| 1 | #include <stdio.h> |
| 2 | #include <string.h> |
| 3 | #include <sys/time.h> |
| 4 | #include <sys/types.h> |
| 5 | #include <sys/stat.h> |
| 6 | #include <dirent.h> |
| 7 | #include <libgen.h> |
| 8 | #include <math.h> |
| 9 | |
| 10 | #include "fio.h" |
| 11 | #include "diskutil.h" |
| 12 | #include "lib/ieee754.h" |
| 13 | #include "json.h" |
| 14 | #include "lib/getrusage.h" |
| 15 | #include "idletime.h" |
| 16 | #include "lib/pow2.h" |
| 17 | #include "lib/output_buffer.h" |
| 18 | #include "helper_thread.h" |
| 19 | |
| 20 | struct fio_mutex *stat_mutex; |
| 21 | |
| 22 | void clear_rusage_stat(struct thread_data *td) |
| 23 | { |
| 24 | struct thread_stat *ts = &td->ts; |
| 25 | |
| 26 | fio_getrusage(&td->ru_start); |
| 27 | ts->usr_time = ts->sys_time = 0; |
| 28 | ts->ctx = 0; |
| 29 | ts->minf = ts->majf = 0; |
| 30 | } |
| 31 | |
| 32 | void update_rusage_stat(struct thread_data *td) |
| 33 | { |
| 34 | struct thread_stat *ts = &td->ts; |
| 35 | |
| 36 | fio_getrusage(&td->ru_end); |
| 37 | ts->usr_time += mtime_since(&td->ru_start.ru_utime, |
| 38 | &td->ru_end.ru_utime); |
| 39 | ts->sys_time += mtime_since(&td->ru_start.ru_stime, |
| 40 | &td->ru_end.ru_stime); |
| 41 | ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw |
| 42 | - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); |
| 43 | ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt; |
| 44 | ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt; |
| 45 | |
| 46 | memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); |
| 47 | } |
| 48 | |
| 49 | /* |
| 50 | * Given a latency, return the index of the corresponding bucket in |
| 51 | * the structure tracking percentiles. |
| 52 | * |
| 53 | * (1) find the group (and error bits) that the value (latency) |
| 54 | * belongs to by looking at its MSB. (2) find the bucket number in the |
| 55 | * group by looking at the index bits. |
| 56 | * |
| 57 | */ |
| 58 | static unsigned int plat_val_to_idx(unsigned int val) |
| 59 | { |
| 60 | unsigned int msb, error_bits, base, offset, idx; |
| 61 | |
| 62 | /* Find MSB starting from bit 0 */ |
| 63 | if (val == 0) |
| 64 | msb = 0; |
| 65 | else |
| 66 | msb = (sizeof(val)*8) - __builtin_clz(val) - 1; |
| 67 | |
| 68 | /* |
| 69 | * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use |
| 70 | * all bits of the sample as index |
| 71 | */ |
| 72 | if (msb <= FIO_IO_U_PLAT_BITS) |
| 73 | return val; |
| 74 | |
| 75 | /* Compute the number of error bits to discard*/ |
| 76 | error_bits = msb - FIO_IO_U_PLAT_BITS; |
| 77 | |
| 78 | /* Compute the number of buckets before the group */ |
| 79 | base = (error_bits + 1) << FIO_IO_U_PLAT_BITS; |
| 80 | |
| 81 | /* |
| 82 | * Discard the error bits and apply the mask to find the |
| 83 | * index for the buckets in the group |
| 84 | */ |
| 85 | offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits); |
| 86 | |
| 87 | /* Make sure the index does not exceed (array size - 1) */ |
| 88 | idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ? |
| 89 | (base + offset) : (FIO_IO_U_PLAT_NR - 1); |
| 90 | |
| 91 | return idx; |
| 92 | } |
| 93 | |
| 94 | /* |
| 95 | * Convert the given index of the bucket array to the value |
| 96 | * represented by the bucket |
| 97 | */ |
| 98 | static unsigned int plat_idx_to_val(unsigned int idx) |
| 99 | { |
| 100 | unsigned int error_bits, k, base; |
| 101 | |
| 102 | assert(idx < FIO_IO_U_PLAT_NR); |
| 103 | |
| 104 | /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use |
| 105 | * all bits of the sample as index */ |
| 106 | if (idx < (FIO_IO_U_PLAT_VAL << 1)) |
| 107 | return idx; |
| 108 | |
| 109 | /* Find the group and compute the minimum value of that group */ |
| 110 | error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1; |
| 111 | base = 1 << (error_bits + FIO_IO_U_PLAT_BITS); |
| 112 | |
| 113 | /* Find its bucket number of the group */ |
| 114 | k = idx % FIO_IO_U_PLAT_VAL; |
| 115 | |
| 116 | /* Return the mean of the range of the bucket */ |
| 117 | return base + ((k + 0.5) * (1 << error_bits)); |
| 118 | } |
| 119 | |
| 120 | static int double_cmp(const void *a, const void *b) |
| 121 | { |
| 122 | const fio_fp64_t fa = *(const fio_fp64_t *) a; |
| 123 | const fio_fp64_t fb = *(const fio_fp64_t *) b; |
| 124 | int cmp = 0; |
| 125 | |
| 126 | if (fa.u.f > fb.u.f) |
| 127 | cmp = 1; |
| 128 | else if (fa.u.f < fb.u.f) |
| 129 | cmp = -1; |
| 130 | |
| 131 | return cmp; |
| 132 | } |
| 133 | |
| 134 | unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, |
| 135 | fio_fp64_t *plist, unsigned int **output, |
| 136 | unsigned int *maxv, unsigned int *minv) |
| 137 | { |
| 138 | unsigned long sum = 0; |
| 139 | unsigned int len, i, j = 0; |
| 140 | unsigned int oval_len = 0; |
| 141 | unsigned int *ovals = NULL; |
| 142 | int is_last; |
| 143 | |
| 144 | *minv = -1U; |
| 145 | *maxv = 0; |
| 146 | |
| 147 | len = 0; |
| 148 | while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) |
| 149 | len++; |
| 150 | |
| 151 | if (!len) |
| 152 | return 0; |
| 153 | |
| 154 | /* |
| 155 | * Sort the percentile list. Note that it may already be sorted if |
| 156 | * we are using the default values, but since it's a short list this |
| 157 | * isn't a worry. Also note that this does not work for NaN values. |
| 158 | */ |
| 159 | if (len > 1) |
| 160 | qsort((void *)plist, len, sizeof(plist[0]), double_cmp); |
| 161 | |
| 162 | /* |
| 163 | * Calculate bucket values, note down max and min values |
| 164 | */ |
| 165 | is_last = 0; |
| 166 | for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) { |
| 167 | sum += io_u_plat[i]; |
| 168 | while (sum >= (plist[j].u.f / 100.0 * nr)) { |
| 169 | assert(plist[j].u.f <= 100.0); |
| 170 | |
| 171 | if (j == oval_len) { |
| 172 | oval_len += 100; |
| 173 | ovals = realloc(ovals, oval_len * sizeof(unsigned int)); |
| 174 | } |
| 175 | |
| 176 | ovals[j] = plat_idx_to_val(i); |
| 177 | if (ovals[j] < *minv) |
| 178 | *minv = ovals[j]; |
| 179 | if (ovals[j] > *maxv) |
| 180 | *maxv = ovals[j]; |
| 181 | |
| 182 | is_last = (j == len - 1); |
| 183 | if (is_last) |
| 184 | break; |
| 185 | |
| 186 | j++; |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | *output = ovals; |
| 191 | return len; |
| 192 | } |
| 193 | |
| 194 | /* |
| 195 | * Find and display the p-th percentile of clat |
| 196 | */ |
| 197 | static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, |
| 198 | fio_fp64_t *plist, unsigned int precision, |
| 199 | struct buf_output *out) |
| 200 | { |
| 201 | unsigned int len, j = 0, minv, maxv; |
| 202 | unsigned int *ovals; |
| 203 | int is_last, per_line, scale_down; |
| 204 | char fmt[32]; |
| 205 | |
| 206 | len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv); |
| 207 | if (!len) |
| 208 | goto out; |
| 209 | |
| 210 | /* |
| 211 | * We default to usecs, but if the value range is such that we |
| 212 | * should scale down to msecs, do that. |
| 213 | */ |
| 214 | if (minv > 2000 && maxv > 99999) { |
| 215 | scale_down = 1; |
| 216 | log_buf(out, " clat percentiles (msec):\n |"); |
| 217 | } else { |
| 218 | scale_down = 0; |
| 219 | log_buf(out, " clat percentiles (usec):\n |"); |
| 220 | } |
| 221 | |
| 222 | snprintf(fmt, sizeof(fmt), "%%1.%uf", precision); |
| 223 | per_line = (80 - 7) / (precision + 14); |
| 224 | |
| 225 | for (j = 0; j < len; j++) { |
| 226 | char fbuf[16], *ptr = fbuf; |
| 227 | |
| 228 | /* for formatting */ |
| 229 | if (j != 0 && (j % per_line) == 0) |
| 230 | log_buf(out, " |"); |
| 231 | |
| 232 | /* end of the list */ |
| 233 | is_last = (j == len - 1); |
| 234 | |
| 235 | if (plist[j].u.f < 10.0) |
| 236 | ptr += sprintf(fbuf, " "); |
| 237 | |
| 238 | snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f); |
| 239 | |
| 240 | if (scale_down) |
| 241 | ovals[j] = (ovals[j] + 999) / 1000; |
| 242 | |
| 243 | log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ','); |
| 244 | |
| 245 | if (is_last) |
| 246 | break; |
| 247 | |
| 248 | if ((j % per_line) == per_line - 1) /* for formatting */ |
| 249 | log_buf(out, "\n"); |
| 250 | } |
| 251 | |
| 252 | out: |
| 253 | if (ovals) |
| 254 | free(ovals); |
| 255 | } |
| 256 | |
| 257 | int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, |
| 258 | double *mean, double *dev) |
| 259 | { |
| 260 | double n = (double) is->samples; |
| 261 | |
| 262 | if (n == 0) |
| 263 | return 0; |
| 264 | |
| 265 | *min = is->min_val; |
| 266 | *max = is->max_val; |
| 267 | *mean = is->mean.u.f; |
| 268 | |
| 269 | if (n > 1.0) |
| 270 | *dev = sqrt(is->S.u.f / (n - 1.0)); |
| 271 | else |
| 272 | *dev = 0; |
| 273 | |
| 274 | return 1; |
| 275 | } |
| 276 | |
| 277 | void show_group_stats(struct group_run_stats *rs, struct buf_output *out) |
| 278 | { |
| 279 | char *p1, *p2, *p3, *p4; |
| 280 | const char *str[] = { " READ", " WRITE" , " TRIM"}; |
| 281 | int i; |
| 282 | |
| 283 | log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid); |
| 284 | |
| 285 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 286 | const int i2p = is_power_of_2(rs->kb_base); |
| 287 | |
| 288 | if (!rs->max_run[i]) |
| 289 | continue; |
| 290 | |
| 291 | p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8); |
| 292 | p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base); |
| 293 | p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base); |
| 294 | p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base); |
| 295 | |
| 296 | log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s," |
| 297 | " mint=%llumsec, maxt=%llumsec\n", |
| 298 | rs->unified_rw_rep ? " MIXED" : str[i], |
| 299 | p1, p2, p3, p4, |
| 300 | (unsigned long long) rs->min_run[i], |
| 301 | (unsigned long long) rs->max_run[i]); |
| 302 | |
| 303 | free(p1); |
| 304 | free(p2); |
| 305 | free(p3); |
| 306 | free(p4); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist) |
| 311 | { |
| 312 | int i; |
| 313 | |
| 314 | /* |
| 315 | * Do depth distribution calculations |
| 316 | */ |
| 317 | for (i = 0; i < FIO_IO_U_MAP_NR; i++) { |
| 318 | if (total) { |
| 319 | io_u_dist[i] = (double) map[i] / (double) total; |
| 320 | io_u_dist[i] *= 100.0; |
| 321 | if (io_u_dist[i] < 0.1 && map[i]) |
| 322 | io_u_dist[i] = 0.1; |
| 323 | } else |
| 324 | io_u_dist[i] = 0.0; |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | static void stat_calc_lat(struct thread_stat *ts, double *dst, |
| 329 | unsigned int *src, int nr) |
| 330 | { |
| 331 | unsigned long total = ddir_rw_sum(ts->total_io_u); |
| 332 | int i; |
| 333 | |
| 334 | /* |
| 335 | * Do latency distribution calculations |
| 336 | */ |
| 337 | for (i = 0; i < nr; i++) { |
| 338 | if (total) { |
| 339 | dst[i] = (double) src[i] / (double) total; |
| 340 | dst[i] *= 100.0; |
| 341 | if (dst[i] < 0.01 && src[i]) |
| 342 | dst[i] = 0.01; |
| 343 | } else |
| 344 | dst[i] = 0.0; |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat) |
| 349 | { |
| 350 | stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR); |
| 351 | } |
| 352 | |
| 353 | void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat) |
| 354 | { |
| 355 | stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR); |
| 356 | } |
| 357 | |
| 358 | static void display_lat(const char *name, unsigned long min, unsigned long max, |
| 359 | double mean, double dev, struct buf_output *out) |
| 360 | { |
| 361 | const char *base = "(usec)"; |
| 362 | char *minp, *maxp; |
| 363 | |
| 364 | if (!usec_to_msec(&min, &max, &mean, &dev)) |
| 365 | base = "(msec)"; |
| 366 | |
| 367 | minp = num2str(min, 6, 1, 0, 0); |
| 368 | maxp = num2str(max, 6, 1, 0, 0); |
| 369 | |
| 370 | log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f," |
| 371 | " stdev=%5.02f\n", name, base, minp, maxp, mean, dev); |
| 372 | |
| 373 | free(minp); |
| 374 | free(maxp); |
| 375 | } |
| 376 | |
| 377 | static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts, |
| 378 | int ddir, struct buf_output *out) |
| 379 | { |
| 380 | const char *str[] = { "read ", "write", "trim" }; |
| 381 | unsigned long min, max, runt; |
| 382 | unsigned long long bw, iops; |
| 383 | double mean, dev; |
| 384 | char *io_p, *bw_p, *iops_p; |
| 385 | int i2p; |
| 386 | |
| 387 | assert(ddir_rw(ddir)); |
| 388 | |
| 389 | if (!ts->runtime[ddir]) |
| 390 | return; |
| 391 | |
| 392 | i2p = is_power_of_2(rs->kb_base); |
| 393 | runt = ts->runtime[ddir]; |
| 394 | |
| 395 | bw = (1000 * ts->io_bytes[ddir]) / runt; |
| 396 | io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8); |
| 397 | bw_p = num2str(bw, 6, 1, i2p, ts->unit_base); |
| 398 | |
| 399 | iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt; |
| 400 | iops_p = num2str(iops, 6, 1, 0, 0); |
| 401 | |
| 402 | log_buf(out, " %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n", |
| 403 | rs->unified_rw_rep ? "mixed" : str[ddir], |
| 404 | io_p, bw_p, iops_p, |
| 405 | (unsigned long long) ts->runtime[ddir]); |
| 406 | |
| 407 | free(io_p); |
| 408 | free(bw_p); |
| 409 | free(iops_p); |
| 410 | |
| 411 | if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) |
| 412 | display_lat("slat", min, max, mean, dev, out); |
| 413 | if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) |
| 414 | display_lat("clat", min, max, mean, dev, out); |
| 415 | if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) |
| 416 | display_lat(" lat", min, max, mean, dev, out); |
| 417 | |
| 418 | if (ts->clat_percentiles) { |
| 419 | show_clat_percentiles(ts->io_u_plat[ddir], |
| 420 | ts->clat_stat[ddir].samples, |
| 421 | ts->percentile_list, |
| 422 | ts->percentile_precision, out); |
| 423 | } |
| 424 | if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { |
| 425 | double p_of_agg = 100.0, fkb_base = (double)rs->kb_base; |
| 426 | const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB"); |
| 427 | |
| 428 | if (rs->unit_base == 1) { |
| 429 | min *= 8.0; |
| 430 | max *= 8.0; |
| 431 | mean *= 8.0; |
| 432 | dev *= 8.0; |
| 433 | } |
| 434 | |
| 435 | if (rs->agg[ddir]) { |
| 436 | p_of_agg = mean * 100 / (double) rs->agg[ddir]; |
| 437 | if (p_of_agg > 100.0) |
| 438 | p_of_agg = 100.0; |
| 439 | } |
| 440 | |
| 441 | if (mean > fkb_base * fkb_base) { |
| 442 | min /= fkb_base; |
| 443 | max /= fkb_base; |
| 444 | mean /= fkb_base; |
| 445 | dev /= fkb_base; |
| 446 | bw_str = (rs->unit_base == 1 ? "Mbit" : "MB"); |
| 447 | } |
| 448 | |
| 449 | log_buf(out, " bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%," |
| 450 | " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max, |
| 451 | p_of_agg, mean, dev); |
| 452 | } |
| 453 | } |
| 454 | |
| 455 | static int show_lat(double *io_u_lat, int nr, const char **ranges, |
| 456 | const char *msg, struct buf_output *out) |
| 457 | { |
| 458 | int new_line = 1, i, line = 0, shown = 0; |
| 459 | |
| 460 | for (i = 0; i < nr; i++) { |
| 461 | if (io_u_lat[i] <= 0.0) |
| 462 | continue; |
| 463 | shown = 1; |
| 464 | if (new_line) { |
| 465 | if (line) |
| 466 | log_buf(out, "\n"); |
| 467 | log_buf(out, " lat (%s) : ", msg); |
| 468 | new_line = 0; |
| 469 | line = 0; |
| 470 | } |
| 471 | if (line) |
| 472 | log_buf(out, ", "); |
| 473 | log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]); |
| 474 | line++; |
| 475 | if (line == 5) |
| 476 | new_line = 1; |
| 477 | } |
| 478 | |
| 479 | if (shown) |
| 480 | log_buf(out, "\n"); |
| 481 | |
| 482 | return shown; |
| 483 | } |
| 484 | |
| 485 | static void show_lat_u(double *io_u_lat_u, struct buf_output *out) |
| 486 | { |
| 487 | const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", |
| 488 | "250=", "500=", "750=", "1000=", }; |
| 489 | |
| 490 | show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out); |
| 491 | } |
| 492 | |
| 493 | static void show_lat_m(double *io_u_lat_m, struct buf_output *out) |
| 494 | { |
| 495 | const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", |
| 496 | "250=", "500=", "750=", "1000=", "2000=", |
| 497 | ">=2000=", }; |
| 498 | |
| 499 | show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out); |
| 500 | } |
| 501 | |
| 502 | static void show_latencies(struct thread_stat *ts, struct buf_output *out) |
| 503 | { |
| 504 | double io_u_lat_u[FIO_IO_U_LAT_U_NR]; |
| 505 | double io_u_lat_m[FIO_IO_U_LAT_M_NR]; |
| 506 | |
| 507 | stat_calc_lat_u(ts, io_u_lat_u); |
| 508 | stat_calc_lat_m(ts, io_u_lat_m); |
| 509 | |
| 510 | show_lat_u(io_u_lat_u, out); |
| 511 | show_lat_m(io_u_lat_m, out); |
| 512 | } |
| 513 | |
| 514 | static int block_state_category(int block_state) |
| 515 | { |
| 516 | switch (block_state) { |
| 517 | case BLOCK_STATE_UNINIT: |
| 518 | return 0; |
| 519 | case BLOCK_STATE_TRIMMED: |
| 520 | case BLOCK_STATE_WRITTEN: |
| 521 | return 1; |
| 522 | case BLOCK_STATE_WRITE_FAILURE: |
| 523 | case BLOCK_STATE_TRIM_FAILURE: |
| 524 | return 2; |
| 525 | default: |
| 526 | /* Silence compile warning on some BSDs and have a return */ |
| 527 | assert(0); |
| 528 | return -1; |
| 529 | } |
| 530 | } |
| 531 | |
| 532 | static int compare_block_infos(const void *bs1, const void *bs2) |
| 533 | { |
| 534 | uint32_t block1 = *(uint32_t *)bs1; |
| 535 | uint32_t block2 = *(uint32_t *)bs2; |
| 536 | int state1 = BLOCK_INFO_STATE(block1); |
| 537 | int state2 = BLOCK_INFO_STATE(block2); |
| 538 | int bscat1 = block_state_category(state1); |
| 539 | int bscat2 = block_state_category(state2); |
| 540 | int cycles1 = BLOCK_INFO_TRIMS(block1); |
| 541 | int cycles2 = BLOCK_INFO_TRIMS(block2); |
| 542 | |
| 543 | if (bscat1 < bscat2) |
| 544 | return -1; |
| 545 | if (bscat1 > bscat2) |
| 546 | return 1; |
| 547 | |
| 548 | if (cycles1 < cycles2) |
| 549 | return -1; |
| 550 | if (cycles1 > cycles2) |
| 551 | return 1; |
| 552 | |
| 553 | if (state1 < state2) |
| 554 | return -1; |
| 555 | if (state1 > state2) |
| 556 | return 1; |
| 557 | |
| 558 | assert(block1 == block2); |
| 559 | return 0; |
| 560 | } |
| 561 | |
| 562 | static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos, |
| 563 | fio_fp64_t *plist, unsigned int **percentiles, |
| 564 | unsigned int *types) |
| 565 | { |
| 566 | int len = 0; |
| 567 | int i, nr_uninit; |
| 568 | |
| 569 | qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos); |
| 570 | |
| 571 | while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) |
| 572 | len++; |
| 573 | |
| 574 | if (!len) |
| 575 | return 0; |
| 576 | |
| 577 | /* |
| 578 | * Sort the percentile list. Note that it may already be sorted if |
| 579 | * we are using the default values, but since it's a short list this |
| 580 | * isn't a worry. Also note that this does not work for NaN values. |
| 581 | */ |
| 582 | if (len > 1) |
| 583 | qsort((void *)plist, len, sizeof(plist[0]), double_cmp); |
| 584 | |
| 585 | nr_uninit = 0; |
| 586 | /* Start only after the uninit entries end */ |
| 587 | for (nr_uninit = 0; |
| 588 | nr_uninit < nr_block_infos |
| 589 | && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT; |
| 590 | nr_uninit ++) |
| 591 | ; |
| 592 | |
| 593 | if (nr_uninit == nr_block_infos) |
| 594 | return 0; |
| 595 | |
| 596 | *percentiles = calloc(len, sizeof(**percentiles)); |
| 597 | |
| 598 | for (i = 0; i < len; i++) { |
| 599 | int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100) |
| 600 | + nr_uninit; |
| 601 | (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]); |
| 602 | } |
| 603 | |
| 604 | memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT); |
| 605 | for (i = 0; i < nr_block_infos; i++) |
| 606 | types[BLOCK_INFO_STATE(block_infos[i])]++; |
| 607 | |
| 608 | return len; |
| 609 | } |
| 610 | |
| 611 | static const char *block_state_names[] = { |
| 612 | [BLOCK_STATE_UNINIT] = "unwritten", |
| 613 | [BLOCK_STATE_TRIMMED] = "trimmed", |
| 614 | [BLOCK_STATE_WRITTEN] = "written", |
| 615 | [BLOCK_STATE_TRIM_FAILURE] = "trim failure", |
| 616 | [BLOCK_STATE_WRITE_FAILURE] = "write failure", |
| 617 | }; |
| 618 | |
| 619 | static void show_block_infos(int nr_block_infos, uint32_t *block_infos, |
| 620 | fio_fp64_t *plist, struct buf_output *out) |
| 621 | { |
| 622 | int len, pos, i; |
| 623 | unsigned int *percentiles = NULL; |
| 624 | unsigned int block_state_counts[BLOCK_STATE_COUNT]; |
| 625 | |
| 626 | len = calc_block_percentiles(nr_block_infos, block_infos, plist, |
| 627 | &percentiles, block_state_counts); |
| 628 | |
| 629 | log_buf(out, " block lifetime percentiles :\n |"); |
| 630 | pos = 0; |
| 631 | for (i = 0; i < len; i++) { |
| 632 | uint32_t block_info = percentiles[i]; |
| 633 | #define LINE_LENGTH 75 |
| 634 | char str[LINE_LENGTH]; |
| 635 | int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c", |
| 636 | plist[i].u.f, block_info, |
| 637 | i == len - 1 ? '\n' : ','); |
| 638 | assert(strln < LINE_LENGTH); |
| 639 | if (pos + strln > LINE_LENGTH) { |
| 640 | pos = 0; |
| 641 | log_buf(out, "\n |"); |
| 642 | } |
| 643 | log_buf(out, "%s", str); |
| 644 | pos += strln; |
| 645 | #undef LINE_LENGTH |
| 646 | } |
| 647 | if (percentiles) |
| 648 | free(percentiles); |
| 649 | |
| 650 | log_buf(out, " states :"); |
| 651 | for (i = 0; i < BLOCK_STATE_COUNT; i++) |
| 652 | log_buf(out, " %s=%u%c", |
| 653 | block_state_names[i], block_state_counts[i], |
| 654 | i == BLOCK_STATE_COUNT - 1 ? '\n' : ','); |
| 655 | } |
| 656 | |
| 657 | static void show_thread_status_normal(struct thread_stat *ts, |
| 658 | struct group_run_stats *rs, |
| 659 | struct buf_output *out) |
| 660 | { |
| 661 | double usr_cpu, sys_cpu; |
| 662 | unsigned long runtime; |
| 663 | double io_u_dist[FIO_IO_U_MAP_NR]; |
| 664 | time_t time_p; |
| 665 | char time_buf[32]; |
| 666 | |
| 667 | if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u)) |
| 668 | return; |
| 669 | |
| 670 | time(&time_p); |
| 671 | os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf)); |
| 672 | |
| 673 | if (!ts->error) { |
| 674 | log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s", |
| 675 | ts->name, ts->groupid, ts->members, |
| 676 | ts->error, (int) ts->pid, time_buf); |
| 677 | } else { |
| 678 | log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s", |
| 679 | ts->name, ts->groupid, ts->members, |
| 680 | ts->error, ts->verror, (int) ts->pid, |
| 681 | time_buf); |
| 682 | } |
| 683 | |
| 684 | if (strlen(ts->description)) |
| 685 | log_buf(out, " Description : [%s]\n", ts->description); |
| 686 | |
| 687 | if (ts->io_bytes[DDIR_READ]) |
| 688 | show_ddir_status(rs, ts, DDIR_READ, out); |
| 689 | if (ts->io_bytes[DDIR_WRITE]) |
| 690 | show_ddir_status(rs, ts, DDIR_WRITE, out); |
| 691 | if (ts->io_bytes[DDIR_TRIM]) |
| 692 | show_ddir_status(rs, ts, DDIR_TRIM, out); |
| 693 | |
| 694 | show_latencies(ts, out); |
| 695 | |
| 696 | runtime = ts->total_run_time; |
| 697 | if (runtime) { |
| 698 | double runt = (double) runtime; |
| 699 | |
| 700 | usr_cpu = (double) ts->usr_time * 100 / runt; |
| 701 | sys_cpu = (double) ts->sys_time * 100 / runt; |
| 702 | } else { |
| 703 | usr_cpu = 0; |
| 704 | sys_cpu = 0; |
| 705 | } |
| 706 | |
| 707 | log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu," |
| 708 | " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu, |
| 709 | (unsigned long long) ts->ctx, |
| 710 | (unsigned long long) ts->majf, |
| 711 | (unsigned long long) ts->minf); |
| 712 | |
| 713 | stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); |
| 714 | log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," |
| 715 | " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], |
| 716 | io_u_dist[1], io_u_dist[2], |
| 717 | io_u_dist[3], io_u_dist[4], |
| 718 | io_u_dist[5], io_u_dist[6]); |
| 719 | |
| 720 | stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); |
| 721 | log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," |
| 722 | " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], |
| 723 | io_u_dist[1], io_u_dist[2], |
| 724 | io_u_dist[3], io_u_dist[4], |
| 725 | io_u_dist[5], io_u_dist[6]); |
| 726 | stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); |
| 727 | log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," |
| 728 | " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], |
| 729 | io_u_dist[1], io_u_dist[2], |
| 730 | io_u_dist[3], io_u_dist[4], |
| 731 | io_u_dist[5], io_u_dist[6]); |
| 732 | log_buf(out, " issued : total=r=%llu/w=%llu/d=%llu," |
| 733 | " short=r=%llu/w=%llu/d=%llu," |
| 734 | " drop=r=%llu/w=%llu/d=%llu\n", |
| 735 | (unsigned long long) ts->total_io_u[0], |
| 736 | (unsigned long long) ts->total_io_u[1], |
| 737 | (unsigned long long) ts->total_io_u[2], |
| 738 | (unsigned long long) ts->short_io_u[0], |
| 739 | (unsigned long long) ts->short_io_u[1], |
| 740 | (unsigned long long) ts->short_io_u[2], |
| 741 | (unsigned long long) ts->drop_io_u[0], |
| 742 | (unsigned long long) ts->drop_io_u[1], |
| 743 | (unsigned long long) ts->drop_io_u[2]); |
| 744 | if (ts->continue_on_error) { |
| 745 | log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n", |
| 746 | (unsigned long long)ts->total_err_count, |
| 747 | ts->first_error, |
| 748 | strerror(ts->first_error)); |
| 749 | } |
| 750 | if (ts->latency_depth) { |
| 751 | log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n", |
| 752 | (unsigned long long)ts->latency_target, |
| 753 | (unsigned long long)ts->latency_window, |
| 754 | ts->latency_percentile.u.f, |
| 755 | ts->latency_depth); |
| 756 | } |
| 757 | |
| 758 | if (ts->nr_block_infos) |
| 759 | show_block_infos(ts->nr_block_infos, ts->block_infos, |
| 760 | ts->percentile_list, out); |
| 761 | } |
| 762 | |
| 763 | static void show_ddir_status_terse(struct thread_stat *ts, |
| 764 | struct group_run_stats *rs, int ddir, |
| 765 | struct buf_output *out) |
| 766 | { |
| 767 | unsigned long min, max; |
| 768 | unsigned long long bw, iops; |
| 769 | unsigned int *ovals = NULL; |
| 770 | double mean, dev; |
| 771 | unsigned int len, minv, maxv; |
| 772 | int i; |
| 773 | |
| 774 | assert(ddir_rw(ddir)); |
| 775 | |
| 776 | iops = bw = 0; |
| 777 | if (ts->runtime[ddir]) { |
| 778 | uint64_t runt = ts->runtime[ddir]; |
| 779 | |
| 780 | bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; |
| 781 | iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; |
| 782 | } |
| 783 | |
| 784 | log_buf(out, ";%llu;%llu;%llu;%llu", |
| 785 | (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops, |
| 786 | (unsigned long long) ts->runtime[ddir]); |
| 787 | |
| 788 | if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) |
| 789 | log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev); |
| 790 | else |
| 791 | log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); |
| 792 | |
| 793 | if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) |
| 794 | log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev); |
| 795 | else |
| 796 | log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); |
| 797 | |
| 798 | if (ts->clat_percentiles) { |
| 799 | len = calc_clat_percentiles(ts->io_u_plat[ddir], |
| 800 | ts->clat_stat[ddir].samples, |
| 801 | ts->percentile_list, &ovals, &maxv, |
| 802 | &minv); |
| 803 | } else |
| 804 | len = 0; |
| 805 | |
| 806 | for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { |
| 807 | if (i >= len) { |
| 808 | log_buf(out, ";0%%=0"); |
| 809 | continue; |
| 810 | } |
| 811 | log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]); |
| 812 | } |
| 813 | |
| 814 | if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) |
| 815 | log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev); |
| 816 | else |
| 817 | log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); |
| 818 | |
| 819 | if (ovals) |
| 820 | free(ovals); |
| 821 | |
| 822 | if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { |
| 823 | double p_of_agg = 100.0; |
| 824 | |
| 825 | if (rs->agg[ddir]) { |
| 826 | p_of_agg = mean * 100 / (double) rs->agg[ddir]; |
| 827 | if (p_of_agg > 100.0) |
| 828 | p_of_agg = 100.0; |
| 829 | } |
| 830 | |
| 831 | log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); |
| 832 | } else |
| 833 | log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); |
| 834 | } |
| 835 | |
| 836 | static void add_ddir_status_json(struct thread_stat *ts, |
| 837 | struct group_run_stats *rs, int ddir, struct json_object *parent) |
| 838 | { |
| 839 | unsigned long min, max; |
| 840 | unsigned long long bw; |
| 841 | unsigned int *ovals = NULL; |
| 842 | double mean, dev, iops; |
| 843 | unsigned int len, minv, maxv; |
| 844 | int i; |
| 845 | const char *ddirname[] = {"read", "write", "trim"}; |
| 846 | struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object; |
| 847 | char buf[120]; |
| 848 | double p_of_agg = 100.0; |
| 849 | |
| 850 | assert(ddir_rw(ddir)); |
| 851 | |
| 852 | if (ts->unified_rw_rep && ddir != DDIR_READ) |
| 853 | return; |
| 854 | |
| 855 | dir_object = json_create_object(); |
| 856 | json_object_add_value_object(parent, |
| 857 | ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object); |
| 858 | |
| 859 | bw = 0; |
| 860 | iops = 0.0; |
| 861 | if (ts->runtime[ddir]) { |
| 862 | uint64_t runt = ts->runtime[ddir]; |
| 863 | |
| 864 | bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; |
| 865 | iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt; |
| 866 | } |
| 867 | |
| 868 | json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10); |
| 869 | json_object_add_value_int(dir_object, "bw", bw); |
| 870 | json_object_add_value_float(dir_object, "iops", iops); |
| 871 | json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]); |
| 872 | json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]); |
| 873 | json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]); |
| 874 | json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]); |
| 875 | |
| 876 | if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { |
| 877 | min = max = 0; |
| 878 | mean = dev = 0.0; |
| 879 | } |
| 880 | tmp_object = json_create_object(); |
| 881 | json_object_add_value_object(dir_object, "slat", tmp_object); |
| 882 | json_object_add_value_int(tmp_object, "min", min); |
| 883 | json_object_add_value_int(tmp_object, "max", max); |
| 884 | json_object_add_value_float(tmp_object, "mean", mean); |
| 885 | json_object_add_value_float(tmp_object, "stddev", dev); |
| 886 | |
| 887 | if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { |
| 888 | min = max = 0; |
| 889 | mean = dev = 0.0; |
| 890 | } |
| 891 | tmp_object = json_create_object(); |
| 892 | json_object_add_value_object(dir_object, "clat", tmp_object); |
| 893 | json_object_add_value_int(tmp_object, "min", min); |
| 894 | json_object_add_value_int(tmp_object, "max", max); |
| 895 | json_object_add_value_float(tmp_object, "mean", mean); |
| 896 | json_object_add_value_float(tmp_object, "stddev", dev); |
| 897 | |
| 898 | if (ts->clat_percentiles) { |
| 899 | len = calc_clat_percentiles(ts->io_u_plat[ddir], |
| 900 | ts->clat_stat[ddir].samples, |
| 901 | ts->percentile_list, &ovals, &maxv, |
| 902 | &minv); |
| 903 | } else |
| 904 | len = 0; |
| 905 | |
| 906 | percentile_object = json_create_object(); |
| 907 | json_object_add_value_object(tmp_object, "percentile", percentile_object); |
| 908 | for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { |
| 909 | if (i >= len) { |
| 910 | json_object_add_value_int(percentile_object, "0.00", 0); |
| 911 | continue; |
| 912 | } |
| 913 | snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f); |
| 914 | json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]); |
| 915 | } |
| 916 | |
| 917 | if (output_format & FIO_OUTPUT_JSON_PLUS) { |
| 918 | clat_bins_object = json_create_object(); |
| 919 | json_object_add_value_object(tmp_object, "bins", clat_bins_object); |
| 920 | for(i = 0; i < FIO_IO_U_PLAT_NR; i++) { |
| 921 | snprintf(buf, sizeof(buf), "%d", i); |
| 922 | json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]); |
| 923 | } |
| 924 | json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS); |
| 925 | json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL); |
| 926 | json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR); |
| 927 | } |
| 928 | |
| 929 | if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { |
| 930 | min = max = 0; |
| 931 | mean = dev = 0.0; |
| 932 | } |
| 933 | tmp_object = json_create_object(); |
| 934 | json_object_add_value_object(dir_object, "lat", tmp_object); |
| 935 | json_object_add_value_int(tmp_object, "min", min); |
| 936 | json_object_add_value_int(tmp_object, "max", max); |
| 937 | json_object_add_value_float(tmp_object, "mean", mean); |
| 938 | json_object_add_value_float(tmp_object, "stddev", dev); |
| 939 | if (ovals) |
| 940 | free(ovals); |
| 941 | |
| 942 | if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { |
| 943 | if (rs->agg[ddir]) { |
| 944 | p_of_agg = mean * 100 / (double) rs->agg[ddir]; |
| 945 | if (p_of_agg > 100.0) |
| 946 | p_of_agg = 100.0; |
| 947 | } |
| 948 | } else { |
| 949 | min = max = 0; |
| 950 | p_of_agg = mean = dev = 0.0; |
| 951 | } |
| 952 | json_object_add_value_int(dir_object, "bw_min", min); |
| 953 | json_object_add_value_int(dir_object, "bw_max", max); |
| 954 | json_object_add_value_float(dir_object, "bw_agg", p_of_agg); |
| 955 | json_object_add_value_float(dir_object, "bw_mean", mean); |
| 956 | json_object_add_value_float(dir_object, "bw_dev", dev); |
| 957 | } |
| 958 | |
| 959 | static void show_thread_status_terse_v2(struct thread_stat *ts, |
| 960 | struct group_run_stats *rs, |
| 961 | struct buf_output *out) |
| 962 | { |
| 963 | double io_u_dist[FIO_IO_U_MAP_NR]; |
| 964 | double io_u_lat_u[FIO_IO_U_LAT_U_NR]; |
| 965 | double io_u_lat_m[FIO_IO_U_LAT_M_NR]; |
| 966 | double usr_cpu, sys_cpu; |
| 967 | int i; |
| 968 | |
| 969 | /* General Info */ |
| 970 | log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error); |
| 971 | /* Log Read Status */ |
| 972 | show_ddir_status_terse(ts, rs, DDIR_READ, out); |
| 973 | /* Log Write Status */ |
| 974 | show_ddir_status_terse(ts, rs, DDIR_WRITE, out); |
| 975 | /* Log Trim Status */ |
| 976 | show_ddir_status_terse(ts, rs, DDIR_TRIM, out); |
| 977 | |
| 978 | /* CPU Usage */ |
| 979 | if (ts->total_run_time) { |
| 980 | double runt = (double) ts->total_run_time; |
| 981 | |
| 982 | usr_cpu = (double) ts->usr_time * 100 / runt; |
| 983 | sys_cpu = (double) ts->sys_time * 100 / runt; |
| 984 | } else { |
| 985 | usr_cpu = 0; |
| 986 | sys_cpu = 0; |
| 987 | } |
| 988 | |
| 989 | log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, |
| 990 | (unsigned long long) ts->ctx, |
| 991 | (unsigned long long) ts->majf, |
| 992 | (unsigned long long) ts->minf); |
| 993 | |
| 994 | /* Calc % distribution of IO depths, usecond, msecond latency */ |
| 995 | stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); |
| 996 | stat_calc_lat_u(ts, io_u_lat_u); |
| 997 | stat_calc_lat_m(ts, io_u_lat_m); |
| 998 | |
| 999 | /* Only show fixed 7 I/O depth levels*/ |
| 1000 | log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", |
| 1001 | io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], |
| 1002 | io_u_dist[4], io_u_dist[5], io_u_dist[6]); |
| 1003 | |
| 1004 | /* Microsecond latency */ |
| 1005 | for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) |
| 1006 | log_buf(out, ";%3.2f%%", io_u_lat_u[i]); |
| 1007 | /* Millisecond latency */ |
| 1008 | for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) |
| 1009 | log_buf(out, ";%3.2f%%", io_u_lat_m[i]); |
| 1010 | /* Additional output if continue_on_error set - default off*/ |
| 1011 | if (ts->continue_on_error) |
| 1012 | log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); |
| 1013 | log_buf(out, "\n"); |
| 1014 | |
| 1015 | /* Additional output if description is set */ |
| 1016 | if (strlen(ts->description)) |
| 1017 | log_buf(out, ";%s", ts->description); |
| 1018 | |
| 1019 | log_buf(out, "\n"); |
| 1020 | } |
| 1021 | |
| 1022 | static void show_thread_status_terse_v3_v4(struct thread_stat *ts, |
| 1023 | struct group_run_stats *rs, int ver, |
| 1024 | struct buf_output *out) |
| 1025 | { |
| 1026 | double io_u_dist[FIO_IO_U_MAP_NR]; |
| 1027 | double io_u_lat_u[FIO_IO_U_LAT_U_NR]; |
| 1028 | double io_u_lat_m[FIO_IO_U_LAT_M_NR]; |
| 1029 | double usr_cpu, sys_cpu; |
| 1030 | int i; |
| 1031 | |
| 1032 | /* General Info */ |
| 1033 | log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string, |
| 1034 | ts->name, ts->groupid, ts->error); |
| 1035 | /* Log Read Status */ |
| 1036 | show_ddir_status_terse(ts, rs, DDIR_READ, out); |
| 1037 | /* Log Write Status */ |
| 1038 | show_ddir_status_terse(ts, rs, DDIR_WRITE, out); |
| 1039 | /* Log Trim Status */ |
| 1040 | if (ver == 4) |
| 1041 | show_ddir_status_terse(ts, rs, DDIR_TRIM, out); |
| 1042 | |
| 1043 | /* CPU Usage */ |
| 1044 | if (ts->total_run_time) { |
| 1045 | double runt = (double) ts->total_run_time; |
| 1046 | |
| 1047 | usr_cpu = (double) ts->usr_time * 100 / runt; |
| 1048 | sys_cpu = (double) ts->sys_time * 100 / runt; |
| 1049 | } else { |
| 1050 | usr_cpu = 0; |
| 1051 | sys_cpu = 0; |
| 1052 | } |
| 1053 | |
| 1054 | log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, |
| 1055 | (unsigned long long) ts->ctx, |
| 1056 | (unsigned long long) ts->majf, |
| 1057 | (unsigned long long) ts->minf); |
| 1058 | |
| 1059 | /* Calc % distribution of IO depths, usecond, msecond latency */ |
| 1060 | stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); |
| 1061 | stat_calc_lat_u(ts, io_u_lat_u); |
| 1062 | stat_calc_lat_m(ts, io_u_lat_m); |
| 1063 | |
| 1064 | /* Only show fixed 7 I/O depth levels*/ |
| 1065 | log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", |
| 1066 | io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], |
| 1067 | io_u_dist[4], io_u_dist[5], io_u_dist[6]); |
| 1068 | |
| 1069 | /* Microsecond latency */ |
| 1070 | for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) |
| 1071 | log_buf(out, ";%3.2f%%", io_u_lat_u[i]); |
| 1072 | /* Millisecond latency */ |
| 1073 | for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) |
| 1074 | log_buf(out, ";%3.2f%%", io_u_lat_m[i]); |
| 1075 | |
| 1076 | /* disk util stats, if any */ |
| 1077 | show_disk_util(1, NULL, out); |
| 1078 | |
| 1079 | /* Additional output if continue_on_error set - default off*/ |
| 1080 | if (ts->continue_on_error) |
| 1081 | log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); |
| 1082 | |
| 1083 | /* Additional output if description is set */ |
| 1084 | if (strlen(ts->description)) |
| 1085 | log_buf(out, ";%s", ts->description); |
| 1086 | |
| 1087 | log_buf(out, "\n"); |
| 1088 | } |
| 1089 | |
| 1090 | void json_add_job_opts(struct json_object *root, const char *name, |
| 1091 | struct flist_head *opt_list, bool num_jobs) |
| 1092 | { |
| 1093 | struct json_object *dir_object; |
| 1094 | struct flist_head *entry; |
| 1095 | struct print_option *p; |
| 1096 | |
| 1097 | if (flist_empty(opt_list)) |
| 1098 | return; |
| 1099 | |
| 1100 | dir_object = json_create_object(); |
| 1101 | json_object_add_value_object(root, name, dir_object); |
| 1102 | |
| 1103 | flist_for_each(entry, opt_list) { |
| 1104 | const char *pos = ""; |
| 1105 | |
| 1106 | p = flist_entry(entry, struct print_option, list); |
| 1107 | if (!num_jobs && !strcmp(p->name, "numjobs")) |
| 1108 | continue; |
| 1109 | if (p->value) |
| 1110 | pos = p->value; |
| 1111 | json_object_add_value_string(dir_object, p->name, pos); |
| 1112 | } |
| 1113 | } |
| 1114 | |
| 1115 | static struct json_object *show_thread_status_json(struct thread_stat *ts, |
| 1116 | struct group_run_stats *rs, |
| 1117 | struct flist_head *opt_list) |
| 1118 | { |
| 1119 | struct json_object *root, *tmp; |
| 1120 | struct jobs_eta *je; |
| 1121 | double io_u_dist[FIO_IO_U_MAP_NR]; |
| 1122 | double io_u_lat_u[FIO_IO_U_LAT_U_NR]; |
| 1123 | double io_u_lat_m[FIO_IO_U_LAT_M_NR]; |
| 1124 | double usr_cpu, sys_cpu; |
| 1125 | int i; |
| 1126 | size_t size; |
| 1127 | |
| 1128 | root = json_create_object(); |
| 1129 | json_object_add_value_string(root, "jobname", ts->name); |
| 1130 | json_object_add_value_int(root, "groupid", ts->groupid); |
| 1131 | json_object_add_value_int(root, "error", ts->error); |
| 1132 | |
| 1133 | /* ETA Info */ |
| 1134 | je = get_jobs_eta(true, &size); |
| 1135 | if (je) { |
| 1136 | json_object_add_value_int(root, "eta", je->eta_sec); |
| 1137 | json_object_add_value_int(root, "elapsed", je->elapsed_sec); |
| 1138 | } |
| 1139 | |
| 1140 | if (opt_list) |
| 1141 | json_add_job_opts(root, "job options", opt_list, true); |
| 1142 | |
| 1143 | add_ddir_status_json(ts, rs, DDIR_READ, root); |
| 1144 | add_ddir_status_json(ts, rs, DDIR_WRITE, root); |
| 1145 | add_ddir_status_json(ts, rs, DDIR_TRIM, root); |
| 1146 | |
| 1147 | /* CPU Usage */ |
| 1148 | if (ts->total_run_time) { |
| 1149 | double runt = (double) ts->total_run_time; |
| 1150 | |
| 1151 | usr_cpu = (double) ts->usr_time * 100 / runt; |
| 1152 | sys_cpu = (double) ts->sys_time * 100 / runt; |
| 1153 | } else { |
| 1154 | usr_cpu = 0; |
| 1155 | sys_cpu = 0; |
| 1156 | } |
| 1157 | json_object_add_value_float(root, "usr_cpu", usr_cpu); |
| 1158 | json_object_add_value_float(root, "sys_cpu", sys_cpu); |
| 1159 | json_object_add_value_int(root, "ctx", ts->ctx); |
| 1160 | json_object_add_value_int(root, "majf", ts->majf); |
| 1161 | json_object_add_value_int(root, "minf", ts->minf); |
| 1162 | |
| 1163 | |
| 1164 | /* Calc % distribution of IO depths, usecond, msecond latency */ |
| 1165 | stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); |
| 1166 | stat_calc_lat_u(ts, io_u_lat_u); |
| 1167 | stat_calc_lat_m(ts, io_u_lat_m); |
| 1168 | |
| 1169 | tmp = json_create_object(); |
| 1170 | json_object_add_value_object(root, "iodepth_level", tmp); |
| 1171 | /* Only show fixed 7 I/O depth levels*/ |
| 1172 | for (i = 0; i < 7; i++) { |
| 1173 | char name[20]; |
| 1174 | if (i < 6) |
| 1175 | snprintf(name, 20, "%d", 1 << i); |
| 1176 | else |
| 1177 | snprintf(name, 20, ">=%d", 1 << i); |
| 1178 | json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]); |
| 1179 | } |
| 1180 | |
| 1181 | tmp = json_create_object(); |
| 1182 | json_object_add_value_object(root, "latency_us", tmp); |
| 1183 | /* Microsecond latency */ |
| 1184 | for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) { |
| 1185 | const char *ranges[] = { "2", "4", "10", "20", "50", "100", |
| 1186 | "250", "500", "750", "1000", }; |
| 1187 | json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]); |
| 1188 | } |
| 1189 | /* Millisecond latency */ |
| 1190 | tmp = json_create_object(); |
| 1191 | json_object_add_value_object(root, "latency_ms", tmp); |
| 1192 | for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) { |
| 1193 | const char *ranges[] = { "2", "4", "10", "20", "50", "100", |
| 1194 | "250", "500", "750", "1000", "2000", |
| 1195 | ">=2000", }; |
| 1196 | json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]); |
| 1197 | } |
| 1198 | |
| 1199 | /* Additional output if continue_on_error set - default off*/ |
| 1200 | if (ts->continue_on_error) { |
| 1201 | json_object_add_value_int(root, "total_err", ts->total_err_count); |
| 1202 | json_object_add_value_int(root, "first_error", ts->first_error); |
| 1203 | } |
| 1204 | |
| 1205 | if (ts->latency_depth) { |
| 1206 | json_object_add_value_int(root, "latency_depth", ts->latency_depth); |
| 1207 | json_object_add_value_int(root, "latency_target", ts->latency_target); |
| 1208 | json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f); |
| 1209 | json_object_add_value_int(root, "latency_window", ts->latency_window); |
| 1210 | } |
| 1211 | |
| 1212 | /* Additional output if description is set */ |
| 1213 | if (strlen(ts->description)) |
| 1214 | json_object_add_value_string(root, "desc", ts->description); |
| 1215 | |
| 1216 | if (ts->nr_block_infos) { |
| 1217 | /* Block error histogram and types */ |
| 1218 | int len; |
| 1219 | unsigned int *percentiles = NULL; |
| 1220 | unsigned int block_state_counts[BLOCK_STATE_COUNT]; |
| 1221 | |
| 1222 | len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos, |
| 1223 | ts->percentile_list, |
| 1224 | &percentiles, block_state_counts); |
| 1225 | |
| 1226 | if (len) { |
| 1227 | struct json_object *block, *percentile_object, *states; |
| 1228 | int state; |
| 1229 | block = json_create_object(); |
| 1230 | json_object_add_value_object(root, "block", block); |
| 1231 | |
| 1232 | percentile_object = json_create_object(); |
| 1233 | json_object_add_value_object(block, "percentiles", |
| 1234 | percentile_object); |
| 1235 | for (i = 0; i < len; i++) { |
| 1236 | char buf[20]; |
| 1237 | snprintf(buf, sizeof(buf), "%f", |
| 1238 | ts->percentile_list[i].u.f); |
| 1239 | json_object_add_value_int(percentile_object, |
| 1240 | (const char *)buf, |
| 1241 | percentiles[i]); |
| 1242 | } |
| 1243 | |
| 1244 | states = json_create_object(); |
| 1245 | json_object_add_value_object(block, "states", states); |
| 1246 | for (state = 0; state < BLOCK_STATE_COUNT; state++) { |
| 1247 | json_object_add_value_int(states, |
| 1248 | block_state_names[state], |
| 1249 | block_state_counts[state]); |
| 1250 | } |
| 1251 | free(percentiles); |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | return root; |
| 1256 | } |
| 1257 | |
| 1258 | static void show_thread_status_terse(struct thread_stat *ts, |
| 1259 | struct group_run_stats *rs, |
| 1260 | struct buf_output *out) |
| 1261 | { |
| 1262 | if (terse_version == 2) |
| 1263 | show_thread_status_terse_v2(ts, rs, out); |
| 1264 | else if (terse_version == 3 || terse_version == 4) |
| 1265 | show_thread_status_terse_v3_v4(ts, rs, terse_version, out); |
| 1266 | else |
| 1267 | log_err("fio: bad terse version!? %d\n", terse_version); |
| 1268 | } |
| 1269 | |
| 1270 | struct json_object *show_thread_status(struct thread_stat *ts, |
| 1271 | struct group_run_stats *rs, |
| 1272 | struct flist_head *opt_list, |
| 1273 | struct buf_output *out) |
| 1274 | { |
| 1275 | struct json_object *ret = NULL; |
| 1276 | |
| 1277 | if (output_format & FIO_OUTPUT_TERSE) |
| 1278 | show_thread_status_terse(ts, rs, out); |
| 1279 | if (output_format & FIO_OUTPUT_JSON) |
| 1280 | ret = show_thread_status_json(ts, rs, opt_list); |
| 1281 | if (output_format & FIO_OUTPUT_NORMAL) |
| 1282 | show_thread_status_normal(ts, rs, out); |
| 1283 | |
| 1284 | return ret; |
| 1285 | } |
| 1286 | |
| 1287 | static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first) |
| 1288 | { |
| 1289 | double mean, S; |
| 1290 | |
| 1291 | if (src->samples == 0) |
| 1292 | return; |
| 1293 | |
| 1294 | dst->min_val = min(dst->min_val, src->min_val); |
| 1295 | dst->max_val = max(dst->max_val, src->max_val); |
| 1296 | |
| 1297 | /* |
| 1298 | * Compute new mean and S after the merge |
| 1299 | * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance |
| 1300 | * #Parallel_algorithm> |
| 1301 | */ |
| 1302 | if (first) { |
| 1303 | mean = src->mean.u.f; |
| 1304 | S = src->S.u.f; |
| 1305 | } else { |
| 1306 | double delta = src->mean.u.f - dst->mean.u.f; |
| 1307 | |
| 1308 | mean = ((src->mean.u.f * src->samples) + |
| 1309 | (dst->mean.u.f * dst->samples)) / |
| 1310 | (dst->samples + src->samples); |
| 1311 | |
| 1312 | S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * |
| 1313 | (dst->samples * src->samples) / |
| 1314 | (dst->samples + src->samples); |
| 1315 | } |
| 1316 | |
| 1317 | dst->samples += src->samples; |
| 1318 | dst->mean.u.f = mean; |
| 1319 | dst->S.u.f = S; |
| 1320 | } |
| 1321 | |
| 1322 | void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) |
| 1323 | { |
| 1324 | int i; |
| 1325 | |
| 1326 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 1327 | if (dst->max_run[i] < src->max_run[i]) |
| 1328 | dst->max_run[i] = src->max_run[i]; |
| 1329 | if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) |
| 1330 | dst->min_run[i] = src->min_run[i]; |
| 1331 | if (dst->max_bw[i] < src->max_bw[i]) |
| 1332 | dst->max_bw[i] = src->max_bw[i]; |
| 1333 | if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) |
| 1334 | dst->min_bw[i] = src->min_bw[i]; |
| 1335 | |
| 1336 | dst->io_kb[i] += src->io_kb[i]; |
| 1337 | dst->agg[i] += src->agg[i]; |
| 1338 | } |
| 1339 | |
| 1340 | if (!dst->kb_base) |
| 1341 | dst->kb_base = src->kb_base; |
| 1342 | if (!dst->unit_base) |
| 1343 | dst->unit_base = src->unit_base; |
| 1344 | } |
| 1345 | |
| 1346 | void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, |
| 1347 | bool first) |
| 1348 | { |
| 1349 | int l, k; |
| 1350 | |
| 1351 | for (l = 0; l < DDIR_RWDIR_CNT; l++) { |
| 1352 | if (!dst->unified_rw_rep) { |
| 1353 | sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first); |
| 1354 | sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first); |
| 1355 | sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first); |
| 1356 | sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first); |
| 1357 | |
| 1358 | dst->io_bytes[l] += src->io_bytes[l]; |
| 1359 | |
| 1360 | if (dst->runtime[l] < src->runtime[l]) |
| 1361 | dst->runtime[l] = src->runtime[l]; |
| 1362 | } else { |
| 1363 | sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first); |
| 1364 | sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first); |
| 1365 | sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first); |
| 1366 | sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first); |
| 1367 | |
| 1368 | dst->io_bytes[0] += src->io_bytes[l]; |
| 1369 | |
| 1370 | if (dst->runtime[0] < src->runtime[l]) |
| 1371 | dst->runtime[0] = src->runtime[l]; |
| 1372 | |
| 1373 | /* |
| 1374 | * We're summing to the same destination, so override |
| 1375 | * 'first' after the first iteration of the loop |
| 1376 | */ |
| 1377 | first = false; |
| 1378 | } |
| 1379 | } |
| 1380 | |
| 1381 | dst->usr_time += src->usr_time; |
| 1382 | dst->sys_time += src->sys_time; |
| 1383 | dst->ctx += src->ctx; |
| 1384 | dst->majf += src->majf; |
| 1385 | dst->minf += src->minf; |
| 1386 | |
| 1387 | for (k = 0; k < FIO_IO_U_MAP_NR; k++) |
| 1388 | dst->io_u_map[k] += src->io_u_map[k]; |
| 1389 | for (k = 0; k < FIO_IO_U_MAP_NR; k++) |
| 1390 | dst->io_u_submit[k] += src->io_u_submit[k]; |
| 1391 | for (k = 0; k < FIO_IO_U_MAP_NR; k++) |
| 1392 | dst->io_u_complete[k] += src->io_u_complete[k]; |
| 1393 | for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) |
| 1394 | dst->io_u_lat_u[k] += src->io_u_lat_u[k]; |
| 1395 | for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) |
| 1396 | dst->io_u_lat_m[k] += src->io_u_lat_m[k]; |
| 1397 | |
| 1398 | for (k = 0; k < DDIR_RWDIR_CNT; k++) { |
| 1399 | if (!dst->unified_rw_rep) { |
| 1400 | dst->total_io_u[k] += src->total_io_u[k]; |
| 1401 | dst->short_io_u[k] += src->short_io_u[k]; |
| 1402 | dst->drop_io_u[k] += src->drop_io_u[k]; |
| 1403 | } else { |
| 1404 | dst->total_io_u[0] += src->total_io_u[k]; |
| 1405 | dst->short_io_u[0] += src->short_io_u[k]; |
| 1406 | dst->drop_io_u[0] += src->drop_io_u[k]; |
| 1407 | } |
| 1408 | } |
| 1409 | |
| 1410 | for (k = 0; k < DDIR_RWDIR_CNT; k++) { |
| 1411 | int m; |
| 1412 | |
| 1413 | for (m = 0; m < FIO_IO_U_PLAT_NR; m++) { |
| 1414 | if (!dst->unified_rw_rep) |
| 1415 | dst->io_u_plat[k][m] += src->io_u_plat[k][m]; |
| 1416 | else |
| 1417 | dst->io_u_plat[0][m] += src->io_u_plat[k][m]; |
| 1418 | } |
| 1419 | } |
| 1420 | |
| 1421 | dst->total_run_time += src->total_run_time; |
| 1422 | dst->total_submit += src->total_submit; |
| 1423 | dst->total_complete += src->total_complete; |
| 1424 | } |
| 1425 | |
| 1426 | void init_group_run_stat(struct group_run_stats *gs) |
| 1427 | { |
| 1428 | int i; |
| 1429 | memset(gs, 0, sizeof(*gs)); |
| 1430 | |
| 1431 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 1432 | gs->min_bw[i] = gs->min_run[i] = ~0UL; |
| 1433 | } |
| 1434 | |
| 1435 | void init_thread_stat(struct thread_stat *ts) |
| 1436 | { |
| 1437 | int j; |
| 1438 | |
| 1439 | memset(ts, 0, sizeof(*ts)); |
| 1440 | |
| 1441 | for (j = 0; j < DDIR_RWDIR_CNT; j++) { |
| 1442 | ts->lat_stat[j].min_val = -1UL; |
| 1443 | ts->clat_stat[j].min_val = -1UL; |
| 1444 | ts->slat_stat[j].min_val = -1UL; |
| 1445 | ts->bw_stat[j].min_val = -1UL; |
| 1446 | } |
| 1447 | ts->groupid = -1; |
| 1448 | } |
| 1449 | |
| 1450 | void __show_run_stats(void) |
| 1451 | { |
| 1452 | struct group_run_stats *runstats, *rs; |
| 1453 | struct thread_data *td; |
| 1454 | struct thread_stat *threadstats, *ts; |
| 1455 | int i, j, k, nr_ts, last_ts, idx; |
| 1456 | int kb_base_warned = 0; |
| 1457 | int unit_base_warned = 0; |
| 1458 | struct json_object *root = NULL; |
| 1459 | struct json_array *array = NULL; |
| 1460 | struct buf_output output[FIO_OUTPUT_NR]; |
| 1461 | struct flist_head **opt_lists; |
| 1462 | |
| 1463 | runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); |
| 1464 | |
| 1465 | for (i = 0; i < groupid + 1; i++) |
| 1466 | init_group_run_stat(&runstats[i]); |
| 1467 | |
| 1468 | /* |
| 1469 | * find out how many threads stats we need. if group reporting isn't |
| 1470 | * enabled, it's one-per-td. |
| 1471 | */ |
| 1472 | nr_ts = 0; |
| 1473 | last_ts = -1; |
| 1474 | for_each_td(td, i) { |
| 1475 | if (!td->o.group_reporting) { |
| 1476 | nr_ts++; |
| 1477 | continue; |
| 1478 | } |
| 1479 | if (last_ts == td->groupid) |
| 1480 | continue; |
| 1481 | |
| 1482 | last_ts = td->groupid; |
| 1483 | nr_ts++; |
| 1484 | } |
| 1485 | |
| 1486 | threadstats = malloc(nr_ts * sizeof(struct thread_stat)); |
| 1487 | opt_lists = malloc(nr_ts * sizeof(struct flist_head *)); |
| 1488 | |
| 1489 | for (i = 0; i < nr_ts; i++) { |
| 1490 | init_thread_stat(&threadstats[i]); |
| 1491 | opt_lists[i] = NULL; |
| 1492 | } |
| 1493 | |
| 1494 | j = 0; |
| 1495 | last_ts = -1; |
| 1496 | idx = 0; |
| 1497 | for_each_td(td, i) { |
| 1498 | if (idx && (!td->o.group_reporting || |
| 1499 | (td->o.group_reporting && last_ts != td->groupid))) { |
| 1500 | idx = 0; |
| 1501 | j++; |
| 1502 | } |
| 1503 | |
| 1504 | last_ts = td->groupid; |
| 1505 | |
| 1506 | ts = &threadstats[j]; |
| 1507 | |
| 1508 | ts->clat_percentiles = td->o.clat_percentiles; |
| 1509 | ts->percentile_precision = td->o.percentile_precision; |
| 1510 | memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); |
| 1511 | opt_lists[j] = &td->opt_list; |
| 1512 | |
| 1513 | idx++; |
| 1514 | ts->members++; |
| 1515 | |
| 1516 | if (ts->groupid == -1) { |
| 1517 | /* |
| 1518 | * These are per-group shared already |
| 1519 | */ |
| 1520 | strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1); |
| 1521 | if (td->o.description) |
| 1522 | strncpy(ts->description, td->o.description, |
| 1523 | FIO_JOBDESC_SIZE - 1); |
| 1524 | else |
| 1525 | memset(ts->description, 0, FIO_JOBDESC_SIZE); |
| 1526 | |
| 1527 | /* |
| 1528 | * If multiple entries in this group, this is |
| 1529 | * the first member. |
| 1530 | */ |
| 1531 | ts->thread_number = td->thread_number; |
| 1532 | ts->groupid = td->groupid; |
| 1533 | |
| 1534 | /* |
| 1535 | * first pid in group, not very useful... |
| 1536 | */ |
| 1537 | ts->pid = td->pid; |
| 1538 | |
| 1539 | ts->kb_base = td->o.kb_base; |
| 1540 | ts->unit_base = td->o.unit_base; |
| 1541 | ts->unified_rw_rep = td->o.unified_rw_rep; |
| 1542 | } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { |
| 1543 | log_info("fio: kb_base differs for jobs in group, using" |
| 1544 | " %u as the base\n", ts->kb_base); |
| 1545 | kb_base_warned = 1; |
| 1546 | } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) { |
| 1547 | log_info("fio: unit_base differs for jobs in group, using" |
| 1548 | " %u as the base\n", ts->unit_base); |
| 1549 | unit_base_warned = 1; |
| 1550 | } |
| 1551 | |
| 1552 | ts->continue_on_error = td->o.continue_on_error; |
| 1553 | ts->total_err_count += td->total_err_count; |
| 1554 | ts->first_error = td->first_error; |
| 1555 | if (!ts->error) { |
| 1556 | if (!td->error && td->o.continue_on_error && |
| 1557 | td->first_error) { |
| 1558 | ts->error = td->first_error; |
| 1559 | ts->verror[sizeof(ts->verror) - 1] = '\0'; |
| 1560 | strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1); |
| 1561 | } else if (td->error) { |
| 1562 | ts->error = td->error; |
| 1563 | ts->verror[sizeof(ts->verror) - 1] = '\0'; |
| 1564 | strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1); |
| 1565 | } |
| 1566 | } |
| 1567 | |
| 1568 | ts->latency_depth = td->latency_qd; |
| 1569 | ts->latency_target = td->o.latency_target; |
| 1570 | ts->latency_percentile = td->o.latency_percentile; |
| 1571 | ts->latency_window = td->o.latency_window; |
| 1572 | |
| 1573 | ts->nr_block_infos = td->ts.nr_block_infos; |
| 1574 | for (k = 0; k < ts->nr_block_infos; k++) |
| 1575 | ts->block_infos[k] = td->ts.block_infos[k]; |
| 1576 | |
| 1577 | sum_thread_stats(ts, &td->ts, idx == 1); |
| 1578 | } |
| 1579 | |
| 1580 | for (i = 0; i < nr_ts; i++) { |
| 1581 | unsigned long long bw; |
| 1582 | |
| 1583 | ts = &threadstats[i]; |
| 1584 | if (ts->groupid == -1) |
| 1585 | continue; |
| 1586 | rs = &runstats[ts->groupid]; |
| 1587 | rs->kb_base = ts->kb_base; |
| 1588 | rs->unit_base = ts->unit_base; |
| 1589 | rs->unified_rw_rep += ts->unified_rw_rep; |
| 1590 | |
| 1591 | for (j = 0; j < DDIR_RWDIR_CNT; j++) { |
| 1592 | if (!ts->runtime[j]) |
| 1593 | continue; |
| 1594 | if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) |
| 1595 | rs->min_run[j] = ts->runtime[j]; |
| 1596 | if (ts->runtime[j] > rs->max_run[j]) |
| 1597 | rs->max_run[j] = ts->runtime[j]; |
| 1598 | |
| 1599 | bw = 0; |
| 1600 | if (ts->runtime[j]) { |
| 1601 | unsigned long runt = ts->runtime[j]; |
| 1602 | unsigned long long kb; |
| 1603 | |
| 1604 | kb = ts->io_bytes[j] / rs->kb_base; |
| 1605 | bw = kb * 1000 / runt; |
| 1606 | } |
| 1607 | if (bw < rs->min_bw[j]) |
| 1608 | rs->min_bw[j] = bw; |
| 1609 | if (bw > rs->max_bw[j]) |
| 1610 | rs->max_bw[j] = bw; |
| 1611 | |
| 1612 | rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | for (i = 0; i < groupid + 1; i++) { |
| 1617 | int ddir; |
| 1618 | |
| 1619 | rs = &runstats[i]; |
| 1620 | |
| 1621 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) { |
| 1622 | if (rs->max_run[ddir]) |
| 1623 | rs->agg[ddir] = (rs->io_kb[ddir] * 1000) / |
| 1624 | rs->max_run[ddir]; |
| 1625 | } |
| 1626 | } |
| 1627 | |
| 1628 | for (i = 0; i < FIO_OUTPUT_NR; i++) |
| 1629 | buf_output_init(&output[i]); |
| 1630 | |
| 1631 | /* |
| 1632 | * don't overwrite last signal output |
| 1633 | */ |
| 1634 | if (output_format & FIO_OUTPUT_NORMAL) |
| 1635 | log_buf(&output[__FIO_OUTPUT_NORMAL], "\n"); |
| 1636 | if (output_format & FIO_OUTPUT_JSON) { |
| 1637 | struct thread_data *global; |
| 1638 | char time_buf[32]; |
| 1639 | time_t time_p; |
| 1640 | |
| 1641 | time(&time_p); |
| 1642 | os_ctime_r((const time_t *) &time_p, time_buf, |
| 1643 | sizeof(time_buf)); |
| 1644 | time_buf[strlen(time_buf) - 1] = '\0'; |
| 1645 | |
| 1646 | root = json_create_object(); |
| 1647 | json_object_add_value_string(root, "fio version", fio_version_string); |
| 1648 | json_object_add_value_int(root, "timestamp", time_p); |
| 1649 | json_object_add_value_string(root, "time", time_buf); |
| 1650 | global = get_global_options(); |
| 1651 | json_add_job_opts(root, "global options", &global->opt_list, false); |
| 1652 | array = json_create_array(); |
| 1653 | json_object_add_value_array(root, "jobs", array); |
| 1654 | } |
| 1655 | |
| 1656 | if (is_backend) |
| 1657 | fio_server_send_job_options(&get_global_options()->opt_list, -1U); |
| 1658 | |
| 1659 | for (i = 0; i < nr_ts; i++) { |
| 1660 | ts = &threadstats[i]; |
| 1661 | rs = &runstats[ts->groupid]; |
| 1662 | |
| 1663 | if (is_backend) { |
| 1664 | fio_server_send_job_options(opt_lists[i], i); |
| 1665 | fio_server_send_ts(ts, rs); |
| 1666 | } else { |
| 1667 | if (output_format & FIO_OUTPUT_TERSE) |
| 1668 | show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]); |
| 1669 | if (output_format & FIO_OUTPUT_JSON) { |
| 1670 | struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]); |
| 1671 | json_array_add_value_object(array, tmp); |
| 1672 | } |
| 1673 | if (output_format & FIO_OUTPUT_NORMAL) |
| 1674 | show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]); |
| 1675 | } |
| 1676 | } |
| 1677 | if (!is_backend && (output_format & FIO_OUTPUT_JSON)) { |
| 1678 | /* disk util stats, if any */ |
| 1679 | show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]); |
| 1680 | |
| 1681 | show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]); |
| 1682 | |
| 1683 | json_print_object(root, &output[__FIO_OUTPUT_JSON]); |
| 1684 | log_buf(&output[__FIO_OUTPUT_JSON], "\n"); |
| 1685 | json_free_object(root); |
| 1686 | } |
| 1687 | |
| 1688 | for (i = 0; i < groupid + 1; i++) { |
| 1689 | rs = &runstats[i]; |
| 1690 | |
| 1691 | rs->groupid = i; |
| 1692 | if (is_backend) |
| 1693 | fio_server_send_gs(rs); |
| 1694 | else if (output_format & FIO_OUTPUT_NORMAL) |
| 1695 | show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]); |
| 1696 | } |
| 1697 | |
| 1698 | if (is_backend) |
| 1699 | fio_server_send_du(); |
| 1700 | else if (output_format & FIO_OUTPUT_NORMAL) { |
| 1701 | show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]); |
| 1702 | show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]); |
| 1703 | } |
| 1704 | |
| 1705 | for (i = 0; i < FIO_OUTPUT_NR; i++) { |
| 1706 | buf_output_flush(&output[i]); |
| 1707 | buf_output_free(&output[i]); |
| 1708 | } |
| 1709 | |
| 1710 | log_info_flush(); |
| 1711 | free(runstats); |
| 1712 | free(threadstats); |
| 1713 | free(opt_lists); |
| 1714 | } |
| 1715 | |
| 1716 | void show_run_stats(void) |
| 1717 | { |
| 1718 | fio_mutex_down(stat_mutex); |
| 1719 | __show_run_stats(); |
| 1720 | fio_mutex_up(stat_mutex); |
| 1721 | } |
| 1722 | |
| 1723 | void __show_running_run_stats(void) |
| 1724 | { |
| 1725 | struct thread_data *td; |
| 1726 | unsigned long long *rt; |
| 1727 | struct timeval tv; |
| 1728 | int i; |
| 1729 | |
| 1730 | fio_mutex_down(stat_mutex); |
| 1731 | |
| 1732 | rt = malloc(thread_number * sizeof(unsigned long long)); |
| 1733 | fio_gettime(&tv, NULL); |
| 1734 | |
| 1735 | for_each_td(td, i) { |
| 1736 | td->update_rusage = 1; |
| 1737 | td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; |
| 1738 | td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; |
| 1739 | td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; |
| 1740 | td->ts.total_run_time = mtime_since(&td->epoch, &tv); |
| 1741 | |
| 1742 | rt[i] = mtime_since(&td->start, &tv); |
| 1743 | if (td_read(td) && td->ts.io_bytes[DDIR_READ]) |
| 1744 | td->ts.runtime[DDIR_READ] += rt[i]; |
| 1745 | if (td_write(td) && td->ts.io_bytes[DDIR_WRITE]) |
| 1746 | td->ts.runtime[DDIR_WRITE] += rt[i]; |
| 1747 | if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM]) |
| 1748 | td->ts.runtime[DDIR_TRIM] += rt[i]; |
| 1749 | } |
| 1750 | |
| 1751 | for_each_td(td, i) { |
| 1752 | if (td->runstate >= TD_EXITED) |
| 1753 | continue; |
| 1754 | if (td->rusage_sem) { |
| 1755 | td->update_rusage = 1; |
| 1756 | fio_mutex_down(td->rusage_sem); |
| 1757 | } |
| 1758 | td->update_rusage = 0; |
| 1759 | } |
| 1760 | |
| 1761 | __show_run_stats(); |
| 1762 | |
| 1763 | for_each_td(td, i) { |
| 1764 | if (td_read(td) && td->ts.io_bytes[DDIR_READ]) |
| 1765 | td->ts.runtime[DDIR_READ] -= rt[i]; |
| 1766 | if (td_write(td) && td->ts.io_bytes[DDIR_WRITE]) |
| 1767 | td->ts.runtime[DDIR_WRITE] -= rt[i]; |
| 1768 | if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM]) |
| 1769 | td->ts.runtime[DDIR_TRIM] -= rt[i]; |
| 1770 | } |
| 1771 | |
| 1772 | free(rt); |
| 1773 | fio_mutex_up(stat_mutex); |
| 1774 | } |
| 1775 | |
| 1776 | static int status_interval_init; |
| 1777 | static struct timeval status_time; |
| 1778 | static int status_file_disabled; |
| 1779 | |
| 1780 | #define FIO_STATUS_FILE "fio-dump-status" |
| 1781 | |
| 1782 | static int check_status_file(void) |
| 1783 | { |
| 1784 | struct stat sb; |
| 1785 | const char *temp_dir; |
| 1786 | char fio_status_file_path[PATH_MAX]; |
| 1787 | |
| 1788 | if (status_file_disabled) |
| 1789 | return 0; |
| 1790 | |
| 1791 | temp_dir = getenv("TMPDIR"); |
| 1792 | if (temp_dir == NULL) { |
| 1793 | temp_dir = getenv("TEMP"); |
| 1794 | if (temp_dir && strlen(temp_dir) >= PATH_MAX) |
| 1795 | temp_dir = NULL; |
| 1796 | } |
| 1797 | if (temp_dir == NULL) |
| 1798 | temp_dir = "/tmp"; |
| 1799 | |
| 1800 | snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE); |
| 1801 | |
| 1802 | if (stat(fio_status_file_path, &sb)) |
| 1803 | return 0; |
| 1804 | |
| 1805 | if (unlink(fio_status_file_path) < 0) { |
| 1806 | log_err("fio: failed to unlink %s: %s\n", fio_status_file_path, |
| 1807 | strerror(errno)); |
| 1808 | log_err("fio: disabling status file updates\n"); |
| 1809 | status_file_disabled = 1; |
| 1810 | } |
| 1811 | |
| 1812 | return 1; |
| 1813 | } |
| 1814 | |
| 1815 | void check_for_running_stats(void) |
| 1816 | { |
| 1817 | if (status_interval) { |
| 1818 | if (!status_interval_init) { |
| 1819 | fio_gettime(&status_time, NULL); |
| 1820 | status_interval_init = 1; |
| 1821 | } else if (mtime_since_now(&status_time) >= status_interval) { |
| 1822 | show_running_run_stats(); |
| 1823 | fio_gettime(&status_time, NULL); |
| 1824 | return; |
| 1825 | } |
| 1826 | } |
| 1827 | if (check_status_file()) { |
| 1828 | show_running_run_stats(); |
| 1829 | return; |
| 1830 | } |
| 1831 | } |
| 1832 | |
| 1833 | static inline void add_stat_sample(struct io_stat *is, unsigned long data) |
| 1834 | { |
| 1835 | double val = data; |
| 1836 | double delta; |
| 1837 | |
| 1838 | if (data > is->max_val) |
| 1839 | is->max_val = data; |
| 1840 | if (data < is->min_val) |
| 1841 | is->min_val = data; |
| 1842 | |
| 1843 | delta = val - is->mean.u.f; |
| 1844 | if (delta) { |
| 1845 | is->mean.u.f += delta / (is->samples + 1.0); |
| 1846 | is->S.u.f += delta * (val - is->mean.u.f); |
| 1847 | } |
| 1848 | |
| 1849 | is->samples++; |
| 1850 | } |
| 1851 | |
| 1852 | static void __add_log_sample(struct io_log *iolog, unsigned long val, |
| 1853 | enum fio_ddir ddir, unsigned int bs, |
| 1854 | unsigned long t, uint64_t offset) |
| 1855 | { |
| 1856 | uint64_t nr_samples = iolog->nr_samples; |
| 1857 | struct io_sample *s; |
| 1858 | |
| 1859 | if (iolog->disabled) |
| 1860 | return; |
| 1861 | |
| 1862 | if (!iolog->nr_samples) |
| 1863 | iolog->avg_last = t; |
| 1864 | |
| 1865 | if (iolog->nr_samples == iolog->max_samples) { |
| 1866 | size_t new_size, new_samples; |
| 1867 | void *new_log; |
| 1868 | |
| 1869 | if (!iolog->max_samples) |
| 1870 | new_samples = DEF_LOG_ENTRIES; |
| 1871 | else |
| 1872 | new_samples = iolog->max_samples * 2; |
| 1873 | |
| 1874 | new_size = new_samples * log_entry_sz(iolog); |
| 1875 | |
| 1876 | if (iolog->log_gz && (new_size > iolog->log_gz)) { |
| 1877 | if (!iolog->log) { |
| 1878 | iolog->log = malloc(new_size); |
| 1879 | iolog->max_samples = new_samples; |
| 1880 | } else if (iolog_flush(iolog, 0)) { |
| 1881 | log_err("fio: failed flushing iolog! Will stop logging.\n"); |
| 1882 | iolog->disabled = 1; |
| 1883 | return; |
| 1884 | } |
| 1885 | nr_samples = iolog->nr_samples; |
| 1886 | } else { |
| 1887 | new_log = realloc(iolog->log, new_size); |
| 1888 | if (!new_log) { |
| 1889 | log_err("fio: failed extending iolog! Will stop logging.\n"); |
| 1890 | iolog->disabled = 1; |
| 1891 | return; |
| 1892 | } |
| 1893 | iolog->log = new_log; |
| 1894 | iolog->max_samples = new_samples; |
| 1895 | } |
| 1896 | } |
| 1897 | |
| 1898 | s = get_sample(iolog, nr_samples); |
| 1899 | |
| 1900 | s->val = val; |
| 1901 | s->time = t; |
| 1902 | io_sample_set_ddir(iolog, s, ddir); |
| 1903 | s->bs = bs; |
| 1904 | |
| 1905 | if (iolog->log_offset) { |
| 1906 | struct io_sample_offset *so = (void *) s; |
| 1907 | |
| 1908 | so->offset = offset; |
| 1909 | } |
| 1910 | |
| 1911 | iolog->nr_samples++; |
| 1912 | } |
| 1913 | |
| 1914 | static inline void reset_io_stat(struct io_stat *ios) |
| 1915 | { |
| 1916 | ios->max_val = ios->min_val = ios->samples = 0; |
| 1917 | ios->mean.u.f = ios->S.u.f = 0; |
| 1918 | } |
| 1919 | |
| 1920 | void reset_io_stats(struct thread_data *td) |
| 1921 | { |
| 1922 | struct thread_stat *ts = &td->ts; |
| 1923 | int i, j; |
| 1924 | |
| 1925 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 1926 | reset_io_stat(&ts->clat_stat[i]); |
| 1927 | reset_io_stat(&ts->slat_stat[i]); |
| 1928 | reset_io_stat(&ts->lat_stat[i]); |
| 1929 | reset_io_stat(&ts->bw_stat[i]); |
| 1930 | reset_io_stat(&ts->iops_stat[i]); |
| 1931 | |
| 1932 | ts->io_bytes[i] = 0; |
| 1933 | ts->runtime[i] = 0; |
| 1934 | |
| 1935 | for (j = 0; j < FIO_IO_U_PLAT_NR; j++) |
| 1936 | ts->io_u_plat[i][j] = 0; |
| 1937 | } |
| 1938 | |
| 1939 | for (i = 0; i < FIO_IO_U_MAP_NR; i++) { |
| 1940 | ts->io_u_map[i] = 0; |
| 1941 | ts->io_u_submit[i] = 0; |
| 1942 | ts->io_u_complete[i] = 0; |
| 1943 | ts->io_u_lat_u[i] = 0; |
| 1944 | ts->io_u_lat_m[i] = 0; |
| 1945 | ts->total_submit = 0; |
| 1946 | ts->total_complete = 0; |
| 1947 | } |
| 1948 | |
| 1949 | for (i = 0; i < 3; i++) { |
| 1950 | ts->total_io_u[i] = 0; |
| 1951 | ts->short_io_u[i] = 0; |
| 1952 | ts->drop_io_u[i] = 0; |
| 1953 | } |
| 1954 | } |
| 1955 | |
| 1956 | static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir, |
| 1957 | unsigned long elapsed, bool log_max) |
| 1958 | { |
| 1959 | /* |
| 1960 | * Note an entry in the log. Use the mean from the logged samples, |
| 1961 | * making sure to properly round up. Only write a log entry if we |
| 1962 | * had actual samples done. |
| 1963 | */ |
| 1964 | if (iolog->avg_window[ddir].samples) { |
| 1965 | unsigned long val; |
| 1966 | |
| 1967 | if (log_max) |
| 1968 | val = iolog->avg_window[ddir].max_val; |
| 1969 | else |
| 1970 | val = iolog->avg_window[ddir].mean.u.f + 0.50; |
| 1971 | |
| 1972 | __add_log_sample(iolog, val, ddir, 0, elapsed, 0); |
| 1973 | } |
| 1974 | |
| 1975 | reset_io_stat(&iolog->avg_window[ddir]); |
| 1976 | } |
| 1977 | |
| 1978 | static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed, |
| 1979 | bool log_max) |
| 1980 | { |
| 1981 | int ddir; |
| 1982 | |
| 1983 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
| 1984 | __add_stat_to_log(iolog, ddir, elapsed, log_max); |
| 1985 | } |
| 1986 | |
| 1987 | static void add_log_sample(struct thread_data *td, struct io_log *iolog, |
| 1988 | unsigned long val, enum fio_ddir ddir, |
| 1989 | unsigned int bs, uint64_t offset) |
| 1990 | { |
| 1991 | unsigned long elapsed, this_window; |
| 1992 | |
| 1993 | if (!ddir_rw(ddir)) |
| 1994 | return; |
| 1995 | |
| 1996 | elapsed = mtime_since_now(&td->epoch); |
| 1997 | |
| 1998 | /* |
| 1999 | * If no time averaging, just add the log sample. |
| 2000 | */ |
| 2001 | if (!iolog->avg_msec) { |
| 2002 | __add_log_sample(iolog, val, ddir, bs, elapsed, offset); |
| 2003 | return; |
| 2004 | } |
| 2005 | |
| 2006 | /* |
| 2007 | * Add the sample. If the time period has passed, then |
| 2008 | * add that entry to the log and clear. |
| 2009 | */ |
| 2010 | add_stat_sample(&iolog->avg_window[ddir], val); |
| 2011 | |
| 2012 | /* |
| 2013 | * If period hasn't passed, adding the above sample is all we |
| 2014 | * need to do. |
| 2015 | */ |
| 2016 | this_window = elapsed - iolog->avg_last; |
| 2017 | if (this_window < iolog->avg_msec) |
| 2018 | return; |
| 2019 | |
| 2020 | _add_stat_to_log(iolog, elapsed, td->o.log_max != 0); |
| 2021 | |
| 2022 | iolog->avg_last = elapsed; |
| 2023 | } |
| 2024 | |
| 2025 | void finalize_logs(struct thread_data *td, bool unit_logs) |
| 2026 | { |
| 2027 | unsigned long elapsed; |
| 2028 | |
| 2029 | elapsed = mtime_since_now(&td->epoch); |
| 2030 | |
| 2031 | if (td->clat_log && unit_logs) |
| 2032 | _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0); |
| 2033 | if (td->slat_log && unit_logs) |
| 2034 | _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0); |
| 2035 | if (td->lat_log && unit_logs) |
| 2036 | _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0); |
| 2037 | if (td->bw_log && (unit_logs == per_unit_log(td->bw_log))) |
| 2038 | _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0); |
| 2039 | if (td->iops_log && (unit_logs == per_unit_log(td->iops_log))) |
| 2040 | _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0); |
| 2041 | } |
| 2042 | |
| 2043 | void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) |
| 2044 | { |
| 2045 | struct io_log *iolog; |
| 2046 | |
| 2047 | if (!ddir_rw(ddir)) |
| 2048 | return; |
| 2049 | |
| 2050 | iolog = agg_io_log[ddir]; |
| 2051 | __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0); |
| 2052 | } |
| 2053 | |
| 2054 | static void add_clat_percentile_sample(struct thread_stat *ts, |
| 2055 | unsigned long usec, enum fio_ddir ddir) |
| 2056 | { |
| 2057 | unsigned int idx = plat_val_to_idx(usec); |
| 2058 | assert(idx < FIO_IO_U_PLAT_NR); |
| 2059 | |
| 2060 | ts->io_u_plat[ddir][idx]++; |
| 2061 | } |
| 2062 | |
| 2063 | void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, |
| 2064 | unsigned long usec, unsigned int bs, uint64_t offset) |
| 2065 | { |
| 2066 | struct thread_stat *ts = &td->ts; |
| 2067 | |
| 2068 | td_io_u_lock(td); |
| 2069 | |
| 2070 | add_stat_sample(&ts->clat_stat[ddir], usec); |
| 2071 | |
| 2072 | if (td->clat_log) |
| 2073 | add_log_sample(td, td->clat_log, usec, ddir, bs, offset); |
| 2074 | |
| 2075 | if (ts->clat_percentiles) |
| 2076 | add_clat_percentile_sample(ts, usec, ddir); |
| 2077 | |
| 2078 | td_io_u_unlock(td); |
| 2079 | } |
| 2080 | |
| 2081 | void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, |
| 2082 | unsigned long usec, unsigned int bs, uint64_t offset) |
| 2083 | { |
| 2084 | struct thread_stat *ts = &td->ts; |
| 2085 | |
| 2086 | if (!ddir_rw(ddir)) |
| 2087 | return; |
| 2088 | |
| 2089 | td_io_u_lock(td); |
| 2090 | |
| 2091 | add_stat_sample(&ts->slat_stat[ddir], usec); |
| 2092 | |
| 2093 | if (td->slat_log) |
| 2094 | add_log_sample(td, td->slat_log, usec, ddir, bs, offset); |
| 2095 | |
| 2096 | td_io_u_unlock(td); |
| 2097 | } |
| 2098 | |
| 2099 | void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, |
| 2100 | unsigned long usec, unsigned int bs, uint64_t offset) |
| 2101 | { |
| 2102 | struct thread_stat *ts = &td->ts; |
| 2103 | |
| 2104 | if (!ddir_rw(ddir)) |
| 2105 | return; |
| 2106 | |
| 2107 | td_io_u_lock(td); |
| 2108 | |
| 2109 | add_stat_sample(&ts->lat_stat[ddir], usec); |
| 2110 | |
| 2111 | if (td->lat_log) |
| 2112 | add_log_sample(td, td->lat_log, usec, ddir, bs, offset); |
| 2113 | |
| 2114 | td_io_u_unlock(td); |
| 2115 | } |
| 2116 | |
| 2117 | void add_bw_sample(struct thread_data *td, struct io_u *io_u, |
| 2118 | unsigned int bytes, unsigned long spent) |
| 2119 | { |
| 2120 | struct thread_stat *ts = &td->ts; |
| 2121 | unsigned long rate; |
| 2122 | |
| 2123 | if (spent) |
| 2124 | rate = bytes * 1000 / spent; |
| 2125 | else |
| 2126 | rate = 0; |
| 2127 | |
| 2128 | td_io_u_lock(td); |
| 2129 | |
| 2130 | add_stat_sample(&ts->bw_stat[io_u->ddir], rate); |
| 2131 | |
| 2132 | if (td->bw_log) |
| 2133 | add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset); |
| 2134 | |
| 2135 | td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir]; |
| 2136 | td_io_u_unlock(td); |
| 2137 | } |
| 2138 | |
| 2139 | static int add_bw_samples(struct thread_data *td, struct timeval *t) |
| 2140 | { |
| 2141 | struct thread_stat *ts = &td->ts; |
| 2142 | unsigned long spent, rate; |
| 2143 | enum fio_ddir ddir; |
| 2144 | |
| 2145 | if (per_unit_log(td->bw_log)) |
| 2146 | return 0; |
| 2147 | |
| 2148 | spent = mtime_since(&td->bw_sample_time, t); |
| 2149 | if (spent < td->o.bw_avg_time && |
| 2150 | td->o.bw_avg_time - spent >= 10) |
| 2151 | return td->o.bw_avg_time - spent; |
| 2152 | |
| 2153 | td_io_u_lock(td); |
| 2154 | |
| 2155 | /* |
| 2156 | * Compute both read and write rates for the interval. |
| 2157 | */ |
| 2158 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { |
| 2159 | uint64_t delta; |
| 2160 | |
| 2161 | delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]; |
| 2162 | if (!delta) |
| 2163 | continue; /* No entries for interval */ |
| 2164 | |
| 2165 | if (spent) |
| 2166 | rate = delta * 1000 / spent / 1024; |
| 2167 | else |
| 2168 | rate = 0; |
| 2169 | |
| 2170 | add_stat_sample(&ts->bw_stat[ddir], rate); |
| 2171 | |
| 2172 | if (td->bw_log) { |
| 2173 | unsigned int bs = 0; |
| 2174 | |
| 2175 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
| 2176 | bs = td->o.min_bs[ddir]; |
| 2177 | |
| 2178 | add_log_sample(td, td->bw_log, rate, ddir, bs, 0); |
| 2179 | } |
| 2180 | |
| 2181 | td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; |
| 2182 | } |
| 2183 | |
| 2184 | timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time); |
| 2185 | |
| 2186 | td_io_u_unlock(td); |
| 2187 | |
| 2188 | if (spent <= td->o.bw_avg_time) |
| 2189 | return td->o.bw_avg_time; |
| 2190 | |
| 2191 | return td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time); |
| 2192 | } |
| 2193 | |
| 2194 | void add_iops_sample(struct thread_data *td, struct io_u *io_u, |
| 2195 | unsigned int bytes) |
| 2196 | { |
| 2197 | struct thread_stat *ts = &td->ts; |
| 2198 | |
| 2199 | td_io_u_lock(td); |
| 2200 | |
| 2201 | add_stat_sample(&ts->iops_stat[io_u->ddir], 1); |
| 2202 | |
| 2203 | if (td->iops_log) |
| 2204 | add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset); |
| 2205 | |
| 2206 | td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir]; |
| 2207 | td_io_u_unlock(td); |
| 2208 | } |
| 2209 | |
| 2210 | static int add_iops_samples(struct thread_data *td, struct timeval *t) |
| 2211 | { |
| 2212 | struct thread_stat *ts = &td->ts; |
| 2213 | unsigned long spent, iops; |
| 2214 | enum fio_ddir ddir; |
| 2215 | |
| 2216 | if (per_unit_log(td->iops_log)) |
| 2217 | return 0; |
| 2218 | |
| 2219 | spent = mtime_since(&td->iops_sample_time, t); |
| 2220 | if (spent < td->o.iops_avg_time && |
| 2221 | td->o.iops_avg_time - spent >= 10) |
| 2222 | return td->o.iops_avg_time - spent; |
| 2223 | |
| 2224 | td_io_u_lock(td); |
| 2225 | |
| 2226 | /* |
| 2227 | * Compute both read and write rates for the interval. |
| 2228 | */ |
| 2229 | for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { |
| 2230 | uint64_t delta; |
| 2231 | |
| 2232 | delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]; |
| 2233 | if (!delta) |
| 2234 | continue; /* No entries for interval */ |
| 2235 | |
| 2236 | if (spent) |
| 2237 | iops = (delta * 1000) / spent; |
| 2238 | else |
| 2239 | iops = 0; |
| 2240 | |
| 2241 | add_stat_sample(&ts->iops_stat[ddir], iops); |
| 2242 | |
| 2243 | if (td->iops_log) { |
| 2244 | unsigned int bs = 0; |
| 2245 | |
| 2246 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
| 2247 | bs = td->o.min_bs[ddir]; |
| 2248 | |
| 2249 | add_log_sample(td, td->iops_log, iops, ddir, bs, 0); |
| 2250 | } |
| 2251 | |
| 2252 | td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; |
| 2253 | } |
| 2254 | |
| 2255 | timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time); |
| 2256 | |
| 2257 | td_io_u_unlock(td); |
| 2258 | |
| 2259 | if (spent <= td->o.iops_avg_time) |
| 2260 | return td->o.iops_avg_time; |
| 2261 | |
| 2262 | return td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time); |
| 2263 | } |
| 2264 | |
| 2265 | /* |
| 2266 | * Returns msecs to next event |
| 2267 | */ |
| 2268 | int calc_log_samples(void) |
| 2269 | { |
| 2270 | struct thread_data *td; |
| 2271 | unsigned int next = ~0U, tmp; |
| 2272 | struct timeval now; |
| 2273 | int i; |
| 2274 | |
| 2275 | fio_gettime(&now, NULL); |
| 2276 | |
| 2277 | for_each_td(td, i) { |
| 2278 | if (!ramp_time_over(td) || |
| 2279 | !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) { |
| 2280 | next = min(td->o.iops_avg_time, td->o.bw_avg_time); |
| 2281 | continue; |
| 2282 | } |
| 2283 | if (!per_unit_log(td->bw_log)) { |
| 2284 | tmp = add_bw_samples(td, &now); |
| 2285 | if (tmp < next) |
| 2286 | next = tmp; |
| 2287 | } |
| 2288 | if (!per_unit_log(td->iops_log)) { |
| 2289 | tmp = add_iops_samples(td, &now); |
| 2290 | if (tmp < next) |
| 2291 | next = tmp; |
| 2292 | } |
| 2293 | } |
| 2294 | |
| 2295 | return next == ~0U ? 0 : next; |
| 2296 | } |
| 2297 | |
| 2298 | void stat_init(void) |
| 2299 | { |
| 2300 | stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
| 2301 | } |
| 2302 | |
| 2303 | void stat_exit(void) |
| 2304 | { |
| 2305 | /* |
| 2306 | * When we have the mutex, we know out-of-band access to it |
| 2307 | * have ended. |
| 2308 | */ |
| 2309 | fio_mutex_down(stat_mutex); |
| 2310 | fio_mutex_remove(stat_mutex); |
| 2311 | } |
| 2312 | |
| 2313 | /* |
| 2314 | * Called from signal handler. Wake up status thread. |
| 2315 | */ |
| 2316 | void show_running_run_stats(void) |
| 2317 | { |
| 2318 | helper_do_stat(); |
| 2319 | } |
| 2320 | |
| 2321 | uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u) |
| 2322 | { |
| 2323 | /* Ignore io_u's which span multiple blocks--they will just get |
| 2324 | * inaccurate counts. */ |
| 2325 | int idx = (io_u->offset - io_u->file->file_offset) |
| 2326 | / td->o.bs[DDIR_TRIM]; |
| 2327 | uint32_t *info = &td->ts.block_infos[idx]; |
| 2328 | assert(idx < td->ts.nr_block_infos); |
| 2329 | return info; |
| 2330 | } |