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
  17 #include <stdio.h>
  18 #include <stdlib.h>
  19 #include <string.h>
  20 #include <sys/time.h>
  21 #include <errno.h>
  22
  23 #define K 1024
  24
  25 static const char       *length_str     = "1MB";
  26 static const char       *routine        = "default";
  27 static int              iterations      = 1;
  28 static bool             use_cycle;
  29 static int              cycle_fd;
  30 static bool             only_prefault;
  31 static bool             no_prefault;
  32
  33 static const struct option options[] = {
  34         OPT_STRING('l', "length", &length_str, "1MB",
  35                     "Specify length of memory to copy. "
  36                     "Available units: B, KB, MB, GB and TB (upper and lower)"),
  37         OPT_STRING('r', "routine", &routine, "default",
  38                     "Specify routine to copy"),
  39         OPT_INTEGER('i', "iterations", &iterations,
  40                     "repeat memcpy() invocation this number of times"),
  41         OPT_BOOLEAN('c', "cycle", &use_cycle,
  42                     "Use cycles event instead of gettimeofday() for measuring"),
  43         OPT_BOOLEAN('o', "only-prefault", &only_prefault,
  44                     "Show only the result with page faults before memcpy()"),
  45         OPT_BOOLEAN('n', "no-prefault", &no_prefault,
  46                     "Show only the result without page faults before memcpy()"),
  47         OPT_END()
  48 };
  49
  50 typedef void *(*memcpy_t)(void *, const void *, size_t);
  51
  52 struct routine {
  53         const char *name;
  54         const char *desc;
  55         memcpy_t fn;
  56 };
  57
  58 struct routine routines[] = {
  59         { "default",
  60           "Default memcpy() provided by glibc",
  61           memcpy },
  62 #ifdef HAVE_ARCH_X86_64_SUPPORT
  63
  64 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
  65 #include "mem-memcpy-x86-64-asm-def.h"
  66 #undef MEMCPY_FN
  67
  68 #endif
  69
  70         { NULL,
  71           NULL,
  72           NULL   }
  73 };
  74
  75 static const char * const bench_mem_memcpy_usage[] = {
  76         "perf bench mem memcpy <options>",
  77         NULL
  78 };
  79
  80 static struct perf_event_attr cycle_attr = {
  81         .type           = PERF_TYPE_HARDWARE,
  82         .config         = PERF_COUNT_HW_CPU_CYCLES
  83 };
  84
  85 static void init_cycle(void)
  86 {
  87         cycle_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1,
  88                                        perf_event_open_cloexec_flag());
  89
  90         if (cycle_fd < 0 && errno == ENOSYS)
  91                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  92         else
  93                 BUG_ON(cycle_fd < 0);
  94 }
  95
  96 static u64 get_cycle(void)
  97 {
  98         int ret;
  99         u64 clk;
 100
 101         ret = read(cycle_fd, &clk, sizeof(u64));
 102         BUG_ON(ret != sizeof(u64));
 103
 104         return clk;
 105 }
 106
 107 static double timeval2double(struct timeval *ts)
 108 {
 109         return (double)ts->tv_sec +
 110                 (double)ts->tv_usec / (double)1000000;
 111 }
 112
 113 static void alloc_mem(void **dst, void **src, size_t length)
 114 {
 115         *dst = zalloc(length);
 116         if (!*dst)
 117                 die("memory allocation failed - maybe length is too large?\n");
 118
 119         *src = zalloc(length);
 120         if (!*src)
 121                 die("memory allocation failed - maybe length is too large?\n");
 122         /* Make sure to always replace the zero pages even if MMAP_THRESH is crossed */
 123         memset(*src, 0, length);
 124 }
 125
 126 static u64 do_memcpy_cycle(memcpy_t fn, size_t len, bool prefault)
 127 {
 128         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 129         void *src = NULL, *dst = NULL;
 130         int i;
 131
 132         alloc_mem(&src, &dst, len);
 133
 134         if (prefault)
 135                 fn(dst, src, len);
 136
 137         cycle_start = get_cycle();
 138         for (i = 0; i < iterations; ++i)
 139                 fn(dst, src, len);
 140         cycle_end = get_cycle();
 141
 142         free(src);
 143         free(dst);
 144         return cycle_end - cycle_start;
 145 }
 146
 147 static double do_memcpy_gettimeofday(memcpy_t fn, size_t len, bool prefault)
 148 {
 149         struct timeval tv_start, tv_end, tv_diff;
 150         void *src = NULL, *dst = NULL;
 151         int i;
 152
 153         alloc_mem(&src, &dst, len);
 154
 155         if (prefault)
 156                 fn(dst, src, len);
 157
 158         BUG_ON(gettimeofday(&tv_start, NULL));
 159         for (i = 0; i < iterations; ++i)
 160                 fn(dst, src, len);
 161         BUG_ON(gettimeofday(&tv_end, NULL));
 162
 163         timersub(&tv_end, &tv_start, &tv_diff);
 164
 165         free(src);
 166         free(dst);
 167         return (double)((double)len / timeval2double(&tv_diff));
 168 }
 169
 170 #define pf (no_prefault ? 0 : 1)
 171
 172 #define print_bps(x) do {                                       \
 173                 if (x < K)                                      \
 174                         printf(" %14lf B/Sec", x);              \
 175                 else if (x < K * K)                             \
 176                         printf(" %14lfd KB/Sec", x / K);        \
 177                 else if (x < K * K * K)                         \
 178                         printf(" %14lf MB/Sec", x / K / K);     \
 179                 else                                            \
 180                         printf(" %14lf GB/Sec", x / K / K / K); \
 181         } while (0)
 182
 183 int bench_mem_memcpy(int argc, const char **argv,
 184                      const char *prefix __maybe_unused)
 185 {
 186         int i;
 187         size_t len;
 188         double result_bps[2];
 189         u64 result_cycle[2];
 190
 191         argc = parse_options(argc, argv, options,
 192                              bench_mem_memcpy_usage, 0);
 193
 194         if (no_prefault && only_prefault) {
 195                 fprintf(stderr, "Invalid options: -o and -n are mutually exclusive\n");
 196                 return 1;
 197         }
 198
 199         if (use_cycle)
 200                 init_cycle();
 201
 202         len = (size_t)perf_atoll((char *)length_str);
 203
 204         result_cycle[0] = result_cycle[1] = 0ULL;
 205         result_bps[0] = result_bps[1] = 0.0;
 206
 207         if ((s64)len <= 0) {
 208                 fprintf(stderr, "Invalid length:%s\n", length_str);
 209                 return 1;
 210         }
 211
 212         /* same to without specifying either of prefault and no-prefault */
 213         if (only_prefault && no_prefault)
 214                 only_prefault = no_prefault = false;
 215
 216         for (i = 0; routines[i].name; i++) {
 217                 if (!strcmp(routines[i].name, routine))
 218                         break;
 219         }
 220         if (!routines[i].name) {
 221                 printf("Unknown routine:%s\n", routine);
 222                 printf("Available routines...\n");
 223                 for (i = 0; routines[i].name; i++) {
 224                         printf("\t%s ... %s\n",
 225                                routines[i].name, routines[i].desc);
 226                 }
 227                 return 1;
 228         }
 229
 230         if (bench_format == BENCH_FORMAT_DEFAULT)
 231                 printf("# Copying %s Bytes ...\n\n", length_str);
 232
 233         if (!only_prefault && !no_prefault) {
 234                 /* show both of results */
 235                 if (use_cycle) {
 236                         result_cycle[0] =
 237                                 do_memcpy_cycle(routines[i].fn, len, false);
 238                         result_cycle[1] =
 239                                 do_memcpy_cycle(routines[i].fn, len, true);
 240                 } else {
 241                         result_bps[0] =
 242                                 do_memcpy_gettimeofday(routines[i].fn,
 243                                                 len, false);
 244                         result_bps[1] =
 245                                 do_memcpy_gettimeofday(routines[i].fn,
 246                                                 len, true);
 247                 }
 248         } else {
 249                 if (use_cycle) {
 250                         result_cycle[pf] =
 251                                 do_memcpy_cycle(routines[i].fn,
 252                                                 len, only_prefault);
 253                 } else {
 254                         result_bps[pf] =
 255                                 do_memcpy_gettimeofday(routines[i].fn,
 256                                                 len, only_prefault);
 257                 }
 258         }
 259
 260         switch (bench_format) {
 261         case BENCH_FORMAT_DEFAULT:
 262                 if (!only_prefault && !no_prefault) {
 263                         if (use_cycle) {
 264                                 printf(" %14lf Cycle/Byte\n",
 265                                         (double)result_cycle[0]
 266                                         / (double)len);
 267                                 printf(" %14lf Cycle/Byte (with prefault)\n",
 268                                         (double)result_cycle[1]
 269                                         / (double)len);
 270                         } else {
 271                                 print_bps(result_bps[0]);
 272                                 printf("\n");
 273                                 print_bps(result_bps[1]);
 274                                 printf(" (with prefault)\n");
 275                         }
 276                 } else {
 277                         if (use_cycle) {
 278                                 printf(" %14lf Cycle/Byte",
 279                                         (double)result_cycle[pf]
 280                                         / (double)len);
 281                         } else
 282                                 print_bps(result_bps[pf]);
 283
 284                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 285                 }
 286                 break;
 287         case BENCH_FORMAT_SIMPLE:
 288                 if (!only_prefault && !no_prefault) {
 289                         if (use_cycle) {
 290                                 printf("%lf %lf\n",
 291                                         (double)result_cycle[0] / (double)len,
 292                                         (double)result_cycle[1] / (double)len);
 293                         } else {
 294                                 printf("%lf %lf\n",
 295                                         result_bps[0], result_bps[1]);
 296                         }
 297                 } else {
 298                         if (use_cycle) {
 299                                 printf("%lf\n", (double)result_cycle[pf]
 300                                         / (double)len);
 301                         } else
 302                                 printf("%lf\n", result_bps[pf]);
 303                 }
 304                 break;
 305         default:
 306                 /* reaching this means there's some disaster: */
 307                 die("unknown format: %d\n", bench_format);
 308                 break;
 309         }
 310
 311         return 0;
 312 }