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 "bench.h"
  14 #include "mem-memcpy-arch.h"
  15
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <string.h>
  19 #include <sys/time.h>
  20 #include <errno.h>
  21
  22 #define K 1024
  23
  24 static const char       *length_str     = "1MB";
  25 static const char       *routine        = "default";
  26 static int              iterations      = 1;
  27 static bool             use_cycle;
  28 static int              cycle_fd;
  29 static bool             only_prefault;
  30 static bool             no_prefault;
  31
  32 static const struct option options[] = {
  33         OPT_STRING('l', "length", &length_str, "1MB",
  34                     "Specify length of memory to copy. "
  35                     "Available units: B, KB, MB, GB and TB (upper and lower)"),
  36         OPT_STRING('r', "routine", &routine, "default",
  37                     "Specify routine to copy"),
  38         OPT_INTEGER('i', "iterations", &iterations,
  39                     "repeat memcpy() invocation this number of times"),
  40         OPT_BOOLEAN('c', "cycle", &use_cycle,
  41                     "Use cycles event instead of gettimeofday() for measuring"),
  42         OPT_BOOLEAN('o', "only-prefault", &only_prefault,
  43                     "Show only the result with page faults before memcpy()"),
  44         OPT_BOOLEAN('n', "no-prefault", &no_prefault,
  45                     "Show only the result without page faults before memcpy()"),
  46         OPT_END()
  47 };
  48
  49 typedef void *(*memcpy_t)(void *, const void *, size_t);
  50
  51 struct routine {
  52         const char *name;
  53         const char *desc;
  54         memcpy_t fn;
  55 };
  56
  57 struct routine routines[] = {
  58         { "default",
  59           "Default memcpy() provided by glibc",
  60           memcpy },
  61 #ifdef HAVE_ARCH_X86_64_SUPPORT
  62
  63 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
  64 #include "mem-memcpy-x86-64-asm-def.h"
  65 #undef MEMCPY_FN
  66
  67 #endif
  68
  69         { NULL,
  70           NULL,
  71           NULL   }
  72 };
  73
  74 static const char * const bench_mem_memcpy_usage[] = {
  75         "perf bench mem memcpy <options>",
  76         NULL
  77 };
  78
  79 static struct perf_event_attr cycle_attr = {
  80         .type           = PERF_TYPE_HARDWARE,
  81         .config         = PERF_COUNT_HW_CPU_CYCLES
  82 };
  83
  84 static void init_cycle(void)
  85 {
  86         cycle_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, 0);
  87
  88         if (cycle_fd < 0 && errno == ENOSYS)
  89                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  90         else
  91                 BUG_ON(cycle_fd < 0);
  92 }
  93
  94 static u64 get_cycle(void)
  95 {
  96         int ret;
  97         u64 clk;
  98
  99         ret = read(cycle_fd, &clk, sizeof(u64));
 100         BUG_ON(ret != sizeof(u64));
 101
 102         return clk;
 103 }
 104
 105 static double timeval2double(struct timeval *ts)
 106 {
 107         return (double)ts->tv_sec +
 108                 (double)ts->tv_usec / (double)1000000;
 109 }
 110
 111 static void alloc_mem(void **dst, void **src, size_t length)
 112 {
 113         *dst = zalloc(length);
 114         if (!*dst)
 115                 die("memory allocation failed - maybe length is too large?\n");
 116
 117         *src = zalloc(length);
 118         if (!*src)
 119                 die("memory allocation failed - maybe length is too large?\n");
 120         /* Make sure to always replace the zero pages even if MMAP_THRESH is crossed */
 121         memset(*src, 0, length);
 122 }
 123
 124 static u64 do_memcpy_cycle(memcpy_t fn, size_t len, bool prefault)
 125 {
 126         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 127         void *src = NULL, *dst = NULL;
 128         int i;
 129
 130         alloc_mem(&src, &dst, len);
 131
 132         if (prefault)
 133                 fn(dst, src, len);
 134
 135         cycle_start = get_cycle();
 136         for (i = 0; i < iterations; ++i)
 137                 fn(dst, src, len);
 138         cycle_end = get_cycle();
 139
 140         free(src);
 141         free(dst);
 142         return cycle_end - cycle_start;
 143 }
 144
 145 static double do_memcpy_gettimeofday(memcpy_t fn, size_t len, bool prefault)
 146 {
 147         struct timeval tv_start, tv_end, tv_diff;
 148         void *src = NULL, *dst = NULL;
 149         int i;
 150
 151         alloc_mem(&src, &dst, len);
 152
 153         if (prefault)
 154                 fn(dst, src, len);
 155
 156         BUG_ON(gettimeofday(&tv_start, NULL));
 157         for (i = 0; i < iterations; ++i)
 158                 fn(dst, src, len);
 159         BUG_ON(gettimeofday(&tv_end, NULL));
 160
 161         timersub(&tv_end, &tv_start, &tv_diff);
 162
 163         free(src);
 164         free(dst);
 165         return (double)((double)len / timeval2double(&tv_diff));
 166 }
 167
 168 #define pf (no_prefault ? 0 : 1)
 169
 170 #define print_bps(x) do {                                       \
 171                 if (x < K)                                      \
 172                         printf(" %14lf B/Sec", x);              \
 173                 else if (x < K * K)                             \
 174                         printf(" %14lfd KB/Sec", x / K);        \
 175                 else if (x < K * K * K)                         \
 176                         printf(" %14lf MB/Sec", x / K / K);     \
 177                 else                                            \
 178                         printf(" %14lf GB/Sec", x / K / K / K); \
 179         } while (0)
 180
 181 int bench_mem_memcpy(int argc, const char **argv,
 182                      const char *prefix __maybe_unused)
 183 {
 184         int i;
 185         size_t len;
 186         double result_bps[2];
 187         u64 result_cycle[2];
 188
 189         argc = parse_options(argc, argv, options,
 190                              bench_mem_memcpy_usage, 0);
 191
 192         if (no_prefault && only_prefault) {
 193                 fprintf(stderr, "Invalid options: -o and -n are mutually exclusive\n");
 194                 return 1;
 195         }
 196
 197         if (use_cycle)
 198                 init_cycle();
 199
 200         len = (size_t)perf_atoll((char *)length_str);
 201
 202         result_cycle[0] = result_cycle[1] = 0ULL;
 203         result_bps[0] = result_bps[1] = 0.0;
 204
 205         if ((s64)len <= 0) {
 206                 fprintf(stderr, "Invalid length:%s\n", length_str);
 207                 return 1;
 208         }
 209
 210         /* same to without specifying either of prefault and no-prefault */
 211         if (only_prefault && no_prefault)
 212                 only_prefault = no_prefault = false;
 213
 214         for (i = 0; routines[i].name; i++) {
 215                 if (!strcmp(routines[i].name, routine))
 216                         break;
 217         }
 218         if (!routines[i].name) {
 219                 printf("Unknown routine:%s\n", routine);
 220                 printf("Available routines...\n");
 221                 for (i = 0; routines[i].name; i++) {
 222                         printf("\t%s ... %s\n",
 223                                routines[i].name, routines[i].desc);
 224                 }
 225                 return 1;
 226         }
 227
 228         if (bench_format == BENCH_FORMAT_DEFAULT)
 229                 printf("# Copying %s Bytes ...\n\n", length_str);
 230
 231         if (!only_prefault && !no_prefault) {
 232                 /* show both of results */
 233                 if (use_cycle) {
 234                         result_cycle[0] =
 235                                 do_memcpy_cycle(routines[i].fn, len, false);
 236                         result_cycle[1] =
 237                                 do_memcpy_cycle(routines[i].fn, len, true);
 238                 } else {
 239                         result_bps[0] =
 240                                 do_memcpy_gettimeofday(routines[i].fn,
 241                                                 len, false);
 242                         result_bps[1] =
 243                                 do_memcpy_gettimeofday(routines[i].fn,
 244                                                 len, true);
 245                 }
 246         } else {
 247                 if (use_cycle) {
 248                         result_cycle[pf] =
 249                                 do_memcpy_cycle(routines[i].fn,
 250                                                 len, only_prefault);
 251                 } else {
 252                         result_bps[pf] =
 253                                 do_memcpy_gettimeofday(routines[i].fn,
 254                                                 len, only_prefault);
 255                 }
 256         }
 257
 258         switch (bench_format) {
 259         case BENCH_FORMAT_DEFAULT:
 260                 if (!only_prefault && !no_prefault) {
 261                         if (use_cycle) {
 262                                 printf(" %14lf Cycle/Byte\n",
 263                                         (double)result_cycle[0]
 264                                         / (double)len);
 265                                 printf(" %14lf Cycle/Byte (with prefault)\n",
 266                                         (double)result_cycle[1]
 267                                         / (double)len);
 268                         } else {
 269                                 print_bps(result_bps[0]);
 270                                 printf("\n");
 271                                 print_bps(result_bps[1]);
 272                                 printf(" (with prefault)\n");
 273                         }
 274                 } else {
 275                         if (use_cycle) {
 276                                 printf(" %14lf Cycle/Byte",
 277                                         (double)result_cycle[pf]
 278                                         / (double)len);
 279                         } else
 280                                 print_bps(result_bps[pf]);
 281
 282                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 283                 }
 284                 break;
 285         case BENCH_FORMAT_SIMPLE:
 286                 if (!only_prefault && !no_prefault) {
 287                         if (use_cycle) {
 288                                 printf("%lf %lf\n",
 289                                         (double)result_cycle[0] / (double)len,
 290                                         (double)result_cycle[1] / (double)len);
 291                         } else {
 292                                 printf("%lf %lf\n",
 293                                         result_bps[0], result_bps[1]);
 294                         }
 295                 } else {
 296                         if (use_cycle) {
 297                                 printf("%lf\n", (double)result_cycle[pf]
 298                                         / (double)len);
 299                         } else
 300                                 printf("%lf\n", result_bps[pf]);
 301                 }
 302                 break;
 303         default:
 304                 /* reaching this means there's some disaster: */
 305                 die("unknown format: %d\n", bench_format);
 306                 break;
 307         }
 308
 309         return 0;
 310 }