server: idle loop support
[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"
009b1be4 49#include "server.h"
ebac4655 50
cfc99db7
JA
51unsigned long page_mask;
52unsigned long page_size;
d529ee19
JA
53
54#define PAGE_ALIGN(buf) \
29d610e1 55 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
56
57int groupid = 0;
58int thread_number = 0;
9cedf167
JA
59int nr_process = 0;
60int nr_thread = 0;
ebac4655 61int shm_id = 0;
53cdc686 62int temp_stall_ts;
d3eeeabc 63unsigned long done_secs = 0;
ebac4655 64
cdd18ad8 65static struct fio_mutex *startup_mutex;
d2bbb8d9 66static struct fio_mutex *writeout_mutex;
6ce15a32 67static volatile int fio_abort;
437c9b71 68static int exit_value;
be4ecfdf 69static pthread_t gtod_thread;
7985701b 70static pthread_t disk_util_thread;
dae5341c 71static struct flist_head *cgroup_list;
6adb38a1 72static char *cgroup_mnt;
ebac4655 73
d3cc4ebf
JA
74unsigned long arch_flags = 0;
75
bb3884d8
JA
76struct io_log *agg_io_log[2];
77
75154845 78#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 79
b29ee5b3 80void td_set_runstate(struct thread_data *td, int runstate)
6ce15a32 81{
d8fab1b6
JA
82 if (td->runstate == runstate)
83 return;
84
5921e80c
JA
85 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
86 td->runstate, runstate);
6ce15a32
JA
87 td->runstate = runstate;
88}
89
cc0df00a 90void fio_terminate_threads(int group_id)
ebac4655 91{
34572e28 92 struct thread_data *td;
ebac4655
JA
93 int i;
94
6f88bcb0
JA
95 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
96
34572e28 97 for_each_td(td, i) {
ebac4655 98 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
6f88bcb0 99 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
5921e80c 100 td->o.name, (int) td->pid);
ad830ed7
JA
101 td->terminate = 1;
102 td->o.start_delay = 0;
103
7a93ab19
JA
104 /*
105 * if the thread is running, just let it exit
106 */
6aa27816
MC
107 if (!td->pid)
108 continue;
109 else if (td->runstate < TD_RAMP)
03e20d68 110 kill(td->pid, SIGTERM);
f63f7f3b
JA
111 else {
112 struct ioengine_ops *ops = td->io_ops;
113
03e20d68
BC
114 if (ops && (ops->flags & FIO_SIGTERM))
115 kill(td->pid, SIGTERM);
f63f7f3b 116 }
ebac4655
JA
117 }
118 }
119}
120
697a606c
JA
121static void sig_int(int sig)
122{
123 if (threads) {
4ceb30d4 124 log_info("\nfio: terminating on signal %d\n", sig);
009b1be4 125 exit_backend = 1;
5ec10eaa 126 fflush(stdout);
deaa1af2 127 exit_value = 128;
cc0df00a 128 fio_terminate_threads(TERMINATE_ALL);
ebac4655
JA
129 }
130}
131
7985701b 132static void *disk_thread_main(void *data)
a5b01f1b 133{
7985701b
JA
134 fio_mutex_up(startup_mutex);
135
136 while (threads) {
137 usleep(DISK_UTIL_MSEC * 1000);
9a793c2f
JA
138 if (!threads)
139 break;
7985701b 140 update_io_ticks();
cf451d1e
JA
141
142 if (is_backend)
143 fio_server_send_status();
144 else
145 print_thread_status();
7985701b
JA
146 }
147
148 return NULL;
a5b01f1b
JA
149}
150
7985701b 151static int create_disk_util_thread(void)
a5b01f1b 152{
7985701b 153 int ret;
a5b01f1b 154
69204d6e 155 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
7985701b
JA
156 if (ret) {
157 log_err("Can't create disk util thread: %s\n", strerror(ret));
158 return 1;
159 }
a5b01f1b 160
7985701b
JA
161 ret = pthread_detach(disk_util_thread);
162 if (ret) {
163 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
164 return 1;
165 }
03e20d68 166
7985701b
JA
167 dprint(FD_MUTEX, "wait on startup_mutex\n");
168 fio_mutex_down(startup_mutex);
169 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
170 return 0;
a5b01f1b
JA
171}
172
697a606c
JA
173static void set_sig_handlers(void)
174{
175 struct sigaction act;
176
697a606c
JA
177 memset(&act, 0, sizeof(act));
178 act.sa_handler = sig_int;
179 act.sa_flags = SA_RESTART;
180 sigaction(SIGINT, &act, NULL);
de79c915 181
5f8f9726 182 memset(&act, 0, sizeof(act));
b20b6be4 183 act.sa_handler = sig_int;
5f8f9726 184 act.sa_flags = SA_RESTART;
03e20d68 185 sigaction(SIGTERM, &act, NULL);
cc0df00a
JA
186
187 if (is_backend) {
188 memset(&act, 0, sizeof(act));
189 act.sa_handler = sig_int;
190 act.sa_flags = SA_RESTART;
191 sigaction(SIGPIPE, &act, NULL);
192 }
697a606c
JA
193}
194
906c8d75
JA
195/*
196 * Check if we are above the minimum rate given.
197 */
581e7141 198static int __check_min_rate(struct thread_data *td, struct timeval *now,
bd93e34b 199 enum fio_ddir ddir)
ebac4655 200{
0904200b 201 unsigned long long bytes = 0;
4e991c23 202 unsigned long iops = 0;
ebac4655
JA
203 unsigned long spent;
204 unsigned long rate;
581e7141
JA
205 unsigned int ratemin = 0;
206 unsigned int rate_iops = 0;
207 unsigned int rate_iops_min = 0;
ebac4655 208
ff58fced
JA
209 assert(ddir_rw(ddir));
210
d982d873
JA
211 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
212 return 0;
213
ebac4655
JA
214 /*
215 * allow a 2 second settle period in the beginning
216 */
217 if (mtime_since(&td->start, now) < 2000)
218 return 0;
219
581e7141
JA
220 iops += td->io_blocks[ddir];
221 bytes += td->this_io_bytes[ddir];
222 ratemin += td->o.ratemin[ddir];
223 rate_iops += td->o.rate_iops[ddir];
224 rate_iops_min += td->o.rate_iops_min[ddir];
0904200b 225
ebac4655
JA
226 /*
227 * if rate blocks is set, sample is running
228 */
581e7141
JA
229 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
230 spent = mtime_since(&td->lastrate[ddir], now);
2dc1bbeb 231 if (spent < td->o.ratecycle)
ebac4655
JA
232 return 0;
233
581e7141 234 if (td->o.rate[ddir]) {
4e991c23
JA
235 /*
236 * check bandwidth specified rate
237 */
581e7141 238 if (bytes < td->rate_bytes[ddir]) {
5ec10eaa 239 log_err("%s: min rate %u not met\n", td->o.name,
581e7141 240 ratemin);
4e991c23
JA
241 return 1;
242 } else {
581e7141
JA
243 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
244 if (rate < ratemin ||
245 bytes < td->rate_bytes[ddir]) {
5ec10eaa 246 log_err("%s: min rate %u not met, got"
b22989b9 247 " %luKB/sec\n", td->o.name,
581e7141 248 ratemin, rate);
4e991c23
JA
249 return 1;
250 }
251 }
413dd459 252 } else {
4e991c23
JA
253 /*
254 * checks iops specified rate
255 */
581e7141 256 if (iops < rate_iops) {
5ec10eaa 257 log_err("%s: min iops rate %u not met\n",
581e7141 258 td->o.name, rate_iops);
413dd459 259 return 1;
4e991c23 260 } else {
581e7141
JA
261 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
262 if (rate < rate_iops_min ||
263 iops < td->rate_blocks[ddir]) {
5ec10eaa
JA
264 log_err("%s: min iops rate %u not met,"
265 " got %lu\n", td->o.name,
581e7141 266 rate_iops_min, rate);
4e991c23 267 }
413dd459 268 }
ebac4655
JA
269 }
270 }
271
581e7141
JA
272 td->rate_bytes[ddir] = bytes;
273 td->rate_blocks[ddir] = iops;
274 memcpy(&td->lastrate[ddir], now, sizeof(*now));
ebac4655
JA
275 return 0;
276}
277
581e7141
JA
278static int check_min_rate(struct thread_data *td, struct timeval *now,
279 unsigned long *bytes_done)
280{
281 int ret = 0;
282
283 if (bytes_done[0])
284 ret |= __check_min_rate(td, now, 0);
285 if (bytes_done[1])
286 ret |= __check_min_rate(td, now, 1);
287
288 return ret;
289}
290
ebac4655
JA
291static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
292{
2dc1bbeb 293 if (!td->o.timeout)
ebac4655 294 return 0;
2dc1bbeb 295 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
ebac4655
JA
296 return 1;
297
298 return 0;
299}
300
906c8d75
JA
301/*
302 * When job exits, we can cancel the in-flight IO if we are using async
303 * io. Attempt to do so.
304 */
ebac4655
JA
305static void cleanup_pending_aio(struct thread_data *td)
306{
01743ee1 307 struct flist_head *entry, *n;
ebac4655
JA
308 struct io_u *io_u;
309 int r;
310
311 /*
312 * get immediately available events, if any
313 */
581e7141 314 r = io_u_queued_complete(td, 0, NULL);
b2fdda43
JA
315 if (r < 0)
316 return;
ebac4655
JA
317
318 /*
319 * now cancel remaining active events
320 */
2866c82d 321 if (td->io_ops->cancel) {
01743ee1
JA
322 flist_for_each_safe(entry, n, &td->io_u_busylist) {
323 io_u = flist_entry(entry, struct io_u, list);
ebac4655 324
0c6e7517
JA
325 /*
326 * if the io_u isn't in flight, then that generally
327 * means someone leaked an io_u. complain but fix
328 * it up, so we don't stall here.
329 */
330 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
331 log_err("fio: non-busy IO on busy list\n");
ebac4655 332 put_io_u(td, io_u);
0c6e7517
JA
333 } else {
334 r = td->io_ops->cancel(td, io_u);
335 if (!r)
336 put_io_u(td, io_u);
337 }
ebac4655
JA
338 }
339 }
340
97601024 341 if (td->cur_depth)
581e7141 342 r = io_u_queued_complete(td, td->cur_depth, NULL);
ebac4655
JA
343}
344
858a3d47
JA
345/*
346 * Helper to handle the final sync of a file. Works just like the normal
347 * io path, just does everything sync.
348 */
349static int fio_io_sync(struct thread_data *td, struct fio_file *f)
350{
351 struct io_u *io_u = __get_io_u(td);
858a3d47
JA
352 int ret;
353
354 if (!io_u)
355 return 1;
356
357 io_u->ddir = DDIR_SYNC;
358 io_u->file = f;
359
360 if (td_io_prep(td, io_u)) {
361 put_io_u(td, io_u);
362 return 1;
363 }
364
755200a3 365requeue:
858a3d47 366 ret = td_io_queue(td, io_u);
36167d82 367 if (ret < 0) {
e1161c32 368 td_verror(td, io_u->error, "td_io_queue");
858a3d47 369 put_io_u(td, io_u);
858a3d47 370 return 1;
36167d82 371 } else if (ret == FIO_Q_QUEUED) {
581e7141 372 if (io_u_queued_complete(td, 1, NULL) < 0)
36167d82 373 return 1;
36167d82
JA
374 } else if (ret == FIO_Q_COMPLETED) {
375 if (io_u->error) {
e1161c32 376 td_verror(td, io_u->error, "td_io_queue");
36167d82
JA
377 return 1;
378 }
858a3d47 379
581e7141 380 if (io_u_sync_complete(td, io_u, NULL) < 0)
b2fdda43 381 return 1;
755200a3
JA
382 } else if (ret == FIO_Q_BUSY) {
383 if (td_io_commit(td))
384 return 1;
385 goto requeue;
858a3d47
JA
386 }
387
388 return 0;
389}
390
687c6a23
JA
391static inline void __update_tv_cache(struct thread_data *td)
392{
393 fio_gettime(&td->tv_cache, NULL);
394}
395
993bf48b
JA
396static inline void update_tv_cache(struct thread_data *td)
397{
398 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
687c6a23 399 __update_tv_cache(td);
993bf48b
JA
400}
401
f2bba182
RR
402static int break_on_this_error(struct thread_data *td, int *retptr)
403{
404 int ret = *retptr;
405
406 if (ret < 0 || td->error) {
407 int err;
408
1ec99eea
JA
409 if (!td->o.continue_on_error)
410 return 1;
f2bba182
RR
411
412 if (ret < 0)
413 err = -ret;
414 else
415 err = td->error;
416
f2bba182
RR
417 if (td_non_fatal_error(err)) {
418 /*
419 * Continue with the I/Os in case of
420 * a non fatal error.
421 */
8a4f8a24 422 update_error_count(td, err);
f2bba182
RR
423 td_clear_error(td);
424 *retptr = 0;
425 return 0;
8a4f8a24
RR
426 } else if (td->o.fill_device && err == ENOSPC) {
427 /*
428 * We expect to hit this error if
429 * fill_device option is set.
430 */
431 td_clear_error(td);
432 td->terminate = 1;
433 return 1;
f2bba182
RR
434 } else {
435 /*
436 * Stop the I/O in case of a fatal
437 * error.
438 */
8a4f8a24 439 update_error_count(td, err);
f2bba182
RR
440 return 1;
441 }
442 }
443
444 return 0;
445}
446
906c8d75 447/*
34403fb1 448 * The main verify engine. Runs over the writes we previously submitted,
906c8d75
JA
449 * reads the blocks back in, and checks the crc/md5 of the data.
450 */
1e97cce9 451static void do_verify(struct thread_data *td)
ebac4655 452{
53cdc686 453 struct fio_file *f;
36167d82 454 struct io_u *io_u;
af52b345
JA
455 int ret, min_events;
456 unsigned int i;
e5b401d4 457
008618a8
JA
458 dprint(FD_VERIFY, "starting loop\n");
459
e5b401d4
JA
460 /*
461 * sync io first and invalidate cache, to make sure we really
462 * read from disk.
463 */
464 for_each_file(td, f, i) {
d6aed795 465 if (!fio_file_open(f))
3d7b485f 466 continue;
b2fdda43
JA
467 if (fio_io_sync(td, f))
468 break;
469 if (file_invalidate_cache(td, f))
470 break;
e5b401d4 471 }
ebac4655 472
b2fdda43
JA
473 if (td->error)
474 return;
475
ebac4655
JA
476 td_set_runstate(td, TD_VERIFYING);
477
36167d82
JA
478 io_u = NULL;
479 while (!td->terminate) {
4950421a 480 int ret2, full;
5451792e 481
993bf48b
JA
482 update_tv_cache(td);
483
484 if (runtime_exceeded(td, &td->tv_cache)) {
687c6a23
JA
485 __update_tv_cache(td);
486 if (runtime_exceeded(td, &td->tv_cache)) {
487 td->terminate = 1;
488 break;
489 }
069c2918 490 }
02bcaa8c 491
dcb4830d
RR
492 io_u = __get_io_u(td);
493 if (!io_u)
494 break;
495
069c2918
JA
496 if (get_next_verify(td, io_u)) {
497 put_io_u(td, io_u);
ebac4655 498 break;
069c2918 499 }
ebac4655 500
069c2918
JA
501 if (td_io_prep(td, io_u)) {
502 put_io_u(td, io_u);
53cdc686 503 break;
069c2918 504 }
d7762cf8 505
e8462bd8
JA
506 if (td->o.verify_async)
507 io_u->end_io = verify_io_u_async;
508 else
509 io_u->end_io = verify_io_u;
53cdc686 510
11786802 511 ret = td_io_queue(td, io_u);
36167d82
JA
512 switch (ret) {
513 case FIO_Q_COMPLETED:
f2bba182 514 if (io_u->error) {
22819ec2 515 ret = -io_u->error;
f2bba182
RR
516 clear_io_u(td, io_u);
517 } else if (io_u->resid) {
36167d82 518 int bytes = io_u->xfer_buflen - io_u->resid;
ebac4655 519
9e9d164e
JA
520 /*
521 * zero read, fail
522 */
523 if (!bytes) {
41d9f0fd 524 td_verror(td, EIO, "full resid");
9e9d164e
JA
525 put_io_u(td, io_u);
526 break;
527 }
8400d9b2 528
36167d82
JA
529 io_u->xfer_buflen = io_u->resid;
530 io_u->xfer_buf += bytes;
8400d9b2
JA
531 io_u->offset += bytes;
532
ff58fced
JA
533 if (ddir_rw(io_u->ddir))
534 td->ts.short_io_u[io_u->ddir]++;
30061b97 535
2b13e716 536 f = io_u->file;
d460eb31 537 if (io_u->offset == f->real_file_size)
8400d9b2
JA
538 goto sync_done;
539
11786802
JA
540 requeue_io_u(td, &io_u);
541 } else {
8400d9b2 542sync_done:
581e7141 543 ret = io_u_sync_complete(td, io_u, NULL);
11786802
JA
544 if (ret < 0)
545 break;
36167d82 546 }
36167d82
JA
547 continue;
548 case FIO_Q_QUEUED:
549 break;
755200a3
JA
550 case FIO_Q_BUSY:
551 requeue_io_u(td, &io_u);
5451792e
JA
552 ret2 = td_io_commit(td);
553 if (ret2 < 0)
554 ret = ret2;
755200a3 555 break;
36167d82
JA
556 default:
557 assert(ret < 0);
e1161c32 558 td_verror(td, -ret, "td_io_queue");
ebac4655
JA
559 break;
560 }
561
f2bba182 562 if (break_on_this_error(td, &ret))
3af6ef39
JA
563 break;
564
ebac4655 565 /*
3af6ef39 566 * if we can queue more, do so. but check if there are
ccc02050
JA
567 * completed io_u's first. Note that we can get BUSY even
568 * without IO queued, if the system is resource starved.
ebac4655 569 */
ccc02050 570 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
4950421a 571 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
572 min_events = min(td->o.iodepth_batch_complete,
573 td->cur_depth);
1f84f1d9 574 if (full && !min_events && td->o.iodepth_batch_complete != 0)
838bc709 575 min_events = 1;
e916b390 576
4950421a
JA
577 do {
578 /*
579 * Reap required number of io units, if any,
580 * and do the verification on them through
581 * the callback handler
582 */
581e7141 583 if (io_u_queued_complete(td, min_events, NULL) < 0) {
4950421a
JA
584 ret = -1;
585 break;
586 }
587 } while (full && (td->cur_depth > td->o.iodepth_low));
588 }
589 if (ret < 0)
ebac4655 590 break;
36167d82 591 }
ebac4655 592
c01c0395
JA
593 if (!td->error) {
594 min_events = td->cur_depth;
595
596 if (min_events)
581e7141 597 ret = io_u_queued_complete(td, min_events, NULL);
c01c0395 598 } else
ebac4655
JA
599 cleanup_pending_aio(td);
600
601 td_set_runstate(td, TD_RUNNING);
008618a8
JA
602
603 dprint(FD_VERIFY, "exiting loop\n");
ebac4655
JA
604}
605
32cd46a0 606/*
906c8d75 607 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
608 * and reaps them, checking for rate and errors along the way.
609 */
ebac4655
JA
610static void do_io(struct thread_data *td)
611{
af52b345
JA
612 unsigned int i;
613 int ret = 0;
ebac4655 614
b29ee5b3
JA
615 if (in_ramp_time(td))
616 td_set_runstate(td, TD_RAMP);
617 else
618 td_set_runstate(td, TD_RUNNING);
5853e5a8 619
457bf399 620 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
a917a8b3 621 (!flist_empty(&td->trim_list)) ||
457bf399 622 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
97601024 623 struct timeval comp_time;
581e7141 624 unsigned long bytes_done[2] = { 0, 0 };
84585003 625 int min_evts = 0;
ebac4655 626 struct io_u *io_u;
4950421a 627 int ret2, full;
ebac4655
JA
628
629 if (td->terminate)
630 break;
631
993bf48b 632 update_tv_cache(td);
97601024 633
993bf48b 634 if (runtime_exceeded(td, &td->tv_cache)) {
687c6a23
JA
635 __update_tv_cache(td);
636 if (runtime_exceeded(td, &td->tv_cache)) {
637 td->terminate = 1;
638 break;
639 }
97601024 640 }
36167d82 641
dcb4830d
RR
642 io_u = get_io_u(td);
643 if (!io_u)
644 break;
645
b67740d3
JA
646 /*
647 * Add verification end_io handler, if asked to verify
648 * a previously written file.
649 */
715d2b3e
JA
650 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
651 !td_rw(td)) {
e8462bd8
JA
652 if (td->o.verify_async)
653 io_u->end_io = verify_io_u_async;
654 else
655 io_u->end_io = verify_io_u;
d8fab1b6 656 td_set_runstate(td, TD_VERIFYING);
b29ee5b3
JA
657 } else if (in_ramp_time(td))
658 td_set_runstate(td, TD_RAMP);
659 else
d8fab1b6 660 td_set_runstate(td, TD_RUNNING);
b67740d3 661
11786802 662 ret = td_io_queue(td, io_u);
36167d82
JA
663 switch (ret) {
664 case FIO_Q_COMPLETED:
f2bba182 665 if (io_u->error) {
5451792e 666 ret = -io_u->error;
f2bba182
RR
667 clear_io_u(td, io_u);
668 } else if (io_u->resid) {
36167d82 669 int bytes = io_u->xfer_buflen - io_u->resid;
d460eb31 670 struct fio_file *f = io_u->file;
36167d82 671
9e9d164e
JA
672 /*
673 * zero read, fail
674 */
675 if (!bytes) {
41d9f0fd 676 td_verror(td, EIO, "full resid");
9e9d164e
JA
677 put_io_u(td, io_u);
678 break;
679 }
680
cec6b55d 681 io_u->xfer_buflen = io_u->resid;
36167d82 682 io_u->xfer_buf += bytes;
5a7c5680
JA
683 io_u->offset += bytes;
684
ff58fced
JA
685 if (ddir_rw(io_u->ddir))
686 td->ts.short_io_u[io_u->ddir]++;
30061b97 687
d460eb31 688 if (io_u->offset == f->real_file_size)
5a7c5680
JA
689 goto sync_done;
690
11786802
JA
691 requeue_io_u(td, &io_u);
692 } else {
5a7c5680 693sync_done:
581e7141
JA
694 if (__should_check_rate(td, 0) ||
695 __should_check_rate(td, 1))
993bf48b
JA
696 fio_gettime(&comp_time, NULL);
697
581e7141
JA
698 ret = io_u_sync_complete(td, io_u, bytes_done);
699 if (ret < 0)
700 break;
cec6b55d 701 }
36167d82
JA
702 break;
703 case FIO_Q_QUEUED:
7e77dd02
JA
704 /*
705 * if the engine doesn't have a commit hook,
706 * the io_u is really queued. if it does have such
707 * a hook, it has to call io_u_queued() itself.
708 */
709 if (td->io_ops->commit == NULL)
710 io_u_queued(td, io_u);
36167d82 711 break;
755200a3
JA
712 case FIO_Q_BUSY:
713 requeue_io_u(td, &io_u);
5451792e
JA
714 ret2 = td_io_commit(td);
715 if (ret2 < 0)
716 ret = ret2;
755200a3 717 break;
36167d82
JA
718 default:
719 assert(ret < 0);
720 put_io_u(td, io_u);
721 break;
ebac4655
JA
722 }
723
f2bba182 724 if (break_on_this_error(td, &ret))
36167d82
JA
725 break;
726
97601024 727 /*
ccc02050
JA
728 * See if we need to complete some commands. Note that we
729 * can get BUSY even without IO queued, if the system is
730 * resource starved.
97601024 731 */
ccc02050 732 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
4950421a 733 if (full || !td->o.iodepth_batch_complete) {
6a96276c
RR
734 min_evts = min(td->o.iodepth_batch_complete,
735 td->cur_depth);
1f84f1d9 736 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
36167d82 737 min_evts = 1;
4950421a 738
581e7141
JA
739 if (__should_check_rate(td, 0) ||
740 __should_check_rate(td, 1))
993bf48b 741 fio_gettime(&comp_time, NULL);
4950421a
JA
742
743 do {
581e7141
JA
744 ret = io_u_queued_complete(td, min_evts, bytes_done);
745 if (ret < 0)
4950421a
JA
746 break;
747
4950421a 748 } while (full && (td->cur_depth > td->o.iodepth_low));
ebac4655
JA
749 }
750
4950421a
JA
751 if (ret < 0)
752 break;
581e7141 753 if (!(bytes_done[0] + bytes_done[1]))
97601024
JA
754 continue;
755
581e7141
JA
756 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
757 if (check_min_rate(td, &comp_time, bytes_done)) {
721938ae 758 if (exitall_on_terminate)
cc0df00a 759 fio_terminate_threads(td->groupid);
721938ae
JA
760 td_verror(td, EIO, "check_min_rate");
761 break;
762 }
ebac4655
JA
763 }
764
2dc1bbeb 765 if (td->o.thinktime) {
9c1f7434
JA
766 unsigned long long b;
767
768 b = td->io_blocks[0] + td->io_blocks[1];
2dc1bbeb 769 if (!(b % td->o.thinktime_blocks)) {
48097d5c
JA
770 int left;
771
2dc1bbeb 772 if (td->o.thinktime_spin)
6dd6f2cd 773 usec_spin(td->o.thinktime_spin);
48097d5c 774
2dc1bbeb 775 left = td->o.thinktime - td->o.thinktime_spin;
48097d5c
JA
776 if (left)
777 usec_sleep(td, left);
778 }
9c1f7434 779 }
ebac4655
JA
780 }
781
d708ac5e
JA
782 if (td->trim_entries)
783 printf("trim entries %ld\n", td->trim_entries);
a917a8b3 784
aa31f1f1
SL
785 if (td->o.fill_device && td->error == ENOSPC) {
786 td->error = 0;
787 td->terminate = 1;
788 }
4d2413c6 789 if (!td->error) {
3d7c391d
JA
790 struct fio_file *f;
791
c01c0395 792 i = td->cur_depth;
d06093e7 793 if (i) {
581e7141 794 ret = io_u_queued_complete(td, i, NULL);
d06093e7
SL
795 if (td->o.fill_device && td->error == ENOSPC)
796 td->error = 0;
797 }
ebac4655 798
2dc1bbeb 799 if (should_fsync(td) && td->o.end_fsync) {
84585003 800 td_set_runstate(td, TD_FSYNCING);
3d7b485f
JA
801
802 for_each_file(td, f, i) {
d6aed795 803 if (!fio_file_open(f))
3d7b485f 804 continue;
858a3d47 805 fio_io_sync(td, f);
3d7b485f 806 }
84585003 807 }
c01c0395
JA
808 } else
809 cleanup_pending_aio(td);
2ba1c290
JA
810
811 /*
812 * stop job if we failed doing any IO
813 */
814 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
815 td->done = 1;
ebac4655
JA
816}
817
ebac4655
JA
818static void cleanup_io_u(struct thread_data *td)
819{
01743ee1 820 struct flist_head *entry, *n;
ebac4655
JA
821 struct io_u *io_u;
822
01743ee1
JA
823 flist_for_each_safe(entry, n, &td->io_u_freelist) {
824 io_u = flist_entry(entry, struct io_u, list);
ebac4655 825
01743ee1 826 flist_del(&io_u->list);
91aea6ec 827 fio_memfree(io_u, sizeof(*io_u));
ebac4655
JA
828 }
829
2f9ade3c 830 free_io_mem(td);
ebac4655
JA
831}
832
833static int init_io_u(struct thread_data *td)
834{
835 struct io_u *io_u;
a00735e6 836 unsigned int max_bs;
eb7ccf38 837 int cl_align, i, max_units;
ebac4655
JA
838 char *p;
839
1e3c09af 840 max_units = td->o.iodepth;
2dc1bbeb 841 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
5ec10eaa
JA
842 td->orig_buffer_size = (unsigned long long) max_bs
843 * (unsigned long long) max_units;
844
845 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
846 unsigned long bs;
74b025b0 847
5ec10eaa
JA
848 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
849 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
850 }
ebac4655 851
c3990645
JA
852 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
853 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
854 return 1;
855 }
856
2f9ade3c
JA
857 if (allocate_io_mem(td))
858 return 1;
ebac4655 859
1c04bf78
JA
860 if (td->o.odirect || td->o.mem_align ||
861 (td->io_ops->flags & FIO_RAWIO))
d529ee19 862 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
b3f378e9
JA
863 else
864 p = td->orig_buffer;
865
eb7ccf38
JA
866 cl_align = os_cache_line_size();
867
ebac4655 868 for (i = 0; i < max_units; i++) {
eb7ccf38
JA
869 void *ptr;
870
db1cd90b
JA
871 if (td->terminate)
872 return 1;
eb7ccf38 873
91aea6ec
JA
874 ptr = fio_memalign(cl_align, sizeof(*io_u));
875 if (!ptr) {
876 log_err("fio: unable to allocate aligned memory\n");
eb7ccf38
JA
877 break;
878 }
879
880 io_u = ptr;
ebac4655 881 memset(io_u, 0, sizeof(*io_u));
01743ee1 882 INIT_FLIST_HEAD(&io_u->list);
d529ee19 883 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
ebac4655 884
b4c5e1ac
JA
885 if (!(td->io_ops->flags & FIO_NOIO)) {
886 io_u->buf = p + max_bs * i;
d529ee19 887 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
e9459e5a 888
3545a109 889 if (td_write(td))
5973cafb 890 io_u_fill_buffer(td, io_u, max_bs);
3545a109 891 if (td_write(td) && td->o.verify_pattern_bytes) {
cbe8d756
RR
892 /*
893 * Fill the buffer with the pattern if we are
894 * going to be doing writes.
895 */
7d9fb455 896 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
cbe8d756 897 }
b4c5e1ac 898 }
6b9cea23 899
b1ff3403 900 io_u->index = i;
0c6e7517 901 io_u->flags = IO_U_F_FREE;
01743ee1 902 flist_add(&io_u->list, &td->io_u_freelist);
ebac4655
JA
903 }
904
905 return 0;
906}
907
da86774e
JA
908static int switch_ioscheduler(struct thread_data *td)
909{
910 char tmp[256], tmp2[128];
911 FILE *f;
912 int ret;
913
ba0fbe10 914 if (td->io_ops->flags & FIO_DISKLESSIO)
f48b467c
JA
915 return 0;
916
da86774e
JA
917 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
918
919 f = fopen(tmp, "r+");
920 if (!f) {
cbf5c1d7 921 if (errno == ENOENT) {
5ec10eaa
JA
922 log_err("fio: os or kernel doesn't support IO scheduler"
923 " switching\n");
cbf5c1d7
JA
924 return 0;
925 }
926 td_verror(td, errno, "fopen iosched");
da86774e
JA
927 return 1;
928 }
929
930 /*
931 * Set io scheduler.
932 */
2dc1bbeb 933 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
da86774e 934 if (ferror(f) || ret != 1) {
e1161c32 935 td_verror(td, errno, "fwrite");
da86774e
JA
936 fclose(f);
937 return 1;
938 }
939
940 rewind(f);
941
942 /*
943 * Read back and check that the selected scheduler is now the default.
944 */
945 ret = fread(tmp, 1, sizeof(tmp), f);
946 if (ferror(f) || ret < 0) {
e1161c32 947 td_verror(td, errno, "fread");
da86774e
JA
948 fclose(f);
949 return 1;
950 }
951
2dc1bbeb 952 sprintf(tmp2, "[%s]", td->o.ioscheduler);
da86774e 953 if (!strstr(tmp, tmp2)) {
2dc1bbeb 954 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
e1161c32 955 td_verror(td, EINVAL, "iosched_switch");
da86774e
JA
956 fclose(f);
957 return 1;
958 }
959
960 fclose(f);
961 return 0;
962}
963
492158cf
JA
964static int keep_running(struct thread_data *td)
965{
966 unsigned long long io_done;
967
20e354ef
JA
968 if (td->done)
969 return 0;
492158cf
JA
970 if (td->o.time_based)
971 return 1;
972 if (td->o.loops) {
973 td->o.loops--;
974 return 1;
975 }
976
5ec10eaa
JA
977 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
978 + td->io_skip_bytes;
492158cf
JA
979 if (io_done < td->o.size)
980 return 1;
981
982 return 0;
983}
984
b29ee5b3 985static void reset_io_counters(struct thread_data *td)
ebac4655 986{
f0505a14 987 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
ebac4655 988 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 989 td->zone_bytes = 0;
581e7141
JA
990 td->rate_bytes[0] = td->rate_bytes[1] = 0;
991 td->rate_blocks[0] = td->rate_blocks[1] = 0;
ebac4655 992
c1324df1
JA
993 td->last_was_sync = 0;
994
2ba1c290
JA
995 /*
996 * reset file done count if we are to start over
997 */
998 if (td->o.time_based || td->o.loops)
48f5abd3 999 td->nr_done_files = 0;
b29ee5b3
JA
1000}
1001
1002void reset_all_stats(struct thread_data *td)
1003{
1004 struct timeval tv;
1005 int i;
1006
1007 reset_io_counters(td);
1008
1009 for (i = 0; i < 2; i++) {
1010 td->io_bytes[i] = 0;
1011 td->io_blocks[i] = 0;
1012 td->io_issues[i] = 0;
1013 td->ts.total_io_u[i] = 0;
1014 }
9b836561 1015
b29ee5b3 1016 fio_gettime(&tv, NULL);
cda99fa0
SSP
1017 td->ts.runtime[0] = 0;
1018 td->ts.runtime[1] = 0;
b29ee5b3
JA
1019 memcpy(&td->epoch, &tv, sizeof(tv));
1020 memcpy(&td->start, &tv, sizeof(tv));
1021}
1022
df9c26b1 1023static void clear_io_state(struct thread_data *td)
b29ee5b3
JA
1024{
1025 struct fio_file *f;
1026 unsigned int i;
b29ee5b3
JA
1027
1028 reset_io_counters(td);
281f9b63 1029
24ffd2c2 1030 close_files(td);
df9c26b1 1031 for_each_file(td, f, i)
d6aed795 1032 fio_file_clear_done(f);
f0153948
JA
1033
1034 /*
1035 * Set the same seed to get repeatable runs
1036 */
1037 td_fill_rand_seeds(td);
ebac4655
JA
1038}
1039
398e8c4d
JA
1040static int exec_string(const char *string)
1041{
1042 int ret, newlen = strlen(string) + 1 + 8;
1043 char *str;
1044
1045 str = malloc(newlen);
1046 sprintf(str, "sh -c %s", string);
1047
1048 ret = system(str);
1049 if (ret == -1)
1050 log_err("fio: exec of cmd <%s> failed\n", str);
1051
1052 free(str);
1053 return ret;
1054}
1055
906c8d75
JA
1056/*
1057 * Entry point for the thread based jobs. The process based jobs end up
1058 * here as well, after a little setup.
1059 */
ebac4655
JA
1060static void *thread_main(void *data)
1061{
cda99fa0 1062 unsigned long long elapsed;
ebac4655 1063 struct thread_data *td = data;
e8462bd8 1064 pthread_condattr_t attr;
a978ba68 1065 int clear_state;
ebac4655 1066
47f767c1 1067 if (!td->o.use_thread) {
ebac4655 1068 setsid();
47f767c1
JA
1069 td->pid = getpid();
1070 } else
1071 td->pid = gettid();
ebac4655 1072
5921e80c 1073 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
ee56ad50 1074
01743ee1
JA
1075 INIT_FLIST_HEAD(&td->io_u_freelist);
1076 INIT_FLIST_HEAD(&td->io_u_busylist);
1077 INIT_FLIST_HEAD(&td->io_u_requeues);
1078 INIT_FLIST_HEAD(&td->io_log_list);
1079 INIT_FLIST_HEAD(&td->io_hist_list);
e8462bd8 1080 INIT_FLIST_HEAD(&td->verify_list);
0d29de83 1081 INIT_FLIST_HEAD(&td->trim_list);
e8462bd8 1082 pthread_mutex_init(&td->io_u_lock, NULL);
bb5d7d0b 1083 td->io_hist_tree = RB_ROOT;
aea47d44 1084
e8462bd8
JA
1085 pthread_condattr_init(&attr);
1086 pthread_cond_init(&td->verify_cond, &attr);
1087 pthread_cond_init(&td->free_cond, &attr);
1088
db1cd90b 1089 td_set_runstate(td, TD_INITIALIZED);
29adda3c 1090 dprint(FD_MUTEX, "up startup_mutex\n");
cdd18ad8 1091 fio_mutex_up(startup_mutex);
29adda3c 1092 dprint(FD_MUTEX, "wait on td->mutex\n");
cdd18ad8 1093 fio_mutex_down(td->mutex);
29adda3c 1094 dprint(FD_MUTEX, "done waiting on td->mutex\n");
db1cd90b
JA
1095
1096 /*
cdd18ad8 1097 * the ->mutex mutex is now no longer used, close it to avoid
db1cd90b
JA
1098 * eating a file descriptor
1099 */
cdd18ad8 1100 fio_mutex_remove(td->mutex);
db1cd90b 1101
dfc6e751
SH
1102 /*
1103 * A new gid requires privilege, so we need to do this before setting
1104 * the uid.
1105 */
e0b0d892
JA
1106 if (td->o.gid != -1U && setgid(td->o.gid)) {
1107 td_verror(td, errno, "setgid");
1108 goto err;
1109 }
dfc6e751
SH
1110 if (td->o.uid != -1U && setuid(td->o.uid)) {
1111 td_verror(td, errno, "setuid");
1112 goto err;
1113 }
e0b0d892 1114
be18a94f
JA
1115 /*
1116 * If we have a gettimeofday() thread, make sure we exclude that
1117 * thread from this job
1118 */
1119 if (td->o.gtod_cpu)
1120 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1121
1122 /*
1123 * Set affinity first, in case it has an impact on the memory
1124 * allocations.
1125 */
1126 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1127 td_verror(td, errno, "cpu_set_affinity");
1128 goto err;
1129 }
1130
d84f8d49
JA
1131 /*
1132 * May alter parameters that init_io_u() will use, so we need to
1133 * do this first.
1134 */
1135 if (init_iolog(td))
1136 goto err;
1137
ebac4655 1138 if (init_io_u(td))
db1cd90b 1139 goto err;
ebac4655 1140
e8462bd8
JA
1141 if (td->o.verify_async && verify_async_init(td))
1142 goto err;
1143
ac684785 1144 if (td->ioprio_set) {
ebac4655 1145 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
e1161c32 1146 td_verror(td, errno, "ioprio_set");
db1cd90b 1147 goto err;
ebac4655
JA
1148 }
1149 }
1150
6adb38a1 1151 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
a696fa2a
JA
1152 goto err;
1153
2dc1bbeb 1154 if (nice(td->o.nice) == -1) {
e1161c32 1155 td_verror(td, errno, "nice");
db1cd90b 1156 goto err;
b6f4d880
JA
1157 }
1158
2dc1bbeb 1159 if (td->o.ioscheduler && switch_ioscheduler(td))
db1cd90b 1160 goto err;
37f56873 1161
2dc1bbeb 1162 if (!td->o.create_serialize && setup_files(td))
ebac4655
JA
1163 goto err;
1164
7d6c5283
JA
1165 if (td_io_init(td))
1166 goto err;
1167
68727076
JA
1168 if (init_random_map(td))
1169 goto err;
1170
2dc1bbeb 1171 if (td->o.exec_prerun) {
398e8c4d 1172 if (exec_string(td->o.exec_prerun))
69cfd7e0
JA
1173 goto err;
1174 }
4e0ba8af 1175
afad68f7
ZY
1176 if (td->o.pre_read) {
1177 if (pre_read_files(td) < 0)
1178 goto err;
1179 }
1180
69008999 1181 fio_gettime(&td->epoch, NULL);
c8aaba19 1182 getrusage(RUSAGE_SELF, &td->ru_start);
69008999 1183
a978ba68 1184 clear_state = 0;
492158cf 1185 while (keep_running(td)) {
02bcaa8c 1186 fio_gettime(&td->start, NULL);
f0505a14
JA
1187 memcpy(&td->stat_sample_time[0], &td->start, sizeof(td->start));
1188 memcpy(&td->stat_sample_time[1], &td->start, sizeof(td->start));
993bf48b 1189 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
ebac4655 1190
cf2dd8d6 1191 if (td->o.ratemin[0] || td->o.ratemin[1])
f0505a14 1192 memcpy(&td->lastrate, &td->stat_sample_time,
5ec10eaa 1193 sizeof(td->lastrate));
ebac4655 1194
df9c26b1
JA
1195 if (clear_state)
1196 clear_io_state(td);
a978ba68 1197
ebac4655
JA
1198 prune_io_piece_log(td);
1199
ba0fbe10 1200 do_io(td);
ebac4655 1201
a978ba68
JA
1202 clear_state = 1;
1203
38d77cae 1204 if (td_read(td) && td->io_bytes[DDIR_READ]) {
0a4957c6 1205 elapsed = utime_since_now(&td->start);
cda99fa0 1206 td->ts.runtime[DDIR_READ] += elapsed;
38d77cae
JA
1207 }
1208 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
0a4957c6 1209 elapsed = utime_since_now(&td->start);
cda99fa0 1210 td->ts.runtime[DDIR_WRITE] += elapsed;
38d77cae 1211 }
5ec10eaa 1212
ebac4655
JA
1213 if (td->error || td->terminate)
1214 break;
1215
e84c73a8
SL
1216 if (!td->o.do_verify ||
1217 td->o.verify == VERIFY_NONE ||
b67740d3 1218 (td->io_ops->flags & FIO_UNIDIR))
ebac4655
JA
1219 continue;
1220
df9c26b1 1221 clear_io_state(td);
a978ba68 1222
02bcaa8c 1223 fio_gettime(&td->start, NULL);
ebac4655
JA
1224
1225 do_verify(td);
1226
cda99fa0 1227 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
1228
1229 if (td->error || td->terminate)
1230 break;
1231 }
1232
36dff966 1233 update_rusage_stat(td);
cda99fa0
SSP
1234 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1235 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
756867bd
JA
1236 td->ts.total_run_time = mtime_since_now(&td->epoch);
1237 td->ts.io_bytes[0] = td->io_bytes[0];
1238 td->ts.io_bytes[1] = td->io_bytes[1];
1239
d2bbb8d9 1240 fio_mutex_down(writeout_mutex);
7b9f733a 1241 if (td->bw_log) {
f484470b 1242 if (td->o.bw_log_file) {
7b9f733a 1243 finish_log_named(td, td->bw_log,
f484470b
JA
1244 td->o.bw_log_file, "bw");
1245 } else
7b9f733a 1246 finish_log(td, td->bw_log, "bw");
e3cedca7 1247 }
7b9f733a 1248 if (td->lat_log) {
02af0988 1249 if (td->o.lat_log_file) {
7b9f733a 1250 finish_log_named(td, td->lat_log,
02af0988
JA
1251 td->o.lat_log_file, "lat");
1252 } else
7b9f733a 1253 finish_log(td, td->lat_log, "lat");
02af0988 1254 }
7b9f733a 1255 if (td->slat_log) {
f484470b 1256 if (td->o.lat_log_file) {
7b9f733a 1257 finish_log_named(td, td->slat_log,
1ca95738 1258 td->o.lat_log_file, "slat");
f484470b 1259 } else
7b9f733a 1260 finish_log(td, td->slat_log, "slat");
e3cedca7 1261 }
7b9f733a 1262 if (td->clat_log) {
f484470b 1263 if (td->o.lat_log_file) {
7b9f733a 1264 finish_log_named(td, td->clat_log,
f484470b
JA
1265 td->o.lat_log_file, "clat");
1266 } else
7b9f733a 1267 finish_log(td, td->clat_log, "clat");
e3cedca7 1268 }
d2bbb8d9 1269 fio_mutex_up(writeout_mutex);
398e8c4d
JA
1270 if (td->o.exec_postrun)
1271 exec_string(td->o.exec_postrun);
ebac4655
JA
1272
1273 if (exitall_on_terminate)
cc0df00a 1274 fio_terminate_threads(td->groupid);
ebac4655
JA
1275
1276err:
5bf13a5a 1277 if (td->error)
4ceb30d4 1278 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
5ec10eaa 1279 td->verror);
b4707ed7
JA
1280
1281 if (td->o.verify_async)
1282 verify_async_exit(td);
1283
24ffd2c2 1284 close_and_free_files(td);
2866c82d 1285 close_ioengine(td);
ebac4655 1286 cleanup_io_u(td);
6adb38a1 1287 cgroup_shutdown(td, &cgroup_mnt);
f29b25a3 1288
d2ce18b5
JA
1289 if (td->o.cpumask_set) {
1290 int ret = fio_cpuset_exit(&td->o.cpumask);
1291
1292 td_verror(td, ret, "fio_cpuset_exit");
1293 }
4e78e405 1294
f29b25a3
JA
1295 /*
1296 * do this very late, it will log file closing as well
1297 */
1298 if (td->o.write_iolog_file)
1299 write_iolog_close(td);
1300
d23bb327 1301 options_mem_free(td);
ebac4655 1302 td_set_runstate(td, TD_EXITED);
43d76807 1303 return (void *) (unsigned long) td->error;
ebac4655
JA
1304}
1305
906c8d75
JA
1306/*
1307 * We cannot pass the td data into a forked process, so attach the td and
1308 * pass it to the thread worker.
1309 */
a6418147 1310static int fork_main(int shmid, int offset)
ebac4655
JA
1311{
1312 struct thread_data *td;
a6418147 1313 void *data, *ret;
ebac4655 1314
65258961 1315#ifndef __hpux
ebac4655
JA
1316 data = shmat(shmid, NULL, 0);
1317 if (data == (void *) -1) {
a6418147
JA
1318 int __err = errno;
1319
ebac4655 1320 perror("shmat");
a6418147 1321 return __err;
ebac4655 1322 }
65258961
JA
1323#else
1324 /*
1325 * HP-UX inherits shm mappings?
1326 */
1327 data = threads;
1328#endif
ebac4655
JA
1329
1330 td = data + offset * sizeof(struct thread_data);
a6418147 1331 ret = thread_main(td);
ebac4655 1332 shmdt(data);
43d76807 1333 return (int) (unsigned long) ret;
ebac4655
JA
1334}
1335
906c8d75
JA
1336/*
1337 * Run over the job map and reap the threads that have exited, if any.
1338 */
ebac4655
JA
1339static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1340{
34572e28 1341 struct thread_data *td;
1f809d15 1342 int i, cputhreads, realthreads, pending, status, ret;
ebac4655
JA
1343
1344 /*
1345 * reap exited threads (TD_EXITED -> TD_REAPED)
1346 */
1f809d15 1347 realthreads = pending = cputhreads = 0;
34572e28 1348 for_each_td(td, i) {
3707f45b 1349 int flags = 0;
a2f77c9f 1350
84585003
JA
1351 /*
1352 * ->io_ops is NULL for a thread that has closed its
1353 * io engine
1354 */
ba0fbe10 1355 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
b990b5c0 1356 cputhreads++;
1f809d15
JA
1357 else
1358 realthreads++;
b990b5c0 1359
a1dedc1f
AC
1360 if (!td->pid) {
1361 pending++;
1362 continue;
1363 }
1364 if (td->runstate == TD_REAPED)
a2f77c9f 1365 continue;
2dc1bbeb 1366 if (td->o.use_thread) {
3707f45b
JA
1367 if (td->runstate == TD_EXITED) {
1368 td_set_runstate(td, TD_REAPED);
1369 goto reaped;
1370 }
1371 continue;
1372 }
a2f77c9f
JA
1373
1374 flags = WNOHANG;
1375 if (td->runstate == TD_EXITED)
1376 flags = 0;
1377
1378 /*
1379 * check if someone quit or got killed in an unusual way
1380 */
1381 ret = waitpid(td->pid, &status, flags);
3707f45b 1382 if (ret < 0) {
a2f77c9f 1383 if (errno == ECHILD) {
5921e80c
JA
1384 log_err("fio: pid=%d disappeared %d\n",
1385 (int) td->pid, td->runstate);
a2f77c9f
JA
1386 td_set_runstate(td, TD_REAPED);
1387 goto reaped;
1388 }
1389 perror("waitpid");
1390 } else if (ret == td->pid) {
1391 if (WIFSIGNALED(status)) {
fab6aa71
JA
1392 int sig = WTERMSIG(status);
1393
03e20d68 1394 if (sig != SIGTERM)
5ec10eaa 1395 log_err("fio: pid=%d, got signal=%d\n",
5921e80c 1396 (int) td->pid, sig);
fab6aa71
JA
1397 td_set_runstate(td, TD_REAPED);
1398 goto reaped;
1399 }
a2f77c9f
JA
1400 if (WIFEXITED(status)) {
1401 if (WEXITSTATUS(status) && !td->error)
1402 td->error = WEXITSTATUS(status);
a2f77c9f 1403
a2f77c9f
JA
1404 td_set_runstate(td, TD_REAPED);
1405 goto reaped;
a6418147
JA
1406 }
1407 }
ebac4655 1408
a2f77c9f
JA
1409 /*
1410 * thread is not dead, continue
1411 */
e48676ba 1412 pending++;
a2f77c9f 1413 continue;
fab6aa71 1414reaped:
ebac4655 1415 (*nr_running)--;
581e7141
JA
1416 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1417 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
6f88bcb0
JA
1418 if (!td->pid)
1419 pending--;
a2f77c9f
JA
1420
1421 if (td->error)
1422 exit_value++;
d3eeeabc
JA
1423
1424 done_secs += mtime_since_now(&td->epoch) / 1000;
ebac4655 1425 }
b990b5c0 1426
1f809d15 1427 if (*nr_running == cputhreads && !pending && realthreads)
cc0df00a 1428 fio_terminate_threads(TERMINATE_ALL);
ebac4655
JA
1429}
1430
be4ecfdf
JA
1431static void *gtod_thread_main(void *data)
1432{
1433 fio_mutex_up(startup_mutex);
1434
1435 /*
1436 * As long as we have jobs around, update the clock. It would be nice
1437 * to have some way of NOT hammering that CPU with gettimeofday(),
1438 * but I'm not sure what to use outside of a simple CPU nop to relax
1439 * it - we don't want to lose precision.
1440 */
1441 while (threads) {
1442 fio_gtod_update();
1443 nop;
1444 }
1445
1446 return NULL;
1447}
1448
1449static int fio_start_gtod_thread(void)
1450{
304a47c7 1451 pthread_attr_t attr;
f42c153d
JA
1452 int ret;
1453
304a47c7
VA
1454 pthread_attr_init(&attr);
1455 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1456 ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1457 pthread_attr_destroy(&attr);
f42c153d
JA
1458 if (ret) {
1459 log_err("Can't create gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1460 return 1;
1461 }
f42c153d
JA
1462
1463 ret = pthread_detach(gtod_thread);
1464 if (ret) {
1465 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
be4ecfdf
JA
1466 return 1;
1467 }
1468
29adda3c 1469 dprint(FD_MUTEX, "wait on startup_mutex\n");
be4ecfdf 1470 fio_mutex_down(startup_mutex);
29adda3c 1471 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
be4ecfdf
JA
1472 return 0;
1473}
1474
906c8d75
JA
1475/*
1476 * Main function for kicking off and reaping jobs, as needed.
1477 */
ebac4655
JA
1478static void run_threads(void)
1479{
ebac4655
JA
1480 struct thread_data *td;
1481 unsigned long spent;
1482 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 1483
2f9ade3c
JA
1484 if (fio_pin_memory())
1485 return;
ebac4655 1486
be4ecfdf
JA
1487 if (fio_gtod_offload && fio_start_gtod_thread())
1488 return;
1489
cc0df00a
JA
1490 set_sig_handlers();
1491
c6ae0a5b 1492 if (!terse_output) {
4ceb30d4 1493 log_info("Starting ");
9cedf167 1494 if (nr_thread)
4ceb30d4 1495 log_info("%d thread%s", nr_thread,
5ec10eaa 1496 nr_thread > 1 ? "s" : "");
9cedf167
JA
1497 if (nr_process) {
1498 if (nr_thread)
1499 printf(" and ");
4ceb30d4 1500 log_info("%d process%s", nr_process,
5ec10eaa 1501 nr_process > 1 ? "es" : "");
9cedf167 1502 }
4ceb30d4 1503 log_info("\n");
c6ae0a5b
JA
1504 fflush(stdout);
1505 }
c04f7ec3 1506
ebac4655
JA
1507 todo = thread_number;
1508 nr_running = 0;
1509 nr_started = 0;
1510 m_rate = t_rate = 0;
1511
34572e28 1512 for_each_td(td, i) {
263e529f 1513 print_status_init(td->thread_number - 1);
ebac4655 1514
067430bb 1515 if (!td->o.create_serialize)
ebac4655
JA
1516 continue;
1517
1518 /*
1519 * do file setup here so it happens sequentially,
1520 * we don't want X number of threads getting their
1521 * client data interspersed on disk
1522 */
53cdc686 1523 if (setup_files(td)) {
5bf13a5a
JA
1524 exit_value++;
1525 if (td->error)
5921e80c
JA
1526 log_err("fio: pid=%d, err=%d/%s\n",
1527 (int) td->pid, td->error, td->verror);
ebac4655
JA
1528 td_set_runstate(td, TD_REAPED);
1529 todo--;
c69e8875
JA
1530 } else {
1531 struct fio_file *f;
2b13e716 1532 unsigned int j;
c69e8875
JA
1533
1534 /*
1535 * for sharing to work, each job must always open
1536 * its own files. so close them, if we opened them
1537 * for creation
1538 */
2b13e716 1539 for_each_file(td, f, j) {
22a57ba8
JA
1540 if (fio_file_open(f))
1541 td_io_close_file(td, f);
22a57ba8 1542 }
5ec10eaa 1543 }
ebac4655
JA
1544 }
1545
a2f77c9f
JA
1546 set_genesis_time();
1547
ebac4655 1548 while (todo) {
fca70358 1549 struct thread_data *map[REAL_MAX_JOBS];
75154845
JA
1550 struct timeval this_start;
1551 int this_jobs = 0, left;
1552
ebac4655
JA
1553 /*
1554 * create threads (TD_NOT_CREATED -> TD_CREATED)
1555 */
34572e28 1556 for_each_td(td, i) {
ebac4655
JA
1557 if (td->runstate != TD_NOT_CREATED)
1558 continue;
1559
1560 /*
1561 * never got a chance to start, killed by other
1562 * thread for some reason
1563 */
1564 if (td->terminate) {
1565 todo--;
1566 continue;
1567 }
1568
2dc1bbeb 1569 if (td->o.start_delay) {
263e529f 1570 spent = mtime_since_genesis();
ebac4655 1571
2dc1bbeb 1572 if (td->o.start_delay * 1000 > spent)
ebac4655
JA
1573 continue;
1574 }
1575
6f88bcb0
JA
1576 if (td->o.stonewall && (nr_started || nr_running)) {
1577 dprint(FD_PROCESS, "%s: stonewall wait\n",
1578 td->o.name);
ebac4655 1579 break;
6f88bcb0 1580 }
ebac4655 1581
067430bb
JA
1582 init_disk_util(td);
1583
75154845
JA
1584 /*
1585 * Set state to created. Thread will transition
1586 * to TD_INITIALIZED when it's done setting up.
1587 */
ebac4655 1588 td_set_runstate(td, TD_CREATED);
75154845 1589 map[this_jobs++] = td;
ebac4655
JA
1590 nr_started++;
1591
2dc1bbeb 1592 if (td->o.use_thread) {
f42c153d
JA
1593 int ret;
1594
ee56ad50 1595 dprint(FD_PROCESS, "will pthread_create\n");
f42c153d
JA
1596 ret = pthread_create(&td->thread, NULL,
1597 thread_main, td);
1598 if (ret) {
1599 log_err("pthread_create: %s\n",
1600 strerror(ret));
ebac4655 1601 nr_started--;
c8bb6faf 1602 break;
ebac4655 1603 }
f42c153d
JA
1604 ret = pthread_detach(td->thread);
1605 if (ret)
1606 log_err("pthread_detach: %s",
1607 strerror(ret));
ebac4655 1608 } else {
5e1d306e 1609 pid_t pid;
ee56ad50 1610 dprint(FD_PROCESS, "will fork\n");
5e1d306e
JA
1611 pid = fork();
1612 if (!pid) {
a6418147
JA
1613 int ret = fork_main(shm_id, i);
1614
5e1d306e
JA
1615 _exit(ret);
1616 } else if (i == fio_debug_jobno)
1617 *fio_debug_jobp = pid;
ebac4655 1618 }
29adda3c 1619 dprint(FD_MUTEX, "wait on startup_mutex\n");
656b1393
JA
1620 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1621 log_err("fio: job startup hung? exiting.\n");
cc0df00a 1622 fio_terminate_threads(TERMINATE_ALL);
656b1393
JA
1623 fio_abort = 1;
1624 nr_started--;
1625 break;
1626 }
29adda3c 1627 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
ebac4655
JA
1628 }
1629
1630 /*
75154845
JA
1631 * Wait for the started threads to transition to
1632 * TD_INITIALIZED.
ebac4655 1633 */
02bcaa8c 1634 fio_gettime(&this_start, NULL);
75154845 1635 left = this_jobs;
6ce15a32 1636 while (left && !fio_abort) {
75154845
JA
1637 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1638 break;
1639
1640 usleep(100000);
1641
1642 for (i = 0; i < this_jobs; i++) {
1643 td = map[i];
1644 if (!td)
1645 continue;
b6f4d880 1646 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1647 map[i] = NULL;
1648 left--;
b6f4d880
JA
1649 } else if (td->runstate >= TD_EXITED) {
1650 map[i] = NULL;
1651 left--;
1652 todo--;
1653 nr_running++; /* work-around... */
75154845
JA
1654 }
1655 }
1656 }
1657
1658 if (left) {
3b70d7e5 1659 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1660 for (i = 0; i < this_jobs; i++) {
1661 td = map[i];
1662 if (!td)
1663 continue;
1664 kill(td->pid, SIGTERM);
1665 }
1666 break;
1667 }
1668
1669 /*
b6f4d880 1670 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1671 */
34572e28 1672 for_each_td(td, i) {
75154845 1673 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1674 continue;
1675
b29ee5b3
JA
1676 if (in_ramp_time(td))
1677 td_set_runstate(td, TD_RAMP);
1678 else
1679 td_set_runstate(td, TD_RUNNING);
ebac4655
JA
1680 nr_running++;
1681 nr_started--;
581e7141
JA
1682 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1683 t_rate += td->o.rate[0] + td->o.rate[1];
75154845 1684 todo--;
cdd18ad8 1685 fio_mutex_up(td->mutex);
ebac4655
JA
1686 }
1687
1688 reap_threads(&nr_running, &t_rate, &m_rate);
1689
1690 if (todo)
1691 usleep(100000);
1692 }
1693
1694 while (nr_running) {
1695 reap_threads(&nr_running, &t_rate, &m_rate);
cc0df00a
JA
1696
1697 if (is_backend)
1698 fio_server_idle_loop();
70e0c316
JA
1699 else
1700 usleep(10000);
ebac4655
JA
1701 }
1702
1703 update_io_ticks();
2f9ade3c 1704 fio_unpin_memory();
ebac4655
JA
1705}
1706
50d16976 1707int exec_run(void)
ebac4655 1708{
a37f69b7 1709 if (nr_clients)
37db14fe 1710 return fio_handle_clients();
07b3232d
JA
1711 if (exec_profile && load_profile(exec_profile))
1712 return 1;
1713
74929ac2
JA
1714 if (!thread_number)
1715 return 0;
1716
bb3884d8
JA
1717 if (write_bw_log) {
1718 setup_log(&agg_io_log[DDIR_READ]);
1719 setup_log(&agg_io_log[DDIR_WRITE]);
1720 }
1721
cdd18ad8 1722 startup_mutex = fio_mutex_init(0);
03e20d68
BC
1723 if (startup_mutex == NULL)
1724 return 1;
d2bbb8d9 1725 writeout_mutex = fio_mutex_init(1);
03e20d68
BC
1726 if (writeout_mutex == NULL)
1727 return 1;
07739b57 1728
a2f77c9f 1729 set_genesis_time();
7985701b 1730 create_disk_util_thread();
ebac4655 1731
dae5341c
JA
1732 cgroup_list = smalloc(sizeof(*cgroup_list));
1733 INIT_FLIST_HEAD(cgroup_list);
1734
ebac4655 1735 run_threads();
6ce15a32 1736
bb3884d8 1737 if (!fio_abort) {
6ce15a32 1738 show_run_stats();
bb3884d8 1739 if (write_bw_log) {
5ec10eaa
JA
1740 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1741 __finish_log(agg_io_log[DDIR_WRITE],
1742 "agg-write_bw.log");
bb3884d8
JA
1743 }
1744 }
ebac4655 1745
dae5341c
JA
1746 cgroup_kill(cgroup_list);
1747 sfree(cgroup_list);
61eb313e 1748 sfree(cgroup_mnt);
39f22027 1749
cdd18ad8 1750 fio_mutex_remove(startup_mutex);
d2bbb8d9 1751 fio_mutex_remove(writeout_mutex);
437c9b71 1752 return exit_value;
ebac4655 1753}
50d16976
JA
1754
1755void reset_fio_state(void)
1756{
1757 groupid = 0;
1758 thread_number = 0;
1759 nr_process = 0;
1760 nr_thread = 0;
1761 done_secs = 0;
1762}
1763
1764int main(int argc, char *argv[], char *envp[])
1765{
1766 long ps;
1767
1768 arch_init(envp);
1769
1770 sinit();
1771
1772 /*
1773 * We need locale for number printing, if it isn't set then just
1774 * go with the US format.
1775 */
1776 if (!getenv("LC_NUMERIC"))
1777 setlocale(LC_NUMERIC, "en_US");
1778
1779 ps = sysconf(_SC_PAGESIZE);
1780 if (ps < 0) {
1781 log_err("Failed to get page size\n");
1782 return 1;
1783 }
1784
1785 page_size = ps;
1786 page_mask = ps - 1;
1787
1788 fio_keywords_init();
1789
1790 if (parse_options(argc, argv))
1791 return 1;
1792
1793 return exec_run();
1794}