Merge branch 'eta_overflow' of https://github.com/sitsofe/fio
[fio.git] / eta.c
1 /*
2  * Status and ETA code
3  */
4 #include <unistd.h>
5 #include <fcntl.h>
6 #include <string.h>
7
8 #include "fio.h"
9 #include "lib/pow2.h"
10
11 static char __run_str[REAL_MAX_JOBS + 1];
12 static char run_str[__THREAD_RUNSTR_SZ(REAL_MAX_JOBS)];
13
14 static void update_condensed_str(char *rstr, char *run_str_condensed)
15 {
16         if (*rstr) {
17                 while (*rstr) {
18                         int nr = 1;
19
20                         *run_str_condensed++ = *rstr++;
21                         while (*(rstr - 1) == *rstr) {
22                                 rstr++;
23                                 nr++;
24                         }
25                         run_str_condensed += sprintf(run_str_condensed, "(%u),", nr);
26                 }
27                 run_str_condensed--;
28         }
29         *run_str_condensed = '\0';
30 }
31
32 /*
33  * Sets the status of the 'td' in the printed status map.
34  */
35 static void check_str_update(struct thread_data *td)
36 {
37         char c = __run_str[td->thread_number - 1];
38
39         switch (td->runstate) {
40         case TD_REAPED:
41                 if (td->error)
42                         c = 'X';
43                 else if (td->sig)
44                         c = 'K';
45                 else
46                         c = '_';
47                 break;
48         case TD_EXITED:
49                 c = 'E';
50                 break;
51         case TD_RAMP:
52                 c = '/';
53                 break;
54         case TD_RUNNING:
55                 if (td_rw(td)) {
56                         if (td_random(td)) {
57                                 if (td->o.rwmix[DDIR_READ] == 100)
58                                         c = 'r';
59                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
60                                         c = 'w';
61                                 else
62                                         c = 'm';
63                         } else {
64                                 if (td->o.rwmix[DDIR_READ] == 100)
65                                         c = 'R';
66                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
67                                         c = 'W';
68                                 else
69                                         c = 'M';
70                         }
71                 } else if (td_read(td)) {
72                         if (td_random(td))
73                                 c = 'r';
74                         else
75                                 c = 'R';
76                 } else if (td_write(td)) {
77                         if (td_random(td))
78                                 c = 'w';
79                         else
80                                 c = 'W';
81                 } else {
82                         if (td_random(td))
83                                 c = 'd';
84                         else
85                                 c = 'D';
86                 }
87                 break;
88         case TD_PRE_READING:
89                 c = 'p';
90                 break;
91         case TD_VERIFYING:
92                 c = 'V';
93                 break;
94         case TD_FSYNCING:
95                 c = 'F';
96                 break;
97         case TD_FINISHING:
98                 c = 'f';
99                 break;
100         case TD_CREATED:
101                 c = 'C';
102                 break;
103         case TD_INITIALIZED:
104         case TD_SETTING_UP:
105                 c = 'I';
106                 break;
107         case TD_NOT_CREATED:
108                 c = 'P';
109                 break;
110         default:
111                 log_err("state %d\n", td->runstate);
112         }
113
114         __run_str[td->thread_number - 1] = c;
115         update_condensed_str(__run_str, run_str);
116 }
117
118 /*
119  * Convert seconds to a printable string.
120  */
121 void eta_to_str(char *str, unsigned long eta_sec)
122 {
123         unsigned int d, h, m, s;
124         int disp_hour = 0;
125
126         if (eta_sec == -1) {
127                 sprintf(str, "--");
128                 return;
129         }
130
131         s = eta_sec % 60;
132         eta_sec /= 60;
133         m = eta_sec % 60;
134         eta_sec /= 60;
135         h = eta_sec % 24;
136         eta_sec /= 24;
137         d = eta_sec;
138
139         if (d) {
140                 disp_hour = 1;
141                 str += sprintf(str, "%02ud:", d);
142         }
143
144         if (h || disp_hour)
145                 str += sprintf(str, "%02uh:", h);
146
147         str += sprintf(str, "%02um:", m);
148         str += sprintf(str, "%02us", s);
149 }
150
151 /*
152  * Best effort calculation of the estimated pending runtime of a job.
153  */
154 static unsigned long thread_eta(struct thread_data *td)
155 {
156         unsigned long long bytes_total, bytes_done;
157         unsigned long eta_sec = 0;
158         unsigned long elapsed;
159         uint64_t timeout;
160
161         elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
162         timeout = td->o.timeout / 1000000UL;
163
164         bytes_total = td->total_io_size;
165
166         if (td->flags & TD_F_NO_PROGRESS)
167                 return -1;
168
169         if (td->o.fill_device && td->o.size  == -1ULL) {
170                 if (!td->fill_device_size || td->fill_device_size == -1ULL)
171                         return 0;
172
173                 bytes_total = td->fill_device_size;
174         }
175
176         if (td->o.zone_size && td->o.zone_skip && bytes_total) {
177                 unsigned int nr_zones;
178                 uint64_t zone_bytes;
179
180                 zone_bytes = bytes_total + td->o.zone_size + td->o.zone_skip;
181                 nr_zones = (zone_bytes - 1) / (td->o.zone_size + td->o.zone_skip);
182                 bytes_total -= nr_zones * td->o.zone_skip;
183         }
184
185         /*
186          * if writing and verifying afterwards, bytes_total will be twice the
187          * size. In a mixed workload, verify phase will be the size of the
188          * first stage writes.
189          */
190         if (td->o.do_verify && td->o.verify && td_write(td)) {
191                 if (td_rw(td)) {
192                         unsigned int perc = 50;
193
194                         if (td->o.rwmix[DDIR_WRITE])
195                                 perc = td->o.rwmix[DDIR_WRITE];
196
197                         bytes_total += (bytes_total * perc) / 100;
198                 } else
199                         bytes_total <<= 1;
200         }
201
202         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
203                 double perc, perc_t;
204
205                 bytes_done = ddir_rw_sum(td->io_bytes);
206
207                 if (bytes_total) {
208                         perc = (double) bytes_done / (double) bytes_total;
209                         if (perc > 1.0)
210                                 perc = 1.0;
211                 } else
212                         perc = 0.0;
213
214                 if (td->o.time_based) {
215                         if (timeout) {
216                                 perc_t = (double) elapsed / (double) timeout;
217                                 if (perc_t < perc)
218                                         perc = perc_t;
219                         } else {
220                                 /*
221                                  * Will never hit, we can't have time_based
222                                  * without a timeout set.
223                                  */
224                                 perc = 0.0;
225                         }
226                 }
227
228                 if (perc == 0.0) {
229                         eta_sec = timeout;
230                 } else {
231                         eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
232                 }
233
234                 if (td->o.timeout &&
235                     eta_sec > (timeout + done_secs - elapsed))
236                         eta_sec = timeout + done_secs - elapsed;
237         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
238                         || td->runstate == TD_INITIALIZED
239                         || td->runstate == TD_SETTING_UP
240                         || td->runstate == TD_RAMP
241                         || td->runstate == TD_PRE_READING) {
242                 int64_t t_eta = 0, r_eta = 0;
243                 unsigned long long rate_bytes;
244
245                 /*
246                  * We can only guess - assume it'll run the full timeout
247                  * if given, otherwise assume it'll run at the specified rate.
248                  */
249                 if (td->o.timeout) {
250                         uint64_t __timeout = td->o.timeout;
251                         uint64_t start_delay = td->o.start_delay;
252                         uint64_t ramp_time = td->o.ramp_time;
253
254                         t_eta = __timeout + start_delay;
255                         if (!td->ramp_time_over) {
256                                 t_eta += ramp_time;
257                         }
258                         t_eta /= 1000000ULL;
259
260                         if ((td->runstate == TD_RAMP) && in_ramp_time(td)) {
261                                 unsigned long ramp_left;
262
263                                 ramp_left = mtime_since_now(&td->epoch);
264                                 ramp_left = (ramp_left + 999) / 1000;
265                                 if (ramp_left <= t_eta)
266                                         t_eta -= ramp_left;
267                         }
268                 }
269                 rate_bytes = 0;
270                 if (td_read(td))
271                         rate_bytes  = td->o.rate[DDIR_READ];
272                 if (td_write(td))
273                         rate_bytes += td->o.rate[DDIR_WRITE];
274                 if (td_trim(td))
275                         rate_bytes += td->o.rate[DDIR_TRIM];
276
277                 if (rate_bytes) {
278                         r_eta = bytes_total / rate_bytes;
279                         r_eta += (td->o.start_delay / 1000000ULL);
280                 }
281
282                 if (r_eta && t_eta)
283                         eta_sec = min(r_eta, t_eta);
284                 else if (r_eta)
285                         eta_sec = r_eta;
286                 else if (t_eta)
287                         eta_sec = t_eta;
288                 else
289                         eta_sec = 0;
290         } else {
291                 /*
292                  * thread is already done or waiting for fsync
293                  */
294                 eta_sec = 0;
295         }
296
297         return eta_sec;
298 }
299
300 static void calc_rate(int unified_rw_rep, unsigned long mtime,
301                       unsigned long long *io_bytes,
302                       unsigned long long *prev_io_bytes, uint64_t *rate)
303 {
304         int i;
305
306         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
307                 unsigned long long diff, this_rate;
308
309                 diff = io_bytes[i] - prev_io_bytes[i];
310                 if (mtime)
311                         this_rate = ((1000 * diff) / mtime) / 1024; /* KiB/s */
312                 else
313                         this_rate = 0;
314
315                 if (unified_rw_rep) {
316                         rate[i] = 0;
317                         rate[0] += this_rate;
318                 } else
319                         rate[i] = this_rate;
320
321                 prev_io_bytes[i] = io_bytes[i];
322         }
323 }
324
325 static void calc_iops(int unified_rw_rep, unsigned long mtime,
326                       unsigned long long *io_iops,
327                       unsigned long long *prev_io_iops, unsigned int *iops)
328 {
329         int i;
330
331         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
332                 unsigned long long diff, this_iops;
333
334                 diff = io_iops[i] - prev_io_iops[i];
335                 if (mtime)
336                         this_iops = (diff * 1000) / mtime;
337                 else
338                         this_iops = 0;
339
340                 if (unified_rw_rep) {
341                         iops[i] = 0;
342                         iops[0] += this_iops;
343                 } else
344                         iops[i] = this_iops;
345
346                 prev_io_iops[i] = io_iops[i];
347         }
348 }
349
350 /*
351  * Allow a little slack - if we're within 95% of the time, allow ETA.
352  */
353 bool eta_time_within_slack(unsigned int time)
354 {
355         return time > ((eta_interval_msec * 95) / 100);
356 }
357
358 /*
359  * Print status of the jobs we know about. This includes rate estimates,
360  * ETA, thread state, etc.
361  */
362 bool calc_thread_status(struct jobs_eta *je, int force)
363 {
364         struct thread_data *td;
365         int i, unified_rw_rep;
366         uint64_t rate_time, disp_time, bw_avg_time, *eta_secs;
367         unsigned long long io_bytes[DDIR_RWDIR_CNT];
368         unsigned long long io_iops[DDIR_RWDIR_CNT];
369         struct timespec now;
370
371         static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
372         static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
373         static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
374         static struct timespec rate_prev_time, disp_prev_time;
375
376         if (!force) {
377                 if (!(output_format & FIO_OUTPUT_NORMAL) &&
378                     f_out == stdout)
379                         return false;
380                 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
381                         return false;
382
383                 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
384                         return false;
385         }
386
387         if (!ddir_rw_sum(rate_io_bytes))
388                 fill_start_time(&rate_prev_time);
389         if (!ddir_rw_sum(disp_io_bytes))
390                 fill_start_time(&disp_prev_time);
391
392         eta_secs = malloc(thread_number * sizeof(uint64_t));
393         memset(eta_secs, 0, thread_number * sizeof(uint64_t));
394
395         je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
396
397         io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
398         io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
399         bw_avg_time = ULONG_MAX;
400         unified_rw_rep = 0;
401         for_each_td(td, i) {
402                 unified_rw_rep += td->o.unified_rw_rep;
403                 if (is_power_of_2(td->o.kb_base))
404                         je->is_pow2 = 1;
405                 je->unit_base = td->o.unit_base;
406                 if (td->o.bw_avg_time < bw_avg_time)
407                         bw_avg_time = td->o.bw_avg_time;
408                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
409                     || td->runstate == TD_FSYNCING
410                     || td->runstate == TD_PRE_READING
411                     || td->runstate == TD_FINISHING) {
412                         je->nr_running++;
413                         if (td_read(td)) {
414                                 je->t_rate[0] += td->o.rate[DDIR_READ];
415                                 je->t_iops[0] += td->o.rate_iops[DDIR_READ];
416                                 je->m_rate[0] += td->o.ratemin[DDIR_READ];
417                                 je->m_iops[0] += td->o.rate_iops_min[DDIR_READ];
418                         }
419                         if (td_write(td)) {
420                                 je->t_rate[1] += td->o.rate[DDIR_WRITE];
421                                 je->t_iops[1] += td->o.rate_iops[DDIR_WRITE];
422                                 je->m_rate[1] += td->o.ratemin[DDIR_WRITE];
423                                 je->m_iops[1] += td->o.rate_iops_min[DDIR_WRITE];
424                         }
425                         if (td_trim(td)) {
426                                 je->t_rate[2] += td->o.rate[DDIR_TRIM];
427                                 je->t_iops[2] += td->o.rate_iops[DDIR_TRIM];
428                                 je->m_rate[2] += td->o.ratemin[DDIR_TRIM];
429                                 je->m_iops[2] += td->o.rate_iops_min[DDIR_TRIM];
430                         }
431
432                         je->files_open += td->nr_open_files;
433                 } else if (td->runstate == TD_RAMP) {
434                         je->nr_running++;
435                         je->nr_ramp++;
436                 } else if (td->runstate == TD_SETTING_UP)
437                         je->nr_setting_up++;
438                 else if (td->runstate < TD_RUNNING)
439                         je->nr_pending++;
440
441                 if (je->elapsed_sec >= 3)
442                         eta_secs[i] = thread_eta(td);
443                 else
444                         eta_secs[i] = INT_MAX;
445
446                 check_str_update(td);
447
448                 if (td->runstate > TD_SETTING_UP) {
449                         int ddir;
450
451                         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
452                                 if (unified_rw_rep) {
453                                         io_bytes[0] += td->io_bytes[ddir];
454                                         io_iops[0] += td->io_blocks[ddir];
455                                 } else {
456                                         io_bytes[ddir] += td->io_bytes[ddir];
457                                         io_iops[ddir] += td->io_blocks[ddir];
458                                 }
459                         }
460                 }
461         }
462
463         if (exitall_on_terminate) {
464                 je->eta_sec = INT_MAX;
465                 for_each_td(td, i) {
466                         if (eta_secs[i] < je->eta_sec)
467                                 je->eta_sec = eta_secs[i];
468                 }
469         } else {
470                 unsigned long eta_stone = 0;
471
472                 je->eta_sec = 0;
473                 for_each_td(td, i) {
474                         if ((td->runstate == TD_NOT_CREATED) && td->o.stonewall)
475                                 eta_stone += eta_secs[i];
476                         else {
477                                 if (eta_secs[i] > je->eta_sec)
478                                         je->eta_sec = eta_secs[i];
479                         }
480                 }
481                 je->eta_sec += eta_stone;
482         }
483
484         free(eta_secs);
485
486         fio_gettime(&now, NULL);
487         rate_time = mtime_since(&rate_prev_time, &now);
488
489         if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
490                 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
491                                 je->rate);
492                 memcpy(&rate_prev_time, &now, sizeof(now));
493                 add_agg_sample(sample_val(je->rate[DDIR_READ]), DDIR_READ, 0);
494                 add_agg_sample(sample_val(je->rate[DDIR_WRITE]), DDIR_WRITE, 0);
495                 add_agg_sample(sample_val(je->rate[DDIR_TRIM]), DDIR_TRIM, 0);
496         }
497
498         disp_time = mtime_since(&disp_prev_time, &now);
499
500         if (!force && !eta_time_within_slack(disp_time))
501                 return false;
502
503         calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
504         calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
505
506         memcpy(&disp_prev_time, &now, sizeof(now));
507
508         if (!force && !je->nr_running && !je->nr_pending)
509                 return false;
510
511         je->nr_threads = thread_number;
512         update_condensed_str(__run_str, run_str);
513         memcpy(je->run_str, run_str, strlen(run_str));
514         return true;
515 }
516
517 void display_thread_status(struct jobs_eta *je)
518 {
519         static struct timespec disp_eta_new_line;
520         static int eta_new_line_init, eta_new_line_pending;
521         static int linelen_last;
522         static int eta_good;
523         char output[REAL_MAX_JOBS + 512], *p = output;
524         char eta_str[128];
525         double perc = 0.0;
526
527         if (je->eta_sec != INT_MAX && je->elapsed_sec) {
528                 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
529                 eta_to_str(eta_str, je->eta_sec);
530         }
531
532         if (eta_new_line_pending) {
533                 eta_new_line_pending = 0;
534                 p += sprintf(p, "\n");
535         }
536
537         p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
538
539         /* rate limits, if any */
540         if (je->m_rate[0] || je->m_rate[1] || je->m_rate[2] ||
541             je->t_rate[0] || je->t_rate[1] || je->t_rate[2]) {
542                 char *tr, *mr;
543
544                 mr = num2str(je->m_rate[0] + je->m_rate[1] + je->m_rate[2],
545                                 je->sig_figs, 0, je->is_pow2, N2S_BYTEPERSEC);
546                 tr = num2str(je->t_rate[0] + je->t_rate[1] + je->t_rate[2],
547                                 je->sig_figs, 0, je->is_pow2, N2S_BYTEPERSEC);
548
549                 p += sprintf(p, ", %s-%s", mr, tr);
550                 free(tr);
551                 free(mr);
552         } else if (je->m_iops[0] || je->m_iops[1] || je->m_iops[2] ||
553                    je->t_iops[0] || je->t_iops[1] || je->t_iops[2]) {
554                 p += sprintf(p, ", %d-%d IOPS",
555                                         je->m_iops[0] + je->m_iops[1] + je->m_iops[2],
556                                         je->t_iops[0] + je->t_iops[1] + je->t_iops[2]);
557         }
558
559         /* current run string, % done, bandwidth, iops, eta */
560         if (je->eta_sec != INT_MAX && je->nr_running) {
561                 char perc_str[32];
562                 char *iops_str[DDIR_RWDIR_CNT];
563                 char *rate_str[DDIR_RWDIR_CNT];
564                 size_t left;
565                 int l;
566                 int ddir;
567
568                 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running ||
569                     je->eta_sec == -1)
570                         strcpy(perc_str, "-.-%");
571                 else {
572                         double mult = 100.0;
573
574                         if (je->nr_setting_up && je->nr_running)
575                                 mult *= (1.0 - (double) je->nr_setting_up / (double) je->nr_running);
576
577                         eta_good = 1;
578                         perc *= mult;
579                         sprintf(perc_str, "%3.1f%%", perc);
580                 }
581
582                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
583                         rate_str[ddir] = num2str(je->rate[ddir], 4,
584                                                 1024, je->is_pow2, je->unit_base);
585                         iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0, N2S_NONE);
586                 }
587
588                 left = sizeof(output) - (p - output) - 2;
589
590                 if (je->rate[DDIR_TRIM] || je->iops[DDIR_TRIM])
591                         l = snprintf(p, left,
592                                 ": [%s][%s][r=%s,w=%s,t=%s][r=%s,w=%s,t=%s IOPS][eta %s]",
593                                 je->run_str, perc_str, rate_str[DDIR_READ],
594                                 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
595                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
596                                 iops_str[DDIR_TRIM], eta_str);
597                 else
598                         l = snprintf(p, left,
599                                 ": [%s][%s][r=%s,w=%s][r=%s,w=%s IOPS][eta %s]",
600                                 je->run_str, perc_str,
601                                 rate_str[DDIR_READ], rate_str[DDIR_WRITE],
602                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
603                                 eta_str);
604                 if (l > left)
605                         l = left;
606                 p += l;
607                 if (l >= 0 && l < linelen_last)
608                         p += sprintf(p, "%*s", linelen_last - l, "");
609                 linelen_last = l;
610
611                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
612                         free(rate_str[ddir]);
613                         free(iops_str[ddir]);
614                 }
615         }
616         p += sprintf(p, "\r");
617
618         printf("%s", output);
619
620         if (!eta_new_line_init) {
621                 fio_gettime(&disp_eta_new_line, NULL);
622                 eta_new_line_init = 1;
623         } else if (eta_new_line && mtime_since_now(&disp_eta_new_line) > eta_new_line) {
624                 fio_gettime(&disp_eta_new_line, NULL);
625                 eta_new_line_pending = 1;
626         }
627
628         fflush(stdout);
629 }
630
631 struct jobs_eta *get_jobs_eta(bool force, size_t *size)
632 {
633         struct jobs_eta *je;
634
635         if (!thread_number)
636                 return NULL;
637
638         *size = sizeof(*je) + THREAD_RUNSTR_SZ + 8;
639         je = malloc(*size);
640         if (!je)
641                 return NULL;
642         memset(je, 0, *size);
643
644         if (!calc_thread_status(je, force)) {
645                 free(je);
646                 return NULL;
647         }
648
649         *size = sizeof(*je) + strlen((char *) je->run_str) + 1;
650         return je;
651 }
652
653 void print_thread_status(void)
654 {
655         struct jobs_eta *je;
656         size_t size;
657
658         je = get_jobs_eta(false, &size);
659         if (je)
660                 display_thread_status(je);
661
662         free(je);
663 }
664
665 void print_status_init(int thr_number)
666 {
667         __run_str[thr_number] = 'P';
668         update_condensed_str(__run_str, run_str);
669 }