sfree(NULL) is ok
[fio.git] / fio.c
CommitLineData
ebac4655
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
aae22ca7 5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
ebac4655 6 *
8e9fe637
JA
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
ebac4655 10 * This program is free software; you can redistribute it and/or modify
8e9fe637
JA
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
ebac4655
JA
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
ebac4655
JA
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
ebac4655
JA
27#include <signal.h>
28#include <time.h>
dbe1125e 29#include <locale.h>
36167d82 30#include <assert.h>
ebac4655
JA
31#include <sys/stat.h>
32#include <sys/wait.h>
33#include <sys/ipc.h>
34#include <sys/shm.h>
ebac4655
JA
35#include <sys/mman.h>
36
37#include "fio.h"
210f43b3 38#include "hash.h"
2e5cdb11 39#include "smalloc.h"
4f5af7b2 40#include "verify.h"
7c9b1bce 41#include "diskutil.h"
a696fa2a 42#include "cgroup.h"
07b3232d 43#include "profile.h"
d54a144d 44#include "lib/rand.h"
ebac4655 45
cfc99db7
JA
46unsigned long page_mask;
47unsigned long page_size;
d529ee19
JA
48
49#define PAGE_ALIGN(buf) \
29d610e1 50 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
51
52int groupid = 0;
53int thread_number = 0;
9cedf167
JA
54int nr_process = 0;
55int nr_thread = 0;
ebac4655 56int shm_id = 0;
53cdc686 57int temp_stall_ts;
d3eeeabc 58unsigned long done_secs = 0;
ebac4655 59
cdd18ad8 60static struct fio_mutex *startup_mutex;
d2bbb8d9 61static struct fio_mutex *writeout_mutex;
6ce15a32 62static volatile int fio_abort;
437c9b71 63static int exit_value;
ccb0fa24 64static struct itimerval itimer;
be4ecfdf 65static pthread_t gtod_thread;
dae5341c 66static struct flist_head *cgroup_list;
6adb38a1 67static char *cgroup_mnt;
ebac4655 68
bb3884d8
JA
69struct io_log *agg_io_log[2];
70
ebac4655 71#define TERMINATE_ALL (-1)
75154845 72#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 73
b29ee5b3 74void td_set_runstate(struct thread_data *td, int runstate)
6ce15a32 75{
d8fab1b6
JA
76 if (td->runstate == runstate)
77 return;
78
5921e80c
JA
79 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
80 td->runstate, runstate);
6ce15a32
JA
81 td->runstate = runstate;
82}
83
390c40e2 84static void terminate_threads(int group_id)
ebac4655 85{
34572e28 86 struct thread_data *td;
ebac4655
JA
87 int i;
88
6f88bcb0
JA
89 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
90
34572e28 91 for_each_td(td, i) {
ebac4655 92 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
6f88bcb0 93 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
5921e80c 94 td->o.name, (int) td->pid);
ad830ed7
JA
95 td->terminate = 1;
96 td->o.start_delay = 0;
97
7a93ab19
JA
98 /*
99 * if the thread is running, just let it exit
100 */
101 if (td->runstate < TD_RUNNING)
c1302d44 102 kill(td->pid, SIGQUIT);
f63f7f3b
JA
103 else {
104 struct ioengine_ops *ops = td->io_ops;
105
106 if (ops && (ops->flags & FIO_SIGQUIT))
107 kill(td->pid, SIGQUIT);
108 }
ebac4655
JA
109 }
110 }
111}
112
ccb0fa24
JA
113static void status_timer_arm(void)
114{
115 itimer.it_value.tv_sec = 0;
116 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
117 setitimer(ITIMER_REAL, &itimer, NULL);
118}
119
02444ad1 120static void sig_alrm(int fio_unused sig)
ebac4655 121{
697a606c 122 if (threads) {
5ec10eaa 123 update_io_ticks();
5ec10eaa 124 print_thread_status();
ccb0fa24 125 status_timer_arm();
697a606c
JA
126 }
127}
128
5f8f9726
JA
129/*
130 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
131 */
132static void sig_quit(int sig)
133{
134}
135
697a606c
JA
136static void sig_int(int sig)
137{
138 if (threads) {
4ceb30d4 139 log_info("\nfio: terminating on signal %d\n", sig);
5ec10eaa
JA
140 fflush(stdout);
141 terminate_threads(TERMINATE_ALL);
ebac4655
JA
142 }
143}
144
02444ad1 145static void sig_ill(int fio_unused sig)
de79c915
JA
146{
147 if (!threads)
148 return;
149
150 log_err("fio: illegal instruction. your cpu does not support "
151 "the sse4.2 instruction for crc32c\n");
152 terminate_threads(TERMINATE_ALL);
153 exit(4);
154}
155
697a606c
JA
156static void set_sig_handlers(void)
157{
158 struct sigaction act;
159
160 memset(&act, 0, sizeof(act));
161 act.sa_handler = sig_alrm;
162 act.sa_flags = SA_RESTART;
163 sigaction(SIGALRM, &act, NULL);
164
165 memset(&act, 0, sizeof(act));
166 act.sa_handler = sig_int;
167 act.sa_flags = SA_RESTART;
168 sigaction(SIGINT, &act, NULL);
de79c915
JA
169
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_ill;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGILL, &act, NULL);
5f8f9726
JA
174
175 memset(&act, 0, sizeof(act));
176 act.sa_handler = sig_quit;
177 act.sa_flags = SA_RESTART;
178 sigaction(SIGQUIT, &act, NULL);
697a606c
JA
179}
180
906c8d75
JA
181/*
182 * Check if we are above the minimum rate given.
183 */
581e7141
JA
184static int __check_min_rate(struct thread_data *td, struct timeval *now,
185 enum td_ddir ddir)
ebac4655 186{
0904200b 187 unsigned long long bytes = 0;
4e991c23 188 unsigned long iops = 0;
ebac4655
JA
189 unsigned long spent;
190 unsigned long rate;
581e7141
JA
191 unsigned int ratemin = 0;
192 unsigned int rate_iops = 0;
193 unsigned int rate_iops_min = 0;
ebac4655 194
d982d873
JA
195 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
196 return 0;
197
ebac4655
JA
198 /*
199 * allow a 2 second settle period in the beginning
200 */
201 if (mtime_since(&td->start, now) < 2000)
202 return 0;
203
581e7141
JA
204 iops += td->io_blocks[ddir];
205 bytes += td->this_io_bytes[ddir];
206 ratemin += td->o.ratemin[ddir];
207 rate_iops += td->o.rate_iops[ddir];
208 rate_iops_min += td->o.rate_iops_min[ddir];
0904200b 209
ebac4655
JA
210 /*
211 * if rate blocks is set, sample is running
212 */
581e7141
JA
213 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
214 spent = mtime_since(&td->lastrate[ddir], now);
2dc1bbeb 215 if (spent < td->o.ratecycle)
ebac4655
JA
216 return 0;
217
581e7141 218 if (td->o.rate[ddir]) {
4e991c23
JA
219 /*
220 * check bandwidth specified rate
221 */
581e7141 222 if (bytes < td->rate_bytes[ddir]) {
5ec10eaa 223 log_err("%s: min rate %u not met\n", td->o.name,
581e7141 224 ratemin);
4e991c23
JA
225 return 1;
226 } else {
581e7141
JA
227 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
228 if (rate < ratemin ||
229 bytes < td->rate_bytes[ddir]) {
5ec10eaa 230 log_err("%s: min rate %u not met, got"
b22989b9 231 " %luKB/sec\n", td->o.name,
581e7141 232 ratemin, rate);
4e991c23
JA
233 return 1;
234 }
235 }
413dd459 236 } else {
4e991c23
JA
237 /*
238 * checks iops specified rate
239 */
581e7141 240 if (iops < rate_iops) {
5ec10eaa 241 log_err("%s: min iops rate %u not met\n",
581e7141 242 td->o.name, rate_iops);
413dd459 243 return 1;
4e991c23 244 } else {
581e7141
JA
245 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
246 if (rate < rate_iops_min ||
247 iops < td->rate_blocks[ddir]) {
5ec10eaa
JA
248 log_err("%s: min iops rate %u not met,"
249 " got %lu\n", td->o.name,
581e7141 250 rate_iops_min, rate);
4e991c23 251 }
413dd459 252 }
ebac4655
JA
253 }
254 }
255
581e7141
JA
256 td->rate_bytes[ddir] = bytes;
257 td->rate_blocks[ddir] = iops;
258 memcpy(&td->lastrate[ddir], now, sizeof(*now));
ebac4655
JA
259 return 0;
260}
261
581e7141
JA
262static int check_min_rate(struct thread_data *td, struct timeval *now,
263 unsigned long *bytes_done)
264{
265 int ret = 0;
266
267 if (bytes_done[0])
268 ret |= __check_min_rate(td, now, 0);
269 if (bytes_done[1])
270 ret |= __check_min_rate(td, now, 1);
271
272 return ret;
273}
274
ebac4655
JA
275static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
276{
2dc1bbeb 277 if (!td->o.timeout)
ebac4655 278 return 0;
2dc1bbeb 279 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
280 return 1;
281
282 return 0;
283}
284
906c8d75
JA
285/*
286 * When job exits, we can cancel the in-flight IO if we are using async
287 * io. Attempt to do so.
288 */
ebac4655
JA
289static void cleanup_pending_aio(struct thread_data *td)
290{
01743ee1 291 struct flist_head *entry, *n;
ebac4655
JA
292 struct io_u *io_u;
293 int r;
294
295 /*
296 * get immediately available events, if any
297 */
581e7141 298 r = io_u_queued_complete(td, 0, NULL);
b2fdda43
JA
299 if (r < 0)
300 return;
ebac4655
JA
301
302 /*
303 * now cancel remaining active events
304 */
2866c82d 305 if (td->io_ops->cancel) {
01743ee1
JA
306 flist_for_each_safe(entry, n, &td->io_u_busylist) {
307 io_u = flist_entry(entry, struct io_u, list);
ebac4655 308
0c6e7517
JA
309 /*
310 * if the io_u isn't in flight, then that generally
311 * means someone leaked an io_u. complain but fix
312 * it up, so we don't stall here.
313 */
314 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
315 log_err("fio: non-busy IO on busy list\n");
ebac4655 316 put_io_u(td, io_u);
0c6e7517
JA
317 } else {
318 r = td->io_ops->cancel(td, io_u);
319 if (!r)
320 put_io_u(td, io_u);
321 }
ebac4655
JA
322 }
323 }
324
97601024 325 if (td->cur_depth)
581e7141 326 r = io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
327}
328
858a3d47
JA
329/*
330 * Helper to handle the final sync of a file. Works just like the normal
331 * io path, just does everything sync.
332 */
333static int fio_io_sync(struct thread_data *td, struct fio_file *f)
334{
335 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
336 int ret;
337
338 if (!io_u)
339 return 1;
340
341 io_u->ddir = DDIR_SYNC;
342 io_u->file = f;
343
344 if (td_io_prep(td, io_u)) {
345 put_io_u(td, io_u);
346 return 1;
347 }
348
755200a3 349requeue:
858a3d47 350 ret = td_io_queue(td, io_u);
36167d82 351 if (ret < 0) {
e1161c32 352 td_verror(td, io_u->error, "td_io_queue");
858a3d47 353 put_io_u(td, io_u);
858a3d47 354 return 1;
36167d82 355 } else if (ret == FIO_Q_QUEUED) {
581e7141 356 if (io_u_queued_complete(td, 1, NULL) < 0)
36167d82 357 return 1;
36167d82
JA
358 } else if (ret == FIO_Q_COMPLETED) {
359 if (io_u->error) {
e1161c32 360 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
361 return 1;
362 }
858a3d47 363
581e7141 364 if (io_u_sync_complete(td, io_u, NULL) < 0)
b2fdda43 365 return 1;
755200a3
JA
366 } else if (ret == FIO_Q_BUSY) {
367 if (td_io_commit(td))
368 return 1;
369 goto requeue;
858a3d47
JA
370 }
371
372 return 0;
373}
374
993bf48b
JA
375static inline void update_tv_cache(struct thread_data *td)
376{
377 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
378 fio_gettime(&td->tv_cache, NULL);
379}
380
f2bba182
RR
381static int break_on_this_error(struct thread_data *td, int *retptr)
382{
383 int ret = *retptr;
384
385 if (ret < 0 || td->error) {
386 int err;
387
1ec99eea
JA
388 if (!td->o.continue_on_error)
389 return 1;
f2bba182
RR
390
391 if (ret < 0)
392 err = -ret;
393 else
394 err = td->error;
395
f2bba182
RR
396 if (td_non_fatal_error(err)) {
397 /*
398 * Continue with the I/Os in case of
399 * a non fatal error.
400 */
8a4f8a24 401 update_error_count(td, err);
f2bba182
RR
402 td_clear_error(td);
403 *retptr = 0;
404 return 0;
8a4f8a24
RR
405 } else if (td->o.fill_device && err == ENOSPC) {
406 /*
407 * We expect to hit this error if
408 * fill_device option is set.
409 */
410 td_clear_error(td);
411 td->terminate = 1;
412 return 1;
f2bba182
RR
413 } else {
414 /*
415 * Stop the I/O in case of a fatal
416 * error.
417 */
8a4f8a24 418 update_error_count(td, err);
f2bba182
RR
419 return 1;
420 }
421 }
422
423 return 0;
424}
425
906c8d75 426/*
34403fb1 427 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
428 * reads the blocks back in, and checks the crc/md5 of the data.
429 */
1e97cce9 430static void do_verify(struct thread_data *td)
ebac4655 431{
53cdc686 432 struct fio_file *f;
36167d82 433 struct io_u *io_u;
af52b345
JA
434 int ret, min_events;
435 unsigned int i;
e5b401d4 436
008618a8
JA
437 dprint(FD_VERIFY, "starting loop\n");
438
e5b401d4
JA
439 /*
440 * sync io first and invalidate cache, to make sure we really
441 * read from disk.
442 */
443 for_each_file(td, f, i) {
d6aed795 444 if (!fio_file_open(f))
3d7b485f 445 continue;
b2fdda43
JA
446 if (fio_io_sync(td, f))
447 break;
448 if (file_invalidate_cache(td, f))
449 break;
e5b401d4 450 }
ebac4655 451
b2fdda43
JA
452 if (td->error)
453 return;
454
ebac4655
JA
455 td_set_runstate(td, TD_VERIFYING);
456
36167d82
JA
457 io_u = NULL;
458 while (!td->terminate) {
4950421a 459 int ret2, full;
5451792e 460
993bf48b
JA
461 update_tv_cache(td);
462
463 if (runtime_exceeded(td, &td->tv_cache)) {
90a87d4b 464 td->terminate = 1;
02bcaa8c 465 break;
069c2918 466 }
02bcaa8c 467
dcb4830d
RR
468 io_u = __get_io_u(td);
469 if (!io_u)
470 break;
471
069c2918
JA
472 if (get_next_verify(td, io_u)) {
473 put_io_u(td, io_u);
ebac4655 474 break;
069c2918 475 }
ebac4655 476
069c2918
JA
477 if (td_io_prep(td, io_u)) {
478 put_io_u(td, io_u);
53cdc686 479 break;
069c2918 480 }
d7762cf8 481
e8462bd8
JA
482 if (td->o.verify_async)
483 io_u->end_io = verify_io_u_async;
484 else
485 io_u->end_io = verify_io_u;
53cdc686 486
11786802 487 ret = td_io_queue(td, io_u);
36167d82
JA
488 switch (ret) {
489 case FIO_Q_COMPLETED:
f2bba182 490 if (io_u->error) {
22819ec2 491 ret = -io_u->error;
f2bba182
RR
492 clear_io_u(td, io_u);
493 } else if (io_u->resid) {
36167d82 494 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 495 struct fio_file *f = io_u->file;
ebac4655 496
9e9d164e
JA
497 /*
498 * zero read, fail
499 */
500 if (!bytes) {
41d9f0fd 501 td_verror(td, EIO, "full resid");
9e9d164e
JA
502 put_io_u(td, io_u);
503 break;
504 }
8400d9b2 505
36167d82
JA
506 io_u->xfer_buflen = io_u->resid;
507 io_u->xfer_buf += bytes;
8400d9b2
JA
508 io_u->offset += bytes;
509
30061b97
JA
510 td->ts.short_io_u[io_u->ddir]++;
511
d460eb31 512 if (io_u->offset == f->real_file_size)
8400d9b2
JA
513 goto sync_done;
514
11786802
JA
515 requeue_io_u(td, &io_u);
516 } else {
8400d9b2 517sync_done:
581e7141 518 ret = io_u_sync_complete(td, io_u, NULL);
11786802
JA
519 if (ret < 0)
520 break;
36167d82 521 }
36167d82
JA
522 continue;
523 case FIO_Q_QUEUED:
524 break;
755200a3
JA
525 case FIO_Q_BUSY:
526 requeue_io_u(td, &io_u);
5451792e
JA
527 ret2 = td_io_commit(td);
528 if (ret2 < 0)
529 ret = ret2;
755200a3 530 break;
36167d82
JA
531 default:
532 assert(ret < 0);
e1161c32 533 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
534 break;
535 }
536
f2bba182 537 if (break_on_this_error(td, &ret))
3af6ef39
JA
538 break;
539
ebac4655 540 /*
3af6ef39
JA
541 * if we can queue more, do so. but check if there are
542 * completed io_u's first.
ebac4655 543 */
4950421a
JA
544 full = queue_full(td) || ret == FIO_Q_BUSY;
545 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
546 min_events = min(td->o.iodepth_batch_complete,
547 td->cur_depth);
4950421a 548 if (full && !min_events)
838bc709 549 min_events = 1;
e916b390 550
4950421a
JA
551 do {
552 /*
553 * Reap required number of io units, if any,
554 * and do the verification on them through
555 * the callback handler
556 */
581e7141 557 if (io_u_queued_complete(td, min_events, NULL) < 0) {
4950421a
JA
558 ret = -1;
559 break;
560 }
561 } while (full && (td->cur_depth > td->o.iodepth_low));
562 }
563 if (ret < 0)
ebac4655 564 break;
36167d82 565 }
ebac4655 566
c01c0395
JA
567 if (!td->error) {
568 min_events = td->cur_depth;
569
570 if (min_events)
581e7141 571 ret = io_u_queued_complete(td, min_events, NULL);
c01c0395 572 } else
ebac4655
JA
573 cleanup_pending_aio(td);
574
575 td_set_runstate(td, TD_RUNNING);
008618a8
JA
576
577 dprint(FD_VERIFY, "exiting loop\n");
ebac4655
JA
578}
579
32cd46a0 580/*
906c8d75 581 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
582 * and reaps them, checking for rate and errors along the way.
583 */
ebac4655
JA
584static void do_io(struct thread_data *td)
585{
af52b345
JA
586 unsigned int i;
587 int ret = 0;
ebac4655 588
b29ee5b3
JA
589 if (in_ramp_time(td))
590 td_set_runstate(td, TD_RAMP);
591 else
592 td_set_runstate(td, TD_RUNNING);
5853e5a8 593
457bf399
GO
594 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
595 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
97601024 596 struct timeval comp_time;
581e7141 597 unsigned long bytes_done[2] = { 0, 0 };
84585003 598 int min_evts = 0;
ebac4655 599 struct io_u *io_u;
4950421a 600 int ret2, full;
ebac4655
JA
601
602 if (td->terminate)
603 break;
604
993bf48b 605 update_tv_cache(td);
97601024 606
993bf48b 607 if (runtime_exceeded(td, &td->tv_cache)) {
90a87d4b 608 td->terminate = 1;
97601024
JA
609 break;
610 }
36167d82 611
dcb4830d
RR
612 io_u = get_io_u(td);
613 if (!io_u)
614 break;
615
b67740d3
JA
616 /*
617 * Add verification end_io handler, if asked to verify
618 * a previously written file.
619 */
715d2b3e
JA
620 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
621 !td_rw(td)) {
e8462bd8
JA
622 if (td->o.verify_async)
623 io_u->end_io = verify_io_u_async;
624 else
625 io_u->end_io = verify_io_u;
d8fab1b6 626 td_set_runstate(td, TD_VERIFYING);
b29ee5b3
JA
627 } else if (in_ramp_time(td))
628 td_set_runstate(td, TD_RAMP);
629 else
d8fab1b6 630 td_set_runstate(td, TD_RUNNING);
b67740d3 631
11786802 632 ret = td_io_queue(td, io_u);
36167d82
JA
633 switch (ret) {
634 case FIO_Q_COMPLETED:
f2bba182 635 if (io_u->error) {
5451792e 636 ret = -io_u->error;
f2bba182
RR
637 clear_io_u(td, io_u);
638 } else if (io_u->resid) {
36167d82 639 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 640 struct fio_file *f = io_u->file;
36167d82 641
9e9d164e
JA
642 /*
643 * zero read, fail
644 */
645 if (!bytes) {
41d9f0fd 646 td_verror(td, EIO, "full resid");
9e9d164e
JA
647 put_io_u(td, io_u);
648 break;
649 }
650
cec6b55d 651 io_u->xfer_buflen = io_u->resid;
36167d82 652 io_u->xfer_buf += bytes;
5a7c5680
JA
653 io_u->offset += bytes;
654
30061b97
JA
655 td->ts.short_io_u[io_u->ddir]++;
656
d460eb31 657 if (io_u->offset == f->real_file_size)
5a7c5680
JA
658 goto sync_done;
659
11786802
JA
660 requeue_io_u(td, &io_u);
661 } else {
5a7c5680 662sync_done:
581e7141
JA
663 if (__should_check_rate(td, 0) ||
664 __should_check_rate(td, 1))
993bf48b
JA
665 fio_gettime(&comp_time, NULL);
666
581e7141
JA
667 ret = io_u_sync_complete(td, io_u, bytes_done);
668 if (ret < 0)
669 break;
cec6b55d 670 }
36167d82
JA
671 break;
672 case FIO_Q_QUEUED:
7e77dd02
JA
673 /*
674 * if the engine doesn't have a commit hook,
675 * the io_u is really queued. if it does have such
676 * a hook, it has to call io_u_queued() itself.
677 */
678 if (td->io_ops->commit == NULL)
679 io_u_queued(td, io_u);
36167d82 680 break;
755200a3
JA
681 case FIO_Q_BUSY:
682 requeue_io_u(td, &io_u);
5451792e
JA
683 ret2 = td_io_commit(td);
684 if (ret2 < 0)
685 ret = ret2;
755200a3 686 break;
36167d82
JA
687 default:
688 assert(ret < 0);
689 put_io_u(td, io_u);
690 break;
ebac4655
JA
691 }
692
f2bba182 693 if (break_on_this_error(td, &ret))
36167d82
JA
694 break;
695
97601024
JA
696 /*
697 * See if we need to complete some commands
698 */
4950421a
JA
699 full = queue_full(td) || ret == FIO_Q_BUSY;
700 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
701 min_evts = min(td->o.iodepth_batch_complete,
702 td->cur_depth);
4950421a 703 if (full && !min_evts)
36167d82 704 min_evts = 1;
4950421a 705
581e7141
JA
706 if (__should_check_rate(td, 0) ||
707 __should_check_rate(td, 1))
993bf48b 708 fio_gettime(&comp_time, NULL);
4950421a
JA
709
710 do {
581e7141
JA
711 ret = io_u_queued_complete(td, min_evts, bytes_done);
712 if (ret < 0)
4950421a
JA
713 break;
714
4950421a 715 } while (full && (td->cur_depth > td->o.iodepth_low));
ebac4655
JA
716 }
717
4950421a
JA
718 if (ret < 0)
719 break;
581e7141 720 if (!(bytes_done[0] + bytes_done[1]))
97601024
JA
721 continue;
722
581e7141
JA
723 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
724 if (check_min_rate(td, &comp_time, bytes_done)) {
721938ae
JA
725 if (exitall_on_terminate)
726 terminate_threads(td->groupid);
727 td_verror(td, EIO, "check_min_rate");
728 break;
729 }
ebac4655
JA
730 }
731
2dc1bbeb 732 if (td->o.thinktime) {
9c1f7434
JA
733 unsigned long long b;
734
735 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 736 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
737 int left;
738
2dc1bbeb 739 if (td->o.thinktime_spin)
6dd6f2cd 740 usec_spin(td->o.thinktime_spin);
48097d5c 741
2dc1bbeb 742 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
743 if (left)
744 usec_sleep(td, left);
745 }
9c1f7434 746 }
ebac4655
JA
747 }
748
aa31f1f1
SL
749 if (td->o.fill_device && td->error == ENOSPC) {
750 td->error = 0;
751 td->terminate = 1;
752 }
4d2413c6 753 if (!td->error) {
3d7c391d
JA
754 struct fio_file *f;
755
c01c0395
JA
756 i = td->cur_depth;
757 if (i)
581e7141 758 ret = io_u_queued_complete(td, i, NULL);
ebac4655 759
2dc1bbeb 760 if (should_fsync(td) && td->o.end_fsync) {
84585003 761 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
762
763 for_each_file(td, f, i) {
d6aed795 764 if (!fio_file_open(f))
3d7b485f 765 continue;
858a3d47 766 fio_io_sync(td, f);
3d7b485f 767 }
84585003 768 }
c01c0395
JA
769 } else
770 cleanup_pending_aio(td);
2ba1c290
JA
771
772 /*
773 * stop job if we failed doing any IO
774 */
775 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
776 td->done = 1;
ebac4655
JA
777}
778
ebac4655
JA
779static void cleanup_io_u(struct thread_data *td)
780{
01743ee1 781 struct flist_head *entry, *n;
ebac4655
JA
782 struct io_u *io_u;
783
01743ee1
JA
784 flist_for_each_safe(entry, n, &td->io_u_freelist) {
785 io_u = flist_entry(entry, struct io_u, list);
ebac4655 786
01743ee1 787 flist_del(&io_u->list);
ebac4655
JA
788 free(io_u);
789 }
790
2f9ade3c 791 free_io_mem(td);
ebac4655
JA
792}
793
794static int init_io_u(struct thread_data *td)
795{
796 struct io_u *io_u;
a00735e6 797 unsigned int max_bs;
eb7ccf38 798 int cl_align, i, max_units;
ebac4655
JA
799 char *p;
800
1e3c09af 801 max_units = td->o.iodepth;
2dc1bbeb 802 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
5ec10eaa
JA
803 td->orig_buffer_size = (unsigned long long) max_bs
804 * (unsigned long long) max_units;
805
806 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
807 unsigned long bs;
74b025b0 808
5ec10eaa
JA
809 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
810 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
811 }
ebac4655 812
c3990645
JA
813 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
814 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
815 return 1;
816 }
817
2f9ade3c
JA
818 if (allocate_io_mem(td))
819 return 1;
ebac4655 820
fc966bef 821 if (td->o.odirect || td->o.mem_align)
d529ee19 822 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
b3f378e9
JA
823 else
824 p = td->orig_buffer;
825
eb7ccf38
JA
826 cl_align = os_cache_line_size();
827
ebac4655 828 for (i = 0; i < max_units; i++) {
eb7ccf38
JA
829 void *ptr;
830
db1cd90b
JA
831 if (td->terminate)
832 return 1;
eb7ccf38
JA
833
834 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
835 log_err("fio: posix_memalign=%s\n", strerror(errno));
836 break;
837 }
838
839 io_u = ptr;
ebac4655 840 memset(io_u, 0, sizeof(*io_u));
01743ee1 841 INIT_FLIST_HEAD(&io_u->list);
d529ee19 842 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
ebac4655 843
b4c5e1ac
JA
844 if (!(td->io_ops->flags & FIO_NOIO)) {
845 io_u->buf = p + max_bs * i;
d529ee19 846 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
e9459e5a 847
5973cafb
JA
848 if (td_write(td) && !td->o.refill_buffers)
849 io_u_fill_buffer(td, io_u, max_bs);
b4c5e1ac 850 }
6b9cea23 851
b1ff3403 852 io_u->index = i;
0c6e7517 853 io_u->flags = IO_U_F_FREE;
01743ee1 854 flist_add(&io_u->list, &td->io_u_freelist);
ebac4655
JA
855 }
856
857 return 0;
858}
859
da86774e
JA
860static int switch_ioscheduler(struct thread_data *td)
861{
862 char tmp[256], tmp2[128];
863 FILE *f;
864 int ret;
865
ba0fbe10 866 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
867 return 0;
868
da86774e
JA
869 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
870
871 f = fopen(tmp, "r+");
872 if (!f) {
cbf5c1d7 873 if (errno == ENOENT) {
5ec10eaa
JA
874 log_err("fio: os or kernel doesn't support IO scheduler"
875 " switching\n");
cbf5c1d7
JA
876 return 0;
877 }
878 td_verror(td, errno, "fopen iosched");
da86774e
JA
879 return 1;
880 }
881
882 /*
883 * Set io scheduler.
884 */
2dc1bbeb 885 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 886 if (ferror(f) || ret != 1) {
e1161c32 887 td_verror(td, errno, "fwrite");
da86774e
JA
888 fclose(f);
889 return 1;
890 }
891
892 rewind(f);
893
894 /*
895 * Read back and check that the selected scheduler is now the default.
896 */
897 ret = fread(tmp, 1, sizeof(tmp), f);
898 if (ferror(f) || ret < 0) {
e1161c32 899 td_verror(td, errno, "fread");
da86774e
JA
900 fclose(f);
901 return 1;
902 }
903
2dc1bbeb 904 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 905 if (!strstr(tmp, tmp2)) {
2dc1bbeb 906 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 907 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
908 fclose(f);
909 return 1;
910 }
911
912 fclose(f);
913 return 0;
914}
915
492158cf
JA
916static int keep_running(struct thread_data *td)
917{
918 unsigned long long io_done;
919
20e354ef
JA
920 if (td->done)
921 return 0;
492158cf
JA
922 if (td->o.time_based)
923 return 1;
924 if (td->o.loops) {
925 td->o.loops--;
926 return 1;
927 }
928
5ec10eaa
JA
929 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
930 + td->io_skip_bytes;
492158cf
JA
931 if (io_done < td->o.size)
932 return 1;
933
934 return 0;
935}
936
b29ee5b3 937static void reset_io_counters(struct thread_data *td)
ebac4655 938{
756867bd 939 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 940 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 941 td->zone_bytes = 0;
581e7141
JA
942 td->rate_bytes[0] = td->rate_bytes[1] = 0;
943 td->rate_blocks[0] = td->rate_blocks[1] = 0;
ebac4655 944
c1324df1
JA
945 td->last_was_sync = 0;
946
2ba1c290
JA
947 /*
948 * reset file done count if we are to start over
949 */
950 if (td->o.time_based || td->o.loops)
48f5abd3 951 td->nr_done_files = 0;
5bfc35d7
JA
952
953 /*
954 * Set the same seed to get repeatable runs
955 */
956 td_fill_rand_seeds(td);
b29ee5b3
JA
957}
958
959void reset_all_stats(struct thread_data *td)
960{
961 struct timeval tv;
962 int i;
963
964 reset_io_counters(td);
965
966 for (i = 0; i < 2; i++) {
967 td->io_bytes[i] = 0;
968 td->io_blocks[i] = 0;
969 td->io_issues[i] = 0;
970 td->ts.total_io_u[i] = 0;
971 }
972
973 fio_gettime(&tv, NULL);
974 memcpy(&td->epoch, &tv, sizeof(tv));
975 memcpy(&td->start, &tv, sizeof(tv));
976}
977
df9c26b1 978static void clear_io_state(struct thread_data *td)
b29ee5b3
JA
979{
980 struct fio_file *f;
981 unsigned int i;
b29ee5b3
JA
982
983 reset_io_counters(td);
281f9b63 984
24ffd2c2 985 close_files(td);
df9c26b1 986 for_each_file(td, f, i)
d6aed795 987 fio_file_clear_done(f);
ebac4655
JA
988}
989
398e8c4d
JA
990static int exec_string(const char *string)
991{
992 int ret, newlen = strlen(string) + 1 + 8;
993 char *str;
994
995 str = malloc(newlen);
996 sprintf(str, "sh -c %s", string);
997
998 ret = system(str);
999 if (ret == -1)
1000 log_err("fio: exec of cmd <%s> failed\n", str);
1001
1002 free(str);
1003 return ret;
1004}
1005
906c8d75
JA
1006/*
1007 * Entry point for the thread based jobs. The process based jobs end up
1008 * here as well, after a little setup.
1009 */
ebac4655
JA
1010static void *thread_main(void *data)
1011{
10927316 1012 unsigned long long runtime[2], elapsed;
ebac4655 1013 struct thread_data *td = data;
e8462bd8 1014 pthread_condattr_t attr;
a978ba68 1015 int clear_state;
ebac4655 1016
2dc1bbeb 1017 if (!td->o.use_thread)
ebac4655
JA
1018 setsid();
1019
1020 td->pid = getpid();
1021
5921e80c 1022 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
ee56ad50 1023
01743ee1
JA
1024 INIT_FLIST_HEAD(&td->io_u_freelist);
1025 INIT_FLIST_HEAD(&td->io_u_busylist);
1026 INIT_FLIST_HEAD(&td->io_u_requeues);
1027 INIT_FLIST_HEAD(&td->io_log_list);
1028 INIT_FLIST_HEAD(&td->io_hist_list);
e8462bd8
JA
1029 INIT_FLIST_HEAD(&td->verify_list);
1030 pthread_mutex_init(&td->io_u_lock, NULL);
bb5d7d0b 1031 td->io_hist_tree = RB_ROOT;
aea47d44 1032
e8462bd8
JA
1033 pthread_condattr_init(&attr);
1034 pthread_cond_init(&td->verify_cond, &attr);
1035 pthread_cond_init(&td->free_cond, &attr);
1036
db1cd90b 1037 td_set_runstate(td, TD_INITIALIZED);
29adda3c 1038 dprint(FD_MUTEX, "up startup_mutex\n");
cdd18ad8 1039 fio_mutex_up(startup_mutex);
29adda3c 1040 dprint(FD_MUTEX, "wait on td->mutex\n");
cdd18ad8 1041 fio_mutex_down(td->mutex);
29adda3c 1042 dprint(FD_MUTEX, "done waiting on td->mutex\n");
db1cd90b
JA
1043
1044 /*
cdd18ad8 1045 * the ->mutex mutex is now no longer used, close it to avoid
db1cd90b
JA
1046 * eating a file descriptor
1047 */
cdd18ad8 1048 fio_mutex_remove(td->mutex);
db1cd90b 1049
e0b0d892
JA
1050 if (td->o.uid != -1U && setuid(td->o.uid)) {
1051 td_verror(td, errno, "setuid");
1052 goto err;
1053 }
1054 if (td->o.gid != -1U && setgid(td->o.gid)) {
1055 td_verror(td, errno, "setgid");
1056 goto err;
1057 }
1058
d84f8d49
JA
1059 /*
1060 * May alter parameters that init_io_u() will use, so we need to
1061 * do this first.
1062 */
1063 if (init_iolog(td))
1064 goto err;
1065
ebac4655 1066 if (init_io_u(td))
db1cd90b 1067 goto err;
ebac4655 1068
e8462bd8
JA
1069 if (td->o.verify_async && verify_async_init(td))
1070 goto err;
1071
1072 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
e1161c32 1073 td_verror(td, errno, "cpu_set_affinity");
db1cd90b 1074 goto err;
ebac4655
JA
1075 }
1076
f484470b
JA
1077 /*
1078 * If we have a gettimeofday() thread, make sure we exclude that
1079 * thread from this job
1080 */
be4ecfdf 1081 if (td->o.gtod_cpu) {
be4ecfdf 1082 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
e8462bd8 1083 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
be4ecfdf
JA
1084 td_verror(td, errno, "cpu_set_affinity");
1085 goto err;
1086 }
1087 }
1088
ac684785 1089 if (td->ioprio_set) {
ebac4655 1090 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 1091 td_verror(td, errno, "ioprio_set");
db1cd90b 1092 goto err;
ebac4655
JA
1093 }
1094 }
1095
6adb38a1 1096 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
a696fa2a
JA
1097 goto err;
1098
2dc1bbeb 1099 if (nice(td->o.nice) == -1) {
e1161c32 1100 td_verror(td, errno, "nice");
db1cd90b 1101 goto err;
b6f4d880
JA
1102 }
1103
2dc1bbeb 1104 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 1105 goto err;
37f56873 1106
2dc1bbeb 1107 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
1108 goto err;
1109
7d6c5283
JA
1110 if (td_io_init(td))
1111 goto err;
1112
68727076
JA
1113 if (init_random_map(td))
1114 goto err;
1115
2dc1bbeb 1116 if (td->o.exec_prerun) {
398e8c4d 1117 if (exec_string(td->o.exec_prerun))
69cfd7e0
JA
1118 goto err;
1119 }
4e0ba8af 1120
afad68f7
ZY
1121 if (td->o.pre_read) {
1122 if (pre_read_files(td) < 0)
1123 goto err;
1124 }
1125
69008999 1126 fio_gettime(&td->epoch, NULL);
756867bd 1127 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999
JA
1128
1129 runtime[0] = runtime[1] = 0;
a978ba68 1130 clear_state = 0;
492158cf 1131 while (keep_running(td)) {
02bcaa8c 1132 fio_gettime(&td->start, NULL);
c5d3d790
JA
1133 memcpy(&td->ts.stat_sample_time[0], &td->start,
1134 sizeof(td->start));
1135 memcpy(&td->ts.stat_sample_time[1], &td->start,
1136 sizeof(td->start));
993bf48b 1137 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
ebac4655 1138
cf2dd8d6 1139 if (td->o.ratemin[0] || td->o.ratemin[1])
5ec10eaa
JA
1140 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1141 sizeof(td->lastrate));
ebac4655 1142
df9c26b1
JA
1143 if (clear_state)
1144 clear_io_state(td);
a978ba68 1145
ebac4655
JA
1146 prune_io_piece_log(td);
1147
ba0fbe10 1148 do_io(td);
ebac4655 1149
a978ba68
JA
1150 clear_state = 1;
1151
38d77cae 1152 if (td_read(td) && td->io_bytes[DDIR_READ]) {
0a4957c6 1153 elapsed = utime_since_now(&td->start);
38d77cae
JA
1154 runtime[DDIR_READ] += elapsed;
1155 }
1156 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
0a4957c6 1157 elapsed = utime_since_now(&td->start);
38d77cae
JA
1158 runtime[DDIR_WRITE] += elapsed;
1159 }
5ec10eaa 1160
ebac4655
JA
1161 if (td->error || td->terminate)
1162 break;
1163
e84c73a8
SL
1164 if (!td->o.do_verify ||
1165 td->o.verify == VERIFY_NONE ||
b67740d3 1166 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
1167 continue;
1168
df9c26b1 1169 clear_io_state(td);
a978ba68 1170
02bcaa8c 1171 fio_gettime(&td->start, NULL);
ebac4655
JA
1172
1173 do_verify(td);
1174
69008999 1175 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
1176
1177 if (td->error || td->terminate)
1178 break;
1179 }
1180
36dff966 1181 update_rusage_stat(td);
47f0cc48
YHJT
1182 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1183 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
756867bd
JA
1184 td->ts.total_run_time = mtime_since_now(&td->epoch);
1185 td->ts.io_bytes[0] = td->io_bytes[0];
1186 td->ts.io_bytes[1] = td->io_bytes[1];
1187
d2bbb8d9 1188 fio_mutex_down(writeout_mutex);
e3cedca7 1189 if (td->ts.bw_log) {
f484470b
JA
1190 if (td->o.bw_log_file) {
1191 finish_log_named(td, td->ts.bw_log,
1192 td->o.bw_log_file, "bw");
1193 } else
e3cedca7
JA
1194 finish_log(td, td->ts.bw_log, "bw");
1195 }
1196 if (td->ts.slat_log) {
f484470b
JA
1197 if (td->o.lat_log_file) {
1198 finish_log_named(td, td->ts.slat_log,
1ca95738 1199 td->o.lat_log_file, "slat");
f484470b 1200 } else
e3cedca7
JA
1201 finish_log(td, td->ts.slat_log, "slat");
1202 }
1203 if (td->ts.clat_log) {
f484470b
JA
1204 if (td->o.lat_log_file) {
1205 finish_log_named(td, td->ts.clat_log,
1206 td->o.lat_log_file, "clat");
1207 } else
e3cedca7
JA
1208 finish_log(td, td->ts.clat_log, "clat");
1209 }
d2bbb8d9 1210 fio_mutex_up(writeout_mutex);
398e8c4d
JA
1211 if (td->o.exec_postrun)
1212 exec_string(td->o.exec_postrun);
ebac4655
JA
1213
1214 if (exitall_on_terminate)
390c40e2 1215 terminate_threads(td->groupid);
ebac4655
JA
1216
1217err:
5bf13a5a 1218 if (td->error)
4ceb30d4 1219 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
5ec10eaa 1220 td->verror);
b4707ed7
JA
1221
1222 if (td->o.verify_async)
1223 verify_async_exit(td);
1224
24ffd2c2 1225 close_and_free_files(td);
2866c82d 1226 close_ioengine(td);
ebac4655 1227 cleanup_io_u(td);
6adb38a1 1228 cgroup_shutdown(td, &cgroup_mnt);
f29b25a3 1229
d2ce18b5
JA
1230 if (td->o.cpumask_set) {
1231 int ret = fio_cpuset_exit(&td->o.cpumask);
1232
1233 td_verror(td, ret, "fio_cpuset_exit");
1234 }
4e78e405 1235
f29b25a3
JA
1236 /*
1237 * do this very late, it will log file closing as well
1238 */
1239 if (td->o.write_iolog_file)
1240 write_iolog_close(td);
1241
d23bb327 1242 options_mem_free(td);
ebac4655 1243 td_set_runstate(td, TD_EXITED);
43d76807 1244 return (void *) (unsigned long) td->error;
ebac4655
JA
1245}
1246
906c8d75
JA
1247/*
1248 * We cannot pass the td data into a forked process, so attach the td and
1249 * pass it to the thread worker.
1250 */
a6418147 1251static int fork_main(int shmid, int offset)
ebac4655
JA
1252{
1253 struct thread_data *td;
a6418147 1254 void *data, *ret;
ebac4655
JA
1255
1256 data = shmat(shmid, NULL, 0);
1257 if (data == (void *) -1) {
a6418147
JA
1258 int __err = errno;
1259
ebac4655 1260 perror("shmat");
a6418147 1261 return __err;
ebac4655
JA
1262 }
1263
1264 td = data + offset * sizeof(struct thread_data);
a6418147 1265 ret = thread_main(td);
ebac4655 1266 shmdt(data);
43d76807 1267 return (int) (unsigned long) ret;
ebac4655
JA
1268}
1269
906c8d75
JA
1270/*
1271 * Run over the job map and reap the threads that have exited, if any.
1272 */
ebac4655
JA
1273static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1274{
34572e28 1275 struct thread_data *td;
1f809d15 1276 int i, cputhreads, realthreads, pending, status, ret;
ebac4655
JA
1277
1278 /*
1279 * reap exited threads (TD_EXITED -> TD_REAPED)
1280 */
1f809d15 1281 realthreads = pending = cputhreads = 0;
34572e28 1282 for_each_td(td, i) {
3707f45b 1283 int flags = 0;
a2f77c9f 1284
84585003
JA
1285 /*
1286 * ->io_ops is NULL for a thread that has closed its
1287 * io engine
1288 */
ba0fbe10 1289 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0 1290 cputhreads++;
1f809d15
JA
1291 else
1292 realthreads++;
b990b5c0 1293
a1dedc1f
AC
1294 if (!td->pid) {
1295 pending++;
1296 continue;
1297 }
1298 if (td->runstate == TD_REAPED)
a2f77c9f 1299 continue;
2dc1bbeb 1300 if (td->o.use_thread) {
3707f45b
JA
1301 if (td->runstate == TD_EXITED) {
1302 td_set_runstate(td, TD_REAPED);
1303 goto reaped;
1304 }
1305 continue;
1306 }
a2f77c9f
JA
1307
1308 flags = WNOHANG;
1309 if (td->runstate == TD_EXITED)
1310 flags = 0;
1311
1312 /*
1313 * check if someone quit or got killed in an unusual way
1314 */
1315 ret = waitpid(td->pid, &status, flags);
3707f45b 1316 if (ret < 0) {
a2f77c9f 1317 if (errno == ECHILD) {
5921e80c
JA
1318 log_err("fio: pid=%d disappeared %d\n",
1319 (int) td->pid, td->runstate);
a2f77c9f
JA
1320 td_set_runstate(td, TD_REAPED);
1321 goto reaped;
1322 }
1323 perror("waitpid");
1324 } else if (ret == td->pid) {
1325 if (WIFSIGNALED(status)) {
fab6aa71
JA
1326 int sig = WTERMSIG(status);
1327
d9244941 1328 if (sig != SIGQUIT)
5ec10eaa 1329 log_err("fio: pid=%d, got signal=%d\n",
5921e80c 1330 (int) td->pid, sig);
fab6aa71
JA
1331 td_set_runstate(td, TD_REAPED);
1332 goto reaped;
1333 }
a2f77c9f
JA
1334 if (WIFEXITED(status)) {
1335 if (WEXITSTATUS(status) && !td->error)
1336 td->error = WEXITSTATUS(status);
a2f77c9f 1337
a2f77c9f
JA
1338 td_set_runstate(td, TD_REAPED);
1339 goto reaped;
a6418147
JA
1340 }
1341 }
ebac4655 1342
a2f77c9f
JA
1343 /*
1344 * thread is not dead, continue
1345 */
e48676ba 1346 pending++;
a2f77c9f 1347 continue;
fab6aa71 1348reaped:
ebac4655 1349 (*nr_running)--;
581e7141
JA
1350 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1351 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
6f88bcb0
JA
1352 if (!td->pid)
1353 pending--;
a2f77c9f
JA
1354
1355 if (td->error)
1356 exit_value++;
d3eeeabc
JA
1357
1358 done_secs += mtime_since_now(&td->epoch) / 1000;
ebac4655 1359 }
b990b5c0 1360
1f809d15 1361 if (*nr_running == cputhreads && !pending && realthreads)
390c40e2 1362 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1363}
1364
be4ecfdf
JA
1365static void *gtod_thread_main(void *data)
1366{
1367 fio_mutex_up(startup_mutex);
1368
1369 /*
1370 * As long as we have jobs around, update the clock. It would be nice
1371 * to have some way of NOT hammering that CPU with gettimeofday(),
1372 * but I'm not sure what to use outside of a simple CPU nop to relax
1373 * it - we don't want to lose precision.
1374 */
1375 while (threads) {
1376 fio_gtod_update();
1377 nop;
1378 }
1379
1380 return NULL;
1381}
1382
1383static int fio_start_gtod_thread(void)
1384{
f42c153d
JA
1385 int ret;
1386
1387 ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1388 if (ret) {
1389 log_err("Can't create gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1390 return 1;
1391 }
f42c153d
JA
1392
1393 ret = pthread_detach(gtod_thread);
1394 if (ret) {
1395 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1396 return 1;
1397 }
1398
29adda3c 1399 dprint(FD_MUTEX, "wait on startup_mutex\n");
be4ecfdf 1400 fio_mutex_down(startup_mutex);
29adda3c 1401 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
be4ecfdf
JA
1402 return 0;
1403}
1404
906c8d75
JA
1405/*
1406 * Main function for kicking off and reaping jobs, as needed.
1407 */
ebac4655
JA
1408static void run_threads(void)
1409{
ebac4655
JA
1410 struct thread_data *td;
1411 unsigned long spent;
1412 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1413
2f9ade3c
JA
1414 if (fio_pin_memory())
1415 return;
ebac4655 1416
be4ecfdf
JA
1417 if (fio_gtod_offload && fio_start_gtod_thread())
1418 return;
1419
c6ae0a5b 1420 if (!terse_output) {
4ceb30d4 1421 log_info("Starting ");
9cedf167 1422 if (nr_thread)
4ceb30d4 1423 log_info("%d thread%s", nr_thread,
5ec10eaa 1424 nr_thread > 1 ? "s" : "");
9cedf167
JA
1425 if (nr_process) {
1426 if (nr_thread)
1427 printf(" and ");
4ceb30d4 1428 log_info("%d process%s", nr_process,
5ec10eaa 1429 nr_process > 1 ? "es" : "");
9cedf167 1430 }
4ceb30d4 1431 log_info("\n");
c6ae0a5b
JA
1432 fflush(stdout);
1433 }
c04f7ec3 1434
697a606c 1435 set_sig_handlers();
4efa970e 1436
ebac4655
JA
1437 todo = thread_number;
1438 nr_running = 0;
1439 nr_started = 0;
1440 m_rate = t_rate = 0;
1441
34572e28 1442 for_each_td(td, i) {
263e529f 1443 print_status_init(td->thread_number - 1);
ebac4655 1444
2dc1bbeb 1445 if (!td->o.create_serialize) {
380cf265 1446 init_disk_util(td);
ebac4655 1447 continue;
380cf265 1448 }
ebac4655
JA
1449
1450 /*
1451 * do file setup here so it happens sequentially,
1452 * we don't want X number of threads getting their
1453 * client data interspersed on disk
1454 */
53cdc686 1455 if (setup_files(td)) {
5bf13a5a
JA
1456 exit_value++;
1457 if (td->error)
5921e80c
JA
1458 log_err("fio: pid=%d, err=%d/%s\n",
1459 (int) td->pid, td->error, td->verror);
ebac4655
JA
1460 td_set_runstate(td, TD_REAPED);
1461 todo--;
c69e8875
JA
1462 } else {
1463 struct fio_file *f;
1464 unsigned int i;
1465
1466 /*
1467 * for sharing to work, each job must always open
1468 * its own files. so close them, if we opened them
1469 * for creation
1470 */
22a57ba8
JA
1471 for_each_file(td, f, i) {
1472 if (fio_file_open(f))
1473 td_io_close_file(td, f);
22a57ba8 1474 }
5ec10eaa 1475 }
380cf265
JA
1476
1477 init_disk_util(td);
ebac4655
JA
1478 }
1479
a2f77c9f
JA
1480 set_genesis_time();
1481
ebac4655 1482 while (todo) {
75154845
JA
1483 struct thread_data *map[MAX_JOBS];
1484 struct timeval this_start;
1485 int this_jobs = 0, left;
1486
ebac4655
JA
1487 /*
1488 * create threads (TD_NOT_CREATED -> TD_CREATED)
1489 */
34572e28 1490 for_each_td(td, i) {
ebac4655
JA
1491 if (td->runstate != TD_NOT_CREATED)
1492 continue;
1493
1494 /*
1495 * never got a chance to start, killed by other
1496 * thread for some reason
1497 */
1498 if (td->terminate) {
1499 todo--;
1500 continue;
1501 }
1502
2dc1bbeb 1503 if (td->o.start_delay) {
263e529f 1504 spent = mtime_since_genesis();
ebac4655 1505
2dc1bbeb 1506 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1507 continue;
1508 }
1509
6f88bcb0
JA
1510 if (td->o.stonewall && (nr_started || nr_running)) {
1511 dprint(FD_PROCESS, "%s: stonewall wait\n",
1512 td->o.name);
ebac4655 1513 break;
6f88bcb0 1514 }
ebac4655 1515
75154845
JA
1516 /*
1517 * Set state to created. Thread will transition
1518 * to TD_INITIALIZED when it's done setting up.
1519 */
ebac4655 1520 td_set_runstate(td, TD_CREATED);
75154845 1521 map[this_jobs++] = td;
ebac4655
JA
1522 nr_started++;
1523
2dc1bbeb 1524 if (td->o.use_thread) {
f42c153d
JA
1525 int ret;
1526
ee56ad50 1527 dprint(FD_PROCESS, "will pthread_create\n");
f42c153d
JA
1528 ret = pthread_create(&td->thread, NULL,
1529 thread_main, td);
1530 if (ret) {
1531 log_err("pthread_create: %s\n",
1532 strerror(ret));
ebac4655 1533 nr_started--;
c8bb6faf 1534 break;
ebac4655 1535 }
f42c153d
JA
1536 ret = pthread_detach(td->thread);
1537 if (ret)
1538 log_err("pthread_detach: %s",
1539 strerror(ret));
ebac4655 1540 } else {
5e1d306e 1541 pid_t pid;
ee56ad50 1542 dprint(FD_PROCESS, "will fork\n");
5e1d306e
JA
1543 pid = fork();
1544 if (!pid) {
a6418147
JA
1545 int ret = fork_main(shm_id, i);
1546
5e1d306e
JA
1547 _exit(ret);
1548 } else if (i == fio_debug_jobno)
1549 *fio_debug_jobp = pid;
ebac4655 1550 }
29adda3c 1551 dprint(FD_MUTEX, "wait on startup_mutex\n");
656b1393
JA
1552 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1553 log_err("fio: job startup hung? exiting.\n");
1554 terminate_threads(TERMINATE_ALL);
1555 fio_abort = 1;
1556 nr_started--;
1557 break;
1558 }
29adda3c 1559 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
ebac4655
JA
1560 }
1561
1562 /*
75154845
JA
1563 * Wait for the started threads to transition to
1564 * TD_INITIALIZED.
ebac4655 1565 */
02bcaa8c 1566 fio_gettime(&this_start, NULL);
75154845 1567 left = this_jobs;
6ce15a32 1568 while (left && !fio_abort) {
75154845
JA
1569 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1570 break;
1571
1572 usleep(100000);
1573
1574 for (i = 0; i < this_jobs; i++) {
1575 td = map[i];
1576 if (!td)
1577 continue;
b6f4d880 1578 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1579 map[i] = NULL;
1580 left--;
b6f4d880
JA
1581 } else if (td->runstate >= TD_EXITED) {
1582 map[i] = NULL;
1583 left--;
1584 todo--;
1585 nr_running++; /* work-around... */
75154845
JA
1586 }
1587 }
1588 }
1589
1590 if (left) {
3b70d7e5 1591 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1592 for (i = 0; i < this_jobs; i++) {
1593 td = map[i];
1594 if (!td)
1595 continue;
1596 kill(td->pid, SIGTERM);
1597 }
1598 break;
1599 }
1600
1601 /*
b6f4d880 1602 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1603 */
34572e28 1604 for_each_td(td, i) {
75154845 1605 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1606 continue;
1607
b29ee5b3
JA
1608 if (in_ramp_time(td))
1609 td_set_runstate(td, TD_RAMP);
1610 else
1611 td_set_runstate(td, TD_RUNNING);
ebac4655
JA
1612 nr_running++;
1613 nr_started--;
581e7141
JA
1614 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1615 t_rate += td->o.rate[0] + td->o.rate[1];
75154845 1616 todo--;
cdd18ad8 1617 fio_mutex_up(td->mutex);
ebac4655
JA
1618 }
1619
1620 reap_threads(&nr_running, &t_rate, &m_rate);
1621
1622 if (todo)
1623 usleep(100000);
1624 }
1625
1626 while (nr_running) {
1627 reap_threads(&nr_running, &t_rate, &m_rate);
1628 usleep(10000);
1629 }
1630
1631 update_io_ticks();
2f9ade3c 1632 fio_unpin_memory();
ebac4655
JA
1633}
1634
ebac4655
JA
1635int main(int argc, char *argv[])
1636{
29d610e1
JA
1637 long ps;
1638
2e5cdb11 1639 sinit();
d54a144d 1640 init_rand(&__fio_rand_state);
2e5cdb11 1641
dbe1125e
JA
1642 /*
1643 * We need locale for number printing, if it isn't set then just
1644 * go with the US format.
1645 */
1646 if (!getenv("LC_NUMERIC"))
1647 setlocale(LC_NUMERIC, "en_US");
1648
29d610e1
JA
1649 ps = sysconf(_SC_PAGESIZE);
1650 if (ps < 0) {
1651 log_err("Failed to get page size\n");
1652 return 1;
1653 }
1654
cfc99db7 1655 page_size = ps;
29d610e1
JA
1656 page_mask = ps - 1;
1657
74929ac2
JA
1658 fio_keywords_init();
1659
1660 if (parse_options(argc, argv))
1661 return 1;
1662
07b3232d
JA
1663 if (exec_profile && load_profile(exec_profile))
1664 return 1;
1665
74929ac2
JA
1666 if (!thread_number)
1667 return 0;
1668
bb3884d8
JA
1669 if (write_bw_log) {
1670 setup_log(&agg_io_log[DDIR_READ]);
1671 setup_log(&agg_io_log[DDIR_WRITE]);
1672 }
1673
cdd18ad8 1674 startup_mutex = fio_mutex_init(0);
d2bbb8d9 1675 writeout_mutex = fio_mutex_init(1);
07739b57 1676
a2f77c9f
JA
1677 set_genesis_time();
1678
ccb0fa24 1679 status_timer_arm();
ebac4655 1680
dae5341c
JA
1681 cgroup_list = smalloc(sizeof(*cgroup_list));
1682 INIT_FLIST_HEAD(cgroup_list);
1683
ebac4655 1684 run_threads();
6ce15a32 1685
bb3884d8 1686 if (!fio_abort) {
6ce15a32 1687 show_run_stats();
bb3884d8 1688 if (write_bw_log) {
5ec10eaa
JA
1689 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1690 __finish_log(agg_io_log[DDIR_WRITE],
1691 "agg-write_bw.log");
bb3884d8
JA
1692 }
1693 }
ebac4655 1694
dae5341c
JA
1695 cgroup_kill(cgroup_list);
1696 sfree(cgroup_list);
61eb313e 1697 sfree(cgroup_mnt);
39f22027 1698
cdd18ad8 1699 fio_mutex_remove(startup_mutex);
d2bbb8d9 1700 fio_mutex_remove(writeout_mutex);
437c9b71 1701 return exit_value;
ebac4655 1702}