1433
1434         if (path)
1435                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1436
1437         return printed;
1438 }
1439
1440 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1441 {
1442         size_t printed = scnprintf(bf, size, "%d", fd);
1443         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1444
1445         if (thread) {
1446                 const char *path = thread__fd_path(thread, fd, trace);
1447
1448                 if (path)
1449                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1450
1451                 thread__put(thread);
1452         }
1453
1454         return printed;
1455 }
1456
1457 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1458                                               struct syscall_arg *arg)
1459 {
1460         int fd = arg->val;
1461         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1462         struct thread_trace *ttrace = thread__priv(arg->thread);
1463
1464         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1465                 zfree(&ttrace->files.table[fd].pathname);
1466
1467         return printed;
1468 }
1469
1470 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1471                                      unsigned long ptr)
1472 {
1473         struct thread_trace *ttrace = thread__priv(thread);
1474
1475         ttrace->filename.ptr = ptr;
1476         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1477 }
1478
1479 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1480 {
1481         struct augmented_arg *augmented_arg = arg->augmented.args;
1482         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1483         /*
1484          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1485          * we would have two strings, each prefixed by its size.
1486          */
1487         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1488
1489         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1490         arg->augmented.size -= consumed;
1491
1492         return printed;
1493 }
1494
1495 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1496                                               struct syscall_arg *arg)
1497 {
1498         unsigned long ptr = arg->val;
1499
1500         if (arg->augmented.args)
1501                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1502
1503         if (!arg->trace->vfs_getname)
1504                 return scnprintf(bf, size, "%#x", ptr);
1505
1506         thread__set_filename_pos(arg->thread, bf, ptr);
1507         return 0;
1508 }
1509
1510 static bool trace__filter_duration(struct trace *trace, double t)
1511 {
1512         return t < (trace->duration_filter * NSEC_PER_MSEC);
1513 }
1514
1515 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1516 {
1517         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1518
1519         return fprintf(fp, "%10.3f ", ts);
1520 }
1521
1522 /*
1523  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1524  * using ttrace->entry_time for a thread that receives a sys_exit without
1525  * first having received a sys_enter ("poll" issued before tracing session
1526  * starts, lost sys_enter exit due to ring buffer overflow).
1527  */
1528 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1529 {
1530         if (tstamp > 0)
1531                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1532
1533         return fprintf(fp, "         ? ");
1534 }
1535
1536 static pid_t workload_pid = -1;
1537 static volatile sig_atomic_t done = false;
1538 static volatile sig_atomic_t interrupted = false;
1539
1540 static void sighandler_interrupt(int sig __maybe_unused)
1541 {
1542         done = interrupted = true;
1543 }
1544
1545 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1546                             void *context __maybe_unused)
1547 {
1548         if (info->si_pid == workload_pid)
1549                 done = true;
1550 }
1551
1552 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1553 {
1554         size_t printed = 0;
1555
1556         if (trace->multiple_threads) {
1557                 if (trace->show_comm)
1558                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1559                 printed += fprintf(fp, "%d ", thread->tid);
1560         }
1561
1562         return printed;
1563 }
1564
1565 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1566                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1567 {
1568         size_t printed = 0;
1569
1570         if (trace->show_tstamp)
1571                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1572         if (trace->show_duration)
1573                 printed += fprintf_duration(duration, duration_calculated, fp);
1574         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1575 }
1576
1577 static int trace__process_event(struct trace *trace, struct machine *machine,
1578                                 union perf_event *event, struct perf_sample *sample)
1579 {
1580         int ret = 0;
1581
1582         switch (event->header.type) {
1583         case PERF_RECORD_LOST:
1584                 color_fprintf(trace->output, PERF_COLOR_RED,
1585                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1586                 ret = machine__process_lost_event(machine, event, sample);
1587                 break;
1588         default:
1589                 ret = machine__process_event(machine, event, sample);
1590                 break;
1591         }
1592
1593         return ret;
1594 }
1595
1596 static int trace__tool_process(struct perf_tool *tool,
1597                                union perf_event *event,
1598                                struct perf_sample *sample,
1599                                struct machine *machine)
1600 {
1601         struct trace *trace = container_of(tool, struct trace, tool);
1602         return trace__process_event(trace, machine, event, sample);
1603 }
1604
1605 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1606 {
1607         struct machine *machine = vmachine;
1608
1609         if (machine->kptr_restrict_warned)
1610                 return NULL;
1611
1612         if (symbol_conf.kptr_restrict) {
1613                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1614                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1615                            "Kernel samples will not be resolved.\n");
1616                 machine->kptr_restrict_warned = true;
1617                 return NULL;
1618         }
1619
1620         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1621 }
1622
1623 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1624 {
1625         int err = symbol__init(NULL);
1626
1627         if (err)
1628                 return err;
1629
1630         trace->host = machine__new_host();
1631         if (trace->host == NULL)
1632                 return -ENOMEM;
1633
1634         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1635         if (err < 0)
1636                 goto out;
1637
1638         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1639                                             evlist->core.threads, trace__tool_process,
1640                                             true, false, 1);
1641 out:
1642         if (err)
1643                 symbol__exit();
1644
1645         return err;
1646 }
1647
1648 static void trace__symbols__exit(struct trace *trace)
1649 {
1650         machine__exit(trace->host);
1651         trace->host = NULL;
1652
1653         symbol__exit();
1654 }
1655
1656 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1657 {
1658         int idx;
1659
1660         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1661                 nr_args = sc->fmt->nr_args;
1662
1663         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1664         if (sc->arg_fmt == NULL)
1665                 return -1;
1666
1667         for (idx = 0; idx < nr_args; ++idx) {
1668                 if (sc->fmt)
1669                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1670         }
1671
1672         sc->nr_args = nr_args;
1673         return 0;
1674 }
1675
1676 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1677         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1678         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1679 };
1680
1681 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1682 {
1683        const struct syscall_arg_fmt *fmt = fmtp;
1684        return strcmp(name, fmt->name);
1685 }
1686
1687 static struct syscall_arg_fmt *
1688 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1689 {
1690        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1691 }
1692
1693 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1694 {
1695        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1696        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1697 }
1698
1699 static struct tep_format_field *
1700 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1701 {
1702         struct tep_format_field *last_field = NULL;
1703         int len;
1704
1705         for (; field; field = field->next, ++arg) {
1706                 last_field = field;
1707
1708                 if (arg->scnprintf)
1709                         continue;
1710
1711                 len = strlen(field->name);
1712
1713                 if (strcmp(field->type, "const char *") == 0 &&
1714                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1715                      strstr(field->name, "path") != NULL))
1716                         arg->scnprintf = SCA_FILENAME;
1717                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1718                         arg->scnprintf = SCA_PTR;
1719                 else if (strcmp(field->type, "pid_t") == 0)
1720                         arg->scnprintf = SCA_PID;
1721                 else if (strcmp(field->type, "umode_t") == 0)
1722                         arg->scnprintf = SCA_MODE_T;
1723                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1724                         arg->scnprintf = SCA_CHAR_ARRAY;
1725                         arg->nr_entries = field->arraylen;
1726                 } else if ((strcmp(field->type, "int") == 0 ||
1727                           strcmp(field->type, "unsigned int") == 0 ||
1728                           strcmp(field->type, "long") == 0) &&
1729                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1730                         /*
1731                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1732                          * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1733                          * 65 int
1734                          * 23 unsigned int
1735                          * 7 unsigned long
1736                          */
1737                         arg->scnprintf = SCA_FD;
1738                } else {
1739                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1740
1741                         if (fmt) {
1742                                 arg->scnprintf = fmt->scnprintf;
1743                                 arg->strtoul   = fmt->strtoul;
1744                         }
1745                 }
1746         }
1747
1748         return last_field;
1749 }
1750
1751 static int syscall__set_arg_fmts(struct syscall *sc)
1752 {
1753         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1754
1755         if (last_field)
1756                 sc->args_size = last_field->offset + last_field->size;
1757
1758         return 0;
1759 }
1760
1761 static int trace__read_syscall_info(struct trace *trace, int id)
1762 {
1763         char tp_name[128];
1764         struct syscall *sc;
1765         const char *name = syscalltbl__name(trace->sctbl, id);
1766
1767 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1768         if (trace->syscalls.table == NULL) {
1769                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1770                 if (trace->syscalls.table == NULL)
1771                         return -ENOMEM;
1772         }
1773 #else
1774         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1775                 // When using libaudit we don't know beforehand what is the max syscall id
1776                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1777
1778                 if (table == NULL)
1779                         return -ENOMEM;
1780
1781                 // Need to memset from offset 0 and +1 members if brand new
1782                 if (trace->syscalls.table == NULL)
1783                         memset(table, 0, (id + 1) * sizeof(*sc));
1784                 else
1785                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1786
1787                 trace->syscalls.table         = table;
1788                 trace->sctbl->syscalls.max_id = id;
1789         }
1790 #endif
1791         sc = trace->syscalls.table + id;
1792         if (sc->nonexistent)
1793                 return -EEXIST;
1794
1795         if (name == NULL) {
1796                 sc->nonexistent = true;
1797                 return -EEXIST;
1798         }
1799
1800         sc->name = name;
1801         sc->fmt  = syscall_fmt__find(sc->name);
1802
1803         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1804         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1805
1806         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1807                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1808                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1809         }
1810
1811         /*
1812          * Fails to read trace point format via sysfs node, so the trace point
1813          * doesn't exist.  Set the 'nonexistent' flag as true.
1814          */
1815         if (IS_ERR(sc->tp_format)) {
1816                 sc->nonexistent = true;
1817                 return PTR_ERR(sc->tp_format);
1818         }
1819
1820         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1821                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1822                 return -ENOMEM;
1823
1824         sc->args = sc->tp_format->format.fields;
1825         /*
1826          * We need to check and discard the first variable '__syscall_nr'
1827          * or 'nr' that mean the syscall number. It is needless here.
1828          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1829          */
1830         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1831                 sc->args = sc->args->next;
1832                 --sc->nr_args;
1833         }
1834
1835         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1836         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1837
1838         return syscall__set_arg_fmts(sc);
1839 }
1840
1841 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1842 {
1843         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1844
1845         if (fmt != NULL) {
1846                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1847                 return 0;
1848         }
1849
1850         return -ENOMEM;
1851 }
1852
1853 static int intcmp(const void *a, const void *b)
1854 {
1855         const int *one = a, *another = b;
1856
1857         return *one - *another;
1858 }
1859
1860 static int trace__validate_ev_qualifier(struct trace *trace)
1861 {
1862         int err = 0;
1863         bool printed_invalid_prefix = false;
1864         struct str_node *pos;
1865         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1866
1867         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1868                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1869
1870         if (trace->ev_qualifier_ids.entries == NULL) {
1871                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1872                        trace->output);
1873                 err = -EINVAL;
1874                 goto out;
1875         }
1876
1877         strlist__for_each_entry(pos, trace->ev_qualifier) {
1878                 const char *sc = pos->s;
1879                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1880
1881                 if (id < 0) {
1882                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1883                         if (id >= 0)
1884                                 goto matches;
1885
1886                         if (!printed_invalid_prefix) {
1887                                 pr_debug("Skipping unknown syscalls: ");
1888                                 printed_invalid_prefix = true;
1889                         } else {
1890                                 pr_debug(", ");
1891                         }
1892
1893                         pr_debug("%s", sc);
1894                         continue;
1895                 }
1896 matches:
1897                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1898                 if (match_next == -1)
1899                         continue;
1900
1901                 while (1) {
1902                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1903                         if (id < 0)
1904                                 break;
1905                         if (nr_allocated == nr_used) {
1906                                 void *entries;
1907
1908                                 nr_allocated += 8;
1909                                 entries = realloc(trace->ev_qualifier_ids.entries,
1910                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1911                                 if (entries == NULL) {
1912                                         err = -ENOMEM;
1913                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1914                                         goto out_free;
1915                                 }
1916                                 trace->ev_qualifier_ids.entries = entries;
1917                         }
1918                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1919                 }
1920         }
1921
1922         trace->ev_qualifier_ids.nr = nr_used;
1923         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1924 out:
1925         if (printed_invalid_prefix)
1926                 pr_debug("\n");
1927         return err;
1928 out_free:
1929         zfree(&trace->ev_qualifier_ids.entries);
1930         trace->ev_qualifier_ids.nr = 0;
1931         goto out;
1932 }
1933
1934 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1935 {
1936         bool in_ev_qualifier;
1937
1938         if (trace->ev_qualifier_ids.nr == 0)
1939                 return true;
1940
1941         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1942                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1943
1944         if (in_ev_qualifier)
1945                return !trace->not_ev_qualifier;
1946
1947         return trace->not_ev_qualifier;
1948 }
1949
1950 /*
1951  * args is to be interpreted as a series of longs but we need to handle
1952  * 8-byte unaligned accesses. args points to raw_data within the event
1953  * and raw_data is guaranteed to be 8-byte unaligned because it is
1954  * preceded by raw_size which is a u32. So we need to copy args to a temp
1955  * variable to read it. Most notably this avoids extended load instructions
1956  * on unaligned addresses
1957  */
1958 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1959 {
1960         unsigned long val;
1961         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1962
1963         memcpy(&val, p, sizeof(val));
1964         return val;
1965 }
1966
1967 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1968                                       struct syscall_arg *arg)
1969 {
1970         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1971                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1972
1973         return scnprintf(bf, size, "arg%d: ", arg->idx);
1974 }
1975
1976 /*
1977  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1978  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1979  * in tools/perf/trace/beauty/mount_flags.c
1980  */
1981 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1982 {
1983         if (fmt && fmt->mask_val)
1984                 return fmt->mask_val(arg, val);
1985
1986         return val;
1987 }
1988
1989 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1990                                              struct syscall_arg *arg, unsigned long val)
1991 {
1992         if (fmt && fmt->scnprintf) {
1993                 arg->val = val;
1994                 if (fmt->parm)
1995                         arg->parm = fmt->parm;
1996                 return fmt->scnprintf(bf, size, arg);
1997         }
1998         return scnprintf(bf, size, "%ld", val);
1999 }
2000
2001 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2002                                       unsigned char *args, void *augmented_args, int augmented_args_size,
2003                                       struct trace *trace, struct thread *thread)
2004 {
2005         size_t printed = 0;
2006         unsigned long val;
2007         u8 bit = 1;
2008         struct syscall_arg arg = {
2009                 .args   = args,
2010                 .augmented = {
2011                         .size = augmented_args_size,
2012                         .args = augmented_args,
2013                 },
2014                 .idx    = 0,
2015                 .mask   = 0,
2016                 .trace  = trace,
2017                 .thread = thread,
2018                 .show_string_prefix = trace->show_string_prefix,
2019         };
2020         struct thread_trace *ttrace = thread__priv(thread);
2021
2022         /*
2023          * Things like fcntl will set this in its 'cmd' formatter to pick the
2024          * right formatter for the return value (an fd? file flags?), which is
2025          * not needed for syscalls that always return a given type, say an fd.
2026          */
2027         ttrace->ret_scnprintf = NULL;
2028
2029         if (sc->args != NULL) {
2030                 struct tep_format_field *field;
2031
2032                 for (field = sc->args; field;
2033                      field = field->next, ++arg.idx, bit <<= 1) {
2034                         if (arg.mask & bit)
2035                                 continue;
2036
2037                         arg.fmt = &sc->arg_fmt[arg.idx];
2038                         val = syscall_arg__val(&arg, arg.idx);
2039                         /*
2040                          * Some syscall args need some mask, most don't and
2041                          * return val untouched.
2042                          */
2043                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2044
2045                         /*
2046                          * Suppress this argument if its value is zero and
2047                          * and we don't have a string associated in an
2048                          * strarray for it.
2049                          */
2050                         if (val == 0 &&
2051                             !trace->show_zeros &&
2052                             !(sc->arg_fmt &&
2053                               (sc->arg_fmt[arg.idx].show_zero ||
2054                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2055                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2056                               sc->arg_fmt[arg.idx].parm))
2057                                 continue;
2058
2059                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2060
2061                         if (trace->show_arg_names)
2062                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2063
2064                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2065                                                                   bf + printed, size - printed, &arg, val);
2066                 }
2067         } else if (IS_ERR(sc->tp_format)) {
2068                 /*
2069                  * If we managed to read the tracepoint /format file, then we
2070                  * may end up not having any args, like with gettid(), so only
2071                  * print the raw args when we didn't manage to read it.
2072                  */
2073                 while (arg.idx < sc->nr_args) {
2074                         if (arg.mask & bit)
2075                                 goto next_arg;
2076                         val = syscall_arg__val(&arg, arg.idx);
2077                         if (printed)
2078                                 printed += scnprintf(bf + printed, size - printed, ", ");
2079                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2080                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2081 next_arg:
2082                         ++arg.idx;
2083                         bit <<= 1;
2084                 }
2085         }
2086
2087         return printed;
2088 }
2089
2090 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2091                                   union perf_event *event,
2092                                   struct perf_sample *sample);
2093
2094 static struct syscall *trace__syscall_info(struct trace *trace,
2095                                            struct evsel *evsel, int id)
2096 {
2097         int err = 0;
2098
2099         if (id < 0) {
2100
2101                 /*
2102                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2103                  * before that, leaving at a higher verbosity level till that is
2104                  * explained. Reproduced with plain ftrace with:
2105                  *
2106                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2107                  * grep "NR -1 " /t/trace_pipe
2108                  *
2109                  * After generating some load on the machine.
2110                  */
2111                 if (verbose > 1) {
2112                         static u64 n;
2113                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2114                                 id, evsel__name(evsel), ++n);
2115                 }
2116                 return NULL;
2117         }
2118
2119         err = -EINVAL;
2120
2121 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2122         if (id > trace->sctbl->syscalls.max_id) {
2123 #else
2124         if (id >= trace->sctbl->syscalls.max_id) {
2125                 /*
2126                  * With libaudit we don't know beforehand what is the max_id,
2127                  * so we let trace__read_syscall_info() figure that out as we
2128                  * go on reading syscalls.
2129                  */
2130                 err = trace__read_syscall_info(trace, id);
2131                 if (err)
2132 #endif
2133                 goto out_cant_read;
2134         }
2135
2136         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2137             (err = trace__read_syscall_info(trace, id)) != 0)
2138                 goto out_cant_read;
2139
2140         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2141                 goto out_cant_read;
2142
2143         return &trace->syscalls.table[id];
2144
2145 out_cant_read:
2146         if (verbose > 0) {
2147                 char sbuf[STRERR_BUFSIZE];
2148                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2149                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2150                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2151                 fputs(" information\n", trace->output);
2152         }
2153         return NULL;
2154 }
2155
2156 struct syscall_stats {
2157         struct stats stats;
2158         u64          nr_failures;
2159         int          max_errno;
2160         u32          *errnos;
2161 };
2162
2163 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2164                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2165 {
2166         struct int_node *inode;
2167         struct syscall_stats *stats;
2168         u64 duration = 0;
2169
2170         inode = intlist__findnew(ttrace->syscall_stats, id);
2171         if (inode == NULL)
2172                 return;
2173
2174         stats = inode->priv;
2175         if (stats == NULL) {
2176                 stats = zalloc(sizeof(*stats));
2177                 if (stats == NULL)
2178                         return;
2179
2180                 init_stats(&stats->stats);
2181                 inode->priv = stats;
2182         }
2183
2184         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2185                 duration = sample->time - ttrace->entry_time;
2186
2187         update_stats(&stats->stats, duration);
2188
2189         if (err < 0) {
2190                 ++stats->nr_failures;
2191
2192                 if (!errno_summary)
2193                         return;
2194
2195                 err = -err;
2196                 if (err > stats->max_errno) {
2197                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2198
2199                         if (new_errnos) {
2200                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2201                         } else {
2202                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2203                                          thread__comm_str(thread), thread->pid_, thread->tid);
2204                                 return;
2205                         }
2206
2207                         stats->errnos = new_errnos;
2208                         stats->max_errno = err;
2209                 }
2210
2211                 ++stats->errnos[err - 1];
2212         }
2213 }
2214
2215 static int trace__printf_interrupted_entry(struct trace *trace)
2216 {
2217         struct thread_trace *ttrace;
2218         size_t printed;
2219         int len;
2220
2221         if (trace->failure_only || trace->current == NULL)
2222                 return 0;
2223
2224         ttrace = thread__priv(trace->current);
2225
2226         if (!ttrace->entry_pending)
2227                 return 0;
2228
2229         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2230         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2231
2232         if (len < trace->args_alignment - 4)
2233                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2234
2235         printed += fprintf(trace->output, " ...\n");
2236
2237         ttrace->entry_pending = false;
2238         ++trace->nr_events_printed;
2239
2240         return printed;
2241 }
2242
2243 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2244                                  struct perf_sample *sample, struct thread *thread)
2245 {
2246         int printed = 0;
2247
2248         if (trace->print_sample) {
2249                 double ts = (double)sample->time / NSEC_PER_MSEC;
2250
2251                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2252                                    evsel__name(evsel), ts,
2253                                    thread__comm_str(thread),
2254                                    sample->pid, sample->tid, sample->cpu);
2255         }
2256
2257         return printed;
2258 }
2259
2260 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2261 {
2262         void *augmented_args = NULL;
2263         /*
2264          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2265          * and there we get all 6 syscall args plus the tracepoint common fields
2266          * that gets calculated at the start and the syscall_nr (another long).
2267          * So we check if that is the case and if so don't look after the
2268          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2269          * which is fixed.
2270          *
2271          * We'll revisit this later to pass s->args_size to the BPF augmenter
2272          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2273          * copies only what we need for each syscall, like what happens when we
2274          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2275          * traffic to just what is needed for each syscall.
2276          */
2277         int args_size = raw_augmented_args_size ?: sc->args_size;
2278
2279         *augmented_args_size = sample->raw_size - args_size;
2280         if (*augmented_args_size > 0)
2281                 augmented_args = sample->raw_data + args_size;
2282
2283         return augmented_args;
2284 }
2285
2286 static void syscall__exit(struct syscall *sc)
2287 {
2288         if (!sc)
2289                 return;
2290
2291         free(sc->arg_fmt);
2292 }
2293
2294 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2295                             union perf_event *event __maybe_unused,
2296                             struct perf_sample *sample)
2297 {
2298         char *msg;
2299         void *args;
2300         int printed = 0;
2301         struct thread *thread;
2302         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2303         int augmented_args_size = 0;
2304         void *augmented_args = NULL;
2305         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2306         struct thread_trace *ttrace;
2307
2308         if (sc == NULL)
2309                 return -1;
2310
2311         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2312         ttrace = thread__trace(thread, trace->output);
2313         if (ttrace == NULL)
2314                 goto out_put;
2315
2316         trace__fprintf_sample(trace, evsel, sample, thread);
2317
2318         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2319
2320         if (ttrace->entry_str == NULL) {
2321                 ttrace->entry_str = malloc(trace__entry_str_size);
2322                 if (!ttrace->entry_str)
2323                         goto out_put;
2324         }
2325
2326         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2327                 trace__printf_interrupted_entry(trace);
2328         /*
2329          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2330          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2331          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2332          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2333          * so when handling, say the openat syscall, we end up getting 6 args for the
2334          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2335          * thinking that the extra 2 u64 args are the augmented filename, so just check
2336          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2337          */
2338         if (evsel != trace->syscalls.events.sys_enter)
2339                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2340         ttrace->entry_time = sample->time;
2341         msg = ttrace->entry_str;
2342         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2343
2344         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2345                                            args, augmented_args, augmented_args_size, trace, thread);
2346
2347         if (sc->is_exit) {
2348                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2349                         int alignment = 0;
2350
2351                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2352                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2353                         if (trace->args_alignment > printed)
2354                                 alignment = trace->args_alignment - printed;
2355                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2356                 }
2357         } else {
2358                 ttrace->entry_pending = true;
2359                 /* See trace__vfs_getname & trace__sys_exit */
2360                 ttrace->filename.pending_open = false;
2361         }
2362
2363         if (trace->current != thread) {
2364                 thread__put(trace->current);
2365                 trace->current = thread__get(thread);
2366         }
2367         err = 0;
2368 out_put:
2369         thread__put(thread);
2370         return err;
2371 }
2372
2373 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2374                                     struct perf_sample *sample)
2375 {
2376         struct thread_trace *ttrace;
2377         struct thread *thread;
2378         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2379         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2380         char msg[1024];
2381         void *args, *augmented_args = NULL;
2382         int augmented_args_size;
2383
2384         if (sc == NULL)
2385                 return -1;
2386
2387         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2388         ttrace = thread__trace(thread, trace->output);
2389         /*
2390          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2391          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2392          */
2393         if (ttrace == NULL)
2394                 goto out_put;
2395
2396         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2397         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2398         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2399         fprintf(trace->output, "%s", msg);
2400         err = 0;
2401 out_put:
2402         thread__put(thread);
2403         return err;
2404 }
2405
2406 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2407                                     struct perf_sample *sample,
2408                                     struct callchain_cursor *cursor)
2409 {
2410         struct addr_location al;
2411         int max_stack = evsel->core.attr.sample_max_stack ?
2412                         evsel->core.attr.sample_max_stack :
2413                         trace->max_stack;
2414         int err;
2415
2416         if (machine__resolve(trace->host, &al, sample) < 0)
2417                 return -1;
2418
2419         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2420         addr_location__put(&al);
2421         return err;
2422 }
2423
2424 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2425 {
2426         /* TODO: user-configurable print_opts */
2427         const unsigned int print_opts = EVSEL__PRINT_SYM |
2428                                         EVSEL__PRINT_DSO |
2429                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2430
2431         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2432 }
2433
2434 static const char *errno_to_name(struct evsel *evsel, int err)
2435 {
2436         struct perf_env *env = evsel__env(evsel);
2437         const char *arch_name = perf_env__arch(env);
2438
2439         return arch_syscalls__strerrno(arch_name, err);
2440 }
2441
2442 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2443                            union perf_event *event __maybe_unused,
2444                            struct perf_sample *sample)
2445 {
2446         long ret;
2447         u64 duration = 0;
2448         bool duration_calculated = false;
2449         struct thread *thread;
2450         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2451         int alignment = trace->args_alignment;
2452         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2453         struct thread_trace *ttrace;
2454
2455         if (sc == NULL)
2456                 return -1;
2457
2458         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2459         ttrace = thread__trace(thread, trace->output);
2460         if (ttrace == NULL)
2461                 goto out_put;
2462
2463         trace__fprintf_sample(trace, evsel, sample, thread);
2464
2465         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2466
2467         if (trace->summary)
2468                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2469
2470         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2471                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2472                 ttrace->filename.pending_open = false;
2473                 ++trace->stats.vfs_getname;
2474         }
2475
2476         if (ttrace->entry_time) {
2477                 duration = sample->time - ttrace->entry_time;
2478                 if (trace__filter_duration(trace, duration))
2479                         goto out;
2480                 duration_calculated = true;
2481         } else if (trace->duration_filter)
2482                 goto out;
2483
2484         if (sample->callchain) {
2485                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2486                 if (callchain_ret == 0) {
2487                         if (callchain_cursor.nr < trace->min_stack)
2488                                 goto out;
2489                         callchain_ret = 1;
2490                 }
2491         }
2492
2493         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2494                 goto out;
2495
2496         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2497
2498         if (ttrace->entry_pending) {
2499                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2500         } else {
2501                 printed += fprintf(trace->output, " ... [");
2502                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2503                 printed += 9;
2504                 printed += fprintf(trace->output, "]: %s()", sc->name);
2505         }
2506
2507         printed++; /* the closing ')' */
2508
2509         if (alignment > printed)
2510                 alignment -= printed;
2511         else
2512                 alignment = 0;
2513
2514         fprintf(trace->output, ")%*s= ", alignment, " ");
2515
2516         if (sc->fmt == NULL) {
2517                 if (ret < 0)
2518                         goto errno_print;
2519 signed_print:
2520                 fprintf(trace->output, "%ld", ret);
2521         } else if (ret < 0) {
2522 errno_print: {
2523                 char bf[STRERR_BUFSIZE];
2524                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2525                            *e = errno_to_name(evsel, -ret);
2526
2527                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2528         }
2529         } else if (ret == 0 && sc->fmt->timeout)
2530                 fprintf(trace->output, "0 (Timeout)");
2531         else if (ttrace->ret_scnprintf) {
2532                 char bf[1024];
2533                 struct syscall_arg arg = {
2534                         .val    = ret,
2535                         .thread = thread,
2536                         .trace  = trace,
2537                 };
2538                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2539                 ttrace->ret_scnprintf = NULL;
2540                 fprintf(trace->output, "%s", bf);
2541         } else if (sc->fmt->hexret)
2542                 fprintf(trace->output, "%#lx", ret);
2543         else if (sc->fmt->errpid) {
2544                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2545
2546                 if (child != NULL) {
2547                         fprintf(trace->output, "%ld", ret);
2548                         if (child->comm_set)
2549                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2550                         thread__put(child);
2551                 }
2552         } else
2553                 goto signed_print;
2554
2555         fputc('\n', trace->output);
2556
2557         /*
2558          * We only consider an 'event' for the sake of --max-events a non-filtered
2559          * sys_enter + sys_exit and other tracepoint events.
2560          */
2561         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2562                 interrupted = true;
2563
2564         if (callchain_ret > 0)
2565                 trace__fprintf_callchain(trace, sample);
2566         else if (callchain_ret < 0)
2567                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2568 out:
2569         ttrace->entry_pending = false;
2570         err = 0;
2571 out_put:
2572         thread__put(thread);
2573         return err;
2574 }
2575
2576 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2577                               union perf_event *event __maybe_unused,
2578                               struct perf_sample *sample)
2579 {
2580         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2581         struct thread_trace *ttrace;
2582         size_t filename_len, entry_str_len, to_move;
2583         ssize_t remaining_space;
2584         char *pos;
2585         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2586
2587         if (!thread)
2588                 goto out;
2589
2590         ttrace = thread__priv(thread);
2591         if (!ttrace)
2592                 goto out_put;
2593
2594         filename_len = strlen(filename);
2595         if (filename_len == 0)
2596                 goto out_put;
2597
2598         if (ttrace->filename.namelen < filename_len) {
2599                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2600
2601                 if (f == NULL)
2602                         goto out_put;
2603
2604                 ttrace->filename.namelen = filename_len;
2605                 ttrace->filename.name = f;
2606         }
2607
2608         strcpy(ttrace->filename.name, filename);
2609         ttrace->filename.pending_open = true;
2610
2611         if (!ttrace->filename.ptr)
2612                 goto out_put;
2613
2614         entry_str_len = strlen(ttrace->entry_str);
2615         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2616         if (remaining_space <= 0)
2617                 goto out_put;
2618
2619         if (filename_len > (size_t)remaining_space) {
2620                 filename += filename_len - remaining_space;
2621                 filename_len = remaining_space;
2622         }
2623
2624         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2625         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2626         memmove(pos + filename_len, pos, to_move);
2627         memcpy(pos, filename, filename_len);
2628
2629         ttrace->filename.ptr = 0;
2630         ttrace->filename.entry_str_pos = 0;
2631 out_put:
2632         thread__put(thread);
2633 out:
2634         return 0;
2635 }
2636
2637 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2638                                      union perf_event *event __maybe_unused,
2639                                      struct perf_sample *sample)
2640 {
2641         u64 runtime = evsel__intval(evsel, sample, "runtime");
2642         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2643         struct thread *thread = machine__findnew_thread(trace->host,
2644                                                         sample->pid,
2645                                                         sample->tid);
2646         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2647
2648         if (ttrace == NULL)
2649                 goto out_dump;
2650
2651         ttrace->runtime_ms += runtime_ms;
2652         trace->runtime_ms += runtime_ms;
2653 out_put:
2654         thread__put(thread);
2655         return 0;
2656
2657 out_dump:
2658         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2659                evsel->name,
2660                evsel__strval(evsel, sample, "comm"),
2661                (pid_t)evsel__intval(evsel, sample, "pid"),
2662                runtime,
2663                evsel__intval(evsel, sample, "vruntime"));
2664         goto out_put;
2665 }
2666
2667 static int bpf_output__printer(enum binary_printer_ops op,
2668                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2669 {
2670         unsigned char ch = (unsigned char)val;
2671
2672         switch (op) {
2673         case BINARY_PRINT_CHAR_DATA:
2674                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2675         case BINARY_PRINT_DATA_BEGIN:
2676         case BINARY_PRINT_LINE_BEGIN:
2677         case BINARY_PRINT_ADDR:
2678         case BINARY_PRINT_NUM_DATA:
2679         case BINARY_PRINT_NUM_PAD:
2680         case BINARY_PRINT_SEP:
2681         case BINARY_PRINT_CHAR_PAD:
2682         case BINARY_PRINT_LINE_END:
2683         case BINARY_PRINT_DATA_END:
2684         default:
2685                 break;
2686         }
2687
2688         return 0;
2689 }
2690
2691 static void bpf_output__fprintf(struct trace *trace,
2692                                 struct perf_sample *sample)
2693 {
2694         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2695                         bpf_output__printer, NULL, trace->output);
2696         ++trace->nr_events_printed;
2697 }
2698
2699 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2700                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2701 {
2702         char bf[2048];
2703         size_t size = sizeof(bf);
2704         struct tep_format_field *field = evsel->tp_format->format.fields;
2705         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2706         size_t printed = 0;
2707         unsigned long val;
2708         u8 bit = 1;
2709         struct syscall_arg syscall_arg = {
2710                 .augmented = {
2711                         .size = augmented_args_size,
2712                         .args = augmented_args,
2713                 },
2714                 .idx    = 0,
2715                 .mask   = 0,
2716                 .trace  = trace,
2717                 .thread = thread,
2718                 .show_string_prefix = trace->show_string_prefix,
2719         };
2720
2721         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2722                 if (syscall_arg.mask & bit)
2723                         continue;
2724
2725                 syscall_arg.len = 0;
2726                 syscall_arg.fmt = arg;
2727                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2728                         int offset = field->offset;
2729
2730                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2731                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2732                                 syscall_arg.len = offset >> 16;
2733                                 offset &= 0xffff;
2734                                 if (tep_field_is_relative(field->flags))
2735                                         offset += field->offset + field->size;
2736                         }
2737
2738                         val = (uintptr_t)(sample->raw_data + offset);
2739                 } else
2740                         val = format_field__intval(field, sample, evsel->needs_swap);
2741                 /*
2742                  * Some syscall args need some mask, most don't and
2743                  * return val untouched.
2744                  */
2745                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2746
2747                 /*
2748                  * Suppress this argument if its value is zero and
2749                  * we don't have a string associated in an
2750                  * strarray for it.
2751                  */
2752                 if (val == 0 &&
2753                     !trace->show_zeros &&
2754                     !((arg->show_zero ||
2755                        arg->scnprintf == SCA_STRARRAY ||
2756                        arg->scnprintf == SCA_STRARRAYS) &&
2757                       arg->parm))
2758                         continue;
2759
2760                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2761
2762                 if (trace->show_arg_names)
2763                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2764
2765                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2766         }
2767
2768         return printed + fprintf(trace->output, "%s", bf);
2769 }
2770
2771 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2772                                 union perf_event *event __maybe_unused,
2773                                 struct perf_sample *sample)
2774 {
2775         struct thread *thread;
2776         int callchain_ret = 0;
2777         /*
2778          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2779          * this event's max_events having been hit and this is an entry coming
2780          * from the ring buffer that we should discard, since the max events
2781          * have already been considered/printed.
2782          */
2783         if (evsel->disabled)
2784                 return 0;
2785
2786         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2787
2788         if (sample->callchain) {
2789                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2790                 if (callchain_ret == 0) {
2791                         if (callchain_cursor.nr < trace->min_stack)
2792                                 goto out;
2793                         callchain_ret = 1;
2794                 }
2795         }
2796
2797         trace__printf_interrupted_entry(trace);
2798         trace__fprintf_tstamp(trace, sample->time, trace->output);
2799
2800         if (trace->trace_syscalls && trace->show_duration)
2801                 fprintf(trace->output, "(         ): ");
2802
2803         if (thread)
2804                 trace__fprintf_comm_tid(trace, thread, trace->output);
2805
2806         if (evsel == trace->syscalls.events.augmented) {
2807                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2808                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2809
2810                 if (sc) {
2811                         fprintf(trace->output, "%s(", sc->name);
2812                         trace__fprintf_sys_enter(trace, evsel, sample);
2813                         fputc(')', trace->output);
2814                         goto newline;
2815                 }
2816
2817                 /*
2818                  * XXX: Not having the associated syscall info or not finding/adding
2819                  *      the thread should never happen, but if it does...
2820                  *      fall thru and print it as a bpf_output event.
2821                  */
2822         }
2823
2824         fprintf(trace->output, "%s(", evsel->name);
2825
2826         if (evsel__is_bpf_output(evsel)) {
2827                 bpf_output__fprintf(trace, sample);
2828         } else if (evsel->tp_format) {
2829                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2830                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2831                         if (trace->libtraceevent_print) {
2832                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2833                                                       sample->raw_data, sample->raw_size,
2834                                                       trace->output);
2835                         } else {
2836                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2837                         }
2838                 }
2839         }
2840
2841 newline:
2842         fprintf(trace->output, ")\n");
2843
2844         if (callchain_ret > 0)
2845                 trace__fprintf_callchain(trace, sample);
2846         else if (callchain_ret < 0)
2847                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2848
2849         ++trace->nr_events_printed;
2850
2851         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2852                 evsel__disable(evsel);
2853                 evsel__close(evsel);
2854         }
2855 out:
2856         thread__put(thread);
2857         return 0;
2858 }
2859
2860 static void print_location(FILE *f, struct perf_sample *sample,
2861                            struct addr_location *al,
2862                            bool print_dso, bool print_sym)
2863 {
2864
2865         if ((verbose > 0 || print_dso) && al->map)
2866                 fprintf(f, "%s@", al->map->dso->long_name);
2867
2868         if ((verbose > 0 || print_sym) && al->sym)
2869                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2870                         al->addr - al->sym->start);
2871         else if (al->map)
2872                 fprintf(f, "0x%" PRIx64, al->addr);
2873         else
2874                 fprintf(f, "0x%" PRIx64, sample->addr);
2875 }
2876
2877 static int trace__pgfault(struct trace *trace,
2878                           struct evsel *evsel,
2879                           union perf_event *event __maybe_unused,
2880                           struct perf_sample *sample)
2881 {
2882         struct thread *thread;
2883         struct addr_location al;
2884         char map_type = 'd';
2885         struct thread_trace *ttrace;
2886         int err = -1;
2887         int callchain_ret = 0;
2888
2889         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2890
2891         if (sample->callchain) {
2892                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2893                 if (callchain_ret == 0) {
2894                         if (callchain_cursor.nr < trace->min_stack)
2895                                 goto out_put;
2896                         callchain_ret = 1;
2897                 }
2898         }
2899
2900         ttrace = thread__trace(thread, trace->output);
2901         if (ttrace == NULL)
2902                 goto out_put;
2903
2904         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2905                 ttrace->pfmaj++;
2906         else
2907                 ttrace->pfmin++;
2908
2909         if (trace->summary_only)
2910                 goto out;
2911
2912         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2913
2914         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2915
2916         fprintf(trace->output, "%sfault [",
2917                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2918                 "maj" : "min");
2919
2920         print_location(trace->output, sample, &al, false, true);
2921
2922         fprintf(trace->output, "] => ");
2923
2924         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2925
2926         if (!al.map) {
2927                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2928
2929                 if (al.map)
2930                         map_type = 'x';
2931                 else
2932                         map_type = '?';
2933         }
2934
2935         print_location(trace->output, sample, &al, true, false);
2936
2937         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2938
2939         if (callchain_ret > 0)
2940                 trace__fprintf_callchain(trace, sample);
2941         else if (callchain_ret < 0)
2942                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2943
2944         ++trace->nr_events_printed;
2945 out:
2946         err = 0;
2947 out_put:
2948         thread__put(thread);
2949         return err;
2950 }
2951
2952 static void trace__set_base_time(struct trace *trace,
2953                                  struct evsel *evsel,
2954                                  struct perf_sample *sample)
2955 {
2956         /*
2957          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2958          * and don't use sample->time unconditionally, we may end up having
2959          * some other event in the future without PERF_SAMPLE_TIME for good
2960          * reason, i.e. we may not be interested in its timestamps, just in
2961          * it taking place, picking some piece of information when it
2962          * appears in our event stream (vfs_getname comes to mind).
2963          */
2964         if (trace->base_time == 0 && !trace->full_time &&
2965             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2966                 trace->base_time = sample->time;
2967 }
2968
2969 static int trace__process_sample(struct perf_tool *tool,
2970                                  union perf_event *event,
2971                                  struct perf_sample *sample,
2972                                  struct evsel *evsel,
2973                                  struct machine *machine __maybe_unused)
2974 {
2975         struct trace *trace = container_of(tool, struct trace, tool);
2976         struct thread *thread;
2977         int err = 0;
2978
2979         tracepoint_handler handler = evsel->handler;
2980
2981         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2982         if (thread && thread__is_filtered(thread))
2983                 goto out;
2984
2985         trace__set_base_time(trace, evsel, sample);
2986
2987         if (handler) {
2988                 ++trace->nr_events;
2989                 handler(trace, evsel, event, sample);
2990         }
2991 out:
2992         thread__put(thread);
2993         return err;
2994 }
2995
2996 static int trace__record(struct trace *trace, int argc, const char **argv)
2997 {
2998         unsigned int rec_argc, i, j;
2999         const char **rec_argv;
3000         const char * const record_args[] = {
3001                 "record",
3002                 "-R",
3003                 "-m", "1024",
3004                 "-c", "1",
3005         };
3006         pid_t pid = getpid();
3007         char *filter = asprintf__tp_filter_pids(1, &pid);
3008         const char * const sc_args[] = { "-e", };
3009         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3010         const char * const majpf_args[] = { "-e", "major-faults" };
3011         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3012         const char * const minpf_args[] = { "-e", "minor-faults" };
3013         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3014         int err = -1;
3015
3016         /* +3 is for the event string below and the pid filter */
3017         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3018                 majpf_args_nr + minpf_args_nr + argc;
3019         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3020
3021         if (rec_argv == NULL || filter == NULL)
3022                 goto out_free;
3023
3024         j = 0;
3025         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3026                 rec_argv[j++] = record_args[i];
3027
3028         if (trace->trace_syscalls) {
3029                 for (i = 0; i < sc_args_nr; i++)
3030                         rec_argv[j++] = sc_args[i];
3031
3032                 /* event string may be different for older kernels - e.g., RHEL6 */
3033                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3034                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3035                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3036                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3037                 else {
3038                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3039                         goto out_free;
3040                 }
3041         }
3042
3043         rec_argv[j++] = "--filter";
3044         rec_argv[j++] = filter;
3045
3046         if (trace->trace_pgfaults & TRACE_PFMAJ)
3047                 for (i = 0; i < majpf_args_nr; i++)
3048                         rec_argv[j++] = majpf_args[i];
3049
3050         if (trace->trace_pgfaults & TRACE_PFMIN)
3051                 for (i = 0; i < minpf_args_nr; i++)
3052                         rec_argv[j++] = minpf_args[i];
3053
3054         for (i = 0; i < (unsigned int)argc; i++)
3055                 rec_argv[j++] = argv[i];
3056
3057         err = cmd_record(j, rec_argv);
3058 out_free:
3059         free(filter);
3060         free(rec_argv);
3061         return err;
3062 }
3063
3064 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3065
3066 static bool evlist__add_vfs_getname(struct evlist *evlist)
3067 {
3068         bool found = false;
3069         struct evsel *evsel, *tmp;
3070         struct parse_events_error err;
3071         int ret;
3072
3073         parse_events_error__init(&err);
3074         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3075         parse_events_error__exit(&err);
3076         if (ret)
3077                 return false;
3078
3079         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3080                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3081                         continue;
3082
3083                 if (evsel__field(evsel, "pathname")) {
3084                         evsel->handler = trace__vfs_getname;
3085                         found = true;
3086                         continue;
3087                 }
3088
3089                 list_del_init(&evsel->core.node);
3090                 evsel->evlist = NULL;
3091                 evsel__delete(evsel);
3092         }
3093
3094         return found;
3095 }
3096
3097 static struct evsel *evsel__new_pgfault(u64 config)
3098 {
3099         struct evsel *evsel;
3100         struct perf_event_attr attr = {
3101                 .type = PERF_TYPE_SOFTWARE,
3102                 .mmap_data = 1,
3103         };
3104
3105         attr.config = config;
3106         attr.sample_period = 1;
3107
3108         event_attr_init(&attr);
3109
3110         evsel = evsel__new(&attr);
3111         if (evsel)
3112                 evsel->handler = trace__pgfault;
3113
3114         return evsel;
3115 }
3116
3117 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3118 {
3119         struct evsel *evsel;
3120
3121         evlist__for_each_entry(evlist, evsel) {
3122                 struct evsel_trace *et = evsel->priv;
3123
3124                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3125                         continue;
3126
3127                 free(et->fmt);
3128                 free(et);
3129         }
3130 }
3131
3132 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3133 {
3134         const u32 type = event->header.type;
3135         struct evsel *evsel;
3136
3137         if (type != PERF_RECORD_SAMPLE) {
3138                 trace__process_event(trace, trace->host, event, sample);
3139                 return;
3140         }
3141
3142         evsel = evlist__id2evsel(trace->evlist, sample->id);
3143         if (evsel == NULL) {
3144                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3145                 return;
3146         }
3147
3148         if (evswitch__discard(&trace->evswitch, evsel))
3149                 return;
3150
3151         trace__set_base_time(trace, evsel, sample);
3152
3153         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3154             sample->raw_data == NULL) {
3155                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3156                        evsel__name(evsel), sample->tid,
3157                        sample->cpu, sample->raw_size);
3158         } else {
3159                 tracepoint_handler handler = evsel->handler;
3160                 handler(trace, evsel, event, sample);
3161         }
3162
3163         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3164                 interrupted = true;
3165 }
3166
3167 static int trace__add_syscall_newtp(struct trace *trace)
3168 {
3169         int ret = -1;
3170         struct evlist *evlist = trace->evlist;
3171         struct evsel *sys_enter, *sys_exit;
3172
3173         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3174         if (sys_enter == NULL)
3175                 goto out;
3176
3177         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3178                 goto out_delete_sys_enter;
3179
3180         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3181         if (sys_exit == NULL)
3182                 goto out_delete_sys_enter;
3183
3184         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3185                 goto out_delete_sys_exit;
3186
3187         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3188         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3189
3190         evlist__add(evlist, sys_enter);
3191         evlist__add(evlist, sys_exit);
3192
3193         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3194                 /*
3195                  * We're interested only in the user space callchain
3196                  * leading to the syscall, allow overriding that for
3197                  * debugging reasons using --kernel_syscall_callchains
3198                  */
3199                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3200         }
3201
3202         trace->syscalls.events.sys_enter = sys_enter;
3203         trace->syscalls.events.sys_exit  = sys_exit;
3204
3205         ret = 0;
3206 out:
3207         return ret;
3208
3209 out_delete_sys_exit:
3210         evsel__delete_priv(sys_exit);
3211 out_delete_sys_enter:
3212         evsel__delete_priv(sys_enter);
3213         goto out;
3214 }
3215
3216 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3217 {
3218         int err = -1;
3219         struct evsel *sys_exit;
3220         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3221                                                 trace->ev_qualifier_ids.nr,
3222                                                 trace->ev_qualifier_ids.entries);
3223
3224         if (filter == NULL)
3225                 goto out_enomem;
3226
3227         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3228                 sys_exit = trace->syscalls.events.sys_exit;
3229                 err = evsel__append_tp_filter(sys_exit, filter);
3230         }
3231
3232         free(filter);
3233 out:
3234         return err;
3235 out_enomem:
3236         errno = ENOMEM;
3237         goto out;
3238 }
3239
3240 #ifdef HAVE_LIBBPF_SUPPORT
3241 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3242 {
3243         if (trace->bpf_obj == NULL)
3244                 return NULL;
3245
3246         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3247 }
3248
3249 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3250 {
3251         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3252 }
3253
3254 static void trace__set_bpf_map_syscalls(struct trace *trace)
3255 {
3256         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3257         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3258 }
3259
3260 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3261 {
3262         struct bpf_program *pos, *prog = NULL;
3263         const char *sec_name;
3264
3265         if (trace->bpf_obj == NULL)
3266                 return NULL;
3267
3268         bpf_object__for_each_program(pos, trace->bpf_obj) {
3269                 sec_name = bpf_program__section_name(pos);
3270                 if (sec_name && !strcmp(sec_name, name)) {
3271                         prog = pos;
3272                         break;
3273                 }
3274         }
3275
3276         return prog;
3277 }
3278
3279 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3280                                                         const char *prog_name, const char *type)
3281 {
3282         struct bpf_program *prog;
3283
3284         if (prog_name == NULL) {
3285                 char default_prog_name[256];
3286                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3287                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3288                 if (prog != NULL)
3289                         goto out_found;
3290                 if (sc->fmt && sc->fmt->alias) {
3291                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3292                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3293                         if (prog != NULL)
3294                                 goto out_found;
3295                 }
3296                 goto out_unaugmented;
3297         }
3298
3299         prog = trace__find_bpf_program_by_title(trace, prog_name);
3300
3301         if (prog != NULL) {
3302 out_found:
3303                 return prog;
3304         }
3305
3306         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3307                  prog_name, type, sc->name);
3308 out_unaugmented:
3309         return trace->syscalls.unaugmented_prog;
3310 }
3311
3312 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3313 {
3314         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3315
3316         if (sc == NULL)
3317                 return;
3318
3319         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3320         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3321 }
3322
3323 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3324 {
3325         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3326         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3327 }
3328
3329 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3330 {
3331         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3332         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3333 }
3334
3335 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3336 {
3337         struct tep_format_field *field, *candidate_field;
3338         int id;
3339
3340         /*
3341          * We're only interested in syscalls that have a pointer:
3342          */
3343         for (field = sc->args; field; field = field->next) {
3344                 if (field->flags & TEP_FIELD_IS_POINTER)
3345                         goto try_to_find_pair;
3346         }
3347
3348         return NULL;
3349
3350 try_to_find_pair:
3351         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3352                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3353                 struct bpf_program *pair_prog;
3354                 bool is_candidate = false;
3355
3356                 if (pair == NULL || pair == sc ||
3357                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3358                         continue;
3359
3360                 for (field = sc->args, candidate_field = pair->args;
3361                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3362                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3363                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3364
3365                         if (is_pointer) {
3366                                if (!candidate_is_pointer) {
3367                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3368                                         continue;
3369                                }
3370                         } else {
3371                                 if (candidate_is_pointer) {
3372                                         // The candidate might copy a pointer we don't have, skip it.
3373                                         goto next_candidate;
3374                                 }
3375                                 continue;
3376                         }
3377
3378                         if (strcmp(field->type, candidate_field->type))
3379                                 goto next_candidate;
3380
3381                         is_candidate = true;
3382                 }
3383
3384                 if (!is_candidate)
3385                         goto next_candidate;
3386
3387                 /*
3388                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3389                  * then it may be collecting that and we then can't use it, as it would collect
3390                  * more than what is common to the two syscalls.
3391                  */
3392                 if (candidate_field) {
3393                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3394                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3395                                         goto next_candidate;
3396                 }
3397
3398                 pair_prog = pair->bpf_prog.sys_enter;
3399                 /*
3400                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3401                  * have been searched for, so search it here and if it returns the
3402                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3403                  * program for a filtered syscall on a non-filtered one.
3404                  *
3405                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3406                  * useful for "renameat2".
3407                  */
3408                 if (pair_prog == NULL) {
3409                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3410                         if (pair_prog == trace->syscalls.unaugmented_prog)
3411                                 goto next_candidate;
3412                 }
3413
3414                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3415                 return pair_prog;
3416         next_candidate:
3417                 continue;
3418         }
3419
3420         return NULL;
3421 }
3422
3423 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3424 {
3425         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3426             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3427         int err = 0, key;
3428
3429         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3430                 int prog_fd;
3431
3432                 if (!trace__syscall_enabled(trace, key))
3433                         continue;
3434
3435                 trace__init_syscall_bpf_progs(trace, key);
3436
3437                 // It'll get at least the "!raw_syscalls:unaugmented"
3438                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3439                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3440                 if (err)
3441                         break;
3442                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3443                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3444                 if (err)
3445                         break;
3446         }
3447
3448         /*
3449          * Now lets do a second pass looking for enabled syscalls without
3450          * an augmenter that have a signature that is a superset of another
3451          * syscall with an augmenter so that we can auto-reuse it.
3452          *
3453          * I.e. if we have an augmenter for the "open" syscall that has
3454          * this signature:
3455          *
3456          *   int open(const char *pathname, int flags, mode_t mode);
3457          *
3458          * I.e. that will collect just the first string argument, then we
3459          * can reuse it for the 'creat' syscall, that has this signature:
3460          *
3461          *   int creat(const char *pathname, mode_t mode);
3462          *
3463          * and for:
3464          *
3465          *   int stat(const char *pathname, struct stat *statbuf);
3466          *   int lstat(const char *pathname, struct stat *statbuf);
3467          *
3468          * Because the 'open' augmenter will collect the first arg as a string,
3469          * and leave alone all the other args, which already helps with
3470          * beautifying 'stat' and 'lstat''s pathname arg.
3471          *
3472          * Then, in time, when 'stat' gets an augmenter that collects both
3473          * first and second arg (this one on the raw_syscalls:sys_exit prog
3474          * array tail call, then that one will be used.
3475          */
3476         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3477                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3478                 struct bpf_program *pair_prog;
3479                 int prog_fd;
3480
3481                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3482                         continue;
3483
3484                 /*
3485                  * For now we're just reusing the sys_enter prog, and if it
3486                  * already has an augmenter, we don't need to find one.
3487                  */
3488                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3489                         continue;
3490
3491                 /*
3492                  * Look at all the other syscalls for one that has a signature
3493                  * that is close enough that we can share:
3494                  */
3495                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3496                 if (pair_prog == NULL)
3497                         continue;
3498
3499                 sc->bpf_prog.sys_enter = pair_prog;
3500
3501                 /*
3502                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3503                  * with the fd for the program we're reusing:
3504                  */
3505                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3506                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3507                 if (err)
3508                         break;
3509         }
3510
3511
3512         return err;
3513 }
3514
3515 static void trace__delete_augmented_syscalls(struct trace *trace)
3516 {
3517         struct evsel *evsel, *tmp;
3518
3519         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3520         evsel__delete(trace->syscalls.events.augmented);
3521         trace->syscalls.events.augmented = NULL;
3522
3523         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3524                 if (evsel->bpf_obj == trace->bpf_obj) {
3525                         evlist__remove(trace->evlist, evsel);
3526                         evsel__delete(evsel);
3527                 }
3528
3529         }
3530
3531         bpf_object__close(trace->bpf_obj);
3532         trace->bpf_obj = NULL;
3533 }
3534 #else // HAVE_LIBBPF_SUPPORT
3535 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3536                                                    const char *name __maybe_unused)
3537 {
3538         return NULL;
3539 }
3540
3541 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3542 {
3543 }
3544
3545 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3546 {
3547 }
3548
3549 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3550                                                             const char *name __maybe_unused)
3551 {
3552         return NULL;
3553 }
3554
3555 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3556 {
3557         return 0;
3558 }
3559
3560 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3561 {
3562 }
3563 #endif // HAVE_LIBBPF_SUPPORT
3564
3565 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3566 {
3567         struct evsel *evsel;
3568
3569         evlist__for_each_entry(trace->evlist, evsel) {
3570                 if (evsel == trace->syscalls.events.augmented ||
3571                     evsel->bpf_obj == trace->bpf_obj)
3572                         continue;
3573
3574                 return false;
3575         }
3576
3577         return true;
3578 }
3579
3580 static int trace__set_ev_qualifier_filter(struct trace *trace)
3581 {
3582         if (trace->syscalls.events.sys_enter)
3583                 return trace__set_ev_qualifier_tp_filter(trace);
3584         return 0;
3585 }
3586
3587 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3588                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3589 {
3590         int err = 0;
3591 #ifdef HAVE_LIBBPF_SUPPORT
3592         bool value = true;
3593         int map_fd = bpf_map__fd(map);
3594         size_t i;
3595
3596         for (i = 0; i < npids; ++i) {
3597                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3598                 if (err)
3599                         break;
3600         }
3601 #endif
3602         return err;
3603 }
3604
3605 static int trace__set_filter_loop_pids(struct trace *trace)
3606 {
3607         unsigned int nr = 1, err;
3608         pid_t pids[32] = {
3609                 getpid(),
3610         };
3611         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3612
3613         while (thread && nr < ARRAY_SIZE(pids)) {
3614                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3615
3616                 if (parent == NULL)
3617                         break;
3618
3619                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3620                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3621                         pids[nr++] = parent->tid;
3622                         break;
3623                 }
3624                 thread = parent;
3625         }
3626
3627         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3628         if (!err && trace->filter_pids.map)
3629                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3630
3631         return err;
3632 }
3633
3634 static int trace__set_filter_pids(struct trace *trace)
3635 {
3636         int err = 0;
3637         /*
3638          * Better not use !target__has_task() here because we need to cover the
3639          * case where no threads were specified in the command line, but a
3640          * workload was, and in that case we will fill in the thread_map when
3641          * we fork the workload in evlist__prepare_workload.
3642          */
3643         if (trace->filter_pids.nr > 0) {
3644                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3645                                                     trace->filter_pids.entries);
3646                 if (!err && trace->filter_pids.map) {
3647                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3648                                                        trace->filter_pids.entries);
3649                 }
3650         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3651                 err = trace__set_filter_loop_pids(trace);
3652         }
3653
3654         return err;
3655 }
3656
3657 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3658 {
3659         struct evlist *evlist = trace->evlist;
3660         struct perf_sample sample;
3661         int err = evlist__parse_sample(evlist, event, &sample);
3662
3663         if (err)
3664                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3665         else
3666                 trace__handle_event(trace, event, &sample);
3667
3668         return 0;
3669 }
3670
3671 static int __trace__flush_events(struct trace *trace)
3672 {
3673         u64 first = ordered_events__first_time(&trace->oe.data);
3674         u64 flush = trace->oe.last - NSEC_PER_SEC;
3675
3676         /* Is there some thing to flush.. */
3677         if (first && first < flush)
3678                 return ordered_events__flush_time(&trace->oe.data, flush);
3679
3680         return 0;
3681 }
3682
3683 static int trace__flush_events(struct trace *trace)
3684 {
3685         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3686 }
3687
3688 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3689 {
3690         int err;
3691
3692         if (!trace->sort_events)
3693                 return __trace__deliver_event(trace, event);
3694
3695         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3696         if (err && err != -1)
3697                 return err;
3698
3699         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3700         if (err)
3701                 return err;
3702
3703         return trace__flush_events(trace);
3704 }
3705
3706 static int ordered_events__deliver_event(struct ordered_events *oe,
3707                                          struct ordered_event *event)
3708 {
3709         struct trace *trace = container_of(oe, struct trace, oe.data);
3710
3711         return __trace__deliver_event(trace, event->event);
3712 }
3713
3714 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3715 {
3716         struct tep_format_field *field;
3717         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3718
3719         if (evsel->tp_format == NULL || fmt == NULL)
3720                 return NULL;
3721
3722         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3723                 if (strcmp(field->name, arg) == 0)
3724                         return fmt;
3725
3726         return NULL;
3727 }
3728
3729 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3730 {
3731         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3732
3733         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3734                 char *right = tok + 1, *right_end;
3735
3736                 if (*right == '=')
3737                         ++right;
3738
3739                 while (isspace(*right))
3740                         ++right;
3741
3742                 if (*right == '\0')
3743                         break;
3744
3745                 while (!isalpha(*left))
3746                         if (++left == tok) {
3747                                 /*
3748                                  * Bail out, can't find the name of the argument that is being
3749                                  * used in the filter, let it try to set this filter, will fail later.
3750                                  */
3751                                 return 0;
3752                         }
3753
3754                 right_end = right + 1;
3755                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3756                         ++right_end;
3757
3758                 if (isalpha(*right)) {
3759                         struct syscall_arg_fmt *fmt;
3760                         int left_size = tok - left,
3761                             right_size = right_end - right;
3762                         char arg[128];
3763
3764                         while (isspace(left[left_size - 1]))
3765                                 --left_size;
3766
3767                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3768
3769                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3770                         if (fmt == NULL) {
3771                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3772                                        arg, evsel->name, evsel->filter);
3773                                 return -1;
3774                         }
3775
3776                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3777                                  arg, (int)(right - tok), tok, right_size, right);
3778
3779                         if (fmt->strtoul) {
3780                                 u64 val;
3781                                 struct syscall_arg syscall_arg = {
3782                                         .parm = fmt->parm,
3783                                 };
3784
3785                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3786                                         char *n, expansion[19];
3787                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3788                                         int expansion_offset = right - new_filter;
3789
3790                                         pr_debug("%s", expansion);
3791
3792                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3793                                                 pr_debug(" out of memory!\n");
3794                                                 free(new_filter);
3795                                                 return -1;
3796                                         }
3797                                         if (new_filter != evsel->filter)
3798                                                 free(new_filter);
3799                                         left = n + expansion_offset + expansion_lenght;
3800                                         new_filter = n;
3801                                 } else {
3802                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3803                                                right_size, right, arg, evsel->name, evsel->filter);
3804                                         return -1;
3805                                 }
3806                         } else {
3807                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3808                                        arg, evsel->name, evsel->filter);
3809                                 return -1;
3810                         }
3811
3812                         pr_debug("\n");
3813                 } else {
3814                         left = right_end;
3815                 }
3816         }
3817
3818         if (new_filter != evsel->filter) {
3819                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3820                 evsel__set_filter(evsel, new_filter);
3821                 free(new_filter);
3822         }
3823
3824         return 0;
3825 }
3826
3827 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3828 {
3829         struct evlist *evlist = trace->evlist;
3830         struct evsel *evsel;
3831
3832         evlist__for_each_entry(evlist, evsel) {
3833                 if (evsel->filter == NULL)
3834                         continue;
3835
3836                 if (trace__expand_filter(trace, evsel)) {
3837                         *err_evsel = evsel;
3838                         return -1;
3839                 }
3840         }
3841
3842         return 0;
3843 }
3844
3845 static int trace__run(struct trace *trace, int argc, const char **argv)
3846 {
3847         struct evlist *evlist = trace->evlist;
3848         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3849         int err = -1, i;
3850         unsigned long before;
3851         const bool forks = argc > 0;
3852         bool draining = false;
3853
3854         trace->live = true;
3855
3856         if (!trace->raw_augmented_syscalls) {
3857                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3858                         goto out_error_raw_syscalls;
3859
3860                 if (trace->trace_syscalls)
3861                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3862         }
3863
3864         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3865                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3866                 if (pgfault_maj == NULL)
3867                         goto out_error_mem;
3868                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3869                 evlist__add(evlist, pgfault_maj);
3870         }
3871
3872         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3873                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3874                 if (pgfault_min == NULL)
3875                         goto out_error_mem;
3876                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3877                 evlist__add(evlist, pgfault_min);
3878         }
3879
3880         /* Enable ignoring missing threads when -u/-p option is defined. */
3881         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3882
3883         if (trace->sched &&
3884             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3885                 goto out_error_sched_stat_runtime;
3886         /*
3887          * If a global cgroup was set, apply it to all the events without an
3888          * explicit cgroup. I.e.:
3889          *
3890          *      trace -G A -e sched:*switch
3891          *
3892          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3893          * _and_ sched:sched_switch to the 'A' cgroup, while:
3894          *
3895          * trace -e sched:*switch -G A
3896          *
3897          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3898          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3899          * a cgroup (on the root cgroup, sys wide, etc).
3900          *
3901          * Multiple cgroups:
3902          *
3903          * trace -G A -e sched:*switch -G B
3904          *
3905          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3906          * to the 'B' cgroup.
3907          *
3908          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3909          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3910          */
3911         if (trace->cgroup)
3912                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3913
3914         err = evlist__create_maps(evlist, &trace->opts.target);
3915         if (err < 0) {
3916                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3917                 goto out_delete_evlist;
3918         }
3919
3920         err = trace__symbols_init(trace, evlist);
3921         if (err < 0) {
3922                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3923                 goto out_delete_evlist;
3924         }
3925
3926         evlist__config(evlist, &trace->opts, &callchain_param);
3927
3928         if (forks) {
3929                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3930                 if (err < 0) {
3931                         fprintf(trace->output, "Couldn't run the workload!\n");
3932                         goto out_delete_evlist;
3933                 }
3934                 workload_pid = evlist->workload.pid;
3935         }
3936
3937         err = evlist__open(evlist);
3938         if (err < 0)
3939                 goto out_error_open;
3940
3941         err = bpf__apply_obj_config();
3942         if (err) {
3943                 char errbuf[BUFSIZ];
3944
3945                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3946                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3947                          errbuf);
3948                 goto out_error_open;
3949         }
3950
3951         err = trace__set_filter_pids(trace);
3952         if (err < 0)
3953                 goto out_error_mem;
3954
3955         if (trace->syscalls.prog_array.sys_enter)
3956                 trace__init_syscalls_bpf_prog_array_maps(trace);
3957
3958         if (trace->ev_qualifier_ids.nr > 0) {
3959                 err = trace__set_ev_qualifier_filter(trace);
3960                 if (err < 0)
3961                         goto out_errno;
3962
3963                 if (trace->syscalls.events.sys_exit) {
3964                         pr_debug("event qualifier tracepoint filter: %s\n",
3965                                  trace->syscalls.events.sys_exit->filter);
3966                 }
3967         }
3968
3969         /*
3970          * If the "close" syscall is not traced, then we will not have the
3971          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3972          * fd->pathname table and were ending up showing the last value set by
3973          * syscalls opening a pathname and associating it with a descriptor or
3974          * reading it from /proc/pid/fd/ in cases where that doesn't make
3975          * sense.
3976          *
3977          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3978          *  not in use.
3979          */
3980         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3981
3982         err = trace__expand_filters(trace, &evsel);
3983         if (err)
3984                 goto out_delete_evlist;
3985         err = evlist__apply_filters(evlist, &evsel);
3986         if (err < 0)
3987                 goto out_error_apply_filters;
3988
3989         if (trace->dump.map)
3990                 bpf_map__fprintf(trace->dump.map, trace->output);
3991
3992         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3993         if (err < 0)
3994                 goto out_error_mmap;
3995
3996         if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
3997                 evlist__enable(evlist);
3998
3999         if (forks)
4000                 evlist__start_workload(evlist);
4001
4002         if (trace->opts.target.initial_delay) {
4003                 usleep(trace->opts.target.initial_delay * 1000);
4004                 evlist__enable(evlist);
4005         }
4006
4007         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4008                 perf_thread_map__nr(evlist->core.threads) > 1 ||
4009                 evlist__first(evlist)->core.attr.inherit;
4010
4011         /*
4012          * Now that we already used evsel->core.attr to ask the kernel to setup the
4013          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4014          * trace__resolve_callchain(), allowing per-event max-stack settings
4015          * to override an explicitly set --max-stack global setting.
4016          */
4017         evlist__for_each_entry(evlist, evsel) {
4018                 if (evsel__has_callchain(evsel) &&
4019                     evsel->core.attr.sample_max_stack == 0)
4020                         evsel->core.attr.sample_max_stack = trace->max_stack;
4021         }
4022 again:
4023         before = trace->nr_events;
4024
4025         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4026                 union perf_event *event;
4027                 struct mmap *md;
4028
4029                 md = &evlist->mmap[i];
4030                 if (perf_mmap__read_init(&md->core) < 0)
4031                         continue;
4032
4033                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4034                         ++trace->nr_events;
4035
4036                         err = trace__deliver_event(trace, event);
4037                         if (err)
4038                                 goto out_disable;
4039
4040                         perf_mmap__consume(&md->core);
4041
4042                         if (interrupted)
4043                                 goto out_disable;
4044
4045                         if (done && !draining) {
4046                                 evlist__disable(evlist);
4047                                 draining = true;
4048                         }
4049                 }
4050                 perf_mmap__read_done(&md->core);
4051         }
4052
4053         if (trace->nr_events == before) {
4054                 int timeout = done ? 100 : -1;
4055
4056                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4057                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4058                                 draining = true;
4059
4060                         goto again;
4061                 } else {
4062                         if (trace__flush_events(trace))
4063                                 goto out_disable;
4064                 }
4065         } else {
4066                 goto again;
4067         }
4068
4069 out_disable:
4070         thread__zput(trace->current);
4071
4072         evlist__disable(evlist);
4073
4074         if (trace->sort_events)
4075                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4076
4077         if (!err) {
4078                 if (trace->summary)
4079                         trace__fprintf_thread_summary(trace, trace->output);
4080
4081                 if (trace->show_tool_stats) {
4082                         fprintf(trace->output, "Stats:\n "
4083                                                " vfs_getname : %" PRIu64 "\n"
4084                                                " proc_getname: %" PRIu64 "\n",
4085                                 trace->stats.vfs_getname,
4086                                 trace->stats.proc_getname);
4087                 }
4088         }
4089
4090 out_delete_evlist:
4091         trace__symbols__exit(trace);
4092         evlist__free_syscall_tp_fields(evlist);
4093         evlist__delete(evlist);
4094         cgroup__put(trace->cgroup);
4095         trace->evlist = NULL;
4096         trace->live = false;
4097         return err;
4098 {
4099         char errbuf[BUFSIZ];
4100
4101 out_error_sched_stat_runtime:
4102         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4103         goto out_error;
4104
4105 out_error_raw_syscalls:
4106         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4107         goto out_error;
4108
4109 out_error_mmap:
4110         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4111         goto out_error;
4112
4113 out_error_open:
4114         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4115
4116 out_error:
4117         fprintf(trace->output, "%s\n", errbuf);
4118         goto out_delete_evlist;
4119
4120 out_error_apply_filters:
4121         fprintf(trace->output,
4122                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4123                 evsel->filter, evsel__name(evsel), errno,
4124                 str_error_r(errno, errbuf, sizeof(errbuf)));
4125         goto out_delete_evlist;
4126 }
4127 out_error_mem:
4128         fprintf(trace->output, "Not enough memory to run!\n");
4129         goto out_delete_evlist;
4130
4131 out_errno:
4132         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4133         goto out_delete_evlist;
4134 }
4135
4136 static int trace__replay(struct trace *trace)
4137 {
4138         const struct evsel_str_handler handlers[] = {
4139                 { "probe:vfs_getname",       trace__vfs_getname, },
4140         };
4141         struct perf_data data = {
4142                 .path  = input_name,
4143                 .mode  = PERF_DATA_MODE_READ,
4144                 .force = trace->force,
4145         };
4146         struct perf_session *session;
4147         struct evsel *evsel;
4148         int err = -1;
4149
4150         trace->tool.sample        = trace__process_sample;
4151         trace->tool.mmap          = perf_event__process_mmap;
4152         trace->tool.mmap2         = perf_event__process_mmap2;
4153         trace->tool.comm          = perf_event__process_comm;
4154         trace->tool.exit          = perf_event__process_exit;
4155         trace->tool.fork          = perf_event__process_fork;
4156         trace->tool.attr          = perf_event__process_attr;
4157         trace->tool.tracing_data  = perf_event__process_tracing_data;
4158         trace->tool.build_id      = perf_event__process_build_id;
4159         trace->tool.namespaces    = perf_event__process_namespaces;
4160
4161         trace->tool.ordered_events = true;
4162         trace->tool.ordering_requires_timestamps = true;
4163
4164         /* add tid to output */
4165         trace->multiple_threads = true;
4166
4167         session = perf_session__new(&data, &trace->tool);
4168         if (IS_ERR(session))
4169                 return PTR_ERR(session);
4170
4171         if (trace->opts.target.pid)
4172                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4173
4174         if (trace->opts.target.tid)
4175                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4176
4177         if (symbol__init(&session->header.env) < 0)
4178                 goto out;
4179
4180         trace->host = &session->machines.host;
4181
4182         err = perf_session__set_tracepoints_handlers(session, handlers);
4183         if (err)
4184                 goto out;
4185
4186         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4187         trace->syscalls.events.sys_enter = evsel;
4188         /* older kernels have syscalls tp versus raw_syscalls */
4189         if (evsel == NULL)
4190                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4191
4192         if (evsel &&
4193             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4194             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4195                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4196                 goto out;
4197         }
4198
4199         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4200         trace->syscalls.events.sys_exit = evsel;
4201         if (evsel == NULL)
4202                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4203         if (evsel &&
4204             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4205             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4206                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4207                 goto out;
4208         }
4209
4210         evlist__for_each_entry(session->evlist, evsel) {
4211                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4212                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4213                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4214                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4215                         evsel->handler = trace__pgfault;
4216         }
4217
4218         setup_pager();
4219
4220         err = perf_session__process_events(session);
4221         if (err)
4222                 pr_err("Failed to process events, error %d", err);
4223
4224         else if (trace->summary)
4225                 trace__fprintf_thread_summary(trace, trace->output);
4226
4227 out:
4228         perf_session__delete(session);
4229
4230         return err;
4231 }
4232
4233 static size_t trace__fprintf_threads_header(FILE *fp)
4234 {
4235         size_t printed;
4236
4237         printed  = fprintf(fp, "\n Summary of events:\n\n");
4238
4239         return printed;
4240 }
4241
4242 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4243         struct syscall_stats *stats;
4244         double               msecs;
4245         int                  syscall;
4246 )
4247 {
4248         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4249         struct syscall_stats *stats = source->priv;
4250
4251         entry->syscall = source->i;
4252         entry->stats   = stats;
4253         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4254 }
4255
4256 static size_t thread__dump_stats(struct thread_trace *ttrace,
4257                                  struct trace *trace, FILE *fp)
4258 {
4259         size_t printed = 0;
4260         struct syscall *sc;
4261         struct rb_node *nd;
4262         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4263
4264         if (syscall_stats == NULL)
4265                 return 0;
4266
4267         printed += fprintf(fp, "\n");
4268
4269         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4270         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4271         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4272
4273         resort_rb__for_each_entry(nd, syscall_stats) {
4274                 struct syscall_stats *stats = syscall_stats_entry->stats;
4275                 if (stats) {
4276                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4277                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4278                         double avg = avg_stats(&stats->stats);
4279                         double pct;
4280                         u64 n = (u64)stats->stats.n;
4281
4282                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4283                         avg /= NSEC_PER_MSEC;
4284
4285                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4286                         printed += fprintf(fp, "   %-15s", sc->name);
4287                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4288                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4289                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4290
4291                         if (trace->errno_summary && stats->nr_failures) {
4292                                 const char *arch_name = perf_env__arch(trace->host->env);
4293                                 int e;
4294
4295                                 for (e = 0; e < stats->max_errno; ++e) {
4296                                         if (stats->errnos[e] != 0)
4297                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4298                                 }
4299                         }
4300                 }
4301         }
4302
4303         resort_rb__delete(syscall_stats);
4304         printed += fprintf(fp, "\n\n");
4305
4306         return printed;
4307 }
4308
4309 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4310 {
4311         size_t printed = 0;
4312         struct thread_trace *ttrace = thread__priv(thread);
4313         double ratio;
4314
4315         if (ttrace == NULL)
4316                 return 0;
4317
4318         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4319
4320         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4321         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4322         printed += fprintf(fp, "%.1f%%", ratio);
4323         if (ttrace->pfmaj)
4324                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4325         if (ttrace->pfmin)
4326                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4327         if (trace->sched)
4328                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4329         else if (fputc('\n', fp) != EOF)
4330                 ++printed;
4331
4332         printed += thread__dump_stats(ttrace, trace, fp);
4333
4334         return printed;
4335 }
4336
4337 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4338 {
4339         return ttrace ? ttrace->nr_events : 0;
4340 }
4341
4342 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4343         struct thread *thread;
4344 )
4345 {
4346         entry->thread = rb_entry(nd, struct thread, rb_node);
4347 }
4348
4349 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4350 {
4351         size_t printed = trace__fprintf_threads_header(fp);
4352         struct rb_node *nd;
4353         int i;
4354
4355         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4356                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4357
4358                 if (threads == NULL) {
4359                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4360                         return 0;
4361                 }
4362
4363                 resort_rb__for_each_entry(nd, threads)
4364                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4365
4366                 resort_rb__delete(threads);
4367         }
4368         return printed;
4369 }
4370
4371 static int trace__set_duration(const struct option *opt, const char *str,
4372                                int unset __maybe_unused)
4373 {
4374         struct trace *trace = opt->value;
4375
4376         trace->duration_filter = atof(str);
4377         return 0;
4378 }
4379
4380 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4381                                               int unset __maybe_unused)
4382 {
4383         int ret = -1;
4384         size_t i;
4385         struct trace *trace = opt->value;
4386         /*
4387          * FIXME: introduce a intarray class, plain parse csv and create a
4388          * { int nr, int entries[] } struct...
4389          */
4390         struct intlist *list = intlist__new(str);
4391
4392         if (list == NULL)
4393                 return -1;
4394
4395         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4396         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4397
4398         if (trace->filter_pids.entries == NULL)
4399                 goto out;
4400
4401         trace->filter_pids.entries[0] = getpid();
4402
4403         for (i = 1; i < trace->filter_pids.nr; ++i)
4404                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4405
4406         intlist__delete(list);
4407         ret = 0;
4408 out:
4409         return ret;
4410 }
4411
4412 static int trace__open_output(struct trace *trace, const char *filename)
4413 {
4414         struct stat st;
4415
4416         if (!stat(filename, &st) && st.st_size) {
4417                 char oldname[PATH_MAX];
4418
4419                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4420                 unlink(oldname);
4421                 rename(filename, oldname);
4422         }
4423
4424         trace->output = fopen(filename, "w");
4425
4426         return trace->output == NULL ? -errno : 0;
4427 }
4428
4429 static int parse_pagefaults(const struct option *opt, const char *str,
4430                             int unset __maybe_unused)
4431 {
4432         int *trace_pgfaults = opt->value;
4433
4434         if (strcmp(str, "all") == 0)
4435                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4436         else if (strcmp(str, "maj") == 0)
4437                 *trace_pgfaults |= TRACE_PFMAJ;
4438         else if (strcmp(str, "min") == 0)
4439                 *trace_pgfaults |= TRACE_PFMIN;
4440         else
4441                 return -1;
4442
4443         return 0;
4444 }
4445
4446 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4447 {
4448         struct evsel *evsel;
4449
4450         evlist__for_each_entry(evlist, evsel) {
4451                 if (evsel->handler == NULL)
4452                         evsel->handler = handler;
4453         }
4454 }
4455
4456 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4457 {
4458         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4459
4460         if (fmt) {
4461                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4462
4463                 if (scfmt) {
4464                         int skip = 0;
4465
4466                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4467                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4468                                 ++skip;
4469
4470                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4471                 }
4472         }
4473 }
4474
4475 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4476 {
4477         struct evsel *evsel;
4478
4479         evlist__for_each_entry(evlist, evsel) {
4480                 if (evsel->priv || !evsel->tp_format)
4481                         continue;
4482
4483                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4484                         evsel__init_tp_arg_scnprintf(evsel);
4485                         continue;
4486                 }
4487
4488                 if (evsel__init_syscall_tp(evsel))
4489                         return -1;
4490
4491                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4492                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4493
4494                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4495                                 return -1;
4496
4497                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4498                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4499                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4500
4501                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4502                                 return -1;
4503
4504                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4505                 }
4506         }
4507
4508         return 0;
4509 }
4510
4511 /*
4512  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4513  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4514  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4515  *
4516  * It'd be better to introduce a parse_options() variant that would return a
4517  * list with the terms it didn't match to an event...
4518  */
4519 static int trace__parse_events_option(const struct option *opt, const char *str,
4520                                       int unset __maybe_unused)
4521 {
4522         struct trace *trace = (struct trace *)opt->value;
4523         const char *s = str;
4524         char *sep = NULL, *lists[2] = { NULL, NULL, };
4525         int len = strlen(str) + 1, err = -1, list, idx;
4526         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4527         char group_name[PATH_MAX];
4528         struct syscall_fmt *fmt;
4529
4530         if (strace_groups_dir == NULL)
4531                 return -1;
4532
4533         if (*s == '!') {
4534                 ++s;
4535                 trace->not_ev_qualifier = true;
4536         }
4537
4538         while (1) {
4539                 if ((sep = strchr(s, ',')) != NULL)
4540                         *sep = '\0';
4541
4542                 list = 0;
4543                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4544                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4545                         list = 1;
4546                         goto do_concat;
4547                 }
4548
4549                 fmt = syscall_fmt__find_by_alias(s);
4550                 if (fmt != NULL) {
4551                         list = 1;
4552                         s = fmt->name;
4553                 } else {
4554                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4555                         if (access(group_name, R_OK) == 0)
4556                                 list = 1;
4557                 }
4558 do_concat:
4559                 if (lists[list]) {
4560                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4561                 } else {
4562                         lists[list] = malloc(len);
4563                         if (lists[list] == NULL)
4564                                 goto out;
4565                         strcpy(lists[list], s);
4566                 }
4567
4568                 if (!sep)
4569                         break;
4570
4571                 *sep = ',';
4572                 s = sep + 1;
4573         }
4574
4575         if (lists[1] != NULL) {
4576                 struct strlist_config slist_config = {
4577                         .dirname = strace_groups_dir,
4578                 };
4579
4580                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4581                 if (trace->ev_qualifier == NULL) {
4582                         fputs("Not enough memory to parse event qualifier", trace->output);
4583                         goto out;
4584                 }
4585
4586                 if (trace__validate_ev_qualifier(trace))
4587                         goto out;
4588                 trace->trace_syscalls = true;
4589         }
4590
4591         err = 0;
4592
4593         if (lists[0]) {
4594                 struct option o = {
4595                         .value = &trace->evlist,
4596                 };
4597                 err = parse_events_option(&o, lists[0], 0);
4598         }
4599 out:
4600         free(strace_groups_dir);
4601         free(lists[0]);
4602         free(lists[1]);
4603         if (sep)
4604                 *sep = ',';
4605
4606         return err;
4607 }
4608
4609 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4610 {
4611         struct trace *trace = opt->value;
4612
4613         if (!list_empty(&trace->evlist->core.entries)) {
4614                 struct option o = {
4615                         .value = &trace->evlist,
4616                 };
4617                 return parse_cgroups(&o, str, unset);
4618         }
4619         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4620
4621         return 0;
4622 }
4623
4624 static int trace__config(const char *var, const char *value, void *arg)
4625 {
4626         struct trace *trace = arg;
4627         int err = 0;
4628
4629         if (!strcmp(var, "trace.add_events")) {
4630                 trace->perfconfig_events = strdup(value);
4631                 if (trace->perfconfig_events == NULL) {
4632                         pr_err("Not enough memory for %s\n", "trace.add_events");
4633                         return -1;
4634                 }
4635         } else if (!strcmp(var, "trace.show_timestamp")) {
4636                 trace->show_tstamp = perf_config_bool(var, value);
4637         } else if (!strcmp(var, "trace.show_duration")) {
4638                 trace->show_duration = perf_config_bool(var, value);
4639         } else if (!strcmp(var, "trace.show_arg_names")) {
4640                 trace->show_arg_names = perf_config_bool(var, value);
4641                 if (!trace->show_arg_names)
4642                         trace->show_zeros = true;
4643         } else if (!strcmp(var, "trace.show_zeros")) {
4644                 bool new_show_zeros = perf_config_bool(var, value);
4645                 if (!trace->show_arg_names && !new_show_zeros) {
4646                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4647                         goto out;
4648                 }
4649                 trace->show_zeros = new_show_zeros;
4650         } else if (!strcmp(var, "trace.show_prefix")) {
4651                 trace->show_string_prefix = perf_config_bool(var, value);
4652         } else if (!strcmp(var, "trace.no_inherit")) {
4653                 trace->opts.no_inherit = perf_config_bool(var, value);
4654         } else if (!strcmp(var, "trace.args_alignment")) {
4655                 int args_alignment = 0;
4656                 if (perf_config_int(&args_alignment, var, value) == 0)
4657                         trace->args_alignment = args_alignment;
4658         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4659                 if (strcasecmp(value, "libtraceevent") == 0)
4660                         trace->libtraceevent_print = true;
4661                 else if (strcasecmp(value, "libbeauty") == 0)
4662                         trace->libtraceevent_print = false;
4663         }
4664 out:
4665         return err;
4666 }
4667
4668 static void trace__exit(struct trace *trace)
4669 {
4670         int i;
4671
4672         strlist__delete(trace->ev_qualifier);
4673         free(trace->ev_qualifier_ids.entries);
4674         if (trace->syscalls.table) {
4675                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4676                         syscall__exit(&trace->syscalls.table[i]);
4677                 free(trace->syscalls.table);
4678         }
4679         syscalltbl__delete(trace->sctbl);
4680         zfree(&trace->perfconfig_events);
4681 }
4682
4683 int cmd_trace(int argc, const char **argv)
4684 {
4685         const char *trace_usage[] = {
4686                 "perf trace [<options>] [<command>]",
4687                 "perf trace [<options>] -- <command> [<options>]",
4688                 "perf trace record [<options>] [<command>]",
4689                 "perf trace record [<options>] -- <command> [<options>]",
4690                 NULL
4691         };
4692         struct trace trace = {
4693                 .opts = {
4694                         .target = {
4695                                 .uid       = UINT_MAX,
4696                                 .uses_mmap = true,
4697                         },
4698                         .user_freq     = UINT_MAX,
4699                         .user_interval = ULLONG_MAX,
4700                         .no_buffering  = true,
4701                         .mmap_pages    = UINT_MAX,
4702                 },
4703                 .output = stderr,
4704                 .show_comm = true,
4705                 .show_tstamp = true,
4706                 .show_duration = true,
4707                 .show_arg_names = true,
4708                 .args_alignment = 70,
4709                 .trace_syscalls = false,
4710                 .kernel_syscallchains = false,
4711                 .max_stack = UINT_MAX,
4712                 .max_events = ULONG_MAX,
4713         };
4714         const char *map_dump_str = NULL;
4715         const char *output_name = NULL;
4716         const struct option trace_options[] = {
4717         OPT_CALLBACK('e', "event", &trace, "event",
4718                      "event/syscall selector. use 'perf list' to list available events",
4719                      trace__parse_events_option),
4720         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4721                      "event filter", parse_filter),
4722         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4723                     "show the thread COMM next to its id"),
4724         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4725         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4726                      trace__parse_events_option),
4727         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4728         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4729         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4730                     "trace events on existing process id"),
4731         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4732                     "trace events on existing thread id"),
4733         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4734                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4735         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4736                     "system-wide collection from all CPUs"),
4737         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4738                     "list of cpus to monitor"),
4739         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4740                     "child tasks do not inherit counters"),
4741         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4742                      "number of mmap data pages", evlist__parse_mmap_pages),
4743         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4744                    "user to profile"),
4745         OPT_CALLBACK(0, "duration", &trace, "float",
4746                      "show only events with duration > N.M ms",
4747                      trace__set_duration),
4748 #ifdef HAVE_LIBBPF_SUPPORT
4749         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4750 #endif
4751         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4752         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4753         OPT_BOOLEAN('T', "time", &trace.full_time,
4754                     "Show full timestamp, not time relative to first start"),
4755         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4756                     "Show only syscalls that failed"),
4757         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4758                     "Show only syscall summary with statistics"),
4759         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4760                     "Show all syscalls and summary with statistics"),
4761         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4762                     "Show errno stats per syscall, use with -s or -S"),
4763         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4764                      "Trace pagefaults", parse_pagefaults, "maj"),
4765         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4766         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4767         OPT_CALLBACK(0, "call-graph", &trace.opts,
4768                      "record_mode[,record_size]", record_callchain_help,
4769                      &record_parse_callchain_opt),
4770         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4771                     "Use libtraceevent to print the tracepoint arguments."),
4772         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4773                     "Show the kernel callchains on the syscall exit path"),
4774         OPT_ULONG(0, "max-events", &trace.max_events,
4775                 "Set the maximum number of events to print, exit after that is reached. "),
4776         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4777                      "Set the minimum stack depth when parsing the callchain, "
4778                      "anything below the specified depth will be ignored."),
4779         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4780                      "Set the maximum stack depth when parsing the callchain, "
4781                      "anything beyond the specified depth will be ignored. "
4782                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4783         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4784                         "Sort batch of events before processing, use if getting out of order events"),
4785         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4786                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4787         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4788                         "per thread proc mmap processing timeout in ms"),
4789         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4790                      trace__parse_cgroups),
4791         OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4792                      "ms to wait before starting measurement after program "
4793                      "start"),
4794         OPTS_EVSWITCH(&trace.evswitch),
4795         OPT_END()
4796         };
4797         bool __maybe_unused max_stack_user_set = true;
4798         bool mmap_pages_user_set = true;
4799         struct evsel *evsel;
4800         const char * const trace_subcommands[] = { "record", NULL };
4801         int err = -1;
4802         char bf[BUFSIZ];
4803         struct sigaction sigchld_act;
4804
4805         signal(SIGSEGV, sighandler_dump_stack);
4806         signal(SIGFPE, sighandler_dump_stack);
4807         signal(SIGINT, sighandler_interrupt);
4808
4809         memset(&sigchld_act, 0, sizeof(sigchld_act));
4810         sigchld_act.sa_flags = SA_SIGINFO;
4811         sigchld_act.sa_sigaction = sighandler_chld;
4812         sigaction(SIGCHLD, &sigchld_act, NULL);
4813
4814         trace.evlist = evlist__new();
4815         trace.sctbl = syscalltbl__new();
4816
4817         if (trace.evlist == NULL || trace.sctbl == NULL) {
4818                 pr_err("Not enough memory to run!\n");
4819                 err = -ENOMEM;
4820                 goto out;
4821         }
4822
4823         /*
4824          * Parsing .perfconfig may entail creating a BPF event, that may need
4825          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4826          * is too small. This affects just this process, not touching the
4827          * global setting. If it fails we'll get something in 'perf trace -v'
4828          * to help diagnose the problem.
4829          */
4830         rlimit__bump_memlock();
4831
4832         err = perf_config(trace__config, &trace);
4833         if (err)
4834                 goto out;
4835
4836         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4837                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4838
4839         /*
4840          * Here we already passed thru trace__parse_events_option() and it has
4841          * already figured out if -e syscall_name, if not but if --event
4842          * foo:bar was used, the user is interested _just_ in those, say,
4843          * tracepoint events, not in the strace-like syscall-name-based mode.
4844          *
4845          * This is important because we need to check if strace-like mode is
4846          * needed to decided if we should filter out the eBPF
4847          * __augmented_syscalls__ code, if it is in the mix, say, via
4848          * .perfconfig trace.add_events, and filter those out.
4849          */
4850         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4851             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4852                 trace.trace_syscalls = true;
4853         }
4854         /*
4855          * Now that we have --verbose figured out, lets see if we need to parse
4856          * events from .perfconfig, so that if those events fail parsing, say some
4857          * BPF program fails, then we'll be able to use --verbose to see what went
4858          * wrong in more detail.
4859          */
4860         if (trace.perfconfig_events != NULL) {
4861                 struct parse_events_error parse_err;
4862
4863                 parse_events_error__init(&parse_err);
4864                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4865                 if (err)
4866                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4867                 parse_events_error__exit(&parse_err);
4868                 if (err)
4869                         goto out;
4870         }
4871
4872         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4873                 usage_with_options_msg(trace_usage, trace_options,
4874                                        "cgroup monitoring only available in system-wide mode");
4875         }
4876
4877         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4878         if (IS_ERR(evsel)) {
4879                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4880                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4881                 goto out;
4882         }
4883
4884         if (evsel) {
4885                 trace.syscalls.events.augmented = evsel;
4886
4887                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4888                 if (evsel == NULL) {
4889                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4890                         goto out;
4891                 }
4892
4893                 if (evsel->bpf_obj == NULL) {
4894                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4895                         goto out;
4896                 }
4897
4898                 trace.bpf_obj = evsel->bpf_obj;
4899
4900                 /*
4901                  * If we have _just_ the augmenter event but don't have a
4902                  * explicit --syscalls, then assume we want all strace-like
4903                  * syscalls:
4904                  */
4905                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4906                         trace.trace_syscalls = true;
4907                 /*
4908                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4909                  * strace-like syscall tracing is not set, then we need to trow
4910                  * away the augmenter, i.e. all the events that were created
4911                  * from that BPF object file.
4912                  *
4913                  * This is more to fix the current .perfconfig trace.add_events
4914                  * style of setting up the strace-like eBPF based syscall point
4915                  * payload augmenter.
4916                  *
4917                  * All this complexity will be avoided by adding an alternative
4918                  * to trace.add_events in the form of
4919                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4920                  * need it.
4921                  *
4922                  * .perfconfig trace.add_events is still useful if we want, for
4923                  * instance, have msr_write.msr in some .perfconfig profile based
4924                  * 'perf trace --config determinism.profile' mode, where for some
4925                  * particular goal/workload type we want a set of events and
4926                  * output mode (with timings, etc) instead of having to add
4927                  * all via the command line.
4928                  *
4929                  * Also --config to specify an alternate .perfconfig file needs
4930                  * to be implemented.
4931                  */
4932                 if (!trace.trace_syscalls) {
4933                         trace__delete_augmented_syscalls(&trace);
4934                 } else {
4935                         trace__set_bpf_map_filtered_pids(&trace);
4936                         trace__set_bpf_map_syscalls(&trace);
4937                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4938                 }
4939         }
4940
4941         err = bpf__setup_stdout(trace.evlist);
4942         if (err) {
4943                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4944                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4945                 goto out;
4946         }
4947
4948         err = -1;
4949
4950         if (map_dump_str) {
4951                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4952                 if (trace.dump.map == NULL) {
4953                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4954                         goto out;
4955                 }
4956         }
4957
4958         if (trace.trace_pgfaults) {
4959                 trace.opts.sample_address = true;
4960                 trace.opts.sample_time = true;
4961         }
4962
4963         if (trace.opts.mmap_pages == UINT_MAX)
4964                 mmap_pages_user_set = false;
4965
4966         if (trace.max_stack == UINT_MAX) {
4967                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4968                 max_stack_user_set = false;
4969         }
4970
4971 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4972         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4973                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4974         }
4975 #endif
4976
4977         if (callchain_param.enabled) {
4978                 if (!mmap_pages_user_set && geteuid() == 0)
4979                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4980
4981                 symbol_conf.use_callchain = true;
4982         }
4983
4984         if (trace.evlist->core.nr_entries > 0) {
4985                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4986                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4987                         perror("failed to set syscalls:* tracepoint fields");
4988                         goto out;
4989                 }
4990         }
4991
4992         if (trace.sort_events) {
4993                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4994                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4995         }
4996
4997         /*
4998          * If we are augmenting syscalls, then combine what we put in the
4999          * __augmented_syscalls__ BPF map with what is in the
5000          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5001          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5002          *
5003          * We'll switch to look at two BPF maps, one for sys_enter and the
5004          * other for sys_exit when we start augmenting the sys_exit paths with
5005          * buffers that are being copied from kernel to userspace, think 'read'
5006          * syscall.
5007          */
5008         if (trace.syscalls.events.augmented) {
5009                 evlist__for_each_entry(trace.evlist, evsel) {
5010                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5011
5012                         if (raw_syscalls_sys_exit) {
5013                                 trace.raw_augmented_syscalls = true;
5014                                 goto init_augmented_syscall_tp;
5015                         }
5016
5017                         if (trace.syscalls.events.augmented->priv == NULL &&
5018                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5019                                 struct evsel *augmented = trace.syscalls.events.augmented;
5020                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5021                                     evsel__init_augmented_syscall_tp_args(augmented))
5022                                         goto out;
5023                                 /*
5024                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5025                                  * Above we made sure we can get from the payload the tp fields
5026                                  * that we get from syscalls:sys_enter tracefs format file.
5027                                  */
5028                                 augmented->handler = trace__sys_enter;
5029                                 /*
5030                                  * Now we do the same for the *syscalls:sys_enter event so that
5031                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5032                                  * as not to filter it, then we'll handle it just like we would
5033                                  * for the BPF_OUTPUT one:
5034                                  */
5035                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5036                                     evsel__init_augmented_syscall_tp_args(evsel))
5037                                         goto out;
5038                                 evsel->handler = trace__sys_enter;
5039                         }
5040
5041                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5042                                 struct syscall_tp *sc;
5043 init_augmented_syscall_tp:
5044                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5045                                         goto out;
5046                                 sc = __evsel__syscall_tp(evsel);
5047                                 /*
5048                                  * For now with BPF raw_augmented we hook into
5049                                  * raw_syscalls:sys_enter and there we get all
5050                                  * 6 syscall args plus the tracepoint common
5051                                  * fields and the syscall_nr (another long).
5052                                  * So we check if that is the case and if so
5053                                  * don't look after the sc->args_size but
5054                                  * always after the full raw_syscalls:sys_enter
5055                                  * payload, which is fixed.
5056                                  *
5057                                  * We'll revisit this later to pass
5058                                  * s->args_size to the BPF augmenter (now
5059                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5060                                  * so that it copies only what we need for each
5061                                  * syscall, like what happens when we use
5062                                  * syscalls:sys_enter_NAME, so that we reduce
5063                                  * the kernel/userspace traffic to just what is
5064                                  * needed for each syscall.
5065                                  */
5066                                 if (trace.raw_augmented_syscalls)
5067                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5068                                 evsel__init_augmented_syscall_tp_ret(evsel);
5069                                 evsel->handler = trace__sys_exit;
5070                         }
5071                 }
5072         }
5073
5074         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5075                 return trace__record(&trace, argc-1, &argv[1]);
5076
5077         /* Using just --errno-summary will trigger --summary */
5078         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5079                 trace.summary_only = true;
5080
5081         /* summary_only implies summary option, but don't overwrite summary if set */
5082         if (trace.summary_only)
5083                 trace.summary = trace.summary_only;
5084
5085         if (output_name != NULL) {
5086                 err = trace__open_output(&trace, output_name);
5087                 if (err < 0) {
5088                         perror("failed to create output file");
5089                         goto out;
5090                 }
5091         }
5092
5093         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5094         if (err)
5095                 goto out_close;
5096
5097         err = target__validate(&trace.opts.target);
5098         if (err) {
5099                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5100                 fprintf(trace.output, "%s", bf);
5101                 goto out_close;
5102         }
5103
5104         err = target__parse_uid(&trace.opts.target);
5105         if (err) {
5106                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5107                 fprintf(trace.output, "%s", bf);
5108                 goto out_close;
5109         }
5110
5111         if (!argc && target__none(&trace.opts.target))
5112                 trace.opts.target.system_wide = true;
5113
5114         if (input_name)
5115                 err = trace__replay(&trace);
5116         else
5117                 err = trace__run(&trace, argc, argv);
5118
5119 out_close:
5120         if (output_name != NULL)
5121                 fclose(trace.output);
5122 out:
5123         trace__exit(&trace);
5124         return err;
5125 }