Fix bad alignment in buffer size allocation
[fio.git] / fio.c
CommitLineData
ebac4655
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
aae22ca7 5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
ebac4655 6 *
8e9fe637
JA
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
ebac4655 10 * This program is free software; you can redistribute it and/or modify
8e9fe637
JA
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
ebac4655
JA
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
ebac4655
JA
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
ebac4655
JA
27#include <signal.h>
28#include <time.h>
dbe1125e 29#include <locale.h>
36167d82 30#include <assert.h>
ebac4655
JA
31#include <sys/stat.h>
32#include <sys/wait.h>
33#include <sys/ipc.h>
34#include <sys/shm.h>
ebac4655
JA
35#include <sys/mman.h>
36
37#include "fio.h"
210f43b3 38#include "hash.h"
ebac4655 39
cfc99db7
JA
40unsigned long page_mask;
41unsigned long page_size;
29d610e1
JA
42#define ALIGN(buf) \
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
44
45int groupid = 0;
46int thread_number = 0;
9cedf167
JA
47int nr_process = 0;
48int nr_thread = 0;
ebac4655 49int shm_id = 0;
53cdc686 50int temp_stall_ts;
ebac4655 51
07739b57 52static struct fio_sem *startup_sem;
6ce15a32 53static volatile int fio_abort;
437c9b71 54static int exit_value;
ebac4655 55
bb3884d8
JA
56struct io_log *agg_io_log[2];
57
ebac4655 58#define TERMINATE_ALL (-1)
75154845 59#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 60
6ce15a32
JA
61static inline void td_set_runstate(struct thread_data *td, int runstate)
62{
63 td->runstate = runstate;
64}
65
390c40e2 66static void terminate_threads(int group_id)
ebac4655 67{
34572e28 68 struct thread_data *td;
ebac4655
JA
69 int i;
70
34572e28 71 for_each_td(td, i) {
ebac4655 72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
7a93ab19
JA
73 /*
74 * if the thread is running, just let it exit
75 */
76 if (td->runstate < TD_RUNNING)
c1302d44 77 kill(td->pid, SIGQUIT);
ebac4655 78 td->terminate = 1;
2dc1bbeb 79 td->o.start_delay = 0;
ebac4655
JA
80 }
81 }
82}
83
84static void sig_handler(int sig)
85{
86 switch (sig) {
87 case SIGALRM:
88 update_io_ticks();
89 disk_util_timer_arm();
90 print_thread_status();
91 break;
92 default:
6ce15a32 93 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 94 fflush(stdout);
390c40e2 95 terminate_threads(TERMINATE_ALL);
ebac4655
JA
96 break;
97 }
98}
99
906c8d75
JA
100/*
101 * Check if we are above the minimum rate given.
102 */
ebac4655
JA
103static int check_min_rate(struct thread_data *td, struct timeval *now)
104{
0904200b 105 unsigned long long bytes = 0;
4e991c23 106 unsigned long iops = 0;
ebac4655
JA
107 unsigned long spent;
108 unsigned long rate;
ebac4655 109
780bf1a1
JA
110 /*
111 * No minimum rate set, always ok
112 */
2dc1bbeb 113 if (!td->o.ratemin && !td->o.rate_iops_min)
780bf1a1
JA
114 return 0;
115
ebac4655
JA
116 /*
117 * allow a 2 second settle period in the beginning
118 */
119 if (mtime_since(&td->start, now) < 2000)
120 return 0;
121
4e991c23
JA
122 if (td_read(td)) {
123 iops += td->io_blocks[DDIR_READ];
0904200b 124 bytes += td->this_io_bytes[DDIR_READ];
4e991c23
JA
125 }
126 if (td_write(td)) {
127 iops += td->io_blocks[DDIR_WRITE];
0904200b 128 bytes += td->this_io_bytes[DDIR_WRITE];
4e991c23 129 }
0904200b 130
ebac4655
JA
131 /*
132 * if rate blocks is set, sample is running
133 */
4e991c23 134 if (td->rate_bytes || td->rate_blocks) {
ebac4655 135 spent = mtime_since(&td->lastrate, now);
2dc1bbeb 136 if (spent < td->o.ratecycle)
ebac4655
JA
137 return 0;
138
2dc1bbeb 139 if (td->o.rate) {
4e991c23
JA
140 /*
141 * check bandwidth specified rate
142 */
143 if (bytes < td->rate_bytes) {
2dc1bbeb 144 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
4e991c23
JA
145 return 1;
146 } else {
147 rate = (bytes - td->rate_bytes) / spent;
2dc1bbeb
JA
148 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
4e991c23
JA
150 return 1;
151 }
152 }
413dd459 153 } else {
4e991c23
JA
154 /*
155 * checks iops specified rate
156 */
2dc1bbeb
JA
157 if (iops < td->o.rate_iops) {
158 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
413dd459 159 return 1;
4e991c23
JA
160 } else {
161 rate = (iops - td->rate_blocks) / spent;
2dc1bbeb
JA
162 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
4e991c23 164 }
413dd459 165 }
ebac4655
JA
166 }
167 }
168
0904200b 169 td->rate_bytes = bytes;
4e991c23 170 td->rate_blocks = iops;
ebac4655
JA
171 memcpy(&td->lastrate, now, sizeof(*now));
172 return 0;
173}
174
175static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176{
2dc1bbeb 177 if (!td->o.timeout)
ebac4655 178 return 0;
2dc1bbeb 179 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
180 return 1;
181
182 return 0;
183}
184
906c8d75
JA
185/*
186 * When job exits, we can cancel the in-flight IO if we are using async
187 * io. Attempt to do so.
188 */
ebac4655
JA
189static void cleanup_pending_aio(struct thread_data *td)
190{
ebac4655 191 struct list_head *entry, *n;
ebac4655
JA
192 struct io_u *io_u;
193 int r;
194
195 /*
196 * get immediately available events, if any
197 */
d7762cf8 198 r = io_u_queued_complete(td, 0);
b2fdda43
JA
199 if (r < 0)
200 return;
ebac4655
JA
201
202 /*
203 * now cancel remaining active events
204 */
2866c82d 205 if (td->io_ops->cancel) {
ebac4655
JA
206 list_for_each_safe(entry, n, &td->io_u_busylist) {
207 io_u = list_entry(entry, struct io_u, list);
208
0c6e7517
JA
209 /*
210 * if the io_u isn't in flight, then that generally
211 * means someone leaked an io_u. complain but fix
212 * it up, so we don't stall here.
213 */
214 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215 log_err("fio: non-busy IO on busy list\n");
ebac4655 216 put_io_u(td, io_u);
0c6e7517
JA
217 } else {
218 r = td->io_ops->cancel(td, io_u);
219 if (!r)
220 put_io_u(td, io_u);
221 }
ebac4655
JA
222 }
223 }
224
97601024 225 if (td->cur_depth)
d7762cf8 226 r = io_u_queued_complete(td, td->cur_depth);
ebac4655
JA
227}
228
858a3d47
JA
229/*
230 * Helper to handle the final sync of a file. Works just like the normal
231 * io path, just does everything sync.
232 */
233static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234{
235 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
236 int ret;
237
238 if (!io_u)
239 return 1;
240
241 io_u->ddir = DDIR_SYNC;
242 io_u->file = f;
243
244 if (td_io_prep(td, io_u)) {
245 put_io_u(td, io_u);
246 return 1;
247 }
248
755200a3 249requeue:
858a3d47 250 ret = td_io_queue(td, io_u);
36167d82 251 if (ret < 0) {
e1161c32 252 td_verror(td, io_u->error, "td_io_queue");
858a3d47 253 put_io_u(td, io_u);
858a3d47 254 return 1;
36167d82 255 } else if (ret == FIO_Q_QUEUED) {
d7762cf8 256 if (io_u_queued_complete(td, 1) < 0)
36167d82 257 return 1;
36167d82
JA
258 } else if (ret == FIO_Q_COMPLETED) {
259 if (io_u->error) {
e1161c32 260 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
261 return 1;
262 }
858a3d47 263
d7762cf8 264 if (io_u_sync_complete(td, io_u) < 0)
b2fdda43 265 return 1;
755200a3
JA
266 } else if (ret == FIO_Q_BUSY) {
267 if (td_io_commit(td))
268 return 1;
269 goto requeue;
858a3d47
JA
270 }
271
272 return 0;
273}
274
906c8d75 275/*
34403fb1 276 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
277 * reads the blocks back in, and checks the crc/md5 of the data.
278 */
1e97cce9 279static void do_verify(struct thread_data *td)
ebac4655 280{
53cdc686 281 struct fio_file *f;
36167d82 282 struct io_u *io_u;
af52b345
JA
283 int ret, min_events;
284 unsigned int i;
e5b401d4
JA
285
286 /*
287 * sync io first and invalidate cache, to make sure we really
288 * read from disk.
289 */
290 for_each_file(td, f, i) {
3d7b485f
JA
291 if (!(f->flags & FIO_FILE_OPEN))
292 continue;
b2fdda43
JA
293 if (fio_io_sync(td, f))
294 break;
295 if (file_invalidate_cache(td, f))
296 break;
e5b401d4 297 }
ebac4655 298
b2fdda43
JA
299 if (td->error)
300 return;
301
ebac4655
JA
302 td_set_runstate(td, TD_VERIFYING);
303
36167d82
JA
304 io_u = NULL;
305 while (!td->terminate) {
5451792e
JA
306 int ret2;
307
ebac4655
JA
308 io_u = __get_io_u(td);
309 if (!io_u)
310 break;
311
069c2918
JA
312 if (runtime_exceeded(td, &io_u->start_time)) {
313 put_io_u(td, io_u);
90a87d4b 314 td->terminate = 1;
02bcaa8c 315 break;
069c2918 316 }
02bcaa8c 317
069c2918
JA
318 if (get_next_verify(td, io_u)) {
319 put_io_u(td, io_u);
ebac4655 320 break;
069c2918 321 }
ebac4655 322
069c2918
JA
323 if (td_io_prep(td, io_u)) {
324 put_io_u(td, io_u);
53cdc686 325 break;
069c2918 326 }
d7762cf8
JA
327
328 io_u->end_io = verify_io_u;
53cdc686 329
11786802 330 ret = td_io_queue(td, io_u);
36167d82
JA
331 switch (ret) {
332 case FIO_Q_COMPLETED:
333 if (io_u->error)
22819ec2 334 ret = -io_u->error;
9e9d164e 335 else if (io_u->resid) {
36167d82 336 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 337 struct fio_file *f = io_u->file;
ebac4655 338
9e9d164e
JA
339 /*
340 * zero read, fail
341 */
342 if (!bytes) {
343 td_verror(td, ENODATA, "full resid");
344 put_io_u(td, io_u);
345 break;
346 }
8400d9b2 347
36167d82
JA
348 io_u->xfer_buflen = io_u->resid;
349 io_u->xfer_buf += bytes;
8400d9b2 350 io_u->offset += bytes;
d460eb31 351 f->last_completed_pos = io_u->offset;
8400d9b2 352
30061b97
JA
353 td->ts.short_io_u[io_u->ddir]++;
354
d460eb31 355 if (io_u->offset == f->real_file_size)
8400d9b2
JA
356 goto sync_done;
357
11786802
JA
358 requeue_io_u(td, &io_u);
359 } else {
8400d9b2 360sync_done:
11786802
JA
361 ret = io_u_sync_complete(td, io_u);
362 if (ret < 0)
363 break;
36167d82 364 }
36167d82
JA
365 continue;
366 case FIO_Q_QUEUED:
367 break;
755200a3
JA
368 case FIO_Q_BUSY:
369 requeue_io_u(td, &io_u);
5451792e
JA
370 ret2 = td_io_commit(td);
371 if (ret2 < 0)
372 ret = ret2;
755200a3 373 break;
36167d82
JA
374 default:
375 assert(ret < 0);
e1161c32 376 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
377 break;
378 }
379
99784632 380 if (ret < 0 || td->error)
3af6ef39
JA
381 break;
382
ebac4655 383 /*
3af6ef39
JA
384 * if we can queue more, do so. but check if there are
385 * completed io_u's first.
ebac4655 386 */
97601024 387 min_events = 0;
e916b390 388 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 389 min_events = 1;
3af6ef39 390
2dc1bbeb
JA
391 if (td->cur_depth > td->o.iodepth_low)
392 min_events = td->cur_depth - td->o.iodepth_low;
e916b390
JA
393 }
394
3af6ef39
JA
395 /*
396 * Reap required number of io units, if any, and do the
397 * verification on them through the callback handler
398 */
d7762cf8 399 if (io_u_queued_complete(td, min_events) < 0)
ebac4655 400 break;
36167d82 401 }
ebac4655 402
c01c0395
JA
403 if (!td->error) {
404 min_events = td->cur_depth;
405
406 if (min_events)
407 ret = io_u_queued_complete(td, min_events);
408 } else
ebac4655
JA
409 cleanup_pending_aio(td);
410
411 td_set_runstate(td, TD_RUNNING);
412}
413
32cd46a0 414/*
906c8d75 415 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
416 * and reaps them, checking for rate and errors along the way.
417 */
ebac4655
JA
418static void do_io(struct thread_data *td)
419{
02bcaa8c 420 struct timeval s;
ebac4655 421 unsigned long usec;
af52b345
JA
422 unsigned int i;
423 int ret = 0;
ebac4655 424
5853e5a8
JA
425 td_set_runstate(td, TD_RUNNING);
426
7bb48f84 427 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
97601024
JA
428 struct timeval comp_time;
429 long bytes_done = 0;
84585003 430 int min_evts = 0;
ebac4655 431 struct io_u *io_u;
5451792e 432 int ret2;
ebac4655
JA
433
434 if (td->terminate)
435 break;
436
3d7c391d 437 io_u = get_io_u(td);
ebac4655
JA
438 if (!io_u)
439 break;
440
441 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
442
443 if (runtime_exceeded(td, &s)) {
444 put_io_u(td, io_u);
90a87d4b 445 td->terminate = 1;
97601024
JA
446 break;
447 }
36167d82 448
11786802 449 ret = td_io_queue(td, io_u);
36167d82
JA
450 switch (ret) {
451 case FIO_Q_COMPLETED:
5451792e
JA
452 if (io_u->error)
453 ret = -io_u->error;
9e9d164e 454 else if (io_u->resid) {
36167d82 455 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 456 struct fio_file *f = io_u->file;
36167d82 457
9e9d164e
JA
458 /*
459 * zero read, fail
460 */
461 if (!bytes) {
462 td_verror(td, ENODATA, "full resid");
463 put_io_u(td, io_u);
464 break;
465 }
466
cec6b55d 467 io_u->xfer_buflen = io_u->resid;
36167d82 468 io_u->xfer_buf += bytes;
5a7c5680 469 io_u->offset += bytes;
d460eb31 470 f->last_completed_pos = io_u->offset;
5a7c5680 471
30061b97
JA
472 td->ts.short_io_u[io_u->ddir]++;
473
d460eb31 474 if (io_u->offset == f->real_file_size)
5a7c5680
JA
475 goto sync_done;
476
11786802
JA
477 requeue_io_u(td, &io_u);
478 } else {
5a7c5680 479sync_done:
11786802
JA
480 fio_gettime(&comp_time, NULL);
481 bytes_done = io_u_sync_complete(td, io_u);
482 if (bytes_done < 0)
483 ret = bytes_done;
cec6b55d 484 }
36167d82
JA
485 break;
486 case FIO_Q_QUEUED:
7e77dd02
JA
487 /*
488 * if the engine doesn't have a commit hook,
489 * the io_u is really queued. if it does have such
490 * a hook, it has to call io_u_queued() itself.
491 */
492 if (td->io_ops->commit == NULL)
493 io_u_queued(td, io_u);
36167d82 494 break;
755200a3
JA
495 case FIO_Q_BUSY:
496 requeue_io_u(td, &io_u);
5451792e
JA
497 ret2 = td_io_commit(td);
498 if (ret2 < 0)
499 ret = ret2;
755200a3 500 break;
36167d82
JA
501 default:
502 assert(ret < 0);
503 put_io_u(td, io_u);
504 break;
ebac4655
JA
505 }
506
99784632 507 if (ret < 0 || td->error)
36167d82
JA
508 break;
509
97601024
JA
510 /*
511 * See if we need to complete some commands
512 */
755200a3 513 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 514 min_evts = 0;
e916b390 515 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 516 min_evts = 1;
ebac4655 517
2dc1bbeb
JA
518 if (td->cur_depth > td->o.iodepth_low)
519 min_evts = td->cur_depth - td->o.iodepth_low;
e916b390
JA
520 }
521
97601024 522 fio_gettime(&comp_time, NULL);
d7762cf8 523 bytes_done = io_u_queued_complete(td, min_evts);
97601024 524 if (bytes_done < 0)
36167d82 525 break;
ebac4655
JA
526 }
527
97601024
JA
528 if (!bytes_done)
529 continue;
530
ebac4655
JA
531 /*
532 * the rate is batched for now, it should work for batches
533 * of completions except the very first one which may look
534 * a little bursty
535 */
97601024 536 usec = utime_since(&s, &comp_time);
ebac4655 537
413dd459 538 rate_throttle(td, usec, bytes_done);
ebac4655 539
97601024 540 if (check_min_rate(td, &comp_time)) {
98aa62d8 541 if (exitall_on_terminate)
390c40e2 542 terminate_threads(td->groupid);
e1161c32 543 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
544 break;
545 }
546
2dc1bbeb 547 if (td->o.thinktime) {
9c1f7434
JA
548 unsigned long long b;
549
550 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 551 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
552 int left;
553
2dc1bbeb
JA
554 if (td->o.thinktime_spin)
555 __usec_sleep(td->o.thinktime_spin);
48097d5c 556
2dc1bbeb 557 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
558 if (left)
559 usec_sleep(td, left);
560 }
9c1f7434 561 }
ebac4655
JA
562 }
563
4d2413c6 564 if (!td->error) {
3d7c391d
JA
565 struct fio_file *f;
566
c01c0395
JA
567 i = td->cur_depth;
568 if (i)
569 ret = io_u_queued_complete(td, i);
ebac4655 570
2dc1bbeb 571 if (should_fsync(td) && td->o.end_fsync) {
84585003 572 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
573
574 for_each_file(td, f, i) {
575 if (!(f->flags & FIO_FILE_OPEN))
576 continue;
858a3d47 577 fio_io_sync(td, f);
3d7b485f 578 }
84585003 579 }
c01c0395
JA
580 } else
581 cleanup_pending_aio(td);
ebac4655
JA
582}
583
ebac4655
JA
584static void cleanup_io_u(struct thread_data *td)
585{
586 struct list_head *entry, *n;
587 struct io_u *io_u;
588
589 list_for_each_safe(entry, n, &td->io_u_freelist) {
590 io_u = list_entry(entry, struct io_u, list);
591
592 list_del(&io_u->list);
593 free(io_u);
594 }
595
2f9ade3c 596 free_io_mem(td);
ebac4655
JA
597}
598
6b9cea23
JA
599/*
600 * "randomly" fill the buffer contents
601 */
e9459e5a 602static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
6b9cea23 603{
210f43b3 604 long *ptr = io_u->buf;
6b9cea23 605
e9459e5a
JA
606 if (!td->o.zero_buffers) {
607 while ((void *) ptr - io_u->buf < max_bs) {
210f43b3 608 *ptr = rand() * GOLDEN_RATIO_PRIME;
e9459e5a
JA
609 ptr++;
610 }
611 } else
612 memset(ptr, 0, max_bs);
6b9cea23
JA
613}
614
ebac4655
JA
615static int init_io_u(struct thread_data *td)
616{
617 struct io_u *io_u;
a00735e6 618 unsigned int max_bs;
ebac4655
JA
619 int i, max_units;
620 char *p;
621
2866c82d 622 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
623 max_units = 1;
624 else
2dc1bbeb 625 max_units = td->o.iodepth;
ebac4655 626
2dc1bbeb 627 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
c3990645 628 td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
74b025b0 629
2dc1bbeb
JA
630 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
631 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
4afbf66f
JA
632 else if (td->orig_buffer_size & page_mask)
633 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
ebac4655 634
c3990645
JA
635 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
636 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
637 return 1;
638 }
639
2f9ade3c
JA
640 if (allocate_io_mem(td))
641 return 1;
ebac4655 642
ebac4655
JA
643 p = ALIGN(td->orig_buffer);
644 for (i = 0; i < max_units; i++) {
db1cd90b
JA
645 if (td->terminate)
646 return 1;
ebac4655
JA
647 io_u = malloc(sizeof(*io_u));
648 memset(io_u, 0, sizeof(*io_u));
649 INIT_LIST_HEAD(&io_u->list);
650
a00735e6 651 io_u->buf = p + max_bs * i;
e9459e5a 652
660a1cb5 653 if (td_write(td))
e9459e5a 654 fill_io_buf(td, io_u, max_bs);
6b9cea23 655
b1ff3403 656 io_u->index = i;
0c6e7517 657 io_u->flags = IO_U_F_FREE;
ebac4655
JA
658 list_add(&io_u->list, &td->io_u_freelist);
659 }
660
433afcb4
JA
661 io_u_init_timeout();
662
ebac4655
JA
663 return 0;
664}
665
da86774e
JA
666static int switch_ioscheduler(struct thread_data *td)
667{
668 char tmp[256], tmp2[128];
669 FILE *f;
670 int ret;
671
ba0fbe10 672 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
673 return 0;
674
da86774e
JA
675 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
676
677 f = fopen(tmp, "r+");
678 if (!f) {
cbf5c1d7
JA
679 if (errno == ENOENT) {
680 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
681 return 0;
682 }
683 td_verror(td, errno, "fopen iosched");
da86774e
JA
684 return 1;
685 }
686
687 /*
688 * Set io scheduler.
689 */
2dc1bbeb 690 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 691 if (ferror(f) || ret != 1) {
e1161c32 692 td_verror(td, errno, "fwrite");
da86774e
JA
693 fclose(f);
694 return 1;
695 }
696
697 rewind(f);
698
699 /*
700 * Read back and check that the selected scheduler is now the default.
701 */
702 ret = fread(tmp, 1, sizeof(tmp), f);
703 if (ferror(f) || ret < 0) {
e1161c32 704 td_verror(td, errno, "fread");
da86774e
JA
705 fclose(f);
706 return 1;
707 }
708
2dc1bbeb 709 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 710 if (!strstr(tmp, tmp2)) {
2dc1bbeb 711 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 712 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
713 fclose(f);
714 return 1;
715 }
716
717 fclose(f);
718 return 0;
719}
720
492158cf
JA
721static int keep_running(struct thread_data *td)
722{
723 unsigned long long io_done;
724
725 if (td->o.time_based)
726 return 1;
727 if (td->o.loops) {
728 td->o.loops--;
729 return 1;
730 }
731
732 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
733 if (io_done < td->o.size)
734 return 1;
735
736 return 0;
737}
738
a978ba68 739static int clear_io_state(struct thread_data *td)
ebac4655 740{
53cdc686 741 struct fio_file *f;
af52b345
JA
742 unsigned int i;
743 int ret;
ebac4655 744
756867bd 745 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 746 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 747 td->zone_bytes = 0;
d7d3b49b 748 td->rate_bytes = 0;
4e991c23 749 td->rate_blocks = 0;
38d77cae 750 td->rw_end_set[0] = td->rw_end_set[1] = 0;
ebac4655 751
c1324df1
JA
752 td->last_was_sync = 0;
753
281f9b63
JA
754 td->nr_done_files = 0;
755
a978ba68
JA
756 for_each_file(td, f, i)
757 td_io_close_file(td, f);
53cdc686 758
a978ba68
JA
759 ret = 0;
760 for_each_file(td, f, i) {
281f9b63 761 f->flags &= ~FIO_FILE_DONE;
a978ba68
JA
762 ret = td_io_open_file(td, f);
763 if (ret)
764 break;
53cdc686 765 }
a978ba68
JA
766
767 return ret;
ebac4655
JA
768}
769
906c8d75
JA
770/*
771 * Entry point for the thread based jobs. The process based jobs end up
772 * here as well, after a little setup.
773 */
ebac4655
JA
774static void *thread_main(void *data)
775{
69008999 776 unsigned long long runtime[2];
ebac4655 777 struct thread_data *td = data;
38d77cae 778 unsigned long elapsed;
a978ba68 779 int clear_state;
ebac4655 780
2dc1bbeb 781 if (!td->o.use_thread)
ebac4655
JA
782 setsid();
783
784 td->pid = getpid();
785
aea47d44
JA
786 INIT_LIST_HEAD(&td->io_u_freelist);
787 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 788 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44 789 INIT_LIST_HEAD(&td->io_log_list);
8de8f047 790 INIT_LIST_HEAD(&td->io_hist_list);
bb5d7d0b 791 td->io_hist_tree = RB_ROOT;
aea47d44 792
db1cd90b
JA
793 td_set_runstate(td, TD_INITIALIZED);
794 fio_sem_up(startup_sem);
795 fio_sem_down(td->mutex);
796
797 /*
798 * the ->mutex semaphore is now no longer used, close it to avoid
799 * eating a file descriptor
800 */
801 fio_sem_remove(td->mutex);
802
d84f8d49
JA
803 /*
804 * May alter parameters that init_io_u() will use, so we need to
805 * do this first.
806 */
807 if (init_iolog(td))
808 goto err;
809
ebac4655 810 if (init_io_u(td))
db1cd90b 811 goto err;
ebac4655 812
375b2695 813 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
e1161c32 814 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 815 goto err;
ebac4655
JA
816 }
817
ebac4655
JA
818 if (td->ioprio) {
819 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 820 td_verror(td, errno, "ioprio_set");
db1cd90b 821 goto err;
ebac4655
JA
822 }
823 }
824
2dc1bbeb 825 if (nice(td->o.nice) == -1) {
e1161c32 826 td_verror(td, errno, "nice");
db1cd90b 827 goto err;
b6f4d880
JA
828 }
829
2dc1bbeb 830 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 831 goto err;
37f56873 832
2dc1bbeb 833 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
834 goto err;
835
7d6c5283
JA
836 if (td_io_init(td))
837 goto err;
838
b5af8293
JA
839 if (open_files(td))
840 goto err;
841
68727076
JA
842 if (init_random_map(td))
843 goto err;
844
2dc1bbeb
JA
845 if (td->o.exec_prerun) {
846 if (system(td->o.exec_prerun) < 0)
69cfd7e0
JA
847 goto err;
848 }
4e0ba8af 849
69008999 850 fio_gettime(&td->epoch, NULL);
433afcb4 851 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
756867bd 852 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
853
854 runtime[0] = runtime[1] = 0;
a978ba68 855 clear_state = 0;
492158cf 856 while (keep_running(td)) {
02bcaa8c 857 fio_gettime(&td->start, NULL);
756867bd 858 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655 859
2dc1bbeb 860 if (td->o.ratemin)
756867bd 861 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655 862
a978ba68
JA
863 if (clear_state && clear_io_state(td))
864 break;
865
ebac4655
JA
866 prune_io_piece_log(td);
867
ba0fbe10 868 do_io(td);
ebac4655 869
a978ba68
JA
870 clear_state = 1;
871
38d77cae
JA
872 if (td_read(td) && td->io_bytes[DDIR_READ]) {
873 if (td->rw_end_set[DDIR_READ])
874 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
875 else
876 elapsed = utime_since_now(&td->start);
877
878 runtime[DDIR_READ] += elapsed;
879 }
880 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
881 if (td->rw_end_set[DDIR_WRITE])
882 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
883 else
884 elapsed = utime_since_now(&td->start);
885
886 runtime[DDIR_WRITE] += elapsed;
887 }
413dd459 888
ebac4655
JA
889 if (td->error || td->terminate)
890 break;
891
2dc1bbeb 892 if (td->o.verify == VERIFY_NONE)
ebac4655
JA
893 continue;
894
a978ba68
JA
895 if (clear_io_state(td))
896 break;
897
02bcaa8c 898 fio_gettime(&td->start, NULL);
ebac4655
JA
899
900 do_verify(td);
901
69008999 902 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
903
904 if (td->error || td->terminate)
905 break;
906 }
907
36dff966 908 update_rusage_stat(td);
756867bd
JA
909 td->ts.runtime[0] = runtime[0] / 1000;
910 td->ts.runtime[1] = runtime[1] / 1000;
911 td->ts.total_run_time = mtime_since_now(&td->epoch);
912 td->ts.io_bytes[0] = td->io_bytes[0];
913 td->ts.io_bytes[1] = td->io_bytes[1];
914
915 if (td->ts.bw_log)
916 finish_log(td, td->ts.bw_log, "bw");
917 if (td->ts.slat_log)
918 finish_log(td, td->ts.slat_log, "slat");
919 if (td->ts.clat_log)
920 finish_log(td, td->ts.clat_log, "clat");
2dc1bbeb 921 if (td->o.write_iolog_file)
843a7413 922 write_iolog_close(td);
2dc1bbeb
JA
923 if (td->o.exec_postrun) {
924 if (system(td->o.exec_postrun) < 0)
925 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
69cfd7e0 926 }
ebac4655
JA
927
928 if (exitall_on_terminate)
390c40e2 929 terminate_threads(td->groupid);
ebac4655
JA
930
931err:
5bf13a5a
JA
932 if (td->error)
933 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 934 close_files(td);
2866c82d 935 close_ioengine(td);
ebac4655 936 cleanup_io_u(td);
d23bb327 937 options_mem_free(td);
ebac4655 938 td_set_runstate(td, TD_EXITED);
43d76807 939 return (void *) (unsigned long) td->error;
ebac4655
JA
940}
941
906c8d75
JA
942/*
943 * We cannot pass the td data into a forked process, so attach the td and
944 * pass it to the thread worker.
945 */
a6418147 946static int fork_main(int shmid, int offset)
ebac4655
JA
947{
948 struct thread_data *td;
a6418147 949 void *data, *ret;
ebac4655
JA
950
951 data = shmat(shmid, NULL, 0);
952 if (data == (void *) -1) {
a6418147
JA
953 int __err = errno;
954
ebac4655 955 perror("shmat");
a6418147 956 return __err;
ebac4655
JA
957 }
958
959 td = data + offset * sizeof(struct thread_data);
a6418147 960 ret = thread_main(td);
ebac4655 961 shmdt(data);
43d76807 962 return (int) (unsigned long) ret;
ebac4655
JA
963}
964
906c8d75
JA
965/*
966 * Run over the job map and reap the threads that have exited, if any.
967 */
ebac4655
JA
968static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
969{
34572e28 970 struct thread_data *td;
fab6aa71 971 int i, cputhreads, pending, status, ret;
ebac4655
JA
972
973 /*
974 * reap exited threads (TD_EXITED -> TD_REAPED)
975 */
4d2413c6 976 pending = cputhreads = 0;
34572e28 977 for_each_td(td, i) {
3707f45b 978 int flags = 0;
a2f77c9f 979
84585003
JA
980 /*
981 * ->io_ops is NULL for a thread that has closed its
982 * io engine
983 */
ba0fbe10 984 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0
JA
985 cputhreads++;
986
a2f77c9f
JA
987 if (!td->pid || td->runstate == TD_REAPED)
988 continue;
2dc1bbeb 989 if (td->o.use_thread) {
3707f45b
JA
990 if (td->runstate == TD_EXITED) {
991 td_set_runstate(td, TD_REAPED);
992 goto reaped;
993 }
994 continue;
995 }
a2f77c9f
JA
996
997 flags = WNOHANG;
998 if (td->runstate == TD_EXITED)
999 flags = 0;
1000
1001 /*
1002 * check if someone quit or got killed in an unusual way
1003 */
1004 ret = waitpid(td->pid, &status, flags);
3707f45b 1005 if (ret < 0) {
a2f77c9f
JA
1006 if (errno == ECHILD) {
1007 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1008 td_set_runstate(td, TD_REAPED);
1009 goto reaped;
1010 }
1011 perror("waitpid");
1012 } else if (ret == td->pid) {
1013 if (WIFSIGNALED(status)) {
fab6aa71
JA
1014 int sig = WTERMSIG(status);
1015
d9244941
JA
1016 if (sig != SIGQUIT)
1017 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
fab6aa71
JA
1018 td_set_runstate(td, TD_REAPED);
1019 goto reaped;
1020 }
a2f77c9f
JA
1021 if (WIFEXITED(status)) {
1022 if (WEXITSTATUS(status) && !td->error)
1023 td->error = WEXITSTATUS(status);
a2f77c9f 1024
a2f77c9f
JA
1025 td_set_runstate(td, TD_REAPED);
1026 goto reaped;
a6418147
JA
1027 }
1028 }
ebac4655 1029
a2f77c9f
JA
1030 /*
1031 * thread is not dead, continue
1032 */
e48676ba 1033 pending++;
a2f77c9f 1034 continue;
fab6aa71 1035reaped:
2dc1bbeb 1036 if (td->o.use_thread) {
3707f45b
JA
1037 long ret;
1038
1039 if (pthread_join(td->thread, (void *) &ret))
1040 perror("pthread_join");
1041 }
1042
ebac4655 1043 (*nr_running)--;
2dc1bbeb
JA
1044 (*m_rate) -= td->o.ratemin;
1045 (*t_rate) -= td->o.rate;
e48676ba 1046 pending--;
a2f77c9f
JA
1047
1048 if (td->error)
1049 exit_value++;
ebac4655 1050 }
b990b5c0 1051
4d2413c6 1052 if (*nr_running == cputhreads && !pending)
390c40e2 1053 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1054}
1055
906c8d75
JA
1056/*
1057 * Main function for kicking off and reaping jobs, as needed.
1058 */
ebac4655
JA
1059static void run_threads(void)
1060{
ebac4655
JA
1061 struct thread_data *td;
1062 unsigned long spent;
1063 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1064
2f9ade3c
JA
1065 if (fio_pin_memory())
1066 return;
ebac4655 1067
c6ae0a5b 1068 if (!terse_output) {
9cedf167
JA
1069 printf("Starting ");
1070 if (nr_thread)
1071 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1072 if (nr_process) {
1073 if (nr_thread)
1074 printf(" and ");
1075 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1076 }
1077 printf("\n");
c6ae0a5b
JA
1078 fflush(stdout);
1079 }
c04f7ec3 1080
4efa970e
JA
1081 signal(SIGINT, sig_handler);
1082 signal(SIGALRM, sig_handler);
1083
ebac4655
JA
1084 todo = thread_number;
1085 nr_running = 0;
1086 nr_started = 0;
1087 m_rate = t_rate = 0;
1088
34572e28 1089 for_each_td(td, i) {
263e529f 1090 print_status_init(td->thread_number - 1);
ebac4655 1091
2dc1bbeb 1092 if (!td->o.create_serialize) {
380cf265 1093 init_disk_util(td);
ebac4655 1094 continue;
380cf265 1095 }
ebac4655
JA
1096
1097 /*
1098 * do file setup here so it happens sequentially,
1099 * we don't want X number of threads getting their
1100 * client data interspersed on disk
1101 */
53cdc686 1102 if (setup_files(td)) {
5bf13a5a
JA
1103 exit_value++;
1104 if (td->error)
1105 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
1106 td_set_runstate(td, TD_REAPED);
1107 todo--;
1108 }
380cf265
JA
1109
1110 init_disk_util(td);
ebac4655
JA
1111 }
1112
a2f77c9f
JA
1113 set_genesis_time();
1114
ebac4655 1115 while (todo) {
75154845
JA
1116 struct thread_data *map[MAX_JOBS];
1117 struct timeval this_start;
1118 int this_jobs = 0, left;
1119
ebac4655
JA
1120 /*
1121 * create threads (TD_NOT_CREATED -> TD_CREATED)
1122 */
34572e28 1123 for_each_td(td, i) {
ebac4655
JA
1124 if (td->runstate != TD_NOT_CREATED)
1125 continue;
1126
1127 /*
1128 * never got a chance to start, killed by other
1129 * thread for some reason
1130 */
1131 if (td->terminate) {
1132 todo--;
1133 continue;
1134 }
1135
2dc1bbeb 1136 if (td->o.start_delay) {
263e529f 1137 spent = mtime_since_genesis();
ebac4655 1138
2dc1bbeb 1139 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1140 continue;
1141 }
1142
2dc1bbeb 1143 if (td->o.stonewall && (nr_started || nr_running))
ebac4655
JA
1144 break;
1145
75154845
JA
1146 /*
1147 * Set state to created. Thread will transition
1148 * to TD_INITIALIZED when it's done setting up.
1149 */
ebac4655 1150 td_set_runstate(td, TD_CREATED);
75154845 1151 map[this_jobs++] = td;
ebac4655
JA
1152 nr_started++;
1153
2dc1bbeb 1154 if (td->o.use_thread) {
ebac4655
JA
1155 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1156 perror("thread_create");
1157 nr_started--;
c8bb6faf 1158 break;
ebac4655
JA
1159 }
1160 } else {
07739b57 1161 if (!fork()) {
a6418147
JA
1162 int ret = fork_main(shm_id, i);
1163
1164 exit(ret);
ebac4655
JA
1165 }
1166 }
07739b57 1167 fio_sem_down(startup_sem);
ebac4655
JA
1168 }
1169
1170 /*
75154845
JA
1171 * Wait for the started threads to transition to
1172 * TD_INITIALIZED.
ebac4655 1173 */
02bcaa8c 1174 fio_gettime(&this_start, NULL);
75154845 1175 left = this_jobs;
6ce15a32 1176 while (left && !fio_abort) {
75154845
JA
1177 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1178 break;
1179
1180 usleep(100000);
1181
1182 for (i = 0; i < this_jobs; i++) {
1183 td = map[i];
1184 if (!td)
1185 continue;
b6f4d880 1186 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1187 map[i] = NULL;
1188 left--;
b6f4d880
JA
1189 } else if (td->runstate >= TD_EXITED) {
1190 map[i] = NULL;
1191 left--;
1192 todo--;
1193 nr_running++; /* work-around... */
75154845
JA
1194 }
1195 }
1196 }
1197
1198 if (left) {
3b70d7e5 1199 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1200 for (i = 0; i < this_jobs; i++) {
1201 td = map[i];
1202 if (!td)
1203 continue;
1204 kill(td->pid, SIGTERM);
1205 }
1206 break;
1207 }
1208
1209 /*
b6f4d880 1210 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1211 */
34572e28 1212 for_each_td(td, i) {
75154845 1213 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1214 continue;
1215
1216 td_set_runstate(td, TD_RUNNING);
1217 nr_running++;
1218 nr_started--;
2dc1bbeb
JA
1219 m_rate += td->o.ratemin;
1220 t_rate += td->o.rate;
75154845 1221 todo--;
07739b57 1222 fio_sem_up(td->mutex);
ebac4655
JA
1223 }
1224
1225 reap_threads(&nr_running, &t_rate, &m_rate);
1226
1227 if (todo)
1228 usleep(100000);
1229 }
1230
1231 while (nr_running) {
1232 reap_threads(&nr_running, &t_rate, &m_rate);
1233 usleep(10000);
1234 }
1235
1236 update_io_ticks();
2f9ade3c 1237 fio_unpin_memory();
ebac4655
JA
1238}
1239
ebac4655
JA
1240int main(int argc, char *argv[])
1241{
29d610e1
JA
1242 long ps;
1243
dbe1125e
JA
1244 /*
1245 * We need locale for number printing, if it isn't set then just
1246 * go with the US format.
1247 */
1248 if (!getenv("LC_NUMERIC"))
1249 setlocale(LC_NUMERIC, "en_US");
1250
ebac4655
JA
1251 if (parse_options(argc, argv))
1252 return 1;
1253
4b472fa3
JA
1254 if (!thread_number)
1255 return 0;
ebac4655 1256
29d610e1
JA
1257 ps = sysconf(_SC_PAGESIZE);
1258 if (ps < 0) {
1259 log_err("Failed to get page size\n");
1260 return 1;
1261 }
1262
cfc99db7 1263 page_size = ps;
29d610e1
JA
1264 page_mask = ps - 1;
1265
bb3884d8
JA
1266 if (write_bw_log) {
1267 setup_log(&agg_io_log[DDIR_READ]);
1268 setup_log(&agg_io_log[DDIR_WRITE]);
1269 }
1270
07739b57
JA
1271 startup_sem = fio_sem_init(0);
1272
a2f77c9f
JA
1273 set_genesis_time();
1274
ebac4655
JA
1275 disk_util_timer_arm();
1276
1277 run_threads();
6ce15a32 1278
bb3884d8 1279 if (!fio_abort) {
6ce15a32 1280 show_run_stats();
bb3884d8
JA
1281 if (write_bw_log) {
1282 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1283 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1284 }
1285 }
ebac4655 1286
07739b57 1287 fio_sem_remove(startup_sem);
437c9b71 1288 return exit_value;
ebac4655 1289}