Fix bug with time_based run
[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"
210f43b3 38#include "hash.h"
ebac4655 39
cfc99db7
JA
40unsigned long page_mask;
41unsigned long page_size;
29d610e1
JA
42#define ALIGN(buf) \
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
44
45int groupid = 0;
46int thread_number = 0;
9cedf167
JA
47int nr_process = 0;
48int nr_thread = 0;
ebac4655 49int shm_id = 0;
53cdc686 50int temp_stall_ts;
ebac4655 51
07739b57 52static struct fio_sem *startup_sem;
6ce15a32 53static volatile int fio_abort;
437c9b71 54static int exit_value;
ebac4655 55
bb3884d8
JA
56struct io_log *agg_io_log[2];
57
ebac4655 58#define TERMINATE_ALL (-1)
75154845 59#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 60
6ce15a32
JA
61static inline void td_set_runstate(struct thread_data *td, int runstate)
62{
63 td->runstate = runstate;
64}
65
390c40e2 66static void terminate_threads(int group_id)
ebac4655 67{
34572e28 68 struct thread_data *td;
ebac4655
JA
69 int i;
70
34572e28 71 for_each_td(td, i) {
ebac4655 72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
7a93ab19
JA
73 /*
74 * if the thread is running, just let it exit
75 */
76 if (td->runstate < TD_RUNNING)
c1302d44 77 kill(td->pid, SIGQUIT);
ebac4655 78 td->terminate = 1;
2dc1bbeb 79 td->o.start_delay = 0;
ebac4655
JA
80 }
81 }
82}
83
84static void sig_handler(int sig)
85{
86 switch (sig) {
87 case SIGALRM:
88 update_io_ticks();
89 disk_util_timer_arm();
90 print_thread_status();
91 break;
92 default:
6ce15a32 93 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 94 fflush(stdout);
390c40e2 95 terminate_threads(TERMINATE_ALL);
ebac4655
JA
96 break;
97 }
98}
99
906c8d75
JA
100/*
101 * Check if we are above the minimum rate given.
102 */
ebac4655
JA
103static int check_min_rate(struct thread_data *td, struct timeval *now)
104{
0904200b 105 unsigned long long bytes = 0;
4e991c23 106 unsigned long iops = 0;
ebac4655
JA
107 unsigned long spent;
108 unsigned long rate;
ebac4655 109
780bf1a1
JA
110 /*
111 * No minimum rate set, always ok
112 */
2dc1bbeb 113 if (!td->o.ratemin && !td->o.rate_iops_min)
780bf1a1
JA
114 return 0;
115
ebac4655
JA
116 /*
117 * allow a 2 second settle period in the beginning
118 */
119 if (mtime_since(&td->start, now) < 2000)
120 return 0;
121
4e991c23
JA
122 if (td_read(td)) {
123 iops += td->io_blocks[DDIR_READ];
0904200b 124 bytes += td->this_io_bytes[DDIR_READ];
4e991c23
JA
125 }
126 if (td_write(td)) {
127 iops += td->io_blocks[DDIR_WRITE];
0904200b 128 bytes += td->this_io_bytes[DDIR_WRITE];
4e991c23 129 }
0904200b 130
ebac4655
JA
131 /*
132 * if rate blocks is set, sample is running
133 */
4e991c23 134 if (td->rate_bytes || td->rate_blocks) {
ebac4655 135 spent = mtime_since(&td->lastrate, now);
2dc1bbeb 136 if (spent < td->o.ratecycle)
ebac4655
JA
137 return 0;
138
2dc1bbeb 139 if (td->o.rate) {
4e991c23
JA
140 /*
141 * check bandwidth specified rate
142 */
143 if (bytes < td->rate_bytes) {
2dc1bbeb 144 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
4e991c23
JA
145 return 1;
146 } else {
147 rate = (bytes - td->rate_bytes) / spent;
2dc1bbeb
JA
148 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
4e991c23
JA
150 return 1;
151 }
152 }
413dd459 153 } else {
4e991c23
JA
154 /*
155 * checks iops specified rate
156 */
2dc1bbeb
JA
157 if (iops < td->o.rate_iops) {
158 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
413dd459 159 return 1;
4e991c23
JA
160 } else {
161 rate = (iops - td->rate_blocks) / spent;
2dc1bbeb
JA
162 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
4e991c23 164 }
413dd459 165 }
ebac4655
JA
166 }
167 }
168
0904200b 169 td->rate_bytes = bytes;
4e991c23 170 td->rate_blocks = iops;
ebac4655
JA
171 memcpy(&td->lastrate, now, sizeof(*now));
172 return 0;
173}
174
175static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176{
2dc1bbeb 177 if (!td->o.timeout)
ebac4655 178 return 0;
2dc1bbeb 179 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
180 return 1;
181
182 return 0;
183}
184
906c8d75
JA
185/*
186 * When job exits, we can cancel the in-flight IO if we are using async
187 * io. Attempt to do so.
188 */
ebac4655
JA
189static void cleanup_pending_aio(struct thread_data *td)
190{
ebac4655 191 struct list_head *entry, *n;
ebac4655
JA
192 struct io_u *io_u;
193 int r;
194
195 /*
196 * get immediately available events, if any
197 */
d7762cf8 198 r = io_u_queued_complete(td, 0);
b2fdda43
JA
199 if (r < 0)
200 return;
ebac4655
JA
201
202 /*
203 * now cancel remaining active events
204 */
2866c82d 205 if (td->io_ops->cancel) {
ebac4655
JA
206 list_for_each_safe(entry, n, &td->io_u_busylist) {
207 io_u = list_entry(entry, struct io_u, list);
208
0c6e7517
JA
209 /*
210 * if the io_u isn't in flight, then that generally
211 * means someone leaked an io_u. complain but fix
212 * it up, so we don't stall here.
213 */
214 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215 log_err("fio: non-busy IO on busy list\n");
ebac4655 216 put_io_u(td, io_u);
0c6e7517
JA
217 } else {
218 r = td->io_ops->cancel(td, io_u);
219 if (!r)
220 put_io_u(td, io_u);
221 }
ebac4655
JA
222 }
223 }
224
97601024 225 if (td->cur_depth)
d7762cf8 226 r = io_u_queued_complete(td, td->cur_depth);
ebac4655
JA
227}
228
858a3d47
JA
229/*
230 * Helper to handle the final sync of a file. Works just like the normal
231 * io path, just does everything sync.
232 */
233static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234{
235 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
236 int ret;
237
238 if (!io_u)
239 return 1;
240
241 io_u->ddir = DDIR_SYNC;
242 io_u->file = f;
243
244 if (td_io_prep(td, io_u)) {
245 put_io_u(td, io_u);
246 return 1;
247 }
248
755200a3 249requeue:
858a3d47 250 ret = td_io_queue(td, io_u);
36167d82 251 if (ret < 0) {
e1161c32 252 td_verror(td, io_u->error, "td_io_queue");
858a3d47 253 put_io_u(td, io_u);
858a3d47 254 return 1;
36167d82 255 } else if (ret == FIO_Q_QUEUED) {
d7762cf8 256 if (io_u_queued_complete(td, 1) < 0)
36167d82 257 return 1;
36167d82
JA
258 } else if (ret == FIO_Q_COMPLETED) {
259 if (io_u->error) {
e1161c32 260 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
261 return 1;
262 }
858a3d47 263
d7762cf8 264 if (io_u_sync_complete(td, io_u) < 0)
b2fdda43 265 return 1;
755200a3
JA
266 } else if (ret == FIO_Q_BUSY) {
267 if (td_io_commit(td))
268 return 1;
269 goto requeue;
858a3d47
JA
270 }
271
272 return 0;
273}
274
906c8d75 275/*
34403fb1 276 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
277 * reads the blocks back in, and checks the crc/md5 of the data.
278 */
1e97cce9 279static void do_verify(struct thread_data *td)
ebac4655 280{
53cdc686 281 struct fio_file *f;
36167d82 282 struct io_u *io_u;
af52b345
JA
283 int ret, min_events;
284 unsigned int i;
e5b401d4
JA
285
286 /*
287 * sync io first and invalidate cache, to make sure we really
288 * read from disk.
289 */
290 for_each_file(td, f, i) {
3d7b485f
JA
291 if (!(f->flags & FIO_FILE_OPEN))
292 continue;
b2fdda43
JA
293 if (fio_io_sync(td, f))
294 break;
295 if (file_invalidate_cache(td, f))
296 break;
e5b401d4 297 }
ebac4655 298
b2fdda43
JA
299 if (td->error)
300 return;
301
ebac4655
JA
302 td_set_runstate(td, TD_VERIFYING);
303
36167d82
JA
304 io_u = NULL;
305 while (!td->terminate) {
5451792e
JA
306 int ret2;
307
ebac4655
JA
308 io_u = __get_io_u(td);
309 if (!io_u)
310 break;
311
069c2918
JA
312 if (runtime_exceeded(td, &io_u->start_time)) {
313 put_io_u(td, io_u);
90a87d4b 314 td->terminate = 1;
02bcaa8c 315 break;
069c2918 316 }
02bcaa8c 317
069c2918
JA
318 if (get_next_verify(td, io_u)) {
319 put_io_u(td, io_u);
ebac4655 320 break;
069c2918 321 }
ebac4655 322
069c2918
JA
323 if (td_io_prep(td, io_u)) {
324 put_io_u(td, io_u);
53cdc686 325 break;
069c2918 326 }
d7762cf8
JA
327
328 io_u->end_io = verify_io_u;
53cdc686 329
11786802 330 ret = td_io_queue(td, io_u);
36167d82
JA
331 switch (ret) {
332 case FIO_Q_COMPLETED:
333 if (io_u->error)
22819ec2 334 ret = -io_u->error;
9e9d164e 335 else if (io_u->resid) {
36167d82 336 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 337 struct fio_file *f = io_u->file;
ebac4655 338
9e9d164e
JA
339 /*
340 * zero read, fail
341 */
342 if (!bytes) {
343 td_verror(td, ENODATA, "full resid");
344 put_io_u(td, io_u);
345 break;
346 }
8400d9b2 347
36167d82
JA
348 io_u->xfer_buflen = io_u->resid;
349 io_u->xfer_buf += bytes;
8400d9b2 350 io_u->offset += bytes;
d460eb31 351 f->last_completed_pos = io_u->offset;
8400d9b2 352
30061b97
JA
353 td->ts.short_io_u[io_u->ddir]++;
354
d460eb31 355 if (io_u->offset == f->real_file_size)
8400d9b2
JA
356 goto sync_done;
357
11786802
JA
358 requeue_io_u(td, &io_u);
359 } else {
8400d9b2 360sync_done:
11786802
JA
361 ret = io_u_sync_complete(td, io_u);
362 if (ret < 0)
363 break;
36167d82 364 }
36167d82
JA
365 continue;
366 case FIO_Q_QUEUED:
367 break;
755200a3
JA
368 case FIO_Q_BUSY:
369 requeue_io_u(td, &io_u);
5451792e
JA
370 ret2 = td_io_commit(td);
371 if (ret2 < 0)
372 ret = ret2;
755200a3 373 break;
36167d82
JA
374 default:
375 assert(ret < 0);
e1161c32 376 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
377 break;
378 }
379
99784632 380 if (ret < 0 || td->error)
3af6ef39
JA
381 break;
382
ebac4655 383 /*
3af6ef39
JA
384 * if we can queue more, do so. but check if there are
385 * completed io_u's first.
ebac4655 386 */
97601024 387 min_events = 0;
e916b390 388 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 389 min_events = 1;
3af6ef39 390
2dc1bbeb
JA
391 if (td->cur_depth > td->o.iodepth_low)
392 min_events = td->cur_depth - td->o.iodepth_low;
e916b390
JA
393 }
394
3af6ef39
JA
395 /*
396 * Reap required number of io units, if any, and do the
397 * verification on them through the callback handler
398 */
d7762cf8 399 if (io_u_queued_complete(td, min_events) < 0)
ebac4655 400 break;
36167d82 401 }
ebac4655 402
c01c0395
JA
403 if (!td->error) {
404 min_events = td->cur_depth;
405
406 if (min_events)
407 ret = io_u_queued_complete(td, min_events);
408 } else
ebac4655
JA
409 cleanup_pending_aio(td);
410
411 td_set_runstate(td, TD_RUNNING);
412}
413
32cd46a0 414/*
906c8d75 415 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
416 * and reaps them, checking for rate and errors along the way.
417 */
ebac4655
JA
418static void do_io(struct thread_data *td)
419{
02bcaa8c 420 struct timeval s;
ebac4655 421 unsigned long usec;
af52b345
JA
422 unsigned int i;
423 int ret = 0;
ebac4655 424
5853e5a8
JA
425 td_set_runstate(td, TD_RUNNING);
426
7bb48f84 427 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
97601024
JA
428 struct timeval comp_time;
429 long bytes_done = 0;
84585003 430 int min_evts = 0;
ebac4655 431 struct io_u *io_u;
5451792e 432 int ret2;
ebac4655
JA
433
434 if (td->terminate)
435 break;
436
3d7c391d 437 io_u = get_io_u(td);
ebac4655
JA
438 if (!io_u)
439 break;
440
441 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
442
443 if (runtime_exceeded(td, &s)) {
444 put_io_u(td, io_u);
90a87d4b 445 td->terminate = 1;
97601024
JA
446 break;
447 }
36167d82 448
11786802 449 ret = td_io_queue(td, io_u);
36167d82
JA
450 switch (ret) {
451 case FIO_Q_COMPLETED:
5451792e
JA
452 if (io_u->error)
453 ret = -io_u->error;
9e9d164e 454 else if (io_u->resid) {
36167d82 455 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 456 struct fio_file *f = io_u->file;
36167d82 457
9e9d164e
JA
458 /*
459 * zero read, fail
460 */
461 if (!bytes) {
462 td_verror(td, ENODATA, "full resid");
463 put_io_u(td, io_u);
464 break;
465 }
466
cec6b55d 467 io_u->xfer_buflen = io_u->resid;
36167d82 468 io_u->xfer_buf += bytes;
5a7c5680 469 io_u->offset += bytes;
d460eb31 470 f->last_completed_pos = io_u->offset;
5a7c5680 471
30061b97
JA
472 td->ts.short_io_u[io_u->ddir]++;
473
d460eb31 474 if (io_u->offset == f->real_file_size)
5a7c5680
JA
475 goto sync_done;
476
11786802
JA
477 requeue_io_u(td, &io_u);
478 } else {
5a7c5680 479sync_done:
11786802
JA
480 fio_gettime(&comp_time, NULL);
481 bytes_done = io_u_sync_complete(td, io_u);
482 if (bytes_done < 0)
483 ret = bytes_done;
cec6b55d 484 }
36167d82
JA
485 break;
486 case FIO_Q_QUEUED:
7e77dd02
JA
487 /*
488 * if the engine doesn't have a commit hook,
489 * the io_u is really queued. if it does have such
490 * a hook, it has to call io_u_queued() itself.
491 */
492 if (td->io_ops->commit == NULL)
493 io_u_queued(td, io_u);
36167d82 494 break;
755200a3
JA
495 case FIO_Q_BUSY:
496 requeue_io_u(td, &io_u);
5451792e
JA
497 ret2 = td_io_commit(td);
498 if (ret2 < 0)
499 ret = ret2;
755200a3 500 break;
36167d82
JA
501 default:
502 assert(ret < 0);
503 put_io_u(td, io_u);
504 break;
ebac4655
JA
505 }
506
99784632 507 if (ret < 0 || td->error)
36167d82
JA
508 break;
509
97601024
JA
510 /*
511 * See if we need to complete some commands
512 */
755200a3 513 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 514 min_evts = 0;
e916b390 515 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 516 min_evts = 1;
ebac4655 517
2dc1bbeb
JA
518 if (td->cur_depth > td->o.iodepth_low)
519 min_evts = td->cur_depth - td->o.iodepth_low;
e916b390
JA
520 }
521
97601024 522 fio_gettime(&comp_time, NULL);
d7762cf8 523 bytes_done = io_u_queued_complete(td, min_evts);
97601024 524 if (bytes_done < 0)
36167d82 525 break;
ebac4655
JA
526 }
527
97601024
JA
528 if (!bytes_done)
529 continue;
530
ebac4655
JA
531 /*
532 * the rate is batched for now, it should work for batches
533 * of completions except the very first one which may look
534 * a little bursty
535 */
97601024 536 usec = utime_since(&s, &comp_time);
ebac4655 537
413dd459 538 rate_throttle(td, usec, bytes_done);
ebac4655 539
97601024 540 if (check_min_rate(td, &comp_time)) {
98aa62d8 541 if (exitall_on_terminate)
390c40e2 542 terminate_threads(td->groupid);
e1161c32 543 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
544 break;
545 }
546
2dc1bbeb 547 if (td->o.thinktime) {
9c1f7434
JA
548 unsigned long long b;
549
550 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 551 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
552 int left;
553
2dc1bbeb
JA
554 if (td->o.thinktime_spin)
555 __usec_sleep(td->o.thinktime_spin);
48097d5c 556
2dc1bbeb 557 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
558 if (left)
559 usec_sleep(td, left);
560 }
9c1f7434 561 }
ebac4655
JA
562 }
563
4d2413c6 564 if (!td->error) {
3d7c391d
JA
565 struct fio_file *f;
566
c01c0395
JA
567 i = td->cur_depth;
568 if (i)
569 ret = io_u_queued_complete(td, i);
ebac4655 570
2dc1bbeb 571 if (should_fsync(td) && td->o.end_fsync) {
84585003 572 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
573
574 for_each_file(td, f, i) {
575 if (!(f->flags & FIO_FILE_OPEN))
576 continue;
858a3d47 577 fio_io_sync(td, f);
3d7b485f 578 }
84585003 579 }
c01c0395
JA
580 } else
581 cleanup_pending_aio(td);
ebac4655
JA
582}
583
ebac4655
JA
584static void cleanup_io_u(struct thread_data *td)
585{
586 struct list_head *entry, *n;
587 struct io_u *io_u;
588
589 list_for_each_safe(entry, n, &td->io_u_freelist) {
590 io_u = list_entry(entry, struct io_u, list);
591
592 list_del(&io_u->list);
593 free(io_u);
594 }
595
2f9ade3c 596 free_io_mem(td);
ebac4655
JA
597}
598
6b9cea23
JA
599/*
600 * "randomly" fill the buffer contents
601 */
e9459e5a 602static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
6b9cea23 603{
210f43b3 604 long *ptr = io_u->buf;
6b9cea23 605
e9459e5a
JA
606 if (!td->o.zero_buffers) {
607 while ((void *) ptr - io_u->buf < max_bs) {
210f43b3 608 *ptr = rand() * GOLDEN_RATIO_PRIME;
e9459e5a
JA
609 ptr++;
610 }
611 } else
612 memset(ptr, 0, max_bs);
6b9cea23
JA
613}
614
ebac4655
JA
615static int init_io_u(struct thread_data *td)
616{
2b4ce34f 617 unsigned long long buf_size;
ebac4655 618 struct io_u *io_u;
a00735e6 619 unsigned int max_bs;
ebac4655
JA
620 int i, max_units;
621 char *p;
622
2866c82d 623 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
624 max_units = 1;
625 else
2dc1bbeb 626 max_units = td->o.iodepth;
ebac4655 627
2dc1bbeb 628 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
2b4ce34f
JA
629 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
630 buf_size += page_mask;
631 if (buf_size != (size_t) buf_size) {
632 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
633 return 1;
634 }
635
636 td->orig_buffer_size = buf_size;
74b025b0 637
2dc1bbeb
JA
638 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
639 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
4afbf66f
JA
640 else if (td->orig_buffer_size & page_mask)
641 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
ebac4655 642
2f9ade3c
JA
643 if (allocate_io_mem(td))
644 return 1;
ebac4655 645
ebac4655
JA
646 p = ALIGN(td->orig_buffer);
647 for (i = 0; i < max_units; i++) {
db1cd90b
JA
648 if (td->terminate)
649 return 1;
ebac4655
JA
650 io_u = malloc(sizeof(*io_u));
651 memset(io_u, 0, sizeof(*io_u));
652 INIT_LIST_HEAD(&io_u->list);
653
a00735e6 654 io_u->buf = p + max_bs * i;
e9459e5a 655
660a1cb5 656 if (td_write(td))
e9459e5a 657 fill_io_buf(td, io_u, max_bs);
6b9cea23 658
b1ff3403 659 io_u->index = i;
0c6e7517 660 io_u->flags = IO_U_F_FREE;
ebac4655
JA
661 list_add(&io_u->list, &td->io_u_freelist);
662 }
663
433afcb4
JA
664 io_u_init_timeout();
665
ebac4655
JA
666 return 0;
667}
668
da86774e
JA
669static int switch_ioscheduler(struct thread_data *td)
670{
671 char tmp[256], tmp2[128];
672 FILE *f;
673 int ret;
674
ba0fbe10 675 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
676 return 0;
677
da86774e
JA
678 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
679
680 f = fopen(tmp, "r+");
681 if (!f) {
cbf5c1d7
JA
682 if (errno == ENOENT) {
683 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
684 return 0;
685 }
686 td_verror(td, errno, "fopen iosched");
da86774e
JA
687 return 1;
688 }
689
690 /*
691 * Set io scheduler.
692 */
2dc1bbeb 693 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 694 if (ferror(f) || ret != 1) {
e1161c32 695 td_verror(td, errno, "fwrite");
da86774e
JA
696 fclose(f);
697 return 1;
698 }
699
700 rewind(f);
701
702 /*
703 * Read back and check that the selected scheduler is now the default.
704 */
705 ret = fread(tmp, 1, sizeof(tmp), f);
706 if (ferror(f) || ret < 0) {
e1161c32 707 td_verror(td, errno, "fread");
da86774e
JA
708 fclose(f);
709 return 1;
710 }
711
2dc1bbeb 712 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 713 if (!strstr(tmp, tmp2)) {
2dc1bbeb 714 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 715 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
716 fclose(f);
717 return 1;
718 }
719
720 fclose(f);
721 return 0;
722}
723
a978ba68 724static int clear_io_state(struct thread_data *td)
ebac4655 725{
53cdc686 726 struct fio_file *f;
af52b345
JA
727 unsigned int i;
728 int ret;
ebac4655 729
756867bd 730 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 731 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 732 td->zone_bytes = 0;
d7d3b49b 733 td->rate_bytes = 0;
4e991c23 734 td->rate_blocks = 0;
38d77cae 735 td->rw_end_set[0] = td->rw_end_set[1] = 0;
ebac4655 736
c1324df1
JA
737 td->last_was_sync = 0;
738
281f9b63
JA
739 td->nr_done_files = 0;
740
a978ba68
JA
741 for_each_file(td, f, i)
742 td_io_close_file(td, f);
53cdc686 743
a978ba68
JA
744 ret = 0;
745 for_each_file(td, f, i) {
281f9b63 746 f->flags &= ~FIO_FILE_DONE;
a978ba68
JA
747 ret = td_io_open_file(td, f);
748 if (ret)
749 break;
53cdc686 750 }
a978ba68
JA
751
752 return ret;
ebac4655
JA
753}
754
906c8d75
JA
755/*
756 * Entry point for the thread based jobs. The process based jobs end up
757 * here as well, after a little setup.
758 */
ebac4655
JA
759static void *thread_main(void *data)
760{
69008999 761 unsigned long long runtime[2];
ebac4655 762 struct thread_data *td = data;
38d77cae 763 unsigned long elapsed;
a978ba68 764 int clear_state;
ebac4655 765
2dc1bbeb 766 if (!td->o.use_thread)
ebac4655
JA
767 setsid();
768
769 td->pid = getpid();
770
aea47d44
JA
771 INIT_LIST_HEAD(&td->io_u_freelist);
772 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 773 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44 774 INIT_LIST_HEAD(&td->io_log_list);
8de8f047 775 INIT_LIST_HEAD(&td->io_hist_list);
bb5d7d0b 776 td->io_hist_tree = RB_ROOT;
aea47d44 777
db1cd90b
JA
778 td_set_runstate(td, TD_INITIALIZED);
779 fio_sem_up(startup_sem);
780 fio_sem_down(td->mutex);
781
782 /*
783 * the ->mutex semaphore is now no longer used, close it to avoid
784 * eating a file descriptor
785 */
786 fio_sem_remove(td->mutex);
787
d84f8d49
JA
788 /*
789 * May alter parameters that init_io_u() will use, so we need to
790 * do this first.
791 */
792 if (init_iolog(td))
793 goto err;
794
ebac4655 795 if (init_io_u(td))
db1cd90b 796 goto err;
ebac4655 797
375b2695 798 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
e1161c32 799 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 800 goto err;
ebac4655
JA
801 }
802
ebac4655
JA
803 if (td->ioprio) {
804 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 805 td_verror(td, errno, "ioprio_set");
db1cd90b 806 goto err;
ebac4655
JA
807 }
808 }
809
2dc1bbeb 810 if (nice(td->o.nice) == -1) {
e1161c32 811 td_verror(td, errno, "nice");
db1cd90b 812 goto err;
b6f4d880
JA
813 }
814
2dc1bbeb 815 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 816 goto err;
37f56873 817
2dc1bbeb 818 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
819 goto err;
820
7d6c5283
JA
821 if (td_io_init(td))
822 goto err;
823
b5af8293
JA
824 if (open_files(td))
825 goto err;
826
68727076
JA
827 if (init_random_map(td))
828 goto err;
829
2dc1bbeb
JA
830 if (td->o.exec_prerun) {
831 if (system(td->o.exec_prerun) < 0)
69cfd7e0
JA
832 goto err;
833 }
4e0ba8af 834
69008999 835 fio_gettime(&td->epoch, NULL);
433afcb4 836 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
756867bd 837 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
838
839 runtime[0] = runtime[1] = 0;
a978ba68 840 clear_state = 0;
cf4464ca 841 while (td->o.time_based || td->o.loops--) {
02bcaa8c 842 fio_gettime(&td->start, NULL);
756867bd 843 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655 844
2dc1bbeb 845 if (td->o.ratemin)
756867bd 846 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655 847
a978ba68
JA
848 if (clear_state && clear_io_state(td))
849 break;
850
ebac4655
JA
851 prune_io_piece_log(td);
852
ba0fbe10 853 do_io(td);
ebac4655 854
a978ba68
JA
855 clear_state = 1;
856
38d77cae
JA
857 if (td_read(td) && td->io_bytes[DDIR_READ]) {
858 if (td->rw_end_set[DDIR_READ])
859 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
860 else
861 elapsed = utime_since_now(&td->start);
862
863 runtime[DDIR_READ] += elapsed;
864 }
865 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
866 if (td->rw_end_set[DDIR_WRITE])
867 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
868 else
869 elapsed = utime_since_now(&td->start);
870
871 runtime[DDIR_WRITE] += elapsed;
872 }
413dd459 873
ebac4655
JA
874 if (td->error || td->terminate)
875 break;
876
2dc1bbeb 877 if (td->o.verify == VERIFY_NONE)
ebac4655
JA
878 continue;
879
a978ba68
JA
880 if (clear_io_state(td))
881 break;
882
02bcaa8c 883 fio_gettime(&td->start, NULL);
ebac4655
JA
884
885 do_verify(td);
886
69008999 887 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
888
889 if (td->error || td->terminate)
890 break;
891 }
892
36dff966 893 update_rusage_stat(td);
756867bd
JA
894 td->ts.runtime[0] = runtime[0] / 1000;
895 td->ts.runtime[1] = runtime[1] / 1000;
896 td->ts.total_run_time = mtime_since_now(&td->epoch);
897 td->ts.io_bytes[0] = td->io_bytes[0];
898 td->ts.io_bytes[1] = td->io_bytes[1];
899
900 if (td->ts.bw_log)
901 finish_log(td, td->ts.bw_log, "bw");
902 if (td->ts.slat_log)
903 finish_log(td, td->ts.slat_log, "slat");
904 if (td->ts.clat_log)
905 finish_log(td, td->ts.clat_log, "clat");
2dc1bbeb 906 if (td->o.write_iolog_file)
843a7413 907 write_iolog_close(td);
2dc1bbeb
JA
908 if (td->o.exec_postrun) {
909 if (system(td->o.exec_postrun) < 0)
910 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
69cfd7e0 911 }
ebac4655
JA
912
913 if (exitall_on_terminate)
390c40e2 914 terminate_threads(td->groupid);
ebac4655
JA
915
916err:
5bf13a5a
JA
917 if (td->error)
918 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 919 close_files(td);
2866c82d 920 close_ioengine(td);
ebac4655 921 cleanup_io_u(td);
d23bb327 922 options_mem_free(td);
ebac4655 923 td_set_runstate(td, TD_EXITED);
43d76807 924 return (void *) (unsigned long) td->error;
ebac4655
JA
925}
926
906c8d75
JA
927/*
928 * We cannot pass the td data into a forked process, so attach the td and
929 * pass it to the thread worker.
930 */
a6418147 931static int fork_main(int shmid, int offset)
ebac4655
JA
932{
933 struct thread_data *td;
a6418147 934 void *data, *ret;
ebac4655
JA
935
936 data = shmat(shmid, NULL, 0);
937 if (data == (void *) -1) {
a6418147
JA
938 int __err = errno;
939
ebac4655 940 perror("shmat");
a6418147 941 return __err;
ebac4655
JA
942 }
943
944 td = data + offset * sizeof(struct thread_data);
a6418147 945 ret = thread_main(td);
ebac4655 946 shmdt(data);
43d76807 947 return (int) (unsigned long) ret;
ebac4655
JA
948}
949
906c8d75
JA
950/*
951 * Run over the job map and reap the threads that have exited, if any.
952 */
ebac4655
JA
953static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
954{
34572e28 955 struct thread_data *td;
fab6aa71 956 int i, cputhreads, pending, status, ret;
ebac4655
JA
957
958 /*
959 * reap exited threads (TD_EXITED -> TD_REAPED)
960 */
4d2413c6 961 pending = cputhreads = 0;
34572e28 962 for_each_td(td, i) {
3707f45b 963 int flags = 0;
a2f77c9f 964
84585003
JA
965 /*
966 * ->io_ops is NULL for a thread that has closed its
967 * io engine
968 */
ba0fbe10 969 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0
JA
970 cputhreads++;
971
a2f77c9f
JA
972 if (!td->pid || td->runstate == TD_REAPED)
973 continue;
2dc1bbeb 974 if (td->o.use_thread) {
3707f45b
JA
975 if (td->runstate == TD_EXITED) {
976 td_set_runstate(td, TD_REAPED);
977 goto reaped;
978 }
979 continue;
980 }
a2f77c9f
JA
981
982 flags = WNOHANG;
983 if (td->runstate == TD_EXITED)
984 flags = 0;
985
986 /*
987 * check if someone quit or got killed in an unusual way
988 */
989 ret = waitpid(td->pid, &status, flags);
3707f45b 990 if (ret < 0) {
a2f77c9f
JA
991 if (errno == ECHILD) {
992 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
993 td_set_runstate(td, TD_REAPED);
994 goto reaped;
995 }
996 perror("waitpid");
997 } else if (ret == td->pid) {
998 if (WIFSIGNALED(status)) {
fab6aa71
JA
999 int sig = WTERMSIG(status);
1000
d9244941
JA
1001 if (sig != SIGQUIT)
1002 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
fab6aa71
JA
1003 td_set_runstate(td, TD_REAPED);
1004 goto reaped;
1005 }
a2f77c9f
JA
1006 if (WIFEXITED(status)) {
1007 if (WEXITSTATUS(status) && !td->error)
1008 td->error = WEXITSTATUS(status);
a2f77c9f 1009
a2f77c9f
JA
1010 td_set_runstate(td, TD_REAPED);
1011 goto reaped;
a6418147
JA
1012 }
1013 }
ebac4655 1014
a2f77c9f
JA
1015 /*
1016 * thread is not dead, continue
1017 */
e48676ba 1018 pending++;
a2f77c9f 1019 continue;
fab6aa71 1020reaped:
2dc1bbeb 1021 if (td->o.use_thread) {
3707f45b
JA
1022 long ret;
1023
1024 if (pthread_join(td->thread, (void *) &ret))
1025 perror("pthread_join");
1026 }
1027
ebac4655 1028 (*nr_running)--;
2dc1bbeb
JA
1029 (*m_rate) -= td->o.ratemin;
1030 (*t_rate) -= td->o.rate;
e48676ba 1031 pending--;
a2f77c9f
JA
1032
1033 if (td->error)
1034 exit_value++;
ebac4655 1035 }
b990b5c0 1036
4d2413c6 1037 if (*nr_running == cputhreads && !pending)
390c40e2 1038 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1039}
1040
906c8d75
JA
1041/*
1042 * Main function for kicking off and reaping jobs, as needed.
1043 */
ebac4655
JA
1044static void run_threads(void)
1045{
ebac4655
JA
1046 struct thread_data *td;
1047 unsigned long spent;
1048 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1049
2f9ade3c
JA
1050 if (fio_pin_memory())
1051 return;
ebac4655 1052
c6ae0a5b 1053 if (!terse_output) {
9cedf167
JA
1054 printf("Starting ");
1055 if (nr_thread)
1056 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1057 if (nr_process) {
1058 if (nr_thread)
1059 printf(" and ");
1060 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1061 }
1062 printf("\n");
c6ae0a5b
JA
1063 fflush(stdout);
1064 }
c04f7ec3 1065
4efa970e
JA
1066 signal(SIGINT, sig_handler);
1067 signal(SIGALRM, sig_handler);
1068
ebac4655
JA
1069 todo = thread_number;
1070 nr_running = 0;
1071 nr_started = 0;
1072 m_rate = t_rate = 0;
1073
34572e28 1074 for_each_td(td, i) {
263e529f 1075 print_status_init(td->thread_number - 1);
ebac4655 1076
2dc1bbeb 1077 if (!td->o.create_serialize) {
380cf265 1078 init_disk_util(td);
ebac4655 1079 continue;
380cf265 1080 }
ebac4655
JA
1081
1082 /*
1083 * do file setup here so it happens sequentially,
1084 * we don't want X number of threads getting their
1085 * client data interspersed on disk
1086 */
53cdc686 1087 if (setup_files(td)) {
5bf13a5a
JA
1088 exit_value++;
1089 if (td->error)
1090 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
1091 td_set_runstate(td, TD_REAPED);
1092 todo--;
1093 }
380cf265
JA
1094
1095 init_disk_util(td);
ebac4655
JA
1096 }
1097
a2f77c9f
JA
1098 set_genesis_time();
1099
ebac4655 1100 while (todo) {
75154845
JA
1101 struct thread_data *map[MAX_JOBS];
1102 struct timeval this_start;
1103 int this_jobs = 0, left;
1104
ebac4655
JA
1105 /*
1106 * create threads (TD_NOT_CREATED -> TD_CREATED)
1107 */
34572e28 1108 for_each_td(td, i) {
ebac4655
JA
1109 if (td->runstate != TD_NOT_CREATED)
1110 continue;
1111
1112 /*
1113 * never got a chance to start, killed by other
1114 * thread for some reason
1115 */
1116 if (td->terminate) {
1117 todo--;
1118 continue;
1119 }
1120
2dc1bbeb 1121 if (td->o.start_delay) {
263e529f 1122 spent = mtime_since_genesis();
ebac4655 1123
2dc1bbeb 1124 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1125 continue;
1126 }
1127
2dc1bbeb 1128 if (td->o.stonewall && (nr_started || nr_running))
ebac4655
JA
1129 break;
1130
75154845
JA
1131 /*
1132 * Set state to created. Thread will transition
1133 * to TD_INITIALIZED when it's done setting up.
1134 */
ebac4655 1135 td_set_runstate(td, TD_CREATED);
75154845 1136 map[this_jobs++] = td;
ebac4655
JA
1137 nr_started++;
1138
2dc1bbeb 1139 if (td->o.use_thread) {
ebac4655
JA
1140 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1141 perror("thread_create");
1142 nr_started--;
c8bb6faf 1143 break;
ebac4655
JA
1144 }
1145 } else {
07739b57 1146 if (!fork()) {
a6418147
JA
1147 int ret = fork_main(shm_id, i);
1148
1149 exit(ret);
ebac4655
JA
1150 }
1151 }
07739b57 1152 fio_sem_down(startup_sem);
ebac4655
JA
1153 }
1154
1155 /*
75154845
JA
1156 * Wait for the started threads to transition to
1157 * TD_INITIALIZED.
ebac4655 1158 */
02bcaa8c 1159 fio_gettime(&this_start, NULL);
75154845 1160 left = this_jobs;
6ce15a32 1161 while (left && !fio_abort) {
75154845
JA
1162 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1163 break;
1164
1165 usleep(100000);
1166
1167 for (i = 0; i < this_jobs; i++) {
1168 td = map[i];
1169 if (!td)
1170 continue;
b6f4d880 1171 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1172 map[i] = NULL;
1173 left--;
b6f4d880
JA
1174 } else if (td->runstate >= TD_EXITED) {
1175 map[i] = NULL;
1176 left--;
1177 todo--;
1178 nr_running++; /* work-around... */
75154845
JA
1179 }
1180 }
1181 }
1182
1183 if (left) {
3b70d7e5 1184 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1185 for (i = 0; i < this_jobs; i++) {
1186 td = map[i];
1187 if (!td)
1188 continue;
1189 kill(td->pid, SIGTERM);
1190 }
1191 break;
1192 }
1193
1194 /*
b6f4d880 1195 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1196 */
34572e28 1197 for_each_td(td, i) {
75154845 1198 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1199 continue;
1200
1201 td_set_runstate(td, TD_RUNNING);
1202 nr_running++;
1203 nr_started--;
2dc1bbeb
JA
1204 m_rate += td->o.ratemin;
1205 t_rate += td->o.rate;
75154845 1206 todo--;
07739b57 1207 fio_sem_up(td->mutex);
ebac4655
JA
1208 }
1209
1210 reap_threads(&nr_running, &t_rate, &m_rate);
1211
1212 if (todo)
1213 usleep(100000);
1214 }
1215
1216 while (nr_running) {
1217 reap_threads(&nr_running, &t_rate, &m_rate);
1218 usleep(10000);
1219 }
1220
1221 update_io_ticks();
2f9ade3c 1222 fio_unpin_memory();
ebac4655
JA
1223}
1224
ebac4655
JA
1225int main(int argc, char *argv[])
1226{
29d610e1
JA
1227 long ps;
1228
dbe1125e
JA
1229 /*
1230 * We need locale for number printing, if it isn't set then just
1231 * go with the US format.
1232 */
1233 if (!getenv("LC_NUMERIC"))
1234 setlocale(LC_NUMERIC, "en_US");
1235
ebac4655
JA
1236 if (parse_options(argc, argv))
1237 return 1;
1238
4b472fa3
JA
1239 if (!thread_number)
1240 return 0;
ebac4655 1241
29d610e1
JA
1242 ps = sysconf(_SC_PAGESIZE);
1243 if (ps < 0) {
1244 log_err("Failed to get page size\n");
1245 return 1;
1246 }
1247
cfc99db7 1248 page_size = ps;
29d610e1
JA
1249 page_mask = ps - 1;
1250
bb3884d8
JA
1251 if (write_bw_log) {
1252 setup_log(&agg_io_log[DDIR_READ]);
1253 setup_log(&agg_io_log[DDIR_WRITE]);
1254 }
1255
07739b57
JA
1256 startup_sem = fio_sem_init(0);
1257
a2f77c9f
JA
1258 set_genesis_time();
1259
ebac4655
JA
1260 disk_util_timer_arm();
1261
1262 run_threads();
6ce15a32 1263
bb3884d8 1264 if (!fio_abort) {
6ce15a32 1265 show_run_stats();
bb3884d8
JA
1266 if (write_bw_log) {
1267 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1268 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1269 }
1270 }
ebac4655 1271
07739b57 1272 fio_sem_remove(startup_sem);
437c9b71 1273 return exit_value;
ebac4655 1274}