Move os/arch/compiler headers into directories
[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"
ebac4655 38
cfc99db7
JA
39unsigned long page_mask;
40unsigned long page_size;
29d610e1
JA
41#define ALIGN(buf) \
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
43
44int groupid = 0;
45int thread_number = 0;
9cedf167
JA
46int nr_process = 0;
47int nr_thread = 0;
ebac4655 48int shm_id = 0;
53cdc686 49int temp_stall_ts;
ebac4655 50
07739b57 51static struct fio_sem *startup_sem;
6ce15a32 52static volatile int fio_abort;
437c9b71 53static int exit_value;
ebac4655 54
bb3884d8
JA
55struct io_log *agg_io_log[2];
56
ebac4655 57#define TERMINATE_ALL (-1)
75154845 58#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 59
6ce15a32
JA
60static inline void td_set_runstate(struct thread_data *td, int runstate)
61{
62 td->runstate = runstate;
63}
64
390c40e2 65static void terminate_threads(int group_id)
ebac4655 66{
34572e28 67 struct thread_data *td;
ebac4655
JA
68 int i;
69
34572e28 70 for_each_td(td, i) {
ebac4655 71 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
c1302d44
JA
72 /*
73 * if the thread is running, just let it exit
74 */
75 if (td->runstate < TD_RUNNING)
76 kill(td->pid, SIGQUIT);
ebac4655 77 td->terminate = 1;
2dc1bbeb 78 td->o.start_delay = 0;
ebac4655
JA
79 }
80 }
81}
82
83static void sig_handler(int sig)
84{
85 switch (sig) {
86 case SIGALRM:
87 update_io_ticks();
88 disk_util_timer_arm();
89 print_thread_status();
90 break;
91 default:
6ce15a32 92 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 93 fflush(stdout);
390c40e2 94 terminate_threads(TERMINATE_ALL);
ebac4655
JA
95 break;
96 }
97}
98
906c8d75
JA
99/*
100 * Check if we are above the minimum rate given.
101 */
ebac4655
JA
102static int check_min_rate(struct thread_data *td, struct timeval *now)
103{
0904200b 104 unsigned long long bytes = 0;
4e991c23 105 unsigned long iops = 0;
ebac4655
JA
106 unsigned long spent;
107 unsigned long rate;
ebac4655 108
780bf1a1
JA
109 /*
110 * No minimum rate set, always ok
111 */
2dc1bbeb 112 if (!td->o.ratemin && !td->o.rate_iops_min)
780bf1a1
JA
113 return 0;
114
ebac4655
JA
115 /*
116 * allow a 2 second settle period in the beginning
117 */
118 if (mtime_since(&td->start, now) < 2000)
119 return 0;
120
4e991c23
JA
121 if (td_read(td)) {
122 iops += td->io_blocks[DDIR_READ];
0904200b 123 bytes += td->this_io_bytes[DDIR_READ];
4e991c23
JA
124 }
125 if (td_write(td)) {
126 iops += td->io_blocks[DDIR_WRITE];
0904200b 127 bytes += td->this_io_bytes[DDIR_WRITE];
4e991c23 128 }
0904200b 129
ebac4655
JA
130 /*
131 * if rate blocks is set, sample is running
132 */
4e991c23 133 if (td->rate_bytes || td->rate_blocks) {
ebac4655 134 spent = mtime_since(&td->lastrate, now);
2dc1bbeb 135 if (spent < td->o.ratecycle)
ebac4655
JA
136 return 0;
137
2dc1bbeb 138 if (td->o.rate) {
4e991c23
JA
139 /*
140 * check bandwidth specified rate
141 */
142 if (bytes < td->rate_bytes) {
2dc1bbeb 143 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
4e991c23
JA
144 return 1;
145 } else {
146 rate = (bytes - td->rate_bytes) / spent;
2dc1bbeb
JA
147 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
148 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
4e991c23
JA
149 return 1;
150 }
151 }
413dd459 152 } else {
4e991c23
JA
153 /*
154 * checks iops specified rate
155 */
2dc1bbeb
JA
156 if (iops < td->o.rate_iops) {
157 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
413dd459 158 return 1;
4e991c23
JA
159 } else {
160 rate = (iops - td->rate_blocks) / spent;
2dc1bbeb
JA
161 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
162 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
4e991c23 163 }
413dd459 164 }
ebac4655
JA
165 }
166 }
167
0904200b 168 td->rate_bytes = bytes;
4e991c23 169 td->rate_blocks = iops;
ebac4655
JA
170 memcpy(&td->lastrate, now, sizeof(*now));
171 return 0;
172}
173
174static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
175{
2dc1bbeb 176 if (!td->o.timeout)
ebac4655 177 return 0;
2dc1bbeb 178 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
179 return 1;
180
181 return 0;
182}
183
906c8d75
JA
184/*
185 * When job exits, we can cancel the in-flight IO if we are using async
186 * io. Attempt to do so.
187 */
ebac4655
JA
188static void cleanup_pending_aio(struct thread_data *td)
189{
ebac4655 190 struct list_head *entry, *n;
ebac4655
JA
191 struct io_u *io_u;
192 int r;
193
194 /*
195 * get immediately available events, if any
196 */
d7762cf8 197 r = io_u_queued_complete(td, 0);
b2fdda43
JA
198 if (r < 0)
199 return;
ebac4655
JA
200
201 /*
202 * now cancel remaining active events
203 */
2866c82d 204 if (td->io_ops->cancel) {
ebac4655
JA
205 list_for_each_safe(entry, n, &td->io_u_busylist) {
206 io_u = list_entry(entry, struct io_u, list);
207
0c6e7517
JA
208 /*
209 * if the io_u isn't in flight, then that generally
210 * means someone leaked an io_u. complain but fix
211 * it up, so we don't stall here.
212 */
213 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
214 log_err("fio: non-busy IO on busy list\n");
ebac4655 215 put_io_u(td, io_u);
0c6e7517
JA
216 } else {
217 r = td->io_ops->cancel(td, io_u);
218 if (!r)
219 put_io_u(td, io_u);
220 }
ebac4655
JA
221 }
222 }
223
97601024 224 if (td->cur_depth)
d7762cf8 225 r = io_u_queued_complete(td, td->cur_depth);
ebac4655
JA
226}
227
858a3d47
JA
228/*
229 * Helper to handle the final sync of a file. Works just like the normal
230 * io path, just does everything sync.
231 */
232static int fio_io_sync(struct thread_data *td, struct fio_file *f)
233{
234 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
235 int ret;
236
237 if (!io_u)
238 return 1;
239
240 io_u->ddir = DDIR_SYNC;
241 io_u->file = f;
242
243 if (td_io_prep(td, io_u)) {
244 put_io_u(td, io_u);
245 return 1;
246 }
247
755200a3 248requeue:
858a3d47 249 ret = td_io_queue(td, io_u);
36167d82 250 if (ret < 0) {
e1161c32 251 td_verror(td, io_u->error, "td_io_queue");
858a3d47 252 put_io_u(td, io_u);
858a3d47 253 return 1;
36167d82 254 } else if (ret == FIO_Q_QUEUED) {
d7762cf8 255 if (io_u_queued_complete(td, 1) < 0)
36167d82 256 return 1;
36167d82
JA
257 } else if (ret == FIO_Q_COMPLETED) {
258 if (io_u->error) {
e1161c32 259 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
260 return 1;
261 }
858a3d47 262
d7762cf8 263 if (io_u_sync_complete(td, io_u) < 0)
b2fdda43 264 return 1;
755200a3
JA
265 } else if (ret == FIO_Q_BUSY) {
266 if (td_io_commit(td))
267 return 1;
268 goto requeue;
858a3d47
JA
269 }
270
271 return 0;
272}
273
906c8d75 274/*
34403fb1 275 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
276 * reads the blocks back in, and checks the crc/md5 of the data.
277 */
1e97cce9 278static void do_verify(struct thread_data *td)
ebac4655 279{
53cdc686 280 struct fio_file *f;
36167d82 281 struct io_u *io_u;
af52b345
JA
282 int ret, min_events;
283 unsigned int i;
e5b401d4
JA
284
285 /*
286 * sync io first and invalidate cache, to make sure we really
287 * read from disk.
288 */
289 for_each_file(td, f, i) {
3d7b485f
JA
290 if (!(f->flags & FIO_FILE_OPEN))
291 continue;
b2fdda43
JA
292 if (fio_io_sync(td, f))
293 break;
294 if (file_invalidate_cache(td, f))
295 break;
e5b401d4 296 }
ebac4655 297
b2fdda43
JA
298 if (td->error)
299 return;
300
ebac4655
JA
301 td_set_runstate(td, TD_VERIFYING);
302
36167d82
JA
303 io_u = NULL;
304 while (!td->terminate) {
5451792e
JA
305 int ret2;
306
ebac4655
JA
307 io_u = __get_io_u(td);
308 if (!io_u)
309 break;
310
069c2918
JA
311 if (runtime_exceeded(td, &io_u->start_time)) {
312 put_io_u(td, io_u);
02bcaa8c 313 break;
069c2918 314 }
02bcaa8c 315
069c2918
JA
316 if (get_next_verify(td, io_u)) {
317 put_io_u(td, io_u);
ebac4655 318 break;
069c2918 319 }
ebac4655 320
069c2918
JA
321 if (td_io_prep(td, io_u)) {
322 put_io_u(td, io_u);
53cdc686 323 break;
069c2918 324 }
d7762cf8
JA
325
326 io_u->end_io = verify_io_u;
53cdc686 327
11786802 328 ret = td_io_queue(td, io_u);
36167d82
JA
329 switch (ret) {
330 case FIO_Q_COMPLETED:
331 if (io_u->error)
22819ec2 332 ret = -io_u->error;
9e9d164e 333 else if (io_u->resid) {
36167d82 334 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 335
9e9d164e
JA
336 /*
337 * zero read, fail
338 */
339 if (!bytes) {
340 td_verror(td, ENODATA, "full resid");
341 put_io_u(td, io_u);
342 break;
343 }
8400d9b2 344
36167d82
JA
345 io_u->xfer_buflen = io_u->resid;
346 io_u->xfer_buf += bytes;
8400d9b2
JA
347 io_u->offset += bytes;
348
349 if (io_u->offset == io_u->file->real_file_size)
350 goto sync_done;
351
11786802
JA
352 requeue_io_u(td, &io_u);
353 } else {
8400d9b2 354sync_done:
11786802
JA
355 ret = io_u_sync_complete(td, io_u);
356 if (ret < 0)
357 break;
36167d82 358 }
36167d82
JA
359 continue;
360 case FIO_Q_QUEUED:
361 break;
755200a3
JA
362 case FIO_Q_BUSY:
363 requeue_io_u(td, &io_u);
5451792e
JA
364 ret2 = td_io_commit(td);
365 if (ret2 < 0)
366 ret = ret2;
755200a3 367 break;
36167d82
JA
368 default:
369 assert(ret < 0);
e1161c32 370 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
371 break;
372 }
373
99784632 374 if (ret < 0 || td->error)
3af6ef39
JA
375 break;
376
ebac4655 377 /*
3af6ef39
JA
378 * if we can queue more, do so. but check if there are
379 * completed io_u's first.
ebac4655 380 */
97601024 381 min_events = 0;
e916b390 382 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 383 min_events = 1;
3af6ef39 384
2dc1bbeb
JA
385 if (td->cur_depth > td->o.iodepth_low)
386 min_events = td->cur_depth - td->o.iodepth_low;
e916b390
JA
387 }
388
3af6ef39
JA
389 /*
390 * Reap required number of io units, if any, and do the
391 * verification on them through the callback handler
392 */
d7762cf8 393 if (io_u_queued_complete(td, min_events) < 0)
ebac4655 394 break;
36167d82 395 }
ebac4655 396
c01c0395
JA
397 if (!td->error) {
398 min_events = td->cur_depth;
399
400 if (min_events)
401 ret = io_u_queued_complete(td, min_events);
402 } else
ebac4655
JA
403 cleanup_pending_aio(td);
404
405 td_set_runstate(td, TD_RUNNING);
406}
407
32cd46a0 408/*
906c8d75 409 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
410 * and reaps them, checking for rate and errors along the way.
411 */
ebac4655
JA
412static void do_io(struct thread_data *td)
413{
02bcaa8c 414 struct timeval s;
ebac4655 415 unsigned long usec;
af52b345
JA
416 unsigned int i;
417 int ret = 0;
ebac4655 418
5853e5a8
JA
419 td_set_runstate(td, TD_RUNNING);
420
7bb48f84 421 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
97601024
JA
422 struct timeval comp_time;
423 long bytes_done = 0;
84585003 424 int min_evts = 0;
ebac4655 425 struct io_u *io_u;
5451792e 426 int ret2;
ebac4655
JA
427
428 if (td->terminate)
429 break;
430
3d7c391d 431 io_u = get_io_u(td);
ebac4655
JA
432 if (!io_u)
433 break;
434
435 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
436
437 if (runtime_exceeded(td, &s)) {
438 put_io_u(td, io_u);
439 break;
440 }
36167d82 441
11786802 442 ret = td_io_queue(td, io_u);
36167d82
JA
443 switch (ret) {
444 case FIO_Q_COMPLETED:
5451792e
JA
445 if (io_u->error)
446 ret = -io_u->error;
9e9d164e 447 else if (io_u->resid) {
36167d82
JA
448 int bytes = io_u->xfer_buflen - io_u->resid;
449
9e9d164e
JA
450 /*
451 * zero read, fail
452 */
453 if (!bytes) {
454 td_verror(td, ENODATA, "full resid");
455 put_io_u(td, io_u);
456 break;
457 }
458
cec6b55d 459 io_u->xfer_buflen = io_u->resid;
36167d82 460 io_u->xfer_buf += bytes;
5a7c5680
JA
461 io_u->offset += bytes;
462
463 if (io_u->offset == io_u->file->real_file_size)
464 goto sync_done;
465
11786802
JA
466 requeue_io_u(td, &io_u);
467 } else {
5a7c5680 468sync_done:
11786802
JA
469 fio_gettime(&comp_time, NULL);
470 bytes_done = io_u_sync_complete(td, io_u);
471 if (bytes_done < 0)
472 ret = bytes_done;
cec6b55d 473 }
36167d82
JA
474 break;
475 case FIO_Q_QUEUED:
7e77dd02
JA
476 /*
477 * if the engine doesn't have a commit hook,
478 * the io_u is really queued. if it does have such
479 * a hook, it has to call io_u_queued() itself.
480 */
481 if (td->io_ops->commit == NULL)
482 io_u_queued(td, io_u);
36167d82 483 break;
755200a3
JA
484 case FIO_Q_BUSY:
485 requeue_io_u(td, &io_u);
5451792e
JA
486 ret2 = td_io_commit(td);
487 if (ret2 < 0)
488 ret = ret2;
755200a3 489 break;
36167d82
JA
490 default:
491 assert(ret < 0);
492 put_io_u(td, io_u);
493 break;
ebac4655
JA
494 }
495
99784632 496 if (ret < 0 || td->error)
36167d82
JA
497 break;
498
97601024
JA
499 /*
500 * See if we need to complete some commands
501 */
755200a3 502 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 503 min_evts = 0;
e916b390 504 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 505 min_evts = 1;
ebac4655 506
2dc1bbeb
JA
507 if (td->cur_depth > td->o.iodepth_low)
508 min_evts = td->cur_depth - td->o.iodepth_low;
e916b390
JA
509 }
510
97601024 511 fio_gettime(&comp_time, NULL);
d7762cf8 512 bytes_done = io_u_queued_complete(td, min_evts);
97601024 513 if (bytes_done < 0)
36167d82 514 break;
ebac4655
JA
515 }
516
97601024
JA
517 if (!bytes_done)
518 continue;
519
ebac4655
JA
520 /*
521 * the rate is batched for now, it should work for batches
522 * of completions except the very first one which may look
523 * a little bursty
524 */
97601024 525 usec = utime_since(&s, &comp_time);
ebac4655 526
413dd459 527 rate_throttle(td, usec, bytes_done);
ebac4655 528
97601024 529 if (check_min_rate(td, &comp_time)) {
98aa62d8 530 if (exitall_on_terminate)
390c40e2 531 terminate_threads(td->groupid);
e1161c32 532 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
533 break;
534 }
535
2dc1bbeb 536 if (td->o.thinktime) {
9c1f7434
JA
537 unsigned long long b;
538
539 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 540 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
541 int left;
542
2dc1bbeb
JA
543 if (td->o.thinktime_spin)
544 __usec_sleep(td->o.thinktime_spin);
48097d5c 545
2dc1bbeb 546 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
547 if (left)
548 usec_sleep(td, left);
549 }
9c1f7434 550 }
ebac4655
JA
551 }
552
4d2413c6 553 if (!td->error) {
3d7c391d
JA
554 struct fio_file *f;
555
c01c0395
JA
556 i = td->cur_depth;
557 if (i)
558 ret = io_u_queued_complete(td, i);
ebac4655 559
2dc1bbeb 560 if (should_fsync(td) && td->o.end_fsync) {
84585003 561 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
562
563 for_each_file(td, f, i) {
564 if (!(f->flags & FIO_FILE_OPEN))
565 continue;
858a3d47 566 fio_io_sync(td, f);
3d7b485f 567 }
84585003 568 }
c01c0395
JA
569 } else
570 cleanup_pending_aio(td);
ebac4655
JA
571}
572
ebac4655
JA
573static void cleanup_io_u(struct thread_data *td)
574{
575 struct list_head *entry, *n;
576 struct io_u *io_u;
577
578 list_for_each_safe(entry, n, &td->io_u_freelist) {
579 io_u = list_entry(entry, struct io_u, list);
580
581 list_del(&io_u->list);
582 free(io_u);
583 }
584
2f9ade3c 585 free_io_mem(td);
ebac4655
JA
586}
587
6b9cea23
JA
588/*
589 * "randomly" fill the buffer contents
590 */
66eeb296 591static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 592{
66eeb296 593 int *ptr = io_u->buf;
6b9cea23
JA
594
595 while ((void *) ptr - io_u->buf < max_bs) {
596 *ptr = rand() * 0x9e370001;
597 ptr++;
598 }
599}
600
ebac4655
JA
601static int init_io_u(struct thread_data *td)
602{
2b4ce34f 603 unsigned long long buf_size;
ebac4655 604 struct io_u *io_u;
a00735e6 605 unsigned int max_bs;
ebac4655
JA
606 int i, max_units;
607 char *p;
608
2866c82d 609 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
610 max_units = 1;
611 else
2dc1bbeb 612 max_units = td->o.iodepth;
ebac4655 613
2dc1bbeb 614 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
2b4ce34f
JA
615 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
616 buf_size += page_mask;
617 if (buf_size != (size_t) buf_size) {
618 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
619 return 1;
620 }
621
622 td->orig_buffer_size = buf_size;
74b025b0 623
2dc1bbeb
JA
624 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
625 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
4afbf66f
JA
626 else if (td->orig_buffer_size & page_mask)
627 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
ebac4655 628
2f9ade3c
JA
629 if (allocate_io_mem(td))
630 return 1;
ebac4655 631
ebac4655
JA
632 p = ALIGN(td->orig_buffer);
633 for (i = 0; i < max_units; i++) {
634 io_u = malloc(sizeof(*io_u));
635 memset(io_u, 0, sizeof(*io_u));
636 INIT_LIST_HEAD(&io_u->list);
637
a00735e6 638 io_u->buf = p + max_bs * i;
6b9cea23 639 if (td_write(td) || td_rw(td))
a00735e6 640 fill_rand_buf(io_u, max_bs);
6b9cea23 641
b1ff3403 642 io_u->index = i;
0c6e7517 643 io_u->flags = IO_U_F_FREE;
ebac4655
JA
644 list_add(&io_u->list, &td->io_u_freelist);
645 }
646
433afcb4
JA
647 io_u_init_timeout();
648
ebac4655
JA
649 return 0;
650}
651
da86774e
JA
652static int switch_ioscheduler(struct thread_data *td)
653{
654 char tmp[256], tmp2[128];
655 FILE *f;
656 int ret;
657
ba0fbe10 658 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
659 return 0;
660
da86774e
JA
661 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
662
663 f = fopen(tmp, "r+");
664 if (!f) {
e1161c32 665 td_verror(td, errno, "fopen");
da86774e
JA
666 return 1;
667 }
668
669 /*
670 * Set io scheduler.
671 */
2dc1bbeb 672 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 673 if (ferror(f) || ret != 1) {
e1161c32 674 td_verror(td, errno, "fwrite");
da86774e
JA
675 fclose(f);
676 return 1;
677 }
678
679 rewind(f);
680
681 /*
682 * Read back and check that the selected scheduler is now the default.
683 */
684 ret = fread(tmp, 1, sizeof(tmp), f);
685 if (ferror(f) || ret < 0) {
e1161c32 686 td_verror(td, errno, "fread");
da86774e
JA
687 fclose(f);
688 return 1;
689 }
690
2dc1bbeb 691 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 692 if (!strstr(tmp, tmp2)) {
2dc1bbeb 693 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 694 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
695 fclose(f);
696 return 1;
697 }
698
699 fclose(f);
700 return 0;
701}
702
a978ba68 703static int clear_io_state(struct thread_data *td)
ebac4655 704{
53cdc686 705 struct fio_file *f;
af52b345
JA
706 unsigned int i;
707 int ret;
ebac4655 708
756867bd 709 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 710 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 711 td->zone_bytes = 0;
d7d3b49b 712 td->rate_bytes = 0;
4e991c23 713 td->rate_blocks = 0;
38d77cae 714 td->rw_end_set[0] = td->rw_end_set[1] = 0;
ebac4655 715
c1324df1
JA
716 td->last_was_sync = 0;
717
a978ba68
JA
718 for_each_file(td, f, i)
719 td_io_close_file(td, f);
53cdc686 720
a978ba68
JA
721 ret = 0;
722 for_each_file(td, f, i) {
723 ret = td_io_open_file(td, f);
724 if (ret)
725 break;
53cdc686 726 }
a978ba68
JA
727
728 return ret;
ebac4655
JA
729}
730
906c8d75
JA
731/*
732 * Entry point for the thread based jobs. The process based jobs end up
733 * here as well, after a little setup.
734 */
ebac4655
JA
735static void *thread_main(void *data)
736{
69008999 737 unsigned long long runtime[2];
ebac4655 738 struct thread_data *td = data;
38d77cae 739 unsigned long elapsed;
a978ba68 740 int clear_state;
ebac4655 741
2dc1bbeb 742 if (!td->o.use_thread)
ebac4655
JA
743 setsid();
744
745 td->pid = getpid();
746
aea47d44
JA
747 INIT_LIST_HEAD(&td->io_u_freelist);
748 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 749 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44 750 INIT_LIST_HEAD(&td->io_log_list);
8de8f047 751 INIT_LIST_HEAD(&td->io_hist_list);
bb5d7d0b 752 td->io_hist_tree = RB_ROOT;
aea47d44 753
ebac4655 754 if (init_io_u(td))
bda4fd9e 755 goto err_sem;
ebac4655
JA
756
757 if (fio_setaffinity(td) == -1) {
e1161c32 758 td_verror(td, errno, "cpu_set_affinity");
bda4fd9e 759 goto err_sem;
ebac4655
JA
760 }
761
aea47d44 762 if (init_iolog(td))
bda4fd9e 763 goto err_sem;
aea47d44 764
ebac4655
JA
765 if (td->ioprio) {
766 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 767 td_verror(td, errno, "ioprio_set");
bda4fd9e 768 goto err_sem;
ebac4655
JA
769 }
770 }
771
2dc1bbeb 772 if (nice(td->o.nice) == -1) {
e1161c32 773 td_verror(td, errno, "nice");
bda4fd9e 774 goto err_sem;
b6f4d880
JA
775 }
776
2dc1bbeb 777 if (td->o.ioscheduler && switch_ioscheduler(td))
bda4fd9e 778 goto err_sem;
da86774e 779
75154845 780 td_set_runstate(td, TD_INITIALIZED);
07739b57
JA
781 fio_sem_up(startup_sem);
782 fio_sem_down(td->mutex);
ebac4655 783
37f56873
JA
784 /*
785 * the ->mutex semaphore is now no longer used, close it to avoid
786 * eating a file descriptor
787 */
788 fio_sem_remove(td->mutex);
789
2dc1bbeb 790 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
791 goto err;
792
7d6c5283
JA
793 if (td_io_init(td))
794 goto err;
795
b5af8293
JA
796 if (open_files(td))
797 goto err;
798
68727076
JA
799 if (init_random_map(td))
800 goto err;
801
2dc1bbeb
JA
802 if (td->o.exec_prerun) {
803 if (system(td->o.exec_prerun) < 0)
69cfd7e0
JA
804 goto err;
805 }
4e0ba8af 806
69008999 807 fio_gettime(&td->epoch, NULL);
433afcb4 808 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
756867bd 809 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
810
811 runtime[0] = runtime[1] = 0;
a978ba68 812 clear_state = 0;
2dc1bbeb 813 while (td->o.loops--) {
02bcaa8c 814 fio_gettime(&td->start, NULL);
756867bd 815 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655 816
2dc1bbeb 817 if (td->o.ratemin)
756867bd 818 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655 819
a978ba68
JA
820 if (clear_state && clear_io_state(td))
821 break;
822
ebac4655
JA
823 prune_io_piece_log(td);
824
ba0fbe10 825 do_io(td);
ebac4655 826
a978ba68
JA
827 clear_state = 1;
828
38d77cae
JA
829 if (td_read(td) && td->io_bytes[DDIR_READ]) {
830 if (td->rw_end_set[DDIR_READ])
831 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
832 else
833 elapsed = utime_since_now(&td->start);
834
835 runtime[DDIR_READ] += elapsed;
836 }
837 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
838 if (td->rw_end_set[DDIR_WRITE])
839 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
840 else
841 elapsed = utime_since_now(&td->start);
842
843 runtime[DDIR_WRITE] += elapsed;
844 }
413dd459 845
ebac4655
JA
846 if (td->error || td->terminate)
847 break;
848
2dc1bbeb 849 if (td->o.verify == VERIFY_NONE)
ebac4655
JA
850 continue;
851
a978ba68
JA
852 if (clear_io_state(td))
853 break;
854
02bcaa8c 855 fio_gettime(&td->start, NULL);
ebac4655
JA
856
857 do_verify(td);
858
69008999 859 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
860
861 if (td->error || td->terminate)
862 break;
863 }
864
36dff966 865 update_rusage_stat(td);
756867bd
JA
866 td->ts.runtime[0] = runtime[0] / 1000;
867 td->ts.runtime[1] = runtime[1] / 1000;
868 td->ts.total_run_time = mtime_since_now(&td->epoch);
869 td->ts.io_bytes[0] = td->io_bytes[0];
870 td->ts.io_bytes[1] = td->io_bytes[1];
871
872 if (td->ts.bw_log)
873 finish_log(td, td->ts.bw_log, "bw");
874 if (td->ts.slat_log)
875 finish_log(td, td->ts.slat_log, "slat");
876 if (td->ts.clat_log)
877 finish_log(td, td->ts.clat_log, "clat");
2dc1bbeb 878 if (td->o.write_iolog_file)
843a7413 879 write_iolog_close(td);
2dc1bbeb
JA
880 if (td->o.exec_postrun) {
881 if (system(td->o.exec_postrun) < 0)
882 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
69cfd7e0 883 }
ebac4655
JA
884
885 if (exitall_on_terminate)
390c40e2 886 terminate_threads(td->groupid);
ebac4655
JA
887
888err:
5bf13a5a
JA
889 if (td->error)
890 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 891 close_files(td);
2866c82d 892 close_ioengine(td);
ebac4655 893 cleanup_io_u(td);
d23bb327 894 options_mem_free(td);
ebac4655 895 td_set_runstate(td, TD_EXITED);
43d76807 896 return (void *) (unsigned long) td->error;
bda4fd9e
JA
897err_sem:
898 fio_sem_up(startup_sem);
899 goto err;
ebac4655
JA
900}
901
906c8d75
JA
902/*
903 * We cannot pass the td data into a forked process, so attach the td and
904 * pass it to the thread worker.
905 */
a6418147 906static int fork_main(int shmid, int offset)
ebac4655
JA
907{
908 struct thread_data *td;
a6418147 909 void *data, *ret;
ebac4655
JA
910
911 data = shmat(shmid, NULL, 0);
912 if (data == (void *) -1) {
a6418147
JA
913 int __err = errno;
914
ebac4655 915 perror("shmat");
a6418147 916 return __err;
ebac4655
JA
917 }
918
919 td = data + offset * sizeof(struct thread_data);
a6418147 920 ret = thread_main(td);
ebac4655 921 shmdt(data);
43d76807 922 return (int) (unsigned long) ret;
ebac4655
JA
923}
924
906c8d75
JA
925/*
926 * Run over the job map and reap the threads that have exited, if any.
927 */
ebac4655
JA
928static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
929{
34572e28 930 struct thread_data *td;
fab6aa71 931 int i, cputhreads, pending, status, ret;
ebac4655
JA
932
933 /*
934 * reap exited threads (TD_EXITED -> TD_REAPED)
935 */
4d2413c6 936 pending = cputhreads = 0;
34572e28 937 for_each_td(td, i) {
3707f45b 938 int flags = 0;
a2f77c9f 939
84585003
JA
940 /*
941 * ->io_ops is NULL for a thread that has closed its
942 * io engine
943 */
ba0fbe10 944 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0
JA
945 cputhreads++;
946
a2f77c9f
JA
947 if (!td->pid || td->runstate == TD_REAPED)
948 continue;
2dc1bbeb 949 if (td->o.use_thread) {
3707f45b
JA
950 if (td->runstate == TD_EXITED) {
951 td_set_runstate(td, TD_REAPED);
952 goto reaped;
953 }
954 continue;
955 }
a2f77c9f
JA
956
957 flags = WNOHANG;
958 if (td->runstate == TD_EXITED)
959 flags = 0;
960
961 /*
962 * check if someone quit or got killed in an unusual way
963 */
964 ret = waitpid(td->pid, &status, flags);
3707f45b 965 if (ret < 0) {
a2f77c9f
JA
966 if (errno == ECHILD) {
967 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
968 td_set_runstate(td, TD_REAPED);
969 goto reaped;
970 }
971 perror("waitpid");
972 } else if (ret == td->pid) {
973 if (WIFSIGNALED(status)) {
fab6aa71
JA
974 int sig = WTERMSIG(status);
975
d9244941
JA
976 if (sig != SIGQUIT)
977 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
fab6aa71
JA
978 td_set_runstate(td, TD_REAPED);
979 goto reaped;
980 }
a2f77c9f
JA
981 if (WIFEXITED(status)) {
982 if (WEXITSTATUS(status) && !td->error)
983 td->error = WEXITSTATUS(status);
a2f77c9f 984
a2f77c9f
JA
985 td_set_runstate(td, TD_REAPED);
986 goto reaped;
a6418147
JA
987 }
988 }
ebac4655 989
a2f77c9f
JA
990 /*
991 * thread is not dead, continue
992 */
993 continue;
fab6aa71 994reaped:
2dc1bbeb 995 if (td->o.use_thread) {
3707f45b
JA
996 long ret;
997
998 if (pthread_join(td->thread, (void *) &ret))
999 perror("pthread_join");
1000 }
1001
ebac4655 1002 (*nr_running)--;
2dc1bbeb
JA
1003 (*m_rate) -= td->o.ratemin;
1004 (*t_rate) -= td->o.rate;
a2f77c9f
JA
1005
1006 if (td->error)
1007 exit_value++;
ebac4655 1008 }
b990b5c0 1009
4d2413c6 1010 if (*nr_running == cputhreads && !pending)
390c40e2 1011 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1012}
1013
906c8d75
JA
1014/*
1015 * Main function for kicking off and reaping jobs, as needed.
1016 */
ebac4655
JA
1017static void run_threads(void)
1018{
ebac4655
JA
1019 struct thread_data *td;
1020 unsigned long spent;
1021 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1022
2f9ade3c
JA
1023 if (fio_pin_memory())
1024 return;
ebac4655 1025
c6ae0a5b 1026 if (!terse_output) {
9cedf167
JA
1027 printf("Starting ");
1028 if (nr_thread)
1029 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1030 if (nr_process) {
1031 if (nr_thread)
1032 printf(" and ");
1033 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1034 }
1035 printf("\n");
c6ae0a5b
JA
1036 fflush(stdout);
1037 }
c04f7ec3 1038
4efa970e
JA
1039 signal(SIGINT, sig_handler);
1040 signal(SIGALRM, sig_handler);
1041
ebac4655
JA
1042 todo = thread_number;
1043 nr_running = 0;
1044 nr_started = 0;
1045 m_rate = t_rate = 0;
1046
34572e28 1047 for_each_td(td, i) {
263e529f 1048 print_status_init(td->thread_number - 1);
ebac4655 1049
2dc1bbeb 1050 if (!td->o.create_serialize) {
380cf265 1051 init_disk_util(td);
ebac4655 1052 continue;
380cf265 1053 }
ebac4655
JA
1054
1055 /*
1056 * do file setup here so it happens sequentially,
1057 * we don't want X number of threads getting their
1058 * client data interspersed on disk
1059 */
53cdc686 1060 if (setup_files(td)) {
5bf13a5a
JA
1061 exit_value++;
1062 if (td->error)
1063 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
1064 td_set_runstate(td, TD_REAPED);
1065 todo--;
1066 }
380cf265
JA
1067
1068 init_disk_util(td);
ebac4655
JA
1069 }
1070
a2f77c9f
JA
1071 set_genesis_time();
1072
ebac4655 1073 while (todo) {
75154845
JA
1074 struct thread_data *map[MAX_JOBS];
1075 struct timeval this_start;
1076 int this_jobs = 0, left;
1077
ebac4655
JA
1078 /*
1079 * create threads (TD_NOT_CREATED -> TD_CREATED)
1080 */
34572e28 1081 for_each_td(td, i) {
ebac4655
JA
1082 if (td->runstate != TD_NOT_CREATED)
1083 continue;
1084
1085 /*
1086 * never got a chance to start, killed by other
1087 * thread for some reason
1088 */
1089 if (td->terminate) {
1090 todo--;
1091 continue;
1092 }
1093
2dc1bbeb 1094 if (td->o.start_delay) {
263e529f 1095 spent = mtime_since_genesis();
ebac4655 1096
2dc1bbeb 1097 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1098 continue;
1099 }
1100
2dc1bbeb 1101 if (td->o.stonewall && (nr_started || nr_running))
ebac4655
JA
1102 break;
1103
75154845
JA
1104 /*
1105 * Set state to created. Thread will transition
1106 * to TD_INITIALIZED when it's done setting up.
1107 */
ebac4655 1108 td_set_runstate(td, TD_CREATED);
75154845 1109 map[this_jobs++] = td;
ebac4655
JA
1110 nr_started++;
1111
2dc1bbeb 1112 if (td->o.use_thread) {
ebac4655
JA
1113 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1114 perror("thread_create");
1115 nr_started--;
c8bb6faf 1116 break;
ebac4655
JA
1117 }
1118 } else {
07739b57 1119 if (!fork()) {
a6418147
JA
1120 int ret = fork_main(shm_id, i);
1121
1122 exit(ret);
ebac4655
JA
1123 }
1124 }
07739b57 1125 fio_sem_down(startup_sem);
ebac4655
JA
1126 }
1127
1128 /*
75154845
JA
1129 * Wait for the started threads to transition to
1130 * TD_INITIALIZED.
ebac4655 1131 */
02bcaa8c 1132 fio_gettime(&this_start, NULL);
75154845 1133 left = this_jobs;
6ce15a32 1134 while (left && !fio_abort) {
75154845
JA
1135 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1136 break;
1137
1138 usleep(100000);
1139
1140 for (i = 0; i < this_jobs; i++) {
1141 td = map[i];
1142 if (!td)
1143 continue;
b6f4d880 1144 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1145 map[i] = NULL;
1146 left--;
b6f4d880
JA
1147 } else if (td->runstate >= TD_EXITED) {
1148 map[i] = NULL;
1149 left--;
1150 todo--;
1151 nr_running++; /* work-around... */
75154845
JA
1152 }
1153 }
1154 }
1155
1156 if (left) {
3b70d7e5 1157 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1158 for (i = 0; i < this_jobs; i++) {
1159 td = map[i];
1160 if (!td)
1161 continue;
1162 kill(td->pid, SIGTERM);
1163 }
1164 break;
1165 }
1166
1167 /*
b6f4d880 1168 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1169 */
34572e28 1170 for_each_td(td, i) {
75154845 1171 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1172 continue;
1173
1174 td_set_runstate(td, TD_RUNNING);
1175 nr_running++;
1176 nr_started--;
2dc1bbeb
JA
1177 m_rate += td->o.ratemin;
1178 t_rate += td->o.rate;
75154845 1179 todo--;
07739b57 1180 fio_sem_up(td->mutex);
ebac4655
JA
1181 }
1182
1183 reap_threads(&nr_running, &t_rate, &m_rate);
1184
1185 if (todo)
1186 usleep(100000);
1187 }
1188
1189 while (nr_running) {
1190 reap_threads(&nr_running, &t_rate, &m_rate);
1191 usleep(10000);
1192 }
1193
1194 update_io_ticks();
2f9ade3c 1195 fio_unpin_memory();
ebac4655
JA
1196}
1197
ebac4655
JA
1198int main(int argc, char *argv[])
1199{
29d610e1
JA
1200 long ps;
1201
dbe1125e
JA
1202 /*
1203 * We need locale for number printing, if it isn't set then just
1204 * go with the US format.
1205 */
1206 if (!getenv("LC_NUMERIC"))
1207 setlocale(LC_NUMERIC, "en_US");
1208
ebac4655
JA
1209 if (parse_options(argc, argv))
1210 return 1;
1211
1212 if (!thread_number) {
3b70d7e5 1213 log_err("Nothing to do\n");
ebac4655
JA
1214 return 1;
1215 }
1216
29d610e1
JA
1217 ps = sysconf(_SC_PAGESIZE);
1218 if (ps < 0) {
1219 log_err("Failed to get page size\n");
1220 return 1;
1221 }
1222
cfc99db7 1223 page_size = ps;
29d610e1
JA
1224 page_mask = ps - 1;
1225
bb3884d8
JA
1226 if (write_bw_log) {
1227 setup_log(&agg_io_log[DDIR_READ]);
1228 setup_log(&agg_io_log[DDIR_WRITE]);
1229 }
1230
07739b57
JA
1231 startup_sem = fio_sem_init(0);
1232
a2f77c9f
JA
1233 set_genesis_time();
1234
ebac4655
JA
1235 disk_util_timer_arm();
1236
1237 run_threads();
6ce15a32 1238
bb3884d8 1239 if (!fio_abort) {
6ce15a32 1240 show_run_stats();
bb3884d8
JA
1241 if (write_bw_log) {
1242 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1243 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1244 }
1245 }
ebac4655 1246
07739b57 1247 fio_sem_remove(startup_sem);
437c9b71 1248 return exit_value;
ebac4655 1249}