tools/perf/builtin-trace.c

   1 #include "builtin.h"
   2 #include "util/evlist.h"
   3 #include "util/parse-options.h"
   4 #include "util/thread_map.h"
   5 #include "event-parse.h"
   6
   7 #include <libaudit.h>
   8 #include <stdlib.h>
   9
  10 static struct syscall_fmt {
  11         const char *name;
  12         const char *alias;
  13         bool       errmsg;
  14         bool       timeout;
  15 } syscall_fmts[] = {
  16         { .name     = "arch_prctl", .errmsg = true, .alias = "prctl", },
  17         { .name     = "fstat",      .errmsg = true, .alias = "newfstat", },
  18         { .name     = "fstatat",    .errmsg = true, .alias = "newfstatat", },
  19         { .name     = "futex",      .errmsg = true, },
  20         { .name     = "poll",       .errmsg = true, .timeout = true, },
  21         { .name     = "ppoll",      .errmsg = true, .timeout = true, },
  22         { .name     = "read",       .errmsg = true, },
  23         { .name     = "recvfrom",   .errmsg = true, },
  24         { .name     = "select",     .errmsg = true, .timeout = true, },
  25         { .name     = "stat",       .errmsg = true, .alias = "newstat", },
  26 };
  27
  28 static int syscall_fmt__cmp(const void *name, const void *fmtp)
  29 {
  30         const struct syscall_fmt *fmt = fmtp;
  31         return strcmp(name, fmt->name);
  32 }
  33
  34 static struct syscall_fmt *syscall_fmt__find(const char *name)
  35 {
  36         const int nmemb = ARRAY_SIZE(syscall_fmts);
  37         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
  38 }
  39
  40 struct syscall {
  41         struct event_format *tp_format;
  42         const char          *name;
  43         struct syscall_fmt  *fmt;
  44 };
  45
  46 struct trace {
  47         int                     audit_machine;
  48         struct {
  49                 int             max;
  50                 struct syscall  *table;
  51         } syscalls;
  52         struct perf_record_opts opts;
  53 };
  54
  55 static bool done = false;
  56
  57 static void sig_handler(int sig __maybe_unused)
  58 {
  59         done = true;
  60 }
  61
  62 static int trace__read_syscall_info(struct trace *trace, int id)
  63 {
  64         char tp_name[128];
  65         struct syscall *sc;
  66         const char *name = audit_syscall_to_name(id, trace->audit_machine);
  67
  68         if (name == NULL)
  69                 return -1;
  70
  71         if (id > trace->syscalls.max) {
  72                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
  73
  74                 if (nsyscalls == NULL)
  75                         return -1;
  76
  77                 if (trace->syscalls.max != -1) {
  78                         memset(nsyscalls + trace->syscalls.max + 1, 0,
  79                                (id - trace->syscalls.max) * sizeof(*sc));
  80                 } else {
  81                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
  82                 }
  83
  84                 trace->syscalls.table = nsyscalls;
  85                 trace->syscalls.max   = id;
  86         }
  87
  88         sc = trace->syscalls.table + id;
  89         sc->name = name;
  90         sc->fmt  = syscall_fmt__find(sc->name);
  91
  92         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
  93         sc->tp_format = event_format__new("syscalls", tp_name);
  94
  95         if (sc->tp_format == NULL && sc->fmt && sc->fmt->alias) {
  96                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
  97                 sc->tp_format = event_format__new("syscalls", tp_name);
  98         }
  99
 100         return sc->tp_format != NULL ? 0 : -1;
 101 }
 102
 103 static size_t syscall__fprintf_args(struct syscall *sc, unsigned long *args, FILE *fp)
 104 {
 105         int i = 0;
 106         size_t printed = 0;
 107
 108         if (sc->tp_format != NULL) {
 109                 struct format_field *field;
 110
 111                 for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 112                         printed += fprintf(fp, "%s%s: %ld", printed ? ", " : "",
 113                                            field->name, args[i++]);
 114                 }
 115         } else {
 116                 while (i < 6) {
 117                         printed += fprintf(fp, "%sarg%d: %ld", printed ? ", " : "", i, args[i]);
 118                         ++i;
 119                 }
 120         }
 121
 122         return printed;
 123 }
 124
 125 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
 126                                   struct perf_sample *sample);
 127
 128 static struct syscall *trace__syscall_info(struct trace *trace,
 129                                            struct perf_evsel *evsel,
 130                                            struct perf_sample *sample)
 131 {
 132         int id = perf_evsel__intval(evsel, sample, "id");
 133
 134         if (id < 0) {
 135                 printf("Invalid syscall %d id, skipping...\n", id);
 136                 return NULL;
 137         }
 138
 139         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
 140             trace__read_syscall_info(trace, id))
 141                 goto out_cant_read;
 142
 143         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
 144                 goto out_cant_read;
 145
 146         return &trace->syscalls.table[id];
 147
 148 out_cant_read:
 149         printf("Problems reading syscall %d information\n", id);
 150         return NULL;
 151 }
 152
 153 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
 154                             struct perf_sample *sample)
 155 {
 156         void *args;
 157         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 158
 159         if (sc == NULL)
 160                 return -1;
 161
 162         args = perf_evsel__rawptr(evsel, sample, "args");
 163         if (args == NULL) {
 164                 printf("Problems reading syscall arguments\n");
 165                 return -1;
 166         }
 167
 168         printf("%s(", sc->name);
 169         syscall__fprintf_args(sc, args, stdout);
 170
 171         return 0;
 172 }
 173
 174 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
 175                            struct perf_sample *sample)
 176 {
 177         int ret;
 178         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 179
 180         if (sc == NULL)
 181                 return -1;
 182
 183         ret = perf_evsel__intval(evsel, sample, "ret");
 184
 185         if (ret < 0 && sc->fmt && sc->fmt->errmsg) {
 186                 char bf[256];
 187                 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
 188                            *e = audit_errno_to_name(-ret);
 189
 190                 printf(") = -1 %s %s", e, emsg);
 191         } else if (ret == 0 && sc->fmt && sc->fmt->timeout)
 192                 printf(") = 0 Timeout");
 193         else
 194                 printf(") = %d", ret);
 195
 196         putchar('\n');
 197         return 0;
 198 }
 199
 200 static int trace__run(struct trace *trace, int argc, const char **argv)
 201 {
 202         struct perf_evlist *evlist = perf_evlist__new(NULL, NULL);
 203         struct perf_evsel *evsel;
 204         int err = -1, i, nr_events = 0, before;
 205         const bool forks = argc > 0;
 206
 207         if (evlist == NULL) {
 208                 printf("Not enough memory to run!\n");
 209                 goto out;
 210         }
 211
 212         if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
 213             perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit)) {
 214                 printf("Couldn't read the raw_syscalls tracepoints information!\n");
 215                 goto out_delete_evlist;
 216         }
 217
 218         err = perf_evlist__create_maps(evlist, &trace->opts.target);
 219         if (err < 0) {
 220                 printf("Problems parsing the target to trace, check your options!\n");
 221                 goto out_delete_evlist;
 222         }
 223
 224         perf_evlist__config_attrs(evlist, &trace->opts);
 225
 226         signal(SIGCHLD, sig_handler);
 227         signal(SIGINT, sig_handler);
 228
 229         if (forks) {
 230                 err = perf_evlist__prepare_workload(evlist, &trace->opts, argv);
 231                 if (err < 0) {
 232                         printf("Couldn't run the workload!\n");
 233                         goto out_delete_evlist;
 234                 }
 235         }
 236
 237         err = perf_evlist__open(evlist);
 238         if (err < 0) {
 239                 printf("Couldn't create the events: %s\n", strerror(errno));
 240                 goto out_delete_evlist;
 241         }
 242
 243         err = perf_evlist__mmap(evlist, UINT_MAX, false);
 244         if (err < 0) {
 245                 printf("Couldn't mmap the events: %s\n", strerror(errno));
 246                 goto out_delete_evlist;
 247         }
 248
 249         perf_evlist__enable(evlist);
 250
 251         if (forks)
 252                 perf_evlist__start_workload(evlist);
 253
 254 again:
 255         before = nr_events;
 256
 257         for (i = 0; i < evlist->nr_mmaps; i++) {
 258                 union perf_event *event;
 259
 260                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 261                         const u32 type = event->header.type;
 262                         tracepoint_handler handler;
 263                         struct perf_sample sample;
 264
 265                         ++nr_events;
 266
 267                         switch (type) {
 268                         case PERF_RECORD_SAMPLE:
 269                                 break;
 270                         case PERF_RECORD_LOST:
 271                                 printf("LOST %" PRIu64 " events!\n", event->lost.lost);
 272                                 continue;
 273                         default:
 274                                 printf("Unexpected %s event, skipping...\n",
 275                                         perf_event__name(type));
 276                                 continue;
 277                         }
 278
 279                         err = perf_evlist__parse_sample(evlist, event, &sample);
 280                         if (err) {
 281                                 printf("Can't parse sample, err = %d, skipping...\n", err);
 282                                 continue;
 283                         }
 284
 285                         evsel = perf_evlist__id2evsel(evlist, sample.id);
 286                         if (evsel == NULL) {
 287                                 printf("Unknown tp ID %" PRIu64 ", skipping...\n", sample.id);
 288                                 continue;
 289                         }
 290
 291                         if (evlist->threads->map[0] == -1 || evlist->threads->nr > 1)
 292                                 printf("%d ", sample.tid);
 293
 294                         if (sample.raw_data == NULL) {
 295                                 printf("%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
 296                                        perf_evsel__name(evsel), sample.tid,
 297                                        sample.cpu, sample.raw_size);
 298                                 continue;
 299                         }
 300
 301                         handler = evsel->handler.func;
 302                         handler(trace, evsel, &sample);
 303                 }
 304         }
 305
 306         if (nr_events == before) {
 307                 if (done)
 308                         goto out_delete_evlist;
 309
 310                 poll(evlist->pollfd, evlist->nr_fds, -1);
 311         }
 312
 313         if (done)
 314                 perf_evlist__disable(evlist);
 315
 316         goto again;
 317
 318 out_delete_evlist:
 319         perf_evlist__delete(evlist);
 320 out:
 321         return err;
 322 }
 323
 324 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
 325 {
 326         const char * const trace_usage[] = {
 327                 "perf trace [<options>] [<command>]",
 328                 "perf trace [<options>] -- <command> [<options>]",
 329                 NULL
 330         };
 331         struct trace trace = {
 332                 .audit_machine = audit_detect_machine(),
 333                 .syscalls = {
 334                         . max = -1,
 335                 },
 336                 .opts = {
 337                         .target = {
 338                                 .uid       = UINT_MAX,
 339                                 .uses_mmap = true,
 340                         },
 341                         .user_freq     = UINT_MAX,
 342                         .user_interval = ULLONG_MAX,
 343                         .no_delay      = true,
 344                         .mmap_pages    = 1024,
 345                 },
 346         };
 347         const struct option trace_options[] = {
 348         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
 349                     "trace events on existing process id"),
 350         OPT_STRING(0, "tid", &trace.opts.target.tid, "tid",
 351                     "trace events on existing thread id"),
 352         OPT_BOOLEAN(0, "all-cpus", &trace.opts.target.system_wide,
 353                     "system-wide collection from all CPUs"),
 354         OPT_STRING(0, "cpu", &trace.opts.target.cpu_list, "cpu",
 355                     "list of cpus to monitor"),
 356         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
 357                     "child tasks do not inherit counters"),
 358         OPT_UINTEGER(0, "mmap-pages", &trace.opts.mmap_pages,
 359                      "number of mmap data pages"),
 360         OPT_STRING(0, "uid", &trace.opts.target.uid_str, "user",
 361                    "user to profile"),
 362         OPT_END()
 363         };
 364         int err;
 365         char bf[BUFSIZ];
 366
 367         argc = parse_options(argc, argv, trace_options, trace_usage, 0);
 368
 369         err = perf_target__validate(&trace.opts.target);
 370         if (err) {
 371                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
 372                 printf("%s", bf);
 373                 return err;
 374         }
 375
 376         err = perf_target__parse_uid(&trace.opts.target);
 377         if (err) {
 378                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
 379                 printf("%s", bf);
 380                 return err;
 381         }
 382
 383         if (!argc && perf_target__none(&trace.opts.target))
 384                 trace.opts.target.system_wide = true;
 385
 386         return trace__run(&trace, argc, argv);
 387 }