Add low water mark for queuing depth
[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
29d610e1
JA
40static unsigned long page_mask;
41#define ALIGN(buf) \
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
43
44int groupid = 0;
45int thread_number = 0;
ebac4655 46int shm_id = 0;
53cdc686 47int temp_stall_ts;
ebac4655 48
bbfd6b00 49static volatile int startup_sem;
6ce15a32 50static volatile int fio_abort;
437c9b71 51static int exit_value;
ebac4655 52
bb3884d8
JA
53struct io_log *agg_io_log[2];
54
ebac4655 55#define TERMINATE_ALL (-1)
75154845 56#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 57
6ce15a32
JA
58static inline void td_set_runstate(struct thread_data *td, int runstate)
59{
60 td->runstate = runstate;
61}
62
63static void terminate_threads(int group_id, int forced_kill)
ebac4655 64{
34572e28 65 struct thread_data *td;
ebac4655
JA
66 int i;
67
34572e28 68 for_each_td(td, i) {
ebac4655
JA
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70 td->terminate = 1;
71 td->start_delay = 0;
6ce15a32
JA
72 if (forced_kill)
73 td_set_runstate(td, TD_EXITED);
ebac4655
JA
74 }
75 }
76}
77
78static void sig_handler(int sig)
79{
80 switch (sig) {
81 case SIGALRM:
82 update_io_ticks();
83 disk_util_timer_arm();
84 print_thread_status();
85 break;
86 default:
6ce15a32 87 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 88 fflush(stdout);
6ce15a32 89 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
90 break;
91 }
92}
93
906c8d75
JA
94/*
95 * Check if we are above the minimum rate given.
96 */
ebac4655
JA
97static int check_min_rate(struct thread_data *td, struct timeval *now)
98{
99 unsigned long spent;
100 unsigned long rate;
101 int ddir = td->ddir;
102
103 /*
104 * allow a 2 second settle period in the beginning
105 */
106 if (mtime_since(&td->start, now) < 2000)
107 return 0;
108
109 /*
110 * if rate blocks is set, sample is running
111 */
112 if (td->rate_bytes) {
113 spent = mtime_since(&td->lastrate, now);
114 if (spent < td->ratecycle)
115 return 0;
116
117 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
118 if (rate < td->ratemin) {
1e97cce9 119 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
120 return 1;
121 }
122 }
123
124 td->rate_bytes = td->this_io_bytes[ddir];
125 memcpy(&td->lastrate, now, sizeof(*now));
126 return 0;
127}
128
129static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
130{
131 if (!td->timeout)
132 return 0;
133 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
134 return 1;
135
136 return 0;
137}
138
906c8d75
JA
139/*
140 * When job exits, we can cancel the in-flight IO if we are using async
141 * io. Attempt to do so.
142 */
ebac4655
JA
143static void cleanup_pending_aio(struct thread_data *td)
144{
ebac4655 145 struct list_head *entry, *n;
ebac4655
JA
146 struct io_u *io_u;
147 int r;
148
149 /*
150 * get immediately available events, if any
151 */
97601024 152 io_u_queued_complete(td, 0, NULL);
ebac4655
JA
153
154 /*
155 * now cancel remaining active events
156 */
2866c82d 157 if (td->io_ops->cancel) {
ebac4655
JA
158 list_for_each_safe(entry, n, &td->io_u_busylist) {
159 io_u = list_entry(entry, struct io_u, list);
160
2866c82d 161 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
162 if (!r)
163 put_io_u(td, io_u);
164 }
165 }
166
97601024
JA
167 if (td->cur_depth)
168 io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
169}
170
858a3d47
JA
171/*
172 * Helper to handle the final sync of a file. Works just like the normal
173 * io path, just does everything sync.
174 */
175static int fio_io_sync(struct thread_data *td, struct fio_file *f)
176{
177 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
178 int ret;
179
180 if (!io_u)
181 return 1;
182
183 io_u->ddir = DDIR_SYNC;
184 io_u->file = f;
185
186 if (td_io_prep(td, io_u)) {
187 put_io_u(td, io_u);
188 return 1;
189 }
190
755200a3 191requeue:
858a3d47 192 ret = td_io_queue(td, io_u);
36167d82 193 if (ret < 0) {
353a7e0e 194 td_verror(td, io_u->error);
858a3d47 195 put_io_u(td, io_u);
858a3d47 196 return 1;
36167d82 197 } else if (ret == FIO_Q_QUEUED) {
97601024 198 if (io_u_queued_complete(td, 1, NULL))
36167d82 199 return 1;
36167d82
JA
200 } else if (ret == FIO_Q_COMPLETED) {
201 if (io_u->error) {
202 td_verror(td, io_u->error);
203 return 1;
204 }
858a3d47 205
97601024 206 io_u_sync_complete(td, io_u, NULL);
755200a3
JA
207 } else if (ret == FIO_Q_BUSY) {
208 if (td_io_commit(td))
209 return 1;
210 goto requeue;
858a3d47
JA
211 }
212
213 return 0;
214}
215
906c8d75
JA
216/*
217 * The main verify engine. Runs over the writes we previusly submitted,
218 * reads the blocks back in, and checks the crc/md5 of the data.
219 */
1e97cce9 220static void do_verify(struct thread_data *td)
ebac4655 221{
53cdc686 222 struct fio_file *f;
36167d82 223 struct io_u *io_u;
3af6ef39 224 int ret, i, min_events;
e5b401d4
JA
225
226 /*
227 * sync io first and invalidate cache, to make sure we really
228 * read from disk.
229 */
230 for_each_file(td, f, i) {
858a3d47 231 fio_io_sync(td, f);
e5b401d4
JA
232 file_invalidate_cache(td, f);
233 }
ebac4655
JA
234
235 td_set_runstate(td, TD_VERIFYING);
236
36167d82
JA
237 io_u = NULL;
238 while (!td->terminate) {
ebac4655
JA
239 io_u = __get_io_u(td);
240 if (!io_u)
241 break;
242
36167d82 243 if (runtime_exceeded(td, &io_u->start_time))
02bcaa8c 244 break;
02bcaa8c 245
36167d82 246 if (get_next_verify(td, io_u))
ebac4655 247 break;
ebac4655 248
36167d82 249 if (td_io_prep(td, io_u))
53cdc686
JA
250 break;
251
36167d82
JA
252requeue:
253 ret = td_io_queue(td, io_u);
53cdc686 254
36167d82
JA
255 switch (ret) {
256 case FIO_Q_COMPLETED:
257 if (io_u->error)
22819ec2 258 ret = -io_u->error;
36167d82
JA
259 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
260 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 261
36167d82
JA
262 io_u->xfer_buflen = io_u->resid;
263 io_u->xfer_buf += bytes;
264 goto requeue;
265 }
99784632
JA
266 ret = io_u_sync_complete(td, io_u, verify_io_u);
267 if (ret)
36167d82 268 break;
36167d82
JA
269 continue;
270 case FIO_Q_QUEUED:
271 break;
755200a3
JA
272 case FIO_Q_BUSY:
273 requeue_io_u(td, &io_u);
274 ret = td_io_commit(td);
275 break;
36167d82
JA
276 default:
277 assert(ret < 0);
22819ec2 278 td_verror(td, -ret);
ebac4655
JA
279 break;
280 }
281
99784632 282 if (ret < 0 || td->error)
3af6ef39
JA
283 break;
284
ebac4655 285 /*
3af6ef39
JA
286 * if we can queue more, do so. but check if there are
287 * completed io_u's first.
ebac4655 288 */
97601024 289 min_events = 0;
e916b390 290 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 291 min_events = 1;
3af6ef39 292
e916b390
JA
293 if (td->cur_depth > td->iodepth_low)
294 min_events = td->cur_depth - td->iodepth_low;
295 }
296
3af6ef39
JA
297 /*
298 * Reap required number of io units, if any, and do the
299 * verification on them through the callback handler
300 */
97601024 301 if (io_u_queued_complete(td, min_events, verify_io_u))
ebac4655 302 break;
36167d82 303 }
ebac4655 304
36167d82
JA
305 if (io_u)
306 put_io_u(td, io_u);
ebac4655
JA
307
308 if (td->cur_depth)
309 cleanup_pending_aio(td);
310
311 td_set_runstate(td, TD_RUNNING);
312}
313
b990b5c0
JA
314/*
315 * Not really an io thread, all it does is burn CPU cycles in the specified
316 * manner.
317 */
318static void do_cpuio(struct thread_data *td)
319{
320 struct timeval e;
321 int split = 100 / td->cpuload;
322 int i = 0;
323
324 while (!td->terminate) {
02bcaa8c 325 fio_gettime(&e, NULL);
b990b5c0
JA
326
327 if (runtime_exceeded(td, &e))
328 break;
329
330 if (!(i % split))
331 __usec_sleep(10000);
332 else
333 usec_sleep(td, 10000);
334
335 i++;
336 }
337}
338
32cd46a0 339/*
906c8d75 340 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
341 * and reaps them, checking for rate and errors along the way.
342 */
ebac4655
JA
343static void do_io(struct thread_data *td)
344{
02bcaa8c 345 struct timeval s;
ebac4655 346 unsigned long usec;
84585003 347 int i, ret = 0;
ebac4655 348
5853e5a8
JA
349 td_set_runstate(td, TD_RUNNING);
350
3951414d 351 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
97601024
JA
352 struct timeval comp_time;
353 long bytes_done = 0;
84585003 354 int min_evts = 0;
ebac4655
JA
355 struct io_u *io_u;
356
357 if (td->terminate)
358 break;
359
3d7c391d 360 io_u = get_io_u(td);
ebac4655
JA
361 if (!io_u)
362 break;
363
364 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
365
366 if (runtime_exceeded(td, &s)) {
367 put_io_u(td, io_u);
368 break;
369 }
cec6b55d 370requeue:
45bee283 371 ret = td_io_queue(td, io_u);
36167d82
JA
372
373 switch (ret) {
374 case FIO_Q_COMPLETED:
375 if (io_u->error) {
376 ret = io_u->error;
377 break;
378 }
379 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
380 int bytes = io_u->xfer_buflen - io_u->resid;
381
cec6b55d 382 io_u->xfer_buflen = io_u->resid;
36167d82 383 io_u->xfer_buf += bytes;
cec6b55d 384 goto requeue;
cec6b55d 385 }
97601024
JA
386 fio_gettime(&comp_time, NULL);
387 bytes_done = io_u_sync_complete(td, io_u, NULL);
99784632
JA
388 if (bytes_done < 0)
389 ret = bytes_done;
36167d82
JA
390 break;
391 case FIO_Q_QUEUED:
7e77dd02
JA
392 /*
393 * if the engine doesn't have a commit hook,
394 * the io_u is really queued. if it does have such
395 * a hook, it has to call io_u_queued() itself.
396 */
397 if (td->io_ops->commit == NULL)
398 io_u_queued(td, io_u);
36167d82 399 break;
755200a3
JA
400 case FIO_Q_BUSY:
401 requeue_io_u(td, &io_u);
402 ret = td_io_commit(td);
403 break;
36167d82
JA
404 default:
405 assert(ret < 0);
406 put_io_u(td, io_u);
407 break;
ebac4655
JA
408 }
409
99784632 410 if (ret < 0 || td->error)
36167d82
JA
411 break;
412
97601024
JA
413 /*
414 * See if we need to complete some commands
415 */
755200a3 416 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 417 min_evts = 0;
e916b390 418 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 419 min_evts = 1;
ebac4655 420
e916b390
JA
421 if (td->cur_depth > td->iodepth_low)
422 min_evts = td->cur_depth - td->iodepth_low;
423 }
424
97601024
JA
425 fio_gettime(&comp_time, NULL);
426 bytes_done = io_u_queued_complete(td, min_evts, NULL);
427 if (bytes_done < 0)
36167d82 428 break;
ebac4655
JA
429 }
430
97601024
JA
431 if (!bytes_done)
432 continue;
433
ebac4655
JA
434 /*
435 * the rate is batched for now, it should work for batches
436 * of completions except the very first one which may look
437 * a little bursty
438 */
97601024 439 usec = utime_since(&s, &comp_time);
ebac4655 440
97601024 441 rate_throttle(td, usec, bytes_done, td->ddir);
ebac4655 442
97601024 443 if (check_min_rate(td, &comp_time)) {
98aa62d8 444 if (exitall_on_terminate)
6ce15a32 445 terminate_threads(td->groupid, 0);
fd841467 446 td_verror(td, ENODATA);
ebac4655
JA
447 break;
448 }
449
9c1f7434
JA
450 if (td->thinktime) {
451 unsigned long long b;
452
453 b = td->io_blocks[0] + td->io_blocks[1];
48097d5c
JA
454 if (!(b % td->thinktime_blocks)) {
455 int left;
456
457 if (td->thinktime_spin)
458 __usec_sleep(td->thinktime_spin);
459
460 left = td->thinktime - td->thinktime_spin;
461 if (left)
462 usec_sleep(td, left);
463 }
9c1f7434 464 }
ebac4655
JA
465 }
466
4d2413c6 467 if (!td->error) {
3d7c391d
JA
468 struct fio_file *f;
469
84585003
JA
470 if (td->cur_depth)
471 cleanup_pending_aio(td);
ebac4655 472
84585003
JA
473 if (should_fsync(td) && td->end_fsync) {
474 td_set_runstate(td, TD_FSYNCING);
475 for_each_file(td, f, i)
858a3d47 476 fio_io_sync(td, f);
84585003 477 }
5853e5a8 478 }
ebac4655
JA
479}
480
ebac4655
JA
481static void cleanup_io_u(struct thread_data *td)
482{
483 struct list_head *entry, *n;
484 struct io_u *io_u;
485
486 list_for_each_safe(entry, n, &td->io_u_freelist) {
487 io_u = list_entry(entry, struct io_u, list);
488
489 list_del(&io_u->list);
490 free(io_u);
491 }
492
2f9ade3c 493 free_io_mem(td);
ebac4655
JA
494}
495
6b9cea23
JA
496/*
497 * "randomly" fill the buffer contents
498 */
66eeb296 499static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 500{
66eeb296 501 int *ptr = io_u->buf;
6b9cea23
JA
502
503 while ((void *) ptr - io_u->buf < max_bs) {
504 *ptr = rand() * 0x9e370001;
505 ptr++;
506 }
507}
508
ebac4655
JA
509static int init_io_u(struct thread_data *td)
510{
511 struct io_u *io_u;
a00735e6 512 unsigned int max_bs;
ebac4655
JA
513 int i, max_units;
514 char *p;
515
2866c82d 516 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
517 return 0;
518
2866c82d 519 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
520 max_units = 1;
521 else
522 max_units = td->iodepth;
523
a00735e6 524 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
74b025b0
JA
525 td->orig_buffer_size = max_bs * max_units;
526
d0bdaf49 527 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
56bb17f2 528 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
74b025b0 529 else
29d610e1 530 td->orig_buffer_size += page_mask;
ebac4655 531
2f9ade3c
JA
532 if (allocate_io_mem(td))
533 return 1;
ebac4655 534
ebac4655
JA
535 p = ALIGN(td->orig_buffer);
536 for (i = 0; i < max_units; i++) {
537 io_u = malloc(sizeof(*io_u));
538 memset(io_u, 0, sizeof(*io_u));
539 INIT_LIST_HEAD(&io_u->list);
540
a00735e6 541 io_u->buf = p + max_bs * i;
6b9cea23 542 if (td_write(td) || td_rw(td))
a00735e6 543 fill_rand_buf(io_u, max_bs);
6b9cea23 544
b1ff3403 545 io_u->index = i;
ebac4655
JA
546 list_add(&io_u->list, &td->io_u_freelist);
547 }
548
549 return 0;
550}
551
da86774e
JA
552static int switch_ioscheduler(struct thread_data *td)
553{
554 char tmp[256], tmp2[128];
555 FILE *f;
556 int ret;
557
f48b467c
JA
558 if (td->io_ops->flags & FIO_CPUIO)
559 return 0;
560
da86774e
JA
561 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
562
563 f = fopen(tmp, "r+");
564 if (!f) {
565 td_verror(td, errno);
566 return 1;
567 }
568
569 /*
570 * Set io scheduler.
571 */
572 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
573 if (ferror(f) || ret != 1) {
574 td_verror(td, errno);
575 fclose(f);
576 return 1;
577 }
578
579 rewind(f);
580
581 /*
582 * Read back and check that the selected scheduler is now the default.
583 */
584 ret = fread(tmp, 1, sizeof(tmp), f);
585 if (ferror(f) || ret < 0) {
586 td_verror(td, errno);
587 fclose(f);
588 return 1;
589 }
590
591 sprintf(tmp2, "[%s]", td->ioscheduler);
592 if (!strstr(tmp, tmp2)) {
3b70d7e5 593 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
594 td_verror(td, EINVAL);
595 fclose(f);
596 return 1;
597 }
598
599 fclose(f);
600 return 0;
601}
602
ebac4655
JA
603static void clear_io_state(struct thread_data *td)
604{
53cdc686
JA
605 struct fio_file *f;
606 int i;
ebac4655 607
079ad09b 608 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 609 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 610 td->zone_bytes = 0;
ebac4655 611
c1324df1
JA
612 td->last_was_sync = 0;
613
53cdc686 614 for_each_file(td, f, i) {
c1324df1
JA
615 f->last_completed_pos = 0;
616
53cdc686
JA
617 f->last_pos = 0;
618 if (td->io_ops->flags & FIO_SYNCIO)
619 lseek(f->fd, SEEK_SET, 0);
620
621 if (f->file_map)
622 memset(f->file_map, 0, f->num_maps * sizeof(long));
623 }
ebac4655
JA
624}
625
906c8d75
JA
626/*
627 * Entry point for the thread based jobs. The process based jobs end up
628 * here as well, after a little setup.
629 */
ebac4655
JA
630static void *thread_main(void *data)
631{
69008999 632 unsigned long long runtime[2];
ebac4655 633 struct thread_data *td = data;
ebac4655
JA
634
635 if (!td->use_thread)
636 setsid();
637
638 td->pid = getpid();
639
aea47d44
JA
640 INIT_LIST_HEAD(&td->io_u_freelist);
641 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 642 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44
JA
643 INIT_LIST_HEAD(&td->io_hist_list);
644 INIT_LIST_HEAD(&td->io_log_list);
645
ebac4655
JA
646 if (init_io_u(td))
647 goto err;
648
649 if (fio_setaffinity(td) == -1) {
650 td_verror(td, errno);
651 goto err;
652 }
653
aea47d44
JA
654 if (init_iolog(td))
655 goto err;
656
ebac4655
JA
657 if (td->ioprio) {
658 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
659 td_verror(td, errno);
660 goto err;
661 }
662 }
663
1056eaad 664 if (nice(td->nice) == -1) {
b6f4d880
JA
665 td_verror(td, errno);
666 goto err;
667 }
668
75154845
JA
669 if (init_random_state(td))
670 goto err;
671
56f9498d
JA
672 if (td->ioscheduler && switch_ioscheduler(td))
673 goto err;
da86774e 674
75154845 675 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
676 fio_sem_up(&startup_sem);
677 fio_sem_down(&td->mutex);
ebac4655 678
53cdc686 679 if (!td->create_serialize && setup_files(td))
ebac4655 680 goto err;
21972cde
JA
681 if (open_files(td))
682 goto err;
ebac4655 683
7d6c5283
JA
684 /*
685 * Do this late, as some IO engines would like to have the
686 * files setup prior to initializing structures.
687 */
688 if (td_io_init(td))
689 goto err;
690
69cfd7e0
JA
691 if (td->exec_prerun) {
692 if (system(td->exec_prerun) < 0)
693 goto err;
694 }
4e0ba8af 695
69008999 696 fio_gettime(&td->epoch, NULL);
079ad09b 697 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
698
699 runtime[0] = runtime[1] = 0;
ebac4655 700 while (td->loops--) {
02bcaa8c 701 fio_gettime(&td->start, NULL);
079ad09b 702 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655
JA
703
704 if (td->ratemin)
079ad09b 705 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655
JA
706
707 clear_io_state(td);
708 prune_io_piece_log(td);
709
2866c82d 710 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
711 do_cpuio(td);
712 else
713 do_io(td);
ebac4655 714
69008999 715 runtime[td->ddir] += utime_since_now(&td->start);
aea47d44 716 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
69008999 717 runtime[td->ddir ^ 1] = runtime[td->ddir];
3d60d1ed 718
ebac4655
JA
719 if (td->error || td->terminate)
720 break;
721
722 if (td->verify == VERIFY_NONE)
723 continue;
724
725 clear_io_state(td);
02bcaa8c 726 fio_gettime(&td->start, NULL);
ebac4655
JA
727
728 do_verify(td);
729
69008999 730 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
731
732 if (td->error || td->terminate)
733 break;
734 }
735
36dff966 736 update_rusage_stat(td);
69008999
JA
737 fio_gettime(&td->end_time, NULL);
738 td->runtime[0] = runtime[0] / 1000;
739 td->runtime[1] = runtime[1] / 1000;
740
079ad09b
JA
741 if (td->ts.bw_log)
742 finish_log(td, td->ts.bw_log, "bw");
743 if (td->ts.slat_log)
744 finish_log(td, td->ts.slat_log, "slat");
745 if (td->ts.clat_log)
746 finish_log(td, td->ts.clat_log, "clat");
076efc7c 747 if (td->write_iolog_file)
843a7413 748 write_iolog_close(td);
69cfd7e0
JA
749 if (td->exec_postrun) {
750 if (system(td->exec_postrun) < 0)
751 log_err("fio: postrun %s failed\n", td->exec_postrun);
752 }
ebac4655
JA
753
754 if (exitall_on_terminate)
6ce15a32 755 terminate_threads(td->groupid, 0);
ebac4655
JA
756
757err:
5bf13a5a
JA
758 if (td->error)
759 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 760 close_files(td);
2866c82d 761 close_ioengine(td);
ebac4655 762 cleanup_io_u(td);
ebac4655 763 td_set_runstate(td, TD_EXITED);
a6418147 764 return (void *) td->error;
ebac4655
JA
765}
766
906c8d75
JA
767/*
768 * We cannot pass the td data into a forked process, so attach the td and
769 * pass it to the thread worker.
770 */
a6418147 771static int fork_main(int shmid, int offset)
ebac4655
JA
772{
773 struct thread_data *td;
a6418147 774 void *data, *ret;
ebac4655
JA
775
776 data = shmat(shmid, NULL, 0);
777 if (data == (void *) -1) {
a6418147
JA
778 int __err = errno;
779
ebac4655 780 perror("shmat");
a6418147 781 return __err;
ebac4655
JA
782 }
783
784 td = data + offset * sizeof(struct thread_data);
a6418147 785 ret = thread_main(td);
ebac4655 786 shmdt(data);
a6418147 787 return (int) ret;
ebac4655
JA
788}
789
906c8d75
JA
790/*
791 * Run over the job map and reap the threads that have exited, if any.
792 */
ebac4655
JA
793static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
794{
34572e28 795 struct thread_data *td;
fab6aa71 796 int i, cputhreads, pending, status, ret;
ebac4655
JA
797
798 /*
799 * reap exited threads (TD_EXITED -> TD_REAPED)
800 */
4d2413c6 801 pending = cputhreads = 0;
34572e28 802 for_each_td(td, i) {
84585003
JA
803 /*
804 * ->io_ops is NULL for a thread that has closed its
805 * io engine
806 */
807 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
808 cputhreads++;
809
fab6aa71
JA
810 if (td->runstate < TD_EXITED) {
811 /*
812 * check if someone quit or got killed in an unusual way
813 */
814 ret = waitpid(td->pid, &status, WNOHANG);
815 if (ret < 0)
816 perror("waitpid");
817 else if ((ret == td->pid) && WIFSIGNALED(status)) {
818 int sig = WTERMSIG(status);
819
820 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
821 td_set_runstate(td, TD_REAPED);
822 goto reaped;
823 }
824 }
825
4d2413c6
JA
826 if (td->runstate != TD_EXITED) {
827 if (td->runstate < TD_RUNNING)
828 pending++;
829
ebac4655 830 continue;
4d2413c6 831 }
ebac4655 832
437c9b71
JA
833 if (td->error)
834 exit_value++;
835
ebac4655
JA
836 td_set_runstate(td, TD_REAPED);
837
838 if (td->use_thread) {
839 long ret;
840
841 if (pthread_join(td->thread, (void *) &ret))
842 perror("thread_join");
a6418147
JA
843 } else {
844 int status;
845
fab6aa71
JA
846 ret = waitpid(td->pid, &status, 0);
847 if (ret < 0)
848 perror("waitpid");
73170f19 849 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
a6418147
JA
850 if (!exit_value)
851 exit_value++;
852 }
853 }
ebac4655 854
fab6aa71 855reaped:
ebac4655
JA
856 (*nr_running)--;
857 (*m_rate) -= td->ratemin;
858 (*t_rate) -= td->rate;
859 }
b990b5c0 860
4d2413c6 861 if (*nr_running == cputhreads && !pending)
6ce15a32 862 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
863}
864
906c8d75
JA
865/*
866 * Main function for kicking off and reaping jobs, as needed.
867 */
ebac4655
JA
868static void run_threads(void)
869{
ebac4655
JA
870 struct thread_data *td;
871 unsigned long spent;
872 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 873
2f9ade3c
JA
874 if (fio_pin_memory())
875 return;
ebac4655 876
c6ae0a5b
JA
877 if (!terse_output) {
878 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
879 fflush(stdout);
880 }
c04f7ec3 881
4efa970e
JA
882 signal(SIGINT, sig_handler);
883 signal(SIGALRM, sig_handler);
884
ebac4655
JA
885 todo = thread_number;
886 nr_running = 0;
887 nr_started = 0;
888 m_rate = t_rate = 0;
889
34572e28 890 for_each_td(td, i) {
263e529f 891 print_status_init(td->thread_number - 1);
ebac4655 892
380cf265
JA
893 if (!td->create_serialize) {
894 init_disk_util(td);
ebac4655 895 continue;
380cf265 896 }
ebac4655
JA
897
898 /*
899 * do file setup here so it happens sequentially,
900 * we don't want X number of threads getting their
901 * client data interspersed on disk
902 */
53cdc686 903 if (setup_files(td)) {
5bf13a5a
JA
904 exit_value++;
905 if (td->error)
906 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
907 td_set_runstate(td, TD_REAPED);
908 todo--;
909 }
380cf265
JA
910
911 init_disk_util(td);
ebac4655
JA
912 }
913
ebac4655 914 while (todo) {
75154845
JA
915 struct thread_data *map[MAX_JOBS];
916 struct timeval this_start;
917 int this_jobs = 0, left;
918
ebac4655
JA
919 /*
920 * create threads (TD_NOT_CREATED -> TD_CREATED)
921 */
34572e28 922 for_each_td(td, i) {
ebac4655
JA
923 if (td->runstate != TD_NOT_CREATED)
924 continue;
925
926 /*
927 * never got a chance to start, killed by other
928 * thread for some reason
929 */
930 if (td->terminate) {
931 todo--;
932 continue;
933 }
934
935 if (td->start_delay) {
263e529f 936 spent = mtime_since_genesis();
ebac4655
JA
937
938 if (td->start_delay * 1000 > spent)
939 continue;
940 }
941
942 if (td->stonewall && (nr_started || nr_running))
943 break;
944
75154845
JA
945 /*
946 * Set state to created. Thread will transition
947 * to TD_INITIALIZED when it's done setting up.
948 */
ebac4655 949 td_set_runstate(td, TD_CREATED);
75154845 950 map[this_jobs++] = td;
bbfd6b00 951 fio_sem_init(&startup_sem, 1);
ebac4655
JA
952 nr_started++;
953
954 if (td->use_thread) {
955 if (pthread_create(&td->thread, NULL, thread_main, td)) {
956 perror("thread_create");
957 nr_started--;
958 }
959 } else {
960 if (fork())
bbfd6b00 961 fio_sem_down(&startup_sem);
ebac4655 962 else {
a6418147
JA
963 int ret = fork_main(shm_id, i);
964
965 exit(ret);
ebac4655
JA
966 }
967 }
968 }
969
970 /*
75154845
JA
971 * Wait for the started threads to transition to
972 * TD_INITIALIZED.
ebac4655 973 */
02bcaa8c 974 fio_gettime(&this_start, NULL);
75154845 975 left = this_jobs;
6ce15a32 976 while (left && !fio_abort) {
75154845
JA
977 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
978 break;
979
980 usleep(100000);
981
982 for (i = 0; i < this_jobs; i++) {
983 td = map[i];
984 if (!td)
985 continue;
b6f4d880 986 if (td->runstate == TD_INITIALIZED) {
75154845
JA
987 map[i] = NULL;
988 left--;
b6f4d880
JA
989 } else if (td->runstate >= TD_EXITED) {
990 map[i] = NULL;
991 left--;
992 todo--;
993 nr_running++; /* work-around... */
75154845
JA
994 }
995 }
996 }
997
998 if (left) {
3b70d7e5 999 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1000 for (i = 0; i < this_jobs; i++) {
1001 td = map[i];
1002 if (!td)
1003 continue;
1004 kill(td->pid, SIGTERM);
1005 }
1006 break;
1007 }
1008
1009 /*
b6f4d880 1010 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1011 */
34572e28 1012 for_each_td(td, i) {
75154845 1013 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1014 continue;
1015
1016 td_set_runstate(td, TD_RUNNING);
1017 nr_running++;
1018 nr_started--;
1019 m_rate += td->ratemin;
1020 t_rate += td->rate;
75154845 1021 todo--;
bbfd6b00 1022 fio_sem_up(&td->mutex);
ebac4655
JA
1023 }
1024
1025 reap_threads(&nr_running, &t_rate, &m_rate);
1026
1027 if (todo)
1028 usleep(100000);
1029 }
1030
1031 while (nr_running) {
1032 reap_threads(&nr_running, &t_rate, &m_rate);
1033 usleep(10000);
1034 }
1035
1036 update_io_ticks();
2f9ade3c 1037 fio_unpin_memory();
ebac4655
JA
1038}
1039
ebac4655
JA
1040int main(int argc, char *argv[])
1041{
29d610e1
JA
1042 long ps;
1043
dbe1125e
JA
1044 /*
1045 * We need locale for number printing, if it isn't set then just
1046 * go with the US format.
1047 */
1048 if (!getenv("LC_NUMERIC"))
1049 setlocale(LC_NUMERIC, "en_US");
1050
ebac4655
JA
1051 if (parse_options(argc, argv))
1052 return 1;
1053
1054 if (!thread_number) {
3b70d7e5 1055 log_err("Nothing to do\n");
ebac4655
JA
1056 return 1;
1057 }
1058
29d610e1
JA
1059 ps = sysconf(_SC_PAGESIZE);
1060 if (ps < 0) {
1061 log_err("Failed to get page size\n");
1062 return 1;
1063 }
1064
1065 page_mask = ps - 1;
1066
bb3884d8
JA
1067 if (write_bw_log) {
1068 setup_log(&agg_io_log[DDIR_READ]);
1069 setup_log(&agg_io_log[DDIR_WRITE]);
1070 }
1071
ebac4655
JA
1072 disk_util_timer_arm();
1073
1074 run_threads();
6ce15a32 1075
bb3884d8 1076 if (!fio_abort) {
6ce15a32 1077 show_run_stats();
bb3884d8
JA
1078 if (write_bw_log) {
1079 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1080 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1081 }
1082 }
ebac4655 1083
437c9b71 1084 return exit_value;
ebac4655 1085}