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