Fix io piece logging to not have O(n) runtime
[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"
38#include "os.h"
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) {
c1302d44
JA
73 /*
74 * if the thread is running, just let it exit
75 */
76 if (td->runstate < TD_RUNNING)
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);
02bcaa8c 314 break;
069c2918 315 }
02bcaa8c 316
069c2918
JA
317 if (get_next_verify(td, io_u)) {
318 put_io_u(td, io_u);
ebac4655 319 break;
069c2918 320 }
ebac4655 321
069c2918
JA
322 if (td_io_prep(td, io_u)) {
323 put_io_u(td, io_u);
53cdc686 324 break;
069c2918 325 }
d7762cf8
JA
326
327 io_u->end_io = verify_io_u;
53cdc686 328
11786802 329 ret = td_io_queue(td, io_u);
36167d82
JA
330 switch (ret) {
331 case FIO_Q_COMPLETED:
332 if (io_u->error)
22819ec2 333 ret = -io_u->error;
9e9d164e 334 else if (io_u->resid) {
36167d82 335 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 336
9e9d164e
JA
337 /*
338 * zero read, fail
339 */
340 if (!bytes) {
341 td_verror(td, ENODATA, "full resid");
342 put_io_u(td, io_u);
343 break;
344 }
36167d82
JA
345 io_u->xfer_buflen = io_u->resid;
346 io_u->xfer_buf += bytes;
11786802
JA
347 requeue_io_u(td, &io_u);
348 } else {
349 ret = io_u_sync_complete(td, io_u);
350 if (ret < 0)
351 break;
36167d82 352 }
36167d82
JA
353 continue;
354 case FIO_Q_QUEUED:
355 break;
755200a3
JA
356 case FIO_Q_BUSY:
357 requeue_io_u(td, &io_u);
5451792e
JA
358 ret2 = td_io_commit(td);
359 if (ret2 < 0)
360 ret = ret2;
755200a3 361 break;
36167d82
JA
362 default:
363 assert(ret < 0);
e1161c32 364 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
365 break;
366 }
367
99784632 368 if (ret < 0 || td->error)
3af6ef39
JA
369 break;
370
ebac4655 371 /*
3af6ef39
JA
372 * if we can queue more, do so. but check if there are
373 * completed io_u's first.
ebac4655 374 */
97601024 375 min_events = 0;
e916b390 376 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 377 min_events = 1;
3af6ef39 378
2dc1bbeb
JA
379 if (td->cur_depth > td->o.iodepth_low)
380 min_events = td->cur_depth - td->o.iodepth_low;
e916b390
JA
381 }
382
3af6ef39
JA
383 /*
384 * Reap required number of io units, if any, and do the
385 * verification on them through the callback handler
386 */
d7762cf8 387 if (io_u_queued_complete(td, min_events) < 0)
ebac4655 388 break;
36167d82 389 }
ebac4655 390
c01c0395
JA
391 if (!td->error) {
392 min_events = td->cur_depth;
393
394 if (min_events)
395 ret = io_u_queued_complete(td, min_events);
396 } else
ebac4655
JA
397 cleanup_pending_aio(td);
398
399 td_set_runstate(td, TD_RUNNING);
400}
401
32cd46a0 402/*
906c8d75 403 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
404 * and reaps them, checking for rate and errors along the way.
405 */
ebac4655
JA
406static void do_io(struct thread_data *td)
407{
02bcaa8c 408 struct timeval s;
ebac4655 409 unsigned long usec;
af52b345
JA
410 unsigned int i;
411 int ret = 0;
ebac4655 412
5853e5a8
JA
413 td_set_runstate(td, TD_RUNNING);
414
3951414d 415 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
97601024
JA
416 struct timeval comp_time;
417 long bytes_done = 0;
84585003 418 int min_evts = 0;
ebac4655 419 struct io_u *io_u;
5451792e 420 int ret2;
ebac4655
JA
421
422 if (td->terminate)
423 break;
424
3d7c391d 425 io_u = get_io_u(td);
ebac4655
JA
426 if (!io_u)
427 break;
428
429 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
430
431 if (runtime_exceeded(td, &s)) {
432 put_io_u(td, io_u);
433 break;
434 }
36167d82 435
11786802 436 ret = td_io_queue(td, io_u);
36167d82
JA
437 switch (ret) {
438 case FIO_Q_COMPLETED:
5451792e
JA
439 if (io_u->error)
440 ret = -io_u->error;
9e9d164e 441 else if (io_u->resid) {
36167d82
JA
442 int bytes = io_u->xfer_buflen - io_u->resid;
443
9e9d164e
JA
444 /*
445 * zero read, fail
446 */
447 if (!bytes) {
448 td_verror(td, ENODATA, "full resid");
449 put_io_u(td, io_u);
450 break;
451 }
452
cec6b55d 453 io_u->xfer_buflen = io_u->resid;
36167d82 454 io_u->xfer_buf += bytes;
11786802
JA
455 requeue_io_u(td, &io_u);
456 } else {
457 fio_gettime(&comp_time, NULL);
458 bytes_done = io_u_sync_complete(td, io_u);
459 if (bytes_done < 0)
460 ret = bytes_done;
cec6b55d 461 }
36167d82
JA
462 break;
463 case FIO_Q_QUEUED:
7e77dd02
JA
464 /*
465 * if the engine doesn't have a commit hook,
466 * the io_u is really queued. if it does have such
467 * a hook, it has to call io_u_queued() itself.
468 */
469 if (td->io_ops->commit == NULL)
470 io_u_queued(td, io_u);
36167d82 471 break;
755200a3
JA
472 case FIO_Q_BUSY:
473 requeue_io_u(td, &io_u);
5451792e
JA
474 ret2 = td_io_commit(td);
475 if (ret2 < 0)
476 ret = ret2;
755200a3 477 break;
36167d82
JA
478 default:
479 assert(ret < 0);
480 put_io_u(td, io_u);
481 break;
ebac4655
JA
482 }
483
99784632 484 if (ret < 0 || td->error)
36167d82
JA
485 break;
486
97601024
JA
487 /*
488 * See if we need to complete some commands
489 */
755200a3 490 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 491 min_evts = 0;
e916b390 492 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 493 min_evts = 1;
ebac4655 494
2dc1bbeb
JA
495 if (td->cur_depth > td->o.iodepth_low)
496 min_evts = td->cur_depth - td->o.iodepth_low;
e916b390
JA
497 }
498
97601024 499 fio_gettime(&comp_time, NULL);
d7762cf8 500 bytes_done = io_u_queued_complete(td, min_evts);
97601024 501 if (bytes_done < 0)
36167d82 502 break;
ebac4655
JA
503 }
504
97601024
JA
505 if (!bytes_done)
506 continue;
507
ebac4655
JA
508 /*
509 * the rate is batched for now, it should work for batches
510 * of completions except the very first one which may look
511 * a little bursty
512 */
97601024 513 usec = utime_since(&s, &comp_time);
ebac4655 514
413dd459 515 rate_throttle(td, usec, bytes_done);
ebac4655 516
97601024 517 if (check_min_rate(td, &comp_time)) {
98aa62d8 518 if (exitall_on_terminate)
390c40e2 519 terminate_threads(td->groupid);
e1161c32 520 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
521 break;
522 }
523
2dc1bbeb 524 if (td->o.thinktime) {
9c1f7434
JA
525 unsigned long long b;
526
527 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 528 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
529 int left;
530
2dc1bbeb
JA
531 if (td->o.thinktime_spin)
532 __usec_sleep(td->o.thinktime_spin);
48097d5c 533
2dc1bbeb 534 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
535 if (left)
536 usec_sleep(td, left);
537 }
9c1f7434 538 }
ebac4655
JA
539 }
540
4d2413c6 541 if (!td->error) {
3d7c391d
JA
542 struct fio_file *f;
543
c01c0395
JA
544 i = td->cur_depth;
545 if (i)
546 ret = io_u_queued_complete(td, i);
ebac4655 547
2dc1bbeb 548 if (should_fsync(td) && td->o.end_fsync) {
84585003 549 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
550
551 for_each_file(td, f, i) {
552 if (!(f->flags & FIO_FILE_OPEN))
553 continue;
858a3d47 554 fio_io_sync(td, f);
3d7b485f 555 }
84585003 556 }
c01c0395
JA
557 } else
558 cleanup_pending_aio(td);
ebac4655
JA
559}
560
ebac4655
JA
561static void cleanup_io_u(struct thread_data *td)
562{
563 struct list_head *entry, *n;
564 struct io_u *io_u;
565
566 list_for_each_safe(entry, n, &td->io_u_freelist) {
567 io_u = list_entry(entry, struct io_u, list);
568
569 list_del(&io_u->list);
570 free(io_u);
571 }
572
2f9ade3c 573 free_io_mem(td);
ebac4655
JA
574}
575
6b9cea23
JA
576/*
577 * "randomly" fill the buffer contents
578 */
66eeb296 579static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 580{
66eeb296 581 int *ptr = io_u->buf;
6b9cea23
JA
582
583 while ((void *) ptr - io_u->buf < max_bs) {
584 *ptr = rand() * 0x9e370001;
585 ptr++;
586 }
587}
588
ebac4655
JA
589static int init_io_u(struct thread_data *td)
590{
2b4ce34f 591 unsigned long long buf_size;
ebac4655 592 struct io_u *io_u;
a00735e6 593 unsigned int max_bs;
ebac4655
JA
594 int i, max_units;
595 char *p;
596
2866c82d 597 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
598 max_units = 1;
599 else
2dc1bbeb 600 max_units = td->o.iodepth;
ebac4655 601
2dc1bbeb 602 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
2b4ce34f
JA
603 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
604 buf_size += page_mask;
605 if (buf_size != (size_t) buf_size) {
606 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
607 return 1;
608 }
609
610 td->orig_buffer_size = buf_size;
74b025b0 611
2dc1bbeb
JA
612 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
613 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
4afbf66f
JA
614 else if (td->orig_buffer_size & page_mask)
615 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
ebac4655 616
2f9ade3c
JA
617 if (allocate_io_mem(td))
618 return 1;
ebac4655 619
ebac4655
JA
620 p = ALIGN(td->orig_buffer);
621 for (i = 0; i < max_units; i++) {
622 io_u = malloc(sizeof(*io_u));
623 memset(io_u, 0, sizeof(*io_u));
624 INIT_LIST_HEAD(&io_u->list);
625
a00735e6 626 io_u->buf = p + max_bs * i;
6b9cea23 627 if (td_write(td) || td_rw(td))
a00735e6 628 fill_rand_buf(io_u, max_bs);
6b9cea23 629
b1ff3403 630 io_u->index = i;
0c6e7517 631 io_u->flags = IO_U_F_FREE;
ebac4655
JA
632 list_add(&io_u->list, &td->io_u_freelist);
633 }
634
433afcb4
JA
635 io_u_init_timeout();
636
ebac4655
JA
637 return 0;
638}
639
da86774e
JA
640static int switch_ioscheduler(struct thread_data *td)
641{
642 char tmp[256], tmp2[128];
643 FILE *f;
644 int ret;
645
ba0fbe10 646 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
647 return 0;
648
da86774e
JA
649 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
650
651 f = fopen(tmp, "r+");
652 if (!f) {
e1161c32 653 td_verror(td, errno, "fopen");
da86774e
JA
654 return 1;
655 }
656
657 /*
658 * Set io scheduler.
659 */
2dc1bbeb 660 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 661 if (ferror(f) || ret != 1) {
e1161c32 662 td_verror(td, errno, "fwrite");
da86774e
JA
663 fclose(f);
664 return 1;
665 }
666
667 rewind(f);
668
669 /*
670 * Read back and check that the selected scheduler is now the default.
671 */
672 ret = fread(tmp, 1, sizeof(tmp), f);
673 if (ferror(f) || ret < 0) {
e1161c32 674 td_verror(td, errno, "fread");
da86774e
JA
675 fclose(f);
676 return 1;
677 }
678
2dc1bbeb 679 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 680 if (!strstr(tmp, tmp2)) {
2dc1bbeb 681 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 682 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
683 fclose(f);
684 return 1;
685 }
686
687 fclose(f);
688 return 0;
689}
690
a978ba68 691static int clear_io_state(struct thread_data *td)
ebac4655 692{
53cdc686 693 struct fio_file *f;
af52b345
JA
694 unsigned int i;
695 int ret;
ebac4655 696
756867bd 697 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 698 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 699 td->zone_bytes = 0;
d7d3b49b 700 td->rate_bytes = 0;
4e991c23 701 td->rate_blocks = 0;
38d77cae 702 td->rw_end_set[0] = td->rw_end_set[1] = 0;
ebac4655 703
c1324df1
JA
704 td->last_was_sync = 0;
705
a978ba68
JA
706 for_each_file(td, f, i)
707 td_io_close_file(td, f);
53cdc686 708
a978ba68
JA
709 ret = 0;
710 for_each_file(td, f, i) {
711 ret = td_io_open_file(td, f);
712 if (ret)
713 break;
53cdc686 714 }
a978ba68
JA
715
716 return ret;
ebac4655
JA
717}
718
906c8d75
JA
719/*
720 * Entry point for the thread based jobs. The process based jobs end up
721 * here as well, after a little setup.
722 */
ebac4655
JA
723static void *thread_main(void *data)
724{
69008999 725 unsigned long long runtime[2];
ebac4655 726 struct thread_data *td = data;
38d77cae 727 unsigned long elapsed;
a978ba68 728 int clear_state;
ebac4655 729
2dc1bbeb 730 if (!td->o.use_thread)
ebac4655
JA
731 setsid();
732
733 td->pid = getpid();
734
aea47d44
JA
735 INIT_LIST_HEAD(&td->io_u_freelist);
736 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 737 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44
JA
738 INIT_LIST_HEAD(&td->io_log_list);
739
ebac4655 740 if (init_io_u(td))
bda4fd9e 741 goto err_sem;
ebac4655
JA
742
743 if (fio_setaffinity(td) == -1) {
e1161c32 744 td_verror(td, errno, "cpu_set_affinity");
bda4fd9e 745 goto err_sem;
ebac4655
JA
746 }
747
aea47d44 748 if (init_iolog(td))
bda4fd9e 749 goto err_sem;
aea47d44 750
ebac4655
JA
751 if (td->ioprio) {
752 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 753 td_verror(td, errno, "ioprio_set");
bda4fd9e 754 goto err_sem;
ebac4655
JA
755 }
756 }
757
2dc1bbeb 758 if (nice(td->o.nice) == -1) {
e1161c32 759 td_verror(td, errno, "nice");
bda4fd9e 760 goto err_sem;
b6f4d880
JA
761 }
762
2dc1bbeb 763 if (td->o.ioscheduler && switch_ioscheduler(td))
bda4fd9e 764 goto err_sem;
da86774e 765
75154845 766 td_set_runstate(td, TD_INITIALIZED);
07739b57
JA
767 fio_sem_up(startup_sem);
768 fio_sem_down(td->mutex);
ebac4655 769
37f56873
JA
770 /*
771 * the ->mutex semaphore is now no longer used, close it to avoid
772 * eating a file descriptor
773 */
774 fio_sem_remove(td->mutex);
775
2dc1bbeb 776 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
777 goto err;
778
7d6c5283
JA
779 if (td_io_init(td))
780 goto err;
781
b5af8293
JA
782 if (open_files(td))
783 goto err;
784
68727076
JA
785 if (init_random_map(td))
786 goto err;
787
2dc1bbeb
JA
788 if (td->o.exec_prerun) {
789 if (system(td->o.exec_prerun) < 0)
69cfd7e0
JA
790 goto err;
791 }
4e0ba8af 792
69008999 793 fio_gettime(&td->epoch, NULL);
433afcb4 794 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
756867bd 795 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
796
797 runtime[0] = runtime[1] = 0;
a978ba68 798 clear_state = 0;
2dc1bbeb 799 while (td->o.loops--) {
02bcaa8c 800 fio_gettime(&td->start, NULL);
756867bd 801 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655 802
2dc1bbeb 803 if (td->o.ratemin)
756867bd 804 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655 805
a978ba68
JA
806 if (clear_state && clear_io_state(td))
807 break;
808
ebac4655
JA
809 prune_io_piece_log(td);
810
ba0fbe10 811 do_io(td);
ebac4655 812
a978ba68
JA
813 clear_state = 1;
814
38d77cae
JA
815 if (td_read(td) && td->io_bytes[DDIR_READ]) {
816 if (td->rw_end_set[DDIR_READ])
817 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
818 else
819 elapsed = utime_since_now(&td->start);
820
821 runtime[DDIR_READ] += elapsed;
822 }
823 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
824 if (td->rw_end_set[DDIR_WRITE])
825 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
826 else
827 elapsed = utime_since_now(&td->start);
828
829 runtime[DDIR_WRITE] += elapsed;
830 }
413dd459 831
ebac4655
JA
832 if (td->error || td->terminate)
833 break;
834
2dc1bbeb 835 if (td->o.verify == VERIFY_NONE)
ebac4655
JA
836 continue;
837
a978ba68
JA
838 if (clear_io_state(td))
839 break;
840
02bcaa8c 841 fio_gettime(&td->start, NULL);
ebac4655
JA
842
843 do_verify(td);
844
69008999 845 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
846
847 if (td->error || td->terminate)
848 break;
849 }
850
36dff966 851 update_rusage_stat(td);
756867bd
JA
852 td->ts.runtime[0] = runtime[0] / 1000;
853 td->ts.runtime[1] = runtime[1] / 1000;
854 td->ts.total_run_time = mtime_since_now(&td->epoch);
855 td->ts.io_bytes[0] = td->io_bytes[0];
856 td->ts.io_bytes[1] = td->io_bytes[1];
857
858 if (td->ts.bw_log)
859 finish_log(td, td->ts.bw_log, "bw");
860 if (td->ts.slat_log)
861 finish_log(td, td->ts.slat_log, "slat");
862 if (td->ts.clat_log)
863 finish_log(td, td->ts.clat_log, "clat");
2dc1bbeb 864 if (td->o.write_iolog_file)
843a7413 865 write_iolog_close(td);
2dc1bbeb
JA
866 if (td->o.exec_postrun) {
867 if (system(td->o.exec_postrun) < 0)
868 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
69cfd7e0 869 }
ebac4655
JA
870
871 if (exitall_on_terminate)
390c40e2 872 terminate_threads(td->groupid);
ebac4655
JA
873
874err:
5bf13a5a
JA
875 if (td->error)
876 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 877 close_files(td);
2866c82d 878 close_ioengine(td);
ebac4655 879 cleanup_io_u(td);
d23bb327 880 options_mem_free(td);
ebac4655 881 td_set_runstate(td, TD_EXITED);
43d76807 882 return (void *) (unsigned long) td->error;
bda4fd9e
JA
883err_sem:
884 fio_sem_up(startup_sem);
885 goto err;
ebac4655
JA
886}
887
906c8d75
JA
888/*
889 * We cannot pass the td data into a forked process, so attach the td and
890 * pass it to the thread worker.
891 */
a6418147 892static int fork_main(int shmid, int offset)
ebac4655
JA
893{
894 struct thread_data *td;
a6418147 895 void *data, *ret;
ebac4655
JA
896
897 data = shmat(shmid, NULL, 0);
898 if (data == (void *) -1) {
a6418147
JA
899 int __err = errno;
900
ebac4655 901 perror("shmat");
a6418147 902 return __err;
ebac4655
JA
903 }
904
905 td = data + offset * sizeof(struct thread_data);
a6418147 906 ret = thread_main(td);
ebac4655 907 shmdt(data);
43d76807 908 return (int) (unsigned long) ret;
ebac4655
JA
909}
910
906c8d75
JA
911/*
912 * Run over the job map and reap the threads that have exited, if any.
913 */
ebac4655
JA
914static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
915{
34572e28 916 struct thread_data *td;
fab6aa71 917 int i, cputhreads, pending, status, ret;
ebac4655
JA
918
919 /*
920 * reap exited threads (TD_EXITED -> TD_REAPED)
921 */
4d2413c6 922 pending = cputhreads = 0;
34572e28 923 for_each_td(td, i) {
3707f45b 924 int flags = 0;
a2f77c9f 925
84585003
JA
926 /*
927 * ->io_ops is NULL for a thread that has closed its
928 * io engine
929 */
ba0fbe10 930 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0
JA
931 cputhreads++;
932
a2f77c9f
JA
933 if (!td->pid || td->runstate == TD_REAPED)
934 continue;
2dc1bbeb 935 if (td->o.use_thread) {
3707f45b
JA
936 if (td->runstate == TD_EXITED) {
937 td_set_runstate(td, TD_REAPED);
938 goto reaped;
939 }
940 continue;
941 }
a2f77c9f
JA
942
943 flags = WNOHANG;
944 if (td->runstate == TD_EXITED)
945 flags = 0;
946
947 /*
948 * check if someone quit or got killed in an unusual way
949 */
950 ret = waitpid(td->pid, &status, flags);
3707f45b 951 if (ret < 0) {
a2f77c9f
JA
952 if (errno == ECHILD) {
953 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
954 td_set_runstate(td, TD_REAPED);
955 goto reaped;
956 }
957 perror("waitpid");
958 } else if (ret == td->pid) {
959 if (WIFSIGNALED(status)) {
fab6aa71
JA
960 int sig = WTERMSIG(status);
961
d9244941
JA
962 if (sig != SIGQUIT)
963 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
fab6aa71
JA
964 td_set_runstate(td, TD_REAPED);
965 goto reaped;
966 }
a2f77c9f
JA
967 if (WIFEXITED(status)) {
968 if (WEXITSTATUS(status) && !td->error)
969 td->error = WEXITSTATUS(status);
a2f77c9f 970
a2f77c9f
JA
971 td_set_runstate(td, TD_REAPED);
972 goto reaped;
a6418147
JA
973 }
974 }
ebac4655 975
a2f77c9f
JA
976 /*
977 * thread is not dead, continue
978 */
979 continue;
fab6aa71 980reaped:
2dc1bbeb 981 if (td->o.use_thread) {
3707f45b
JA
982 long ret;
983
984 if (pthread_join(td->thread, (void *) &ret))
985 perror("pthread_join");
986 }
987
ebac4655 988 (*nr_running)--;
2dc1bbeb
JA
989 (*m_rate) -= td->o.ratemin;
990 (*t_rate) -= td->o.rate;
a2f77c9f
JA
991
992 if (td->error)
993 exit_value++;
ebac4655 994 }
b990b5c0 995
4d2413c6 996 if (*nr_running == cputhreads && !pending)
390c40e2 997 terminate_threads(TERMINATE_ALL);
ebac4655
JA
998}
999
906c8d75
JA
1000/*
1001 * Main function for kicking off and reaping jobs, as needed.
1002 */
ebac4655
JA
1003static void run_threads(void)
1004{
ebac4655
JA
1005 struct thread_data *td;
1006 unsigned long spent;
1007 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1008
2f9ade3c
JA
1009 if (fio_pin_memory())
1010 return;
ebac4655 1011
c6ae0a5b 1012 if (!terse_output) {
9cedf167
JA
1013 printf("Starting ");
1014 if (nr_thread)
1015 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1016 if (nr_process) {
1017 if (nr_thread)
1018 printf(" and ");
1019 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1020 }
1021 printf("\n");
c6ae0a5b
JA
1022 fflush(stdout);
1023 }
c04f7ec3 1024
4efa970e
JA
1025 signal(SIGINT, sig_handler);
1026 signal(SIGALRM, sig_handler);
1027
ebac4655
JA
1028 todo = thread_number;
1029 nr_running = 0;
1030 nr_started = 0;
1031 m_rate = t_rate = 0;
1032
34572e28 1033 for_each_td(td, i) {
263e529f 1034 print_status_init(td->thread_number - 1);
ebac4655 1035
2dc1bbeb 1036 if (!td->o.create_serialize) {
380cf265 1037 init_disk_util(td);
ebac4655 1038 continue;
380cf265 1039 }
ebac4655
JA
1040
1041 /*
1042 * do file setup here so it happens sequentially,
1043 * we don't want X number of threads getting their
1044 * client data interspersed on disk
1045 */
53cdc686 1046 if (setup_files(td)) {
5bf13a5a
JA
1047 exit_value++;
1048 if (td->error)
1049 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
1050 td_set_runstate(td, TD_REAPED);
1051 todo--;
1052 }
380cf265
JA
1053
1054 init_disk_util(td);
ebac4655
JA
1055 }
1056
a2f77c9f
JA
1057 set_genesis_time();
1058
ebac4655 1059 while (todo) {
75154845
JA
1060 struct thread_data *map[MAX_JOBS];
1061 struct timeval this_start;
1062 int this_jobs = 0, left;
1063
ebac4655
JA
1064 /*
1065 * create threads (TD_NOT_CREATED -> TD_CREATED)
1066 */
34572e28 1067 for_each_td(td, i) {
ebac4655
JA
1068 if (td->runstate != TD_NOT_CREATED)
1069 continue;
1070
1071 /*
1072 * never got a chance to start, killed by other
1073 * thread for some reason
1074 */
1075 if (td->terminate) {
1076 todo--;
1077 continue;
1078 }
1079
2dc1bbeb 1080 if (td->o.start_delay) {
263e529f 1081 spent = mtime_since_genesis();
ebac4655 1082
2dc1bbeb 1083 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1084 continue;
1085 }
1086
2dc1bbeb 1087 if (td->o.stonewall && (nr_started || nr_running))
ebac4655
JA
1088 break;
1089
75154845
JA
1090 /*
1091 * Set state to created. Thread will transition
1092 * to TD_INITIALIZED when it's done setting up.
1093 */
ebac4655 1094 td_set_runstate(td, TD_CREATED);
75154845 1095 map[this_jobs++] = td;
ebac4655
JA
1096 nr_started++;
1097
2dc1bbeb 1098 if (td->o.use_thread) {
ebac4655
JA
1099 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1100 perror("thread_create");
1101 nr_started--;
c8bb6faf 1102 break;
ebac4655
JA
1103 }
1104 } else {
07739b57 1105 if (!fork()) {
a6418147
JA
1106 int ret = fork_main(shm_id, i);
1107
1108 exit(ret);
ebac4655
JA
1109 }
1110 }
07739b57 1111 fio_sem_down(startup_sem);
ebac4655
JA
1112 }
1113
1114 /*
75154845
JA
1115 * Wait for the started threads to transition to
1116 * TD_INITIALIZED.
ebac4655 1117 */
02bcaa8c 1118 fio_gettime(&this_start, NULL);
75154845 1119 left = this_jobs;
6ce15a32 1120 while (left && !fio_abort) {
75154845
JA
1121 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1122 break;
1123
1124 usleep(100000);
1125
1126 for (i = 0; i < this_jobs; i++) {
1127 td = map[i];
1128 if (!td)
1129 continue;
b6f4d880 1130 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1131 map[i] = NULL;
1132 left--;
b6f4d880
JA
1133 } else if (td->runstate >= TD_EXITED) {
1134 map[i] = NULL;
1135 left--;
1136 todo--;
1137 nr_running++; /* work-around... */
75154845
JA
1138 }
1139 }
1140 }
1141
1142 if (left) {
3b70d7e5 1143 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1144 for (i = 0; i < this_jobs; i++) {
1145 td = map[i];
1146 if (!td)
1147 continue;
1148 kill(td->pid, SIGTERM);
1149 }
1150 break;
1151 }
1152
1153 /*
b6f4d880 1154 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1155 */
34572e28 1156 for_each_td(td, i) {
75154845 1157 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1158 continue;
1159
1160 td_set_runstate(td, TD_RUNNING);
1161 nr_running++;
1162 nr_started--;
2dc1bbeb
JA
1163 m_rate += td->o.ratemin;
1164 t_rate += td->o.rate;
75154845 1165 todo--;
07739b57 1166 fio_sem_up(td->mutex);
ebac4655
JA
1167 }
1168
1169 reap_threads(&nr_running, &t_rate, &m_rate);
1170
1171 if (todo)
1172 usleep(100000);
1173 }
1174
1175 while (nr_running) {
1176 reap_threads(&nr_running, &t_rate, &m_rate);
1177 usleep(10000);
1178 }
1179
1180 update_io_ticks();
2f9ade3c 1181 fio_unpin_memory();
ebac4655
JA
1182}
1183
ebac4655
JA
1184int main(int argc, char *argv[])
1185{
29d610e1
JA
1186 long ps;
1187
dbe1125e
JA
1188 /*
1189 * We need locale for number printing, if it isn't set then just
1190 * go with the US format.
1191 */
1192 if (!getenv("LC_NUMERIC"))
1193 setlocale(LC_NUMERIC, "en_US");
1194
ebac4655
JA
1195 if (parse_options(argc, argv))
1196 return 1;
1197
1198 if (!thread_number) {
3b70d7e5 1199 log_err("Nothing to do\n");
ebac4655
JA
1200 return 1;
1201 }
1202
29d610e1
JA
1203 ps = sysconf(_SC_PAGESIZE);
1204 if (ps < 0) {
1205 log_err("Failed to get page size\n");
1206 return 1;
1207 }
1208
cfc99db7 1209 page_size = ps;
29d610e1
JA
1210 page_mask = ps - 1;
1211
bb3884d8
JA
1212 if (write_bw_log) {
1213 setup_log(&agg_io_log[DDIR_READ]);
1214 setup_log(&agg_io_log[DDIR_WRITE]);
1215 }
1216
07739b57
JA
1217 startup_sem = fio_sem_init(0);
1218
a2f77c9f
JA
1219 set_genesis_time();
1220
ebac4655
JA
1221 disk_util_timer_arm();
1222
1223 run_threads();
6ce15a32 1224
bb3884d8 1225 if (!fio_abort) {
6ce15a32 1226 show_run_stats();
bb3884d8
JA
1227 if (write_bw_log) {
1228 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1229 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1230 }
1231 }
ebac4655 1232
07739b57 1233 fio_sem_remove(startup_sem);
437c9b71 1234 return exit_value;
ebac4655 1235}