Add support for limiting only rate in only one direction
[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"
2e5cdb11 39#include "smalloc.h"
4f5af7b2 40#include "verify.h"
7c9b1bce 41#include "diskutil.h"
ebac4655 42
cfc99db7
JA
43unsigned long page_mask;
44unsigned long page_size;
29d610e1
JA
45#define ALIGN(buf) \
46 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
47
48int groupid = 0;
49int thread_number = 0;
9cedf167
JA
50int nr_process = 0;
51int nr_thread = 0;
ebac4655 52int shm_id = 0;
53cdc686 53int temp_stall_ts;
d3eeeabc 54unsigned long done_secs = 0;
ebac4655 55
cdd18ad8 56static struct fio_mutex *startup_mutex;
d2bbb8d9 57static struct fio_mutex *writeout_mutex;
6ce15a32 58static volatile int fio_abort;
437c9b71 59static int exit_value;
ccb0fa24 60static struct itimerval itimer;
be4ecfdf 61static pthread_t gtod_thread;
ebac4655 62
bb3884d8
JA
63struct io_log *agg_io_log[2];
64
ebac4655 65#define TERMINATE_ALL (-1)
75154845 66#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 67
b29ee5b3 68void td_set_runstate(struct thread_data *td, int runstate)
6ce15a32 69{
d8fab1b6
JA
70 if (td->runstate == runstate)
71 return;
72
5921e80c
JA
73 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
74 td->runstate, runstate);
6ce15a32
JA
75 td->runstate = runstate;
76}
77
390c40e2 78static void terminate_threads(int group_id)
ebac4655 79{
34572e28 80 struct thread_data *td;
ebac4655
JA
81 int i;
82
6f88bcb0
JA
83 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
84
34572e28 85 for_each_td(td, i) {
ebac4655 86 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
6f88bcb0 87 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
5921e80c 88 td->o.name, (int) td->pid);
ad830ed7
JA
89 td->terminate = 1;
90 td->o.start_delay = 0;
91
7a93ab19
JA
92 /*
93 * if the thread is running, just let it exit
94 */
95 if (td->runstate < TD_RUNNING)
c1302d44 96 kill(td->pid, SIGQUIT);
f63f7f3b
JA
97 else {
98 struct ioengine_ops *ops = td->io_ops;
99
100 if (ops && (ops->flags & FIO_SIGQUIT))
101 kill(td->pid, SIGQUIT);
102 }
ebac4655
JA
103 }
104 }
105}
106
ccb0fa24
JA
107static void status_timer_arm(void)
108{
109 itimer.it_value.tv_sec = 0;
110 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
111 setitimer(ITIMER_REAL, &itimer, NULL);
112}
113
02444ad1 114static void sig_alrm(int fio_unused sig)
ebac4655 115{
697a606c 116 if (threads) {
5ec10eaa 117 update_io_ticks();
5ec10eaa 118 print_thread_status();
ccb0fa24 119 status_timer_arm();
697a606c
JA
120 }
121}
122
5f8f9726
JA
123/*
124 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
125 */
126static void sig_quit(int sig)
127{
128}
129
697a606c
JA
130static void sig_int(int sig)
131{
132 if (threads) {
5ec10eaa
JA
133 printf("\nfio: terminating on signal %d\n", sig);
134 fflush(stdout);
135 terminate_threads(TERMINATE_ALL);
ebac4655
JA
136 }
137}
138
02444ad1 139static void sig_ill(int fio_unused sig)
de79c915
JA
140{
141 if (!threads)
142 return;
143
144 log_err("fio: illegal instruction. your cpu does not support "
145 "the sse4.2 instruction for crc32c\n");
146 terminate_threads(TERMINATE_ALL);
147 exit(4);
148}
149
697a606c
JA
150static void set_sig_handlers(void)
151{
152 struct sigaction act;
153
154 memset(&act, 0, sizeof(act));
155 act.sa_handler = sig_alrm;
156 act.sa_flags = SA_RESTART;
157 sigaction(SIGALRM, &act, NULL);
158
159 memset(&act, 0, sizeof(act));
160 act.sa_handler = sig_int;
161 act.sa_flags = SA_RESTART;
162 sigaction(SIGINT, &act, NULL);
de79c915
JA
163
164 memset(&act, 0, sizeof(act));
165 act.sa_handler = sig_ill;
166 act.sa_flags = SA_RESTART;
167 sigaction(SIGILL, &act, NULL);
5f8f9726
JA
168
169 memset(&act, 0, sizeof(act));
170 act.sa_handler = sig_quit;
171 act.sa_flags = SA_RESTART;
172 sigaction(SIGQUIT, &act, NULL);
697a606c
JA
173}
174
906c8d75
JA
175/*
176 * Check if we are above the minimum rate given.
177 */
581e7141
JA
178static int __check_min_rate(struct thread_data *td, struct timeval *now,
179 enum td_ddir ddir)
ebac4655 180{
0904200b 181 unsigned long long bytes = 0;
4e991c23 182 unsigned long iops = 0;
ebac4655
JA
183 unsigned long spent;
184 unsigned long rate;
581e7141
JA
185 unsigned int ratemin = 0;
186 unsigned int rate_iops = 0;
187 unsigned int rate_iops_min = 0;
ebac4655
JA
188
189 /*
190 * allow a 2 second settle period in the beginning
191 */
192 if (mtime_since(&td->start, now) < 2000)
193 return 0;
194
581e7141
JA
195 iops += td->io_blocks[ddir];
196 bytes += td->this_io_bytes[ddir];
197 ratemin += td->o.ratemin[ddir];
198 rate_iops += td->o.rate_iops[ddir];
199 rate_iops_min += td->o.rate_iops_min[ddir];
0904200b 200
ebac4655
JA
201 /*
202 * if rate blocks is set, sample is running
203 */
581e7141
JA
204 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
205 spent = mtime_since(&td->lastrate[ddir], now);
2dc1bbeb 206 if (spent < td->o.ratecycle)
ebac4655
JA
207 return 0;
208
581e7141 209 if (td->o.rate[ddir]) {
4e991c23
JA
210 /*
211 * check bandwidth specified rate
212 */
581e7141 213 if (bytes < td->rate_bytes[ddir]) {
5ec10eaa 214 log_err("%s: min rate %u not met\n", td->o.name,
581e7141 215 ratemin);
4e991c23
JA
216 return 1;
217 } else {
581e7141
JA
218 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
219 if (rate < ratemin ||
220 bytes < td->rate_bytes[ddir]) {
5ec10eaa
JA
221 log_err("%s: min rate %u not met, got"
222 " %luKiB/sec\n", td->o.name,
581e7141 223 ratemin, rate);
4e991c23
JA
224 return 1;
225 }
226 }
413dd459 227 } else {
4e991c23
JA
228 /*
229 * checks iops specified rate
230 */
581e7141 231 if (iops < rate_iops) {
5ec10eaa 232 log_err("%s: min iops rate %u not met\n",
581e7141 233 td->o.name, rate_iops);
413dd459 234 return 1;
4e991c23 235 } else {
581e7141
JA
236 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
237 if (rate < rate_iops_min ||
238 iops < td->rate_blocks[ddir]) {
5ec10eaa
JA
239 log_err("%s: min iops rate %u not met,"
240 " got %lu\n", td->o.name,
581e7141 241 rate_iops_min, rate);
4e991c23 242 }
413dd459 243 }
ebac4655
JA
244 }
245 }
246
581e7141
JA
247 td->rate_bytes[ddir] = bytes;
248 td->rate_blocks[ddir] = iops;
249 memcpy(&td->lastrate[ddir], now, sizeof(*now));
ebac4655
JA
250 return 0;
251}
252
581e7141
JA
253static int check_min_rate(struct thread_data *td, struct timeval *now,
254 unsigned long *bytes_done)
255{
256 int ret = 0;
257
258 if (bytes_done[0])
259 ret |= __check_min_rate(td, now, 0);
260 if (bytes_done[1])
261 ret |= __check_min_rate(td, now, 1);
262
263 return ret;
264}
265
ebac4655
JA
266static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
267{
2dc1bbeb 268 if (!td->o.timeout)
ebac4655 269 return 0;
2dc1bbeb 270 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
271 return 1;
272
273 return 0;
274}
275
906c8d75
JA
276/*
277 * When job exits, we can cancel the in-flight IO if we are using async
278 * io. Attempt to do so.
279 */
ebac4655
JA
280static void cleanup_pending_aio(struct thread_data *td)
281{
01743ee1 282 struct flist_head *entry, *n;
ebac4655
JA
283 struct io_u *io_u;
284 int r;
285
286 /*
287 * get immediately available events, if any
288 */
581e7141 289 r = io_u_queued_complete(td, 0, NULL);
b2fdda43
JA
290 if (r < 0)
291 return;
ebac4655
JA
292
293 /*
294 * now cancel remaining active events
295 */
2866c82d 296 if (td->io_ops->cancel) {
01743ee1
JA
297 flist_for_each_safe(entry, n, &td->io_u_busylist) {
298 io_u = flist_entry(entry, struct io_u, list);
ebac4655 299
0c6e7517
JA
300 /*
301 * if the io_u isn't in flight, then that generally
302 * means someone leaked an io_u. complain but fix
303 * it up, so we don't stall here.
304 */
305 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
306 log_err("fio: non-busy IO on busy list\n");
ebac4655 307 put_io_u(td, io_u);
0c6e7517
JA
308 } else {
309 r = td->io_ops->cancel(td, io_u);
310 if (!r)
311 put_io_u(td, io_u);
312 }
ebac4655
JA
313 }
314 }
315
97601024 316 if (td->cur_depth)
581e7141 317 r = io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
318}
319
858a3d47
JA
320/*
321 * Helper to handle the final sync of a file. Works just like the normal
322 * io path, just does everything sync.
323 */
324static int fio_io_sync(struct thread_data *td, struct fio_file *f)
325{
326 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
327 int ret;
328
329 if (!io_u)
330 return 1;
331
332 io_u->ddir = DDIR_SYNC;
333 io_u->file = f;
334
335 if (td_io_prep(td, io_u)) {
336 put_io_u(td, io_u);
337 return 1;
338 }
339
755200a3 340requeue:
858a3d47 341 ret = td_io_queue(td, io_u);
36167d82 342 if (ret < 0) {
e1161c32 343 td_verror(td, io_u->error, "td_io_queue");
858a3d47 344 put_io_u(td, io_u);
858a3d47 345 return 1;
36167d82 346 } else if (ret == FIO_Q_QUEUED) {
581e7141 347 if (io_u_queued_complete(td, 1, NULL) < 0)
36167d82 348 return 1;
36167d82
JA
349 } else if (ret == FIO_Q_COMPLETED) {
350 if (io_u->error) {
e1161c32 351 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
352 return 1;
353 }
858a3d47 354
581e7141 355 if (io_u_sync_complete(td, io_u, NULL) < 0)
b2fdda43 356 return 1;
755200a3
JA
357 } else if (ret == FIO_Q_BUSY) {
358 if (td_io_commit(td))
359 return 1;
360 goto requeue;
858a3d47
JA
361 }
362
363 return 0;
364}
365
993bf48b
JA
366static inline void update_tv_cache(struct thread_data *td)
367{
368 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
369 fio_gettime(&td->tv_cache, NULL);
370}
371
906c8d75 372/*
34403fb1 373 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
374 * reads the blocks back in, and checks the crc/md5 of the data.
375 */
1e97cce9 376static void do_verify(struct thread_data *td)
ebac4655 377{
53cdc686 378 struct fio_file *f;
36167d82 379 struct io_u *io_u;
af52b345
JA
380 int ret, min_events;
381 unsigned int i;
e5b401d4
JA
382
383 /*
384 * sync io first and invalidate cache, to make sure we really
385 * read from disk.
386 */
387 for_each_file(td, f, i) {
d6aed795 388 if (!fio_file_open(f))
3d7b485f 389 continue;
b2fdda43
JA
390 if (fio_io_sync(td, f))
391 break;
392 if (file_invalidate_cache(td, f))
393 break;
e5b401d4 394 }
ebac4655 395
b2fdda43
JA
396 if (td->error)
397 return;
398
ebac4655
JA
399 td_set_runstate(td, TD_VERIFYING);
400
36167d82
JA
401 io_u = NULL;
402 while (!td->terminate) {
4950421a 403 int ret2, full;
5451792e 404
ebac4655
JA
405 io_u = __get_io_u(td);
406 if (!io_u)
407 break;
408
993bf48b
JA
409 update_tv_cache(td);
410
411 if (runtime_exceeded(td, &td->tv_cache)) {
069c2918 412 put_io_u(td, io_u);
90a87d4b 413 td->terminate = 1;
02bcaa8c 414 break;
069c2918 415 }
02bcaa8c 416
069c2918
JA
417 if (get_next_verify(td, io_u)) {
418 put_io_u(td, io_u);
ebac4655 419 break;
069c2918 420 }
ebac4655 421
069c2918
JA
422 if (td_io_prep(td, io_u)) {
423 put_io_u(td, io_u);
53cdc686 424 break;
069c2918 425 }
d7762cf8
JA
426
427 io_u->end_io = verify_io_u;
53cdc686 428
11786802 429 ret = td_io_queue(td, io_u);
36167d82
JA
430 switch (ret) {
431 case FIO_Q_COMPLETED:
432 if (io_u->error)
22819ec2 433 ret = -io_u->error;
9e9d164e 434 else if (io_u->resid) {
36167d82 435 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 436 struct fio_file *f = io_u->file;
ebac4655 437
9e9d164e
JA
438 /*
439 * zero read, fail
440 */
441 if (!bytes) {
41d9f0fd 442 td_verror(td, EIO, "full resid");
9e9d164e
JA
443 put_io_u(td, io_u);
444 break;
445 }
8400d9b2 446
36167d82
JA
447 io_u->xfer_buflen = io_u->resid;
448 io_u->xfer_buf += bytes;
8400d9b2
JA
449 io_u->offset += bytes;
450
30061b97
JA
451 td->ts.short_io_u[io_u->ddir]++;
452
d460eb31 453 if (io_u->offset == f->real_file_size)
8400d9b2
JA
454 goto sync_done;
455
11786802
JA
456 requeue_io_u(td, &io_u);
457 } else {
8400d9b2 458sync_done:
581e7141 459 ret = io_u_sync_complete(td, io_u, NULL);
11786802
JA
460 if (ret < 0)
461 break;
36167d82 462 }
36167d82
JA
463 continue;
464 case FIO_Q_QUEUED:
465 break;
755200a3
JA
466 case FIO_Q_BUSY:
467 requeue_io_u(td, &io_u);
5451792e
JA
468 ret2 = td_io_commit(td);
469 if (ret2 < 0)
470 ret = ret2;
755200a3 471 break;
36167d82
JA
472 default:
473 assert(ret < 0);
e1161c32 474 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
475 break;
476 }
477
99784632 478 if (ret < 0 || td->error)
3af6ef39
JA
479 break;
480
ebac4655 481 /*
3af6ef39
JA
482 * if we can queue more, do so. but check if there are
483 * completed io_u's first.
ebac4655 484 */
4950421a
JA
485 full = queue_full(td) || ret == FIO_Q_BUSY;
486 if (full || !td->o.iodepth_batch_complete) {
487 min_events = td->o.iodepth_batch_complete;
488 if (full && !min_events)
838bc709 489 min_events = 1;
e916b390 490
4950421a
JA
491 do {
492 /*
493 * Reap required number of io units, if any,
494 * and do the verification on them through
495 * the callback handler
496 */
581e7141 497 if (io_u_queued_complete(td, min_events, NULL) < 0) {
4950421a
JA
498 ret = -1;
499 break;
500 }
501 } while (full && (td->cur_depth > td->o.iodepth_low));
502 }
503 if (ret < 0)
ebac4655 504 break;
36167d82 505 }
ebac4655 506
c01c0395
JA
507 if (!td->error) {
508 min_events = td->cur_depth;
509
510 if (min_events)
581e7141 511 ret = io_u_queued_complete(td, min_events, NULL);
c01c0395 512 } else
ebac4655
JA
513 cleanup_pending_aio(td);
514
515 td_set_runstate(td, TD_RUNNING);
516}
517
32cd46a0 518/*
906c8d75 519 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
520 * and reaps them, checking for rate and errors along the way.
521 */
ebac4655
JA
522static void do_io(struct thread_data *td)
523{
af52b345
JA
524 unsigned int i;
525 int ret = 0;
ebac4655 526
b29ee5b3
JA
527 if (in_ramp_time(td))
528 td_set_runstate(td, TD_RAMP);
529 else
530 td_set_runstate(td, TD_RUNNING);
5853e5a8 531
7bb48f84 532 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
97601024 533 struct timeval comp_time;
581e7141 534 unsigned long bytes_done[2] = { 0, 0 };
84585003 535 int min_evts = 0;
ebac4655 536 struct io_u *io_u;
4950421a 537 int ret2, full;
ebac4655
JA
538
539 if (td->terminate)
540 break;
541
3d7c391d 542 io_u = get_io_u(td);
ebac4655
JA
543 if (!io_u)
544 break;
545
993bf48b 546 update_tv_cache(td);
97601024 547
993bf48b 548 if (runtime_exceeded(td, &td->tv_cache)) {
97601024 549 put_io_u(td, io_u);
90a87d4b 550 td->terminate = 1;
97601024
JA
551 break;
552 }
36167d82 553
b67740d3
JA
554 /*
555 * Add verification end_io handler, if asked to verify
556 * a previously written file.
557 */
cc3fc855 558 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
b67740d3 559 io_u->end_io = verify_io_u;
d8fab1b6 560 td_set_runstate(td, TD_VERIFYING);
b29ee5b3
JA
561 } else if (in_ramp_time(td))
562 td_set_runstate(td, TD_RAMP);
563 else
d8fab1b6 564 td_set_runstate(td, TD_RUNNING);
b67740d3 565
11786802 566 ret = td_io_queue(td, io_u);
36167d82
JA
567 switch (ret) {
568 case FIO_Q_COMPLETED:
5451792e
JA
569 if (io_u->error)
570 ret = -io_u->error;
9e9d164e 571 else if (io_u->resid) {
36167d82 572 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 573 struct fio_file *f = io_u->file;
36167d82 574
9e9d164e
JA
575 /*
576 * zero read, fail
577 */
578 if (!bytes) {
41d9f0fd 579 td_verror(td, EIO, "full resid");
9e9d164e
JA
580 put_io_u(td, io_u);
581 break;
582 }
583
cec6b55d 584 io_u->xfer_buflen = io_u->resid;
36167d82 585 io_u->xfer_buf += bytes;
5a7c5680
JA
586 io_u->offset += bytes;
587
30061b97
JA
588 td->ts.short_io_u[io_u->ddir]++;
589
d460eb31 590 if (io_u->offset == f->real_file_size)
5a7c5680
JA
591 goto sync_done;
592
11786802
JA
593 requeue_io_u(td, &io_u);
594 } else {
5a7c5680 595sync_done:
581e7141
JA
596 if (__should_check_rate(td, 0) ||
597 __should_check_rate(td, 1))
993bf48b
JA
598 fio_gettime(&comp_time, NULL);
599
581e7141
JA
600 ret = io_u_sync_complete(td, io_u, bytes_done);
601 if (ret < 0)
602 break;
cec6b55d 603 }
36167d82
JA
604 break;
605 case FIO_Q_QUEUED:
7e77dd02
JA
606 /*
607 * if the engine doesn't have a commit hook,
608 * the io_u is really queued. if it does have such
609 * a hook, it has to call io_u_queued() itself.
610 */
611 if (td->io_ops->commit == NULL)
612 io_u_queued(td, io_u);
36167d82 613 break;
755200a3
JA
614 case FIO_Q_BUSY:
615 requeue_io_u(td, &io_u);
5451792e
JA
616 ret2 = td_io_commit(td);
617 if (ret2 < 0)
618 ret = ret2;
755200a3 619 break;
36167d82
JA
620 default:
621 assert(ret < 0);
622 put_io_u(td, io_u);
623 break;
ebac4655
JA
624 }
625
99784632 626 if (ret < 0 || td->error)
36167d82
JA
627 break;
628
97601024
JA
629 /*
630 * See if we need to complete some commands
631 */
4950421a
JA
632 full = queue_full(td) || ret == FIO_Q_BUSY;
633 if (full || !td->o.iodepth_batch_complete) {
634 min_evts = td->o.iodepth_batch_complete;
635 if (full && !min_evts)
36167d82 636 min_evts = 1;
4950421a 637
581e7141
JA
638 if (__should_check_rate(td, 0) ||
639 __should_check_rate(td, 1))
993bf48b 640 fio_gettime(&comp_time, NULL);
4950421a
JA
641
642 do {
581e7141
JA
643 ret = io_u_queued_complete(td, min_evts, bytes_done);
644 if (ret < 0)
4950421a
JA
645 break;
646
4950421a 647 } while (full && (td->cur_depth > td->o.iodepth_low));
ebac4655
JA
648 }
649
4950421a
JA
650 if (ret < 0)
651 break;
581e7141 652 if (!(bytes_done[0] + bytes_done[1]))
97601024
JA
653 continue;
654
581e7141
JA
655 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
656 if (check_min_rate(td, &comp_time, bytes_done)) {
721938ae
JA
657 if (exitall_on_terminate)
658 terminate_threads(td->groupid);
659 td_verror(td, EIO, "check_min_rate");
660 break;
661 }
ebac4655
JA
662 }
663
2dc1bbeb 664 if (td->o.thinktime) {
9c1f7434
JA
665 unsigned long long b;
666
667 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 668 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
669 int left;
670
2dc1bbeb 671 if (td->o.thinktime_spin)
6dd6f2cd 672 usec_spin(td->o.thinktime_spin);
48097d5c 673
2dc1bbeb 674 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
675 if (left)
676 usec_sleep(td, left);
677 }
9c1f7434 678 }
ebac4655
JA
679 }
680
aa31f1f1
SL
681 if (td->o.fill_device && td->error == ENOSPC) {
682 td->error = 0;
683 td->terminate = 1;
684 }
4d2413c6 685 if (!td->error) {
3d7c391d
JA
686 struct fio_file *f;
687
c01c0395
JA
688 i = td->cur_depth;
689 if (i)
581e7141 690 ret = io_u_queued_complete(td, i, NULL);
ebac4655 691
2dc1bbeb 692 if (should_fsync(td) && td->o.end_fsync) {
84585003 693 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
694
695 for_each_file(td, f, i) {
d6aed795 696 if (!fio_file_open(f))
3d7b485f 697 continue;
858a3d47 698 fio_io_sync(td, f);
3d7b485f 699 }
84585003 700 }
c01c0395
JA
701 } else
702 cleanup_pending_aio(td);
2ba1c290
JA
703
704 /*
705 * stop job if we failed doing any IO
706 */
707 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
708 td->done = 1;
ebac4655
JA
709}
710
ebac4655
JA
711static void cleanup_io_u(struct thread_data *td)
712{
01743ee1 713 struct flist_head *entry, *n;
ebac4655
JA
714 struct io_u *io_u;
715
01743ee1
JA
716 flist_for_each_safe(entry, n, &td->io_u_freelist) {
717 io_u = flist_entry(entry, struct io_u, list);
ebac4655 718
01743ee1 719 flist_del(&io_u->list);
ebac4655
JA
720 free(io_u);
721 }
722
2f9ade3c 723 free_io_mem(td);
ebac4655
JA
724}
725
726static int init_io_u(struct thread_data *td)
727{
728 struct io_u *io_u;
a00735e6 729 unsigned int max_bs;
eb7ccf38 730 int cl_align, i, max_units;
ebac4655
JA
731 char *p;
732
1e3c09af 733 max_units = td->o.iodepth;
2dc1bbeb 734 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
5ec10eaa
JA
735 td->orig_buffer_size = (unsigned long long) max_bs
736 * (unsigned long long) max_units;
737
738 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
739 unsigned long bs;
74b025b0 740
5ec10eaa
JA
741 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
742 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
743 }
ebac4655 744
c3990645
JA
745 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
746 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
747 return 1;
748 }
749
2f9ade3c
JA
750 if (allocate_io_mem(td))
751 return 1;
ebac4655 752
b3f378e9
JA
753 if (td->o.odirect)
754 p = ALIGN(td->orig_buffer);
755 else
756 p = td->orig_buffer;
757
eb7ccf38
JA
758 cl_align = os_cache_line_size();
759
ebac4655 760 for (i = 0; i < max_units; i++) {
eb7ccf38
JA
761 void *ptr;
762
db1cd90b
JA
763 if (td->terminate)
764 return 1;
eb7ccf38
JA
765
766 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
767 log_err("fio: posix_memalign=%s\n", strerror(errno));
768 break;
769 }
770
771 io_u = ptr;
ebac4655 772 memset(io_u, 0, sizeof(*io_u));
01743ee1 773 INIT_FLIST_HEAD(&io_u->list);
ebac4655 774
b4c5e1ac
JA
775 if (!(td->io_ops->flags & FIO_NOIO)) {
776 io_u->buf = p + max_bs * i;
e9459e5a 777
5973cafb
JA
778 if (td_write(td) && !td->o.refill_buffers)
779 io_u_fill_buffer(td, io_u, max_bs);
b4c5e1ac 780 }
6b9cea23 781
b1ff3403 782 io_u->index = i;
0c6e7517 783 io_u->flags = IO_U_F_FREE;
01743ee1 784 flist_add(&io_u->list, &td->io_u_freelist);
ebac4655
JA
785 }
786
787 return 0;
788}
789
da86774e
JA
790static int switch_ioscheduler(struct thread_data *td)
791{
792 char tmp[256], tmp2[128];
793 FILE *f;
794 int ret;
795
ba0fbe10 796 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
797 return 0;
798
da86774e
JA
799 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
800
801 f = fopen(tmp, "r+");
802 if (!f) {
cbf5c1d7 803 if (errno == ENOENT) {
5ec10eaa
JA
804 log_err("fio: os or kernel doesn't support IO scheduler"
805 " switching\n");
cbf5c1d7
JA
806 return 0;
807 }
808 td_verror(td, errno, "fopen iosched");
da86774e
JA
809 return 1;
810 }
811
812 /*
813 * Set io scheduler.
814 */
2dc1bbeb 815 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 816 if (ferror(f) || ret != 1) {
e1161c32 817 td_verror(td, errno, "fwrite");
da86774e
JA
818 fclose(f);
819 return 1;
820 }
821
822 rewind(f);
823
824 /*
825 * Read back and check that the selected scheduler is now the default.
826 */
827 ret = fread(tmp, 1, sizeof(tmp), f);
828 if (ferror(f) || ret < 0) {
e1161c32 829 td_verror(td, errno, "fread");
da86774e
JA
830 fclose(f);
831 return 1;
832 }
833
2dc1bbeb 834 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 835 if (!strstr(tmp, tmp2)) {
2dc1bbeb 836 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 837 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
838 fclose(f);
839 return 1;
840 }
841
842 fclose(f);
843 return 0;
844}
845
492158cf
JA
846static int keep_running(struct thread_data *td)
847{
848 unsigned long long io_done;
849
20e354ef
JA
850 if (td->done)
851 return 0;
492158cf
JA
852 if (td->o.time_based)
853 return 1;
854 if (td->o.loops) {
855 td->o.loops--;
856 return 1;
857 }
858
5ec10eaa
JA
859 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
860 + td->io_skip_bytes;
492158cf
JA
861 if (io_done < td->o.size)
862 return 1;
863
864 return 0;
865}
866
b29ee5b3 867static void reset_io_counters(struct thread_data *td)
ebac4655 868{
756867bd 869 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 870 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 871 td->zone_bytes = 0;
581e7141
JA
872 td->rate_bytes[0] = td->rate_bytes[1] = 0;
873 td->rate_blocks[0] = td->rate_blocks[1] = 0;
ebac4655 874
c1324df1
JA
875 td->last_was_sync = 0;
876
2ba1c290
JA
877 /*
878 * reset file done count if we are to start over
879 */
880 if (td->o.time_based || td->o.loops)
48f5abd3 881 td->nr_done_files = 0;
5bfc35d7
JA
882
883 /*
884 * Set the same seed to get repeatable runs
885 */
886 td_fill_rand_seeds(td);
b29ee5b3
JA
887}
888
889void reset_all_stats(struct thread_data *td)
890{
891 struct timeval tv;
892 int i;
893
894 reset_io_counters(td);
895
896 for (i = 0; i < 2; i++) {
897 td->io_bytes[i] = 0;
898 td->io_blocks[i] = 0;
899 td->io_issues[i] = 0;
900 td->ts.total_io_u[i] = 0;
901 }
902
903 fio_gettime(&tv, NULL);
904 memcpy(&td->epoch, &tv, sizeof(tv));
905 memcpy(&td->start, &tv, sizeof(tv));
906}
907
df9c26b1 908static void clear_io_state(struct thread_data *td)
b29ee5b3
JA
909{
910 struct fio_file *f;
911 unsigned int i;
b29ee5b3
JA
912
913 reset_io_counters(td);
281f9b63 914
24ffd2c2 915 close_files(td);
df9c26b1 916 for_each_file(td, f, i)
d6aed795 917 fio_file_clear_done(f);
ebac4655
JA
918}
919
398e8c4d
JA
920static int exec_string(const char *string)
921{
922 int ret, newlen = strlen(string) + 1 + 8;
923 char *str;
924
925 str = malloc(newlen);
926 sprintf(str, "sh -c %s", string);
927
928 ret = system(str);
929 if (ret == -1)
930 log_err("fio: exec of cmd <%s> failed\n", str);
931
932 free(str);
933 return ret;
934}
935
906c8d75
JA
936/*
937 * Entry point for the thread based jobs. The process based jobs end up
938 * here as well, after a little setup.
939 */
ebac4655
JA
940static void *thread_main(void *data)
941{
10927316 942 unsigned long long runtime[2], elapsed;
ebac4655 943 struct thread_data *td = data;
a978ba68 944 int clear_state;
ebac4655 945
2dc1bbeb 946 if (!td->o.use_thread)
ebac4655
JA
947 setsid();
948
949 td->pid = getpid();
950
5921e80c 951 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
ee56ad50 952
01743ee1
JA
953 INIT_FLIST_HEAD(&td->io_u_freelist);
954 INIT_FLIST_HEAD(&td->io_u_busylist);
955 INIT_FLIST_HEAD(&td->io_u_requeues);
956 INIT_FLIST_HEAD(&td->io_log_list);
957 INIT_FLIST_HEAD(&td->io_hist_list);
bb5d7d0b 958 td->io_hist_tree = RB_ROOT;
aea47d44 959
db1cd90b 960 td_set_runstate(td, TD_INITIALIZED);
29adda3c 961 dprint(FD_MUTEX, "up startup_mutex\n");
cdd18ad8 962 fio_mutex_up(startup_mutex);
29adda3c 963 dprint(FD_MUTEX, "wait on td->mutex\n");
cdd18ad8 964 fio_mutex_down(td->mutex);
29adda3c 965 dprint(FD_MUTEX, "done waiting on td->mutex\n");
db1cd90b
JA
966
967 /*
cdd18ad8 968 * the ->mutex mutex is now no longer used, close it to avoid
db1cd90b
JA
969 * eating a file descriptor
970 */
cdd18ad8 971 fio_mutex_remove(td->mutex);
db1cd90b 972
d84f8d49
JA
973 /*
974 * May alter parameters that init_io_u() will use, so we need to
975 * do this first.
976 */
977 if (init_iolog(td))
978 goto err;
979
ebac4655 980 if (init_io_u(td))
db1cd90b 981 goto err;
ebac4655 982
375b2695 983 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
e1161c32 984 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 985 goto err;
ebac4655
JA
986 }
987
f484470b
JA
988 /*
989 * If we have a gettimeofday() thread, make sure we exclude that
990 * thread from this job
991 */
be4ecfdf 992 if (td->o.gtod_cpu) {
be4ecfdf
JA
993 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
994 if (fio_setaffinity(td) == -1) {
995 td_verror(td, errno, "cpu_set_affinity");
996 goto err;
997 }
998 }
999
ac684785 1000 if (td->ioprio_set) {
ebac4655 1001 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 1002 td_verror(td, errno, "ioprio_set");
db1cd90b 1003 goto err;
ebac4655
JA
1004 }
1005 }
1006
2dc1bbeb 1007 if (nice(td->o.nice) == -1) {
e1161c32 1008 td_verror(td, errno, "nice");
db1cd90b 1009 goto err;
b6f4d880
JA
1010 }
1011
2dc1bbeb 1012 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 1013 goto err;
37f56873 1014
2dc1bbeb 1015 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
1016 goto err;
1017
7d6c5283
JA
1018 if (td_io_init(td))
1019 goto err;
1020
68727076
JA
1021 if (init_random_map(td))
1022 goto err;
1023
2dc1bbeb 1024 if (td->o.exec_prerun) {
398e8c4d 1025 if (exec_string(td->o.exec_prerun))
69cfd7e0
JA
1026 goto err;
1027 }
4e0ba8af 1028
afad68f7
ZY
1029 if (td->o.pre_read) {
1030 if (pre_read_files(td) < 0)
1031 goto err;
1032 }
1033
69008999 1034 fio_gettime(&td->epoch, NULL);
756867bd 1035 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
1036
1037 runtime[0] = runtime[1] = 0;
a978ba68 1038 clear_state = 0;
492158cf 1039 while (keep_running(td)) {
02bcaa8c 1040 fio_gettime(&td->start, NULL);
756867bd 1041 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
993bf48b 1042 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
ebac4655 1043
2dc1bbeb 1044 if (td->o.ratemin)
5ec10eaa
JA
1045 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1046 sizeof(td->lastrate));
ebac4655 1047
df9c26b1
JA
1048 if (clear_state)
1049 clear_io_state(td);
a978ba68 1050
ebac4655
JA
1051 prune_io_piece_log(td);
1052
ba0fbe10 1053 do_io(td);
ebac4655 1054
a978ba68
JA
1055 clear_state = 1;
1056
38d77cae 1057 if (td_read(td) && td->io_bytes[DDIR_READ]) {
0a4957c6 1058 elapsed = utime_since_now(&td->start);
38d77cae
JA
1059 runtime[DDIR_READ] += elapsed;
1060 }
1061 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
0a4957c6 1062 elapsed = utime_since_now(&td->start);
38d77cae
JA
1063 runtime[DDIR_WRITE] += elapsed;
1064 }
5ec10eaa 1065
ebac4655
JA
1066 if (td->error || td->terminate)
1067 break;
1068
e84c73a8
SL
1069 if (!td->o.do_verify ||
1070 td->o.verify == VERIFY_NONE ||
b67740d3 1071 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
1072 continue;
1073
df9c26b1 1074 clear_io_state(td);
a978ba68 1075
02bcaa8c 1076 fio_gettime(&td->start, NULL);
ebac4655
JA
1077
1078 do_verify(td);
1079
69008999 1080 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
1081
1082 if (td->error || td->terminate)
1083 break;
1084 }
1085
36dff966 1086 update_rusage_stat(td);
47f0cc48
YHJT
1087 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1088 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
756867bd
JA
1089 td->ts.total_run_time = mtime_since_now(&td->epoch);
1090 td->ts.io_bytes[0] = td->io_bytes[0];
1091 td->ts.io_bytes[1] = td->io_bytes[1];
1092
d2bbb8d9 1093 fio_mutex_down(writeout_mutex);
e3cedca7 1094 if (td->ts.bw_log) {
f484470b
JA
1095 if (td->o.bw_log_file) {
1096 finish_log_named(td, td->ts.bw_log,
1097 td->o.bw_log_file, "bw");
1098 } else
e3cedca7
JA
1099 finish_log(td, td->ts.bw_log, "bw");
1100 }
1101 if (td->ts.slat_log) {
f484470b
JA
1102 if (td->o.lat_log_file) {
1103 finish_log_named(td, td->ts.slat_log,
1ca95738 1104 td->o.lat_log_file, "slat");
f484470b 1105 } else
e3cedca7
JA
1106 finish_log(td, td->ts.slat_log, "slat");
1107 }
1108 if (td->ts.clat_log) {
f484470b
JA
1109 if (td->o.lat_log_file) {
1110 finish_log_named(td, td->ts.clat_log,
1111 td->o.lat_log_file, "clat");
1112 } else
e3cedca7
JA
1113 finish_log(td, td->ts.clat_log, "clat");
1114 }
d2bbb8d9 1115 fio_mutex_up(writeout_mutex);
398e8c4d
JA
1116 if (td->o.exec_postrun)
1117 exec_string(td->o.exec_postrun);
ebac4655
JA
1118
1119 if (exitall_on_terminate)
390c40e2 1120 terminate_threads(td->groupid);
ebac4655
JA
1121
1122err:
5bf13a5a 1123 if (td->error)
5921e80c 1124 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
5ec10eaa 1125 td->verror);
24ffd2c2 1126 close_and_free_files(td);
2866c82d 1127 close_ioengine(td);
ebac4655 1128 cleanup_io_u(td);
f29b25a3 1129
d2ce18b5
JA
1130 if (td->o.cpumask_set) {
1131 int ret = fio_cpuset_exit(&td->o.cpumask);
1132
1133 td_verror(td, ret, "fio_cpuset_exit");
1134 }
4e78e405 1135
f29b25a3
JA
1136 /*
1137 * do this very late, it will log file closing as well
1138 */
1139 if (td->o.write_iolog_file)
1140 write_iolog_close(td);
1141
d23bb327 1142 options_mem_free(td);
ebac4655 1143 td_set_runstate(td, TD_EXITED);
43d76807 1144 return (void *) (unsigned long) td->error;
ebac4655
JA
1145}
1146
906c8d75
JA
1147/*
1148 * We cannot pass the td data into a forked process, so attach the td and
1149 * pass it to the thread worker.
1150 */
a6418147 1151static int fork_main(int shmid, int offset)
ebac4655
JA
1152{
1153 struct thread_data *td;
a6418147 1154 void *data, *ret;
ebac4655
JA
1155
1156 data = shmat(shmid, NULL, 0);
1157 if (data == (void *) -1) {
a6418147
JA
1158 int __err = errno;
1159
ebac4655 1160 perror("shmat");
a6418147 1161 return __err;
ebac4655
JA
1162 }
1163
1164 td = data + offset * sizeof(struct thread_data);
a6418147 1165 ret = thread_main(td);
ebac4655 1166 shmdt(data);
43d76807 1167 return (int) (unsigned long) ret;
ebac4655
JA
1168}
1169
906c8d75
JA
1170/*
1171 * Run over the job map and reap the threads that have exited, if any.
1172 */
ebac4655
JA
1173static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1174{
34572e28 1175 struct thread_data *td;
1f809d15 1176 int i, cputhreads, realthreads, pending, status, ret;
ebac4655
JA
1177
1178 /*
1179 * reap exited threads (TD_EXITED -> TD_REAPED)
1180 */
1f809d15 1181 realthreads = pending = cputhreads = 0;
34572e28 1182 for_each_td(td, i) {
3707f45b 1183 int flags = 0;
a2f77c9f 1184
84585003
JA
1185 /*
1186 * ->io_ops is NULL for a thread that has closed its
1187 * io engine
1188 */
ba0fbe10 1189 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0 1190 cputhreads++;
1f809d15
JA
1191 else
1192 realthreads++;
b990b5c0 1193
a1dedc1f
AC
1194 if (!td->pid) {
1195 pending++;
1196 continue;
1197 }
1198 if (td->runstate == TD_REAPED)
a2f77c9f 1199 continue;
2dc1bbeb 1200 if (td->o.use_thread) {
3707f45b
JA
1201 if (td->runstate == TD_EXITED) {
1202 td_set_runstate(td, TD_REAPED);
1203 goto reaped;
1204 }
1205 continue;
1206 }
a2f77c9f
JA
1207
1208 flags = WNOHANG;
1209 if (td->runstate == TD_EXITED)
1210 flags = 0;
1211
1212 /*
1213 * check if someone quit or got killed in an unusual way
1214 */
1215 ret = waitpid(td->pid, &status, flags);
3707f45b 1216 if (ret < 0) {
a2f77c9f 1217 if (errno == ECHILD) {
5921e80c
JA
1218 log_err("fio: pid=%d disappeared %d\n",
1219 (int) td->pid, td->runstate);
a2f77c9f
JA
1220 td_set_runstate(td, TD_REAPED);
1221 goto reaped;
1222 }
1223 perror("waitpid");
1224 } else if (ret == td->pid) {
1225 if (WIFSIGNALED(status)) {
fab6aa71
JA
1226 int sig = WTERMSIG(status);
1227
d9244941 1228 if (sig != SIGQUIT)
5ec10eaa 1229 log_err("fio: pid=%d, got signal=%d\n",
5921e80c 1230 (int) td->pid, sig);
fab6aa71
JA
1231 td_set_runstate(td, TD_REAPED);
1232 goto reaped;
1233 }
a2f77c9f
JA
1234 if (WIFEXITED(status)) {
1235 if (WEXITSTATUS(status) && !td->error)
1236 td->error = WEXITSTATUS(status);
a2f77c9f 1237
a2f77c9f
JA
1238 td_set_runstate(td, TD_REAPED);
1239 goto reaped;
a6418147
JA
1240 }
1241 }
ebac4655 1242
a2f77c9f
JA
1243 /*
1244 * thread is not dead, continue
1245 */
e48676ba 1246 pending++;
a2f77c9f 1247 continue;
fab6aa71 1248reaped:
ebac4655 1249 (*nr_running)--;
581e7141
JA
1250 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1251 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
6f88bcb0
JA
1252 if (!td->pid)
1253 pending--;
a2f77c9f
JA
1254
1255 if (td->error)
1256 exit_value++;
d3eeeabc
JA
1257
1258 done_secs += mtime_since_now(&td->epoch) / 1000;
ebac4655 1259 }
b990b5c0 1260
1f809d15 1261 if (*nr_running == cputhreads && !pending && realthreads)
390c40e2 1262 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1263}
1264
be4ecfdf
JA
1265static void *gtod_thread_main(void *data)
1266{
1267 fio_mutex_up(startup_mutex);
1268
1269 /*
1270 * As long as we have jobs around, update the clock. It would be nice
1271 * to have some way of NOT hammering that CPU with gettimeofday(),
1272 * but I'm not sure what to use outside of a simple CPU nop to relax
1273 * it - we don't want to lose precision.
1274 */
1275 while (threads) {
1276 fio_gtod_update();
1277 nop;
1278 }
1279
1280 return NULL;
1281}
1282
1283static int fio_start_gtod_thread(void)
1284{
f42c153d
JA
1285 int ret;
1286
1287 ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1288 if (ret) {
1289 log_err("Can't create gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1290 return 1;
1291 }
f42c153d
JA
1292
1293 ret = pthread_detach(gtod_thread);
1294 if (ret) {
1295 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1296 return 1;
1297 }
1298
29adda3c 1299 dprint(FD_MUTEX, "wait on startup_mutex\n");
be4ecfdf 1300 fio_mutex_down(startup_mutex);
29adda3c 1301 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
be4ecfdf
JA
1302 return 0;
1303}
1304
906c8d75
JA
1305/*
1306 * Main function for kicking off and reaping jobs, as needed.
1307 */
ebac4655
JA
1308static void run_threads(void)
1309{
ebac4655
JA
1310 struct thread_data *td;
1311 unsigned long spent;
1312 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1313
2f9ade3c
JA
1314 if (fio_pin_memory())
1315 return;
ebac4655 1316
be4ecfdf
JA
1317 if (fio_gtod_offload && fio_start_gtod_thread())
1318 return;
1319
c6ae0a5b 1320 if (!terse_output) {
9cedf167
JA
1321 printf("Starting ");
1322 if (nr_thread)
5ec10eaa
JA
1323 printf("%d thread%s", nr_thread,
1324 nr_thread > 1 ? "s" : "");
9cedf167
JA
1325 if (nr_process) {
1326 if (nr_thread)
1327 printf(" and ");
5ec10eaa
JA
1328 printf("%d process%s", nr_process,
1329 nr_process > 1 ? "es" : "");
9cedf167
JA
1330 }
1331 printf("\n");
c6ae0a5b
JA
1332 fflush(stdout);
1333 }
c04f7ec3 1334
697a606c 1335 set_sig_handlers();
4efa970e 1336
ebac4655
JA
1337 todo = thread_number;
1338 nr_running = 0;
1339 nr_started = 0;
1340 m_rate = t_rate = 0;
1341
34572e28 1342 for_each_td(td, i) {
263e529f 1343 print_status_init(td->thread_number - 1);
ebac4655 1344
2dc1bbeb 1345 if (!td->o.create_serialize) {
380cf265 1346 init_disk_util(td);
ebac4655 1347 continue;
380cf265 1348 }
ebac4655
JA
1349
1350 /*
1351 * do file setup here so it happens sequentially,
1352 * we don't want X number of threads getting their
1353 * client data interspersed on disk
1354 */
53cdc686 1355 if (setup_files(td)) {
5bf13a5a
JA
1356 exit_value++;
1357 if (td->error)
5921e80c
JA
1358 log_err("fio: pid=%d, err=%d/%s\n",
1359 (int) td->pid, td->error, td->verror);
ebac4655
JA
1360 td_set_runstate(td, TD_REAPED);
1361 todo--;
c69e8875
JA
1362 } else {
1363 struct fio_file *f;
1364 unsigned int i;
1365
1366 /*
1367 * for sharing to work, each job must always open
1368 * its own files. so close them, if we opened them
1369 * for creation
1370 */
1371 for_each_file(td, f, i)
1372 td_io_close_file(td, f);
5ec10eaa 1373 }
380cf265
JA
1374
1375 init_disk_util(td);
ebac4655
JA
1376 }
1377
a2f77c9f
JA
1378 set_genesis_time();
1379
ebac4655 1380 while (todo) {
75154845
JA
1381 struct thread_data *map[MAX_JOBS];
1382 struct timeval this_start;
1383 int this_jobs = 0, left;
1384
ebac4655
JA
1385 /*
1386 * create threads (TD_NOT_CREATED -> TD_CREATED)
1387 */
34572e28 1388 for_each_td(td, i) {
ebac4655
JA
1389 if (td->runstate != TD_NOT_CREATED)
1390 continue;
1391
1392 /*
1393 * never got a chance to start, killed by other
1394 * thread for some reason
1395 */
1396 if (td->terminate) {
1397 todo--;
1398 continue;
1399 }
1400
2dc1bbeb 1401 if (td->o.start_delay) {
263e529f 1402 spent = mtime_since_genesis();
ebac4655 1403
2dc1bbeb 1404 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1405 continue;
1406 }
1407
6f88bcb0
JA
1408 if (td->o.stonewall && (nr_started || nr_running)) {
1409 dprint(FD_PROCESS, "%s: stonewall wait\n",
1410 td->o.name);
ebac4655 1411 break;
6f88bcb0 1412 }
ebac4655 1413
75154845
JA
1414 /*
1415 * Set state to created. Thread will transition
1416 * to TD_INITIALIZED when it's done setting up.
1417 */
ebac4655 1418 td_set_runstate(td, TD_CREATED);
75154845 1419 map[this_jobs++] = td;
ebac4655
JA
1420 nr_started++;
1421
2dc1bbeb 1422 if (td->o.use_thread) {
f42c153d
JA
1423 int ret;
1424
ee56ad50 1425 dprint(FD_PROCESS, "will pthread_create\n");
f42c153d
JA
1426 ret = pthread_create(&td->thread, NULL,
1427 thread_main, td);
1428 if (ret) {
1429 log_err("pthread_create: %s\n",
1430 strerror(ret));
ebac4655 1431 nr_started--;
c8bb6faf 1432 break;
ebac4655 1433 }
f42c153d
JA
1434 ret = pthread_detach(td->thread);
1435 if (ret)
1436 log_err("pthread_detach: %s",
1437 strerror(ret));
ebac4655 1438 } else {
5e1d306e 1439 pid_t pid;
ee56ad50 1440 dprint(FD_PROCESS, "will fork\n");
5e1d306e
JA
1441 pid = fork();
1442 if (!pid) {
a6418147
JA
1443 int ret = fork_main(shm_id, i);
1444
5e1d306e
JA
1445 _exit(ret);
1446 } else if (i == fio_debug_jobno)
1447 *fio_debug_jobp = pid;
ebac4655 1448 }
29adda3c 1449 dprint(FD_MUTEX, "wait on startup_mutex\n");
cdd18ad8 1450 fio_mutex_down(startup_mutex);
29adda3c 1451 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
ebac4655
JA
1452 }
1453
1454 /*
75154845
JA
1455 * Wait for the started threads to transition to
1456 * TD_INITIALIZED.
ebac4655 1457 */
02bcaa8c 1458 fio_gettime(&this_start, NULL);
75154845 1459 left = this_jobs;
6ce15a32 1460 while (left && !fio_abort) {
75154845
JA
1461 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1462 break;
1463
1464 usleep(100000);
1465
1466 for (i = 0; i < this_jobs; i++) {
1467 td = map[i];
1468 if (!td)
1469 continue;
b6f4d880 1470 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1471 map[i] = NULL;
1472 left--;
b6f4d880
JA
1473 } else if (td->runstate >= TD_EXITED) {
1474 map[i] = NULL;
1475 left--;
1476 todo--;
1477 nr_running++; /* work-around... */
75154845
JA
1478 }
1479 }
1480 }
1481
1482 if (left) {
3b70d7e5 1483 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1484 for (i = 0; i < this_jobs; i++) {
1485 td = map[i];
1486 if (!td)
1487 continue;
1488 kill(td->pid, SIGTERM);
1489 }
1490 break;
1491 }
1492
1493 /*
b6f4d880 1494 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1495 */
34572e28 1496 for_each_td(td, i) {
75154845 1497 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1498 continue;
1499
b29ee5b3
JA
1500 if (in_ramp_time(td))
1501 td_set_runstate(td, TD_RAMP);
1502 else
1503 td_set_runstate(td, TD_RUNNING);
ebac4655
JA
1504 nr_running++;
1505 nr_started--;
581e7141
JA
1506 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1507 t_rate += td->o.rate[0] + td->o.rate[1];
75154845 1508 todo--;
cdd18ad8 1509 fio_mutex_up(td->mutex);
ebac4655
JA
1510 }
1511
1512 reap_threads(&nr_running, &t_rate, &m_rate);
1513
1514 if (todo)
1515 usleep(100000);
1516 }
1517
1518 while (nr_running) {
1519 reap_threads(&nr_running, &t_rate, &m_rate);
1520 usleep(10000);
1521 }
1522
1523 update_io_ticks();
2f9ade3c 1524 fio_unpin_memory();
ebac4655
JA
1525}
1526
ebac4655
JA
1527int main(int argc, char *argv[])
1528{
29d610e1
JA
1529 long ps;
1530
2e5cdb11
JA
1531 sinit();
1532
dbe1125e
JA
1533 /*
1534 * We need locale for number printing, if it isn't set then just
1535 * go with the US format.
1536 */
1537 if (!getenv("LC_NUMERIC"))
1538 setlocale(LC_NUMERIC, "en_US");
1539
ebac4655
JA
1540 if (parse_options(argc, argv))
1541 return 1;
1542
4b472fa3
JA
1543 if (!thread_number)
1544 return 0;
ebac4655 1545
29d610e1
JA
1546 ps = sysconf(_SC_PAGESIZE);
1547 if (ps < 0) {
1548 log_err("Failed to get page size\n");
1549 return 1;
1550 }
1551
cfc99db7 1552 page_size = ps;
29d610e1
JA
1553 page_mask = ps - 1;
1554
bb3884d8
JA
1555 if (write_bw_log) {
1556 setup_log(&agg_io_log[DDIR_READ]);
1557 setup_log(&agg_io_log[DDIR_WRITE]);
1558 }
1559
cdd18ad8 1560 startup_mutex = fio_mutex_init(0);
d2bbb8d9 1561 writeout_mutex = fio_mutex_init(1);
07739b57 1562
a2f77c9f
JA
1563 set_genesis_time();
1564
ccb0fa24 1565 status_timer_arm();
ebac4655
JA
1566
1567 run_threads();
6ce15a32 1568
bb3884d8 1569 if (!fio_abort) {
6ce15a32 1570 show_run_stats();
bb3884d8 1571 if (write_bw_log) {
5ec10eaa
JA
1572 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1573 __finish_log(agg_io_log[DDIR_WRITE],
1574 "agg-write_bw.log");
bb3884d8
JA
1575 }
1576 }
ebac4655 1577
cdd18ad8 1578 fio_mutex_remove(startup_mutex);
d2bbb8d9 1579 fio_mutex_remove(writeout_mutex);
437c9b71 1580 return exit_value;
ebac4655 1581}