RDMA engine server mode flow fix ups.
[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                 eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
229
230                 if (td->o.timeout &&
231                     eta_sec > (timeout + done_secs - elapsed))
232                         eta_sec = timeout + done_secs - elapsed;
233         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
234                         || td->runstate == TD_INITIALIZED
235                         || td->runstate == TD_SETTING_UP
236                         || td->runstate == TD_RAMP
237                         || td->runstate == TD_PRE_READING) {
238                 int t_eta = 0, r_eta = 0;
239                 unsigned long long rate_bytes;
240
241                 /*
242                  * We can only guess - assume it'll run the full timeout
243                  * if given, otherwise assume it'll run at the specified rate.
244                  */
245                 if (td->o.timeout) {
246                         uint64_t __timeout = td->o.timeout;
247                         uint64_t start_delay = td->o.start_delay;
248                         uint64_t ramp_time = td->o.ramp_time;
249
250                         t_eta = __timeout + start_delay + ramp_time;
251                         t_eta /= 1000000ULL;
252
253                         if (in_ramp_time(td)) {
254                                 unsigned long ramp_left;
255
256                                 ramp_left = mtime_since_now(&td->epoch);
257                                 ramp_left = (ramp_left + 999) / 1000;
258                                 if (ramp_left <= t_eta)
259                                         t_eta -= ramp_left;
260                         }
261                 }
262                 rate_bytes = ddir_rw_sum(td->o.rate);
263                 if (rate_bytes) {
264                         r_eta = (bytes_total / 1024) / rate_bytes;
265                         r_eta += (td->o.start_delay / 1000000ULL);
266                 }
267
268                 if (r_eta && t_eta)
269                         eta_sec = min(r_eta, t_eta);
270                 else if (r_eta)
271                         eta_sec = r_eta;
272                 else if (t_eta)
273                         eta_sec = t_eta;
274                 else
275                         eta_sec = 0;
276         } else {
277                 /*
278                  * thread is already done or waiting for fsync
279                  */
280                 eta_sec = 0;
281         }
282
283         return eta_sec;
284 }
285
286 static void calc_rate(int unified_rw_rep, unsigned long mtime,
287                       unsigned long long *io_bytes,
288                       unsigned long long *prev_io_bytes, unsigned int *rate)
289 {
290         int i;
291
292         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
293                 unsigned long long diff;
294
295                 diff = io_bytes[i] - prev_io_bytes[i];
296                 if (unified_rw_rep) {
297                         rate[i] = 0;
298                         rate[0] += ((1000 * diff) / mtime) / 1024;
299                 } else
300                         rate[i] = ((1000 * diff) / mtime) / 1024;
301
302                 prev_io_bytes[i] = io_bytes[i];
303         }
304 }
305
306 static void calc_iops(int unified_rw_rep, unsigned long mtime,
307                       unsigned long long *io_iops,
308                       unsigned long long *prev_io_iops, unsigned int *iops)
309 {
310         int i;
311
312         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
313                 unsigned long long diff;
314
315                 diff = io_iops[i] - prev_io_iops[i];
316                 if (unified_rw_rep) {
317                         iops[i] = 0;
318                         iops[0] += (diff * 1000) / mtime;
319                 } else
320                         iops[i] = (diff * 1000) / mtime;
321
322                 prev_io_iops[i] = io_iops[i];
323         }
324 }
325
326 /*
327  * Print status of the jobs we know about. This includes rate estimates,
328  * ETA, thread state, etc.
329  */
330 int calc_thread_status(struct jobs_eta *je, int force)
331 {
332         struct thread_data *td;
333         int i, unified_rw_rep;
334         unsigned long rate_time, disp_time, bw_avg_time, *eta_secs;
335         unsigned long long io_bytes[DDIR_RWDIR_CNT];
336         unsigned long long io_iops[DDIR_RWDIR_CNT];
337         struct timeval now;
338
339         static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
340         static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
341         static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
342         static struct timeval rate_prev_time, disp_prev_time;
343
344         if (!force) {
345                 if (output_format != FIO_OUTPUT_NORMAL &&
346                     f_out == stdout)
347                         return 0;
348                 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
349                         return 0;
350
351                 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
352                         return 0;
353         }
354
355         if (!ddir_rw_sum(rate_io_bytes))
356                 fill_start_time(&rate_prev_time);
357         if (!ddir_rw_sum(disp_io_bytes))
358                 fill_start_time(&disp_prev_time);
359
360         eta_secs = malloc(thread_number * sizeof(unsigned long));
361         memset(eta_secs, 0, thread_number * sizeof(unsigned long));
362
363         je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
364
365         io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
366         io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
367         bw_avg_time = ULONG_MAX;
368         unified_rw_rep = 0;
369         for_each_td(td, i) {
370                 unified_rw_rep += td->o.unified_rw_rep;
371                 if (is_power_of_2(td->o.kb_base))
372                         je->is_pow2 = 1;
373                 je->unit_base = td->o.unit_base;
374                 if (td->o.bw_avg_time < bw_avg_time)
375                         bw_avg_time = td->o.bw_avg_time;
376                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
377                     || td->runstate == TD_FSYNCING
378                     || td->runstate == TD_PRE_READING
379                     || td->runstate == TD_FINISHING) {
380                         je->nr_running++;
381                         if (td_read(td)) {
382                                 je->t_rate[0] += td->o.rate[DDIR_READ];
383                                 je->t_iops[0] += td->o.rate_iops[DDIR_READ];
384                                 je->m_rate[0] += td->o.ratemin[DDIR_READ];
385                                 je->m_iops[0] += td->o.rate_iops_min[DDIR_READ];
386                         }
387                         if (td_write(td)) {
388                                 je->t_rate[1] += td->o.rate[DDIR_WRITE];
389                                 je->t_iops[1] += td->o.rate_iops[DDIR_WRITE];
390                                 je->m_rate[1] += td->o.ratemin[DDIR_WRITE];
391                                 je->m_iops[1] += td->o.rate_iops_min[DDIR_WRITE];
392                         }
393                         if (td_trim(td)) {
394                                 je->t_rate[2] += td->o.rate[DDIR_TRIM];
395                                 je->t_iops[2] += td->o.rate_iops[DDIR_TRIM];
396                                 je->m_rate[2] += td->o.ratemin[DDIR_TRIM];
397                                 je->m_iops[2] += td->o.rate_iops_min[DDIR_TRIM];
398                         }
399
400                         je->files_open += td->nr_open_files;
401                 } else if (td->runstate == TD_RAMP) {
402                         je->nr_running++;
403                         je->nr_ramp++;
404                 } else if (td->runstate == TD_SETTING_UP)
405                         je->nr_setting_up++;
406                 else if (td->runstate < TD_RUNNING)
407                         je->nr_pending++;
408
409                 if (je->elapsed_sec >= 3)
410                         eta_secs[i] = thread_eta(td);
411                 else
412                         eta_secs[i] = INT_MAX;
413
414                 check_str_update(td);
415
416                 if (td->runstate > TD_SETTING_UP) {
417                         int ddir;
418
419                         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
420                                 if (unified_rw_rep) {
421                                         io_bytes[0] += td->io_bytes[ddir];
422                                         io_iops[0] += td->io_blocks[ddir];
423                                 } else {
424                                         io_bytes[ddir] += td->io_bytes[ddir];
425                                         io_iops[ddir] += td->io_blocks[ddir];
426                                 }
427                         }
428                 }
429         }
430
431         if (exitall_on_terminate)
432                 je->eta_sec = INT_MAX;
433         else
434                 je->eta_sec = 0;
435
436         for_each_td(td, i) {
437                 if (exitall_on_terminate) {
438                         if (eta_secs[i] < je->eta_sec)
439                                 je->eta_sec = eta_secs[i];
440                 } else {
441                         if (eta_secs[i] > je->eta_sec)
442                                 je->eta_sec = eta_secs[i];
443                 }
444         }
445
446         free(eta_secs);
447
448         fio_gettime(&now, NULL);
449         rate_time = mtime_since(&rate_prev_time, &now);
450
451         if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
452                 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
453                                 je->rate);
454                 memcpy(&rate_prev_time, &now, sizeof(now));
455                 add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
456                 add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
457                 add_agg_sample(je->rate[DDIR_TRIM], DDIR_TRIM, 0);
458         }
459
460         disp_time = mtime_since(&disp_prev_time, &now);
461
462         /*
463          * Allow a little slack, the target is to print it every 1000 msecs
464          */
465         if (!force && disp_time < 900)
466                 return 0;
467
468         calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
469         calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
470
471         memcpy(&disp_prev_time, &now, sizeof(now));
472
473         if (!force && !je->nr_running && !je->nr_pending)
474                 return 0;
475
476         je->nr_threads = thread_number;
477         update_condensed_str(__run_str, run_str);
478         memcpy(je->run_str, run_str, strlen(run_str));
479         return 1;
480 }
481
482 void display_thread_status(struct jobs_eta *je)
483 {
484         static struct timeval disp_eta_new_line;
485         static int eta_new_line_init, eta_new_line_pending;
486         static int linelen_last;
487         static int eta_good;
488         char output[REAL_MAX_JOBS + 512], *p = output;
489         char eta_str[128];
490         double perc = 0.0;
491
492         if (je->eta_sec != INT_MAX && je->elapsed_sec) {
493                 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
494                 eta_to_str(eta_str, je->eta_sec);
495         }
496
497         if (eta_new_line_pending) {
498                 eta_new_line_pending = 0;
499                 p += sprintf(p, "\n");
500         }
501
502         p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
503         if (je->m_rate[0] || je->m_rate[1] || je->t_rate[0] || je->t_rate[1]) {
504                 char *tr, *mr;
505
506                 mr = num2str(je->m_rate[0] + je->m_rate[1], 4, 0, je->is_pow2, 8);
507                 tr = num2str(je->t_rate[0] + je->t_rate[1], 4, 0, je->is_pow2, 8);
508                 p += sprintf(p, ", CR=%s/%s KB/s", tr, mr);
509                 free(tr);
510                 free(mr);
511         } else if (je->m_iops[0] || je->m_iops[1] || je->t_iops[0] || je->t_iops[1]) {
512                 p += sprintf(p, ", CR=%d/%d IOPS",
513                                         je->t_iops[0] + je->t_iops[1],
514                                         je->m_iops[0] + je->m_iops[1]);
515         }
516         if (je->eta_sec != INT_MAX && je->nr_running) {
517                 char perc_str[32];
518                 char *iops_str[DDIR_RWDIR_CNT];
519                 char *rate_str[DDIR_RWDIR_CNT];
520                 size_t left;
521                 int l;
522                 int ddir;
523
524                 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running ||
525                     je->eta_sec == -1)
526                         strcpy(perc_str, "-.-% done");
527                 else {
528                         double mult = 100.0;
529
530                         if (je->nr_setting_up && je->nr_running)
531                                 mult *= (1.0 - (double) je->nr_setting_up / (double) je->nr_running);
532
533                         eta_good = 1;
534                         perc *= mult;
535                         sprintf(perc_str, "%3.1f%% done", perc);
536                 }
537
538                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
539                         rate_str[ddir] = num2str(je->rate[ddir], 5,
540                                                 1024, je->is_pow2, je->unit_base);
541                         iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0, 0);
542                 }
543
544                 left = sizeof(output) - (p - output) - 1;
545
546                 l = snprintf(p, left, ": [%s] [%s] [%s/%s/%s /s] [%s/%s/%s iops] [eta %s]",
547                                 je->run_str, perc_str, rate_str[DDIR_READ],
548                                 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
549                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
550                                 iops_str[DDIR_TRIM], eta_str);
551                 p += l;
552                 if (l >= 0 && l < linelen_last)
553                         p += sprintf(p, "%*s", linelen_last - l, "");
554                 linelen_last = l;
555
556                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
557                         free(rate_str[ddir]);
558                         free(iops_str[ddir]);
559                 }
560         }
561         p += sprintf(p, "\r");
562
563         printf("%s", output);
564
565         if (!eta_new_line_init) {
566                 fio_gettime(&disp_eta_new_line, NULL);
567                 eta_new_line_init = 1;
568         } else if (eta_new_line && mtime_since_now(&disp_eta_new_line) > eta_new_line) {
569                 fio_gettime(&disp_eta_new_line, NULL);
570                 eta_new_line_pending = 1;
571         }
572
573         fflush(stdout);
574 }
575
576 struct jobs_eta *get_jobs_eta(int force, size_t *size)
577 {
578         struct jobs_eta *je;
579
580         if (!thread_number)
581                 return NULL;
582
583         *size = sizeof(*je) + THREAD_RUNSTR_SZ;
584         je = malloc(*size);
585         if (!je)
586                 return NULL;
587         memset(je, 0, *size);
588
589         if (!calc_thread_status(je, force)) {
590                 free(je);
591                 return NULL;
592         }
593
594         *size = sizeof(*je) + strlen((char *) je->run_str) + 1;
595         return je;
596 }
597
598 void print_thread_status(void)
599 {
600         struct jobs_eta *je;
601         size_t size;
602
603         je = get_jobs_eta(0, &size);
604         if (je)
605                 display_thread_status(je);
606
607         free(je);
608 }
609
610 void print_status_init(int thr_number)
611 {
612         __run_str[thr_number] = 'P';
613         update_condensed_str(__run_str, run_str);
614 }