Improve submission latency calculation
[fio.git] / fio.c
CommitLineData
ebac4655
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
aae22ca7 5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
ebac4655 6 *
8e9fe637
JA
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
ebac4655 10 * This program is free software; you can redistribute it and/or modify
8e9fe637
JA
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
ebac4655
JA
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
ebac4655
JA
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
ebac4655
JA
27#include <signal.h>
28#include <time.h>
dbe1125e 29#include <locale.h>
36167d82 30#include <assert.h>
ebac4655
JA
31#include <sys/stat.h>
32#include <sys/wait.h>
33#include <sys/ipc.h>
34#include <sys/shm.h>
ebac4655
JA
35#include <sys/mman.h>
36
37#include "fio.h"
38#include "os.h"
39
29d610e1
JA
40static unsigned long page_mask;
41#define ALIGN(buf) \
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
43
44int groupid = 0;
45int thread_number = 0;
ebac4655 46int shm_id = 0;
53cdc686 47int temp_stall_ts;
ebac4655 48
bbfd6b00 49static volatile int startup_sem;
6ce15a32 50static volatile int fio_abort;
437c9b71 51static int exit_value;
ebac4655 52
bb3884d8
JA
53struct io_log *agg_io_log[2];
54
ebac4655 55#define TERMINATE_ALL (-1)
75154845 56#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 57
6ce15a32
JA
58static inline void td_set_runstate(struct thread_data *td, int runstate)
59{
60 td->runstate = runstate;
61}
62
63static void terminate_threads(int group_id, int forced_kill)
ebac4655 64{
34572e28 65 struct thread_data *td;
ebac4655
JA
66 int i;
67
34572e28 68 for_each_td(td, i) {
ebac4655
JA
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70 td->terminate = 1;
71 td->start_delay = 0;
6ce15a32
JA
72 if (forced_kill)
73 td_set_runstate(td, TD_EXITED);
ebac4655
JA
74 }
75 }
76}
77
78static void sig_handler(int sig)
79{
80 switch (sig) {
81 case SIGALRM:
82 update_io_ticks();
83 disk_util_timer_arm();
84 print_thread_status();
85 break;
86 default:
6ce15a32 87 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 88 fflush(stdout);
6ce15a32 89 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
90 break;
91 }
92}
93
906c8d75
JA
94/*
95 * Check if we are above the minimum rate given.
96 */
ebac4655
JA
97static int check_min_rate(struct thread_data *td, struct timeval *now)
98{
99 unsigned long spent;
100 unsigned long rate;
101 int ddir = td->ddir;
102
103 /*
104 * allow a 2 second settle period in the beginning
105 */
106 if (mtime_since(&td->start, now) < 2000)
107 return 0;
108
109 /*
110 * if rate blocks is set, sample is running
111 */
112 if (td->rate_bytes) {
113 spent = mtime_since(&td->lastrate, now);
114 if (spent < td->ratecycle)
115 return 0;
116
117 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
118 if (rate < td->ratemin) {
1e97cce9 119 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
120 return 1;
121 }
122 }
123
124 td->rate_bytes = td->this_io_bytes[ddir];
125 memcpy(&td->lastrate, now, sizeof(*now));
126 return 0;
127}
128
129static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
130{
131 if (!td->timeout)
132 return 0;
133 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
134 return 1;
135
136 return 0;
137}
138
906c8d75
JA
139/*
140 * When job exits, we can cancel the in-flight IO if we are using async
141 * io. Attempt to do so.
142 */
ebac4655
JA
143static void cleanup_pending_aio(struct thread_data *td)
144{
ebac4655 145 struct list_head *entry, *n;
ebac4655
JA
146 struct io_u *io_u;
147 int r;
148
149 /*
150 * get immediately available events, if any
151 */
97601024 152 io_u_queued_complete(td, 0, NULL);
ebac4655
JA
153
154 /*
155 * now cancel remaining active events
156 */
2866c82d 157 if (td->io_ops->cancel) {
ebac4655
JA
158 list_for_each_safe(entry, n, &td->io_u_busylist) {
159 io_u = list_entry(entry, struct io_u, list);
160
2866c82d 161 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
162 if (!r)
163 put_io_u(td, io_u);
164 }
165 }
166
97601024
JA
167 if (td->cur_depth)
168 io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
169}
170
858a3d47
JA
171/*
172 * Helper to handle the final sync of a file. Works just like the normal
173 * io path, just does everything sync.
174 */
175static int fio_io_sync(struct thread_data *td, struct fio_file *f)
176{
177 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
178 int ret;
179
180 if (!io_u)
181 return 1;
182
183 io_u->ddir = DDIR_SYNC;
184 io_u->file = f;
185
186 if (td_io_prep(td, io_u)) {
187 put_io_u(td, io_u);
188 return 1;
189 }
190
755200a3 191requeue:
858a3d47 192 ret = td_io_queue(td, io_u);
36167d82 193 if (ret < 0) {
353a7e0e 194 td_verror(td, io_u->error);
858a3d47 195 put_io_u(td, io_u);
858a3d47 196 return 1;
36167d82 197 } else if (ret == FIO_Q_QUEUED) {
97601024 198 if (io_u_queued_complete(td, 1, NULL))
36167d82 199 return 1;
36167d82
JA
200 } else if (ret == FIO_Q_COMPLETED) {
201 if (io_u->error) {
202 td_verror(td, io_u->error);
203 return 1;
204 }
858a3d47 205
97601024 206 io_u_sync_complete(td, io_u, NULL);
755200a3
JA
207 } else if (ret == FIO_Q_BUSY) {
208 if (td_io_commit(td))
209 return 1;
210 goto requeue;
858a3d47
JA
211 }
212
213 return 0;
214}
215
906c8d75
JA
216/*
217 * The main verify engine. Runs over the writes we previusly submitted,
218 * reads the blocks back in, and checks the crc/md5 of the data.
219 */
1e97cce9 220static void do_verify(struct thread_data *td)
ebac4655 221{
53cdc686 222 struct fio_file *f;
36167d82 223 struct io_u *io_u;
3af6ef39 224 int ret, i, min_events;
e5b401d4
JA
225
226 /*
227 * sync io first and invalidate cache, to make sure we really
228 * read from disk.
229 */
230 for_each_file(td, f, i) {
858a3d47 231 fio_io_sync(td, f);
e5b401d4
JA
232 file_invalidate_cache(td, f);
233 }
ebac4655
JA
234
235 td_set_runstate(td, TD_VERIFYING);
236
36167d82
JA
237 io_u = NULL;
238 while (!td->terminate) {
ebac4655
JA
239 io_u = __get_io_u(td);
240 if (!io_u)
241 break;
242
36167d82 243 if (runtime_exceeded(td, &io_u->start_time))
02bcaa8c 244 break;
02bcaa8c 245
36167d82 246 if (get_next_verify(td, io_u))
ebac4655 247 break;
ebac4655 248
36167d82 249 if (td_io_prep(td, io_u))
53cdc686
JA
250 break;
251
36167d82
JA
252requeue:
253 ret = td_io_queue(td, io_u);
53cdc686 254
36167d82
JA
255 switch (ret) {
256 case FIO_Q_COMPLETED:
257 if (io_u->error)
22819ec2 258 ret = -io_u->error;
36167d82
JA
259 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
260 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 261
36167d82
JA
262 io_u->xfer_buflen = io_u->resid;
263 io_u->xfer_buf += bytes;
264 goto requeue;
265 }
99784632
JA
266 ret = io_u_sync_complete(td, io_u, verify_io_u);
267 if (ret)
36167d82 268 break;
36167d82
JA
269 continue;
270 case FIO_Q_QUEUED:
271 break;
755200a3
JA
272 case FIO_Q_BUSY:
273 requeue_io_u(td, &io_u);
274 ret = td_io_commit(td);
275 break;
36167d82
JA
276 default:
277 assert(ret < 0);
22819ec2 278 td_verror(td, -ret);
ebac4655
JA
279 break;
280 }
281
99784632 282 if (ret < 0 || td->error)
3af6ef39
JA
283 break;
284
ebac4655 285 /*
3af6ef39
JA
286 * if we can queue more, do so. but check if there are
287 * completed io_u's first.
ebac4655 288 */
97601024 289 min_events = 0;
755200a3 290 if (queue_full(td) || ret == FIO_Q_BUSY)
3af6ef39 291 min_events = 1;
3af6ef39
JA
292
293 /*
294 * Reap required number of io units, if any, and do the
295 * verification on them through the callback handler
296 */
97601024 297 if (io_u_queued_complete(td, min_events, verify_io_u))
ebac4655 298 break;
36167d82 299 }
ebac4655 300
36167d82
JA
301 if (io_u)
302 put_io_u(td, io_u);
ebac4655
JA
303
304 if (td->cur_depth)
305 cleanup_pending_aio(td);
306
307 td_set_runstate(td, TD_RUNNING);
308}
309
b990b5c0
JA
310/*
311 * Not really an io thread, all it does is burn CPU cycles in the specified
312 * manner.
313 */
314static void do_cpuio(struct thread_data *td)
315{
316 struct timeval e;
317 int split = 100 / td->cpuload;
318 int i = 0;
319
320 while (!td->terminate) {
02bcaa8c 321 fio_gettime(&e, NULL);
b990b5c0
JA
322
323 if (runtime_exceeded(td, &e))
324 break;
325
326 if (!(i % split))
327 __usec_sleep(10000);
328 else
329 usec_sleep(td, 10000);
330
331 i++;
332 }
333}
334
32cd46a0 335/*
906c8d75 336 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
337 * and reaps them, checking for rate and errors along the way.
338 */
ebac4655
JA
339static void do_io(struct thread_data *td)
340{
02bcaa8c 341 struct timeval s;
ebac4655 342 unsigned long usec;
84585003 343 int i, ret = 0;
ebac4655 344
5853e5a8
JA
345 td_set_runstate(td, TD_RUNNING);
346
3951414d 347 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
97601024
JA
348 struct timeval comp_time;
349 long bytes_done = 0;
84585003 350 int min_evts = 0;
ebac4655
JA
351 struct io_u *io_u;
352
353 if (td->terminate)
354 break;
355
3d7c391d 356 io_u = get_io_u(td);
ebac4655
JA
357 if (!io_u)
358 break;
359
360 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
361
362 if (runtime_exceeded(td, &s)) {
363 put_io_u(td, io_u);
364 break;
365 }
cec6b55d 366requeue:
45bee283 367 ret = td_io_queue(td, io_u);
36167d82
JA
368
369 switch (ret) {
370 case FIO_Q_COMPLETED:
371 if (io_u->error) {
372 ret = io_u->error;
373 break;
374 }
375 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
376 int bytes = io_u->xfer_buflen - io_u->resid;
377
cec6b55d 378 io_u->xfer_buflen = io_u->resid;
36167d82 379 io_u->xfer_buf += bytes;
cec6b55d 380 goto requeue;
cec6b55d 381 }
97601024
JA
382 fio_gettime(&comp_time, NULL);
383 bytes_done = io_u_sync_complete(td, io_u, NULL);
99784632
JA
384 if (bytes_done < 0)
385 ret = bytes_done;
36167d82
JA
386 break;
387 case FIO_Q_QUEUED:
7e77dd02
JA
388 /*
389 * if the engine doesn't have a commit hook,
390 * the io_u is really queued. if it does have such
391 * a hook, it has to call io_u_queued() itself.
392 */
393 if (td->io_ops->commit == NULL)
394 io_u_queued(td, io_u);
36167d82 395 break;
755200a3
JA
396 case FIO_Q_BUSY:
397 requeue_io_u(td, &io_u);
398 ret = td_io_commit(td);
399 break;
36167d82
JA
400 default:
401 assert(ret < 0);
402 put_io_u(td, io_u);
403 break;
ebac4655
JA
404 }
405
99784632 406 if (ret < 0 || td->error)
36167d82
JA
407 break;
408
97601024
JA
409 /*
410 * See if we need to complete some commands
411 */
755200a3 412 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 413 min_evts = 0;
755200a3 414 if (queue_full(td) || ret == FIO_Q_BUSY)
36167d82 415 min_evts = 1;
ebac4655 416
97601024
JA
417 fio_gettime(&comp_time, NULL);
418 bytes_done = io_u_queued_complete(td, min_evts, NULL);
419 if (bytes_done < 0)
36167d82 420 break;
ebac4655
JA
421 }
422
97601024
JA
423 if (!bytes_done)
424 continue;
425
ebac4655
JA
426 /*
427 * the rate is batched for now, it should work for batches
428 * of completions except the very first one which may look
429 * a little bursty
430 */
97601024 431 usec = utime_since(&s, &comp_time);
ebac4655 432
97601024 433 rate_throttle(td, usec, bytes_done, td->ddir);
ebac4655 434
97601024 435 if (check_min_rate(td, &comp_time)) {
98aa62d8 436 if (exitall_on_terminate)
6ce15a32 437 terminate_threads(td->groupid, 0);
fd841467 438 td_verror(td, ENODATA);
ebac4655
JA
439 break;
440 }
441
9c1f7434
JA
442 if (td->thinktime) {
443 unsigned long long b;
444
445 b = td->io_blocks[0] + td->io_blocks[1];
48097d5c
JA
446 if (!(b % td->thinktime_blocks)) {
447 int left;
448
449 if (td->thinktime_spin)
450 __usec_sleep(td->thinktime_spin);
451
452 left = td->thinktime - td->thinktime_spin;
453 if (left)
454 usec_sleep(td, left);
455 }
9c1f7434 456 }
ebac4655
JA
457 }
458
4d2413c6 459 if (!td->error) {
3d7c391d
JA
460 struct fio_file *f;
461
84585003
JA
462 if (td->cur_depth)
463 cleanup_pending_aio(td);
ebac4655 464
84585003
JA
465 if (should_fsync(td) && td->end_fsync) {
466 td_set_runstate(td, TD_FSYNCING);
467 for_each_file(td, f, i)
858a3d47 468 fio_io_sync(td, f);
84585003 469 }
5853e5a8 470 }
ebac4655
JA
471}
472
ebac4655
JA
473static void cleanup_io_u(struct thread_data *td)
474{
475 struct list_head *entry, *n;
476 struct io_u *io_u;
477
478 list_for_each_safe(entry, n, &td->io_u_freelist) {
479 io_u = list_entry(entry, struct io_u, list);
480
481 list_del(&io_u->list);
482 free(io_u);
483 }
484
2f9ade3c 485 free_io_mem(td);
ebac4655
JA
486}
487
6b9cea23
JA
488/*
489 * "randomly" fill the buffer contents
490 */
66eeb296 491static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 492{
66eeb296 493 int *ptr = io_u->buf;
6b9cea23
JA
494
495 while ((void *) ptr - io_u->buf < max_bs) {
496 *ptr = rand() * 0x9e370001;
497 ptr++;
498 }
499}
500
ebac4655
JA
501static int init_io_u(struct thread_data *td)
502{
503 struct io_u *io_u;
a00735e6 504 unsigned int max_bs;
ebac4655
JA
505 int i, max_units;
506 char *p;
507
2866c82d 508 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
509 return 0;
510
2866c82d 511 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
512 max_units = 1;
513 else
514 max_units = td->iodepth;
515
a00735e6 516 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
74b025b0
JA
517 td->orig_buffer_size = max_bs * max_units;
518
d0bdaf49 519 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
56bb17f2 520 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
74b025b0 521 else
29d610e1 522 td->orig_buffer_size += page_mask;
ebac4655 523
2f9ade3c
JA
524 if (allocate_io_mem(td))
525 return 1;
ebac4655 526
ebac4655
JA
527 p = ALIGN(td->orig_buffer);
528 for (i = 0; i < max_units; i++) {
529 io_u = malloc(sizeof(*io_u));
530 memset(io_u, 0, sizeof(*io_u));
531 INIT_LIST_HEAD(&io_u->list);
532
a00735e6 533 io_u->buf = p + max_bs * i;
6b9cea23 534 if (td_write(td) || td_rw(td))
a00735e6 535 fill_rand_buf(io_u, max_bs);
6b9cea23 536
b1ff3403 537 io_u->index = i;
ebac4655
JA
538 list_add(&io_u->list, &td->io_u_freelist);
539 }
540
541 return 0;
542}
543
da86774e
JA
544static int switch_ioscheduler(struct thread_data *td)
545{
546 char tmp[256], tmp2[128];
547 FILE *f;
548 int ret;
549
f48b467c
JA
550 if (td->io_ops->flags & FIO_CPUIO)
551 return 0;
552
da86774e
JA
553 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
554
555 f = fopen(tmp, "r+");
556 if (!f) {
557 td_verror(td, errno);
558 return 1;
559 }
560
561 /*
562 * Set io scheduler.
563 */
564 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
565 if (ferror(f) || ret != 1) {
566 td_verror(td, errno);
567 fclose(f);
568 return 1;
569 }
570
571 rewind(f);
572
573 /*
574 * Read back and check that the selected scheduler is now the default.
575 */
576 ret = fread(tmp, 1, sizeof(tmp), f);
577 if (ferror(f) || ret < 0) {
578 td_verror(td, errno);
579 fclose(f);
580 return 1;
581 }
582
583 sprintf(tmp2, "[%s]", td->ioscheduler);
584 if (!strstr(tmp, tmp2)) {
3b70d7e5 585 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
586 td_verror(td, EINVAL);
587 fclose(f);
588 return 1;
589 }
590
591 fclose(f);
592 return 0;
593}
594
ebac4655
JA
595static void clear_io_state(struct thread_data *td)
596{
53cdc686
JA
597 struct fio_file *f;
598 int i;
ebac4655 599
ebac4655
JA
600 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
601 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 602 td->zone_bytes = 0;
ebac4655 603
c1324df1
JA
604 td->last_was_sync = 0;
605
53cdc686 606 for_each_file(td, f, i) {
c1324df1
JA
607 f->last_completed_pos = 0;
608
53cdc686
JA
609 f->last_pos = 0;
610 if (td->io_ops->flags & FIO_SYNCIO)
611 lseek(f->fd, SEEK_SET, 0);
612
613 if (f->file_map)
614 memset(f->file_map, 0, f->num_maps * sizeof(long));
615 }
ebac4655
JA
616}
617
906c8d75
JA
618/*
619 * Entry point for the thread based jobs. The process based jobs end up
620 * here as well, after a little setup.
621 */
ebac4655
JA
622static void *thread_main(void *data)
623{
69008999 624 unsigned long long runtime[2];
ebac4655 625 struct thread_data *td = data;
ebac4655
JA
626
627 if (!td->use_thread)
628 setsid();
629
630 td->pid = getpid();
631
aea47d44
JA
632 INIT_LIST_HEAD(&td->io_u_freelist);
633 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 634 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44
JA
635 INIT_LIST_HEAD(&td->io_hist_list);
636 INIT_LIST_HEAD(&td->io_log_list);
637
ebac4655
JA
638 if (init_io_u(td))
639 goto err;
640
641 if (fio_setaffinity(td) == -1) {
642 td_verror(td, errno);
643 goto err;
644 }
645
aea47d44
JA
646 if (init_iolog(td))
647 goto err;
648
ebac4655
JA
649 if (td->ioprio) {
650 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
651 td_verror(td, errno);
652 goto err;
653 }
654 }
655
1056eaad 656 if (nice(td->nice) == -1) {
b6f4d880
JA
657 td_verror(td, errno);
658 goto err;
659 }
660
75154845
JA
661 if (init_random_state(td))
662 goto err;
663
56f9498d
JA
664 if (td->ioscheduler && switch_ioscheduler(td))
665 goto err;
da86774e 666
75154845 667 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
668 fio_sem_up(&startup_sem);
669 fio_sem_down(&td->mutex);
ebac4655 670
53cdc686 671 if (!td->create_serialize && setup_files(td))
ebac4655 672 goto err;
21972cde
JA
673 if (open_files(td))
674 goto err;
ebac4655 675
7d6c5283
JA
676 /*
677 * Do this late, as some IO engines would like to have the
678 * files setup prior to initializing structures.
679 */
680 if (td_io_init(td))
681 goto err;
682
69cfd7e0
JA
683 if (td->exec_prerun) {
684 if (system(td->exec_prerun) < 0)
685 goto err;
686 }
4e0ba8af 687
69008999 688 fio_gettime(&td->epoch, NULL);
36dff966 689 getrusage(RUSAGE_SELF, &td->ru_start);
69008999
JA
690
691 runtime[0] = runtime[1] = 0;
ebac4655 692 while (td->loops--) {
02bcaa8c 693 fio_gettime(&td->start, NULL);
ebac4655
JA
694 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
695
696 if (td->ratemin)
697 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
698
699 clear_io_state(td);
700 prune_io_piece_log(td);
701
2866c82d 702 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
703 do_cpuio(td);
704 else
705 do_io(td);
ebac4655 706
69008999 707 runtime[td->ddir] += utime_since_now(&td->start);
aea47d44 708 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
69008999 709 runtime[td->ddir ^ 1] = runtime[td->ddir];
3d60d1ed 710
ebac4655
JA
711 if (td->error || td->terminate)
712 break;
713
714 if (td->verify == VERIFY_NONE)
715 continue;
716
717 clear_io_state(td);
02bcaa8c 718 fio_gettime(&td->start, NULL);
ebac4655
JA
719
720 do_verify(td);
721
69008999 722 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
723
724 if (td->error || td->terminate)
725 break;
726 }
727
36dff966 728 update_rusage_stat(td);
69008999
JA
729 fio_gettime(&td->end_time, NULL);
730 td->runtime[0] = runtime[0] / 1000;
731 td->runtime[1] = runtime[1] / 1000;
732
ebac4655
JA
733 if (td->bw_log)
734 finish_log(td, td->bw_log, "bw");
735 if (td->slat_log)
736 finish_log(td, td->slat_log, "slat");
737 if (td->clat_log)
738 finish_log(td, td->clat_log, "clat");
076efc7c 739 if (td->write_iolog_file)
843a7413 740 write_iolog_close(td);
69cfd7e0
JA
741 if (td->exec_postrun) {
742 if (system(td->exec_postrun) < 0)
743 log_err("fio: postrun %s failed\n", td->exec_postrun);
744 }
ebac4655
JA
745
746 if (exitall_on_terminate)
6ce15a32 747 terminate_threads(td->groupid, 0);
ebac4655
JA
748
749err:
5bf13a5a
JA
750 if (td->error)
751 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 752 close_files(td);
2866c82d 753 close_ioengine(td);
ebac4655 754 cleanup_io_u(td);
ebac4655 755 td_set_runstate(td, TD_EXITED);
a6418147 756 return (void *) td->error;
ebac4655
JA
757}
758
906c8d75
JA
759/*
760 * We cannot pass the td data into a forked process, so attach the td and
761 * pass it to the thread worker.
762 */
a6418147 763static int fork_main(int shmid, int offset)
ebac4655
JA
764{
765 struct thread_data *td;
a6418147 766 void *data, *ret;
ebac4655
JA
767
768 data = shmat(shmid, NULL, 0);
769 if (data == (void *) -1) {
a6418147
JA
770 int __err = errno;
771
ebac4655 772 perror("shmat");
a6418147 773 return __err;
ebac4655
JA
774 }
775
776 td = data + offset * sizeof(struct thread_data);
a6418147 777 ret = thread_main(td);
ebac4655 778 shmdt(data);
a6418147 779 return (int) ret;
ebac4655
JA
780}
781
906c8d75
JA
782/*
783 * Run over the job map and reap the threads that have exited, if any.
784 */
ebac4655
JA
785static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
786{
34572e28 787 struct thread_data *td;
fab6aa71 788 int i, cputhreads, pending, status, ret;
ebac4655
JA
789
790 /*
791 * reap exited threads (TD_EXITED -> TD_REAPED)
792 */
4d2413c6 793 pending = cputhreads = 0;
34572e28 794 for_each_td(td, i) {
84585003
JA
795 /*
796 * ->io_ops is NULL for a thread that has closed its
797 * io engine
798 */
799 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
800 cputhreads++;
801
fab6aa71
JA
802 if (td->runstate < TD_EXITED) {
803 /*
804 * check if someone quit or got killed in an unusual way
805 */
806 ret = waitpid(td->pid, &status, WNOHANG);
807 if (ret < 0)
808 perror("waitpid");
809 else if ((ret == td->pid) && WIFSIGNALED(status)) {
810 int sig = WTERMSIG(status);
811
812 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
813 td_set_runstate(td, TD_REAPED);
814 goto reaped;
815 }
816 }
817
4d2413c6
JA
818 if (td->runstate != TD_EXITED) {
819 if (td->runstate < TD_RUNNING)
820 pending++;
821
ebac4655 822 continue;
4d2413c6 823 }
ebac4655 824
437c9b71
JA
825 if (td->error)
826 exit_value++;
827
ebac4655
JA
828 td_set_runstate(td, TD_REAPED);
829
830 if (td->use_thread) {
831 long ret;
832
833 if (pthread_join(td->thread, (void *) &ret))
834 perror("thread_join");
a6418147
JA
835 } else {
836 int status;
837
fab6aa71
JA
838 ret = waitpid(td->pid, &status, 0);
839 if (ret < 0)
840 perror("waitpid");
73170f19 841 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
a6418147
JA
842 if (!exit_value)
843 exit_value++;
844 }
845 }
ebac4655 846
fab6aa71 847reaped:
ebac4655
JA
848 (*nr_running)--;
849 (*m_rate) -= td->ratemin;
850 (*t_rate) -= td->rate;
851 }
b990b5c0 852
4d2413c6 853 if (*nr_running == cputhreads && !pending)
6ce15a32 854 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
855}
856
906c8d75
JA
857/*
858 * Main function for kicking off and reaping jobs, as needed.
859 */
ebac4655
JA
860static void run_threads(void)
861{
ebac4655
JA
862 struct thread_data *td;
863 unsigned long spent;
864 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 865
2f9ade3c
JA
866 if (fio_pin_memory())
867 return;
ebac4655 868
c6ae0a5b
JA
869 if (!terse_output) {
870 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
871 fflush(stdout);
872 }
c04f7ec3 873
4efa970e
JA
874 signal(SIGINT, sig_handler);
875 signal(SIGALRM, sig_handler);
876
ebac4655
JA
877 todo = thread_number;
878 nr_running = 0;
879 nr_started = 0;
880 m_rate = t_rate = 0;
881
34572e28 882 for_each_td(td, i) {
263e529f 883 print_status_init(td->thread_number - 1);
ebac4655 884
380cf265
JA
885 if (!td->create_serialize) {
886 init_disk_util(td);
ebac4655 887 continue;
380cf265 888 }
ebac4655
JA
889
890 /*
891 * do file setup here so it happens sequentially,
892 * we don't want X number of threads getting their
893 * client data interspersed on disk
894 */
53cdc686 895 if (setup_files(td)) {
5bf13a5a
JA
896 exit_value++;
897 if (td->error)
898 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
899 td_set_runstate(td, TD_REAPED);
900 todo--;
901 }
380cf265
JA
902
903 init_disk_util(td);
ebac4655
JA
904 }
905
ebac4655 906 while (todo) {
75154845
JA
907 struct thread_data *map[MAX_JOBS];
908 struct timeval this_start;
909 int this_jobs = 0, left;
910
ebac4655
JA
911 /*
912 * create threads (TD_NOT_CREATED -> TD_CREATED)
913 */
34572e28 914 for_each_td(td, i) {
ebac4655
JA
915 if (td->runstate != TD_NOT_CREATED)
916 continue;
917
918 /*
919 * never got a chance to start, killed by other
920 * thread for some reason
921 */
922 if (td->terminate) {
923 todo--;
924 continue;
925 }
926
927 if (td->start_delay) {
263e529f 928 spent = mtime_since_genesis();
ebac4655
JA
929
930 if (td->start_delay * 1000 > spent)
931 continue;
932 }
933
934 if (td->stonewall && (nr_started || nr_running))
935 break;
936
75154845
JA
937 /*
938 * Set state to created. Thread will transition
939 * to TD_INITIALIZED when it's done setting up.
940 */
ebac4655 941 td_set_runstate(td, TD_CREATED);
75154845 942 map[this_jobs++] = td;
bbfd6b00 943 fio_sem_init(&startup_sem, 1);
ebac4655
JA
944 nr_started++;
945
946 if (td->use_thread) {
947 if (pthread_create(&td->thread, NULL, thread_main, td)) {
948 perror("thread_create");
949 nr_started--;
950 }
951 } else {
952 if (fork())
bbfd6b00 953 fio_sem_down(&startup_sem);
ebac4655 954 else {
a6418147
JA
955 int ret = fork_main(shm_id, i);
956
957 exit(ret);
ebac4655
JA
958 }
959 }
960 }
961
962 /*
75154845
JA
963 * Wait for the started threads to transition to
964 * TD_INITIALIZED.
ebac4655 965 */
02bcaa8c 966 fio_gettime(&this_start, NULL);
75154845 967 left = this_jobs;
6ce15a32 968 while (left && !fio_abort) {
75154845
JA
969 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
970 break;
971
972 usleep(100000);
973
974 for (i = 0; i < this_jobs; i++) {
975 td = map[i];
976 if (!td)
977 continue;
b6f4d880 978 if (td->runstate == TD_INITIALIZED) {
75154845
JA
979 map[i] = NULL;
980 left--;
b6f4d880
JA
981 } else if (td->runstate >= TD_EXITED) {
982 map[i] = NULL;
983 left--;
984 todo--;
985 nr_running++; /* work-around... */
75154845
JA
986 }
987 }
988 }
989
990 if (left) {
3b70d7e5 991 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
992 for (i = 0; i < this_jobs; i++) {
993 td = map[i];
994 if (!td)
995 continue;
996 kill(td->pid, SIGTERM);
997 }
998 break;
999 }
1000
1001 /*
b6f4d880 1002 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1003 */
34572e28 1004 for_each_td(td, i) {
75154845 1005 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1006 continue;
1007
1008 td_set_runstate(td, TD_RUNNING);
1009 nr_running++;
1010 nr_started--;
1011 m_rate += td->ratemin;
1012 t_rate += td->rate;
75154845 1013 todo--;
bbfd6b00 1014 fio_sem_up(&td->mutex);
ebac4655
JA
1015 }
1016
1017 reap_threads(&nr_running, &t_rate, &m_rate);
1018
1019 if (todo)
1020 usleep(100000);
1021 }
1022
1023 while (nr_running) {
1024 reap_threads(&nr_running, &t_rate, &m_rate);
1025 usleep(10000);
1026 }
1027
1028 update_io_ticks();
2f9ade3c 1029 fio_unpin_memory();
ebac4655
JA
1030}
1031
ebac4655
JA
1032int main(int argc, char *argv[])
1033{
29d610e1
JA
1034 long ps;
1035
dbe1125e
JA
1036 /*
1037 * We need locale for number printing, if it isn't set then just
1038 * go with the US format.
1039 */
1040 if (!getenv("LC_NUMERIC"))
1041 setlocale(LC_NUMERIC, "en_US");
1042
ebac4655
JA
1043 if (parse_options(argc, argv))
1044 return 1;
1045
1046 if (!thread_number) {
3b70d7e5 1047 log_err("Nothing to do\n");
ebac4655
JA
1048 return 1;
1049 }
1050
29d610e1
JA
1051 ps = sysconf(_SC_PAGESIZE);
1052 if (ps < 0) {
1053 log_err("Failed to get page size\n");
1054 return 1;
1055 }
1056
1057 page_mask = ps - 1;
1058
bb3884d8
JA
1059 if (write_bw_log) {
1060 setup_log(&agg_io_log[DDIR_READ]);
1061 setup_log(&agg_io_log[DDIR_WRITE]);
1062 }
1063
ebac4655
JA
1064 disk_util_timer_arm();
1065
1066 run_threads();
6ce15a32 1067
bb3884d8 1068 if (!fio_abort) {
6ce15a32 1069 show_run_stats();
bb3884d8
JA
1070 if (write_bw_log) {
1071 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1072 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1073 }
1074 }
ebac4655 1075
437c9b71 1076 return exit_value;
ebac4655 1077}