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 int trace__read_syscall_info(struct trace *trace, int id)
  56 {
  57         char tp_name[128];
  58         struct syscall *sc;
  59         const char *name = audit_syscall_to_name(id, trace->audit_machine);
  60
  61         if (name == NULL)
  62                 return -1;
  63
  64         if (id > trace->syscalls.max) {
  65                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
  66
  67                 if (nsyscalls == NULL)
  68                         return -1;
  69
  70                 if (trace->syscalls.max != -1) {
  71                         memset(nsyscalls + trace->syscalls.max + 1, 0,
  72                                (id - trace->syscalls.max) * sizeof(*sc));
  73                 } else {
  74                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
  75                 }
  76
  77                 trace->syscalls.table = nsyscalls;
  78                 trace->syscalls.max   = id;
  79         }
  80
  81         sc = trace->syscalls.table + id;
  82         sc->name = name;
  83         sc->fmt  = syscall_fmt__find(sc->name);
  84
  85         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
  86         sc->tp_format = event_format__new("syscalls", tp_name);
  87
  88         if (sc->tp_format == NULL && sc->fmt && sc->fmt->alias) {
  89                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
  90                 sc->tp_format = event_format__new("syscalls", tp_name);
  91         }
  92
  93         return sc->tp_format != NULL ? 0 : -1;
  94 }
  95
  96 static size_t syscall__fprintf_args(struct syscall *sc, unsigned long *args, FILE *fp)
  97 {
  98         int i = 0;
  99         size_t printed = 0;
 100
 101         if (sc->tp_format != NULL) {
 102                 struct format_field *field;
 103
 104                 for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 105                         printed += fprintf(fp, "%s%s: %ld", printed ? ", " : "",
 106                                            field->name, args[i++]);
 107                 }
 108         } else {
 109                 while (i < 6) {
 110                         printed += fprintf(fp, "%sarg%d: %ld", printed ? ", " : "", i, args[i]);
 111                         ++i;
 112                 }
 113         }
 114
 115         return printed;
 116 }
 117
 118 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
 119                                   struct perf_sample *sample);
 120
 121 static struct syscall *trace__syscall_info(struct trace *trace,
 122                                            struct perf_evsel *evsel,
 123                                            struct perf_sample *sample)
 124 {
 125         int id = perf_evsel__intval(evsel, sample, "id");
 126
 127         if (id < 0) {
 128                 printf("Invalid syscall %d id, skipping...\n", id);
 129                 return NULL;
 130         }
 131
 132         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
 133             trace__read_syscall_info(trace, id))
 134                 goto out_cant_read;
 135
 136         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
 137                 goto out_cant_read;
 138
 139         return &trace->syscalls.table[id];
 140
 141 out_cant_read:
 142         printf("Problems reading syscall %d information\n", id);
 143         return NULL;
 144 }
 145
 146 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
 147                             struct perf_sample *sample)
 148 {
 149         void *args;
 150         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 151
 152         if (sc == NULL)
 153                 return -1;
 154
 155         args = perf_evsel__rawptr(evsel, sample, "args");
 156         if (args == NULL) {
 157                 printf("Problems reading syscall arguments\n");
 158                 return -1;
 159         }
 160
 161         printf("%s(", sc->name);
 162         syscall__fprintf_args(sc, args, stdout);
 163
 164         return 0;
 165 }
 166
 167 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
 168                            struct perf_sample *sample)
 169 {
 170         int ret;
 171         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 172
 173         if (sc == NULL)
 174                 return -1;
 175
 176         ret = perf_evsel__intval(evsel, sample, "ret");
 177
 178         if (ret < 0 && sc->fmt && sc->fmt->errmsg) {
 179                 char bf[256];
 180                 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
 181                            *e = audit_errno_to_name(-ret);
 182
 183                 printf(") = -1 %s %s", e, emsg);
 184         } else if (ret == 0 && sc->fmt && sc->fmt->timeout)
 185                 printf(") = 0 Timeout");
 186         else
 187                 printf(") = %d", ret);
 188
 189         putchar('\n');
 190         return 0;
 191 }
 192
 193 static int trace__run(struct trace *trace)
 194 {
 195         struct perf_evlist *evlist = perf_evlist__new(NULL, NULL);
 196         struct perf_evsel *evsel;
 197         int err = -1, i, nr_events = 0, before;
 198
 199         if (evlist == NULL) {
 200                 printf("Not enough memory to run!\n");
 201                 goto out;
 202         }
 203
 204         if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
 205             perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit)) {
 206                 printf("Couldn't read the raw_syscalls tracepoints information!\n");
 207                 goto out_delete_evlist;
 208         }
 209
 210         err = perf_evlist__create_maps(evlist, &trace->opts.target);
 211         if (err < 0) {
 212                 printf("Problems parsing the target to trace, check your options!\n");
 213                 goto out_delete_evlist;
 214         }
 215
 216         perf_evlist__config_attrs(evlist, &trace->opts);
 217
 218         err = perf_evlist__open(evlist);
 219         if (err < 0) {
 220                 printf("Couldn't create the events: %s\n", strerror(errno));
 221                 goto out_delete_evlist;
 222         }
 223
 224         err = perf_evlist__mmap(evlist, UINT_MAX, false);
 225         if (err < 0) {
 226                 printf("Couldn't mmap the events: %s\n", strerror(errno));
 227                 goto out_delete_evlist;
 228         }
 229
 230         perf_evlist__enable(evlist);
 231 again:
 232         before = nr_events;
 233
 234         for (i = 0; i < evlist->nr_mmaps; i++) {
 235                 union perf_event *event;
 236
 237                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 238                         const u32 type = event->header.type;
 239                         tracepoint_handler handler;
 240                         struct perf_sample sample;
 241
 242                         ++nr_events;
 243
 244                         switch (type) {
 245                         case PERF_RECORD_SAMPLE:
 246                                 break;
 247                         case PERF_RECORD_LOST:
 248                                 printf("LOST %" PRIu64 " events!\n", event->lost.lost);
 249                                 continue;
 250                         default:
 251                                 printf("Unexpected %s event, skipping...\n",
 252                                         perf_event__name(type));
 253                                 continue;
 254                         }
 255
 256                         err = perf_evlist__parse_sample(evlist, event, &sample);
 257                         if (err) {
 258                                 printf("Can't parse sample, err = %d, skipping...\n", err);
 259                                 continue;
 260                         }
 261
 262                         evsel = perf_evlist__id2evsel(evlist, sample.id);
 263                         if (evsel == NULL) {
 264                                 printf("Unknown tp ID %" PRIu64 ", skipping...\n", sample.id);
 265                                 continue;
 266                         }
 267
 268                         if (evlist->threads->map[0] == -1 || evlist->threads->nr > 1)
 269                                 printf("%d ", sample.tid);
 270
 271                         if (sample.raw_data == NULL) {
 272                                 printf("%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
 273                                        perf_evsel__name(evsel), sample.tid,
 274                                        sample.cpu, sample.raw_size);
 275                                 continue;
 276                         }
 277
 278                         handler = evsel->handler.func;
 279                         handler(trace, evsel, &sample);
 280                 }
 281         }
 282
 283         if (nr_events == before)
 284                 poll(evlist->pollfd, evlist->nr_fds, -1);
 285
 286         goto again;
 287
 288 out_delete_evlist:
 289         perf_evlist__delete(evlist);
 290 out:
 291         return err;
 292 }
 293
 294 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
 295 {
 296         const char * const trace_usage[] = {
 297                 "perf trace [<options>]",
 298                 NULL
 299         };
 300         struct trace trace = {
 301                 .audit_machine = audit_detect_machine(),
 302                 .syscalls = {
 303                         . max = -1,
 304                 },
 305                 .opts = {
 306                         .target = {
 307                                 .uid       = UINT_MAX,
 308                                 .uses_mmap = true,
 309                         },
 310                         .user_freq     = UINT_MAX,
 311                         .user_interval = ULLONG_MAX,
 312                         .no_delay      = true,
 313                         .mmap_pages    = 1024,
 314                 },
 315         };
 316         const struct option trace_options[] = {
 317         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
 318                     "trace events on existing process id"),
 319         OPT_STRING(0, "tid", &trace.opts.target.tid, "tid",
 320                     "trace events on existing thread id"),
 321         OPT_BOOLEAN(0, "all-cpus", &trace.opts.target.system_wide,
 322                     "system-wide collection from all CPUs"),
 323         OPT_STRING(0, "cpu", &trace.opts.target.cpu_list, "cpu",
 324                     "list of cpus to monitor"),
 325         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
 326                     "child tasks do not inherit counters"),
 327         OPT_UINTEGER(0, "mmap-pages", &trace.opts.mmap_pages,
 328                      "number of mmap data pages"),
 329         OPT_STRING(0, "uid", &trace.opts.target.uid_str, "user",
 330                    "user to profile"),
 331         OPT_END()
 332         };
 333         int err;
 334
 335         argc = parse_options(argc, argv, trace_options, trace_usage, 0);
 336         if (argc)
 337                 usage_with_options(trace_usage, trace_options);
 338
 339         err = perf_target__parse_uid(&trace.opts.target);
 340         if (err) {
 341                 char bf[BUFSIZ];
 342                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
 343                 printf("%s", bf);
 344                 return err;
 345         }
 346
 347         return trace__run(&trace);
 348 }