projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
parse: use MIN/MAXDOUBLE instead of some representation of NAN
[fio.git] / backend.c
CommitLineData
2e1df07d
JA		 1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
6 *
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
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 */
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
27#include <limits.h>
28#include <signal.h>
29#include <time.h>
30#include <locale.h>
31#include <assert.h>
32#include <time.h>
e43606c2 33#include <inttypes.h>
2e1df07d
JA		 34#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
ec5c6b12 37#ifndef FIO_NO_HAVE_SHM_H
2e1df07d 38#include <sys/shm.h>
ec5c6b12 39#endif
2e1df07d
JA		 40#include <sys/mman.h>
41
42#include "fio.h"
43#include "hash.h"
44#include "smalloc.h"
45#include "verify.h"
46#include "trim.h"
47#include "diskutil.h"
48#include "cgroup.h"
49#include "profile.h"
50#include "lib/rand.h"
51#include "memalign.h"
52#include "server.h"
44404c5a 53#include "lib/getrusage.h"
f2a2ce0e 54#include "idletime.h"
2e1df07d
JA		 55
56static pthread_t disk_util_thread;
9ec7779f 57static struct fio_mutex *disk_thread_mutex;
2e1df07d
JA		 58static struct fio_mutex *startup_mutex;
59static struct fio_mutex *writeout_mutex;
60static struct flist_head *cgroup_list;
61static char *cgroup_mnt;
62static int exit_value;
63static volatile int fio_abort;
64
6eaf09d6 65struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 66
a3efc919
JA		 67int groupid = 0;
68unsigned int thread_number = 0;
108fea77 69unsigned int stat_number = 0;
a3efc919
JA		 70unsigned int nr_process = 0;
71unsigned int nr_thread = 0;
72int shm_id = 0;
73int temp_stall_ts;
74unsigned long done_secs = 0;
27357187 75volatile int disk_util_exit = 0;
a3efc919 76
2e1df07d 77#define PAGE_ALIGN(buf) \
e43606c2 78 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA		 79
80#define JOB_START_TIMEOUT (5 * 1000)
81
82static void sig_int(int sig)
83{
84 if (threads) {
85 if (is_backend)
86 fio_server_got_signal(sig);
87 else {
88 log_info("\nfio: terminating on signal %d\n", sig);
89 fflush(stdout);
90 exit_value = 128;
91 }
92
93 fio_terminate_threads(TERMINATE_ALL);
94 }
95}
96
4c6d91e8
JA		 97static void sig_show_status(int sig)
98{
99 show_running_run_stats();
100}
101
2e1df07d
JA		 102static void set_sig_handlers(void)
103{
104 struct sigaction act;
105
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGINT, &act, NULL);
110
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGTERM, &act, NULL);
115
2f694507
BC		 116/* Windows uses SIGBREAK as a quit signal from other applications */
117#ifdef WIN32
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGBREAK, &act, NULL);
122#endif
123
4c6d91e8
JA		 124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_show_status;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGUSR1, &act, NULL);
128
2e1df07d
JA		 129 if (is_backend) {
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_int;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGPIPE, &act, NULL);
134 }
135}
136
137/*
138 * Check if we are above the minimum rate given.
139 */
140static int __check_min_rate(struct thread_data *td, struct timeval *now,
141 enum fio_ddir ddir)
142{
143 unsigned long long bytes = 0;
144 unsigned long iops = 0;
145 unsigned long spent;
146 unsigned long rate;
147 unsigned int ratemin = 0;
148 unsigned int rate_iops = 0;
149 unsigned int rate_iops_min = 0;
150
151 assert(ddir_rw(ddir));
152
153 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
154 return 0;
155
156 /*
157 * allow a 2 second settle period in the beginning
158 */
159 if (mtime_since(&td->start, now) < 2000)
160 return 0;
161
162 iops += td->this_io_blocks[ddir];
163 bytes += td->this_io_bytes[ddir];
164 ratemin += td->o.ratemin[ddir];
165 rate_iops += td->o.rate_iops[ddir];
166 rate_iops_min += td->o.rate_iops_min[ddir];
167
168 /*
169 * if rate blocks is set, sample is running
170 */
171 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
172 spent = mtime_since(&td->lastrate[ddir], now);
173 if (spent < td->o.ratecycle)
174 return 0;
175
176 if (td->o.rate[ddir]) {
177 /*
178 * check bandwidth specified rate
179 */
180 if (bytes < td->rate_bytes[ddir]) {
181 log_err("%s: min rate %u not met\n", td->o.name,
182 ratemin);
183 return 1;
184 } else {
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
186 if (rate < ratemin ||
187 bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met, got"
189 " %luKB/sec\n", td->o.name,
190 ratemin, rate);
191 return 1;
192 }
193 }
194 } else {
195 /*
196 * checks iops specified rate
197 */
198 if (iops < rate_iops) {
199 log_err("%s: min iops rate %u not met\n",
200 td->o.name, rate_iops);
201 return 1;
202 } else {
203 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
204 if (rate < rate_iops_min ||
205 iops < td->rate_blocks[ddir]) {
206 log_err("%s: min iops rate %u not met,"
207 " got %lu\n", td->o.name,
208 rate_iops_min, rate);
209 }
210 }
211 }
212 }
213
214 td->rate_bytes[ddir] = bytes;
215 td->rate_blocks[ddir] = iops;
216 memcpy(&td->lastrate[ddir], now, sizeof(*now));
217 return 0;
218}
219
220static int check_min_rate(struct thread_data *td, struct timeval *now,
100f49f1 221 uint64_t *bytes_done)
2e1df07d
JA		 222{
223 int ret = 0;
224
6eaf09d6
SL		 225 if (bytes_done[DDIR_READ])
226 ret |= __check_min_rate(td, now, DDIR_READ);
227 if (bytes_done[DDIR_WRITE])
228 ret |= __check_min_rate(td, now, DDIR_WRITE);
229 if (bytes_done[DDIR_TRIM])
230 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA		 231
232 return ret;
233}
234
235/*
236 * When job exits, we can cancel the in-flight IO if we are using async
237 * io. Attempt to do so.
238 */
239static void cleanup_pending_aio(struct thread_data *td)
240{
241 struct flist_head *entry, *n;
242 struct io_u *io_u;
243 int r;
244
245 /*
246 * get immediately available events, if any
247 */
248 r = io_u_queued_complete(td, 0, NULL);
249 if (r < 0)
250 return;
251
252 /*
253 * now cancel remaining active events
254 */
255 if (td->io_ops->cancel) {
256 flist_for_each_safe(entry, n, &td->io_u_busylist) {
257 io_u = flist_entry(entry, struct io_u, list);
258
259 /*
260 * if the io_u isn't in flight, then that generally
261 * means someone leaked an io_u. complain but fix
262 * it up, so we don't stall here.
263 */
264 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
265 log_err("fio: non-busy IO on busy list\n");
266 put_io_u(td, io_u);
267 } else {
268 r = td->io_ops->cancel(td, io_u);
269 if (!r)
270 put_io_u(td, io_u);
271 }
272 }
273 }
274
275 if (td->cur_depth)
276 r = io_u_queued_complete(td, td->cur_depth, NULL);
277}
278
279/*
280 * Helper to handle the final sync of a file. Works just like the normal
281 * io path, just does everything sync.
282 */
283static int fio_io_sync(struct thread_data *td, struct fio_file *f)
284{
285 struct io_u *io_u = __get_io_u(td);
286 int ret;
287
288 if (!io_u)
289 return 1;
290
291 io_u->ddir = DDIR_SYNC;
292 io_u->file = f;
293
294 if (td_io_prep(td, io_u)) {
295 put_io_u(td, io_u);
296 return 1;
297 }
298
299requeue:
300 ret = td_io_queue(td, io_u);
301 if (ret < 0) {
302 td_verror(td, io_u->error, "td_io_queue");
303 put_io_u(td, io_u);
304 return 1;
305 } else if (ret == FIO_Q_QUEUED) {
306 if (io_u_queued_complete(td, 1, NULL) < 0)
307 return 1;
308 } else if (ret == FIO_Q_COMPLETED) {
309 if (io_u->error) {
310 td_verror(td, io_u->error, "td_io_queue");
311 return 1;
312 }
313
314 if (io_u_sync_complete(td, io_u, NULL) < 0)
315 return 1;
316 } else if (ret == FIO_Q_BUSY) {
317 if (td_io_commit(td))
318 return 1;
319 goto requeue;
320 }
321
322 return 0;
323}
a3efc919 324
2e1df07d
JA		 325static inline void __update_tv_cache(struct thread_data *td)
326{
327 fio_gettime(&td->tv_cache, NULL);
328}
329
330static inline void update_tv_cache(struct thread_data *td)
331{
332 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
333 __update_tv_cache(td);
334}
335
336static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
337{
338 if (in_ramp_time(td))
339 return 0;
340 if (!td->o.timeout)
341 return 0;
342 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
343 return 1;
344
345 return 0;
346}
347
348static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
349 int *retptr)
350{
351 int ret = *retptr;
352
353 if (ret < 0 || td->error) {
8b28bd41
DM		 354 int err = td->error;
355 enum error_type_bit eb;
2e1df07d
JA		 356
357 if (ret < 0)
358 err = -ret;
2e1df07d 359
8b28bd41
DM		 360 eb = td_error_type(ddir, err);
361 if (!(td->o.continue_on_error & (1 << eb)))
2e1df07d
JA		 362 return 1;
363
8b28bd41 364 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA		 365 /*
366 * Continue with the I/Os in case of
367 * a non fatal error.
368 */
369 update_error_count(td, err);
370 td_clear_error(td);
371 *retptr = 0;
372 return 0;
373 } else if (td->o.fill_device && err == ENOSPC) {
374 /*
375 * We expect to hit this error if
376 * fill_device option is set.
377 */
378 td_clear_error(td);
379 td->terminate = 1;
380 return 1;
381 } else {
382 /*
383 * Stop the I/O in case of a fatal
384 * error.
385 */
386 update_error_count(td, err);
387 return 1;
388 }
389 }
390
391 return 0;
392}
393
2e1df07d
JA		 394/*
395 * The main verify engine. Runs over the writes we previously submitted,
396 * reads the blocks back in, and checks the crc/md5 of the data.
397 */
100f49f1 398static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d 399{
100f49f1 400 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA		 401 struct fio_file *f;
402 struct io_u *io_u;
403 int ret, min_events;
404 unsigned int i;
405
406 dprint(FD_VERIFY, "starting loop\n");
407
408 /*
409 * sync io first and invalidate cache, to make sure we really
410 * read from disk.
411 */
412 for_each_file(td, f, i) {
413 if (!fio_file_open(f))
414 continue;
415 if (fio_io_sync(td, f))
416 break;
417 if (file_invalidate_cache(td, f))
418 break;
419 }
420
421 if (td->error)
422 return;
423
424 td_set_runstate(td, TD_VERIFYING);
425
426 io_u = NULL;
427 while (!td->terminate) {
fbccf46c 428 enum fio_ddir ddir;
2e1df07d
JA		 429 int ret2, full;
430
431 update_tv_cache(td);
432
433 if (runtime_exceeded(td, &td->tv_cache)) {
434 __update_tv_cache(td);
435 if (runtime_exceeded(td, &td->tv_cache)) {
436 td->terminate = 1;
437 break;
438 }
439 }
440
9e684a49
DE		 441 if (flow_threshold_exceeded(td))
442 continue;
443
44cbc6da
JA		 444 if (!td->o.experimental_verify) {
445 io_u = __get_io_u(td);
446 if (!io_u)
447 break;
2e1df07d 448
44cbc6da
JA		 449 if (get_next_verify(td, io_u)) {
450 put_io_u(td, io_u);
451 break;
452 }
2e1df07d 453
44cbc6da
JA		 454 if (td_io_prep(td, io_u)) {
455 put_io_u(td, io_u);
456 break;
457 }
458 } else {
100f49f1
JA		 459 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
460 break;
461
bcd5abfa
JA		 462 while ((io_u = get_io_u(td)) != NULL) {
463 /*
464 * We are only interested in the places where
465 * we wrote or trimmed IOs. Turn those into
466 * reads for verification purposes.
467 */
468 if (io_u->ddir == DDIR_READ) {
469 /*
470 * Pretend we issued it for rwmix
471 * accounting
472 */
473 td->io_issues[DDIR_READ]++;
474 put_io_u(td, io_u);
475 continue;
476 } else if (io_u->ddir == DDIR_TRIM) {
477 io_u->ddir = DDIR_READ;
478 io_u->flags |= IO_U_F_TRIMMED;
479 break;
480 } else if (io_u->ddir == DDIR_WRITE) {
481 io_u->ddir = DDIR_READ;
482 break;
483 } else {
484 put_io_u(td, io_u);
485 continue;
486 }
487 }
44cbc6da 488
bcd5abfa 489 if (!io_u)
44cbc6da 490 break;
2e1df07d
JA		 491 }
492
493 if (td->o.verify_async)
494 io_u->end_io = verify_io_u_async;
495 else
496 io_u->end_io = verify_io_u;
497
fbccf46c
JA		 498 ddir = io_u->ddir;
499
2e1df07d
JA		 500 ret = td_io_queue(td, io_u);
501 switch (ret) {
502 case FIO_Q_COMPLETED:
503 if (io_u->error) {
504 ret = -io_u->error;
505 clear_io_u(td, io_u);
506 } else if (io_u->resid) {
507 int bytes = io_u->xfer_buflen - io_u->resid;
508
509 /*
510 * zero read, fail
511 */
512 if (!bytes) {
513 td_verror(td, EIO, "full resid");
514 put_io_u(td, io_u);
515 break;
516 }
517
518 io_u->xfer_buflen = io_u->resid;
519 io_u->xfer_buf += bytes;
520 io_u->offset += bytes;
521
522 if (ddir_rw(io_u->ddir))
523 td->ts.short_io_u[io_u->ddir]++;
524
525 f = io_u->file;
526 if (io_u->offset == f->real_file_size)
527 goto sync_done;
528
529 requeue_io_u(td, &io_u);
530 } else {
531sync_done:
100f49f1 532 ret = io_u_sync_complete(td, io_u, bytes_done);
2e1df07d
JA		 533 if (ret < 0)
534 break;
535 }
536 continue;
537 case FIO_Q_QUEUED:
538 break;
539 case FIO_Q_BUSY:
540 requeue_io_u(td, &io_u);
541 ret2 = td_io_commit(td);
542 if (ret2 < 0)
543 ret = ret2;
544 break;
545 default:
546 assert(ret < 0);
547 td_verror(td, -ret, "td_io_queue");
548 break;
549 }
550
fbccf46c 551 if (break_on_this_error(td, ddir, &ret))
2e1df07d
JA		 552 break;
553
554 /*
555 * if we can queue more, do so. but check if there are
556 * completed io_u's first. Note that we can get BUSY even
557 * without IO queued, if the system is resource starved.
558 */
559 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
560 if (full || !td->o.iodepth_batch_complete) {
561 min_events = min(td->o.iodepth_batch_complete,
562 td->cur_depth);
8a74b56d
JA		 563 /*
564 * if the queue is full, we MUST reap at least 1 event
565 */
566 if (full && !min_events)
2e1df07d
JA		 567 min_events = 1;
568
569 do {
570 /*
571 * Reap required number of io units, if any,
572 * and do the verification on them through
573 * the callback handler
574 */
100f49f1 575 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
2e1df07d
JA		 576 ret = -1;
577 break;
578 }
579 } while (full && (td->cur_depth > td->o.iodepth_low));
580 }
581 if (ret < 0)
582 break;
583 }
584
585 if (!td->error) {
586 min_events = td->cur_depth;
587
588 if (min_events)
589 ret = io_u_queued_complete(td, min_events, NULL);
590 } else
591 cleanup_pending_aio(td);
592
593 td_set_runstate(td, TD_RUNNING);
594
595 dprint(FD_VERIFY, "exiting loop\n");
596}
597
f7078f7b
JA		 598static int io_bytes_exceeded(struct thread_data *td)
599{
600 unsigned long long bytes;
601
602 if (td_rw(td))
6eaf09d6 603 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
f7078f7b 604 else if (td_write(td))
6eaf09d6
SL		 605 bytes = td->this_io_bytes[DDIR_WRITE];
606 else if (td_read(td))
607 bytes = td->this_io_bytes[DDIR_READ];
f7078f7b 608 else
6eaf09d6 609 bytes = td->this_io_bytes[DDIR_TRIM];
f7078f7b
JA		 610
611 return bytes >= td->o.size;
612}
613
2e1df07d
JA		 614/*
615 * Main IO worker function. It retrieves io_u's to process and queues
616 * and reaps them, checking for rate and errors along the way.
100f49f1
JA		 617 *
618 * Returns number of bytes written and trimmed.
2e1df07d 619 */
100f49f1 620static uint64_t do_io(struct thread_data *td)
2e1df07d 621{
100f49f1 622 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA		 623 unsigned int i;
624 int ret = 0;
625
626 if (in_ramp_time(td))
627 td_set_runstate(td, TD_RAMP);
628 else
629 td_set_runstate(td, TD_RUNNING);
630
f7078f7b 631 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
c04e4661
DE		 632 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
633 td->o.time_based) {
2e1df07d 634 struct timeval comp_time;
2e1df07d
JA		 635 int min_evts = 0;
636 struct io_u *io_u;
637 int ret2, full;
638 enum fio_ddir ddir;
639
7d7803fa 640 if (td->terminate || td->done)
2e1df07d
JA		 641 break;
642
643 update_tv_cache(td);
644
645 if (runtime_exceeded(td, &td->tv_cache)) {
646 __update_tv_cache(td);
647 if (runtime_exceeded(td, &td->tv_cache)) {
648 td->terminate = 1;
649 break;
650 }
651 }
652
9e684a49
DE		 653 if (flow_threshold_exceeded(td))
654 continue;
655
2e1df07d
JA		 656 io_u = get_io_u(td);
657 if (!io_u)
658 break;
659
660 ddir = io_u->ddir;
661
662 /*
82af2a7c
JA		 663 * Add verification end_io handler if:
664 * - Asked to verify (!td_rw(td))
665 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA		 666 */
667 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 668 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
2e1df07d
JA		 669 if (td->o.verify_async)
670 io_u->end_io = verify_io_u_async;
671 else
672 io_u->end_io = verify_io_u;
673 td_set_runstate(td, TD_VERIFYING);
674 } else if (in_ramp_time(td))
675 td_set_runstate(td, TD_RAMP);
676 else
677 td_set_runstate(td, TD_RUNNING);
678
679 ret = td_io_queue(td, io_u);
680 switch (ret) {
681 case FIO_Q_COMPLETED:
682 if (io_u->error) {
683 ret = -io_u->error;
684 clear_io_u(td, io_u);
685 } else if (io_u->resid) {
686 int bytes = io_u->xfer_buflen - io_u->resid;
687 struct fio_file *f = io_u->file;
688
689 /*
690 * zero read, fail
691 */
692 if (!bytes) {
693 td_verror(td, EIO, "full resid");
694 put_io_u(td, io_u);
695 break;
696 }
697
698 io_u->xfer_buflen = io_u->resid;
699 io_u->xfer_buf += bytes;
700 io_u->offset += bytes;
701
702 if (ddir_rw(io_u->ddir))
703 td->ts.short_io_u[io_u->ddir]++;
704
705 if (io_u->offset == f->real_file_size)
706 goto sync_done;
707
708 requeue_io_u(td, &io_u);
709 } else {
710sync_done:
6eaf09d6
SL		 711 if (__should_check_rate(td, DDIR_READ) ||
712 __should_check_rate(td, DDIR_WRITE) ||
713 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA		 714 fio_gettime(&comp_time, NULL);
715
716 ret = io_u_sync_complete(td, io_u, bytes_done);
717 if (ret < 0)
718 break;
719 }
720 break;
721 case FIO_Q_QUEUED:
722 /*
723 * if the engine doesn't have a commit hook,
724 * the io_u is really queued. if it does have such
725 * a hook, it has to call io_u_queued() itself.
726 */
727 if (td->io_ops->commit == NULL)
728 io_u_queued(td, io_u);
729 break;
730 case FIO_Q_BUSY:
731 requeue_io_u(td, &io_u);
732 ret2 = td_io_commit(td);
733 if (ret2 < 0)
734 ret = ret2;
735 break;
736 default:
737 assert(ret < 0);
738 put_io_u(td, io_u);
739 break;
740 }
741
742 if (break_on_this_error(td, ddir, &ret))
743 break;
744
745 /*
746 * See if we need to complete some commands. Note that we
747 * can get BUSY even without IO queued, if the system is
748 * resource starved.
749 */
750 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
751 if (full || !td->o.iodepth_batch_complete) {
752 min_evts = min(td->o.iodepth_batch_complete,
753 td->cur_depth);
8a74b56d
JA		 754 /*
755 * if the queue is full, we MUST reap at least 1 event
756 */
757 if (full && !min_evts)
2e1df07d
JA		 758 min_evts = 1;
759
6eaf09d6
SL		 760 if (__should_check_rate(td, DDIR_READ) ||
761 __should_check_rate(td, DDIR_WRITE) ||
762 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA		 763 fio_gettime(&comp_time, NULL);
764
765 do {
766 ret = io_u_queued_complete(td, min_evts, bytes_done);
767 if (ret < 0)
768 break;
769
770 } while (full && (td->cur_depth > td->o.iodepth_low));
771 }
772
773 if (ret < 0)
774 break;
d5abee06 775 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA		 776 continue;
777
778 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
779 if (check_min_rate(td, &comp_time, bytes_done)) {
780 if (exitall_on_terminate)
781 fio_terminate_threads(td->groupid);
782 td_verror(td, EIO, "check_min_rate");
783 break;
784 }
785 }
786
787 if (td->o.thinktime) {
788 unsigned long long b;
789
342f4be4 790 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA		 791 if (!(b % td->o.thinktime_blocks)) {
792 int left;
793
794 if (td->o.thinktime_spin)
795 usec_spin(td->o.thinktime_spin);
796
797 left = td->o.thinktime - td->o.thinktime_spin;
798 if (left)
799 usec_sleep(td, left);
800 }
801 }
802 }
803
804 if (td->trim_entries)
805 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
806
807 if (td->o.fill_device && td->error == ENOSPC) {
808 td->error = 0;
809 td->terminate = 1;
810 }
811 if (!td->error) {
812 struct fio_file *f;
813
814 i = td->cur_depth;
815 if (i) {
816 ret = io_u_queued_complete(td, i, NULL);
817 if (td->o.fill_device && td->error == ENOSPC)
818 td->error = 0;
819 }
820
821 if (should_fsync(td) && td->o.end_fsync) {
822 td_set_runstate(td, TD_FSYNCING);
823
824 for_each_file(td, f, i) {
825 if (!fio_file_open(f))
826 continue;
827 fio_io_sync(td, f);
828 }
829 }
830 } else
831 cleanup_pending_aio(td);
832
833 /*
834 * stop job if we failed doing any IO
835 */
342f4be4 836 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 837 td->done = 1;
100f49f1
JA		 838
839 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
2e1df07d
JA		 840}
841
842static void cleanup_io_u(struct thread_data *td)
843{
844 struct flist_head *entry, *n;
845 struct io_u *io_u;
846
847 flist_for_each_safe(entry, n, &td->io_u_freelist) {
848 io_u = flist_entry(entry, struct io_u, list);
849
850 flist_del(&io_u->list);
c73ed246
JA		 851
852 if (td->io_ops->io_u_free)
853 td->io_ops->io_u_free(td, io_u);
854
2e1df07d
JA		 855 fio_memfree(io_u, sizeof(*io_u));
856 }
857
858 free_io_mem(td);
859}
860
861static int init_io_u(struct thread_data *td)
862{
863 struct io_u *io_u;
9c42684e 864 unsigned int max_bs, min_write;
2e1df07d 865 int cl_align, i, max_units;
59d8d0f5 866 int data_xfer = 1;
2e1df07d
JA		 867 char *p;
868
869 max_units = td->o.iodepth;
870 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
6eaf09d6 871 max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
9c42684e 872 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA		 873 td->orig_buffer_size = (unsigned long long) max_bs
874 * (unsigned long long) max_units;
875
88045e04 876 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA		 877 data_xfer = 0;
878
2e1df07d
JA		 879 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
880 unsigned long bs;
881
882 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
883 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
884 }
885
886 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
887 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
888 return 1;
889 }
890
59d8d0f5 891 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA		 892 return 1;
893
894 if (td->o.odirect || td->o.mem_align ||
895 (td->io_ops->flags & FIO_RAWIO))
896 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
897 else
898 p = td->orig_buffer;
899
900 cl_align = os_cache_line_size();
901
902 for (i = 0; i < max_units; i++) {
903 void *ptr;
904
905 if (td->terminate)
906 return 1;
907
908 ptr = fio_memalign(cl_align, sizeof(*io_u));
909 if (!ptr) {
910 log_err("fio: unable to allocate aligned memory\n");
911 break;
912 }
913
914 io_u = ptr;
915 memset(io_u, 0, sizeof(*io_u));
916 INIT_FLIST_HEAD(&io_u->list);
917 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
918
59d8d0f5 919 if (data_xfer) {
2e1df07d
JA		 920 io_u->buf = p;
921 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
922
923 if (td_write(td))
9c42684e 924 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA		 925 if (td_write(td) && td->o.verify_pattern_bytes) {
926 /*
927 * Fill the buffer with the pattern if we are
928 * going to be doing writes.
929 */
930 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
931 }
932 }
933
934 io_u->index = i;
935 io_u->flags = IO_U_F_FREE;
936 flist_add(&io_u->list, &td->io_u_freelist);
c73ed246
JA		 937
938 if (td->io_ops->io_u_init) {
939 int ret = td->io_ops->io_u_init(td, io_u);
940
941 if (ret) {
942 log_err("fio: failed to init engine data: %d\n", ret);
943 return 1;
944 }
945 }
946
2e1df07d
JA		 947 p += max_bs;
948 }
949
950 return 0;
951}
952
953static int switch_ioscheduler(struct thread_data *td)
954{
955 char tmp[256], tmp2[128];
956 FILE *f;
957 int ret;
958
959 if (td->io_ops->flags & FIO_DISKLESSIO)
960 return 0;
961
962 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
963
964 f = fopen(tmp, "r+");
965 if (!f) {
966 if (errno == ENOENT) {
967 log_err("fio: os or kernel doesn't support IO scheduler"
968 " switching\n");
969 return 0;
970 }
971 td_verror(td, errno, "fopen iosched");
972 return 1;
973 }
974
975 /*
976 * Set io scheduler.
977 */
978 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
979 if (ferror(f) || ret != 1) {
980 td_verror(td, errno, "fwrite");
981 fclose(f);
982 return 1;
983 }
984
985 rewind(f);
986
987 /*
988 * Read back and check that the selected scheduler is now the default.
989 */
990 ret = fread(tmp, 1, sizeof(tmp), f);
991 if (ferror(f) || ret < 0) {
992 td_verror(td, errno, "fread");
993 fclose(f);
994 return 1;
995 }
996
997 sprintf(tmp2, "[%s]", td->o.ioscheduler);
998 if (!strstr(tmp, tmp2)) {
999 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1000 td_verror(td, EINVAL, "iosched_switch");
1001 fclose(f);
1002 return 1;
1003 }
1004
1005 fclose(f);
1006 return 0;
1007}
1008
1009static int keep_running(struct thread_data *td)
1010{
2e1df07d
JA		 1011 if (td->done)
1012 return 0;
1013 if (td->o.time_based)
1014 return 1;
1015 if (td->o.loops) {
1016 td->o.loops--;
1017 return 1;
1018 }
1019
342f4be4 1020 if (ddir_rw_sum(td->io_bytes) < td->o.size)
2e1df07d
JA		 1021 return 1;
1022
1023 return 0;
1024}
1025
1026static int exec_string(const char *string)
1027{
1028 int ret, newlen = strlen(string) + 1 + 8;
1029 char *str;
1030
1031 str = malloc(newlen);
1032 sprintf(str, "sh -c %s", string);
1033
1034 ret = system(str);
1035 if (ret == -1)
1036 log_err("fio: exec of cmd <%s> failed\n", str);
1037
1038 free(str);
1039 return ret;
1040}
1041
1042/*
1043 * Entry point for the thread based jobs. The process based jobs end up
1044 * here as well, after a little setup.
1045 */
1046static void *thread_main(void *data)
1047{
1048 unsigned long long elapsed;
1049 struct thread_data *td = data;
1050 pthread_condattr_t attr;
1051 int clear_state;
1052
1053 if (!td->o.use_thread) {
1054 setsid();
1055 td->pid = getpid();
1056 } else
1057 td->pid = gettid();
1058
5d879392
JA		 1059 fio_local_clock_init(td->o.use_thread);
1060
2e1df07d
JA		 1061 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1062
1063 INIT_FLIST_HEAD(&td->io_u_freelist);
1064 INIT_FLIST_HEAD(&td->io_u_busylist);
1065 INIT_FLIST_HEAD(&td->io_u_requeues);
1066 INIT_FLIST_HEAD(&td->io_log_list);
1067 INIT_FLIST_HEAD(&td->io_hist_list);
1068 INIT_FLIST_HEAD(&td->verify_list);
1069 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1070 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA		 1071 pthread_mutex_init(&td->io_u_lock, NULL);
1072 td->io_hist_tree = RB_ROOT;
1073
1074 pthread_condattr_init(&attr);
1075 pthread_cond_init(&td->verify_cond, &attr);
1076 pthread_cond_init(&td->free_cond, &attr);
1077
1078 td_set_runstate(td, TD_INITIALIZED);
1079 dprint(FD_MUTEX, "up startup_mutex\n");
1080 fio_mutex_up(startup_mutex);
1081 dprint(FD_MUTEX, "wait on td->mutex\n");
1082 fio_mutex_down(td->mutex);
1083 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1084
1085 /*
1086 * the ->mutex mutex is now no longer used, close it to avoid
1087 * eating a file descriptor
1088 */
1089 fio_mutex_remove(td->mutex);
1090
1091 /*
1092 * A new gid requires privilege, so we need to do this before setting
1093 * the uid.
1094 */
1095 if (td->o.gid != -1U && setgid(td->o.gid)) {
1096 td_verror(td, errno, "setgid");
1097 goto err;
1098 }
1099 if (td->o.uid != -1U && setuid(td->o.uid)) {
1100 td_verror(td, errno, "setuid");
1101 goto err;
1102 }
1103
1104 /*
1105 * If we have a gettimeofday() thread, make sure we exclude that
1106 * thread from this job
1107 */
1108 if (td->o.gtod_cpu)
1109 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1110
1111 /*
1112 * Set affinity first, in case it has an impact on the memory
1113 * allocations.
1114 */
1115 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1116 td_verror(td, errno, "cpu_set_affinity");
1117 goto err;
1118 }
1119
67bf9823 1120#ifdef CONFIG_LIBNUMA
d0b937ed
YR		 1121 /* numa node setup */
1122 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1123 int ret;
1124
1125 if (numa_available() < 0) {
1126 td_verror(td, errno, "Does not support NUMA API\n");
1127 goto err;
1128 }
1129
1130 if (td->o.numa_cpumask_set) {
1131 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1132 if (ret == -1) {
1133 td_verror(td, errno, \
1134 "numa_run_on_node_mask failed\n");
1135 goto err;
1136 }
1137 }
1138
1139 if (td->o.numa_memmask_set) {
1140
1141 switch (td->o.numa_mem_mode) {
1142 case MPOL_INTERLEAVE:
1143 numa_set_interleave_mask(td->o.numa_memnodesmask);
1144 break;
1145 case MPOL_BIND:
1146 numa_set_membind(td->o.numa_memnodesmask);
1147 break;
1148 case MPOL_LOCAL:
1149 numa_set_localalloc();
1150 break;
1151 case MPOL_PREFERRED:
1152 numa_set_preferred(td->o.numa_mem_prefer_node);
1153 break;
1154 case MPOL_DEFAULT:
1155 default:
1156 break;
1157 }
1158
1159 }
1160 }
1161#endif
1162
2e1df07d
JA		 1163 /*
1164 * May alter parameters that init_io_u() will use, so we need to
1165 * do this first.
1166 */
1167 if (init_iolog(td))
1168 goto err;
1169
1170 if (init_io_u(td))
1171 goto err;
1172
1173 if (td->o.verify_async && verify_async_init(td))
1174 goto err;
1175
1176 if (td->ioprio_set) {
1177 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1178 td_verror(td, errno, "ioprio_set");
1179 goto err;
1180 }
1181 }
1182
5d89ff79 1183 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA		 1184 goto err;
1185
649c10c5
BC		 1186 errno = 0;
1187 if (nice(td->o.nice) == -1 && errno != 0) {
2e1df07d
JA		 1188 td_verror(td, errno, "nice");
1189 goto err;
1190 }
1191
1192 if (td->o.ioscheduler && switch_ioscheduler(td))
1193 goto err;
1194
1195 if (!td->o.create_serialize && setup_files(td))
1196 goto err;
1197
1198 if (td_io_init(td))
1199 goto err;
1200
1201 if (init_random_map(td))
1202 goto err;
1203
1204 if (td->o.exec_prerun) {
1205 if (exec_string(td->o.exec_prerun))
1206 goto err;
1207 }
1208
1209 if (td->o.pre_read) {
1210 if (pre_read_files(td) < 0)
1211 goto err;
1212 }
1213
1214 fio_gettime(&td->epoch, NULL);
44404c5a 1215 fio_getrusage(&td->ru_start);
2e1df07d
JA		 1216 clear_state = 0;
1217 while (keep_running(td)) {
100f49f1
JA		 1218 uint64_t verify_bytes;
1219
2e1df07d
JA		 1220 fio_gettime(&td->start, NULL);
1221 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1222 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1223 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1224
6eaf09d6
SL		 1225 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1226 td->o.ratemin[DDIR_TRIM]) {
1227 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
2e1df07d 1228 sizeof(td->bw_sample_time));
6eaf09d6
SL		 1229 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1230 sizeof(td->bw_sample_time));
1231 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
2e1df07d
JA		 1232 sizeof(td->bw_sample_time));
1233 }
1234
1235 if (clear_state)
1236 clear_io_state(td);
1237
1238 prune_io_piece_log(td);
1239
100f49f1 1240 verify_bytes = do_io(td);
2e1df07d
JA		 1241
1242 clear_state = 1;
1243
1244 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1245 elapsed = utime_since_now(&td->start);
1246 td->ts.runtime[DDIR_READ] += elapsed;
1247 }
1248 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1249 elapsed = utime_since_now(&td->start);
1250 td->ts.runtime[DDIR_WRITE] += elapsed;
1251 }
6eaf09d6
SL		 1252 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1253 elapsed = utime_since_now(&td->start);
1254 td->ts.runtime[DDIR_TRIM] += elapsed;
1255 }
2e1df07d
JA		 1256
1257 if (td->error || td->terminate)
1258 break;
1259
1260 if (!td->o.do_verify ||
1261 td->o.verify == VERIFY_NONE ||
1262 (td->io_ops->flags & FIO_UNIDIR))
1263 continue;
1264
1265 clear_io_state(td);
1266
1267 fio_gettime(&td->start, NULL);
1268
100f49f1 1269 do_verify(td, verify_bytes);
2e1df07d
JA		 1270
1271 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1272
1273 if (td->error || td->terminate)
1274 break;
1275 }
1276
1277 update_rusage_stat(td);
6eaf09d6
SL		 1278 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1279 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1280 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
2e1df07d 1281 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL		 1282 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1283 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1284 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d
JA		 1285
1286 fio_mutex_down(writeout_mutex);
1287 if (td->bw_log) {
1288 if (td->o.bw_log_file) {
1289 finish_log_named(td, td->bw_log,
1290 td->o.bw_log_file, "bw");
1291 } else
1292 finish_log(td, td->bw_log, "bw");
1293 }
1294 if (td->lat_log) {
1295 if (td->o.lat_log_file) {
1296 finish_log_named(td, td->lat_log,
1297 td->o.lat_log_file, "lat");
1298 } else
1299 finish_log(td, td->lat_log, "lat");
1300 }
1301 if (td->slat_log) {
1302 if (td->o.lat_log_file) {
1303 finish_log_named(td, td->slat_log,
1304 td->o.lat_log_file, "slat");
1305 } else
1306 finish_log(td, td->slat_log, "slat");
1307 }
1308 if (td->clat_log) {
1309 if (td->o.lat_log_file) {
1310 finish_log_named(td, td->clat_log,
1311 td->o.lat_log_file, "clat");
1312 } else
1313 finish_log(td, td->clat_log, "clat");
1314 }
1315 if (td->iops_log) {
1316 if (td->o.iops_log_file) {
1317 finish_log_named(td, td->iops_log,
1318 td->o.iops_log_file, "iops");
1319 } else
1320 finish_log(td, td->iops_log, "iops");
1321 }
1322
1323 fio_mutex_up(writeout_mutex);
1324 if (td->o.exec_postrun)
1325 exec_string(td->o.exec_postrun);
1326
1327 if (exitall_on_terminate)
1328 fio_terminate_threads(td->groupid);
1329
1330err:
1331 if (td->error)
1332 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1333 td->verror);
1334
1335 if (td->o.verify_async)
1336 verify_async_exit(td);
1337
1338 close_and_free_files(td);
2e1df07d 1339 cleanup_io_u(td);
32dbca2c 1340 close_ioengine(td);
2e1df07d
JA		 1341 cgroup_shutdown(td, &cgroup_mnt);
1342
1343 if (td->o.cpumask_set) {
1344 int ret = fio_cpuset_exit(&td->o.cpumask);
1345
1346 td_verror(td, ret, "fio_cpuset_exit");
1347 }
1348
1349 /*
1350 * do this very late, it will log file closing as well
1351 */
1352 if (td->o.write_iolog_file)
1353 write_iolog_close(td);
1354
1355 td_set_runstate(td, TD_EXITED);
e43606c2 1356 return (void *) (uintptr_t) td->error;
2e1df07d
JA		 1357}
1358
1359
1360/*
1361 * We cannot pass the td data into a forked process, so attach the td and
1362 * pass it to the thread worker.
1363 */
1364static int fork_main(int shmid, int offset)
1365{
1366 struct thread_data *td;
1367 void *data, *ret;
1368
1369#ifndef __hpux
1370 data = shmat(shmid, NULL, 0);
1371 if (data == (void *) -1) {
1372 int __err = errno;
1373
1374 perror("shmat");
1375 return __err;
1376 }
1377#else
1378 /*
1379 * HP-UX inherits shm mappings?
1380 */
1381 data = threads;
1382#endif
1383
1384 td = data + offset * sizeof(struct thread_data);
1385 ret = thread_main(td);
1386 shmdt(data);
e43606c2 1387 return (int) (uintptr_t) ret;
2e1df07d
JA		 1388}
1389
1390/*
1391 * Run over the job map and reap the threads that have exited, if any.
1392 */
1393static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1394 unsigned int *m_rate)
1395{
1396 struct thread_data *td;
1397 unsigned int cputhreads, realthreads, pending;
1398 int i, status, ret;
1399
1400 /*
1401 * reap exited threads (TD_EXITED -> TD_REAPED)
1402 */
1403 realthreads = pending = cputhreads = 0;
1404 for_each_td(td, i) {
1405 int flags = 0;
1406
1407 /*
1408 * ->io_ops is NULL for a thread that has closed its
1409 * io engine
1410 */
1411 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1412 cputhreads++;
1413 else
1414 realthreads++;
1415
1416 if (!td->pid) {
1417 pending++;
1418 continue;
1419 }
1420 if (td->runstate == TD_REAPED)
1421 continue;
1422 if (td->o.use_thread) {
1423 if (td->runstate == TD_EXITED) {
1424 td_set_runstate(td, TD_REAPED);
1425 goto reaped;
1426 }
1427 continue;
1428 }
1429
1430 flags = WNOHANG;
1431 if (td->runstate == TD_EXITED)
1432 flags = 0;
1433
1434 /*
1435 * check if someone quit or got killed in an unusual way
1436 */
1437 ret = waitpid(td->pid, &status, flags);
1438 if (ret < 0) {
1439 if (errno == ECHILD) {
1440 log_err("fio: pid=%d disappeared %d\n",
1441 (int) td->pid, td->runstate);
a5e371a6 1442 td->sig = ECHILD;
2e1df07d
JA		 1443 td_set_runstate(td, TD_REAPED);
1444 goto reaped;
1445 }
1446 perror("waitpid");
1447 } else if (ret == td->pid) {
1448 if (WIFSIGNALED(status)) {
1449 int sig = WTERMSIG(status);
1450
36d80bc7 1451 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA		 1452 log_err("fio: pid=%d, got signal=%d\n",
1453 (int) td->pid, sig);
a5e371a6 1454 td->sig = sig;
2e1df07d
JA		 1455 td_set_runstate(td, TD_REAPED);
1456 goto reaped;
1457 }
1458 if (WIFEXITED(status)) {
1459 if (WEXITSTATUS(status) && !td->error)
1460 td->error = WEXITSTATUS(status);
1461
1462 td_set_runstate(td, TD_REAPED);
1463 goto reaped;
1464 }
1465 }
1466
1467 /*
1468 * thread is not dead, continue
1469 */
1470 pending++;
1471 continue;
1472reaped:
1473 (*nr_running)--;
342f4be4
JA		 1474 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1475 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA		 1476 if (!td->pid)
1477 pending--;
1478
1479 if (td->error)
1480 exit_value++;
1481
1482 done_secs += mtime_since_now(&td->epoch) / 1000;
1483 }
1484
1485 if (*nr_running == cputhreads && !pending && realthreads)
1486 fio_terminate_threads(TERMINATE_ALL);
1487}
1488
2e1df07d
JA		 1489/*
1490 * Main function for kicking off and reaping jobs, as needed.
1491 */
1492static void run_threads(void)
1493{
1494 struct thread_data *td;
1495 unsigned long spent;
1496 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1497
1498 if (fio_pin_memory())
1499 return;
1500
1501 if (fio_gtod_offload && fio_start_gtod_thread())
1502 return;
f2a2ce0e
HL		 1503
1504 fio_idle_prof_init();
2e1df07d
JA		 1505
1506 set_sig_handlers();
1507
f3afa57e 1508 if (output_format == FIO_OUTPUT_NORMAL) {
2e1df07d
JA		 1509 log_info("Starting ");
1510 if (nr_thread)
1511 log_info("%d thread%s", nr_thread,
1512 nr_thread > 1 ? "s" : "");
1513 if (nr_process) {
1514 if (nr_thread)
1515 log_info(" and ");
1516 log_info("%d process%s", nr_process,
1517 nr_process > 1 ? "es" : "");
1518 }
1519 log_info("\n");
1520 fflush(stdout);
1521 }
1522
1523 todo = thread_number;
1524 nr_running = 0;
1525 nr_started = 0;
1526 m_rate = t_rate = 0;
1527
1528 for_each_td(td, i) {
1529 print_status_init(td->thread_number - 1);
1530
1531 if (!td->o.create_serialize)
1532 continue;
1533
1534 /*
1535 * do file setup here so it happens sequentially,
1536 * we don't want X number of threads getting their
1537 * client data interspersed on disk
1538 */
1539 if (setup_files(td)) {
1540 exit_value++;
1541 if (td->error)
1542 log_err("fio: pid=%d, err=%d/%s\n",
1543 (int) td->pid, td->error, td->verror);
1544 td_set_runstate(td, TD_REAPED);
1545 todo--;
1546 } else {
1547 struct fio_file *f;
1548 unsigned int j;
1549
1550 /*
1551 * for sharing to work, each job must always open
1552 * its own files. so close them, if we opened them
1553 * for creation
1554 */
1555 for_each_file(td, f, j) {
1556 if (fio_file_open(f))
1557 td_io_close_file(td, f);
1558 }
1559 }
1560 }
1561
f2a2ce0e
HL		 1562 /* start idle threads before io threads start to run */
1563 fio_idle_prof_start();
1564
2e1df07d
JA		 1565 set_genesis_time();
1566
1567 while (todo) {
1568 struct thread_data *map[REAL_MAX_JOBS];
1569 struct timeval this_start;
1570 int this_jobs = 0, left;
1571
1572 /*
1573 * create threads (TD_NOT_CREATED -> TD_CREATED)
1574 */
1575 for_each_td(td, i) {
1576 if (td->runstate != TD_NOT_CREATED)
1577 continue;
1578
1579 /*
1580 * never got a chance to start, killed by other
1581 * thread for some reason
1582 */
1583 if (td->terminate) {
1584 todo--;
1585 continue;
1586 }
1587
1588 if (td->o.start_delay) {
1589 spent = mtime_since_genesis();
1590
1591 if (td->o.start_delay * 1000 > spent)
1592 continue;
1593 }
1594
1595 if (td->o.stonewall && (nr_started || nr_running)) {
1596 dprint(FD_PROCESS, "%s: stonewall wait\n",
1597 td->o.name);
1598 break;
1599 }
1600
1601 init_disk_util(td);
1602
1603 /*
1604 * Set state to created. Thread will transition
1605 * to TD_INITIALIZED when it's done setting up.
1606 */
1607 td_set_runstate(td, TD_CREATED);
1608 map[this_jobs++] = td;
1609 nr_started++;
1610
1611 if (td->o.use_thread) {
1612 int ret;
1613
1614 dprint(FD_PROCESS, "will pthread_create\n");
1615 ret = pthread_create(&td->thread, NULL,
1616 thread_main, td);
1617 if (ret) {
1618 log_err("pthread_create: %s\n",
1619 strerror(ret));
1620 nr_started--;
1621 break;
1622 }
1623 ret = pthread_detach(td->thread);
1624 if (ret)
1625 log_err("pthread_detach: %s",
1626 strerror(ret));
1627 } else {
1628 pid_t pid;
1629 dprint(FD_PROCESS, "will fork\n");
1630 pid = fork();
1631 if (!pid) {
1632 int ret = fork_main(shm_id, i);
1633
1634 _exit(ret);
1635 } else if (i == fio_debug_jobno)
1636 *fio_debug_jobp = pid;
1637 }
1638 dprint(FD_MUTEX, "wait on startup_mutex\n");
1639 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1640 log_err("fio: job startup hung? exiting.\n");
1641 fio_terminate_threads(TERMINATE_ALL);
1642 fio_abort = 1;
1643 nr_started--;
1644 break;
1645 }
1646 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1647 }
1648
1649 /*
1650 * Wait for the started threads to transition to
1651 * TD_INITIALIZED.
1652 */
1653 fio_gettime(&this_start, NULL);
1654 left = this_jobs;
1655 while (left && !fio_abort) {
1656 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1657 break;
1658
1659 usleep(100000);
1660
1661 for (i = 0; i < this_jobs; i++) {
1662 td = map[i];
1663 if (!td)
1664 continue;
1665 if (td->runstate == TD_INITIALIZED) {
1666 map[i] = NULL;
1667 left--;
1668 } else if (td->runstate >= TD_EXITED) {
1669 map[i] = NULL;
1670 left--;
1671 todo--;
1672 nr_running++; /* work-around... */
1673 }
1674 }
1675 }
1676
1677 if (left) {
4e87c37a
JA		 1678 log_err("fio: %d job%s failed to start\n", left,
1679 left > 1 ? "s" : "");
2e1df07d
JA		 1680 for (i = 0; i < this_jobs; i++) {
1681 td = map[i];
1682 if (!td)
1683 continue;
1684 kill(td->pid, SIGTERM);
1685 }
1686 break;
1687 }
1688
1689 /*
1690 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1691 */
1692 for_each_td(td, i) {
1693 if (td->runstate != TD_INITIALIZED)
1694 continue;
1695
1696 if (in_ramp_time(td))
1697 td_set_runstate(td, TD_RAMP);
1698 else
1699 td_set_runstate(td, TD_RUNNING);
1700 nr_running++;
1701 nr_started--;
342f4be4
JA		 1702 m_rate += ddir_rw_sum(td->o.ratemin);
1703 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA		 1704 todo--;
1705 fio_mutex_up(td->mutex);
1706 }
1707
1708 reap_threads(&nr_running, &t_rate, &m_rate);
1709
1710 if (todo) {
1711 if (is_backend)
1712 fio_server_idle_loop();
1713 else
1714 usleep(100000);
1715 }
1716 }
1717
1718 while (nr_running) {
1719 reap_threads(&nr_running, &t_rate, &m_rate);
1720
1721 if (is_backend)
1722 fio_server_idle_loop();
1723 else
1724 usleep(10000);
1725 }
1726
f2a2ce0e
HL		 1727 fio_idle_prof_stop();
1728
2e1df07d
JA		 1729 update_io_ticks();
1730 fio_unpin_memory();
1731}
1732
9ec7779f
JA		 1733void wait_for_disk_thread_exit(void)
1734{
1735 fio_mutex_down(disk_thread_mutex);
1736}
1737
27357187
JA		 1738static void free_disk_util(void)
1739{
1740 disk_util_start_exit();
1741 wait_for_disk_thread_exit();
1742 disk_util_prune_entries();
1743}
1744
2e1df07d
JA		 1745static void *disk_thread_main(void *data)
1746{
9ec7779f
JA		 1747 int ret = 0;
1748
2e1df07d
JA		 1749 fio_mutex_up(startup_mutex);
1750
9ec7779f 1751 while (threads && !ret) {
2e1df07d
JA		 1752 usleep(DISK_UTIL_MSEC * 1000);
1753 if (!threads)
1754 break;
9ec7779f 1755 ret = update_io_ticks();
2e1df07d
JA		 1756
1757 if (!is_backend)
1758 print_thread_status();
1759 }
1760
9ec7779f 1761 fio_mutex_up(disk_thread_mutex);
2e1df07d
JA		 1762 return NULL;
1763}
1764
1765static int create_disk_util_thread(void)
1766{
1767 int ret;
1768
9ec7779f
JA		 1769 setup_disk_util();
1770
521da527 1771 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
9ec7779f 1772
2e1df07d
JA		 1773 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1774 if (ret) {
9ec7779f 1775 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1776 log_err("Can't create disk util thread: %s\n", strerror(ret));
1777 return 1;
1778 }
1779
1780 ret = pthread_detach(disk_util_thread);
1781 if (ret) {
9ec7779f 1782 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1783 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1784 return 1;
1785 }
1786
1787 dprint(FD_MUTEX, "wait on startup_mutex\n");
1788 fio_mutex_down(startup_mutex);
1789 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1790 return 0;
1791}
1792
2e1df07d
JA		 1793int fio_backend(void)
1794{
1795 struct thread_data *td;
1796 int i;
1797
1798 if (exec_profile) {
1799 if (load_profile(exec_profile))
1800 return 1;
1801 free(exec_profile);
1802 exec_profile = NULL;
1803 }
1804 if (!thread_number)
1805 return 0;
1806
1807 if (write_bw_log) {
1808 setup_log(&agg_io_log[DDIR_READ], 0);
1809 setup_log(&agg_io_log[DDIR_WRITE], 0);
6eaf09d6 1810 setup_log(&agg_io_log[DDIR_TRIM], 0);
2e1df07d
JA		 1811 }
1812
521da527 1813 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA		 1814 if (startup_mutex == NULL)
1815 return 1;
521da527 1816 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2e1df07d
JA		 1817 if (writeout_mutex == NULL)
1818 return 1;
1819
1820 set_genesis_time();
1821 create_disk_util_thread();
1822
1823 cgroup_list = smalloc(sizeof(*cgroup_list));
1824 INIT_FLIST_HEAD(cgroup_list);
1825
1826 run_threads();
1827
1828 if (!fio_abort) {
1829 show_run_stats();
1830 if (write_bw_log) {
1831 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1832 __finish_log(agg_io_log[DDIR_WRITE],
1833 "agg-write_bw.log");
6eaf09d6
SL		 1834 __finish_log(agg_io_log[DDIR_TRIM],
1835 "agg-write_bw.log");
2e1df07d
JA		 1836 }
1837 }
1838
1839 for_each_td(td, i)
1840 fio_options_free(td);
1841
a462baef 1842 free_disk_util();
2e1df07d
JA		 1843 cgroup_kill(cgroup_list);
1844 sfree(cgroup_list);
1845 sfree(cgroup_mnt);
1846
1847 fio_mutex_remove(startup_mutex);
1848 fio_mutex_remove(writeout_mutex);
9ec7779f 1849 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1850 return exit_value;
1851}
fio - Flexible IO Tester