| 1 | #include <unistd.h> |
| 2 | |
| 3 | #include "fio.h" |
| 4 | #include "target.h" |
| 5 | #include "smalloc.h" |
| 6 | #include "stat.h" |
| 7 | |
| 8 | void lat_fatal(struct thread_data *td, unsigned long long tnsec, |
| 9 | unsigned long long max_nsec) |
| 10 | { |
| 11 | if (!td->error) |
| 12 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
| 13 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
| 14 | } |
| 15 | |
| 16 | static void lat_ios_note(struct thread_data *td) |
| 17 | { |
| 18 | int i; |
| 19 | |
| 20 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 21 | td->latency_ios[i] = td->io_blocks[i]; |
| 22 | } |
| 23 | |
| 24 | static void lat_new_cycle(struct thread_data *td) |
| 25 | { |
| 26 | fio_gettime(&td->latency_ts, NULL); |
| 27 | lat_ios_note(td); |
| 28 | td->latency_failed = 0; |
| 29 | } |
| 30 | |
| 31 | /* |
| 32 | * We had an IO outside the latency target. Reduce the queue depth. If we |
| 33 | * are at QD=1, then it's time to give up. |
| 34 | */ |
| 35 | static bool __lat_target_failed(struct thread_data *td) |
| 36 | { |
| 37 | if (td->latency_qd == 1) |
| 38 | return true; |
| 39 | |
| 40 | td->latency_qd_high = td->latency_qd; |
| 41 | |
| 42 | if (td->latency_qd == td->latency_qd_low) |
| 43 | td->latency_qd_low--; |
| 44 | |
| 45 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
| 46 | |
| 47 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); |
| 48 | |
| 49 | /* |
| 50 | * When we ramp QD down, quiesce existing IO to prevent |
| 51 | * a storm of ramp downs due to pending higher depth. |
| 52 | */ |
| 53 | io_u_quiesce(td); |
| 54 | lat_new_cycle(td); |
| 55 | return false; |
| 56 | } |
| 57 | |
| 58 | bool lat_target_failed(struct thread_data *td) |
| 59 | { |
| 60 | if (td->o.latency_percentile.u.f == 100.0) |
| 61 | return __lat_target_failed(td); |
| 62 | |
| 63 | td->latency_failed++; |
| 64 | return false; |
| 65 | } |
| 66 | |
| 67 | static void lat_step_init(struct thread_data *td) |
| 68 | { |
| 69 | struct thread_options *o = &td->o; |
| 70 | |
| 71 | fio_gettime(&td->latency_ts, NULL); |
| 72 | td->latency_state = IOD_STATE_PROBE_RAMP; |
| 73 | td->latency_step = 0; |
| 74 | td->latency_qd = td->o.iodepth; |
| 75 | dprint(FD_RATE, "Stepped: %d-%d/%d,%d/%d\n", o->lat_step_low, |
| 76 | o->lat_step_high, o->lat_step_inc, |
| 77 | o->lat_step_ramp, o->lat_step_run); |
| 78 | } |
| 79 | |
| 80 | void lat_target_init(struct thread_data *td) |
| 81 | { |
| 82 | td->latency_end_run = 0; |
| 83 | |
| 84 | if (td->o.latency_target) { |
| 85 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); |
| 86 | fio_gettime(&td->latency_ts, NULL); |
| 87 | td->latency_qd = 1; |
| 88 | td->latency_qd_high = td->o.iodepth; |
| 89 | td->latency_qd_low = 1; |
| 90 | lat_ios_note(td); |
| 91 | } else if (td->o.iodepth_mode == IOD_STEPPED) |
| 92 | lat_step_init(td); |
| 93 | else |
| 94 | td->latency_qd = td->o.iodepth; |
| 95 | } |
| 96 | |
| 97 | void lat_target_reset(struct thread_data *td) |
| 98 | { |
| 99 | if (td->o.latency_target && !td->latency_end_run) |
| 100 | lat_target_init(td); |
| 101 | } |
| 102 | |
| 103 | static void lat_target_success(struct thread_data *td) |
| 104 | { |
| 105 | const unsigned int qd = td->latency_qd; |
| 106 | struct thread_options *o = &td->o; |
| 107 | |
| 108 | td->latency_qd_low = td->latency_qd; |
| 109 | |
| 110 | /* |
| 111 | * If we haven't failed yet, we double up to a failing value instead |
| 112 | * of bisecting from highest possible queue depth. If we have set |
| 113 | * a limit other than td->o.iodepth, bisect between that. |
| 114 | */ |
| 115 | if (td->latency_qd_high != o->iodepth) |
| 116 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
| 117 | else |
| 118 | td->latency_qd *= 2; |
| 119 | |
| 120 | if (td->latency_qd > o->iodepth) |
| 121 | td->latency_qd = o->iodepth; |
| 122 | |
| 123 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); |
| 124 | |
| 125 | /* |
| 126 | * Same as last one, we are done. Let it run a latency cycle, so |
| 127 | * we get only the results from the targeted depth. |
| 128 | */ |
| 129 | if (td->latency_qd == qd) { |
| 130 | if (td->latency_end_run) { |
| 131 | dprint(FD_RATE, "We are done\n"); |
| 132 | td->done = 1; |
| 133 | } else { |
| 134 | dprint(FD_RATE, "Quiesce and final run\n"); |
| 135 | io_u_quiesce(td); |
| 136 | td->latency_end_run = 1; |
| 137 | reset_all_stats(td); |
| 138 | reset_io_stats(td); |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | lat_new_cycle(td); |
| 143 | } |
| 144 | |
| 145 | void __lat_target_check(struct thread_data *td) |
| 146 | { |
| 147 | uint64_t usec_window; |
| 148 | uint64_t ios; |
| 149 | double success_ios; |
| 150 | |
| 151 | usec_window = utime_since_now(&td->latency_ts); |
| 152 | if (usec_window < td->o.latency_window) |
| 153 | return; |
| 154 | |
| 155 | ios = ddir_rw_sum(td->io_blocks) - ddir_rw_sum(td->latency_ios); |
| 156 | success_ios = (double) (ios - td->latency_failed) / (double) ios; |
| 157 | success_ios *= 100.0; |
| 158 | |
| 159 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); |
| 160 | |
| 161 | if (success_ios >= td->o.latency_percentile.u.f) |
| 162 | lat_target_success(td); |
| 163 | else |
| 164 | __lat_target_failed(td); |
| 165 | } |
| 166 | |
| 167 | static void lat_clear_rate(struct thread_data *td) |
| 168 | { |
| 169 | int i; |
| 170 | |
| 171 | td->flags &= ~TD_F_CHECK_RATE; |
| 172 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 173 | td->o.rate_iops[i] = 0; |
| 174 | } |
| 175 | |
| 176 | /* |
| 177 | * Returns true if we're done stepping |
| 178 | */ |
| 179 | static bool lat_step_recalc(struct thread_data *td) |
| 180 | { |
| 181 | struct thread_options *o = &td->o; |
| 182 | unsigned int cur, perc; |
| 183 | |
| 184 | cur = td->latency_step * o->lat_step_inc; |
| 185 | if (cur >= o->lat_step_high) |
| 186 | return true; |
| 187 | |
| 188 | perc = (td->latency_step + 1) * o->lat_step_inc; |
| 189 | if (perc < 100) { |
| 190 | int i; |
| 191 | |
| 192 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 193 | unsigned int this_iops; |
| 194 | |
| 195 | this_iops = (perc * td->latency_iops[i]) / 100; |
| 196 | td->o.rate_iops[i] = this_iops; |
| 197 | } |
| 198 | setup_rate(td); |
| 199 | td->flags |= TD_F_CHECK_RATE; |
| 200 | td->latency_qd = td->o.iodepth * 100 / o->lat_step_high; |
| 201 | } else { |
| 202 | td->latency_qd = td->o.iodepth * perc / o->lat_step_high; |
| 203 | lat_clear_rate(td); |
| 204 | } |
| 205 | |
| 206 | dprint(FD_RATE, "Stepped: step=%d, perc=%d, qd=%d\n", td->latency_step, |
| 207 | perc, td->latency_qd); |
| 208 | return false; |
| 209 | } |
| 210 | |
| 211 | static void lat_step_reset(struct thread_data *td) |
| 212 | { |
| 213 | struct thread_stat *ts = &td->ts; |
| 214 | struct io_stat *ios = &ts->clat_stat[DDIR_RWDIR_CNT]; |
| 215 | |
| 216 | ios->max_val = ios->min_val = ios->samples = 0; |
| 217 | ios->mean.u.f = ios->S.u.f = 0; |
| 218 | |
| 219 | lat_clear_rate(td); |
| 220 | reset_all_stats(td); |
| 221 | reset_io_stats(td); |
| 222 | } |
| 223 | |
| 224 | static uint64_t lat_iops_since(struct thread_data *td, uint64_t msec, |
| 225 | enum fio_ddir ddir) |
| 226 | { |
| 227 | if (msec) { |
| 228 | uint64_t ios; |
| 229 | |
| 230 | ios = td->io_blocks[ddir] - td->latency_ios[ddir]; |
| 231 | return (ios * 1000) / msec; |
| 232 | } |
| 233 | |
| 234 | return 0; |
| 235 | } |
| 236 | |
| 237 | static void lat_step_add_sample(struct thread_data *td, uint64_t msec) |
| 238 | { |
| 239 | struct thread_stat *ts = &td->ts; |
| 240 | unsigned long long min, max; |
| 241 | struct lat_step_stats *ls; |
| 242 | double mean[DDIR_RWDIR_CNT], dev; |
| 243 | int i; |
| 244 | |
| 245 | if (td->nr_lat_stats == ARRAY_SIZE(td->ts.step_stats)) { |
| 246 | log_err("fio: ts->step_stats too small, dropping entries\n"); |
| 247 | return; |
| 248 | } |
| 249 | |
| 250 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 251 | calc_lat(&ts->clat_stat[i], &min, &max, &mean[i], &dev); |
| 252 | |
| 253 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 254 | ls = &td->ts.step_stats[td->nr_lat_stats]; |
| 255 | |
| 256 | ls->iops[i] = lat_iops_since(td, msec, i); |
| 257 | ls->avg[i].u.f = mean[i]; |
| 258 | } |
| 259 | |
| 260 | td->nr_lat_stats++; |
| 261 | } |
| 262 | |
| 263 | bool __lat_ts_has_stats(struct thread_stat *ts, enum fio_ddir ddir) |
| 264 | { |
| 265 | int i; |
| 266 | |
| 267 | for (i = 0; i < ARRAY_SIZE(ts->step_stats); i++) { |
| 268 | struct lat_step_stats *ls = &ts->step_stats[i]; |
| 269 | |
| 270 | if (ls->iops[ddir]) |
| 271 | return true; |
| 272 | } |
| 273 | |
| 274 | return false; |
| 275 | } |
| 276 | |
| 277 | bool lat_ts_has_stats(struct thread_stat *ts) |
| 278 | { |
| 279 | int i; |
| 280 | |
| 281 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 282 | if (__lat_ts_has_stats(ts, i)) |
| 283 | return true; |
| 284 | |
| 285 | return false; |
| 286 | } |
| 287 | |
| 288 | void lat_step_report(struct thread_stat *ts, struct buf_output *out) |
| 289 | { |
| 290 | int i, j; |
| 291 | |
| 292 | for (i = 0; i < ARRAY_SIZE(ts->step_stats); i++) { |
| 293 | struct lat_step_stats *ls = &ts->step_stats[i]; |
| 294 | |
| 295 | for (j = 0; j < DDIR_RWDIR_CNT; j++) { |
| 296 | if (!ls->iops[j]) |
| 297 | continue; |
| 298 | |
| 299 | if (!j) |
| 300 | __log_buf(out, " [%2d] ", i); |
| 301 | else |
| 302 | __log_buf(out, " "); |
| 303 | |
| 304 | __log_buf(out, "%5s: iops=%llu, lat=%.1f nsec\n", |
| 305 | io_ddir_name(j), |
| 306 | (unsigned long long) ls->iops[j], |
| 307 | ls->avg[j].u.f); |
| 308 | } |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | static void lat_next_state(struct thread_data *td, int new_state) |
| 313 | { |
| 314 | td->latency_state = new_state; |
| 315 | fio_gettime(&td->latency_ts, NULL); |
| 316 | } |
| 317 | |
| 318 | bool lat_step_check(struct thread_data *td) |
| 319 | { |
| 320 | struct thread_options *o = &td->o; |
| 321 | uint64_t msec; |
| 322 | |
| 323 | msec = mtime_since_now(&td->latency_ts); |
| 324 | |
| 325 | switch (td->latency_state) { |
| 326 | case IOD_STATE_PROBE_RAMP: |
| 327 | if (msec < o->lat_step_ramp) |
| 328 | break; |
| 329 | |
| 330 | lat_step_reset(td); |
| 331 | lat_ios_note(td); |
| 332 | |
| 333 | lat_next_state(td, IOD_STATE_PROBE_RUN); |
| 334 | break; |
| 335 | case IOD_STATE_PROBE_RUN: { |
| 336 | int i; |
| 337 | |
| 338 | if (msec < o->lat_step_run) |
| 339 | break; |
| 340 | |
| 341 | io_u_quiesce(td); |
| 342 | |
| 343 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
| 344 | td->latency_iops[i] = lat_iops_since(td, msec, i); |
| 345 | |
| 346 | lat_step_reset(td); |
| 347 | lat_step_recalc(td); |
| 348 | |
| 349 | io_u_quiesce(td); |
| 350 | lat_next_state(td, IOD_STATE_RAMP); |
| 351 | break; |
| 352 | } |
| 353 | case IOD_STATE_RAMP: |
| 354 | if (msec < o->lat_step_ramp) |
| 355 | break; |
| 356 | |
| 357 | lat_ios_note(td); |
| 358 | lat_next_state(td, IOD_STATE_RUN); |
| 359 | break; |
| 360 | case IOD_STATE_RUN: |
| 361 | if (msec < o->lat_step_run) |
| 362 | break; |
| 363 | |
| 364 | io_u_quiesce(td); |
| 365 | fio_gettime(&td->latency_ts, NULL); |
| 366 | td->latency_step++; |
| 367 | |
| 368 | lat_step_add_sample(td, msec); |
| 369 | lat_step_reset(td); |
| 370 | |
| 371 | if (!lat_step_recalc(td)) |
| 372 | break; |
| 373 | |
| 374 | td->done = 1; |
| 375 | lat_next_state(td, IOD_STATE_DONE); |
| 376 | break; |
| 377 | }; |
| 378 | |
| 379 | return td->latency_state == IOD_STATE_DONE; |
| 380 | } |