Verify the pattern we've written
[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
b67740d3
JA
449 /*
450 * Add verification end_io handler, if asked to verify
451 * a previously written file.
452 */
453 if (td->o.verify != VERIFY_NONE)
454 io_u->end_io = verify_io_u;
455
11786802 456 ret = td_io_queue(td, io_u);
36167d82
JA
457 switch (ret) {
458 case FIO_Q_COMPLETED:
5451792e
JA
459 if (io_u->error)
460 ret = -io_u->error;
9e9d164e 461 else if (io_u->resid) {
36167d82 462 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 463 struct fio_file *f = io_u->file;
36167d82 464
9e9d164e
JA
465 /*
466 * zero read, fail
467 */
468 if (!bytes) {
469 td_verror(td, ENODATA, "full resid");
470 put_io_u(td, io_u);
471 break;
472 }
473
cec6b55d 474 io_u->xfer_buflen = io_u->resid;
36167d82 475 io_u->xfer_buf += bytes;
5a7c5680 476 io_u->offset += bytes;
d460eb31 477 f->last_completed_pos = io_u->offset;
5a7c5680 478
30061b97
JA
479 td->ts.short_io_u[io_u->ddir]++;
480
d460eb31 481 if (io_u->offset == f->real_file_size)
5a7c5680
JA
482 goto sync_done;
483
11786802
JA
484 requeue_io_u(td, &io_u);
485 } else {
5a7c5680 486sync_done:
11786802
JA
487 fio_gettime(&comp_time, NULL);
488 bytes_done = io_u_sync_complete(td, io_u);
489 if (bytes_done < 0)
490 ret = bytes_done;
cec6b55d 491 }
36167d82
JA
492 break;
493 case FIO_Q_QUEUED:
7e77dd02
JA
494 /*
495 * if the engine doesn't have a commit hook,
496 * the io_u is really queued. if it does have such
497 * a hook, it has to call io_u_queued() itself.
498 */
499 if (td->io_ops->commit == NULL)
500 io_u_queued(td, io_u);
36167d82 501 break;
755200a3
JA
502 case FIO_Q_BUSY:
503 requeue_io_u(td, &io_u);
5451792e
JA
504 ret2 = td_io_commit(td);
505 if (ret2 < 0)
506 ret = ret2;
755200a3 507 break;
36167d82
JA
508 default:
509 assert(ret < 0);
510 put_io_u(td, io_u);
511 break;
ebac4655
JA
512 }
513
99784632 514 if (ret < 0 || td->error)
36167d82
JA
515 break;
516
97601024
JA
517 /*
518 * See if we need to complete some commands
519 */
755200a3 520 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 521 min_evts = 0;
e916b390 522 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 523 min_evts = 1;
ebac4655 524
2dc1bbeb
JA
525 if (td->cur_depth > td->o.iodepth_low)
526 min_evts = td->cur_depth - td->o.iodepth_low;
e916b390
JA
527 }
528
97601024 529 fio_gettime(&comp_time, NULL);
d7762cf8 530 bytes_done = io_u_queued_complete(td, min_evts);
97601024 531 if (bytes_done < 0)
36167d82 532 break;
ebac4655
JA
533 }
534
97601024
JA
535 if (!bytes_done)
536 continue;
537
ebac4655
JA
538 /*
539 * the rate is batched for now, it should work for batches
540 * of completions except the very first one which may look
541 * a little bursty
542 */
97601024 543 usec = utime_since(&s, &comp_time);
ebac4655 544
413dd459 545 rate_throttle(td, usec, bytes_done);
ebac4655 546
97601024 547 if (check_min_rate(td, &comp_time)) {
98aa62d8 548 if (exitall_on_terminate)
390c40e2 549 terminate_threads(td->groupid);
e1161c32 550 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
551 break;
552 }
553
2dc1bbeb 554 if (td->o.thinktime) {
9c1f7434
JA
555 unsigned long long b;
556
557 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 558 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
559 int left;
560
2dc1bbeb
JA
561 if (td->o.thinktime_spin)
562 __usec_sleep(td->o.thinktime_spin);
48097d5c 563
2dc1bbeb 564 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
565 if (left)
566 usec_sleep(td, left);
567 }
9c1f7434 568 }
ebac4655
JA
569 }
570
4d2413c6 571 if (!td->error) {
3d7c391d
JA
572 struct fio_file *f;
573
c01c0395
JA
574 i = td->cur_depth;
575 if (i)
576 ret = io_u_queued_complete(td, i);
ebac4655 577
2dc1bbeb 578 if (should_fsync(td) && td->o.end_fsync) {
84585003 579 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
580
581 for_each_file(td, f, i) {
582 if (!(f->flags & FIO_FILE_OPEN))
583 continue;
858a3d47 584 fio_io_sync(td, f);
3d7b485f 585 }
84585003 586 }
c01c0395
JA
587 } else
588 cleanup_pending_aio(td);
ebac4655
JA
589}
590
ebac4655
JA
591static void cleanup_io_u(struct thread_data *td)
592{
593 struct list_head *entry, *n;
594 struct io_u *io_u;
595
596 list_for_each_safe(entry, n, &td->io_u_freelist) {
597 io_u = list_entry(entry, struct io_u, list);
598
599 list_del(&io_u->list);
600 free(io_u);
601 }
602
2f9ade3c 603 free_io_mem(td);
ebac4655
JA
604}
605
6b9cea23
JA
606/*
607 * "randomly" fill the buffer contents
608 */
e9459e5a 609static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
6b9cea23 610{
210f43b3 611 long *ptr = io_u->buf;
6b9cea23 612
e9459e5a
JA
613 if (!td->o.zero_buffers) {
614 while ((void *) ptr - io_u->buf < max_bs) {
210f43b3 615 *ptr = rand() * GOLDEN_RATIO_PRIME;
e9459e5a
JA
616 ptr++;
617 }
618 } else
619 memset(ptr, 0, max_bs);
6b9cea23
JA
620}
621
ebac4655
JA
622static int init_io_u(struct thread_data *td)
623{
624 struct io_u *io_u;
a00735e6 625 unsigned int max_bs;
ebac4655
JA
626 int i, max_units;
627 char *p;
628
2866c82d 629 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
630 max_units = 1;
631 else
2dc1bbeb 632 max_units = td->o.iodepth;
ebac4655 633
2dc1bbeb 634 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
c3990645 635 td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
74b025b0 636
2dc1bbeb
JA
637 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
638 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
ebac4655 639
c3990645
JA
640 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
641 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
642 return 1;
643 }
644
2f9ade3c
JA
645 if (allocate_io_mem(td))
646 return 1;
ebac4655 647
b3f378e9
JA
648 if (td->o.odirect)
649 p = ALIGN(td->orig_buffer);
650 else
651 p = td->orig_buffer;
652
ebac4655 653 for (i = 0; i < max_units; i++) {
db1cd90b
JA
654 if (td->terminate)
655 return 1;
ebac4655
JA
656 io_u = malloc(sizeof(*io_u));
657 memset(io_u, 0, sizeof(*io_u));
658 INIT_LIST_HEAD(&io_u->list);
659
a00735e6 660 io_u->buf = p + max_bs * i;
e9459e5a 661
660a1cb5 662 if (td_write(td))
e9459e5a 663 fill_io_buf(td, io_u, max_bs);
6b9cea23 664
b1ff3403 665 io_u->index = i;
0c6e7517 666 io_u->flags = IO_U_F_FREE;
ebac4655
JA
667 list_add(&io_u->list, &td->io_u_freelist);
668 }
669
433afcb4
JA
670 io_u_init_timeout();
671
ebac4655
JA
672 return 0;
673}
674
da86774e
JA
675static int switch_ioscheduler(struct thread_data *td)
676{
677 char tmp[256], tmp2[128];
678 FILE *f;
679 int ret;
680
ba0fbe10 681 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
682 return 0;
683
da86774e
JA
684 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
685
686 f = fopen(tmp, "r+");
687 if (!f) {
cbf5c1d7
JA
688 if (errno == ENOENT) {
689 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
690 return 0;
691 }
692 td_verror(td, errno, "fopen iosched");
da86774e
JA
693 return 1;
694 }
695
696 /*
697 * Set io scheduler.
698 */
2dc1bbeb 699 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 700 if (ferror(f) || ret != 1) {
e1161c32 701 td_verror(td, errno, "fwrite");
da86774e
JA
702 fclose(f);
703 return 1;
704 }
705
706 rewind(f);
707
708 /*
709 * Read back and check that the selected scheduler is now the default.
710 */
711 ret = fread(tmp, 1, sizeof(tmp), f);
712 if (ferror(f) || ret < 0) {
e1161c32 713 td_verror(td, errno, "fread");
da86774e
JA
714 fclose(f);
715 return 1;
716 }
717
2dc1bbeb 718 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 719 if (!strstr(tmp, tmp2)) {
2dc1bbeb 720 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 721 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
722 fclose(f);
723 return 1;
724 }
725
726 fclose(f);
727 return 0;
728}
729
492158cf
JA
730static int keep_running(struct thread_data *td)
731{
732 unsigned long long io_done;
733
20e354ef
JA
734 if (td->done)
735 return 0;
492158cf
JA
736 if (td->o.time_based)
737 return 1;
738 if (td->o.loops) {
739 td->o.loops--;
740 return 1;
741 }
742
48f5abd3 743 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
492158cf
JA
744 if (io_done < td->o.size)
745 return 1;
746
747 return 0;
748}
749
a978ba68 750static int clear_io_state(struct thread_data *td)
ebac4655 751{
53cdc686 752 struct fio_file *f;
af52b345
JA
753 unsigned int i;
754 int ret;
ebac4655 755
756867bd 756 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 757 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 758 td->zone_bytes = 0;
d7d3b49b 759 td->rate_bytes = 0;
4e991c23 760 td->rate_blocks = 0;
38d77cae 761 td->rw_end_set[0] = td->rw_end_set[1] = 0;
ebac4655 762
c1324df1
JA
763 td->last_was_sync = 0;
764
48f5abd3
JA
765 if (td->o.time_based)
766 td->nr_done_files = 0;
281f9b63 767
a978ba68
JA
768 for_each_file(td, f, i)
769 td_io_close_file(td, f);
53cdc686 770
a978ba68
JA
771 ret = 0;
772 for_each_file(td, f, i) {
281f9b63 773 f->flags &= ~FIO_FILE_DONE;
a978ba68
JA
774 ret = td_io_open_file(td, f);
775 if (ret)
776 break;
53cdc686 777 }
a978ba68
JA
778
779 return ret;
ebac4655
JA
780}
781
906c8d75
JA
782/*
783 * Entry point for the thread based jobs. The process based jobs end up
784 * here as well, after a little setup.
785 */
ebac4655
JA
786static void *thread_main(void *data)
787{
69008999 788 unsigned long long runtime[2];
ebac4655 789 struct thread_data *td = data;
38d77cae 790 unsigned long elapsed;
a978ba68 791 int clear_state;
ebac4655 792
2dc1bbeb 793 if (!td->o.use_thread)
ebac4655
JA
794 setsid();
795
796 td->pid = getpid();
797
aea47d44
JA
798 INIT_LIST_HEAD(&td->io_u_freelist);
799 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 800 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44 801 INIT_LIST_HEAD(&td->io_log_list);
8de8f047 802 INIT_LIST_HEAD(&td->io_hist_list);
bb5d7d0b 803 td->io_hist_tree = RB_ROOT;
aea47d44 804
db1cd90b
JA
805 td_set_runstate(td, TD_INITIALIZED);
806 fio_sem_up(startup_sem);
807 fio_sem_down(td->mutex);
808
809 /*
810 * the ->mutex semaphore is now no longer used, close it to avoid
811 * eating a file descriptor
812 */
813 fio_sem_remove(td->mutex);
814
d84f8d49
JA
815 /*
816 * May alter parameters that init_io_u() will use, so we need to
817 * do this first.
818 */
819 if (init_iolog(td))
820 goto err;
821
ebac4655 822 if (init_io_u(td))
db1cd90b 823 goto err;
ebac4655 824
375b2695 825 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
e1161c32 826 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 827 goto err;
ebac4655
JA
828 }
829
ebac4655
JA
830 if (td->ioprio) {
831 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 832 td_verror(td, errno, "ioprio_set");
db1cd90b 833 goto err;
ebac4655
JA
834 }
835 }
836
2dc1bbeb 837 if (nice(td->o.nice) == -1) {
e1161c32 838 td_verror(td, errno, "nice");
db1cd90b 839 goto err;
b6f4d880
JA
840 }
841
2dc1bbeb 842 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 843 goto err;
37f56873 844
2dc1bbeb 845 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
846 goto err;
847
7d6c5283
JA
848 if (td_io_init(td))
849 goto err;
850
b5af8293
JA
851 if (open_files(td))
852 goto err;
853
68727076
JA
854 if (init_random_map(td))
855 goto err;
856
2dc1bbeb
JA
857 if (td->o.exec_prerun) {
858 if (system(td->o.exec_prerun) < 0)
69cfd7e0
JA
859 goto err;
860 }
4e0ba8af 861
69008999 862 fio_gettime(&td->epoch, NULL);
433afcb4 863 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
756867bd 864 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
865
866 runtime[0] = runtime[1] = 0;
a978ba68 867 clear_state = 0;
492158cf 868 while (keep_running(td)) {
02bcaa8c 869 fio_gettime(&td->start, NULL);
756867bd 870 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655 871
2dc1bbeb 872 if (td->o.ratemin)
756867bd 873 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655 874
a978ba68
JA
875 if (clear_state && clear_io_state(td))
876 break;
877
ebac4655
JA
878 prune_io_piece_log(td);
879
ba0fbe10 880 do_io(td);
ebac4655 881
a978ba68
JA
882 clear_state = 1;
883
38d77cae
JA
884 if (td_read(td) && td->io_bytes[DDIR_READ]) {
885 if (td->rw_end_set[DDIR_READ])
886 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
887 else
888 elapsed = utime_since_now(&td->start);
889
890 runtime[DDIR_READ] += elapsed;
891 }
892 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
893 if (td->rw_end_set[DDIR_WRITE])
894 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
895 else
896 elapsed = utime_since_now(&td->start);
897
898 runtime[DDIR_WRITE] += elapsed;
899 }
413dd459 900
ebac4655
JA
901 if (td->error || td->terminate)
902 break;
903
b67740d3
JA
904 if (td->o.verify == VERIFY_NONE ||
905 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
906 continue;
907
a978ba68
JA
908 if (clear_io_state(td))
909 break;
910
02bcaa8c 911 fio_gettime(&td->start, NULL);
ebac4655
JA
912
913 do_verify(td);
914
69008999 915 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
916
917 if (td->error || td->terminate)
918 break;
919 }
920
36dff966 921 update_rusage_stat(td);
47f0cc48
YHJT
922 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
923 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
756867bd
JA
924 td->ts.total_run_time = mtime_since_now(&td->epoch);
925 td->ts.io_bytes[0] = td->io_bytes[0];
926 td->ts.io_bytes[1] = td->io_bytes[1];
927
928 if (td->ts.bw_log)
929 finish_log(td, td->ts.bw_log, "bw");
930 if (td->ts.slat_log)
931 finish_log(td, td->ts.slat_log, "slat");
932 if (td->ts.clat_log)
933 finish_log(td, td->ts.clat_log, "clat");
2dc1bbeb
JA
934 if (td->o.exec_postrun) {
935 if (system(td->o.exec_postrun) < 0)
936 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
69cfd7e0 937 }
ebac4655
JA
938
939 if (exitall_on_terminate)
390c40e2 940 terminate_threads(td->groupid);
ebac4655
JA
941
942err:
5bf13a5a
JA
943 if (td->error)
944 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 945 close_files(td);
2866c82d 946 close_ioengine(td);
ebac4655 947 cleanup_io_u(td);
f29b25a3
JA
948
949 /*
950 * do this very late, it will log file closing as well
951 */
952 if (td->o.write_iolog_file)
953 write_iolog_close(td);
954
d23bb327 955 options_mem_free(td);
ebac4655 956 td_set_runstate(td, TD_EXITED);
43d76807 957 return (void *) (unsigned long) td->error;
ebac4655
JA
958}
959
906c8d75
JA
960/*
961 * We cannot pass the td data into a forked process, so attach the td and
962 * pass it to the thread worker.
963 */
a6418147 964static int fork_main(int shmid, int offset)
ebac4655
JA
965{
966 struct thread_data *td;
a6418147 967 void *data, *ret;
ebac4655
JA
968
969 data = shmat(shmid, NULL, 0);
970 if (data == (void *) -1) {
a6418147
JA
971 int __err = errno;
972
ebac4655 973 perror("shmat");
a6418147 974 return __err;
ebac4655
JA
975 }
976
977 td = data + offset * sizeof(struct thread_data);
a6418147 978 ret = thread_main(td);
ebac4655 979 shmdt(data);
43d76807 980 return (int) (unsigned long) ret;
ebac4655
JA
981}
982
906c8d75
JA
983/*
984 * Run over the job map and reap the threads that have exited, if any.
985 */
ebac4655
JA
986static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
987{
34572e28 988 struct thread_data *td;
fab6aa71 989 int i, cputhreads, pending, status, ret;
ebac4655
JA
990
991 /*
992 * reap exited threads (TD_EXITED -> TD_REAPED)
993 */
4d2413c6 994 pending = cputhreads = 0;
34572e28 995 for_each_td(td, i) {
3707f45b 996 int flags = 0;
a2f77c9f 997
84585003
JA
998 /*
999 * ->io_ops is NULL for a thread that has closed its
1000 * io engine
1001 */
ba0fbe10 1002 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0
JA
1003 cputhreads++;
1004
a2f77c9f
JA
1005 if (!td->pid || td->runstate == TD_REAPED)
1006 continue;
2dc1bbeb 1007 if (td->o.use_thread) {
3707f45b
JA
1008 if (td->runstate == TD_EXITED) {
1009 td_set_runstate(td, TD_REAPED);
1010 goto reaped;
1011 }
1012 continue;
1013 }
a2f77c9f
JA
1014
1015 flags = WNOHANG;
1016 if (td->runstate == TD_EXITED)
1017 flags = 0;
1018
1019 /*
1020 * check if someone quit or got killed in an unusual way
1021 */
1022 ret = waitpid(td->pid, &status, flags);
3707f45b 1023 if (ret < 0) {
a2f77c9f
JA
1024 if (errno == ECHILD) {
1025 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1026 td_set_runstate(td, TD_REAPED);
1027 goto reaped;
1028 }
1029 perror("waitpid");
1030 } else if (ret == td->pid) {
1031 if (WIFSIGNALED(status)) {
fab6aa71
JA
1032 int sig = WTERMSIG(status);
1033
d9244941
JA
1034 if (sig != SIGQUIT)
1035 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
fab6aa71
JA
1036 td_set_runstate(td, TD_REAPED);
1037 goto reaped;
1038 }
a2f77c9f
JA
1039 if (WIFEXITED(status)) {
1040 if (WEXITSTATUS(status) && !td->error)
1041 td->error = WEXITSTATUS(status);
a2f77c9f 1042
a2f77c9f
JA
1043 td_set_runstate(td, TD_REAPED);
1044 goto reaped;
a6418147
JA
1045 }
1046 }
ebac4655 1047
a2f77c9f
JA
1048 /*
1049 * thread is not dead, continue
1050 */
e48676ba 1051 pending++;
a2f77c9f 1052 continue;
fab6aa71 1053reaped:
2dc1bbeb 1054 if (td->o.use_thread) {
3707f45b
JA
1055 long ret;
1056
1057 if (pthread_join(td->thread, (void *) &ret))
1058 perror("pthread_join");
1059 }
1060
ebac4655 1061 (*nr_running)--;
2dc1bbeb
JA
1062 (*m_rate) -= td->o.ratemin;
1063 (*t_rate) -= td->o.rate;
e48676ba 1064 pending--;
a2f77c9f
JA
1065
1066 if (td->error)
1067 exit_value++;
ebac4655 1068 }
b990b5c0 1069
4d2413c6 1070 if (*nr_running == cputhreads && !pending)
390c40e2 1071 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1072}
1073
906c8d75
JA
1074/*
1075 * Main function for kicking off and reaping jobs, as needed.
1076 */
ebac4655
JA
1077static void run_threads(void)
1078{
ebac4655
JA
1079 struct thread_data *td;
1080 unsigned long spent;
1081 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1082
2f9ade3c
JA
1083 if (fio_pin_memory())
1084 return;
ebac4655 1085
c6ae0a5b 1086 if (!terse_output) {
9cedf167
JA
1087 printf("Starting ");
1088 if (nr_thread)
1089 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1090 if (nr_process) {
1091 if (nr_thread)
1092 printf(" and ");
1093 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1094 }
1095 printf("\n");
c6ae0a5b
JA
1096 fflush(stdout);
1097 }
c04f7ec3 1098
4efa970e
JA
1099 signal(SIGINT, sig_handler);
1100 signal(SIGALRM, sig_handler);
1101
ebac4655
JA
1102 todo = thread_number;
1103 nr_running = 0;
1104 nr_started = 0;
1105 m_rate = t_rate = 0;
1106
34572e28 1107 for_each_td(td, i) {
263e529f 1108 print_status_init(td->thread_number - 1);
ebac4655 1109
2dc1bbeb 1110 if (!td->o.create_serialize) {
380cf265 1111 init_disk_util(td);
ebac4655 1112 continue;
380cf265 1113 }
ebac4655
JA
1114
1115 /*
1116 * do file setup here so it happens sequentially,
1117 * we don't want X number of threads getting their
1118 * client data interspersed on disk
1119 */
53cdc686 1120 if (setup_files(td)) {
5bf13a5a
JA
1121 exit_value++;
1122 if (td->error)
1123 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
1124 td_set_runstate(td, TD_REAPED);
1125 todo--;
1126 }
380cf265
JA
1127
1128 init_disk_util(td);
ebac4655
JA
1129 }
1130
a2f77c9f
JA
1131 set_genesis_time();
1132
ebac4655 1133 while (todo) {
75154845
JA
1134 struct thread_data *map[MAX_JOBS];
1135 struct timeval this_start;
1136 int this_jobs = 0, left;
1137
ebac4655
JA
1138 /*
1139 * create threads (TD_NOT_CREATED -> TD_CREATED)
1140 */
34572e28 1141 for_each_td(td, i) {
ebac4655
JA
1142 if (td->runstate != TD_NOT_CREATED)
1143 continue;
1144
1145 /*
1146 * never got a chance to start, killed by other
1147 * thread for some reason
1148 */
1149 if (td->terminate) {
1150 todo--;
1151 continue;
1152 }
1153
2dc1bbeb 1154 if (td->o.start_delay) {
263e529f 1155 spent = mtime_since_genesis();
ebac4655 1156
2dc1bbeb 1157 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1158 continue;
1159 }
1160
2dc1bbeb 1161 if (td->o.stonewall && (nr_started || nr_running))
ebac4655
JA
1162 break;
1163
75154845
JA
1164 /*
1165 * Set state to created. Thread will transition
1166 * to TD_INITIALIZED when it's done setting up.
1167 */
ebac4655 1168 td_set_runstate(td, TD_CREATED);
75154845 1169 map[this_jobs++] = td;
ebac4655
JA
1170 nr_started++;
1171
2dc1bbeb 1172 if (td->o.use_thread) {
ebac4655
JA
1173 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1174 perror("thread_create");
1175 nr_started--;
c8bb6faf 1176 break;
ebac4655
JA
1177 }
1178 } else {
07739b57 1179 if (!fork()) {
a6418147
JA
1180 int ret = fork_main(shm_id, i);
1181
1182 exit(ret);
ebac4655
JA
1183 }
1184 }
07739b57 1185 fio_sem_down(startup_sem);
ebac4655
JA
1186 }
1187
1188 /*
75154845
JA
1189 * Wait for the started threads to transition to
1190 * TD_INITIALIZED.
ebac4655 1191 */
02bcaa8c 1192 fio_gettime(&this_start, NULL);
75154845 1193 left = this_jobs;
6ce15a32 1194 while (left && !fio_abort) {
75154845
JA
1195 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1196 break;
1197
1198 usleep(100000);
1199
1200 for (i = 0; i < this_jobs; i++) {
1201 td = map[i];
1202 if (!td)
1203 continue;
b6f4d880 1204 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1205 map[i] = NULL;
1206 left--;
b6f4d880
JA
1207 } else if (td->runstate >= TD_EXITED) {
1208 map[i] = NULL;
1209 left--;
1210 todo--;
1211 nr_running++; /* work-around... */
75154845
JA
1212 }
1213 }
1214 }
1215
1216 if (left) {
3b70d7e5 1217 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1218 for (i = 0; i < this_jobs; i++) {
1219 td = map[i];
1220 if (!td)
1221 continue;
1222 kill(td->pid, SIGTERM);
1223 }
1224 break;
1225 }
1226
1227 /*
b6f4d880 1228 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1229 */
34572e28 1230 for_each_td(td, i) {
75154845 1231 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1232 continue;
1233
1234 td_set_runstate(td, TD_RUNNING);
1235 nr_running++;
1236 nr_started--;
2dc1bbeb
JA
1237 m_rate += td->o.ratemin;
1238 t_rate += td->o.rate;
75154845 1239 todo--;
07739b57 1240 fio_sem_up(td->mutex);
ebac4655
JA
1241 }
1242
1243 reap_threads(&nr_running, &t_rate, &m_rate);
1244
1245 if (todo)
1246 usleep(100000);
1247 }
1248
1249 while (nr_running) {
1250 reap_threads(&nr_running, &t_rate, &m_rate);
1251 usleep(10000);
1252 }
1253
1254 update_io_ticks();
2f9ade3c 1255 fio_unpin_memory();
ebac4655
JA
1256}
1257
ebac4655
JA
1258int main(int argc, char *argv[])
1259{
29d610e1
JA
1260 long ps;
1261
dbe1125e
JA
1262 /*
1263 * We need locale for number printing, if it isn't set then just
1264 * go with the US format.
1265 */
1266 if (!getenv("LC_NUMERIC"))
1267 setlocale(LC_NUMERIC, "en_US");
1268
ebac4655
JA
1269 if (parse_options(argc, argv))
1270 return 1;
1271
4b472fa3
JA
1272 if (!thread_number)
1273 return 0;
ebac4655 1274
29d610e1
JA
1275 ps = sysconf(_SC_PAGESIZE);
1276 if (ps < 0) {
1277 log_err("Failed to get page size\n");
1278 return 1;
1279 }
1280
cfc99db7 1281 page_size = ps;
29d610e1
JA
1282 page_mask = ps - 1;
1283
bb3884d8
JA
1284 if (write_bw_log) {
1285 setup_log(&agg_io_log[DDIR_READ]);
1286 setup_log(&agg_io_log[DDIR_WRITE]);
1287 }
1288
07739b57
JA
1289 startup_sem = fio_sem_init(0);
1290
a2f77c9f
JA
1291 set_genesis_time();
1292
ebac4655
JA
1293 disk_util_timer_arm();
1294
1295 run_threads();
6ce15a32 1296
bb3884d8 1297 if (!fio_abort) {
6ce15a32 1298 show_run_stats();
bb3884d8
JA
1299 if (write_bw_log) {
1300 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1301 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1302 }
1303 }
ebac4655 1304
07739b57 1305 fio_sem_remove(startup_sem);
437c9b71 1306 return exit_value;
ebac4655 1307}