IO engine callback need not dump possible values
[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{
0904200b 99 unsigned long long bytes = 0;
ebac4655
JA
100 unsigned long spent;
101 unsigned long rate;
ebac4655
JA
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
0904200b
JA
109 if (td_read(td))
110 bytes += td->this_io_bytes[DDIR_READ];
111 if (td_write(td))
112 bytes += td->this_io_bytes[DDIR_WRITE];
113
ebac4655
JA
114 /*
115 * if rate blocks is set, sample is running
116 */
117 if (td->rate_bytes) {
118 spent = mtime_since(&td->lastrate, now);
119 if (spent < td->ratecycle)
120 return 0;
121
413dd459
JA
122 if (bytes < td->rate_bytes) {
123 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
ebac4655 124 return 1;
413dd459
JA
125 } else {
126 rate = (bytes - td->rate_bytes) / spent;
127 if (rate < td->ratemin || bytes < td->rate_bytes) {
128 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
129 return 1;
130 }
ebac4655
JA
131 }
132 }
133
0904200b 134 td->rate_bytes = bytes;
ebac4655
JA
135 memcpy(&td->lastrate, now, sizeof(*now));
136 return 0;
137}
138
139static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
140{
141 if (!td->timeout)
142 return 0;
143 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
144 return 1;
145
146 return 0;
147}
148
906c8d75
JA
149/*
150 * When job exits, we can cancel the in-flight IO if we are using async
151 * io. Attempt to do so.
152 */
ebac4655
JA
153static void cleanup_pending_aio(struct thread_data *td)
154{
ebac4655 155 struct list_head *entry, *n;
ebac4655
JA
156 struct io_u *io_u;
157 int r;
158
159 /*
160 * get immediately available events, if any
161 */
b2fdda43
JA
162 r = io_u_queued_complete(td, 0, NULL);
163 if (r < 0)
164 return;
ebac4655
JA
165
166 /*
167 * now cancel remaining active events
168 */
2866c82d 169 if (td->io_ops->cancel) {
ebac4655
JA
170 list_for_each_safe(entry, n, &td->io_u_busylist) {
171 io_u = list_entry(entry, struct io_u, list);
172
0c6e7517
JA
173 /*
174 * if the io_u isn't in flight, then that generally
175 * means someone leaked an io_u. complain but fix
176 * it up, so we don't stall here.
177 */
178 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
179 log_err("fio: non-busy IO on busy list\n");
ebac4655 180 put_io_u(td, io_u);
0c6e7517
JA
181 } else {
182 r = td->io_ops->cancel(td, io_u);
183 if (!r)
184 put_io_u(td, io_u);
185 }
ebac4655
JA
186 }
187 }
188
97601024 189 if (td->cur_depth)
b2fdda43 190 r = io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
191}
192
858a3d47
JA
193/*
194 * Helper to handle the final sync of a file. Works just like the normal
195 * io path, just does everything sync.
196 */
197static int fio_io_sync(struct thread_data *td, struct fio_file *f)
198{
199 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
200 int ret;
201
202 if (!io_u)
203 return 1;
204
205 io_u->ddir = DDIR_SYNC;
206 io_u->file = f;
207
208 if (td_io_prep(td, io_u)) {
209 put_io_u(td, io_u);
210 return 1;
211 }
212
755200a3 213requeue:
858a3d47 214 ret = td_io_queue(td, io_u);
36167d82 215 if (ret < 0) {
e1161c32 216 td_verror(td, io_u->error, "td_io_queue");
858a3d47 217 put_io_u(td, io_u);
858a3d47 218 return 1;
36167d82 219 } else if (ret == FIO_Q_QUEUED) {
49db69aa 220 if (io_u_queued_complete(td, 1, NULL) < 0)
36167d82 221 return 1;
36167d82
JA
222 } else if (ret == FIO_Q_COMPLETED) {
223 if (io_u->error) {
e1161c32 224 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
225 return 1;
226 }
858a3d47 227
b2fdda43
JA
228 if (io_u_sync_complete(td, io_u, NULL) < 0)
229 return 1;
755200a3
JA
230 } else if (ret == FIO_Q_BUSY) {
231 if (td_io_commit(td))
232 return 1;
233 goto requeue;
858a3d47
JA
234 }
235
236 return 0;
237}
238
906c8d75
JA
239/*
240 * The main verify engine. Runs over the writes we previusly submitted,
241 * reads the blocks back in, and checks the crc/md5 of the data.
242 */
1e97cce9 243static void do_verify(struct thread_data *td)
ebac4655 244{
53cdc686 245 struct fio_file *f;
36167d82 246 struct io_u *io_u;
3af6ef39 247 int ret, i, min_events;
e5b401d4
JA
248
249 /*
250 * sync io first and invalidate cache, to make sure we really
251 * read from disk.
252 */
253 for_each_file(td, f, i) {
b2fdda43
JA
254 if (fio_io_sync(td, f))
255 break;
256 if (file_invalidate_cache(td, f))
257 break;
e5b401d4 258 }
ebac4655 259
b2fdda43
JA
260 if (td->error)
261 return;
262
ebac4655
JA
263 td_set_runstate(td, TD_VERIFYING);
264
36167d82
JA
265 io_u = NULL;
266 while (!td->terminate) {
ebac4655
JA
267 io_u = __get_io_u(td);
268 if (!io_u)
269 break;
270
069c2918
JA
271 if (runtime_exceeded(td, &io_u->start_time)) {
272 put_io_u(td, io_u);
02bcaa8c 273 break;
069c2918 274 }
02bcaa8c 275
069c2918
JA
276 if (get_next_verify(td, io_u)) {
277 put_io_u(td, io_u);
ebac4655 278 break;
069c2918 279 }
ebac4655 280
069c2918
JA
281 if (td_io_prep(td, io_u)) {
282 put_io_u(td, io_u);
53cdc686 283 break;
069c2918 284 }
36167d82
JA
285requeue:
286 ret = td_io_queue(td, io_u);
53cdc686 287
36167d82
JA
288 switch (ret) {
289 case FIO_Q_COMPLETED:
290 if (io_u->error)
22819ec2 291 ret = -io_u->error;
36167d82
JA
292 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
293 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 294
36167d82
JA
295 io_u->xfer_buflen = io_u->resid;
296 io_u->xfer_buf += bytes;
297 goto requeue;
298 }
99784632 299 ret = io_u_sync_complete(td, io_u, verify_io_u);
49db69aa 300 if (ret < 0)
36167d82 301 break;
36167d82
JA
302 continue;
303 case FIO_Q_QUEUED:
304 break;
755200a3
JA
305 case FIO_Q_BUSY:
306 requeue_io_u(td, &io_u);
307 ret = td_io_commit(td);
308 break;
36167d82
JA
309 default:
310 assert(ret < 0);
e1161c32 311 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
312 break;
313 }
314
99784632 315 if (ret < 0 || td->error)
3af6ef39
JA
316 break;
317
ebac4655 318 /*
3af6ef39
JA
319 * if we can queue more, do so. but check if there are
320 * completed io_u's first.
ebac4655 321 */
97601024 322 min_events = 0;
e916b390 323 if (queue_full(td) || ret == FIO_Q_BUSY) {
3af6ef39 324 min_events = 1;
3af6ef39 325
e916b390
JA
326 if (td->cur_depth > td->iodepth_low)
327 min_events = td->cur_depth - td->iodepth_low;
328 }
329
3af6ef39
JA
330 /*
331 * Reap required number of io units, if any, and do the
332 * verification on them through the callback handler
333 */
49db69aa 334 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
ebac4655 335 break;
36167d82 336 }
ebac4655 337
ebac4655
JA
338 if (td->cur_depth)
339 cleanup_pending_aio(td);
340
341 td_set_runstate(td, TD_RUNNING);
342}
343
b990b5c0
JA
344/*
345 * Not really an io thread, all it does is burn CPU cycles in the specified
346 * manner.
347 */
348static void do_cpuio(struct thread_data *td)
349{
350 struct timeval e;
351 int split = 100 / td->cpuload;
352 int i = 0;
353
354 while (!td->terminate) {
02bcaa8c 355 fio_gettime(&e, NULL);
b990b5c0
JA
356
357 if (runtime_exceeded(td, &e))
358 break;
359
360 if (!(i % split))
361 __usec_sleep(10000);
362 else
363 usec_sleep(td, 10000);
364
365 i++;
366 }
367}
368
32cd46a0 369/*
906c8d75 370 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
371 * and reaps them, checking for rate and errors along the way.
372 */
ebac4655
JA
373static void do_io(struct thread_data *td)
374{
02bcaa8c 375 struct timeval s;
ebac4655 376 unsigned long usec;
84585003 377 int i, ret = 0;
ebac4655 378
5853e5a8
JA
379 td_set_runstate(td, TD_RUNNING);
380
3951414d 381 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
97601024
JA
382 struct timeval comp_time;
383 long bytes_done = 0;
84585003 384 int min_evts = 0;
ebac4655
JA
385 struct io_u *io_u;
386
387 if (td->terminate)
388 break;
389
3d7c391d 390 io_u = get_io_u(td);
ebac4655
JA
391 if (!io_u)
392 break;
393
394 memcpy(&s, &io_u->start_time, sizeof(s));
97601024
JA
395
396 if (runtime_exceeded(td, &s)) {
397 put_io_u(td, io_u);
398 break;
399 }
cec6b55d 400requeue:
45bee283 401 ret = td_io_queue(td, io_u);
36167d82
JA
402
403 switch (ret) {
404 case FIO_Q_COMPLETED:
405 if (io_u->error) {
406 ret = io_u->error;
407 break;
408 }
409 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
410 int bytes = io_u->xfer_buflen - io_u->resid;
411
cec6b55d 412 io_u->xfer_buflen = io_u->resid;
36167d82 413 io_u->xfer_buf += bytes;
cec6b55d 414 goto requeue;
cec6b55d 415 }
97601024
JA
416 fio_gettime(&comp_time, NULL);
417 bytes_done = io_u_sync_complete(td, io_u, NULL);
99784632
JA
418 if (bytes_done < 0)
419 ret = bytes_done;
36167d82
JA
420 break;
421 case FIO_Q_QUEUED:
7e77dd02
JA
422 /*
423 * if the engine doesn't have a commit hook,
424 * the io_u is really queued. if it does have such
425 * a hook, it has to call io_u_queued() itself.
426 */
427 if (td->io_ops->commit == NULL)
428 io_u_queued(td, io_u);
36167d82 429 break;
755200a3
JA
430 case FIO_Q_BUSY:
431 requeue_io_u(td, &io_u);
432 ret = td_io_commit(td);
433 break;
36167d82
JA
434 default:
435 assert(ret < 0);
436 put_io_u(td, io_u);
437 break;
ebac4655
JA
438 }
439
99784632 440 if (ret < 0 || td->error)
36167d82
JA
441 break;
442
97601024
JA
443 /*
444 * See if we need to complete some commands
445 */
755200a3 446 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
97601024 447 min_evts = 0;
e916b390 448 if (queue_full(td) || ret == FIO_Q_BUSY) {
36167d82 449 min_evts = 1;
ebac4655 450
e916b390
JA
451 if (td->cur_depth > td->iodepth_low)
452 min_evts = td->cur_depth - td->iodepth_low;
453 }
454
97601024
JA
455 fio_gettime(&comp_time, NULL);
456 bytes_done = io_u_queued_complete(td, min_evts, NULL);
457 if (bytes_done < 0)
36167d82 458 break;
ebac4655
JA
459 }
460
97601024
JA
461 if (!bytes_done)
462 continue;
463
ebac4655
JA
464 /*
465 * the rate is batched for now, it should work for batches
466 * of completions except the very first one which may look
467 * a little bursty
468 */
97601024 469 usec = utime_since(&s, &comp_time);
ebac4655 470
413dd459 471 rate_throttle(td, usec, bytes_done);
ebac4655 472
97601024 473 if (check_min_rate(td, &comp_time)) {
98aa62d8 474 if (exitall_on_terminate)
6ce15a32 475 terminate_threads(td->groupid, 0);
e1161c32 476 td_verror(td, ENODATA, "check_min_rate");
ebac4655
JA
477 break;
478 }
479
9c1f7434
JA
480 if (td->thinktime) {
481 unsigned long long b;
482
483 b = td->io_blocks[0] + td->io_blocks[1];
48097d5c
JA
484 if (!(b % td->thinktime_blocks)) {
485 int left;
486
487 if (td->thinktime_spin)
488 __usec_sleep(td->thinktime_spin);
489
490 left = td->thinktime - td->thinktime_spin;
491 if (left)
492 usec_sleep(td, left);
493 }
9c1f7434 494 }
ebac4655
JA
495 }
496
4d2413c6 497 if (!td->error) {
3d7c391d
JA
498 struct fio_file *f;
499
84585003
JA
500 if (td->cur_depth)
501 cleanup_pending_aio(td);
ebac4655 502
84585003
JA
503 if (should_fsync(td) && td->end_fsync) {
504 td_set_runstate(td, TD_FSYNCING);
505 for_each_file(td, f, i)
858a3d47 506 fio_io_sync(td, f);
84585003 507 }
5853e5a8 508 }
ebac4655
JA
509}
510
ebac4655
JA
511static void cleanup_io_u(struct thread_data *td)
512{
513 struct list_head *entry, *n;
514 struct io_u *io_u;
515
516 list_for_each_safe(entry, n, &td->io_u_freelist) {
517 io_u = list_entry(entry, struct io_u, list);
518
519 list_del(&io_u->list);
520 free(io_u);
521 }
522
2f9ade3c 523 free_io_mem(td);
ebac4655
JA
524}
525
6b9cea23
JA
526/*
527 * "randomly" fill the buffer contents
528 */
66eeb296 529static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 530{
66eeb296 531 int *ptr = io_u->buf;
6b9cea23
JA
532
533 while ((void *) ptr - io_u->buf < max_bs) {
534 *ptr = rand() * 0x9e370001;
535 ptr++;
536 }
537}
538
ebac4655
JA
539static int init_io_u(struct thread_data *td)
540{
541 struct io_u *io_u;
a00735e6 542 unsigned int max_bs;
ebac4655
JA
543 int i, max_units;
544 char *p;
545
2866c82d 546 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
547 return 0;
548
2866c82d 549 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
550 max_units = 1;
551 else
552 max_units = td->iodepth;
553
a00735e6 554 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
74b025b0
JA
555 td->orig_buffer_size = max_bs * max_units;
556
d0bdaf49 557 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
56bb17f2 558 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
74b025b0 559 else
29d610e1 560 td->orig_buffer_size += page_mask;
ebac4655 561
2f9ade3c
JA
562 if (allocate_io_mem(td))
563 return 1;
ebac4655 564
ebac4655
JA
565 p = ALIGN(td->orig_buffer);
566 for (i = 0; i < max_units; i++) {
567 io_u = malloc(sizeof(*io_u));
568 memset(io_u, 0, sizeof(*io_u));
569 INIT_LIST_HEAD(&io_u->list);
570
a00735e6 571 io_u->buf = p + max_bs * i;
6b9cea23 572 if (td_write(td) || td_rw(td))
a00735e6 573 fill_rand_buf(io_u, max_bs);
6b9cea23 574
b1ff3403 575 io_u->index = i;
0c6e7517 576 io_u->flags = IO_U_F_FREE;
ebac4655
JA
577 list_add(&io_u->list, &td->io_u_freelist);
578 }
579
433afcb4
JA
580 io_u_init_timeout();
581
ebac4655
JA
582 return 0;
583}
584
da86774e
JA
585static int switch_ioscheduler(struct thread_data *td)
586{
587 char tmp[256], tmp2[128];
588 FILE *f;
589 int ret;
590
f48b467c
JA
591 if (td->io_ops->flags & FIO_CPUIO)
592 return 0;
593
da86774e
JA
594 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
595
596 f = fopen(tmp, "r+");
597 if (!f) {
e1161c32 598 td_verror(td, errno, "fopen");
da86774e
JA
599 return 1;
600 }
601
602 /*
603 * Set io scheduler.
604 */
605 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
606 if (ferror(f) || ret != 1) {
e1161c32 607 td_verror(td, errno, "fwrite");
da86774e
JA
608 fclose(f);
609 return 1;
610 }
611
612 rewind(f);
613
614 /*
615 * Read back and check that the selected scheduler is now the default.
616 */
617 ret = fread(tmp, 1, sizeof(tmp), f);
618 if (ferror(f) || ret < 0) {
e1161c32 619 td_verror(td, errno, "fread");
da86774e
JA
620 fclose(f);
621 return 1;
622 }
623
624 sprintf(tmp2, "[%s]", td->ioscheduler);
625 if (!strstr(tmp, tmp2)) {
3b70d7e5 626 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
e1161c32 627 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
628 fclose(f);
629 return 1;
630 }
631
632 fclose(f);
633 return 0;
634}
635
ebac4655
JA
636static void clear_io_state(struct thread_data *td)
637{
53cdc686
JA
638 struct fio_file *f;
639 int i;
ebac4655 640
079ad09b 641 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 642 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 643 td->zone_bytes = 0;
ebac4655 644
c1324df1
JA
645 td->last_was_sync = 0;
646
53cdc686 647 for_each_file(td, f, i) {
c1324df1
JA
648 f->last_completed_pos = 0;
649
53cdc686
JA
650 f->last_pos = 0;
651 if (td->io_ops->flags & FIO_SYNCIO)
652 lseek(f->fd, SEEK_SET, 0);
653
654 if (f->file_map)
655 memset(f->file_map, 0, f->num_maps * sizeof(long));
656 }
ebac4655
JA
657}
658
906c8d75
JA
659/*
660 * Entry point for the thread based jobs. The process based jobs end up
661 * here as well, after a little setup.
662 */
ebac4655
JA
663static void *thread_main(void *data)
664{
69008999 665 unsigned long long runtime[2];
ebac4655 666 struct thread_data *td = data;
ebac4655
JA
667
668 if (!td->use_thread)
669 setsid();
670
671 td->pid = getpid();
672
aea47d44
JA
673 INIT_LIST_HEAD(&td->io_u_freelist);
674 INIT_LIST_HEAD(&td->io_u_busylist);
755200a3 675 INIT_LIST_HEAD(&td->io_u_requeues);
aea47d44
JA
676 INIT_LIST_HEAD(&td->io_hist_list);
677 INIT_LIST_HEAD(&td->io_log_list);
678
ebac4655
JA
679 if (init_io_u(td))
680 goto err;
681
682 if (fio_setaffinity(td) == -1) {
e1161c32 683 td_verror(td, errno, "cpu_set_affinity");
ebac4655
JA
684 goto err;
685 }
686
aea47d44
JA
687 if (init_iolog(td))
688 goto err;
689
ebac4655
JA
690 if (td->ioprio) {
691 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 692 td_verror(td, errno, "ioprio_set");
ebac4655
JA
693 goto err;
694 }
695 }
696
1056eaad 697 if (nice(td->nice) == -1) {
e1161c32 698 td_verror(td, errno, "nice");
b6f4d880
JA
699 goto err;
700 }
701
75154845
JA
702 if (init_random_state(td))
703 goto err;
704
56f9498d
JA
705 if (td->ioscheduler && switch_ioscheduler(td))
706 goto err;
da86774e 707
75154845 708 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
709 fio_sem_up(&startup_sem);
710 fio_sem_down(&td->mutex);
ebac4655 711
53cdc686 712 if (!td->create_serialize && setup_files(td))
ebac4655 713 goto err;
21972cde
JA
714 if (open_files(td))
715 goto err;
ebac4655 716
7d6c5283
JA
717 /*
718 * Do this late, as some IO engines would like to have the
719 * files setup prior to initializing structures.
720 */
721 if (td_io_init(td))
722 goto err;
723
69cfd7e0
JA
724 if (td->exec_prerun) {
725 if (system(td->exec_prerun) < 0)
726 goto err;
727 }
4e0ba8af 728
69008999 729 fio_gettime(&td->epoch, NULL);
433afcb4 730 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
079ad09b 731 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
732
733 runtime[0] = runtime[1] = 0;
ebac4655 734 while (td->loops--) {
02bcaa8c 735 fio_gettime(&td->start, NULL);
079ad09b 736 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
ebac4655
JA
737
738 if (td->ratemin)
079ad09b 739 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
ebac4655
JA
740
741 clear_io_state(td);
742 prune_io_piece_log(td);
743
2866c82d 744 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
745 do_cpuio(td);
746 else
747 do_io(td);
ebac4655 748
413dd459
JA
749 if (td_read(td) && td->io_bytes[DDIR_READ])
750 runtime[DDIR_READ] += utime_since_now(&td->start);
751 if (td_write(td) && td->io_bytes[DDIR_WRITE])
752 runtime[DDIR_WRITE] += utime_since_now(&td->start);
753
ebac4655
JA
754 if (td->error || td->terminate)
755 break;
756
757 if (td->verify == VERIFY_NONE)
758 continue;
759
760 clear_io_state(td);
02bcaa8c 761 fio_gettime(&td->start, NULL);
ebac4655
JA
762
763 do_verify(td);
764
69008999 765 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
766
767 if (td->error || td->terminate)
768 break;
769 }
770
36dff966 771 update_rusage_stat(td);
69008999
JA
772 fio_gettime(&td->end_time, NULL);
773 td->runtime[0] = runtime[0] / 1000;
774 td->runtime[1] = runtime[1] / 1000;
775
079ad09b
JA
776 if (td->ts.bw_log)
777 finish_log(td, td->ts.bw_log, "bw");
778 if (td->ts.slat_log)
779 finish_log(td, td->ts.slat_log, "slat");
780 if (td->ts.clat_log)
781 finish_log(td, td->ts.clat_log, "clat");
076efc7c 782 if (td->write_iolog_file)
843a7413 783 write_iolog_close(td);
69cfd7e0
JA
784 if (td->exec_postrun) {
785 if (system(td->exec_postrun) < 0)
786 log_err("fio: postrun %s failed\n", td->exec_postrun);
787 }
ebac4655
JA
788
789 if (exitall_on_terminate)
6ce15a32 790 terminate_threads(td->groupid, 0);
ebac4655
JA
791
792err:
5bf13a5a
JA
793 if (td->error)
794 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
53cdc686 795 close_files(td);
2866c82d 796 close_ioengine(td);
ebac4655 797 cleanup_io_u(td);
ebac4655 798 td_set_runstate(td, TD_EXITED);
43d76807 799 return (void *) (unsigned long) td->error;
ebac4655
JA
800}
801
906c8d75
JA
802/*
803 * We cannot pass the td data into a forked process, so attach the td and
804 * pass it to the thread worker.
805 */
a6418147 806static int fork_main(int shmid, int offset)
ebac4655
JA
807{
808 struct thread_data *td;
a6418147 809 void *data, *ret;
ebac4655
JA
810
811 data = shmat(shmid, NULL, 0);
812 if (data == (void *) -1) {
a6418147
JA
813 int __err = errno;
814
ebac4655 815 perror("shmat");
a6418147 816 return __err;
ebac4655
JA
817 }
818
819 td = data + offset * sizeof(struct thread_data);
a6418147 820 ret = thread_main(td);
ebac4655 821 shmdt(data);
43d76807 822 return (int) (unsigned long) ret;
ebac4655
JA
823}
824
906c8d75
JA
825/*
826 * Run over the job map and reap the threads that have exited, if any.
827 */
ebac4655
JA
828static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
829{
34572e28 830 struct thread_data *td;
fab6aa71 831 int i, cputhreads, pending, status, ret;
ebac4655
JA
832
833 /*
834 * reap exited threads (TD_EXITED -> TD_REAPED)
835 */
4d2413c6 836 pending = cputhreads = 0;
34572e28 837 for_each_td(td, i) {
3707f45b 838 int flags = 0;
a2f77c9f 839
84585003
JA
840 /*
841 * ->io_ops is NULL for a thread that has closed its
842 * io engine
843 */
844 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
845 cputhreads++;
846
a2f77c9f
JA
847 if (!td->pid || td->runstate == TD_REAPED)
848 continue;
3707f45b
JA
849 if (td->use_thread) {
850 if (td->runstate == TD_EXITED) {
851 td_set_runstate(td, TD_REAPED);
852 goto reaped;
853 }
854 continue;
855 }
a2f77c9f
JA
856
857 flags = WNOHANG;
858 if (td->runstate == TD_EXITED)
859 flags = 0;
860
861 /*
862 * check if someone quit or got killed in an unusual way
863 */
864 ret = waitpid(td->pid, &status, flags);
3707f45b 865 if (ret < 0) {
a2f77c9f
JA
866 if (errno == ECHILD) {
867 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
868 td_set_runstate(td, TD_REAPED);
869 goto reaped;
870 }
871 perror("waitpid");
872 } else if (ret == td->pid) {
873 if (WIFSIGNALED(status)) {
fab6aa71
JA
874 int sig = WTERMSIG(status);
875
876 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
877 td_set_runstate(td, TD_REAPED);
878 goto reaped;
879 }
a2f77c9f
JA
880 if (WIFEXITED(status)) {
881 if (WEXITSTATUS(status) && !td->error)
882 td->error = WEXITSTATUS(status);
a2f77c9f 883
a2f77c9f
JA
884 td_set_runstate(td, TD_REAPED);
885 goto reaped;
a6418147
JA
886 }
887 }
ebac4655 888
a2f77c9f
JA
889 /*
890 * thread is not dead, continue
891 */
892 continue;
fab6aa71 893reaped:
3707f45b
JA
894 if (td->use_thread) {
895 long ret;
896
897 if (pthread_join(td->thread, (void *) &ret))
898 perror("pthread_join");
899 }
900
ebac4655
JA
901 (*nr_running)--;
902 (*m_rate) -= td->ratemin;
903 (*t_rate) -= td->rate;
a2f77c9f
JA
904
905 if (td->error)
906 exit_value++;
ebac4655 907 }
b990b5c0 908
4d2413c6 909 if (*nr_running == cputhreads && !pending)
6ce15a32 910 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
911}
912
906c8d75
JA
913/*
914 * Main function for kicking off and reaping jobs, as needed.
915 */
ebac4655
JA
916static void run_threads(void)
917{
ebac4655
JA
918 struct thread_data *td;
919 unsigned long spent;
920 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 921
2f9ade3c
JA
922 if (fio_pin_memory())
923 return;
ebac4655 924
c6ae0a5b
JA
925 if (!terse_output) {
926 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
927 fflush(stdout);
928 }
c04f7ec3 929
4efa970e
JA
930 signal(SIGINT, sig_handler);
931 signal(SIGALRM, sig_handler);
932
ebac4655
JA
933 todo = thread_number;
934 nr_running = 0;
935 nr_started = 0;
936 m_rate = t_rate = 0;
937
34572e28 938 for_each_td(td, i) {
263e529f 939 print_status_init(td->thread_number - 1);
ebac4655 940
380cf265
JA
941 if (!td->create_serialize) {
942 init_disk_util(td);
ebac4655 943 continue;
380cf265 944 }
ebac4655
JA
945
946 /*
947 * do file setup here so it happens sequentially,
948 * we don't want X number of threads getting their
949 * client data interspersed on disk
950 */
53cdc686 951 if (setup_files(td)) {
5bf13a5a
JA
952 exit_value++;
953 if (td->error)
954 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
ebac4655
JA
955 td_set_runstate(td, TD_REAPED);
956 todo--;
957 }
380cf265
JA
958
959 init_disk_util(td);
ebac4655
JA
960 }
961
a2f77c9f
JA
962 set_genesis_time();
963
ebac4655 964 while (todo) {
75154845
JA
965 struct thread_data *map[MAX_JOBS];
966 struct timeval this_start;
967 int this_jobs = 0, left;
968
ebac4655
JA
969 /*
970 * create threads (TD_NOT_CREATED -> TD_CREATED)
971 */
34572e28 972 for_each_td(td, i) {
ebac4655
JA
973 if (td->runstate != TD_NOT_CREATED)
974 continue;
975
976 /*
977 * never got a chance to start, killed by other
978 * thread for some reason
979 */
980 if (td->terminate) {
981 todo--;
982 continue;
983 }
984
985 if (td->start_delay) {
263e529f 986 spent = mtime_since_genesis();
ebac4655
JA
987
988 if (td->start_delay * 1000 > spent)
989 continue;
990 }
991
992 if (td->stonewall && (nr_started || nr_running))
993 break;
994
75154845
JA
995 /*
996 * Set state to created. Thread will transition
997 * to TD_INITIALIZED when it's done setting up.
998 */
ebac4655 999 td_set_runstate(td, TD_CREATED);
75154845 1000 map[this_jobs++] = td;
bbfd6b00 1001 fio_sem_init(&startup_sem, 1);
ebac4655
JA
1002 nr_started++;
1003
1004 if (td->use_thread) {
1005 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1006 perror("thread_create");
1007 nr_started--;
1008 }
1009 } else {
1010 if (fork())
bbfd6b00 1011 fio_sem_down(&startup_sem);
ebac4655 1012 else {
a6418147
JA
1013 int ret = fork_main(shm_id, i);
1014
1015 exit(ret);
ebac4655
JA
1016 }
1017 }
1018 }
1019
1020 /*
75154845
JA
1021 * Wait for the started threads to transition to
1022 * TD_INITIALIZED.
ebac4655 1023 */
02bcaa8c 1024 fio_gettime(&this_start, NULL);
75154845 1025 left = this_jobs;
6ce15a32 1026 while (left && !fio_abort) {
75154845
JA
1027 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1028 break;
1029
1030 usleep(100000);
1031
1032 for (i = 0; i < this_jobs; i++) {
1033 td = map[i];
1034 if (!td)
1035 continue;
b6f4d880 1036 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1037 map[i] = NULL;
1038 left--;
b6f4d880
JA
1039 } else if (td->runstate >= TD_EXITED) {
1040 map[i] = NULL;
1041 left--;
1042 todo--;
1043 nr_running++; /* work-around... */
75154845
JA
1044 }
1045 }
1046 }
1047
1048 if (left) {
3b70d7e5 1049 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1050 for (i = 0; i < this_jobs; i++) {
1051 td = map[i];
1052 if (!td)
1053 continue;
1054 kill(td->pid, SIGTERM);
1055 }
1056 break;
1057 }
1058
1059 /*
b6f4d880 1060 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1061 */
34572e28 1062 for_each_td(td, i) {
75154845 1063 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1064 continue;
1065
1066 td_set_runstate(td, TD_RUNNING);
1067 nr_running++;
1068 nr_started--;
1069 m_rate += td->ratemin;
1070 t_rate += td->rate;
75154845 1071 todo--;
bbfd6b00 1072 fio_sem_up(&td->mutex);
ebac4655
JA
1073 }
1074
1075 reap_threads(&nr_running, &t_rate, &m_rate);
1076
1077 if (todo)
1078 usleep(100000);
1079 }
1080
1081 while (nr_running) {
1082 reap_threads(&nr_running, &t_rate, &m_rate);
1083 usleep(10000);
1084 }
1085
1086 update_io_ticks();
2f9ade3c 1087 fio_unpin_memory();
ebac4655
JA
1088}
1089
ebac4655
JA
1090int main(int argc, char *argv[])
1091{
29d610e1
JA
1092 long ps;
1093
dbe1125e
JA
1094 /*
1095 * We need locale for number printing, if it isn't set then just
1096 * go with the US format.
1097 */
1098 if (!getenv("LC_NUMERIC"))
1099 setlocale(LC_NUMERIC, "en_US");
1100
ebac4655
JA
1101 if (parse_options(argc, argv))
1102 return 1;
1103
1104 if (!thread_number) {
3b70d7e5 1105 log_err("Nothing to do\n");
ebac4655
JA
1106 return 1;
1107 }
1108
29d610e1
JA
1109 ps = sysconf(_SC_PAGESIZE);
1110 if (ps < 0) {
1111 log_err("Failed to get page size\n");
1112 return 1;
1113 }
1114
1115 page_mask = ps - 1;
1116
bb3884d8
JA
1117 if (write_bw_log) {
1118 setup_log(&agg_io_log[DDIR_READ]);
1119 setup_log(&agg_io_log[DDIR_WRITE]);
1120 }
1121
a2f77c9f
JA
1122 set_genesis_time();
1123
ebac4655
JA
1124 disk_util_timer_arm();
1125
1126 run_threads();
6ce15a32 1127
bb3884d8 1128 if (!fio_abort) {
6ce15a32 1129 show_run_stats();
bb3884d8
JA
1130 if (write_bw_log) {
1131 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1132 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1133 }
1134 }
ebac4655 1135
437c9b71 1136 return exit_value;
ebac4655 1137}