| 1 | /* |
| 2 | * Clock functions |
| 3 | */ |
| 4 | |
| 5 | #include <unistd.h> |
| 6 | #include <math.h> |
| 7 | #include <sys/time.h> |
| 8 | #include <time.h> |
| 9 | |
| 10 | #include "fio.h" |
| 11 | #include "smalloc.h" |
| 12 | |
| 13 | #include "hash.h" |
| 14 | |
| 15 | #ifdef ARCH_HAVE_CPU_CLOCK |
| 16 | static unsigned long cycles_per_usec; |
| 17 | static unsigned long last_cycles; |
| 18 | int tsc_reliable = 0; |
| 19 | #endif |
| 20 | static struct timeval last_tv; |
| 21 | static int last_tv_valid; |
| 22 | |
| 23 | enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE; |
| 24 | int fio_clock_source_set = 0; |
| 25 | enum fio_cs fio_clock_source_inited = CS_INVAL; |
| 26 | |
| 27 | #ifdef FIO_DEBUG_TIME |
| 28 | |
| 29 | #define HASH_BITS 8 |
| 30 | #define HASH_SIZE (1 << HASH_BITS) |
| 31 | |
| 32 | static struct flist_head hash[HASH_SIZE]; |
| 33 | static int gtod_inited; |
| 34 | |
| 35 | struct gtod_log { |
| 36 | struct flist_head list; |
| 37 | void *caller; |
| 38 | unsigned long calls; |
| 39 | }; |
| 40 | |
| 41 | static struct gtod_log *find_hash(void *caller) |
| 42 | { |
| 43 | unsigned long h = hash_ptr(caller, HASH_BITS); |
| 44 | struct flist_head *entry; |
| 45 | |
| 46 | flist_for_each(entry, &hash[h]) { |
| 47 | struct gtod_log *log = flist_entry(entry, struct gtod_log, |
| 48 | list); |
| 49 | |
| 50 | if (log->caller == caller) |
| 51 | return log; |
| 52 | } |
| 53 | |
| 54 | return NULL; |
| 55 | } |
| 56 | |
| 57 | static struct gtod_log *find_log(void *caller) |
| 58 | { |
| 59 | struct gtod_log *log = find_hash(caller); |
| 60 | |
| 61 | if (!log) { |
| 62 | unsigned long h; |
| 63 | |
| 64 | log = malloc(sizeof(*log)); |
| 65 | INIT_FLIST_HEAD(&log->list); |
| 66 | log->caller = caller; |
| 67 | log->calls = 0; |
| 68 | |
| 69 | h = hash_ptr(caller, HASH_BITS); |
| 70 | flist_add_tail(&log->list, &hash[h]); |
| 71 | } |
| 72 | |
| 73 | return log; |
| 74 | } |
| 75 | |
| 76 | static void gtod_log_caller(void *caller) |
| 77 | { |
| 78 | if (gtod_inited) { |
| 79 | struct gtod_log *log = find_log(caller); |
| 80 | |
| 81 | log->calls++; |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | static void fio_exit fio_dump_gtod(void) |
| 86 | { |
| 87 | unsigned long total_calls = 0; |
| 88 | int i; |
| 89 | |
| 90 | for (i = 0; i < HASH_SIZE; i++) { |
| 91 | struct flist_head *entry; |
| 92 | struct gtod_log *log; |
| 93 | |
| 94 | flist_for_each(entry, &hash[i]) { |
| 95 | log = flist_entry(entry, struct gtod_log, list); |
| 96 | |
| 97 | printf("function %p, calls %lu\n", log->caller, |
| 98 | log->calls); |
| 99 | total_calls += log->calls; |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | printf("Total %lu gettimeofday\n", total_calls); |
| 104 | } |
| 105 | |
| 106 | static void fio_init gtod_init(void) |
| 107 | { |
| 108 | int i; |
| 109 | |
| 110 | for (i = 0; i < HASH_SIZE; i++) |
| 111 | INIT_FLIST_HEAD(&hash[i]); |
| 112 | |
| 113 | gtod_inited = 1; |
| 114 | } |
| 115 | |
| 116 | #endif /* FIO_DEBUG_TIME */ |
| 117 | |
| 118 | #ifdef FIO_DEBUG_TIME |
| 119 | void fio_gettime(struct timeval *tp, void *caller) |
| 120 | #else |
| 121 | void fio_gettime(struct timeval *tp, void fio_unused *caller) |
| 122 | #endif |
| 123 | { |
| 124 | #ifdef FIO_DEBUG_TIME |
| 125 | if (!caller) |
| 126 | caller = __builtin_return_address(0); |
| 127 | |
| 128 | gtod_log_caller(caller); |
| 129 | #endif |
| 130 | if (fio_tv) { |
| 131 | memcpy(tp, fio_tv, sizeof(*tp)); |
| 132 | return; |
| 133 | } |
| 134 | |
| 135 | switch (fio_clock_source) { |
| 136 | case CS_GTOD: |
| 137 | gettimeofday(tp, NULL); |
| 138 | break; |
| 139 | case CS_CGETTIME: { |
| 140 | struct timespec ts; |
| 141 | |
| 142 | #ifdef FIO_HAVE_CLOCK_MONOTONIC |
| 143 | if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) { |
| 144 | #else |
| 145 | if (clock_gettime(CLOCK_REALTIME, &ts) < 0) { |
| 146 | #endif |
| 147 | log_err("fio: clock_gettime fails\n"); |
| 148 | assert(0); |
| 149 | } |
| 150 | |
| 151 | tp->tv_sec = ts.tv_sec; |
| 152 | tp->tv_usec = ts.tv_nsec / 1000; |
| 153 | break; |
| 154 | } |
| 155 | #ifdef ARCH_HAVE_CPU_CLOCK |
| 156 | case CS_CPUCLOCK: { |
| 157 | unsigned long long usecs, t; |
| 158 | |
| 159 | t = get_cpu_clock(); |
| 160 | if (t < last_cycles) { |
| 161 | dprint(FD_TIME, "CPU clock going back in time\n"); |
| 162 | t = last_cycles; |
| 163 | } |
| 164 | |
| 165 | usecs = t / cycles_per_usec; |
| 166 | tp->tv_sec = usecs / 1000000; |
| 167 | tp->tv_usec = usecs % 1000000; |
| 168 | last_cycles = t; |
| 169 | break; |
| 170 | } |
| 171 | #endif |
| 172 | default: |
| 173 | log_err("fio: invalid clock source %d\n", fio_clock_source); |
| 174 | break; |
| 175 | } |
| 176 | |
| 177 | /* |
| 178 | * If Linux is using the tsc clock on non-synced processors, |
| 179 | * sometimes time can appear to drift backwards. Fix that up. |
| 180 | */ |
| 181 | if (last_tv_valid) { |
| 182 | if (tp->tv_sec < last_tv.tv_sec) |
| 183 | tp->tv_sec = last_tv.tv_sec; |
| 184 | else if (last_tv.tv_sec == tp->tv_sec && |
| 185 | tp->tv_usec < last_tv.tv_usec) |
| 186 | tp->tv_usec = last_tv.tv_usec; |
| 187 | } |
| 188 | last_tv_valid = 1; |
| 189 | memcpy(&last_tv, tp, sizeof(*tp)); |
| 190 | } |
| 191 | |
| 192 | #ifdef ARCH_HAVE_CPU_CLOCK |
| 193 | static unsigned long get_cycles_per_usec(void) |
| 194 | { |
| 195 | struct timeval s, e; |
| 196 | unsigned long long c_s, c_e; |
| 197 | |
| 198 | gettimeofday(&s, NULL); |
| 199 | c_s = get_cpu_clock(); |
| 200 | do { |
| 201 | unsigned long long elapsed; |
| 202 | |
| 203 | gettimeofday(&e, NULL); |
| 204 | elapsed = utime_since(&s, &e); |
| 205 | if (elapsed >= 1280) { |
| 206 | c_e = get_cpu_clock(); |
| 207 | break; |
| 208 | } |
| 209 | } while (1); |
| 210 | |
| 211 | return (c_e - c_s + 127) >> 7; |
| 212 | } |
| 213 | |
| 214 | #define NR_TIME_ITERS 50 |
| 215 | |
| 216 | static void calibrate_cpu_clock(void) |
| 217 | { |
| 218 | double delta, mean, S; |
| 219 | unsigned long avg, cycles[NR_TIME_ITERS]; |
| 220 | int i, samples; |
| 221 | |
| 222 | cycles[0] = get_cycles_per_usec(); |
| 223 | S = delta = mean = 0.0; |
| 224 | for (i = 0; i < NR_TIME_ITERS; i++) { |
| 225 | cycles[i] = get_cycles_per_usec(); |
| 226 | delta = cycles[i] - mean; |
| 227 | if (delta) { |
| 228 | mean += delta / (i + 1.0); |
| 229 | S += delta * (cycles[i] - mean); |
| 230 | } |
| 231 | } |
| 232 | |
| 233 | S = sqrt(S / (NR_TIME_ITERS - 1.0)); |
| 234 | |
| 235 | samples = avg = 0; |
| 236 | for (i = 0; i < NR_TIME_ITERS; i++) { |
| 237 | double this = cycles[i]; |
| 238 | |
| 239 | if ((fmax(this, mean) - fmin(this, mean)) > S) |
| 240 | continue; |
| 241 | samples++; |
| 242 | avg += this; |
| 243 | } |
| 244 | |
| 245 | S /= (double) NR_TIME_ITERS; |
| 246 | mean /= 10.0; |
| 247 | |
| 248 | for (i = 0; i < NR_TIME_ITERS; i++) |
| 249 | dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10); |
| 250 | |
| 251 | avg /= samples; |
| 252 | avg = (avg + 9) / 10; |
| 253 | dprint(FD_TIME, "avg: %lu\n", avg); |
| 254 | dprint(FD_TIME, "mean=%f, S=%f\n", mean, S); |
| 255 | |
| 256 | cycles_per_usec = avg; |
| 257 | } |
| 258 | #else |
| 259 | static void calibrate_cpu_clock(void) |
| 260 | { |
| 261 | } |
| 262 | #endif |
| 263 | |
| 264 | void fio_clock_init(void) |
| 265 | { |
| 266 | if (fio_clock_source == fio_clock_source_inited) |
| 267 | return; |
| 268 | |
| 269 | last_tv_valid = 0; |
| 270 | fio_clock_source_inited = fio_clock_source; |
| 271 | calibrate_cpu_clock(); |
| 272 | |
| 273 | /* |
| 274 | * If the arch sets tsc_reliable != 0, then it must be good enough |
| 275 | * to use as THE clock source. For x86 CPUs, this means the TSC |
| 276 | * runs at a constant rate and is synced across CPU cores. |
| 277 | */ |
| 278 | if (tsc_reliable) { |
| 279 | if (!fio_clock_source_set) |
| 280 | fio_clock_source = CS_CPUCLOCK; |
| 281 | } else if (fio_clock_source == CS_CPUCLOCK) |
| 282 | log_info("fio: clocksource=cpu may not be reliable\n"); |
| 283 | } |
| 284 | |
| 285 | unsigned long long utime_since(struct timeval *s, struct timeval *e) |
| 286 | { |
| 287 | long sec, usec; |
| 288 | unsigned long long ret; |
| 289 | |
| 290 | sec = e->tv_sec - s->tv_sec; |
| 291 | usec = e->tv_usec - s->tv_usec; |
| 292 | if (sec > 0 && usec < 0) { |
| 293 | sec--; |
| 294 | usec += 1000000; |
| 295 | } |
| 296 | |
| 297 | /* |
| 298 | * time warp bug on some kernels? |
| 299 | */ |
| 300 | if (sec < 0 || (sec == 0 && usec < 0)) |
| 301 | return 0; |
| 302 | |
| 303 | ret = sec * 1000000ULL + usec; |
| 304 | |
| 305 | return ret; |
| 306 | } |
| 307 | |
| 308 | unsigned long long utime_since_now(struct timeval *s) |
| 309 | { |
| 310 | struct timeval t; |
| 311 | |
| 312 | fio_gettime(&t, NULL); |
| 313 | return utime_since(s, &t); |
| 314 | } |
| 315 | |
| 316 | unsigned long mtime_since(struct timeval *s, struct timeval *e) |
| 317 | { |
| 318 | long sec, usec, ret; |
| 319 | |
| 320 | sec = e->tv_sec - s->tv_sec; |
| 321 | usec = e->tv_usec - s->tv_usec; |
| 322 | if (sec > 0 && usec < 0) { |
| 323 | sec--; |
| 324 | usec += 1000000; |
| 325 | } |
| 326 | |
| 327 | if (sec < 0 || (sec == 0 && usec < 0)) |
| 328 | return 0; |
| 329 | |
| 330 | sec *= 1000UL; |
| 331 | usec /= 1000UL; |
| 332 | ret = sec + usec; |
| 333 | |
| 334 | return ret; |
| 335 | } |
| 336 | |
| 337 | unsigned long mtime_since_now(struct timeval *s) |
| 338 | { |
| 339 | struct timeval t; |
| 340 | void *p = __builtin_return_address(0); |
| 341 | |
| 342 | fio_gettime(&t, p); |
| 343 | return mtime_since(s, &t); |
| 344 | } |
| 345 | |
| 346 | unsigned long time_since_now(struct timeval *s) |
| 347 | { |
| 348 | return mtime_since_now(s) / 1000; |
| 349 | } |