Add gettid() for proper thread affinity on Linux
[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>
9b836561 27#include <limits.h>
ebac4655
JA
28#include <signal.h>
29#include <time.h>
dbe1125e 30#include <locale.h>
36167d82 31#include <assert.h>
a5b01f1b 32#include <time.h>
ebac4655
JA
33#include <sys/stat.h>
34#include <sys/wait.h>
35#include <sys/ipc.h>
36#include <sys/shm.h>
ebac4655
JA
37#include <sys/mman.h>
38
39#include "fio.h"
210f43b3 40#include "hash.h"
2e5cdb11 41#include "smalloc.h"
4f5af7b2 42#include "verify.h"
a917a8b3 43#include "trim.h"
7c9b1bce 44#include "diskutil.h"
a696fa2a 45#include "cgroup.h"
07b3232d 46#include "profile.h"
d54a144d 47#include "lib/rand.h"
91aea6ec 48#include "memalign.h"
ebac4655 49
cfc99db7
JA
50unsigned long page_mask;
51unsigned long page_size;
d529ee19
JA
52
53#define PAGE_ALIGN(buf) \
29d610e1 54 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
55
56int groupid = 0;
57int thread_number = 0;
9cedf167
JA
58int nr_process = 0;
59int nr_thread = 0;
ebac4655 60int shm_id = 0;
53cdc686 61int temp_stall_ts;
d3eeeabc 62unsigned long done_secs = 0;
ebac4655 63
cdd18ad8 64static struct fio_mutex *startup_mutex;
d2bbb8d9 65static struct fio_mutex *writeout_mutex;
6ce15a32 66static volatile int fio_abort;
437c9b71 67static int exit_value;
be4ecfdf 68static pthread_t gtod_thread;
7985701b 69static pthread_t disk_util_thread;
dae5341c 70static struct flist_head *cgroup_list;
6adb38a1 71static char *cgroup_mnt;
ebac4655 72
bb3884d8
JA
73struct io_log *agg_io_log[2];
74
ebac4655 75#define TERMINATE_ALL (-1)
75154845 76#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 77
b29ee5b3 78void td_set_runstate(struct thread_data *td, int runstate)
6ce15a32 79{
d8fab1b6
JA
80 if (td->runstate == runstate)
81 return;
82
5921e80c
JA
83 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
84 td->runstate, runstate);
6ce15a32
JA
85 td->runstate = runstate;
86}
87
390c40e2 88static void terminate_threads(int group_id)
ebac4655 89{
34572e28 90 struct thread_data *td;
ebac4655
JA
91 int i;
92
6f88bcb0
JA
93 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
94
34572e28 95 for_each_td(td, i) {
ebac4655 96 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
6f88bcb0 97 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
5921e80c 98 td->o.name, (int) td->pid);
ad830ed7
JA
99 td->terminate = 1;
100 td->o.start_delay = 0;
101
7a93ab19
JA
102 /*
103 * if the thread is running, just let it exit
104 */
f097728b 105 if (td->runstate < TD_RAMP)
03e20d68 106 kill(td->pid, SIGTERM);
f63f7f3b
JA
107 else {
108 struct ioengine_ops *ops = td->io_ops;
109
03e20d68
BC
110 if (ops && (ops->flags & FIO_SIGTERM))
111 kill(td->pid, SIGTERM);
f63f7f3b 112 }
ebac4655
JA
113 }
114 }
115}
116
697a606c
JA
117static void sig_int(int sig)
118{
119 if (threads) {
4ceb30d4 120 log_info("\nfio: terminating on signal %d\n", sig);
5ec10eaa 121 fflush(stdout);
deaa1af2 122 exit_value = 128;
5ec10eaa 123 terminate_threads(TERMINATE_ALL);
ebac4655
JA
124 }
125}
126
7985701b 127static void *disk_thread_main(void *data)
a5b01f1b 128{
7985701b
JA
129 fio_mutex_up(startup_mutex);
130
131 while (threads) {
132 usleep(DISK_UTIL_MSEC * 1000);
9a793c2f
JA
133 if (!threads)
134 break;
7985701b
JA
135 update_io_ticks();
136 print_thread_status();
137 }
138
139 return NULL;
a5b01f1b
JA
140}
141
7985701b 142static int create_disk_util_thread(void)
a5b01f1b 143{
7985701b 144 int ret;
a5b01f1b 145
69204d6e 146 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
7985701b
JA
147 if (ret) {
148 log_err("Can't create disk util thread: %s\n", strerror(ret));
149 return 1;
150 }
a5b01f1b 151
7985701b
JA
152 ret = pthread_detach(disk_util_thread);
153 if (ret) {
154 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
155 return 1;
156 }
03e20d68 157
7985701b
JA
158 dprint(FD_MUTEX, "wait on startup_mutex\n");
159 fio_mutex_down(startup_mutex);
160 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
161 return 0;
a5b01f1b
JA
162}
163
697a606c
JA
164static void set_sig_handlers(void)
165{
166 struct sigaction act;
167
697a606c
JA
168 memset(&act, 0, sizeof(act));
169 act.sa_handler = sig_int;
170 act.sa_flags = SA_RESTART;
171 sigaction(SIGINT, &act, NULL);
de79c915 172
5f8f9726 173 memset(&act, 0, sizeof(act));
b20b6be4 174 act.sa_handler = sig_int;
5f8f9726 175 act.sa_flags = SA_RESTART;
03e20d68 176 sigaction(SIGTERM, &act, NULL);
697a606c
JA
177}
178
906c8d75
JA
179/*
180 * Check if we are above the minimum rate given.
181 */
581e7141 182static int __check_min_rate(struct thread_data *td, struct timeval *now,
bd93e34b 183 enum fio_ddir ddir)
ebac4655 184{
0904200b 185 unsigned long long bytes = 0;
4e991c23 186 unsigned long iops = 0;
ebac4655
JA
187 unsigned long spent;
188 unsigned long rate;
581e7141
JA
189 unsigned int ratemin = 0;
190 unsigned int rate_iops = 0;
191 unsigned int rate_iops_min = 0;
ebac4655 192
ff58fced
JA
193 assert(ddir_rw(ddir));
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
687c6a23
JA
375static inline void __update_tv_cache(struct thread_data *td)
376{
377 fio_gettime(&td->tv_cache, NULL);
378}
379
993bf48b
JA
380static inline void update_tv_cache(struct thread_data *td)
381{
382 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
687c6a23 383 __update_tv_cache(td);
993bf48b
JA
384}
385
f2bba182
RR
386static int break_on_this_error(struct thread_data *td, int *retptr)
387{
388 int ret = *retptr;
389
390 if (ret < 0 || td->error) {
391 int err;
392
1ec99eea
JA
393 if (!td->o.continue_on_error)
394 return 1;
f2bba182
RR
395
396 if (ret < 0)
397 err = -ret;
398 else
399 err = td->error;
400
f2bba182
RR
401 if (td_non_fatal_error(err)) {
402 /*
403 * Continue with the I/Os in case of
404 * a non fatal error.
405 */
8a4f8a24 406 update_error_count(td, err);
f2bba182
RR
407 td_clear_error(td);
408 *retptr = 0;
409 return 0;
8a4f8a24
RR
410 } else if (td->o.fill_device && err == ENOSPC) {
411 /*
412 * We expect to hit this error if
413 * fill_device option is set.
414 */
415 td_clear_error(td);
416 td->terminate = 1;
417 return 1;
f2bba182
RR
418 } else {
419 /*
420 * Stop the I/O in case of a fatal
421 * error.
422 */
8a4f8a24 423 update_error_count(td, err);
f2bba182
RR
424 return 1;
425 }
426 }
427
428 return 0;
429}
430
906c8d75 431/*
34403fb1 432 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
433 * reads the blocks back in, and checks the crc/md5 of the data.
434 */
1e97cce9 435static void do_verify(struct thread_data *td)
ebac4655 436{
53cdc686 437 struct fio_file *f;
36167d82 438 struct io_u *io_u;
af52b345
JA
439 int ret, min_events;
440 unsigned int i;
e5b401d4 441
008618a8
JA
442 dprint(FD_VERIFY, "starting loop\n");
443
e5b401d4
JA
444 /*
445 * sync io first and invalidate cache, to make sure we really
446 * read from disk.
447 */
448 for_each_file(td, f, i) {
d6aed795 449 if (!fio_file_open(f))
3d7b485f 450 continue;
b2fdda43
JA
451 if (fio_io_sync(td, f))
452 break;
453 if (file_invalidate_cache(td, f))
454 break;
e5b401d4 455 }
ebac4655 456
b2fdda43
JA
457 if (td->error)
458 return;
459
ebac4655
JA
460 td_set_runstate(td, TD_VERIFYING);
461
36167d82
JA
462 io_u = NULL;
463 while (!td->terminate) {
4950421a 464 int ret2, full;
5451792e 465
993bf48b
JA
466 update_tv_cache(td);
467
468 if (runtime_exceeded(td, &td->tv_cache)) {
687c6a23
JA
469 __update_tv_cache(td);
470 if (runtime_exceeded(td, &td->tv_cache)) {
471 td->terminate = 1;
472 break;
473 }
069c2918 474 }
02bcaa8c 475
dcb4830d
RR
476 io_u = __get_io_u(td);
477 if (!io_u)
478 break;
479
069c2918
JA
480 if (get_next_verify(td, io_u)) {
481 put_io_u(td, io_u);
ebac4655 482 break;
069c2918 483 }
ebac4655 484
069c2918
JA
485 if (td_io_prep(td, io_u)) {
486 put_io_u(td, io_u);
53cdc686 487 break;
069c2918 488 }
d7762cf8 489
e8462bd8
JA
490 if (td->o.verify_async)
491 io_u->end_io = verify_io_u_async;
492 else
493 io_u->end_io = verify_io_u;
53cdc686 494
11786802 495 ret = td_io_queue(td, io_u);
36167d82
JA
496 switch (ret) {
497 case FIO_Q_COMPLETED:
f2bba182 498 if (io_u->error) {
22819ec2 499 ret = -io_u->error;
f2bba182
RR
500 clear_io_u(td, io_u);
501 } else if (io_u->resid) {
36167d82 502 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 503
9e9d164e
JA
504 /*
505 * zero read, fail
506 */
507 if (!bytes) {
41d9f0fd 508 td_verror(td, EIO, "full resid");
9e9d164e
JA
509 put_io_u(td, io_u);
510 break;
511 }
8400d9b2 512
36167d82
JA
513 io_u->xfer_buflen = io_u->resid;
514 io_u->xfer_buf += bytes;
8400d9b2
JA
515 io_u->offset += bytes;
516
ff58fced
JA
517 if (ddir_rw(io_u->ddir))
518 td->ts.short_io_u[io_u->ddir]++;
30061b97 519
2b13e716 520 f = io_u->file;
d460eb31 521 if (io_u->offset == f->real_file_size)
8400d9b2
JA
522 goto sync_done;
523
11786802
JA
524 requeue_io_u(td, &io_u);
525 } else {
8400d9b2 526sync_done:
581e7141 527 ret = io_u_sync_complete(td, io_u, NULL);
11786802
JA
528 if (ret < 0)
529 break;
36167d82 530 }
36167d82
JA
531 continue;
532 case FIO_Q_QUEUED:
533 break;
755200a3
JA
534 case FIO_Q_BUSY:
535 requeue_io_u(td, &io_u);
5451792e
JA
536 ret2 = td_io_commit(td);
537 if (ret2 < 0)
538 ret = ret2;
755200a3 539 break;
36167d82
JA
540 default:
541 assert(ret < 0);
e1161c32 542 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
543 break;
544 }
545
f2bba182 546 if (break_on_this_error(td, &ret))
3af6ef39
JA
547 break;
548
ebac4655 549 /*
3af6ef39
JA
550 * if we can queue more, do so. but check if there are
551 * completed io_u's first.
ebac4655 552 */
4950421a
JA
553 full = queue_full(td) || ret == FIO_Q_BUSY;
554 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
555 min_events = min(td->o.iodepth_batch_complete,
556 td->cur_depth);
4950421a 557 if (full && !min_events)
838bc709 558 min_events = 1;
e916b390 559
4950421a
JA
560 do {
561 /*
562 * Reap required number of io units, if any,
563 * and do the verification on them through
564 * the callback handler
565 */
581e7141 566 if (io_u_queued_complete(td, min_events, NULL) < 0) {
4950421a
JA
567 ret = -1;
568 break;
569 }
570 } while (full && (td->cur_depth > td->o.iodepth_low));
571 }
572 if (ret < 0)
ebac4655 573 break;
36167d82 574 }
ebac4655 575
c01c0395
JA
576 if (!td->error) {
577 min_events = td->cur_depth;
578
579 if (min_events)
581e7141 580 ret = io_u_queued_complete(td, min_events, NULL);
c01c0395 581 } else
ebac4655
JA
582 cleanup_pending_aio(td);
583
584 td_set_runstate(td, TD_RUNNING);
008618a8
JA
585
586 dprint(FD_VERIFY, "exiting loop\n");
ebac4655
JA
587}
588
32cd46a0 589/*
906c8d75 590 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
591 * and reaps them, checking for rate and errors along the way.
592 */
ebac4655
JA
593static void do_io(struct thread_data *td)
594{
af52b345
JA
595 unsigned int i;
596 int ret = 0;
ebac4655 597
b29ee5b3
JA
598 if (in_ramp_time(td))
599 td_set_runstate(td, TD_RAMP);
600 else
601 td_set_runstate(td, TD_RUNNING);
5853e5a8 602
457bf399 603 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
a917a8b3 604 (!flist_empty(&td->trim_list)) ||
457bf399 605 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
97601024 606 struct timeval comp_time;
581e7141 607 unsigned long bytes_done[2] = { 0, 0 };
84585003 608 int min_evts = 0;
ebac4655 609 struct io_u *io_u;
4950421a 610 int ret2, full;
ebac4655
JA
611
612 if (td->terminate)
613 break;
614
993bf48b 615 update_tv_cache(td);
97601024 616
993bf48b 617 if (runtime_exceeded(td, &td->tv_cache)) {
687c6a23
JA
618 __update_tv_cache(td);
619 if (runtime_exceeded(td, &td->tv_cache)) {
620 td->terminate = 1;
621 break;
622 }
97601024 623 }
36167d82 624
dcb4830d
RR
625 io_u = get_io_u(td);
626 if (!io_u)
627 break;
628
b67740d3
JA
629 /*
630 * Add verification end_io handler, if asked to verify
631 * a previously written file.
632 */
715d2b3e
JA
633 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
634 !td_rw(td)) {
e8462bd8
JA
635 if (td->o.verify_async)
636 io_u->end_io = verify_io_u_async;
637 else
638 io_u->end_io = verify_io_u;
d8fab1b6 639 td_set_runstate(td, TD_VERIFYING);
b29ee5b3
JA
640 } else if (in_ramp_time(td))
641 td_set_runstate(td, TD_RAMP);
642 else
d8fab1b6 643 td_set_runstate(td, TD_RUNNING);
b67740d3 644
11786802 645 ret = td_io_queue(td, io_u);
36167d82
JA
646 switch (ret) {
647 case FIO_Q_COMPLETED:
f2bba182 648 if (io_u->error) {
5451792e 649 ret = -io_u->error;
f2bba182
RR
650 clear_io_u(td, io_u);
651 } else if (io_u->resid) {
36167d82 652 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 653 struct fio_file *f = io_u->file;
36167d82 654
9e9d164e
JA
655 /*
656 * zero read, fail
657 */
658 if (!bytes) {
41d9f0fd 659 td_verror(td, EIO, "full resid");
9e9d164e
JA
660 put_io_u(td, io_u);
661 break;
662 }
663
cec6b55d 664 io_u->xfer_buflen = io_u->resid;
36167d82 665 io_u->xfer_buf += bytes;
5a7c5680
JA
666 io_u->offset += bytes;
667
ff58fced
JA
668 if (ddir_rw(io_u->ddir))
669 td->ts.short_io_u[io_u->ddir]++;
30061b97 670
d460eb31 671 if (io_u->offset == f->real_file_size)
5a7c5680
JA
672 goto sync_done;
673
11786802
JA
674 requeue_io_u(td, &io_u);
675 } else {
5a7c5680 676sync_done:
581e7141
JA
677 if (__should_check_rate(td, 0) ||
678 __should_check_rate(td, 1))
993bf48b
JA
679 fio_gettime(&comp_time, NULL);
680
581e7141
JA
681 ret = io_u_sync_complete(td, io_u, bytes_done);
682 if (ret < 0)
683 break;
cec6b55d 684 }
36167d82
JA
685 break;
686 case FIO_Q_QUEUED:
7e77dd02
JA
687 /*
688 * if the engine doesn't have a commit hook,
689 * the io_u is really queued. if it does have such
690 * a hook, it has to call io_u_queued() itself.
691 */
692 if (td->io_ops->commit == NULL)
693 io_u_queued(td, io_u);
36167d82 694 break;
755200a3
JA
695 case FIO_Q_BUSY:
696 requeue_io_u(td, &io_u);
5451792e
JA
697 ret2 = td_io_commit(td);
698 if (ret2 < 0)
699 ret = ret2;
755200a3 700 break;
36167d82
JA
701 default:
702 assert(ret < 0);
703 put_io_u(td, io_u);
704 break;
ebac4655
JA
705 }
706
f2bba182 707 if (break_on_this_error(td, &ret))
36167d82
JA
708 break;
709
97601024
JA
710 /*
711 * See if we need to complete some commands
712 */
4950421a
JA
713 full = queue_full(td) || ret == FIO_Q_BUSY;
714 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
715 min_evts = min(td->o.iodepth_batch_complete,
716 td->cur_depth);
4950421a 717 if (full && !min_evts)
36167d82 718 min_evts = 1;
4950421a 719
581e7141
JA
720 if (__should_check_rate(td, 0) ||
721 __should_check_rate(td, 1))
993bf48b 722 fio_gettime(&comp_time, NULL);
4950421a
JA
723
724 do {
581e7141
JA
725 ret = io_u_queued_complete(td, min_evts, bytes_done);
726 if (ret < 0)
4950421a
JA
727 break;
728
4950421a 729 } while (full && (td->cur_depth > td->o.iodepth_low));
ebac4655
JA
730 }
731
4950421a
JA
732 if (ret < 0)
733 break;
581e7141 734 if (!(bytes_done[0] + bytes_done[1]))
97601024
JA
735 continue;
736
581e7141
JA
737 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
738 if (check_min_rate(td, &comp_time, bytes_done)) {
721938ae
JA
739 if (exitall_on_terminate)
740 terminate_threads(td->groupid);
741 td_verror(td, EIO, "check_min_rate");
742 break;
743 }
ebac4655
JA
744 }
745
2dc1bbeb 746 if (td->o.thinktime) {
9c1f7434
JA
747 unsigned long long b;
748
749 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 750 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
751 int left;
752
2dc1bbeb 753 if (td->o.thinktime_spin)
6dd6f2cd 754 usec_spin(td->o.thinktime_spin);
48097d5c 755
2dc1bbeb 756 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
757 if (left)
758 usec_sleep(td, left);
759 }
9c1f7434 760 }
ebac4655
JA
761 }
762
d708ac5e
JA
763 if (td->trim_entries)
764 printf("trim entries %ld\n", td->trim_entries);
a917a8b3 765
aa31f1f1
SL
766 if (td->o.fill_device && td->error == ENOSPC) {
767 td->error = 0;
768 td->terminate = 1;
769 }
4d2413c6 770 if (!td->error) {
3d7c391d
JA
771 struct fio_file *f;
772
c01c0395 773 i = td->cur_depth;
d06093e7 774 if (i) {
581e7141 775 ret = io_u_queued_complete(td, i, NULL);
d06093e7
SL
776 if (td->o.fill_device && td->error == ENOSPC)
777 td->error = 0;
778 }
ebac4655 779
2dc1bbeb 780 if (should_fsync(td) && td->o.end_fsync) {
84585003 781 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
782
783 for_each_file(td, f, i) {
d6aed795 784 if (!fio_file_open(f))
3d7b485f 785 continue;
858a3d47 786 fio_io_sync(td, f);
3d7b485f 787 }
84585003 788 }
c01c0395
JA
789 } else
790 cleanup_pending_aio(td);
2ba1c290
JA
791
792 /*
793 * stop job if we failed doing any IO
794 */
795 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
796 td->done = 1;
ebac4655
JA
797}
798
ebac4655
JA
799static void cleanup_io_u(struct thread_data *td)
800{
01743ee1 801 struct flist_head *entry, *n;
ebac4655
JA
802 struct io_u *io_u;
803
01743ee1
JA
804 flist_for_each_safe(entry, n, &td->io_u_freelist) {
805 io_u = flist_entry(entry, struct io_u, list);
ebac4655 806
01743ee1 807 flist_del(&io_u->list);
91aea6ec 808 fio_memfree(io_u, sizeof(*io_u));
ebac4655
JA
809 }
810
2f9ade3c 811 free_io_mem(td);
ebac4655
JA
812}
813
814static int init_io_u(struct thread_data *td)
815{
816 struct io_u *io_u;
a00735e6 817 unsigned int max_bs;
eb7ccf38 818 int cl_align, i, max_units;
ebac4655
JA
819 char *p;
820
1e3c09af 821 max_units = td->o.iodepth;
2dc1bbeb 822 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
5ec10eaa
JA
823 td->orig_buffer_size = (unsigned long long) max_bs
824 * (unsigned long long) max_units;
825
826 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
827 unsigned long bs;
74b025b0 828
5ec10eaa
JA
829 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
830 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
831 }
ebac4655 832
c3990645
JA
833 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
834 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
835 return 1;
836 }
837
2f9ade3c
JA
838 if (allocate_io_mem(td))
839 return 1;
ebac4655 840
1c04bf78
JA
841 if (td->o.odirect || td->o.mem_align ||
842 (td->io_ops->flags & FIO_RAWIO))
d529ee19 843 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
b3f378e9
JA
844 else
845 p = td->orig_buffer;
846
eb7ccf38
JA
847 cl_align = os_cache_line_size();
848
ebac4655 849 for (i = 0; i < max_units; i++) {
eb7ccf38
JA
850 void *ptr;
851
db1cd90b
JA
852 if (td->terminate)
853 return 1;
eb7ccf38 854
91aea6ec
JA
855 ptr = fio_memalign(cl_align, sizeof(*io_u));
856 if (!ptr) {
857 log_err("fio: unable to allocate aligned memory\n");
eb7ccf38
JA
858 break;
859 }
860
861 io_u = ptr;
ebac4655 862 memset(io_u, 0, sizeof(*io_u));
01743ee1 863 INIT_FLIST_HEAD(&io_u->list);
d529ee19 864 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
ebac4655 865
b4c5e1ac
JA
866 if (!(td->io_ops->flags & FIO_NOIO)) {
867 io_u->buf = p + max_bs * i;
d529ee19 868 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
e9459e5a 869
5973cafb
JA
870 if (td_write(td) && !td->o.refill_buffers)
871 io_u_fill_buffer(td, io_u, max_bs);
cbe8d756
RR
872 else if (td_write(td) && td->o.verify_pattern_bytes) {
873 /*
874 * Fill the buffer with the pattern if we are
875 * going to be doing writes.
876 */
7d9fb455 877 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
cbe8d756 878 }
b4c5e1ac 879 }
6b9cea23 880
b1ff3403 881 io_u->index = i;
0c6e7517 882 io_u->flags = IO_U_F_FREE;
01743ee1 883 flist_add(&io_u->list, &td->io_u_freelist);
ebac4655
JA
884 }
885
886 return 0;
887}
888
da86774e
JA
889static int switch_ioscheduler(struct thread_data *td)
890{
891 char tmp[256], tmp2[128];
892 FILE *f;
893 int ret;
894
ba0fbe10 895 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
896 return 0;
897
da86774e
JA
898 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
899
900 f = fopen(tmp, "r+");
901 if (!f) {
cbf5c1d7 902 if (errno == ENOENT) {
5ec10eaa
JA
903 log_err("fio: os or kernel doesn't support IO scheduler"
904 " switching\n");
cbf5c1d7
JA
905 return 0;
906 }
907 td_verror(td, errno, "fopen iosched");
da86774e
JA
908 return 1;
909 }
910
911 /*
912 * Set io scheduler.
913 */
2dc1bbeb 914 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 915 if (ferror(f) || ret != 1) {
e1161c32 916 td_verror(td, errno, "fwrite");
da86774e
JA
917 fclose(f);
918 return 1;
919 }
920
921 rewind(f);
922
923 /*
924 * Read back and check that the selected scheduler is now the default.
925 */
926 ret = fread(tmp, 1, sizeof(tmp), f);
927 if (ferror(f) || ret < 0) {
e1161c32 928 td_verror(td, errno, "fread");
da86774e
JA
929 fclose(f);
930 return 1;
931 }
932
2dc1bbeb 933 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 934 if (!strstr(tmp, tmp2)) {
2dc1bbeb 935 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 936 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
937 fclose(f);
938 return 1;
939 }
940
941 fclose(f);
942 return 0;
943}
944
492158cf
JA
945static int keep_running(struct thread_data *td)
946{
947 unsigned long long io_done;
948
20e354ef
JA
949 if (td->done)
950 return 0;
492158cf
JA
951 if (td->o.time_based)
952 return 1;
953 if (td->o.loops) {
954 td->o.loops--;
955 return 1;
956 }
957
5ec10eaa
JA
958 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
959 + td->io_skip_bytes;
492158cf
JA
960 if (io_done < td->o.size)
961 return 1;
962
963 return 0;
964}
965
b29ee5b3 966static void reset_io_counters(struct thread_data *td)
ebac4655 967{
756867bd 968 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
ebac4655 969 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 970 td->zone_bytes = 0;
581e7141
JA
971 td->rate_bytes[0] = td->rate_bytes[1] = 0;
972 td->rate_blocks[0] = td->rate_blocks[1] = 0;
ebac4655 973
c1324df1
JA
974 td->last_was_sync = 0;
975
2ba1c290
JA
976 /*
977 * reset file done count if we are to start over
978 */
979 if (td->o.time_based || td->o.loops)
48f5abd3 980 td->nr_done_files = 0;
b29ee5b3
JA
981}
982
983void reset_all_stats(struct thread_data *td)
984{
985 struct timeval tv;
986 int i;
987
988 reset_io_counters(td);
989
990 for (i = 0; i < 2; i++) {
991 td->io_bytes[i] = 0;
992 td->io_blocks[i] = 0;
993 td->io_issues[i] = 0;
994 td->ts.total_io_u[i] = 0;
995 }
9b836561 996
b29ee5b3 997 fio_gettime(&tv, NULL);
cda99fa0
SSP
998 td->ts.runtime[0] = 0;
999 td->ts.runtime[1] = 0;
b29ee5b3
JA
1000 memcpy(&td->epoch, &tv, sizeof(tv));
1001 memcpy(&td->start, &tv, sizeof(tv));
1002}
1003
df9c26b1 1004static void clear_io_state(struct thread_data *td)
b29ee5b3
JA
1005{
1006 struct fio_file *f;
1007 unsigned int i;
b29ee5b3
JA
1008
1009 reset_io_counters(td);
281f9b63 1010
24ffd2c2 1011 close_files(td);
df9c26b1 1012 for_each_file(td, f, i)
d6aed795 1013 fio_file_clear_done(f);
f0153948
JA
1014
1015 /*
1016 * Set the same seed to get repeatable runs
1017 */
1018 td_fill_rand_seeds(td);
ebac4655
JA
1019}
1020
398e8c4d
JA
1021static int exec_string(const char *string)
1022{
1023 int ret, newlen = strlen(string) + 1 + 8;
1024 char *str;
1025
1026 str = malloc(newlen);
1027 sprintf(str, "sh -c %s", string);
1028
1029 ret = system(str);
1030 if (ret == -1)
1031 log_err("fio: exec of cmd <%s> failed\n", str);
1032
1033 free(str);
1034 return ret;
1035}
1036
906c8d75
JA
1037/*
1038 * Entry point for the thread based jobs. The process based jobs end up
1039 * here as well, after a little setup.
1040 */
ebac4655
JA
1041static void *thread_main(void *data)
1042{
cda99fa0 1043 unsigned long long elapsed;
ebac4655 1044 struct thread_data *td = data;
e8462bd8 1045 pthread_condattr_t attr;
a978ba68 1046 int clear_state;
ebac4655 1047
47f767c1 1048 if (!td->o.use_thread) {
ebac4655 1049 setsid();
47f767c1
JA
1050 td->pid = getpid();
1051 } else
1052 td->pid = gettid();
ebac4655 1053
5921e80c 1054 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
ee56ad50 1055
01743ee1
JA
1056 INIT_FLIST_HEAD(&td->io_u_freelist);
1057 INIT_FLIST_HEAD(&td->io_u_busylist);
1058 INIT_FLIST_HEAD(&td->io_u_requeues);
1059 INIT_FLIST_HEAD(&td->io_log_list);
1060 INIT_FLIST_HEAD(&td->io_hist_list);
e8462bd8 1061 INIT_FLIST_HEAD(&td->verify_list);
0d29de83 1062 INIT_FLIST_HEAD(&td->trim_list);
e8462bd8 1063 pthread_mutex_init(&td->io_u_lock, NULL);
bb5d7d0b 1064 td->io_hist_tree = RB_ROOT;
aea47d44 1065
e8462bd8
JA
1066 pthread_condattr_init(&attr);
1067 pthread_cond_init(&td->verify_cond, &attr);
1068 pthread_cond_init(&td->free_cond, &attr);
1069
db1cd90b 1070 td_set_runstate(td, TD_INITIALIZED);
29adda3c 1071 dprint(FD_MUTEX, "up startup_mutex\n");
cdd18ad8 1072 fio_mutex_up(startup_mutex);
29adda3c 1073 dprint(FD_MUTEX, "wait on td->mutex\n");
cdd18ad8 1074 fio_mutex_down(td->mutex);
29adda3c 1075 dprint(FD_MUTEX, "done waiting on td->mutex\n");
db1cd90b
JA
1076
1077 /*
cdd18ad8 1078 * the ->mutex mutex is now no longer used, close it to avoid
db1cd90b
JA
1079 * eating a file descriptor
1080 */
cdd18ad8 1081 fio_mutex_remove(td->mutex);
db1cd90b 1082
dfc6e751
SH
1083 /*
1084 * A new gid requires privilege, so we need to do this before setting
1085 * the uid.
1086 */
e0b0d892
JA
1087 if (td->o.gid != -1U && setgid(td->o.gid)) {
1088 td_verror(td, errno, "setgid");
1089 goto err;
1090 }
dfc6e751
SH
1091 if (td->o.uid != -1U && setuid(td->o.uid)) {
1092 td_verror(td, errno, "setuid");
1093 goto err;
1094 }
e0b0d892 1095
be18a94f
JA
1096 /*
1097 * If we have a gettimeofday() thread, make sure we exclude that
1098 * thread from this job
1099 */
1100 if (td->o.gtod_cpu)
1101 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1102
1103 /*
1104 * Set affinity first, in case it has an impact on the memory
1105 * allocations.
1106 */
1107 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1108 td_verror(td, errno, "cpu_set_affinity");
1109 goto err;
1110 }
1111
d84f8d49
JA
1112 /*
1113 * May alter parameters that init_io_u() will use, so we need to
1114 * do this first.
1115 */
1116 if (init_iolog(td))
1117 goto err;
1118
ebac4655 1119 if (init_io_u(td))
db1cd90b 1120 goto err;
ebac4655 1121
e8462bd8
JA
1122 if (td->o.verify_async && verify_async_init(td))
1123 goto err;
1124
ac684785 1125 if (td->ioprio_set) {
ebac4655 1126 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 1127 td_verror(td, errno, "ioprio_set");
db1cd90b 1128 goto err;
ebac4655
JA
1129 }
1130 }
1131
6adb38a1 1132 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
a696fa2a
JA
1133 goto err;
1134
2dc1bbeb 1135 if (nice(td->o.nice) == -1) {
e1161c32 1136 td_verror(td, errno, "nice");
db1cd90b 1137 goto err;
b6f4d880
JA
1138 }
1139
2dc1bbeb 1140 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 1141 goto err;
37f56873 1142
2dc1bbeb 1143 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
1144 goto err;
1145
7d6c5283
JA
1146 if (td_io_init(td))
1147 goto err;
1148
68727076
JA
1149 if (init_random_map(td))
1150 goto err;
1151
2dc1bbeb 1152 if (td->o.exec_prerun) {
398e8c4d 1153 if (exec_string(td->o.exec_prerun))
69cfd7e0
JA
1154 goto err;
1155 }
4e0ba8af 1156
afad68f7
ZY
1157 if (td->o.pre_read) {
1158 if (pre_read_files(td) < 0)
1159 goto err;
1160 }
1161
69008999 1162 fio_gettime(&td->epoch, NULL);
756867bd 1163 getrusage(RUSAGE_SELF, &td->ts.ru_start);
69008999 1164
a978ba68 1165 clear_state = 0;
492158cf 1166 while (keep_running(td)) {
02bcaa8c 1167 fio_gettime(&td->start, NULL);
c5d3d790
JA
1168 memcpy(&td->ts.stat_sample_time[0], &td->start,
1169 sizeof(td->start));
1170 memcpy(&td->ts.stat_sample_time[1], &td->start,
1171 sizeof(td->start));
993bf48b 1172 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
ebac4655 1173
cf2dd8d6 1174 if (td->o.ratemin[0] || td->o.ratemin[1])
5ec10eaa
JA
1175 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1176 sizeof(td->lastrate));
ebac4655 1177
df9c26b1
JA
1178 if (clear_state)
1179 clear_io_state(td);
a978ba68 1180
ebac4655
JA
1181 prune_io_piece_log(td);
1182
ba0fbe10 1183 do_io(td);
ebac4655 1184
a978ba68
JA
1185 clear_state = 1;
1186
38d77cae 1187 if (td_read(td) && td->io_bytes[DDIR_READ]) {
0a4957c6 1188 elapsed = utime_since_now(&td->start);
cda99fa0 1189 td->ts.runtime[DDIR_READ] += elapsed;
38d77cae
JA
1190 }
1191 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
0a4957c6 1192 elapsed = utime_since_now(&td->start);
cda99fa0 1193 td->ts.runtime[DDIR_WRITE] += elapsed;
38d77cae 1194 }
5ec10eaa 1195
ebac4655
JA
1196 if (td->error || td->terminate)
1197 break;
1198
e84c73a8
SL
1199 if (!td->o.do_verify ||
1200 td->o.verify == VERIFY_NONE ||
b67740d3 1201 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
1202 continue;
1203
df9c26b1 1204 clear_io_state(td);
a978ba68 1205
02bcaa8c 1206 fio_gettime(&td->start, NULL);
ebac4655
JA
1207
1208 do_verify(td);
1209
cda99fa0 1210 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
1211
1212 if (td->error || td->terminate)
1213 break;
1214 }
1215
36dff966 1216 update_rusage_stat(td);
cda99fa0
SSP
1217 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1218 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
756867bd
JA
1219 td->ts.total_run_time = mtime_since_now(&td->epoch);
1220 td->ts.io_bytes[0] = td->io_bytes[0];
1221 td->ts.io_bytes[1] = td->io_bytes[1];
1222
d2bbb8d9 1223 fio_mutex_down(writeout_mutex);
e3cedca7 1224 if (td->ts.bw_log) {
f484470b
JA
1225 if (td->o.bw_log_file) {
1226 finish_log_named(td, td->ts.bw_log,
1227 td->o.bw_log_file, "bw");
1228 } else
e3cedca7
JA
1229 finish_log(td, td->ts.bw_log, "bw");
1230 }
02af0988
JA
1231 if (td->ts.lat_log) {
1232 if (td->o.lat_log_file) {
1233 finish_log_named(td, td->ts.lat_log,
1234 td->o.lat_log_file, "lat");
1235 } else
1236 finish_log(td, td->ts.lat_log, "lat");
1237 }
e3cedca7 1238 if (td->ts.slat_log) {
f484470b
JA
1239 if (td->o.lat_log_file) {
1240 finish_log_named(td, td->ts.slat_log,
1ca95738 1241 td->o.lat_log_file, "slat");
f484470b 1242 } else
e3cedca7
JA
1243 finish_log(td, td->ts.slat_log, "slat");
1244 }
1245 if (td->ts.clat_log) {
f484470b
JA
1246 if (td->o.lat_log_file) {
1247 finish_log_named(td, td->ts.clat_log,
1248 td->o.lat_log_file, "clat");
1249 } else
e3cedca7
JA
1250 finish_log(td, td->ts.clat_log, "clat");
1251 }
d2bbb8d9 1252 fio_mutex_up(writeout_mutex);
398e8c4d
JA
1253 if (td->o.exec_postrun)
1254 exec_string(td->o.exec_postrun);
ebac4655
JA
1255
1256 if (exitall_on_terminate)
390c40e2 1257 terminate_threads(td->groupid);
ebac4655
JA
1258
1259err:
5bf13a5a 1260 if (td->error)
4ceb30d4 1261 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
5ec10eaa 1262 td->verror);
b4707ed7
JA
1263
1264 if (td->o.verify_async)
1265 verify_async_exit(td);
1266
24ffd2c2 1267 close_and_free_files(td);
2866c82d 1268 close_ioengine(td);
ebac4655 1269 cleanup_io_u(td);
6adb38a1 1270 cgroup_shutdown(td, &cgroup_mnt);
f29b25a3 1271
d2ce18b5
JA
1272 if (td->o.cpumask_set) {
1273 int ret = fio_cpuset_exit(&td->o.cpumask);
1274
1275 td_verror(td, ret, "fio_cpuset_exit");
1276 }
4e78e405 1277
f29b25a3
JA
1278 /*
1279 * do this very late, it will log file closing as well
1280 */
1281 if (td->o.write_iolog_file)
1282 write_iolog_close(td);
1283
d23bb327 1284 options_mem_free(td);
ebac4655 1285 td_set_runstate(td, TD_EXITED);
43d76807 1286 return (void *) (unsigned long) td->error;
ebac4655
JA
1287}
1288
906c8d75
JA
1289/*
1290 * We cannot pass the td data into a forked process, so attach the td and
1291 * pass it to the thread worker.
1292 */
a6418147 1293static int fork_main(int shmid, int offset)
ebac4655
JA
1294{
1295 struct thread_data *td;
a6418147 1296 void *data, *ret;
ebac4655 1297
65258961 1298#ifndef __hpux
ebac4655
JA
1299 data = shmat(shmid, NULL, 0);
1300 if (data == (void *) -1) {
a6418147
JA
1301 int __err = errno;
1302
ebac4655 1303 perror("shmat");
a6418147 1304 return __err;
ebac4655 1305 }
65258961
JA
1306#else
1307 /*
1308 * HP-UX inherits shm mappings?
1309 */
1310 data = threads;
1311#endif
ebac4655
JA
1312
1313 td = data + offset * sizeof(struct thread_data);
a6418147 1314 ret = thread_main(td);
ebac4655 1315 shmdt(data);
43d76807 1316 return (int) (unsigned long) ret;
ebac4655
JA
1317}
1318
906c8d75
JA
1319/*
1320 * Run over the job map and reap the threads that have exited, if any.
1321 */
ebac4655
JA
1322static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1323{
34572e28 1324 struct thread_data *td;
1f809d15 1325 int i, cputhreads, realthreads, pending, status, ret;
ebac4655
JA
1326
1327 /*
1328 * reap exited threads (TD_EXITED -> TD_REAPED)
1329 */
1f809d15 1330 realthreads = pending = cputhreads = 0;
34572e28 1331 for_each_td(td, i) {
3707f45b 1332 int flags = 0;
a2f77c9f 1333
84585003
JA
1334 /*
1335 * ->io_ops is NULL for a thread that has closed its
1336 * io engine
1337 */
ba0fbe10 1338 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0 1339 cputhreads++;
1f809d15
JA
1340 else
1341 realthreads++;
b990b5c0 1342
a1dedc1f
AC
1343 if (!td->pid) {
1344 pending++;
1345 continue;
1346 }
1347 if (td->runstate == TD_REAPED)
a2f77c9f 1348 continue;
2dc1bbeb 1349 if (td->o.use_thread) {
3707f45b
JA
1350 if (td->runstate == TD_EXITED) {
1351 td_set_runstate(td, TD_REAPED);
1352 goto reaped;
1353 }
1354 continue;
1355 }
a2f77c9f
JA
1356
1357 flags = WNOHANG;
1358 if (td->runstate == TD_EXITED)
1359 flags = 0;
1360
1361 /*
1362 * check if someone quit or got killed in an unusual way
1363 */
1364 ret = waitpid(td->pid, &status, flags);
3707f45b 1365 if (ret < 0) {
a2f77c9f 1366 if (errno == ECHILD) {
5921e80c
JA
1367 log_err("fio: pid=%d disappeared %d\n",
1368 (int) td->pid, td->runstate);
a2f77c9f
JA
1369 td_set_runstate(td, TD_REAPED);
1370 goto reaped;
1371 }
1372 perror("waitpid");
1373 } else if (ret == td->pid) {
1374 if (WIFSIGNALED(status)) {
fab6aa71
JA
1375 int sig = WTERMSIG(status);
1376
03e20d68 1377 if (sig != SIGTERM)
5ec10eaa 1378 log_err("fio: pid=%d, got signal=%d\n",
5921e80c 1379 (int) td->pid, sig);
fab6aa71
JA
1380 td_set_runstate(td, TD_REAPED);
1381 goto reaped;
1382 }
a2f77c9f
JA
1383 if (WIFEXITED(status)) {
1384 if (WEXITSTATUS(status) && !td->error)
1385 td->error = WEXITSTATUS(status);
a2f77c9f 1386
a2f77c9f
JA
1387 td_set_runstate(td, TD_REAPED);
1388 goto reaped;
a6418147
JA
1389 }
1390 }
ebac4655 1391
a2f77c9f
JA
1392 /*
1393 * thread is not dead, continue
1394 */
e48676ba 1395 pending++;
a2f77c9f 1396 continue;
fab6aa71 1397reaped:
ebac4655 1398 (*nr_running)--;
581e7141
JA
1399 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1400 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
6f88bcb0
JA
1401 if (!td->pid)
1402 pending--;
a2f77c9f
JA
1403
1404 if (td->error)
1405 exit_value++;
d3eeeabc
JA
1406
1407 done_secs += mtime_since_now(&td->epoch) / 1000;
ebac4655 1408 }
b990b5c0 1409
1f809d15 1410 if (*nr_running == cputhreads && !pending && realthreads)
390c40e2 1411 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1412}
1413
be4ecfdf
JA
1414static void *gtod_thread_main(void *data)
1415{
1416 fio_mutex_up(startup_mutex);
1417
1418 /*
1419 * As long as we have jobs around, update the clock. It would be nice
1420 * to have some way of NOT hammering that CPU with gettimeofday(),
1421 * but I'm not sure what to use outside of a simple CPU nop to relax
1422 * it - we don't want to lose precision.
1423 */
1424 while (threads) {
1425 fio_gtod_update();
1426 nop;
1427 }
1428
1429 return NULL;
1430}
1431
1432static int fio_start_gtod_thread(void)
1433{
304a47c7 1434 pthread_attr_t attr;
f42c153d
JA
1435 int ret;
1436
304a47c7
VA
1437 pthread_attr_init(&attr);
1438 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1439 ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1440 pthread_attr_destroy(&attr);
f42c153d
JA
1441 if (ret) {
1442 log_err("Can't create gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1443 return 1;
1444 }
f42c153d
JA
1445
1446 ret = pthread_detach(gtod_thread);
1447 if (ret) {
1448 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1449 return 1;
1450 }
1451
29adda3c 1452 dprint(FD_MUTEX, "wait on startup_mutex\n");
be4ecfdf 1453 fio_mutex_down(startup_mutex);
29adda3c 1454 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
be4ecfdf
JA
1455 return 0;
1456}
1457
906c8d75
JA
1458/*
1459 * Main function for kicking off and reaping jobs, as needed.
1460 */
ebac4655
JA
1461static void run_threads(void)
1462{
ebac4655
JA
1463 struct thread_data *td;
1464 unsigned long spent;
1465 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1466
2f9ade3c
JA
1467 if (fio_pin_memory())
1468 return;
ebac4655 1469
be4ecfdf
JA
1470 if (fio_gtod_offload && fio_start_gtod_thread())
1471 return;
1472
c6ae0a5b 1473 if (!terse_output) {
4ceb30d4 1474 log_info("Starting ");
9cedf167 1475 if (nr_thread)
4ceb30d4 1476 log_info("%d thread%s", nr_thread,
5ec10eaa 1477 nr_thread > 1 ? "s" : "");
9cedf167
JA
1478 if (nr_process) {
1479 if (nr_thread)
1480 printf(" and ");
4ceb30d4 1481 log_info("%d process%s", nr_process,
5ec10eaa 1482 nr_process > 1 ? "es" : "");
9cedf167 1483 }
4ceb30d4 1484 log_info("\n");
c6ae0a5b
JA
1485 fflush(stdout);
1486 }
c04f7ec3 1487
697a606c 1488 set_sig_handlers();
4efa970e 1489
ebac4655
JA
1490 todo = thread_number;
1491 nr_running = 0;
1492 nr_started = 0;
1493 m_rate = t_rate = 0;
1494
34572e28 1495 for_each_td(td, i) {
263e529f 1496 print_status_init(td->thread_number - 1);
ebac4655 1497
2dc1bbeb 1498 if (!td->o.create_serialize) {
380cf265 1499 init_disk_util(td);
ebac4655 1500 continue;
380cf265 1501 }
ebac4655
JA
1502
1503 /*
1504 * do file setup here so it happens sequentially,
1505 * we don't want X number of threads getting their
1506 * client data interspersed on disk
1507 */
53cdc686 1508 if (setup_files(td)) {
5bf13a5a
JA
1509 exit_value++;
1510 if (td->error)
5921e80c
JA
1511 log_err("fio: pid=%d, err=%d/%s\n",
1512 (int) td->pid, td->error, td->verror);
ebac4655
JA
1513 td_set_runstate(td, TD_REAPED);
1514 todo--;
c69e8875
JA
1515 } else {
1516 struct fio_file *f;
2b13e716 1517 unsigned int j;
c69e8875
JA
1518
1519 /*
1520 * for sharing to work, each job must always open
1521 * its own files. so close them, if we opened them
1522 * for creation
1523 */
2b13e716 1524 for_each_file(td, f, j) {
22a57ba8
JA
1525 if (fio_file_open(f))
1526 td_io_close_file(td, f);
22a57ba8 1527 }
5ec10eaa 1528 }
380cf265
JA
1529
1530 init_disk_util(td);
ebac4655
JA
1531 }
1532
a2f77c9f
JA
1533 set_genesis_time();
1534
ebac4655 1535 while (todo) {
fca70358 1536 struct thread_data *map[REAL_MAX_JOBS];
75154845
JA
1537 struct timeval this_start;
1538 int this_jobs = 0, left;
1539
ebac4655
JA
1540 /*
1541 * create threads (TD_NOT_CREATED -> TD_CREATED)
1542 */
34572e28 1543 for_each_td(td, i) {
ebac4655
JA
1544 if (td->runstate != TD_NOT_CREATED)
1545 continue;
1546
1547 /*
1548 * never got a chance to start, killed by other
1549 * thread for some reason
1550 */
1551 if (td->terminate) {
1552 todo--;
1553 continue;
1554 }
1555
2dc1bbeb 1556 if (td->o.start_delay) {
263e529f 1557 spent = mtime_since_genesis();
ebac4655 1558
2dc1bbeb 1559 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1560 continue;
1561 }
1562
6f88bcb0
JA
1563 if (td->o.stonewall && (nr_started || nr_running)) {
1564 dprint(FD_PROCESS, "%s: stonewall wait\n",
1565 td->o.name);
ebac4655 1566 break;
6f88bcb0 1567 }
ebac4655 1568
75154845
JA
1569 /*
1570 * Set state to created. Thread will transition
1571 * to TD_INITIALIZED when it's done setting up.
1572 */
ebac4655 1573 td_set_runstate(td, TD_CREATED);
75154845 1574 map[this_jobs++] = td;
ebac4655
JA
1575 nr_started++;
1576
2dc1bbeb 1577 if (td->o.use_thread) {
f42c153d
JA
1578 int ret;
1579
ee56ad50 1580 dprint(FD_PROCESS, "will pthread_create\n");
f42c153d
JA
1581 ret = pthread_create(&td->thread, NULL,
1582 thread_main, td);
1583 if (ret) {
1584 log_err("pthread_create: %s\n",
1585 strerror(ret));
ebac4655 1586 nr_started--;
c8bb6faf 1587 break;
ebac4655 1588 }
f42c153d
JA
1589 ret = pthread_detach(td->thread);
1590 if (ret)
1591 log_err("pthread_detach: %s",
1592 strerror(ret));
ebac4655 1593 } else {
5e1d306e 1594 pid_t pid;
ee56ad50 1595 dprint(FD_PROCESS, "will fork\n");
5e1d306e
JA
1596 pid = fork();
1597 if (!pid) {
a6418147
JA
1598 int ret = fork_main(shm_id, i);
1599
5e1d306e
JA
1600 _exit(ret);
1601 } else if (i == fio_debug_jobno)
1602 *fio_debug_jobp = pid;
ebac4655 1603 }
29adda3c 1604 dprint(FD_MUTEX, "wait on startup_mutex\n");
656b1393
JA
1605 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1606 log_err("fio: job startup hung? exiting.\n");
1607 terminate_threads(TERMINATE_ALL);
1608 fio_abort = 1;
1609 nr_started--;
1610 break;
1611 }
29adda3c 1612 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
ebac4655
JA
1613 }
1614
1615 /*
75154845
JA
1616 * Wait for the started threads to transition to
1617 * TD_INITIALIZED.
ebac4655 1618 */
02bcaa8c 1619 fio_gettime(&this_start, NULL);
75154845 1620 left = this_jobs;
6ce15a32 1621 while (left && !fio_abort) {
75154845
JA
1622 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1623 break;
1624
1625 usleep(100000);
1626
1627 for (i = 0; i < this_jobs; i++) {
1628 td = map[i];
1629 if (!td)
1630 continue;
b6f4d880 1631 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1632 map[i] = NULL;
1633 left--;
b6f4d880
JA
1634 } else if (td->runstate >= TD_EXITED) {
1635 map[i] = NULL;
1636 left--;
1637 todo--;
1638 nr_running++; /* work-around... */
75154845
JA
1639 }
1640 }
1641 }
1642
1643 if (left) {
3b70d7e5 1644 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1645 for (i = 0; i < this_jobs; i++) {
1646 td = map[i];
1647 if (!td)
1648 continue;
1649 kill(td->pid, SIGTERM);
1650 }
1651 break;
1652 }
1653
1654 /*
b6f4d880 1655 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1656 */
34572e28 1657 for_each_td(td, i) {
75154845 1658 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1659 continue;
1660
b29ee5b3
JA
1661 if (in_ramp_time(td))
1662 td_set_runstate(td, TD_RAMP);
1663 else
1664 td_set_runstate(td, TD_RUNNING);
ebac4655
JA
1665 nr_running++;
1666 nr_started--;
581e7141
JA
1667 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1668 t_rate += td->o.rate[0] + td->o.rate[1];
75154845 1669 todo--;
cdd18ad8 1670 fio_mutex_up(td->mutex);
ebac4655
JA
1671 }
1672
1673 reap_threads(&nr_running, &t_rate, &m_rate);
1674
1675 if (todo)
1676 usleep(100000);
1677 }
1678
1679 while (nr_running) {
1680 reap_threads(&nr_running, &t_rate, &m_rate);
1681 usleep(10000);
1682 }
1683
1684 update_io_ticks();
2f9ade3c 1685 fio_unpin_memory();
ebac4655
JA
1686}
1687
ebac4655
JA
1688int main(int argc, char *argv[])
1689{
29d610e1
JA
1690 long ps;
1691
2e5cdb11 1692 sinit();
d54a144d 1693 init_rand(&__fio_rand_state);
2e5cdb11 1694
dbe1125e
JA
1695 /*
1696 * We need locale for number printing, if it isn't set then just
1697 * go with the US format.
1698 */
1699 if (!getenv("LC_NUMERIC"))
1700 setlocale(LC_NUMERIC, "en_US");
1701
29d610e1
JA
1702 ps = sysconf(_SC_PAGESIZE);
1703 if (ps < 0) {
1704 log_err("Failed to get page size\n");
1705 return 1;
1706 }
1707
cfc99db7 1708 page_size = ps;
29d610e1
JA
1709 page_mask = ps - 1;
1710
74929ac2
JA
1711 fio_keywords_init();
1712
1713 if (parse_options(argc, argv))
1714 return 1;
1715
07b3232d
JA
1716 if (exec_profile && load_profile(exec_profile))
1717 return 1;
1718
74929ac2
JA
1719 if (!thread_number)
1720 return 0;
1721
bb3884d8
JA
1722 if (write_bw_log) {
1723 setup_log(&agg_io_log[DDIR_READ]);
1724 setup_log(&agg_io_log[DDIR_WRITE]);
1725 }
1726
cdd18ad8 1727 startup_mutex = fio_mutex_init(0);
03e20d68
BC
1728 if (startup_mutex == NULL)
1729 return 1;
d2bbb8d9 1730 writeout_mutex = fio_mutex_init(1);
03e20d68
BC
1731 if (writeout_mutex == NULL)
1732 return 1;
07739b57 1733
a2f77c9f 1734 set_genesis_time();
7985701b 1735 create_disk_util_thread();
ebac4655 1736
dae5341c
JA
1737 cgroup_list = smalloc(sizeof(*cgroup_list));
1738 INIT_FLIST_HEAD(cgroup_list);
1739
ebac4655 1740 run_threads();
6ce15a32 1741
bb3884d8 1742 if (!fio_abort) {
6ce15a32 1743 show_run_stats();
bb3884d8 1744 if (write_bw_log) {
5ec10eaa
JA
1745 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1746 __finish_log(agg_io_log[DDIR_WRITE],
1747 "agg-write_bw.log");
bb3884d8
JA
1748 }
1749 }
ebac4655 1750
dae5341c
JA
1751 cgroup_kill(cgroup_list);
1752 sfree(cgroup_list);
61eb313e 1753 sfree(cgroup_mnt);
39f22027 1754
cdd18ad8 1755 fio_mutex_remove(startup_mutex);
d2bbb8d9 1756 fio_mutex_remove(writeout_mutex);
437c9b71 1757 return exit_value;
ebac4655 1758}