tools/perf/bench/inject-buildid.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <stdlib.h>
   3 #include <stddef.h>
   4 #include <ftw.h>
   5 #include <fcntl.h>
   6 #include <errno.h>
   7 #include <unistd.h>
   8 #include <pthread.h>
   9 #include <sys/mman.h>
  10 #include <sys/wait.h>
  11 #include <linux/kernel.h>
  12 #include <linux/time64.h>
  13 #include <linux/list.h>
  14 #include <linux/err.h>
  15 #include <internal/lib.h>
  16 #include <subcmd/parse-options.h>
  17
  18 #include "bench.h"
  19 #include "util/data.h"
  20 #include "util/stat.h"
  21 #include "util/debug.h"
  22 #include "util/event.h"
  23 #include "util/symbol.h"
  24 #include "util/session.h"
  25 #include "util/build-id.h"
  26 #include "util/synthetic-events.h"
  27
  28 #define MMAP_DEV_MAJOR  8
  29 #define DSO_MMAP_RATIO  4
  30
  31 static unsigned int iterations = 100;
  32 static unsigned int nr_mmaps   = 100;
  33 static unsigned int nr_samples = 100;  /* samples per mmap */
  34
  35 static u64 bench_sample_type;
  36 static u16 bench_id_hdr_size;
  37
  38 struct bench_data {
  39         int                     pid;
  40         int                     input_pipe[2];
  41         int                     output_pipe[2];
  42         pthread_t               th;
  43 };
  44
  45 struct bench_dso {
  46         struct list_head        list;
  47         char                    *name;
  48         int                     ino;
  49 };
  50
  51 static int nr_dsos;
  52 static struct bench_dso *dsos;
  53
  54 extern int cmd_inject(int argc, const char *argv[]);
  55
  56 static const struct option options[] = {
  57         OPT_UINTEGER('i', "iterations", &iterations,
  58                      "Number of iterations used to compute average (default: 100)"),
  59         OPT_UINTEGER('m', "nr-mmaps", &nr_mmaps,
  60                      "Number of mmap events for each iteration (default: 100)"),
  61         OPT_UINTEGER('n', "nr-samples", &nr_samples,
  62                      "Number of sample events per mmap event (default: 100)"),
  63         OPT_INCR('v', "verbose", &verbose,
  64                  "be more verbose (show iteration count, DSO name, etc)"),
  65         OPT_END()
  66 };
  67
  68 static const char *const bench_usage[] = {
  69         "perf bench internals inject-build-id <options>",
  70         NULL
  71 };
  72
  73 /*
  74  * Helper for collect_dso that adds the given file as a dso to dso_list
  75  * if it contains a build-id.  Stops after collecting 4 times more than
  76  * we need (for MMAP2 events).
  77  */
  78 static int add_dso(const char *fpath, const struct stat *sb __maybe_unused,
  79                    int typeflag, struct FTW *ftwbuf __maybe_unused)
  80 {
  81         struct bench_dso *dso = &dsos[nr_dsos];
  82         struct build_id bid;
  83
  84         if (typeflag == FTW_D || typeflag == FTW_SL)
  85                 return 0;
  86
  87         if (filename__read_build_id(fpath, &bid) < 0)
  88                 return 0;
  89
  90         dso->name = realpath(fpath, NULL);
  91         if (dso->name == NULL)
  92                 return -1;
  93
  94         dso->ino = nr_dsos++;
  95         pr_debug2("  Adding DSO: %s\n", fpath);
  96
  97         /* stop if we collected enough DSOs */
  98         if ((unsigned int)nr_dsos == DSO_MMAP_RATIO * nr_mmaps)
  99                 return 1;
 100
 101         return 0;
 102 }
 103
 104 static void collect_dso(void)
 105 {
 106         dsos = calloc(nr_mmaps * DSO_MMAP_RATIO, sizeof(*dsos));
 107         if (dsos == NULL) {
 108                 printf("  Memory allocation failed\n");
 109                 exit(1);
 110         }
 111
 112         if (nftw("/usr/lib/", add_dso, 10, FTW_PHYS) < 0)
 113                 return;
 114
 115         pr_debug("  Collected %d DSOs\n", nr_dsos);
 116 }
 117
 118 static void release_dso(void)
 119 {
 120         int i;
 121
 122         for (i = 0; i < nr_dsos; i++) {
 123                 struct bench_dso *dso = &dsos[i];
 124
 125                 free(dso->name);
 126         }
 127         free(dsos);
 128 }
 129
 130 /* Fake address used by mmap and sample events */
 131 static u64 dso_map_addr(struct bench_dso *dso)
 132 {
 133         return 0x400000ULL + dso->ino * 8192ULL;
 134 }
 135
 136 static u32 synthesize_attr(struct bench_data *data)
 137 {
 138         union perf_event event;
 139
 140         memset(&event, 0, sizeof(event.attr) + sizeof(u64));
 141
 142         event.header.type = PERF_RECORD_HEADER_ATTR;
 143         event.header.size = sizeof(event.attr) + sizeof(u64);
 144
 145         event.attr.attr.type = PERF_TYPE_SOFTWARE;
 146         event.attr.attr.config = PERF_COUNT_SW_TASK_CLOCK;
 147         event.attr.attr.exclude_kernel = 1;
 148         event.attr.attr.sample_id_all = 1;
 149         event.attr.attr.sample_type = bench_sample_type;
 150
 151         return writen(data->input_pipe[1], &event, event.header.size);
 152 }
 153
 154 static u32 synthesize_fork(struct bench_data *data)
 155 {
 156         union perf_event event;
 157
 158         memset(&event, 0, sizeof(event.fork) + bench_id_hdr_size);
 159
 160         event.header.type = PERF_RECORD_FORK;
 161         event.header.misc = PERF_RECORD_MISC_FORK_EXEC;
 162         event.header.size = sizeof(event.fork) + bench_id_hdr_size;
 163
 164         event.fork.ppid = 1;
 165         event.fork.ptid = 1;
 166         event.fork.pid = data->pid;
 167         event.fork.tid = data->pid;
 168
 169         return writen(data->input_pipe[1], &event, event.header.size);
 170 }
 171
 172 static u32 synthesize_mmap(struct bench_data *data, struct bench_dso *dso,
 173                            u64 timestamp)
 174 {
 175         union perf_event event;
 176         size_t len = offsetof(struct perf_record_mmap2, filename);
 177         u64 *id_hdr_ptr = (void *)&event;
 178         int ts_idx;
 179
 180         len += roundup(strlen(dso->name) + 1, 8) + bench_id_hdr_size;
 181
 182         memset(&event, 0, min(len, sizeof(event.mmap2)));
 183
 184         event.header.type = PERF_RECORD_MMAP2;
 185         event.header.misc = PERF_RECORD_MISC_USER;
 186         event.header.size = len;
 187
 188         event.mmap2.pid = data->pid;
 189         event.mmap2.tid = data->pid;
 190         event.mmap2.maj = MMAP_DEV_MAJOR;
 191         event.mmap2.ino = dso->ino;
 192
 193         strcpy(event.mmap2.filename, dso->name);
 194
 195         event.mmap2.start = dso_map_addr(dso);
 196         event.mmap2.len = 4096;
 197         event.mmap2.prot = PROT_EXEC;
 198
 199         if (len > sizeof(event.mmap2)) {
 200                 /* write mmap2 event first */
 201                 writen(data->input_pipe[1], &event, len - bench_id_hdr_size);
 202                 /* zero-fill sample id header */
 203                 memset(id_hdr_ptr, 0, bench_id_hdr_size);
 204                 /* put timestamp in the right position */
 205                 ts_idx = (bench_id_hdr_size / sizeof(u64)) - 2;
 206                 id_hdr_ptr[ts_idx] = timestamp;
 207                 writen(data->input_pipe[1], id_hdr_ptr, bench_id_hdr_size);
 208         } else {
 209                 ts_idx = (len / sizeof(u64)) - 2;
 210                 id_hdr_ptr[ts_idx] = timestamp;
 211                 writen(data->input_pipe[1], &event, len);
 212         }
 213         return len;
 214 }
 215
 216 static u32 synthesize_sample(struct bench_data *data, struct bench_dso *dso,
 217                              u64 timestamp)
 218 {
 219         union perf_event event;
 220         struct perf_sample sample = {
 221                 .tid = data->pid,
 222                 .pid = data->pid,
 223                 .ip = dso_map_addr(dso),
 224                 .time = timestamp,
 225         };
 226
 227         event.header.type = PERF_RECORD_SAMPLE;
 228         event.header.misc = PERF_RECORD_MISC_USER;
 229         event.header.size = perf_event__sample_event_size(&sample, bench_sample_type, 0);
 230
 231         perf_event__synthesize_sample(&event, bench_sample_type, 0, &sample);
 232
 233         return writen(data->input_pipe[1], &event, event.header.size);
 234 }
 235
 236 static u32 synthesize_flush(struct bench_data *data)
 237 {
 238         struct perf_event_header header = {
 239                 .size = sizeof(header),
 240                 .type = PERF_RECORD_FINISHED_ROUND,
 241         };
 242
 243         return writen(data->input_pipe[1], &header, header.size);
 244 }
 245
 246 static void *data_reader(void *arg)
 247 {
 248         struct bench_data *data = arg;
 249         char buf[8192];
 250         int flag;
 251         int n;
 252
 253         flag = fcntl(data->output_pipe[0], F_GETFL);
 254         fcntl(data->output_pipe[0], F_SETFL, flag | O_NONBLOCK);
 255
 256         /* read out data from child */
 257         while (true) {
 258                 n = read(data->output_pipe[0], buf, sizeof(buf));
 259                 if (n > 0)
 260                         continue;
 261                 if (n == 0)
 262                         break;
 263
 264                 if (errno != EINTR && errno != EAGAIN)
 265                         break;
 266
 267                 usleep(100);
 268         }
 269
 270         close(data->output_pipe[0]);
 271         return NULL;
 272 }
 273
 274 static int setup_injection(struct bench_data *data, bool build_id_all)
 275 {
 276         int ready_pipe[2];
 277         int dev_null_fd;
 278         char buf;
 279
 280         if (pipe(ready_pipe) < 0)
 281                 return -1;
 282
 283         if (pipe(data->input_pipe) < 0)
 284                 return -1;
 285
 286         if (pipe(data->output_pipe) < 0)
 287                 return -1;
 288
 289         data->pid = fork();
 290         if (data->pid < 0)
 291                 return -1;
 292
 293         if (data->pid == 0) {
 294                 const char **inject_argv;
 295                 int inject_argc = 2;
 296
 297                 close(data->input_pipe[1]);
 298                 close(data->output_pipe[0]);
 299                 close(ready_pipe[0]);
 300
 301                 dup2(data->input_pipe[0], STDIN_FILENO);
 302                 close(data->input_pipe[0]);
 303                 dup2(data->output_pipe[1], STDOUT_FILENO);
 304                 close(data->output_pipe[1]);
 305
 306                 dev_null_fd = open("/dev/null", O_WRONLY);
 307                 if (dev_null_fd < 0)
 308                         exit(1);
 309
 310                 dup2(dev_null_fd, STDERR_FILENO);
 311
 312                 if (build_id_all)
 313                         inject_argc++;
 314
 315                 inject_argv = calloc(inject_argc + 1, sizeof(*inject_argv));
 316                 if (inject_argv == NULL)
 317                         exit(1);
 318
 319                 inject_argv[0] = strdup("inject");
 320                 inject_argv[1] = strdup("-b");
 321                 if (build_id_all)
 322                         inject_argv[2] = strdup("--buildid-all");
 323
 324                 /* signal that we're ready to go */
 325                 close(ready_pipe[1]);
 326
 327                 cmd_inject(inject_argc, inject_argv);
 328
 329                 exit(0);
 330         }
 331
 332         pthread_create(&data->th, NULL, data_reader, data);
 333
 334         close(ready_pipe[1]);
 335         close(data->input_pipe[0]);
 336         close(data->output_pipe[1]);
 337
 338         /* wait for child ready */
 339         if (read(ready_pipe[0], &buf, 1) < 0)
 340                 return -1;
 341         close(ready_pipe[0]);
 342
 343         return 0;
 344 }
 345
 346 static int inject_build_id(struct bench_data *data, u64 *max_rss)
 347 {
 348         int status;
 349         unsigned int i, k;
 350         struct rusage rusage;
 351         u64 len = 0;
 352
 353         /* this makes the child to run */
 354         if (perf_header__write_pipe(data->input_pipe[1]) < 0)
 355                 return -1;
 356
 357         len += synthesize_attr(data);
 358         len += synthesize_fork(data);
 359
 360         for (i = 0; i < nr_mmaps; i++) {
 361                 int idx = rand() % (nr_dsos - 1);
 362                 struct bench_dso *dso = &dsos[idx];
 363                 u64 timestamp = rand() % 1000000;
 364
 365                 pr_debug2("   [%d] injecting: %s\n", i+1, dso->name);
 366                 len += synthesize_mmap(data, dso, timestamp);
 367
 368                 for (k = 0; k < nr_samples; k++)
 369                         len += synthesize_sample(data, dso, timestamp + k * 1000);
 370
 371                 if ((i + 1) % 10 == 0)
 372                         len += synthesize_flush(data);
 373         }
 374
 375         /* this makes the child to finish */
 376         close(data->input_pipe[1]);
 377
 378         wait4(data->pid, &status, 0, &rusage);
 379         *max_rss = rusage.ru_maxrss;
 380
 381         pr_debug("   Child %d exited with %d\n", data->pid, status);
 382
 383         return 0;
 384 }
 385
 386 static void do_inject_loop(struct bench_data *data, bool build_id_all)
 387 {
 388         unsigned int i;
 389         struct stats time_stats, mem_stats;
 390         double time_average, time_stddev;
 391         double mem_average, mem_stddev;
 392
 393         init_stats(&time_stats);
 394         init_stats(&mem_stats);
 395
 396         pr_debug("  Build-id%s injection benchmark\n", build_id_all ? "-all" : "");
 397
 398         for (i = 0; i < iterations; i++) {
 399                 struct timeval start, end, diff;
 400                 u64 runtime_us, max_rss;
 401
 402                 pr_debug("  Iteration #%d\n", i+1);
 403
 404                 if (setup_injection(data, build_id_all) < 0) {
 405                         printf("  Build-id injection setup failed\n");
 406                         break;
 407                 }
 408
 409                 gettimeofday(&start, NULL);
 410                 if (inject_build_id(data, &max_rss) < 0) {
 411                         printf("  Build-id injection failed\n");
 412                         break;
 413                 }
 414
 415                 gettimeofday(&end, NULL);
 416                 timersub(&end, &start, &diff);
 417                 runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
 418                 update_stats(&time_stats, runtime_us);
 419                 update_stats(&mem_stats, max_rss);
 420
 421                 pthread_join(data->th, NULL);
 422         }
 423
 424         time_average = avg_stats(&time_stats) / USEC_PER_MSEC;
 425         time_stddev = stddev_stats(&time_stats) / USEC_PER_MSEC;
 426         printf("  Average build-id%s injection took: %.3f msec (+- %.3f msec)\n",
 427                build_id_all ? "-all" : "", time_average, time_stddev);
 428
 429         /* each iteration, it processes MMAP2 + BUILD_ID + nr_samples * SAMPLE */
 430         time_average = avg_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
 431         time_stddev = stddev_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
 432         printf("  Average time per event: %.3f usec (+- %.3f usec)\n",
 433                 time_average, time_stddev);
 434
 435         mem_average = avg_stats(&mem_stats);
 436         mem_stddev = stddev_stats(&mem_stats);
 437         printf("  Average memory usage: %.0f KB (+- %.0f KB)\n",
 438                 mem_average, mem_stddev);
 439 }
 440
 441 static int do_inject_loops(struct bench_data *data)
 442 {
 443
 444         srand(time(NULL));
 445         symbol__init(NULL);
 446
 447         bench_sample_type  = PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_IP;
 448         bench_sample_type |= PERF_SAMPLE_TID | PERF_SAMPLE_TIME;
 449         bench_id_hdr_size  = 32;
 450
 451         collect_dso();
 452         if (nr_dsos == 0) {
 453                 printf("  Cannot collect DSOs for injection\n");
 454                 return -1;
 455         }
 456
 457         do_inject_loop(data, false);
 458         do_inject_loop(data, true);
 459
 460         release_dso();
 461         return 0;
 462 }
 463
 464 int bench_inject_build_id(int argc, const char **argv)
 465 {
 466         struct bench_data data;
 467
 468         argc = parse_options(argc, argv, options, bench_usage, 0);
 469         if (argc) {
 470                 usage_with_options(bench_usage, options);
 471                 exit(EXIT_FAILURE);
 472         }
 473
 474         return do_inject_loops(&data);
 475 }
 476