tools/perf/bench/mem-memcpy.c

   1 /*
   2  * mem-memcpy.c
   3  *
   4  * memcpy: Simple memory copy in various ways
   5  *
   6  * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
   7  */
   8
   9 #include "../perf.h"
  10 #include "../util/util.h"
  11 #include "../util/parse-options.h"
  12 #include "../util/header.h"
  13 #include "../util/cloexec.h"
  14 #include "bench.h"
  15 #include "mem-memcpy-arch.h"
  16 #include "mem-memset-arch.h"
  17
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <string.h>
  21 #include <sys/time.h>
  22 #include <errno.h>
  23
  24 #define K 1024
  25
  26 static const char       *length_str     = "1MB";
  27 static const char       *routine        = "default";
  28 static int              iterations      = 1;
  29 static bool             use_cycle;
  30 static int              cycle_fd;
  31 static bool             only_prefault;
  32 static bool             no_prefault;
  33
  34 static const struct option options[] = {
  35         OPT_STRING('l', "length", &length_str, "1MB",
  36                     "Specify length of memory to copy. "
  37                     "Available units: B, KB, MB, GB and TB (upper and lower)"),
  38         OPT_STRING('r', "routine", &routine, "default",
  39                     "Specify routine to copy"),
  40         OPT_INTEGER('i', "iterations", &iterations,
  41                     "repeat memcpy() invocation this number of times"),
  42         OPT_BOOLEAN('c', "cycle", &use_cycle,
  43                     "Use cycles event instead of gettimeofday() for measuring"),
  44         OPT_BOOLEAN('o', "only-prefault", &only_prefault,
  45                     "Show only the result with page faults before memcpy()"),
  46         OPT_BOOLEAN('n', "no-prefault", &no_prefault,
  47                     "Show only the result without page faults before memcpy()"),
  48         OPT_END()
  49 };
  50
  51 typedef void *(*memcpy_t)(void *, const void *, size_t);
  52 typedef void *(*memset_t)(void *, int, size_t);
  53
  54 struct routine {
  55         const char *name;
  56         const char *desc;
  57         union {
  58                 memcpy_t memcpy;
  59                 memset_t memset;
  60         } fn;
  61 };
  62
  63 struct routine memcpy_routines[] = {
  64         { .name = "default",
  65           .desc = "Default memcpy() provided by glibc",
  66           .fn.memcpy = memcpy },
  67 #ifdef HAVE_ARCH_X86_64_SUPPORT
  68
  69 #define MEMCPY_FN(_fn, _name, _desc) {.name = _name, .desc = _desc, .fn.memcpy = _fn},
  70 #include "mem-memcpy-x86-64-asm-def.h"
  71 #undef MEMCPY_FN
  72
  73 #endif
  74
  75         { NULL,
  76           NULL,
  77           {NULL}   }
  78 };
  79
  80 static const char * const bench_mem_memcpy_usage[] = {
  81         "perf bench mem memcpy <options>",
  82         NULL
  83 };
  84
  85 static struct perf_event_attr cycle_attr = {
  86         .type           = PERF_TYPE_HARDWARE,
  87         .config         = PERF_COUNT_HW_CPU_CYCLES
  88 };
  89
  90 static void init_cycle(void)
  91 {
  92         cycle_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1,
  93                                        perf_event_open_cloexec_flag());
  94
  95         if (cycle_fd < 0 && errno == ENOSYS)
  96                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  97         else
  98                 BUG_ON(cycle_fd < 0);
  99 }
 100
 101 static u64 get_cycle(void)
 102 {
 103         int ret;
 104         u64 clk;
 105
 106         ret = read(cycle_fd, &clk, sizeof(u64));
 107         BUG_ON(ret != sizeof(u64));
 108
 109         return clk;
 110 }
 111
 112 static double timeval2double(struct timeval *ts)
 113 {
 114         return (double)ts->tv_sec +
 115                 (double)ts->tv_usec / (double)1000000;
 116 }
 117
 118 #define pf (no_prefault ? 0 : 1)
 119
 120 #define print_bps(x) do {                                       \
 121                 if (x < K)                                      \
 122                         printf(" %14lf B/Sec", x);              \
 123                 else if (x < K * K)                             \
 124                         printf(" %14lfd KB/Sec", x / K);        \
 125                 else if (x < K * K * K)                         \
 126                         printf(" %14lf MB/Sec", x / K / K);     \
 127                 else                                            \
 128                         printf(" %14lf GB/Sec", x / K / K / K); \
 129         } while (0)
 130
 131 struct bench_mem_info {
 132         const struct routine *routines;
 133         u64 (*do_cycle)(const struct routine *r, size_t len, bool prefault);
 134         double (*do_gettimeofday)(const struct routine *r, size_t len, bool prefault);
 135         const char *const *usage;
 136 };
 137
 138 static int bench_mem_common(int argc, const char **argv,
 139                      const char *prefix __maybe_unused,
 140                      struct bench_mem_info *info)
 141 {
 142         int i;
 143         size_t len;
 144         double totallen;
 145         double result_bps[2];
 146         u64 result_cycle[2];
 147
 148         argc = parse_options(argc, argv, options,
 149                              info->usage, 0);
 150
 151         if (no_prefault && only_prefault) {
 152                 fprintf(stderr, "Invalid options: -o and -n are mutually exclusive\n");
 153                 return 1;
 154         }
 155
 156         if (use_cycle)
 157                 init_cycle();
 158
 159         len = (size_t)perf_atoll((char *)length_str);
 160         totallen = (double)len * iterations;
 161
 162         result_cycle[0] = result_cycle[1] = 0ULL;
 163         result_bps[0] = result_bps[1] = 0.0;
 164
 165         if ((s64)len <= 0) {
 166                 fprintf(stderr, "Invalid length:%s\n", length_str);
 167                 return 1;
 168         }
 169
 170         /* same to without specifying either of prefault and no-prefault */
 171         if (only_prefault && no_prefault)
 172                 only_prefault = no_prefault = false;
 173
 174         for (i = 0; info->routines[i].name; i++) {
 175                 if (!strcmp(info->routines[i].name, routine))
 176                         break;
 177         }
 178         if (!info->routines[i].name) {
 179                 printf("Unknown routine:%s\n", routine);
 180                 printf("Available routines...\n");
 181                 for (i = 0; info->routines[i].name; i++) {
 182                         printf("\t%s ... %s\n",
 183                                info->routines[i].name, info->routines[i].desc);
 184                 }
 185                 return 1;
 186         }
 187
 188         if (bench_format == BENCH_FORMAT_DEFAULT)
 189                 printf("# Copying %s Bytes ...\n\n", length_str);
 190
 191         if (!only_prefault && !no_prefault) {
 192                 /* show both of results */
 193                 if (use_cycle) {
 194                         result_cycle[0] =
 195                                 info->do_cycle(&info->routines[i], len, false);
 196                         result_cycle[1] =
 197                                 info->do_cycle(&info->routines[i], len, true);
 198                 } else {
 199                         result_bps[0] =
 200                                 info->do_gettimeofday(&info->routines[i],
 201                                                 len, false);
 202                         result_bps[1] =
 203                                 info->do_gettimeofday(&info->routines[i],
 204                                                 len, true);
 205                 }
 206         } else {
 207                 if (use_cycle) {
 208                         result_cycle[pf] =
 209                                 info->do_cycle(&info->routines[i],
 210                                                 len, only_prefault);
 211                 } else {
 212                         result_bps[pf] =
 213                                 info->do_gettimeofday(&info->routines[i],
 214                                                 len, only_prefault);
 215                 }
 216         }
 217
 218         switch (bench_format) {
 219         case BENCH_FORMAT_DEFAULT:
 220                 if (!only_prefault && !no_prefault) {
 221                         if (use_cycle) {
 222                                 printf(" %14lf Cycle/Byte\n",
 223                                         (double)result_cycle[0]
 224                                         / totallen);
 225                                 printf(" %14lf Cycle/Byte (with prefault)\n",
 226                                         (double)result_cycle[1]
 227                                         / totallen);
 228                         } else {
 229                                 print_bps(result_bps[0]);
 230                                 printf("\n");
 231                                 print_bps(result_bps[1]);
 232                                 printf(" (with prefault)\n");
 233                         }
 234                 } else {
 235                         if (use_cycle) {
 236                                 printf(" %14lf Cycle/Byte",
 237                                         (double)result_cycle[pf]
 238                                         / totallen);
 239                         } else
 240                                 print_bps(result_bps[pf]);
 241
 242                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 243                 }
 244                 break;
 245         case BENCH_FORMAT_SIMPLE:
 246                 if (!only_prefault && !no_prefault) {
 247                         if (use_cycle) {
 248                                 printf("%lf %lf\n",
 249                                         (double)result_cycle[0] / totallen,
 250                                         (double)result_cycle[1] / totallen);
 251                         } else {
 252                                 printf("%lf %lf\n",
 253                                         result_bps[0], result_bps[1]);
 254                         }
 255                 } else {
 256                         if (use_cycle) {
 257                                 printf("%lf\n", (double)result_cycle[pf]
 258                                         / totallen);
 259                         } else
 260                                 printf("%lf\n", result_bps[pf]);
 261                 }
 262                 break;
 263         default:
 264                 /* reaching this means there's some disaster: */
 265                 die("unknown format: %d\n", bench_format);
 266                 break;
 267         }
 268
 269         return 0;
 270 }
 271
 272 static void memcpy_alloc_mem(void **dst, void **src, size_t length)
 273 {
 274         *dst = zalloc(length);
 275         if (!*dst)
 276                 die("memory allocation failed - maybe length is too large?\n");
 277
 278         *src = zalloc(length);
 279         if (!*src)
 280                 die("memory allocation failed - maybe length is too large?\n");
 281         /* Make sure to always replace the zero pages even if MMAP_THRESH is crossed */
 282         memset(*src, 0, length);
 283 }
 284
 285 static u64 do_memcpy_cycle(const struct routine *r, size_t len, bool prefault)
 286 {
 287         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 288         void *src = NULL, *dst = NULL;
 289         memcpy_t fn = r->fn.memcpy;
 290         int i;
 291
 292         memcpy_alloc_mem(&dst, &src, len);
 293
 294         if (prefault)
 295                 fn(dst, src, len);
 296
 297         cycle_start = get_cycle();
 298         for (i = 0; i < iterations; ++i)
 299                 fn(dst, src, len);
 300         cycle_end = get_cycle();
 301
 302         free(src);
 303         free(dst);
 304         return cycle_end - cycle_start;
 305 }
 306
 307 static double do_memcpy_gettimeofday(const struct routine *r, size_t len,
 308                                      bool prefault)
 309 {
 310         struct timeval tv_start, tv_end, tv_diff;
 311         memcpy_t fn = r->fn.memcpy;
 312         void *src = NULL, *dst = NULL;
 313         int i;
 314
 315         memcpy_alloc_mem(&dst, &src, len);
 316
 317         if (prefault)
 318                 fn(dst, src, len);
 319
 320         BUG_ON(gettimeofday(&tv_start, NULL));
 321         for (i = 0; i < iterations; ++i)
 322                 fn(dst, src, len);
 323         BUG_ON(gettimeofday(&tv_end, NULL));
 324
 325         timersub(&tv_end, &tv_start, &tv_diff);
 326
 327         free(src);
 328         free(dst);
 329         return (double)(((double)len * iterations) / timeval2double(&tv_diff));
 330 }
 331
 332 int bench_mem_memcpy(int argc, const char **argv,
 333                      const char *prefix __maybe_unused)
 334 {
 335         struct bench_mem_info info = {
 336                 .routines = memcpy_routines,
 337                 .do_cycle = do_memcpy_cycle,
 338                 .do_gettimeofday = do_memcpy_gettimeofday,
 339                 .usage = bench_mem_memcpy_usage,
 340         };
 341
 342         return bench_mem_common(argc, argv, prefix, &info);
 343 }
 344
 345 static void memset_alloc_mem(void **dst, size_t length)
 346 {
 347         *dst = zalloc(length);
 348         if (!*dst)
 349                 die("memory allocation failed - maybe length is too large?\n");
 350 }
 351
 352 static u64 do_memset_cycle(const struct routine *r, size_t len, bool prefault)
 353 {
 354         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 355         memset_t fn = r->fn.memset;
 356         void *dst = NULL;
 357         int i;
 358
 359         memset_alloc_mem(&dst, len);
 360
 361         if (prefault)
 362                 fn(dst, -1, len);
 363
 364         cycle_start = get_cycle();
 365         for (i = 0; i < iterations; ++i)
 366                 fn(dst, i, len);
 367         cycle_end = get_cycle();
 368
 369         free(dst);
 370         return cycle_end - cycle_start;
 371 }
 372
 373 static double do_memset_gettimeofday(const struct routine *r, size_t len,
 374                                      bool prefault)
 375 {
 376         struct timeval tv_start, tv_end, tv_diff;
 377         memset_t fn = r->fn.memset;
 378         void *dst = NULL;
 379         int i;
 380
 381         memset_alloc_mem(&dst, len);
 382
 383         if (prefault)
 384                 fn(dst, -1, len);
 385
 386         BUG_ON(gettimeofday(&tv_start, NULL));
 387         for (i = 0; i < iterations; ++i)
 388                 fn(dst, i, len);
 389         BUG_ON(gettimeofday(&tv_end, NULL));
 390
 391         timersub(&tv_end, &tv_start, &tv_diff);
 392
 393         free(dst);
 394         return (double)(((double)len * iterations) / timeval2double(&tv_diff));
 395 }
 396
 397 static const char * const bench_mem_memset_usage[] = {
 398         "perf bench mem memset <options>",
 399         NULL
 400 };
 401
 402 static const struct routine memset_routines[] = {
 403         { .name ="default",
 404           .desc = "Default memset() provided by glibc",
 405           .fn.memset = memset },
 406 #ifdef HAVE_ARCH_X86_64_SUPPORT
 407
 408 #define MEMSET_FN(_fn, _name, _desc) { .name = _name, .desc = _desc, .fn.memset = _fn },
 409 #include "mem-memset-x86-64-asm-def.h"
 410 #undef MEMSET_FN
 411
 412 #endif
 413
 414         { .name = NULL,
 415           .desc = NULL,
 416           .fn.memset = NULL   }
 417 };
 418
 419 int bench_mem_memset(int argc, const char **argv,
 420                      const char *prefix __maybe_unused)
 421 {
 422         struct bench_mem_info info = {
 423                 .routines = memset_routines,
 424                 .do_cycle = do_memset_cycle,
 425                 .do_gettimeofday = do_memset_gettimeofday,
 426                 .usage = bench_mem_memset_usage,
 427         };
 428
 429         return bench_mem_common(argc, argv, prefix, &info);
 430 }