Fix O_DIRECT memory alignment
[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;
d529ee19
JA
45
46#define PAGE_ALIGN(buf) \
29d610e1 47 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
48
49int groupid = 0;
50int thread_number = 0;
9cedf167
JA
51int nr_process = 0;
52int nr_thread = 0;
ebac4655 53int shm_id = 0;
53cdc686 54int temp_stall_ts;
d3eeeabc 55unsigned long done_secs = 0;
ebac4655 56
cdd18ad8 57static struct fio_mutex *startup_mutex;
d2bbb8d9 58static struct fio_mutex *writeout_mutex;
6ce15a32 59static volatile int fio_abort;
437c9b71 60static int exit_value;
ccb0fa24 61static struct itimerval itimer;
be4ecfdf 62static pthread_t gtod_thread;
ebac4655 63
bb3884d8
JA
64struct io_log *agg_io_log[2];
65
ebac4655 66#define TERMINATE_ALL (-1)
75154845 67#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 68
b29ee5b3 69void td_set_runstate(struct thread_data *td, int runstate)
6ce15a32 70{
d8fab1b6
JA
71 if (td->runstate == runstate)
72 return;
73
5921e80c
JA
74 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
75 td->runstate, runstate);
6ce15a32
JA
76 td->runstate = runstate;
77}
78
390c40e2 79static void terminate_threads(int group_id)
ebac4655 80{
34572e28 81 struct thread_data *td;
ebac4655
JA
82 int i;
83
6f88bcb0
JA
84 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
85
34572e28 86 for_each_td(td, i) {
ebac4655 87 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
6f88bcb0 88 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
5921e80c 89 td->o.name, (int) td->pid);
ad830ed7
JA
90 td->terminate = 1;
91 td->o.start_delay = 0;
92
7a93ab19
JA
93 /*
94 * if the thread is running, just let it exit
95 */
96 if (td->runstate < TD_RUNNING)
c1302d44 97 kill(td->pid, SIGQUIT);
f63f7f3b
JA
98 else {
99 struct ioengine_ops *ops = td->io_ops;
100
101 if (ops && (ops->flags & FIO_SIGQUIT))
102 kill(td->pid, SIGQUIT);
103 }
ebac4655
JA
104 }
105 }
106}
107
ccb0fa24
JA
108static void status_timer_arm(void)
109{
110 itimer.it_value.tv_sec = 0;
111 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
112 setitimer(ITIMER_REAL, &itimer, NULL);
113}
114
02444ad1 115static void sig_alrm(int fio_unused sig)
ebac4655 116{
697a606c 117 if (threads) {
5ec10eaa 118 update_io_ticks();
5ec10eaa 119 print_thread_status();
ccb0fa24 120 status_timer_arm();
697a606c
JA
121 }
122}
123
5f8f9726
JA
124/*
125 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
126 */
127static void sig_quit(int sig)
128{
129}
130
697a606c
JA
131static void sig_int(int sig)
132{
133 if (threads) {
5ec10eaa
JA
134 printf("\nfio: terminating on signal %d\n", sig);
135 fflush(stdout);
136 terminate_threads(TERMINATE_ALL);
ebac4655
JA
137 }
138}
139
02444ad1 140static void sig_ill(int fio_unused sig)
de79c915
JA
141{
142 if (!threads)
143 return;
144
145 log_err("fio: illegal instruction. your cpu does not support "
146 "the sse4.2 instruction for crc32c\n");
147 terminate_threads(TERMINATE_ALL);
148 exit(4);
149}
150
697a606c
JA
151static void set_sig_handlers(void)
152{
153 struct sigaction act;
154
155 memset(&act, 0, sizeof(act));
156 act.sa_handler = sig_alrm;
157 act.sa_flags = SA_RESTART;
158 sigaction(SIGALRM, &act, NULL);
159
160 memset(&act, 0, sizeof(act));
161 act.sa_handler = sig_int;
162 act.sa_flags = SA_RESTART;
163 sigaction(SIGINT, &act, NULL);
de79c915
JA
164
165 memset(&act, 0, sizeof(act));
166 act.sa_handler = sig_ill;
167 act.sa_flags = SA_RESTART;
168 sigaction(SIGILL, &act, NULL);
5f8f9726
JA
169
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_quit;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGQUIT, &act, NULL);
697a606c
JA
174}
175
906c8d75
JA
176/*
177 * Check if we are above the minimum rate given.
178 */
581e7141
JA
179static int __check_min_rate(struct thread_data *td, struct timeval *now,
180 enum td_ddir ddir)
ebac4655 181{
0904200b 182 unsigned long long bytes = 0;
4e991c23 183 unsigned long iops = 0;
ebac4655
JA
184 unsigned long spent;
185 unsigned long rate;
581e7141
JA
186 unsigned int ratemin = 0;
187 unsigned int rate_iops = 0;
188 unsigned int rate_iops_min = 0;
ebac4655 189
d982d873
JA
190 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
191 return 0;
192
ebac4655
JA
193 /*
194 * allow a 2 second settle period in the beginning
195 */
196 if (mtime_since(&td->start, now) < 2000)
197 return 0;
198
581e7141
JA
199 iops += td->io_blocks[ddir];
200 bytes += td->this_io_bytes[ddir];
201 ratemin += td->o.ratemin[ddir];
202 rate_iops += td->o.rate_iops[ddir];
203 rate_iops_min += td->o.rate_iops_min[ddir];
0904200b 204
ebac4655
JA
205 /*
206 * if rate blocks is set, sample is running
207 */
581e7141
JA
208 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
209 spent = mtime_since(&td->lastrate[ddir], now);
2dc1bbeb 210 if (spent < td->o.ratecycle)
ebac4655
JA
211 return 0;
212
581e7141 213 if (td->o.rate[ddir]) {
4e991c23
JA
214 /*
215 * check bandwidth specified rate
216 */
581e7141 217 if (bytes < td->rate_bytes[ddir]) {
5ec10eaa 218 log_err("%s: min rate %u not met\n", td->o.name,
581e7141 219 ratemin);
4e991c23
JA
220 return 1;
221 } else {
581e7141
JA
222 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
223 if (rate < ratemin ||
224 bytes < td->rate_bytes[ddir]) {
5ec10eaa
JA
225 log_err("%s: min rate %u not met, got"
226 " %luKiB/sec\n", td->o.name,
581e7141 227 ratemin, rate);
4e991c23
JA
228 return 1;
229 }
230 }
413dd459 231 } else {
4e991c23
JA
232 /*
233 * checks iops specified rate
234 */
581e7141 235 if (iops < rate_iops) {
5ec10eaa 236 log_err("%s: min iops rate %u not met\n",
581e7141 237 td->o.name, rate_iops);
413dd459 238 return 1;
4e991c23 239 } else {
581e7141
JA
240 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
241 if (rate < rate_iops_min ||
242 iops < td->rate_blocks[ddir]) {
5ec10eaa
JA
243 log_err("%s: min iops rate %u not met,"
244 " got %lu\n", td->o.name,
581e7141 245 rate_iops_min, rate);
4e991c23 246 }
413dd459 247 }
ebac4655
JA
248 }
249 }
250
581e7141
JA
251 td->rate_bytes[ddir] = bytes;
252 td->rate_blocks[ddir] = iops;
253 memcpy(&td->lastrate[ddir], now, sizeof(*now));
ebac4655
JA
254 return 0;
255}
256
581e7141
JA
257static int check_min_rate(struct thread_data *td, struct timeval *now,
258 unsigned long *bytes_done)
259{
260 int ret = 0;
261
262 if (bytes_done[0])
263 ret |= __check_min_rate(td, now, 0);
264 if (bytes_done[1])
265 ret |= __check_min_rate(td, now, 1);
266
267 return ret;
268}
269
ebac4655
JA
270static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
271{
2dc1bbeb 272 if (!td->o.timeout)
ebac4655 273 return 0;
2dc1bbeb 274 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
275 return 1;
276
277 return 0;
278}
279
906c8d75
JA
280/*
281 * When job exits, we can cancel the in-flight IO if we are using async
282 * io. Attempt to do so.
283 */
ebac4655
JA
284static void cleanup_pending_aio(struct thread_data *td)
285{
01743ee1 286 struct flist_head *entry, *n;
ebac4655
JA
287 struct io_u *io_u;
288 int r;
289
290 /*
291 * get immediately available events, if any
292 */
581e7141 293 r = io_u_queued_complete(td, 0, NULL);
b2fdda43
JA
294 if (r < 0)
295 return;
ebac4655
JA
296
297 /*
298 * now cancel remaining active events
299 */
2866c82d 300 if (td->io_ops->cancel) {
01743ee1
JA
301 flist_for_each_safe(entry, n, &td->io_u_busylist) {
302 io_u = flist_entry(entry, struct io_u, list);
ebac4655 303
0c6e7517
JA
304 /*
305 * if the io_u isn't in flight, then that generally
306 * means someone leaked an io_u. complain but fix
307 * it up, so we don't stall here.
308 */
309 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
310 log_err("fio: non-busy IO on busy list\n");
ebac4655 311 put_io_u(td, io_u);
0c6e7517
JA
312 } else {
313 r = td->io_ops->cancel(td, io_u);
314 if (!r)
315 put_io_u(td, io_u);
316 }
ebac4655
JA
317 }
318 }
319
97601024 320 if (td->cur_depth)
581e7141 321 r = io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
322}
323
858a3d47
JA
324/*
325 * Helper to handle the final sync of a file. Works just like the normal
326 * io path, just does everything sync.
327 */
328static int fio_io_sync(struct thread_data *td, struct fio_file *f)
329{
330 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
331 int ret;
332
333 if (!io_u)
334 return 1;
335
336 io_u->ddir = DDIR_SYNC;
337 io_u->file = f;
338
339 if (td_io_prep(td, io_u)) {
340 put_io_u(td, io_u);
341 return 1;
342 }
343
755200a3 344requeue:
858a3d47 345 ret = td_io_queue(td, io_u);
36167d82 346 if (ret < 0) {
e1161c32 347 td_verror(td, io_u->error, "td_io_queue");
858a3d47 348 put_io_u(td, io_u);
858a3d47 349 return 1;
36167d82 350 } else if (ret == FIO_Q_QUEUED) {
581e7141 351 if (io_u_queued_complete(td, 1, NULL) < 0)
36167d82 352 return 1;
36167d82
JA
353 } else if (ret == FIO_Q_COMPLETED) {
354 if (io_u->error) {
e1161c32 355 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
356 return 1;
357 }
858a3d47 358
581e7141 359 if (io_u_sync_complete(td, io_u, NULL) < 0)
b2fdda43 360 return 1;
755200a3
JA
361 } else if (ret == FIO_Q_BUSY) {
362 if (td_io_commit(td))
363 return 1;
364 goto requeue;
858a3d47
JA
365 }
366
367 return 0;
368}
369
993bf48b
JA
370static inline void update_tv_cache(struct thread_data *td)
371{
372 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
373 fio_gettime(&td->tv_cache, NULL);
374}
375
f2bba182
RR
376static int break_on_this_error(struct thread_data *td, int *retptr)
377{
378 int ret = *retptr;
379
380 if (ret < 0 || td->error) {
381 int err;
382
1ec99eea
JA
383 if (!td->o.continue_on_error)
384 return 1;
f2bba182
RR
385
386 if (ret < 0)
387 err = -ret;
388 else
389 err = td->error;
390
f2bba182
RR
391 if (td_non_fatal_error(err)) {
392 /*
393 * Continue with the I/Os in case of
394 * a non fatal error.
395 */
8a4f8a24 396 update_error_count(td, err);
f2bba182
RR
397 td_clear_error(td);
398 *retptr = 0;
399 return 0;
8a4f8a24
RR
400 } else if (td->o.fill_device && err == ENOSPC) {
401 /*
402 * We expect to hit this error if
403 * fill_device option is set.
404 */
405 td_clear_error(td);
406 td->terminate = 1;
407 return 1;
f2bba182
RR
408 } else {
409 /*
410 * Stop the I/O in case of a fatal
411 * error.
412 */
8a4f8a24 413 update_error_count(td, err);
f2bba182
RR
414 return 1;
415 }
416 }
417
418 return 0;
419}
420
906c8d75 421/*
34403fb1 422 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
423 * reads the blocks back in, and checks the crc/md5 of the data.
424 */
1e97cce9 425static void do_verify(struct thread_data *td)
ebac4655 426{
53cdc686 427 struct fio_file *f;
36167d82 428 struct io_u *io_u;
af52b345
JA
429 int ret, min_events;
430 unsigned int i;
e5b401d4
JA
431
432 /*
433 * sync io first and invalidate cache, to make sure we really
434 * read from disk.
435 */
436 for_each_file(td, f, i) {
d6aed795 437 if (!fio_file_open(f))
3d7b485f 438 continue;
b2fdda43
JA
439 if (fio_io_sync(td, f))
440 break;
441 if (file_invalidate_cache(td, f))
442 break;
e5b401d4 443 }
ebac4655 444
b2fdda43
JA
445 if (td->error)
446 return;
447
ebac4655
JA
448 td_set_runstate(td, TD_VERIFYING);
449
36167d82
JA
450 io_u = NULL;
451 while (!td->terminate) {
4950421a 452 int ret2, full;
5451792e 453
993bf48b
JA
454 update_tv_cache(td);
455
456 if (runtime_exceeded(td, &td->tv_cache)) {
90a87d4b 457 td->terminate = 1;
02bcaa8c 458 break;
069c2918 459 }
02bcaa8c 460
dcb4830d
RR
461 io_u = __get_io_u(td);
462 if (!io_u)
463 break;
464
069c2918
JA
465 if (get_next_verify(td, io_u)) {
466 put_io_u(td, io_u);
ebac4655 467 break;
069c2918 468 }
ebac4655 469
069c2918
JA
470 if (td_io_prep(td, io_u)) {
471 put_io_u(td, io_u);
53cdc686 472 break;
069c2918 473 }
d7762cf8
JA
474
475 io_u->end_io = verify_io_u;
53cdc686 476
11786802 477 ret = td_io_queue(td, io_u);
36167d82
JA
478 switch (ret) {
479 case FIO_Q_COMPLETED:
f2bba182 480 if (io_u->error) {
22819ec2 481 ret = -io_u->error;
f2bba182
RR
482 clear_io_u(td, io_u);
483 } else if (io_u->resid) {
36167d82 484 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 485 struct fio_file *f = io_u->file;
ebac4655 486
9e9d164e
JA
487 /*
488 * zero read, fail
489 */
490 if (!bytes) {
41d9f0fd 491 td_verror(td, EIO, "full resid");
9e9d164e
JA
492 put_io_u(td, io_u);
493 break;
494 }
8400d9b2 495
36167d82
JA
496 io_u->xfer_buflen = io_u->resid;
497 io_u->xfer_buf += bytes;
8400d9b2
JA
498 io_u->offset += bytes;
499
30061b97
JA
500 td->ts.short_io_u[io_u->ddir]++;
501
d460eb31 502 if (io_u->offset == f->real_file_size)
8400d9b2
JA
503 goto sync_done;
504
11786802
JA
505 requeue_io_u(td, &io_u);
506 } else {
8400d9b2 507sync_done:
581e7141 508 ret = io_u_sync_complete(td, io_u, NULL);
11786802
JA
509 if (ret < 0)
510 break;
36167d82 511 }
36167d82
JA
512 continue;
513 case FIO_Q_QUEUED:
514 break;
755200a3
JA
515 case FIO_Q_BUSY:
516 requeue_io_u(td, &io_u);
5451792e
JA
517 ret2 = td_io_commit(td);
518 if (ret2 < 0)
519 ret = ret2;
755200a3 520 break;
36167d82
JA
521 default:
522 assert(ret < 0);
e1161c32 523 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
524 break;
525 }
526
f2bba182 527 if (break_on_this_error(td, &ret))
3af6ef39
JA
528 break;
529
ebac4655 530 /*
3af6ef39
JA
531 * if we can queue more, do so. but check if there are
532 * completed io_u's first.
ebac4655 533 */
4950421a
JA
534 full = queue_full(td) || ret == FIO_Q_BUSY;
535 if (full || !td->o.iodepth_batch_complete) {
536 min_events = td->o.iodepth_batch_complete;
537 if (full && !min_events)
838bc709 538 min_events = 1;
e916b390 539
4950421a
JA
540 do {
541 /*
542 * Reap required number of io units, if any,
543 * and do the verification on them through
544 * the callback handler
545 */
581e7141 546 if (io_u_queued_complete(td, min_events, NULL) < 0) {
4950421a
JA
547 ret = -1;
548 break;
549 }
550 } while (full && (td->cur_depth > td->o.iodepth_low));
551 }
552 if (ret < 0)
ebac4655 553 break;
36167d82 554 }
ebac4655 555
c01c0395
JA
556 if (!td->error) {
557 min_events = td->cur_depth;
558
559 if (min_events)
581e7141 560 ret = io_u_queued_complete(td, min_events, NULL);
c01c0395 561 } else
ebac4655
JA
562 cleanup_pending_aio(td);
563
564 td_set_runstate(td, TD_RUNNING);
565}
566
32cd46a0 567/*
906c8d75 568 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
569 * and reaps them, checking for rate and errors along the way.
570 */
ebac4655
JA
571static void do_io(struct thread_data *td)
572{
af52b345
JA
573 unsigned int i;
574 int ret = 0;
ebac4655 575
b29ee5b3
JA
576 if (in_ramp_time(td))
577 td_set_runstate(td, TD_RAMP);
578 else
579 td_set_runstate(td, TD_RUNNING);
5853e5a8 580
7bb48f84 581 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
97601024 582 struct timeval comp_time;
581e7141 583 unsigned long bytes_done[2] = { 0, 0 };
84585003 584 int min_evts = 0;
ebac4655 585 struct io_u *io_u;
4950421a 586 int ret2, full;
ebac4655
JA
587
588 if (td->terminate)
589 break;
590
993bf48b 591 update_tv_cache(td);
97601024 592
993bf48b 593 if (runtime_exceeded(td, &td->tv_cache)) {
90a87d4b 594 td->terminate = 1;
97601024
JA
595 break;
596 }
36167d82 597
dcb4830d
RR
598 io_u = get_io_u(td);
599 if (!io_u)
600 break;
601
b67740d3
JA
602 /*
603 * Add verification end_io handler, if asked to verify
604 * a previously written file.
605 */
cc3fc855 606 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
b67740d3 607 io_u->end_io = verify_io_u;
d8fab1b6 608 td_set_runstate(td, TD_VERIFYING);
b29ee5b3
JA
609 } else if (in_ramp_time(td))
610 td_set_runstate(td, TD_RAMP);
611 else
d8fab1b6 612 td_set_runstate(td, TD_RUNNING);
b67740d3 613
11786802 614 ret = td_io_queue(td, io_u);
36167d82
JA
615 switch (ret) {
616 case FIO_Q_COMPLETED:
f2bba182 617 if (io_u->error) {
5451792e 618 ret = -io_u->error;
f2bba182
RR
619 clear_io_u(td, io_u);
620 } else if (io_u->resid) {
36167d82 621 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 622 struct fio_file *f = io_u->file;
36167d82 623
9e9d164e
JA
624 /*
625 * zero read, fail
626 */
627 if (!bytes) {
41d9f0fd 628 td_verror(td, EIO, "full resid");
9e9d164e
JA
629 put_io_u(td, io_u);
630 break;
631 }
632
cec6b55d 633 io_u->xfer_buflen = io_u->resid;
36167d82 634 io_u->xfer_buf += bytes;
5a7c5680
JA
635 io_u->offset += bytes;
636
30061b97
JA
637 td->ts.short_io_u[io_u->ddir]++;
638
d460eb31 639 if (io_u->offset == f->real_file_size)
5a7c5680
JA
640 goto sync_done;
641
11786802
JA
642 requeue_io_u(td, &io_u);
643 } else {
5a7c5680 644sync_done:
581e7141
JA
645 if (__should_check_rate(td, 0) ||
646 __should_check_rate(td, 1))
993bf48b
JA
647 fio_gettime(&comp_time, NULL);
648
581e7141
JA
649 ret = io_u_sync_complete(td, io_u, bytes_done);
650 if (ret < 0)
651 break;
cec6b55d 652 }
36167d82
JA
653 break;
654 case FIO_Q_QUEUED:
7e77dd02
JA
655 /*
656 * if the engine doesn't have a commit hook,
657 * the io_u is really queued. if it does have such
658 * a hook, it has to call io_u_queued() itself.
659 */
660 if (td->io_ops->commit == NULL)
661 io_u_queued(td, io_u);
36167d82 662 break;
755200a3
JA
663 case FIO_Q_BUSY:
664 requeue_io_u(td, &io_u);
5451792e
JA
665 ret2 = td_io_commit(td);
666 if (ret2 < 0)
667 ret = ret2;
755200a3 668 break;
36167d82
JA
669 default:
670 assert(ret < 0);
671 put_io_u(td, io_u);
672 break;
ebac4655
JA
673 }
674
f2bba182 675 if (break_on_this_error(td, &ret))
36167d82
JA
676 break;
677
97601024
JA
678 /*
679 * See if we need to complete some commands
680 */
4950421a
JA
681 full = queue_full(td) || ret == FIO_Q_BUSY;
682 if (full || !td->o.iodepth_batch_complete) {
683 min_evts = td->o.iodepth_batch_complete;
684 if (full && !min_evts)
36167d82 685 min_evts = 1;
4950421a 686
581e7141
JA
687 if (__should_check_rate(td, 0) ||
688 __should_check_rate(td, 1))
993bf48b 689 fio_gettime(&comp_time, NULL);
4950421a
JA
690
691 do {
581e7141
JA
692 ret = io_u_queued_complete(td, min_evts, bytes_done);
693 if (ret < 0)
4950421a
JA
694 break;
695
4950421a 696 } while (full && (td->cur_depth > td->o.iodepth_low));
ebac4655
JA
697 }
698
4950421a
JA
699 if (ret < 0)
700 break;
581e7141 701 if (!(bytes_done[0] + bytes_done[1]))
97601024
JA
702 continue;
703
581e7141
JA
704 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
705 if (check_min_rate(td, &comp_time, bytes_done)) {
721938ae
JA
706 if (exitall_on_terminate)
707 terminate_threads(td->groupid);
708 td_verror(td, EIO, "check_min_rate");
709 break;
710 }
ebac4655
JA
711 }
712
2dc1bbeb 713 if (td->o.thinktime) {
9c1f7434
JA
714 unsigned long long b;
715
716 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 717 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
718 int left;
719
2dc1bbeb 720 if (td->o.thinktime_spin)
6dd6f2cd 721 usec_spin(td->o.thinktime_spin);
48097d5c 722
2dc1bbeb 723 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
724 if (left)
725 usec_sleep(td, left);
726 }
9c1f7434 727 }
ebac4655
JA
728 }
729
aa31f1f1
SL
730 if (td->o.fill_device && td->error == ENOSPC) {
731 td->error = 0;
732 td->terminate = 1;
733 }
4d2413c6 734 if (!td->error) {
3d7c391d
JA
735 struct fio_file *f;
736
c01c0395
JA
737 i = td->cur_depth;
738 if (i)
581e7141 739 ret = io_u_queued_complete(td, i, NULL);
ebac4655 740
2dc1bbeb 741 if (should_fsync(td) && td->o.end_fsync) {
84585003 742 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
743
744 for_each_file(td, f, i) {
d6aed795 745 if (!fio_file_open(f))
3d7b485f 746 continue;
858a3d47 747 fio_io_sync(td, f);
3d7b485f 748 }
84585003 749 }
c01c0395
JA
750 } else
751 cleanup_pending_aio(td);
2ba1c290
JA
752
753 /*
754 * stop job if we failed doing any IO
755 */
756 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
757 td->done = 1;
ebac4655
JA
758}
759
ebac4655
JA
760static void cleanup_io_u(struct thread_data *td)
761{
01743ee1 762 struct flist_head *entry, *n;
ebac4655
JA
763 struct io_u *io_u;
764
01743ee1
JA
765 flist_for_each_safe(entry, n, &td->io_u_freelist) {
766 io_u = flist_entry(entry, struct io_u, list);
ebac4655 767
01743ee1 768 flist_del(&io_u->list);
ebac4655
JA
769 free(io_u);
770 }
771
2f9ade3c 772 free_io_mem(td);
ebac4655
JA
773}
774
775static int init_io_u(struct thread_data *td)
776{
777 struct io_u *io_u;
a00735e6 778 unsigned int max_bs;
eb7ccf38 779 int cl_align, i, max_units;
ebac4655
JA
780 char *p;
781
1e3c09af 782 max_units = td->o.iodepth;
2dc1bbeb 783 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
5ec10eaa
JA
784 td->orig_buffer_size = (unsigned long long) max_bs
785 * (unsigned long long) max_units;
786
787 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
788 unsigned long bs;
74b025b0 789
5ec10eaa
JA
790 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
791 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
792 }
ebac4655 793
c3990645
JA
794 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
795 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
796 return 1;
797 }
798
2f9ade3c
JA
799 if (allocate_io_mem(td))
800 return 1;
ebac4655 801
fc966bef 802 if (td->o.odirect || td->o.mem_align)
d529ee19 803 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
b3f378e9
JA
804 else
805 p = td->orig_buffer;
806
eb7ccf38
JA
807 cl_align = os_cache_line_size();
808
ebac4655 809 for (i = 0; i < max_units; i++) {
eb7ccf38
JA
810 void *ptr;
811
db1cd90b
JA
812 if (td->terminate)
813 return 1;
eb7ccf38
JA
814
815 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
816 log_err("fio: posix_memalign=%s\n", strerror(errno));
817 break;
818 }
819
820 io_u = ptr;
ebac4655 821 memset(io_u, 0, sizeof(*io_u));
01743ee1 822 INIT_FLIST_HEAD(&io_u->list);
d529ee19 823 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
ebac4655 824
b4c5e1ac
JA
825 if (!(td->io_ops->flags & FIO_NOIO)) {
826 io_u->buf = p + max_bs * i;
d529ee19 827 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
e9459e5a 828
5973cafb
JA
829 if (td_write(td) && !td->o.refill_buffers)
830 io_u_fill_buffer(td, io_u, max_bs);
b4c5e1ac 831 }
6b9cea23 832
b1ff3403 833 io_u->index = i;
0c6e7517 834 io_u->flags = IO_U_F_FREE;
01743ee1 835 flist_add(&io_u->list, &td->io_u_freelist);
ebac4655
JA
836 }
837
838 return 0;
839}
840
da86774e
JA
841static int switch_ioscheduler(struct thread_data *td)
842{
843 char tmp[256], tmp2[128];
844 FILE *f;
845 int ret;
846
ba0fbe10 847 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
848 return 0;
849
da86774e
JA
850 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
851
852 f = fopen(tmp, "r+");
853 if (!f) {
cbf5c1d7 854 if (errno == ENOENT) {
5ec10eaa
JA
855 log_err("fio: os or kernel doesn't support IO scheduler"
856 " switching\n");
cbf5c1d7
JA
857 return 0;
858 }
859 td_verror(td, errno, "fopen iosched");
da86774e
JA
860 return 1;
861 }
862
863 /*
864 * Set io scheduler.
865 */
2dc1bbeb 866 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 867 if (ferror(f) || ret != 1) {
e1161c32 868 td_verror(td, errno, "fwrite");
da86774e
JA
869 fclose(f);
870 return 1;
871 }
872
873 rewind(f);
874
875 /*
876 * Read back and check that the selected scheduler is now the default.
877 */
878 ret = fread(tmp, 1, sizeof(tmp), f);
879 if (ferror(f) || ret < 0) {
e1161c32 880 td_verror(td, errno, "fread");
da86774e
JA
881 fclose(f);
882 return 1;
883 }
884
2dc1bbeb 885 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 886 if (!strstr(tmp, tmp2)) {
2dc1bbeb 887 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 888 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
889 fclose(f);
890 return 1;
891 }
892
893 fclose(f);
894 return 0;
895}
896
492158cf
JA
897static int keep_running(struct thread_data *td)
898{
899 unsigned long long io_done;
900
20e354ef
JA
901 if (td->done)
902 return 0;
492158cf
JA
903 if (td->o.time_based)
904 return 1;
905 if (td->o.loops) {
906 td->o.loops--;
907 return 1;
908 }
909
5ec10eaa
JA
910 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
911 + td->io_skip_bytes;
492158cf
JA
912 if (io_done < td->o.size)
913 return 1;
914
915 return 0;
916}
917
b29ee5b3 918static void reset_io_counters(struct thread_data *td)
ebac4655 919{
756867bd 920 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 921 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 922 td->zone_bytes = 0;
581e7141
JA
923 td->rate_bytes[0] = td->rate_bytes[1] = 0;
924 td->rate_blocks[0] = td->rate_blocks[1] = 0;
ebac4655 925
c1324df1
JA
926 td->last_was_sync = 0;
927
2ba1c290
JA
928 /*
929 * reset file done count if we are to start over
930 */
931 if (td->o.time_based || td->o.loops)
48f5abd3 932 td->nr_done_files = 0;
5bfc35d7
JA
933
934 /*
935 * Set the same seed to get repeatable runs
936 */
937 td_fill_rand_seeds(td);
b29ee5b3
JA
938}
939
940void reset_all_stats(struct thread_data *td)
941{
942 struct timeval tv;
943 int i;
944
945 reset_io_counters(td);
946
947 for (i = 0; i < 2; i++) {
948 td->io_bytes[i] = 0;
949 td->io_blocks[i] = 0;
950 td->io_issues[i] = 0;
951 td->ts.total_io_u[i] = 0;
952 }
953
954 fio_gettime(&tv, NULL);
955 memcpy(&td->epoch, &tv, sizeof(tv));
956 memcpy(&td->start, &tv, sizeof(tv));
957}
958
df9c26b1 959static void clear_io_state(struct thread_data *td)
b29ee5b3
JA
960{
961 struct fio_file *f;
962 unsigned int i;
b29ee5b3
JA
963
964 reset_io_counters(td);
281f9b63 965
24ffd2c2 966 close_files(td);
df9c26b1 967 for_each_file(td, f, i)
d6aed795 968 fio_file_clear_done(f);
ebac4655
JA
969}
970
398e8c4d
JA
971static int exec_string(const char *string)
972{
973 int ret, newlen = strlen(string) + 1 + 8;
974 char *str;
975
976 str = malloc(newlen);
977 sprintf(str, "sh -c %s", string);
978
979 ret = system(str);
980 if (ret == -1)
981 log_err("fio: exec of cmd <%s> failed\n", str);
982
983 free(str);
984 return ret;
985}
986
906c8d75
JA
987/*
988 * Entry point for the thread based jobs. The process based jobs end up
989 * here as well, after a little setup.
990 */
ebac4655
JA
991static void *thread_main(void *data)
992{
10927316 993 unsigned long long runtime[2], elapsed;
ebac4655 994 struct thread_data *td = data;
a978ba68 995 int clear_state;
ebac4655 996
2dc1bbeb 997 if (!td->o.use_thread)
ebac4655
JA
998 setsid();
999
1000 td->pid = getpid();
1001
5921e80c 1002 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
ee56ad50 1003
01743ee1
JA
1004 INIT_FLIST_HEAD(&td->io_u_freelist);
1005 INIT_FLIST_HEAD(&td->io_u_busylist);
1006 INIT_FLIST_HEAD(&td->io_u_requeues);
1007 INIT_FLIST_HEAD(&td->io_log_list);
1008 INIT_FLIST_HEAD(&td->io_hist_list);
bb5d7d0b 1009 td->io_hist_tree = RB_ROOT;
aea47d44 1010
db1cd90b 1011 td_set_runstate(td, TD_INITIALIZED);
29adda3c 1012 dprint(FD_MUTEX, "up startup_mutex\n");
cdd18ad8 1013 fio_mutex_up(startup_mutex);
29adda3c 1014 dprint(FD_MUTEX, "wait on td->mutex\n");
cdd18ad8 1015 fio_mutex_down(td->mutex);
29adda3c 1016 dprint(FD_MUTEX, "done waiting on td->mutex\n");
db1cd90b
JA
1017
1018 /*
cdd18ad8 1019 * the ->mutex mutex is now no longer used, close it to avoid
db1cd90b
JA
1020 * eating a file descriptor
1021 */
cdd18ad8 1022 fio_mutex_remove(td->mutex);
db1cd90b 1023
d84f8d49
JA
1024 /*
1025 * May alter parameters that init_io_u() will use, so we need to
1026 * do this first.
1027 */
1028 if (init_iolog(td))
1029 goto err;
1030
ebac4655 1031 if (init_io_u(td))
db1cd90b 1032 goto err;
ebac4655 1033
375b2695 1034 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
e1161c32 1035 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 1036 goto err;
ebac4655
JA
1037 }
1038
f484470b
JA
1039 /*
1040 * If we have a gettimeofday() thread, make sure we exclude that
1041 * thread from this job
1042 */
be4ecfdf 1043 if (td->o.gtod_cpu) {
be4ecfdf
JA
1044 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1045 if (fio_setaffinity(td) == -1) {
1046 td_verror(td, errno, "cpu_set_affinity");
1047 goto err;
1048 }
1049 }
1050
ac684785 1051 if (td->ioprio_set) {
ebac4655 1052 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 1053 td_verror(td, errno, "ioprio_set");
db1cd90b 1054 goto err;
ebac4655
JA
1055 }
1056 }
1057
2dc1bbeb 1058 if (nice(td->o.nice) == -1) {
e1161c32 1059 td_verror(td, errno, "nice");
db1cd90b 1060 goto err;
b6f4d880
JA
1061 }
1062
2dc1bbeb 1063 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 1064 goto err;
37f56873 1065
2dc1bbeb 1066 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
1067 goto err;
1068
7d6c5283
JA
1069 if (td_io_init(td))
1070 goto err;
1071
68727076
JA
1072 if (init_random_map(td))
1073 goto err;
1074
2dc1bbeb 1075 if (td->o.exec_prerun) {
398e8c4d 1076 if (exec_string(td->o.exec_prerun))
69cfd7e0
JA
1077 goto err;
1078 }
4e0ba8af 1079
afad68f7
ZY
1080 if (td->o.pre_read) {
1081 if (pre_read_files(td) < 0)
1082 goto err;
1083 }
1084
69008999 1085 fio_gettime(&td->epoch, NULL);
756867bd 1086 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
1087
1088 runtime[0] = runtime[1] = 0;
a978ba68 1089 clear_state = 0;
492158cf 1090 while (keep_running(td)) {
02bcaa8c 1091 fio_gettime(&td->start, NULL);
756867bd 1092 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
993bf48b 1093 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
ebac4655 1094
cf2dd8d6 1095 if (td->o.ratemin[0] || td->o.ratemin[1])
5ec10eaa
JA
1096 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1097 sizeof(td->lastrate));
ebac4655 1098
df9c26b1
JA
1099 if (clear_state)
1100 clear_io_state(td);
a978ba68 1101
ebac4655
JA
1102 prune_io_piece_log(td);
1103
ba0fbe10 1104 do_io(td);
ebac4655 1105
a978ba68
JA
1106 clear_state = 1;
1107
38d77cae 1108 if (td_read(td) && td->io_bytes[DDIR_READ]) {
0a4957c6 1109 elapsed = utime_since_now(&td->start);
38d77cae
JA
1110 runtime[DDIR_READ] += elapsed;
1111 }
1112 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
0a4957c6 1113 elapsed = utime_since_now(&td->start);
38d77cae
JA
1114 runtime[DDIR_WRITE] += elapsed;
1115 }
5ec10eaa 1116
ebac4655
JA
1117 if (td->error || td->terminate)
1118 break;
1119
e84c73a8
SL
1120 if (!td->o.do_verify ||
1121 td->o.verify == VERIFY_NONE ||
b67740d3 1122 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
1123 continue;
1124
df9c26b1 1125 clear_io_state(td);
a978ba68 1126
02bcaa8c 1127 fio_gettime(&td->start, NULL);
ebac4655
JA
1128
1129 do_verify(td);
1130
69008999 1131 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
1132
1133 if (td->error || td->terminate)
1134 break;
1135 }
1136
36dff966 1137 update_rusage_stat(td);
47f0cc48
YHJT
1138 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1139 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
756867bd
JA
1140 td->ts.total_run_time = mtime_since_now(&td->epoch);
1141 td->ts.io_bytes[0] = td->io_bytes[0];
1142 td->ts.io_bytes[1] = td->io_bytes[1];
1143
d2bbb8d9 1144 fio_mutex_down(writeout_mutex);
e3cedca7 1145 if (td->ts.bw_log) {
f484470b
JA
1146 if (td->o.bw_log_file) {
1147 finish_log_named(td, td->ts.bw_log,
1148 td->o.bw_log_file, "bw");
1149 } else
e3cedca7
JA
1150 finish_log(td, td->ts.bw_log, "bw");
1151 }
1152 if (td->ts.slat_log) {
f484470b
JA
1153 if (td->o.lat_log_file) {
1154 finish_log_named(td, td->ts.slat_log,
1ca95738 1155 td->o.lat_log_file, "slat");
f484470b 1156 } else
e3cedca7
JA
1157 finish_log(td, td->ts.slat_log, "slat");
1158 }
1159 if (td->ts.clat_log) {
f484470b
JA
1160 if (td->o.lat_log_file) {
1161 finish_log_named(td, td->ts.clat_log,
1162 td->o.lat_log_file, "clat");
1163 } else
e3cedca7
JA
1164 finish_log(td, td->ts.clat_log, "clat");
1165 }
d2bbb8d9 1166 fio_mutex_up(writeout_mutex);
398e8c4d
JA
1167 if (td->o.exec_postrun)
1168 exec_string(td->o.exec_postrun);
ebac4655
JA
1169
1170 if (exitall_on_terminate)
390c40e2 1171 terminate_threads(td->groupid);
ebac4655
JA
1172
1173err:
5bf13a5a 1174 if (td->error)
5921e80c 1175 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
5ec10eaa 1176 td->verror);
24ffd2c2 1177 close_and_free_files(td);
2866c82d 1178 close_ioengine(td);
ebac4655 1179 cleanup_io_u(td);
f29b25a3 1180
d2ce18b5
JA
1181 if (td->o.cpumask_set) {
1182 int ret = fio_cpuset_exit(&td->o.cpumask);
1183
1184 td_verror(td, ret, "fio_cpuset_exit");
1185 }
4e78e405 1186
f29b25a3
JA
1187 /*
1188 * do this very late, it will log file closing as well
1189 */
1190 if (td->o.write_iolog_file)
1191 write_iolog_close(td);
1192
d23bb327 1193 options_mem_free(td);
ebac4655 1194 td_set_runstate(td, TD_EXITED);
43d76807 1195 return (void *) (unsigned long) td->error;
ebac4655
JA
1196}
1197
906c8d75
JA
1198/*
1199 * We cannot pass the td data into a forked process, so attach the td and
1200 * pass it to the thread worker.
1201 */
a6418147 1202static int fork_main(int shmid, int offset)
ebac4655
JA
1203{
1204 struct thread_data *td;
a6418147 1205 void *data, *ret;
ebac4655
JA
1206
1207 data = shmat(shmid, NULL, 0);
1208 if (data == (void *) -1) {
a6418147
JA
1209 int __err = errno;
1210
ebac4655 1211 perror("shmat");
a6418147 1212 return __err;
ebac4655
JA
1213 }
1214
1215 td = data + offset * sizeof(struct thread_data);
a6418147 1216 ret = thread_main(td);
ebac4655 1217 shmdt(data);
43d76807 1218 return (int) (unsigned long) ret;
ebac4655
JA
1219}
1220
906c8d75
JA
1221/*
1222 * Run over the job map and reap the threads that have exited, if any.
1223 */
ebac4655
JA
1224static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1225{
34572e28 1226 struct thread_data *td;
1f809d15 1227 int i, cputhreads, realthreads, pending, status, ret;
ebac4655
JA
1228
1229 /*
1230 * reap exited threads (TD_EXITED -> TD_REAPED)
1231 */
1f809d15 1232 realthreads = pending = cputhreads = 0;
34572e28 1233 for_each_td(td, i) {
3707f45b 1234 int flags = 0;
a2f77c9f 1235
84585003
JA
1236 /*
1237 * ->io_ops is NULL for a thread that has closed its
1238 * io engine
1239 */
ba0fbe10 1240 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0 1241 cputhreads++;
1f809d15
JA
1242 else
1243 realthreads++;
b990b5c0 1244
a1dedc1f
AC
1245 if (!td->pid) {
1246 pending++;
1247 continue;
1248 }
1249 if (td->runstate == TD_REAPED)
a2f77c9f 1250 continue;
2dc1bbeb 1251 if (td->o.use_thread) {
3707f45b
JA
1252 if (td->runstate == TD_EXITED) {
1253 td_set_runstate(td, TD_REAPED);
1254 goto reaped;
1255 }
1256 continue;
1257 }
a2f77c9f
JA
1258
1259 flags = WNOHANG;
1260 if (td->runstate == TD_EXITED)
1261 flags = 0;
1262
1263 /*
1264 * check if someone quit or got killed in an unusual way
1265 */
1266 ret = waitpid(td->pid, &status, flags);
3707f45b 1267 if (ret < 0) {
a2f77c9f 1268 if (errno == ECHILD) {
5921e80c
JA
1269 log_err("fio: pid=%d disappeared %d\n",
1270 (int) td->pid, td->runstate);
a2f77c9f
JA
1271 td_set_runstate(td, TD_REAPED);
1272 goto reaped;
1273 }
1274 perror("waitpid");
1275 } else if (ret == td->pid) {
1276 if (WIFSIGNALED(status)) {
fab6aa71
JA
1277 int sig = WTERMSIG(status);
1278
d9244941 1279 if (sig != SIGQUIT)
5ec10eaa 1280 log_err("fio: pid=%d, got signal=%d\n",
5921e80c 1281 (int) td->pid, sig);
fab6aa71
JA
1282 td_set_runstate(td, TD_REAPED);
1283 goto reaped;
1284 }
a2f77c9f
JA
1285 if (WIFEXITED(status)) {
1286 if (WEXITSTATUS(status) && !td->error)
1287 td->error = WEXITSTATUS(status);
a2f77c9f 1288
a2f77c9f
JA
1289 td_set_runstate(td, TD_REAPED);
1290 goto reaped;
a6418147
JA
1291 }
1292 }
ebac4655 1293
a2f77c9f
JA
1294 /*
1295 * thread is not dead, continue
1296 */
e48676ba 1297 pending++;
a2f77c9f 1298 continue;
fab6aa71 1299reaped:
ebac4655 1300 (*nr_running)--;
581e7141
JA
1301 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1302 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
6f88bcb0
JA
1303 if (!td->pid)
1304 pending--;
a2f77c9f
JA
1305
1306 if (td->error)
1307 exit_value++;
d3eeeabc
JA
1308
1309 done_secs += mtime_since_now(&td->epoch) / 1000;
ebac4655 1310 }
b990b5c0 1311
1f809d15 1312 if (*nr_running == cputhreads && !pending && realthreads)
390c40e2 1313 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1314}
1315
be4ecfdf
JA
1316static void *gtod_thread_main(void *data)
1317{
1318 fio_mutex_up(startup_mutex);
1319
1320 /*
1321 * As long as we have jobs around, update the clock. It would be nice
1322 * to have some way of NOT hammering that CPU with gettimeofday(),
1323 * but I'm not sure what to use outside of a simple CPU nop to relax
1324 * it - we don't want to lose precision.
1325 */
1326 while (threads) {
1327 fio_gtod_update();
1328 nop;
1329 }
1330
1331 return NULL;
1332}
1333
1334static int fio_start_gtod_thread(void)
1335{
f42c153d
JA
1336 int ret;
1337
1338 ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1339 if (ret) {
1340 log_err("Can't create gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1341 return 1;
1342 }
f42c153d
JA
1343
1344 ret = pthread_detach(gtod_thread);
1345 if (ret) {
1346 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1347 return 1;
1348 }
1349
29adda3c 1350 dprint(FD_MUTEX, "wait on startup_mutex\n");
be4ecfdf 1351 fio_mutex_down(startup_mutex);
29adda3c 1352 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
be4ecfdf
JA
1353 return 0;
1354}
1355
906c8d75
JA
1356/*
1357 * Main function for kicking off and reaping jobs, as needed.
1358 */
ebac4655
JA
1359static void run_threads(void)
1360{
ebac4655
JA
1361 struct thread_data *td;
1362 unsigned long spent;
1363 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1364
2f9ade3c
JA
1365 if (fio_pin_memory())
1366 return;
ebac4655 1367
be4ecfdf
JA
1368 if (fio_gtod_offload && fio_start_gtod_thread())
1369 return;
1370
c6ae0a5b 1371 if (!terse_output) {
9cedf167
JA
1372 printf("Starting ");
1373 if (nr_thread)
5ec10eaa
JA
1374 printf("%d thread%s", nr_thread,
1375 nr_thread > 1 ? "s" : "");
9cedf167
JA
1376 if (nr_process) {
1377 if (nr_thread)
1378 printf(" and ");
5ec10eaa
JA
1379 printf("%d process%s", nr_process,
1380 nr_process > 1 ? "es" : "");
9cedf167
JA
1381 }
1382 printf("\n");
c6ae0a5b
JA
1383 fflush(stdout);
1384 }
c04f7ec3 1385
697a606c 1386 set_sig_handlers();
4efa970e 1387
ebac4655
JA
1388 todo = thread_number;
1389 nr_running = 0;
1390 nr_started = 0;
1391 m_rate = t_rate = 0;
1392
34572e28 1393 for_each_td(td, i) {
263e529f 1394 print_status_init(td->thread_number - 1);
ebac4655 1395
2dc1bbeb 1396 if (!td->o.create_serialize) {
380cf265 1397 init_disk_util(td);
ebac4655 1398 continue;
380cf265 1399 }
ebac4655
JA
1400
1401 /*
1402 * do file setup here so it happens sequentially,
1403 * we don't want X number of threads getting their
1404 * client data interspersed on disk
1405 */
53cdc686 1406 if (setup_files(td)) {
5bf13a5a
JA
1407 exit_value++;
1408 if (td->error)
5921e80c
JA
1409 log_err("fio: pid=%d, err=%d/%s\n",
1410 (int) td->pid, td->error, td->verror);
ebac4655
JA
1411 td_set_runstate(td, TD_REAPED);
1412 todo--;
c69e8875
JA
1413 } else {
1414 struct fio_file *f;
1415 unsigned int i;
1416
1417 /*
1418 * for sharing to work, each job must always open
1419 * its own files. so close them, if we opened them
1420 * for creation
1421 */
22a57ba8
JA
1422 for_each_file(td, f, i) {
1423 if (fio_file_open(f))
1424 td_io_close_file(td, f);
22a57ba8 1425 }
5ec10eaa 1426 }
380cf265
JA
1427
1428 init_disk_util(td);
ebac4655
JA
1429 }
1430
a2f77c9f
JA
1431 set_genesis_time();
1432
ebac4655 1433 while (todo) {
75154845
JA
1434 struct thread_data *map[MAX_JOBS];
1435 struct timeval this_start;
1436 int this_jobs = 0, left;
1437
ebac4655
JA
1438 /*
1439 * create threads (TD_NOT_CREATED -> TD_CREATED)
1440 */
34572e28 1441 for_each_td(td, i) {
ebac4655
JA
1442 if (td->runstate != TD_NOT_CREATED)
1443 continue;
1444
1445 /*
1446 * never got a chance to start, killed by other
1447 * thread for some reason
1448 */
1449 if (td->terminate) {
1450 todo--;
1451 continue;
1452 }
1453
2dc1bbeb 1454 if (td->o.start_delay) {
263e529f 1455 spent = mtime_since_genesis();
ebac4655 1456
2dc1bbeb 1457 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1458 continue;
1459 }
1460
6f88bcb0
JA
1461 if (td->o.stonewall && (nr_started || nr_running)) {
1462 dprint(FD_PROCESS, "%s: stonewall wait\n",
1463 td->o.name);
ebac4655 1464 break;
6f88bcb0 1465 }
ebac4655 1466
75154845
JA
1467 /*
1468 * Set state to created. Thread will transition
1469 * to TD_INITIALIZED when it's done setting up.
1470 */
ebac4655 1471 td_set_runstate(td, TD_CREATED);
75154845 1472 map[this_jobs++] = td;
ebac4655
JA
1473 nr_started++;
1474
2dc1bbeb 1475 if (td->o.use_thread) {
f42c153d
JA
1476 int ret;
1477
ee56ad50 1478 dprint(FD_PROCESS, "will pthread_create\n");
f42c153d
JA
1479 ret = pthread_create(&td->thread, NULL,
1480 thread_main, td);
1481 if (ret) {
1482 log_err("pthread_create: %s\n",
1483 strerror(ret));
ebac4655 1484 nr_started--;
c8bb6faf 1485 break;
ebac4655 1486 }
f42c153d
JA
1487 ret = pthread_detach(td->thread);
1488 if (ret)
1489 log_err("pthread_detach: %s",
1490 strerror(ret));
ebac4655 1491 } else {
5e1d306e 1492 pid_t pid;
ee56ad50 1493 dprint(FD_PROCESS, "will fork\n");
5e1d306e
JA
1494 pid = fork();
1495 if (!pid) {
a6418147
JA
1496 int ret = fork_main(shm_id, i);
1497
5e1d306e
JA
1498 _exit(ret);
1499 } else if (i == fio_debug_jobno)
1500 *fio_debug_jobp = pid;
ebac4655 1501 }
29adda3c 1502 dprint(FD_MUTEX, "wait on startup_mutex\n");
656b1393
JA
1503 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1504 log_err("fio: job startup hung? exiting.\n");
1505 terminate_threads(TERMINATE_ALL);
1506 fio_abort = 1;
1507 nr_started--;
1508 break;
1509 }
29adda3c 1510 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
ebac4655
JA
1511 }
1512
1513 /*
75154845
JA
1514 * Wait for the started threads to transition to
1515 * TD_INITIALIZED.
ebac4655 1516 */
02bcaa8c 1517 fio_gettime(&this_start, NULL);
75154845 1518 left = this_jobs;
6ce15a32 1519 while (left && !fio_abort) {
75154845
JA
1520 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1521 break;
1522
1523 usleep(100000);
1524
1525 for (i = 0; i < this_jobs; i++) {
1526 td = map[i];
1527 if (!td)
1528 continue;
b6f4d880 1529 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1530 map[i] = NULL;
1531 left--;
b6f4d880
JA
1532 } else if (td->runstate >= TD_EXITED) {
1533 map[i] = NULL;
1534 left--;
1535 todo--;
1536 nr_running++; /* work-around... */
75154845
JA
1537 }
1538 }
1539 }
1540
1541 if (left) {
3b70d7e5 1542 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1543 for (i = 0; i < this_jobs; i++) {
1544 td = map[i];
1545 if (!td)
1546 continue;
1547 kill(td->pid, SIGTERM);
1548 }
1549 break;
1550 }
1551
1552 /*
b6f4d880 1553 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1554 */
34572e28 1555 for_each_td(td, i) {
75154845 1556 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1557 continue;
1558
b29ee5b3
JA
1559 if (in_ramp_time(td))
1560 td_set_runstate(td, TD_RAMP);
1561 else
1562 td_set_runstate(td, TD_RUNNING);
ebac4655
JA
1563 nr_running++;
1564 nr_started--;
581e7141
JA
1565 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1566 t_rate += td->o.rate[0] + td->o.rate[1];
75154845 1567 todo--;
cdd18ad8 1568 fio_mutex_up(td->mutex);
ebac4655
JA
1569 }
1570
1571 reap_threads(&nr_running, &t_rate, &m_rate);
1572
1573 if (todo)
1574 usleep(100000);
1575 }
1576
1577 while (nr_running) {
1578 reap_threads(&nr_running, &t_rate, &m_rate);
1579 usleep(10000);
1580 }
1581
1582 update_io_ticks();
2f9ade3c 1583 fio_unpin_memory();
ebac4655
JA
1584}
1585
ebac4655
JA
1586int main(int argc, char *argv[])
1587{
29d610e1
JA
1588 long ps;
1589
2e5cdb11
JA
1590 sinit();
1591
dbe1125e
JA
1592 /*
1593 * We need locale for number printing, if it isn't set then just
1594 * go with the US format.
1595 */
1596 if (!getenv("LC_NUMERIC"))
1597 setlocale(LC_NUMERIC, "en_US");
1598
ebac4655
JA
1599 if (parse_options(argc, argv))
1600 return 1;
1601
4b472fa3
JA
1602 if (!thread_number)
1603 return 0;
ebac4655 1604
29d610e1
JA
1605 ps = sysconf(_SC_PAGESIZE);
1606 if (ps < 0) {
1607 log_err("Failed to get page size\n");
1608 return 1;
1609 }
1610
cfc99db7 1611 page_size = ps;
29d610e1
JA
1612 page_mask = ps - 1;
1613
bb3884d8
JA
1614 if (write_bw_log) {
1615 setup_log(&agg_io_log[DDIR_READ]);
1616 setup_log(&agg_io_log[DDIR_WRITE]);
1617 }
1618
cdd18ad8 1619 startup_mutex = fio_mutex_init(0);
d2bbb8d9 1620 writeout_mutex = fio_mutex_init(1);
07739b57 1621
a2f77c9f
JA
1622 set_genesis_time();
1623
ccb0fa24 1624 status_timer_arm();
ebac4655
JA
1625
1626 run_threads();
6ce15a32 1627
bb3884d8 1628 if (!fio_abort) {
6ce15a32 1629 show_run_stats();
bb3884d8 1630 if (write_bw_log) {
5ec10eaa
JA
1631 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1632 __finish_log(agg_io_log[DDIR_WRITE],
1633 "agg-write_bw.log");
bb3884d8
JA
1634 }
1635 }
ebac4655 1636
cdd18ad8 1637 fio_mutex_remove(startup_mutex);
d2bbb8d9 1638 fio_mutex_remove(writeout_mutex);
437c9b71 1639 return exit_value;
ebac4655 1640}