libaio: don't call io_destroy(), let exit_aio() take care of it
[fio.git] / backend.c
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>
33 #include <inttypes.h>
34 #include <sys/stat.h>
35 #include <sys/wait.h>
36 #include <sys/ipc.h>
37 #include <sys/mman.h>
38
39 #include "fio.h"
40 #ifndef FIO_NO_HAVE_SHM_H
41 #include <sys/shm.h>
42 #endif
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"
53 #include "lib/getrusage.h"
54 #include "idletime.h"
55 #include "err.h"
56 #include "lib/tp.h"
57 #include "workqueue.h"
58
59 static pthread_t helper_thread;
60 static pthread_mutex_t helper_lock;
61 pthread_cond_t helper_cond;
62 int helper_do_stat = 0;
63
64 static struct fio_mutex *startup_mutex;
65 static struct flist_head *cgroup_list;
66 static char *cgroup_mnt;
67 static int exit_value;
68 static volatile int fio_abort;
69 static unsigned int nr_process = 0;
70 static unsigned int nr_thread = 0;
71
72 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
73
74 int groupid = 0;
75 unsigned int thread_number = 0;
76 unsigned int stat_number = 0;
77 int shm_id = 0;
78 int temp_stall_ts;
79 unsigned long done_secs = 0;
80 volatile int helper_exit = 0;
81
82 #define PAGE_ALIGN(buf) \
83         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
84
85 #define JOB_START_TIMEOUT       (5 * 1000)
86
87 static void sig_int(int sig)
88 {
89         if (threads) {
90                 if (is_backend)
91                         fio_server_got_signal(sig);
92                 else {
93                         log_info("\nfio: terminating on signal %d\n", sig);
94                         log_info_flush();
95                         exit_value = 128;
96                 }
97
98                 fio_terminate_threads(TERMINATE_ALL);
99         }
100 }
101
102 static void sig_show_status(int sig)
103 {
104         show_running_run_stats();
105 }
106
107 static void set_sig_handlers(void)
108 {
109         struct sigaction act;
110
111         memset(&act, 0, sizeof(act));
112         act.sa_handler = sig_int;
113         act.sa_flags = SA_RESTART;
114         sigaction(SIGINT, &act, NULL);
115
116         memset(&act, 0, sizeof(act));
117         act.sa_handler = sig_int;
118         act.sa_flags = SA_RESTART;
119         sigaction(SIGTERM, &act, NULL);
120
121 /* Windows uses SIGBREAK as a quit signal from other applications */
122 #ifdef WIN32
123         memset(&act, 0, sizeof(act));
124         act.sa_handler = sig_int;
125         act.sa_flags = SA_RESTART;
126         sigaction(SIGBREAK, &act, NULL);
127 #endif
128
129         memset(&act, 0, sizeof(act));
130         act.sa_handler = sig_show_status;
131         act.sa_flags = SA_RESTART;
132         sigaction(SIGUSR1, &act, NULL);
133
134         if (is_backend) {
135                 memset(&act, 0, sizeof(act));
136                 act.sa_handler = sig_int;
137                 act.sa_flags = SA_RESTART;
138                 sigaction(SIGPIPE, &act, NULL);
139         }
140 }
141
142 /*
143  * Check if we are above the minimum rate given.
144  */
145 static int __check_min_rate(struct thread_data *td, struct timeval *now,
146                             enum fio_ddir ddir)
147 {
148         unsigned long long bytes = 0;
149         unsigned long iops = 0;
150         unsigned long spent;
151         unsigned long rate;
152         unsigned int ratemin = 0;
153         unsigned int rate_iops = 0;
154         unsigned int rate_iops_min = 0;
155
156         assert(ddir_rw(ddir));
157
158         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
159                 return 0;
160
161         /*
162          * allow a 2 second settle period in the beginning
163          */
164         if (mtime_since(&td->start, now) < 2000)
165                 return 0;
166
167         iops += td->this_io_blocks[ddir];
168         bytes += td->this_io_bytes[ddir];
169         ratemin += td->o.ratemin[ddir];
170         rate_iops += td->o.rate_iops[ddir];
171         rate_iops_min += td->o.rate_iops_min[ddir];
172
173         /*
174          * if rate blocks is set, sample is running
175          */
176         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
177                 spent = mtime_since(&td->lastrate[ddir], now);
178                 if (spent < td->o.ratecycle)
179                         return 0;
180
181                 if (td->o.rate[ddir]) {
182                         /*
183                          * check bandwidth specified rate
184                          */
185                         if (bytes < td->rate_bytes[ddir]) {
186                                 log_err("%s: min rate %u not met\n", td->o.name,
187                                                                 ratemin);
188                                 return 1;
189                         } else {
190                                 if (spent)
191                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
192                                 else
193                                         rate = 0;
194
195                                 if (rate < ratemin ||
196                                     bytes < td->rate_bytes[ddir]) {
197                                         log_err("%s: min rate %u not met, got"
198                                                 " %luKB/sec\n", td->o.name,
199                                                         ratemin, rate);
200                                         return 1;
201                                 }
202                         }
203                 } else {
204                         /*
205                          * checks iops specified rate
206                          */
207                         if (iops < rate_iops) {
208                                 log_err("%s: min iops rate %u not met\n",
209                                                 td->o.name, rate_iops);
210                                 return 1;
211                         } else {
212                                 if (spent)
213                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
214                                 else
215                                         rate = 0;
216
217                                 if (rate < rate_iops_min ||
218                                     iops < td->rate_blocks[ddir]) {
219                                         log_err("%s: min iops rate %u not met,"
220                                                 " got %lu\n", td->o.name,
221                                                         rate_iops_min, rate);
222                                 }
223                         }
224                 }
225         }
226
227         td->rate_bytes[ddir] = bytes;
228         td->rate_blocks[ddir] = iops;
229         memcpy(&td->lastrate[ddir], now, sizeof(*now));
230         return 0;
231 }
232
233 static int check_min_rate(struct thread_data *td, struct timeval *now)
234 {
235         int ret = 0;
236
237         if (td->bytes_done[DDIR_READ])
238                 ret |= __check_min_rate(td, now, DDIR_READ);
239         if (td->bytes_done[DDIR_WRITE])
240                 ret |= __check_min_rate(td, now, DDIR_WRITE);
241         if (td->bytes_done[DDIR_TRIM])
242                 ret |= __check_min_rate(td, now, DDIR_TRIM);
243
244         return ret;
245 }
246
247 /*
248  * When job exits, we can cancel the in-flight IO if we are using async
249  * io. Attempt to do so.
250  */
251 static void cleanup_pending_aio(struct thread_data *td)
252 {
253         int r;
254
255         /*
256          * get immediately available events, if any
257          */
258         r = io_u_queued_complete(td, 0);
259         if (r < 0)
260                 return;
261
262         /*
263          * now cancel remaining active events
264          */
265         if (td->io_ops->cancel) {
266                 struct io_u *io_u;
267                 int i;
268
269                 io_u_qiter(&td->io_u_all, io_u, i) {
270                         if (io_u->flags & IO_U_F_FLIGHT) {
271                                 r = td->io_ops->cancel(td, io_u);
272                                 if (!r)
273                                         put_io_u(td, io_u);
274                         }
275                 }
276         }
277
278         if (td->cur_depth)
279                 r = io_u_queued_complete(td, td->cur_depth);
280 }
281
282 /*
283  * Helper to handle the final sync of a file. Works just like the normal
284  * io path, just does everything sync.
285  */
286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
287 {
288         struct io_u *io_u = __get_io_u(td);
289         int ret;
290
291         if (!io_u)
292                 return 1;
293
294         io_u->ddir = DDIR_SYNC;
295         io_u->file = f;
296
297         if (td_io_prep(td, io_u)) {
298                 put_io_u(td, io_u);
299                 return 1;
300         }
301
302 requeue:
303         ret = td_io_queue(td, io_u);
304         if (ret < 0) {
305                 td_verror(td, io_u->error, "td_io_queue");
306                 put_io_u(td, io_u);
307                 return 1;
308         } else if (ret == FIO_Q_QUEUED) {
309                 if (io_u_queued_complete(td, 1) < 0)
310                         return 1;
311         } else if (ret == FIO_Q_COMPLETED) {
312                 if (io_u->error) {
313                         td_verror(td, io_u->error, "td_io_queue");
314                         return 1;
315                 }
316
317                 if (io_u_sync_complete(td, io_u) < 0)
318                         return 1;
319         } else if (ret == FIO_Q_BUSY) {
320                 if (td_io_commit(td))
321                         return 1;
322                 goto requeue;
323         }
324
325         return 0;
326 }
327
328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
329 {
330         int ret;
331
332         if (fio_file_open(f))
333                 return fio_io_sync(td, f);
334
335         if (td_io_open_file(td, f))
336                 return 1;
337
338         ret = fio_io_sync(td, f);
339         td_io_close_file(td, f);
340         return ret;
341 }
342
343 static inline void __update_tv_cache(struct thread_data *td)
344 {
345         fio_gettime(&td->tv_cache, NULL);
346 }
347
348 static inline void update_tv_cache(struct thread_data *td)
349 {
350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351                 __update_tv_cache(td);
352 }
353
354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
355 {
356         if (in_ramp_time(td))
357                 return 0;
358         if (!td->o.timeout)
359                 return 0;
360         if (utime_since(&td->epoch, t) >= td->o.timeout)
361                 return 1;
362
363         return 0;
364 }
365
366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
367                                int *retptr)
368 {
369         int ret = *retptr;
370
371         if (ret < 0 || td->error) {
372                 int err = td->error;
373                 enum error_type_bit eb;
374
375                 if (ret < 0)
376                         err = -ret;
377
378                 eb = td_error_type(ddir, err);
379                 if (!(td->o.continue_on_error & (1 << eb)))
380                         return 1;
381
382                 if (td_non_fatal_error(td, eb, err)) {
383                         /*
384                          * Continue with the I/Os in case of
385                          * a non fatal error.
386                          */
387                         update_error_count(td, err);
388                         td_clear_error(td);
389                         *retptr = 0;
390                         return 0;
391                 } else if (td->o.fill_device && err == ENOSPC) {
392                         /*
393                          * We expect to hit this error if
394                          * fill_device option is set.
395                          */
396                         td_clear_error(td);
397                         fio_mark_td_terminate(td);
398                         return 1;
399                 } else {
400                         /*
401                          * Stop the I/O in case of a fatal
402                          * error.
403                          */
404                         update_error_count(td, err);
405                         return 1;
406                 }
407         }
408
409         return 0;
410 }
411
412 static void check_update_rusage(struct thread_data *td)
413 {
414         if (td->update_rusage) {
415                 td->update_rusage = 0;
416                 update_rusage_stat(td);
417                 fio_mutex_up(td->rusage_sem);
418         }
419 }
420
421 static int wait_for_completions(struct thread_data *td, struct timeval *time)
422 {
423         const int full = queue_full(td);
424         int min_evts = 0;
425         int ret;
426
427         /*
428          * if the queue is full, we MUST reap at least 1 event
429          */
430         min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
431         if (full && !min_evts)
432                 min_evts = 1;
433
434         if (time && (__should_check_rate(td, DDIR_READ) ||
435             __should_check_rate(td, DDIR_WRITE) ||
436             __should_check_rate(td, DDIR_TRIM)))
437                 fio_gettime(time, NULL);
438
439         do {
440                 ret = io_u_queued_complete(td, min_evts);
441                 if (ret < 0)
442                         break;
443         } while (full && (td->cur_depth > td->o.iodepth_low));
444
445         return ret;
446 }
447
448 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
449                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
450                    struct timeval *comp_time)
451 {
452         int ret2;
453
454         switch (*ret) {
455         case FIO_Q_COMPLETED:
456                 if (io_u->error) {
457                         *ret = -io_u->error;
458                         clear_io_u(td, io_u);
459                 } else if (io_u->resid) {
460                         int bytes = io_u->xfer_buflen - io_u->resid;
461                         struct fio_file *f = io_u->file;
462
463                         if (bytes_issued)
464                                 *bytes_issued += bytes;
465
466                         if (!from_verify)
467                                 trim_io_piece(td, io_u);
468
469                         /*
470                          * zero read, fail
471                          */
472                         if (!bytes) {
473                                 if (!from_verify)
474                                         unlog_io_piece(td, io_u);
475                                 td_verror(td, EIO, "full resid");
476                                 put_io_u(td, io_u);
477                                 break;
478                         }
479
480                         io_u->xfer_buflen = io_u->resid;
481                         io_u->xfer_buf += bytes;
482                         io_u->offset += bytes;
483
484                         if (ddir_rw(io_u->ddir))
485                                 td->ts.short_io_u[io_u->ddir]++;
486
487                         f = io_u->file;
488                         if (io_u->offset == f->real_file_size)
489                                 goto sync_done;
490
491                         requeue_io_u(td, &io_u);
492                 } else {
493 sync_done:
494                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
495                             __should_check_rate(td, DDIR_WRITE) ||
496                             __should_check_rate(td, DDIR_TRIM)))
497                                 fio_gettime(comp_time, NULL);
498
499                         *ret = io_u_sync_complete(td, io_u);
500                         if (*ret < 0)
501                                 break;
502                 }
503                 return 0;
504         case FIO_Q_QUEUED:
505                 /*
506                  * if the engine doesn't have a commit hook,
507                  * the io_u is really queued. if it does have such
508                  * a hook, it has to call io_u_queued() itself.
509                  */
510                 if (td->io_ops->commit == NULL)
511                         io_u_queued(td, io_u);
512                 if (bytes_issued)
513                         *bytes_issued += io_u->xfer_buflen;
514                 break;
515         case FIO_Q_BUSY:
516                 if (!from_verify)
517                         unlog_io_piece(td, io_u);
518                 requeue_io_u(td, &io_u);
519                 ret2 = td_io_commit(td);
520                 if (ret2 < 0)
521                         *ret = ret2;
522                 break;
523         default:
524                 assert(ret < 0);
525                 td_verror(td, -(*ret), "td_io_queue");
526                 break;
527         }
528
529         if (break_on_this_error(td, ddir, ret))
530                 return 1;
531
532         return 0;
533 }
534
535 /*
536  * The main verify engine. Runs over the writes we previously submitted,
537  * reads the blocks back in, and checks the crc/md5 of the data.
538  */
539 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
540 {
541         struct fio_file *f;
542         struct io_u *io_u;
543         int ret, min_events;
544         unsigned int i;
545
546         dprint(FD_VERIFY, "starting loop\n");
547
548         /*
549          * sync io first and invalidate cache, to make sure we really
550          * read from disk.
551          */
552         for_each_file(td, f, i) {
553                 if (!fio_file_open(f))
554                         continue;
555                 if (fio_io_sync(td, f))
556                         break;
557                 if (file_invalidate_cache(td, f))
558                         break;
559         }
560
561         check_update_rusage(td);
562
563         if (td->error)
564                 return;
565
566         td_set_runstate(td, TD_VERIFYING);
567
568         io_u = NULL;
569         while (!td->terminate) {
570                 enum fio_ddir ddir;
571                 int full;
572
573                 update_tv_cache(td);
574                 check_update_rusage(td);
575
576                 if (runtime_exceeded(td, &td->tv_cache)) {
577                         __update_tv_cache(td);
578                         if (runtime_exceeded(td, &td->tv_cache)) {
579                                 fio_mark_td_terminate(td);
580                                 break;
581                         }
582                 }
583
584                 if (flow_threshold_exceeded(td))
585                         continue;
586
587                 if (!td->o.experimental_verify) {
588                         io_u = __get_io_u(td);
589                         if (!io_u)
590                                 break;
591
592                         if (get_next_verify(td, io_u)) {
593                                 put_io_u(td, io_u);
594                                 break;
595                         }
596
597                         if (td_io_prep(td, io_u)) {
598                                 put_io_u(td, io_u);
599                                 break;
600                         }
601                 } else {
602                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
603                                 break;
604
605                         while ((io_u = get_io_u(td)) != NULL) {
606                                 if (IS_ERR(io_u)) {
607                                         io_u = NULL;
608                                         ret = FIO_Q_BUSY;
609                                         goto reap;
610                                 }
611
612                                 /*
613                                  * We are only interested in the places where
614                                  * we wrote or trimmed IOs. Turn those into
615                                  * reads for verification purposes.
616                                  */
617                                 if (io_u->ddir == DDIR_READ) {
618                                         /*
619                                          * Pretend we issued it for rwmix
620                                          * accounting
621                                          */
622                                         td->io_issues[DDIR_READ]++;
623                                         put_io_u(td, io_u);
624                                         continue;
625                                 } else if (io_u->ddir == DDIR_TRIM) {
626                                         io_u->ddir = DDIR_READ;
627                                         io_u_set(io_u, IO_U_F_TRIMMED);
628                                         break;
629                                 } else if (io_u->ddir == DDIR_WRITE) {
630                                         io_u->ddir = DDIR_READ;
631                                         break;
632                                 } else {
633                                         put_io_u(td, io_u);
634                                         continue;
635                                 }
636                         }
637
638                         if (!io_u)
639                                 break;
640                 }
641
642                 if (verify_state_should_stop(td, io_u)) {
643                         put_io_u(td, io_u);
644                         break;
645                 }
646
647                 if (td->o.verify_async)
648                         io_u->end_io = verify_io_u_async;
649                 else
650                         io_u->end_io = verify_io_u;
651
652                 ddir = io_u->ddir;
653                 if (!td->o.disable_slat)
654                         fio_gettime(&io_u->start_time, NULL);
655
656                 ret = td_io_queue(td, io_u);
657
658                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
659                         break;
660
661                 /*
662                  * if we can queue more, do so. but check if there are
663                  * completed io_u's first. Note that we can get BUSY even
664                  * without IO queued, if the system is resource starved.
665                  */
666 reap:
667                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
668                 if (full || !td->o.iodepth_batch_complete)
669                         ret = wait_for_completions(td, NULL);
670
671                 if (ret < 0)
672                         break;
673         }
674
675         check_update_rusage(td);
676
677         if (!td->error) {
678                 min_events = td->cur_depth;
679
680                 if (min_events)
681                         ret = io_u_queued_complete(td, min_events);
682         } else
683                 cleanup_pending_aio(td);
684
685         td_set_runstate(td, TD_RUNNING);
686
687         dprint(FD_VERIFY, "exiting loop\n");
688 }
689
690 static unsigned int exceeds_number_ios(struct thread_data *td)
691 {
692         unsigned long long number_ios;
693
694         if (!td->o.number_ios)
695                 return 0;
696
697         number_ios = ddir_rw_sum(td->io_blocks);
698         number_ios += td->io_u_queued + td->io_u_in_flight;
699
700         return number_ios >= (td->o.number_ios * td->loops);
701 }
702
703 static int io_issue_bytes_exceeded(struct thread_data *td)
704 {
705         unsigned long long bytes, limit;
706
707         if (td_rw(td))
708                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
709         else if (td_write(td))
710                 bytes = td->io_issue_bytes[DDIR_WRITE];
711         else if (td_read(td))
712                 bytes = td->io_issue_bytes[DDIR_READ];
713         else
714                 bytes = td->io_issue_bytes[DDIR_TRIM];
715
716         if (td->o.io_limit)
717                 limit = td->o.io_limit;
718         else
719                 limit = td->o.size;
720
721         limit *= td->loops;
722         return bytes >= limit || exceeds_number_ios(td);
723 }
724
725 static int io_complete_bytes_exceeded(struct thread_data *td)
726 {
727         unsigned long long bytes, limit;
728
729         if (td_rw(td))
730                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
731         else if (td_write(td))
732                 bytes = td->this_io_bytes[DDIR_WRITE];
733         else if (td_read(td))
734                 bytes = td->this_io_bytes[DDIR_READ];
735         else
736                 bytes = td->this_io_bytes[DDIR_TRIM];
737
738         if (td->o.io_limit)
739                 limit = td->o.io_limit;
740         else
741                 limit = td->o.size;
742
743         limit *= td->loops;
744         return bytes >= limit || exceeds_number_ios(td);
745 }
746
747 /*
748  * Main IO worker function. It retrieves io_u's to process and queues
749  * and reaps them, checking for rate and errors along the way.
750  *
751  * Returns number of bytes written and trimmed.
752  */
753 static uint64_t do_io(struct thread_data *td)
754 {
755         unsigned int i;
756         int ret = 0;
757         uint64_t total_bytes, bytes_issued = 0;
758
759         if (in_ramp_time(td))
760                 td_set_runstate(td, TD_RAMP);
761         else
762                 td_set_runstate(td, TD_RUNNING);
763
764         lat_target_init(td);
765
766         total_bytes = td->o.size;
767         /*
768         * Allow random overwrite workloads to write up to io_limit
769         * before starting verification phase as 'size' doesn't apply.
770         */
771         if (td_write(td) && td_random(td) && td->o.norandommap)
772                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
773         /*
774          * If verify_backlog is enabled, we'll run the verify in this
775          * handler as well. For that case, we may need up to twice the
776          * amount of bytes.
777          */
778         if (td->o.verify != VERIFY_NONE &&
779            (td_write(td) && td->o.verify_backlog))
780                 total_bytes += td->o.size;
781
782         /* In trimwrite mode, each byte is trimmed and then written, so
783          * allow total_bytes to be twice as big */
784         if (td_trimwrite(td))
785                 total_bytes += td->total_io_size;
786
787         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
788                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
789                 td->o.time_based) {
790                 struct timeval comp_time;
791                 struct io_u *io_u;
792                 int full;
793                 enum fio_ddir ddir;
794
795                 check_update_rusage(td);
796
797                 if (td->terminate || td->done)
798                         break;
799
800                 update_tv_cache(td);
801
802                 if (runtime_exceeded(td, &td->tv_cache)) {
803                         __update_tv_cache(td);
804                         if (runtime_exceeded(td, &td->tv_cache)) {
805                                 fio_mark_td_terminate(td);
806                                 break;
807                         }
808                 }
809
810                 if (flow_threshold_exceeded(td))
811                         continue;
812
813                 if (bytes_issued >= total_bytes)
814                         break;
815
816                 io_u = get_io_u(td);
817                 if (IS_ERR_OR_NULL(io_u)) {
818                         int err = PTR_ERR(io_u);
819
820                         io_u = NULL;
821                         if (err == -EBUSY) {
822                                 ret = FIO_Q_BUSY;
823                                 goto reap;
824                         }
825                         if (td->o.latency_target)
826                                 goto reap;
827                         break;
828                 }
829
830                 ddir = io_u->ddir;
831
832                 /*
833                  * Add verification end_io handler if:
834                  *      - Asked to verify (!td_rw(td))
835                  *      - Or the io_u is from our verify list (mixed write/ver)
836                  */
837                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
838                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
839
840                         if (!td->o.verify_pattern_bytes) {
841                                 io_u->rand_seed = __rand(&td->verify_state);
842                                 if (sizeof(int) != sizeof(long *))
843                                         io_u->rand_seed *= __rand(&td->verify_state);
844                         }
845
846                         if (verify_state_should_stop(td, io_u)) {
847                                 put_io_u(td, io_u);
848                                 break;
849                         }
850
851                         if (td->o.verify_async)
852                                 io_u->end_io = verify_io_u_async;
853                         else
854                                 io_u->end_io = verify_io_u;
855                         td_set_runstate(td, TD_VERIFYING);
856                 } else if (in_ramp_time(td))
857                         td_set_runstate(td, TD_RAMP);
858                 else
859                         td_set_runstate(td, TD_RUNNING);
860
861                 /*
862                  * Always log IO before it's issued, so we know the specific
863                  * order of it. The logged unit will track when the IO has
864                  * completed.
865                  */
866                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
867                     td->o.do_verify &&
868                     td->o.verify != VERIFY_NONE &&
869                     !td->o.experimental_verify)
870                         log_io_piece(td, io_u);
871
872                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
873                         if (td->error)
874                                 break;
875                         ret = workqueue_enqueue(&td->io_wq, io_u);
876                 } else {
877                         ret = td_io_queue(td, io_u);
878
879                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 1, &comp_time))
880                                 break;
881
882                         /*
883                          * See if we need to complete some commands. Note that
884                          * we can get BUSY even without IO queued, if the
885                          * system is resource starved.
886                          */
887 reap:
888                         full = queue_full(td) ||
889                                 (ret == FIO_Q_BUSY && td->cur_depth);
890                         if (full || !td->o.iodepth_batch_complete)
891                                 ret = wait_for_completions(td, &comp_time);
892                 }
893                 if (ret < 0)
894                         break;
895                 if (!ddir_rw_sum(td->bytes_done) &&
896                     !(td->io_ops->flags & FIO_NOIO))
897                         continue;
898
899                 if (!in_ramp_time(td) && should_check_rate(td)) {
900                         if (check_min_rate(td, &comp_time)) {
901                                 if (exitall_on_terminate)
902                                         fio_terminate_threads(td->groupid);
903                                 td_verror(td, EIO, "check_min_rate");
904                                 break;
905                         }
906                 }
907                 if (!in_ramp_time(td) && td->o.latency_target)
908                         lat_target_check(td);
909
910                 if (td->o.thinktime) {
911                         unsigned long long b;
912
913                         b = ddir_rw_sum(td->io_blocks);
914                         if (!(b % td->o.thinktime_blocks)) {
915                                 int left;
916
917                                 io_u_quiesce(td);
918
919                                 if (td->o.thinktime_spin)
920                                         usec_spin(td->o.thinktime_spin);
921
922                                 left = td->o.thinktime - td->o.thinktime_spin;
923                                 if (left)
924                                         usec_sleep(td, left);
925                         }
926                 }
927         }
928
929         check_update_rusage(td);
930
931         if (td->trim_entries)
932                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
933
934         if (td->o.fill_device && td->error == ENOSPC) {
935                 td->error = 0;
936                 fio_mark_td_terminate(td);
937         }
938         if (!td->error) {
939                 struct fio_file *f;
940
941                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
942                         workqueue_flush(&td->io_wq);
943                         i = 0;
944                 } else
945                         i = td->cur_depth;
946
947                 if (i) {
948                         ret = io_u_queued_complete(td, i);
949                         if (td->o.fill_device && td->error == ENOSPC)
950                                 td->error = 0;
951                 }
952
953                 if (should_fsync(td) && td->o.end_fsync) {
954                         td_set_runstate(td, TD_FSYNCING);
955
956                         for_each_file(td, f, i) {
957                                 if (!fio_file_fsync(td, f))
958                                         continue;
959
960                                 log_err("fio: end_fsync failed for file %s\n",
961                                                                 f->file_name);
962                         }
963                 }
964         } else
965                 cleanup_pending_aio(td);
966
967         /*
968          * stop job if we failed doing any IO
969          */
970         if (!ddir_rw_sum(td->this_io_bytes))
971                 td->done = 1;
972
973         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
974 }
975
976 static void cleanup_io_u(struct thread_data *td)
977 {
978         struct io_u *io_u;
979
980         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
981
982                 if (td->io_ops->io_u_free)
983                         td->io_ops->io_u_free(td, io_u);
984
985                 fio_memfree(io_u, sizeof(*io_u));
986         }
987
988         free_io_mem(td);
989
990         io_u_rexit(&td->io_u_requeues);
991         io_u_qexit(&td->io_u_freelist);
992         io_u_qexit(&td->io_u_all);
993
994         if (td->last_write_comp)
995                 sfree(td->last_write_comp);
996 }
997
998 static int init_io_u(struct thread_data *td)
999 {
1000         struct io_u *io_u;
1001         unsigned int max_bs, min_write;
1002         int cl_align, i, max_units;
1003         int data_xfer = 1, err;
1004         char *p;
1005
1006         max_units = td->o.iodepth;
1007         max_bs = td_max_bs(td);
1008         min_write = td->o.min_bs[DDIR_WRITE];
1009         td->orig_buffer_size = (unsigned long long) max_bs
1010                                         * (unsigned long long) max_units;
1011
1012         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1013                 data_xfer = 0;
1014
1015         err = 0;
1016         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1017         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1018         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1019
1020         if (err) {
1021                 log_err("fio: failed setting up IO queues\n");
1022                 return 1;
1023         }
1024
1025         /*
1026          * if we may later need to do address alignment, then add any
1027          * possible adjustment here so that we don't cause a buffer
1028          * overflow later. this adjustment may be too much if we get
1029          * lucky and the allocator gives us an aligned address.
1030          */
1031         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1032             (td->io_ops->flags & FIO_RAWIO))
1033                 td->orig_buffer_size += page_mask + td->o.mem_align;
1034
1035         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1036                 unsigned long bs;
1037
1038                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1039                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1040         }
1041
1042         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1043                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1044                 return 1;
1045         }
1046
1047         if (data_xfer && allocate_io_mem(td))
1048                 return 1;
1049
1050         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1051             (td->io_ops->flags & FIO_RAWIO))
1052                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1053         else
1054                 p = td->orig_buffer;
1055
1056         cl_align = os_cache_line_size();
1057
1058         for (i = 0; i < max_units; i++) {
1059                 void *ptr;
1060
1061                 if (td->terminate)
1062                         return 1;
1063
1064                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1065                 if (!ptr) {
1066                         log_err("fio: unable to allocate aligned memory\n");
1067                         break;
1068                 }
1069
1070                 io_u = ptr;
1071                 memset(io_u, 0, sizeof(*io_u));
1072                 INIT_FLIST_HEAD(&io_u->verify_list);
1073                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1074
1075                 if (data_xfer) {
1076                         io_u->buf = p;
1077                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1078
1079                         if (td_write(td))
1080                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1081                         if (td_write(td) && td->o.verify_pattern_bytes) {
1082                                 /*
1083                                  * Fill the buffer with the pattern if we are
1084                                  * going to be doing writes.
1085                                  */
1086                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1087                         }
1088                 }
1089
1090                 io_u->index = i;
1091                 io_u->flags = IO_U_F_FREE;
1092                 io_u_qpush(&td->io_u_freelist, io_u);
1093
1094                 /*
1095                  * io_u never leaves this stack, used for iteration of all
1096                  * io_u buffers.
1097                  */
1098                 io_u_qpush(&td->io_u_all, io_u);
1099
1100                 if (td->io_ops->io_u_init) {
1101                         int ret = td->io_ops->io_u_init(td, io_u);
1102
1103                         if (ret) {
1104                                 log_err("fio: failed to init engine data: %d\n", ret);
1105                                 return 1;
1106                         }
1107                 }
1108
1109                 p += max_bs;
1110         }
1111
1112         if (td->o.verify != VERIFY_NONE) {
1113                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1114                 if (!td->last_write_comp) {
1115                         log_err("fio: failed to alloc write comp data\n");
1116                         return 1;
1117                 }
1118         }
1119
1120         return 0;
1121 }
1122
1123 static int switch_ioscheduler(struct thread_data *td)
1124 {
1125         char tmp[256], tmp2[128];
1126         FILE *f;
1127         int ret;
1128
1129         if (td->io_ops->flags & FIO_DISKLESSIO)
1130                 return 0;
1131
1132         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1133
1134         f = fopen(tmp, "r+");
1135         if (!f) {
1136                 if (errno == ENOENT) {
1137                         log_err("fio: os or kernel doesn't support IO scheduler"
1138                                 " switching\n");
1139                         return 0;
1140                 }
1141                 td_verror(td, errno, "fopen iosched");
1142                 return 1;
1143         }
1144
1145         /*
1146          * Set io scheduler.
1147          */
1148         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1149         if (ferror(f) || ret != 1) {
1150                 td_verror(td, errno, "fwrite");
1151                 fclose(f);
1152                 return 1;
1153         }
1154
1155         rewind(f);
1156
1157         /*
1158          * Read back and check that the selected scheduler is now the default.
1159          */
1160         ret = fread(tmp, sizeof(tmp), 1, f);
1161         if (ferror(f) || ret < 0) {
1162                 td_verror(td, errno, "fread");
1163                 fclose(f);
1164                 return 1;
1165         }
1166         tmp[sizeof(tmp) - 1] = '\0';
1167
1168
1169         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1170         if (!strstr(tmp, tmp2)) {
1171                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1172                 td_verror(td, EINVAL, "iosched_switch");
1173                 fclose(f);
1174                 return 1;
1175         }
1176
1177         fclose(f);
1178         return 0;
1179 }
1180
1181 static int keep_running(struct thread_data *td)
1182 {
1183         unsigned long long limit;
1184
1185         if (td->done)
1186                 return 0;
1187         if (td->o.time_based)
1188                 return 1;
1189         if (td->o.loops) {
1190                 td->o.loops--;
1191                 return 1;
1192         }
1193         if (exceeds_number_ios(td))
1194                 return 0;
1195
1196         if (td->o.io_limit)
1197                 limit = td->o.io_limit;
1198         else
1199                 limit = td->o.size;
1200
1201         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1202                 uint64_t diff;
1203
1204                 /*
1205                  * If the difference is less than the minimum IO size, we
1206                  * are done.
1207                  */
1208                 diff = limit - ddir_rw_sum(td->io_bytes);
1209                 if (diff < td_max_bs(td))
1210                         return 0;
1211
1212                 if (fio_files_done(td))
1213                         return 0;
1214
1215                 return 1;
1216         }
1217
1218         return 0;
1219 }
1220
1221 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1222 {
1223         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1224         char *str;
1225
1226         str = malloc(newlen);
1227         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1228
1229         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1230         ret = system(str);
1231         if (ret == -1)
1232                 log_err("fio: exec of cmd <%s> failed\n", str);
1233
1234         free(str);
1235         return ret;
1236 }
1237
1238 /*
1239  * Dry run to compute correct state of numberio for verification.
1240  */
1241 static uint64_t do_dry_run(struct thread_data *td)
1242 {
1243         td_set_runstate(td, TD_RUNNING);
1244
1245         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1246                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1247                 struct io_u *io_u;
1248                 int ret;
1249
1250                 if (td->terminate || td->done)
1251                         break;
1252
1253                 io_u = get_io_u(td);
1254                 if (!io_u)
1255                         break;
1256
1257                 io_u_set(io_u, IO_U_F_FLIGHT);
1258                 io_u->error = 0;
1259                 io_u->resid = 0;
1260                 if (ddir_rw(acct_ddir(io_u)))
1261                         td->io_issues[acct_ddir(io_u)]++;
1262                 if (ddir_rw(io_u->ddir)) {
1263                         io_u_mark_depth(td, 1);
1264                         td->ts.total_io_u[io_u->ddir]++;
1265                 }
1266
1267                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1268                     td->o.do_verify &&
1269                     td->o.verify != VERIFY_NONE &&
1270                     !td->o.experimental_verify)
1271                         log_io_piece(td, io_u);
1272
1273                 ret = io_u_sync_complete(td, io_u);
1274                 (void) ret;
1275         }
1276
1277         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1278 }
1279
1280 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1281 {
1282         const enum fio_ddir ddir = io_u->ddir;
1283         int ret;
1284
1285         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1286
1287         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1288
1289         td->cur_depth++;
1290
1291         ret = td_io_queue(td, io_u);
1292
1293         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1294
1295         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1296
1297         if (ret == FIO_Q_QUEUED)
1298                 ret = io_u_queued_complete(td, 1);
1299
1300         td->cur_depth--;
1301 }
1302
1303 /*
1304  * Entry point for the thread based jobs. The process based jobs end up
1305  * here as well, after a little setup.
1306  */
1307 static void *thread_main(void *data)
1308 {
1309         unsigned long long elapsed;
1310         struct thread_data *td = data;
1311         struct thread_options *o = &td->o;
1312         pthread_condattr_t attr;
1313         int clear_state;
1314         int ret;
1315
1316         if (!o->use_thread) {
1317                 setsid();
1318                 td->pid = getpid();
1319         } else
1320                 td->pid = gettid();
1321
1322         fio_local_clock_init(o->use_thread);
1323
1324         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1325
1326         if (is_backend)
1327                 fio_server_send_start(td);
1328
1329         INIT_FLIST_HEAD(&td->io_log_list);
1330         INIT_FLIST_HEAD(&td->io_hist_list);
1331         INIT_FLIST_HEAD(&td->verify_list);
1332         INIT_FLIST_HEAD(&td->trim_list);
1333         INIT_FLIST_HEAD(&td->next_rand_list);
1334         pthread_mutex_init(&td->io_u_lock, NULL);
1335         td->io_hist_tree = RB_ROOT;
1336
1337         pthread_condattr_init(&attr);
1338         pthread_cond_init(&td->verify_cond, &attr);
1339         pthread_cond_init(&td->free_cond, &attr);
1340
1341         td_set_runstate(td, TD_INITIALIZED);
1342         dprint(FD_MUTEX, "up startup_mutex\n");
1343         fio_mutex_up(startup_mutex);
1344         dprint(FD_MUTEX, "wait on td->mutex\n");
1345         fio_mutex_down(td->mutex);
1346         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1347
1348         /*
1349          * A new gid requires privilege, so we need to do this before setting
1350          * the uid.
1351          */
1352         if (o->gid != -1U && setgid(o->gid)) {
1353                 td_verror(td, errno, "setgid");
1354                 goto err;
1355         }
1356         if (o->uid != -1U && setuid(o->uid)) {
1357                 td_verror(td, errno, "setuid");
1358                 goto err;
1359         }
1360
1361         /*
1362          * If we have a gettimeofday() thread, make sure we exclude that
1363          * thread from this job
1364          */
1365         if (o->gtod_cpu)
1366                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1367
1368         /*
1369          * Set affinity first, in case it has an impact on the memory
1370          * allocations.
1371          */
1372         if (fio_option_is_set(o, cpumask)) {
1373                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1374                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1375                         if (!ret) {
1376                                 log_err("fio: no CPUs set\n");
1377                                 log_err("fio: Try increasing number of available CPUs\n");
1378                                 td_verror(td, EINVAL, "cpus_split");
1379                                 goto err;
1380                         }
1381                 }
1382                 ret = fio_setaffinity(td->pid, o->cpumask);
1383                 if (ret == -1) {
1384                         td_verror(td, errno, "cpu_set_affinity");
1385                         goto err;
1386                 }
1387         }
1388
1389 #ifdef CONFIG_LIBNUMA
1390         /* numa node setup */
1391         if (fio_option_is_set(o, numa_cpunodes) ||
1392             fio_option_is_set(o, numa_memnodes)) {
1393                 struct bitmask *mask;
1394
1395                 if (numa_available() < 0) {
1396                         td_verror(td, errno, "Does not support NUMA API\n");
1397                         goto err;
1398                 }
1399
1400                 if (fio_option_is_set(o, numa_cpunodes)) {
1401                         mask = numa_parse_nodestring(o->numa_cpunodes);
1402                         ret = numa_run_on_node_mask(mask);
1403                         numa_free_nodemask(mask);
1404                         if (ret == -1) {
1405                                 td_verror(td, errno, \
1406                                         "numa_run_on_node_mask failed\n");
1407                                 goto err;
1408                         }
1409                 }
1410
1411                 if (fio_option_is_set(o, numa_memnodes)) {
1412                         mask = NULL;
1413                         if (o->numa_memnodes)
1414                                 mask = numa_parse_nodestring(o->numa_memnodes);
1415
1416                         switch (o->numa_mem_mode) {
1417                         case MPOL_INTERLEAVE:
1418                                 numa_set_interleave_mask(mask);
1419                                 break;
1420                         case MPOL_BIND:
1421                                 numa_set_membind(mask);
1422                                 break;
1423                         case MPOL_LOCAL:
1424                                 numa_set_localalloc();
1425                                 break;
1426                         case MPOL_PREFERRED:
1427                                 numa_set_preferred(o->numa_mem_prefer_node);
1428                                 break;
1429                         case MPOL_DEFAULT:
1430                         default:
1431                                 break;
1432                         }
1433
1434                         if (mask)
1435                                 numa_free_nodemask(mask);
1436
1437                 }
1438         }
1439 #endif
1440
1441         if (fio_pin_memory(td))
1442                 goto err;
1443
1444         /*
1445          * May alter parameters that init_io_u() will use, so we need to
1446          * do this first.
1447          */
1448         if (init_iolog(td))
1449                 goto err;
1450
1451         if (init_io_u(td))
1452                 goto err;
1453
1454         if (o->verify_async && verify_async_init(td))
1455                 goto err;
1456
1457         if (fio_option_is_set(o, ioprio) ||
1458             fio_option_is_set(o, ioprio_class)) {
1459                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1460                 if (ret == -1) {
1461                         td_verror(td, errno, "ioprio_set");
1462                         goto err;
1463                 }
1464         }
1465
1466         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1467                 goto err;
1468
1469         errno = 0;
1470         if (nice(o->nice) == -1 && errno != 0) {
1471                 td_verror(td, errno, "nice");
1472                 goto err;
1473         }
1474
1475         if (o->ioscheduler && switch_ioscheduler(td))
1476                 goto err;
1477
1478         if (!o->create_serialize && setup_files(td))
1479                 goto err;
1480
1481         if (td_io_init(td))
1482                 goto err;
1483
1484         if (init_random_map(td))
1485                 goto err;
1486
1487         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1488                 goto err;
1489
1490         if (o->pre_read) {
1491                 if (pre_read_files(td) < 0)
1492                         goto err;
1493         }
1494
1495         if (td->flags & TD_F_COMPRESS_LOG)
1496                 tp_init(&td->tp_data);
1497
1498         fio_verify_init(td);
1499
1500         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1501             workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1502                 goto err;
1503
1504         fio_gettime(&td->epoch, NULL);
1505         fio_getrusage(&td->ru_start);
1506         clear_state = 0;
1507         while (keep_running(td)) {
1508                 uint64_t verify_bytes;
1509
1510                 fio_gettime(&td->start, NULL);
1511                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1512                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1513                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1514
1515                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1516                                 o->ratemin[DDIR_TRIM]) {
1517                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1518                                                 sizeof(td->bw_sample_time));
1519                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1520                                                 sizeof(td->bw_sample_time));
1521                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1522                                                 sizeof(td->bw_sample_time));
1523                 }
1524
1525                 if (clear_state)
1526                         clear_io_state(td);
1527
1528                 prune_io_piece_log(td);
1529
1530                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1531                         verify_bytes = do_dry_run(td);
1532                 else
1533                         verify_bytes = do_io(td);
1534
1535                 clear_state = 1;
1536
1537                 /*
1538                  * Make sure we've successfully updated the rusage stats
1539                  * before waiting on the stat mutex. Otherwise we could have
1540                  * the stat thread holding stat mutex and waiting for
1541                  * the rusage_sem, which would never get upped because
1542                  * this thread is waiting for the stat mutex.
1543                  */
1544                 check_update_rusage(td);
1545
1546                 fio_mutex_down(stat_mutex);
1547                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1548                         elapsed = mtime_since_now(&td->start);
1549                         td->ts.runtime[DDIR_READ] += elapsed;
1550                 }
1551                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1552                         elapsed = mtime_since_now(&td->start);
1553                         td->ts.runtime[DDIR_WRITE] += elapsed;
1554                 }
1555                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1556                         elapsed = mtime_since_now(&td->start);
1557                         td->ts.runtime[DDIR_TRIM] += elapsed;
1558                 }
1559                 fio_gettime(&td->start, NULL);
1560                 fio_mutex_up(stat_mutex);
1561
1562                 if (td->error || td->terminate)
1563                         break;
1564
1565                 if (!o->do_verify ||
1566                     o->verify == VERIFY_NONE ||
1567                     (td->io_ops->flags & FIO_UNIDIR))
1568                         continue;
1569
1570                 clear_io_state(td);
1571
1572                 fio_gettime(&td->start, NULL);
1573
1574                 do_verify(td, verify_bytes);
1575
1576                 /*
1577                  * See comment further up for why this is done here.
1578                  */
1579                 check_update_rusage(td);
1580
1581                 fio_mutex_down(stat_mutex);
1582                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1583                 fio_gettime(&td->start, NULL);
1584                 fio_mutex_up(stat_mutex);
1585
1586                 if (td->error || td->terminate)
1587                         break;
1588         }
1589
1590         update_rusage_stat(td);
1591         td->ts.total_run_time = mtime_since_now(&td->epoch);
1592         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1593         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1594         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1595
1596         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1597             (td->o.verify != VERIFY_NONE && td_write(td))) {
1598                 struct all_io_list *state;
1599                 size_t sz;
1600
1601                 state = get_all_io_list(td->thread_number, &sz);
1602                 if (state) {
1603                         __verify_save_state(state, "local");
1604                         free(state);
1605                 }
1606         }
1607
1608         fio_unpin_memory(td);
1609
1610         fio_writeout_logs(td);
1611
1612         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1613                 workqueue_exit(&td->io_wq);
1614
1615         if (td->flags & TD_F_COMPRESS_LOG)
1616                 tp_exit(&td->tp_data);
1617
1618         if (o->exec_postrun)
1619                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1620
1621         if (exitall_on_terminate)
1622                 fio_terminate_threads(td->groupid);
1623
1624 err:
1625         if (td->error)
1626                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1627                                                         td->verror);
1628
1629         if (o->verify_async)
1630                 verify_async_exit(td);
1631
1632         close_and_free_files(td);
1633         cleanup_io_u(td);
1634         close_ioengine(td);
1635         cgroup_shutdown(td, &cgroup_mnt);
1636         verify_free_state(td);
1637
1638         if (fio_option_is_set(o, cpumask)) {
1639                 ret = fio_cpuset_exit(&o->cpumask);
1640                 if (ret)
1641                         td_verror(td, ret, "fio_cpuset_exit");
1642         }
1643
1644         /*
1645          * do this very late, it will log file closing as well
1646          */
1647         if (o->write_iolog_file)
1648                 write_iolog_close(td);
1649
1650         fio_mutex_remove(td->mutex);
1651         td->mutex = NULL;
1652
1653         td_set_runstate(td, TD_EXITED);
1654
1655         /*
1656          * Do this last after setting our runstate to exited, so we
1657          * know that the stat thread is signaled.
1658          */
1659         check_update_rusage(td);
1660
1661         return (void *) (uintptr_t) td->error;
1662 }
1663
1664
1665 /*
1666  * We cannot pass the td data into a forked process, so attach the td and
1667  * pass it to the thread worker.
1668  */
1669 static int fork_main(int shmid, int offset)
1670 {
1671         struct thread_data *td;
1672         void *data, *ret;
1673
1674 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1675         data = shmat(shmid, NULL, 0);
1676         if (data == (void *) -1) {
1677                 int __err = errno;
1678
1679                 perror("shmat");
1680                 return __err;
1681         }
1682 #else
1683         /*
1684          * HP-UX inherits shm mappings?
1685          */
1686         data = threads;
1687 #endif
1688
1689         td = data + offset * sizeof(struct thread_data);
1690         ret = thread_main(td);
1691         shmdt(data);
1692         return (int) (uintptr_t) ret;
1693 }
1694
1695 static void dump_td_info(struct thread_data *td)
1696 {
1697         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1698                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1699                         (unsigned long) time_since_now(&td->terminate_time));
1700 }
1701
1702 /*
1703  * Run over the job map and reap the threads that have exited, if any.
1704  */
1705 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1706                          unsigned int *m_rate)
1707 {
1708         struct thread_data *td;
1709         unsigned int cputhreads, realthreads, pending;
1710         int i, status, ret;
1711
1712         /*
1713          * reap exited threads (TD_EXITED -> TD_REAPED)
1714          */
1715         realthreads = pending = cputhreads = 0;
1716         for_each_td(td, i) {
1717                 int flags = 0;
1718
1719                 /*
1720                  * ->io_ops is NULL for a thread that has closed its
1721                  * io engine
1722                  */
1723                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1724                         cputhreads++;
1725                 else
1726                         realthreads++;
1727
1728                 if (!td->pid) {
1729                         pending++;
1730                         continue;
1731                 }
1732                 if (td->runstate == TD_REAPED)
1733                         continue;
1734                 if (td->o.use_thread) {
1735                         if (td->runstate == TD_EXITED) {
1736                                 td_set_runstate(td, TD_REAPED);
1737                                 goto reaped;
1738                         }
1739                         continue;
1740                 }
1741
1742                 flags = WNOHANG;
1743                 if (td->runstate == TD_EXITED)
1744                         flags = 0;
1745
1746                 /*
1747                  * check if someone quit or got killed in an unusual way
1748                  */
1749                 ret = waitpid(td->pid, &status, flags);
1750                 if (ret < 0) {
1751                         if (errno == ECHILD) {
1752                                 log_err("fio: pid=%d disappeared %d\n",
1753                                                 (int) td->pid, td->runstate);
1754                                 td->sig = ECHILD;
1755                                 td_set_runstate(td, TD_REAPED);
1756                                 goto reaped;
1757                         }
1758                         perror("waitpid");
1759                 } else if (ret == td->pid) {
1760                         if (WIFSIGNALED(status)) {
1761                                 int sig = WTERMSIG(status);
1762
1763                                 if (sig != SIGTERM && sig != SIGUSR2)
1764                                         log_err("fio: pid=%d, got signal=%d\n",
1765                                                         (int) td->pid, sig);
1766                                 td->sig = sig;
1767                                 td_set_runstate(td, TD_REAPED);
1768                                 goto reaped;
1769                         }
1770                         if (WIFEXITED(status)) {
1771                                 if (WEXITSTATUS(status) && !td->error)
1772                                         td->error = WEXITSTATUS(status);
1773
1774                                 td_set_runstate(td, TD_REAPED);
1775                                 goto reaped;
1776                         }
1777                 }
1778
1779                 /*
1780                  * If the job is stuck, do a forceful timeout of it and
1781                  * move on.
1782                  */
1783                 if (td->terminate &&
1784                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1785                         dump_td_info(td);
1786                         td_set_runstate(td, TD_REAPED);
1787                         goto reaped;
1788                 }
1789
1790                 /*
1791                  * thread is not dead, continue
1792                  */
1793                 pending++;
1794                 continue;
1795 reaped:
1796                 (*nr_running)--;
1797                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1798                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1799                 if (!td->pid)
1800                         pending--;
1801
1802                 if (td->error)
1803                         exit_value++;
1804
1805                 done_secs += mtime_since_now(&td->epoch) / 1000;
1806                 profile_td_exit(td);
1807         }
1808
1809         if (*nr_running == cputhreads && !pending && realthreads)
1810                 fio_terminate_threads(TERMINATE_ALL);
1811 }
1812
1813 static int __check_trigger_file(void)
1814 {
1815         struct stat sb;
1816
1817         if (!trigger_file)
1818                 return 0;
1819
1820         if (stat(trigger_file, &sb))
1821                 return 0;
1822
1823         if (unlink(trigger_file) < 0)
1824                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1825                                                         strerror(errno));
1826
1827         return 1;
1828 }
1829
1830 static int trigger_timedout(void)
1831 {
1832         if (trigger_timeout)
1833                 return time_since_genesis() >= trigger_timeout;
1834
1835         return 0;
1836 }
1837
1838 void exec_trigger(const char *cmd)
1839 {
1840         int ret;
1841
1842         if (!cmd)
1843                 return;
1844
1845         ret = system(cmd);
1846         if (ret == -1)
1847                 log_err("fio: failed executing %s trigger\n", cmd);
1848 }
1849
1850 void check_trigger_file(void)
1851 {
1852         if (__check_trigger_file() || trigger_timedout()) {
1853                 if (nr_clients)
1854                         fio_clients_send_trigger(trigger_remote_cmd);
1855                 else {
1856                         verify_save_state();
1857                         fio_terminate_threads(TERMINATE_ALL);
1858                         exec_trigger(trigger_cmd);
1859                 }
1860         }
1861 }
1862
1863 static int fio_verify_load_state(struct thread_data *td)
1864 {
1865         int ret;
1866
1867         if (!td->o.verify_state)
1868                 return 0;
1869
1870         if (is_backend) {
1871                 void *data;
1872
1873                 ret = fio_server_get_verify_state(td->o.name,
1874                                         td->thread_number - 1, &data);
1875                 if (!ret)
1876                         verify_convert_assign_state(td, data);
1877         } else
1878                 ret = verify_load_state(td, "local");
1879
1880         return ret;
1881 }
1882
1883 static void do_usleep(unsigned int usecs)
1884 {
1885         check_for_running_stats();
1886         check_trigger_file();
1887         usleep(usecs);
1888 }
1889
1890 /*
1891  * Main function for kicking off and reaping jobs, as needed.
1892  */
1893 static void run_threads(void)
1894 {
1895         struct thread_data *td;
1896         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1897         uint64_t spent;
1898
1899         if (fio_gtod_offload && fio_start_gtod_thread())
1900                 return;
1901
1902         fio_idle_prof_init();
1903
1904         set_sig_handlers();
1905
1906         nr_thread = nr_process = 0;
1907         for_each_td(td, i) {
1908                 if (td->o.use_thread)
1909                         nr_thread++;
1910                 else
1911                         nr_process++;
1912         }
1913
1914         if (output_format == FIO_OUTPUT_NORMAL) {
1915                 log_info("Starting ");
1916                 if (nr_thread)
1917                         log_info("%d thread%s", nr_thread,
1918                                                 nr_thread > 1 ? "s" : "");
1919                 if (nr_process) {
1920                         if (nr_thread)
1921                                 log_info(" and ");
1922                         log_info("%d process%s", nr_process,
1923                                                 nr_process > 1 ? "es" : "");
1924                 }
1925                 log_info("\n");
1926                 log_info_flush();
1927         }
1928
1929         todo = thread_number;
1930         nr_running = 0;
1931         nr_started = 0;
1932         m_rate = t_rate = 0;
1933
1934         for_each_td(td, i) {
1935                 print_status_init(td->thread_number - 1);
1936
1937                 if (!td->o.create_serialize)
1938                         continue;
1939
1940                 if (fio_verify_load_state(td))
1941                         goto reap;
1942
1943                 /*
1944                  * do file setup here so it happens sequentially,
1945                  * we don't want X number of threads getting their
1946                  * client data interspersed on disk
1947                  */
1948                 if (setup_files(td)) {
1949 reap:
1950                         exit_value++;
1951                         if (td->error)
1952                                 log_err("fio: pid=%d, err=%d/%s\n",
1953                                         (int) td->pid, td->error, td->verror);
1954                         td_set_runstate(td, TD_REAPED);
1955                         todo--;
1956                 } else {
1957                         struct fio_file *f;
1958                         unsigned int j;
1959
1960                         /*
1961                          * for sharing to work, each job must always open
1962                          * its own files. so close them, if we opened them
1963                          * for creation
1964                          */
1965                         for_each_file(td, f, j) {
1966                                 if (fio_file_open(f))
1967                                         td_io_close_file(td, f);
1968                         }
1969                 }
1970         }
1971
1972         /* start idle threads before io threads start to run */
1973         fio_idle_prof_start();
1974
1975         set_genesis_time();
1976
1977         while (todo) {
1978                 struct thread_data *map[REAL_MAX_JOBS];
1979                 struct timeval this_start;
1980                 int this_jobs = 0, left;
1981
1982                 /*
1983                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1984                  */
1985                 for_each_td(td, i) {
1986                         if (td->runstate != TD_NOT_CREATED)
1987                                 continue;
1988
1989                         /*
1990                          * never got a chance to start, killed by other
1991                          * thread for some reason
1992                          */
1993                         if (td->terminate) {
1994                                 todo--;
1995                                 continue;
1996                         }
1997
1998                         if (td->o.start_delay) {
1999                                 spent = utime_since_genesis();
2000
2001                                 if (td->o.start_delay > spent)
2002                                         continue;
2003                         }
2004
2005                         if (td->o.stonewall && (nr_started || nr_running)) {
2006                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2007                                                         td->o.name);
2008                                 break;
2009                         }
2010
2011                         init_disk_util(td);
2012
2013                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2014                         td->update_rusage = 0;
2015
2016                         /*
2017                          * Set state to created. Thread will transition
2018                          * to TD_INITIALIZED when it's done setting up.
2019                          */
2020                         td_set_runstate(td, TD_CREATED);
2021                         map[this_jobs++] = td;
2022                         nr_started++;
2023
2024                         if (td->o.use_thread) {
2025                                 int ret;
2026
2027                                 dprint(FD_PROCESS, "will pthread_create\n");
2028                                 ret = pthread_create(&td->thread, NULL,
2029                                                         thread_main, td);
2030                                 if (ret) {
2031                                         log_err("pthread_create: %s\n",
2032                                                         strerror(ret));
2033                                         nr_started--;
2034                                         break;
2035                                 }
2036                                 ret = pthread_detach(td->thread);
2037                                 if (ret)
2038                                         log_err("pthread_detach: %s",
2039                                                         strerror(ret));
2040                         } else {
2041                                 pid_t pid;
2042                                 dprint(FD_PROCESS, "will fork\n");
2043                                 pid = fork();
2044                                 if (!pid) {
2045                                         int ret = fork_main(shm_id, i);
2046
2047                                         _exit(ret);
2048                                 } else if (i == fio_debug_jobno)
2049                                         *fio_debug_jobp = pid;
2050                         }
2051                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2052                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2053                                 log_err("fio: job startup hung? exiting.\n");
2054                                 fio_terminate_threads(TERMINATE_ALL);
2055                                 fio_abort = 1;
2056                                 nr_started--;
2057                                 break;
2058                         }
2059                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2060                 }
2061
2062                 /*
2063                  * Wait for the started threads to transition to
2064                  * TD_INITIALIZED.
2065                  */
2066                 fio_gettime(&this_start, NULL);
2067                 left = this_jobs;
2068                 while (left && !fio_abort) {
2069                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2070                                 break;
2071
2072                         do_usleep(100000);
2073
2074                         for (i = 0; i < this_jobs; i++) {
2075                                 td = map[i];
2076                                 if (!td)
2077                                         continue;
2078                                 if (td->runstate == TD_INITIALIZED) {
2079                                         map[i] = NULL;
2080                                         left--;
2081                                 } else if (td->runstate >= TD_EXITED) {
2082                                         map[i] = NULL;
2083                                         left--;
2084                                         todo--;
2085                                         nr_running++; /* work-around... */
2086                                 }
2087                         }
2088                 }
2089
2090                 if (left) {
2091                         log_err("fio: %d job%s failed to start\n", left,
2092                                         left > 1 ? "s" : "");
2093                         for (i = 0; i < this_jobs; i++) {
2094                                 td = map[i];
2095                                 if (!td)
2096                                         continue;
2097                                 kill(td->pid, SIGTERM);
2098                         }
2099                         break;
2100                 }
2101
2102                 /*
2103                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2104                  */
2105                 for_each_td(td, i) {
2106                         if (td->runstate != TD_INITIALIZED)
2107                                 continue;
2108
2109                         if (in_ramp_time(td))
2110                                 td_set_runstate(td, TD_RAMP);
2111                         else
2112                                 td_set_runstate(td, TD_RUNNING);
2113                         nr_running++;
2114                         nr_started--;
2115                         m_rate += ddir_rw_sum(td->o.ratemin);
2116                         t_rate += ddir_rw_sum(td->o.rate);
2117                         todo--;
2118                         fio_mutex_up(td->mutex);
2119                 }
2120
2121                 reap_threads(&nr_running, &t_rate, &m_rate);
2122
2123                 if (todo)
2124                         do_usleep(100000);
2125         }
2126
2127         while (nr_running) {
2128                 reap_threads(&nr_running, &t_rate, &m_rate);
2129                 do_usleep(10000);
2130         }
2131
2132         fio_idle_prof_stop();
2133
2134         update_io_ticks();
2135 }
2136
2137 static void wait_for_helper_thread_exit(void)
2138 {
2139         void *ret;
2140
2141         helper_exit = 1;
2142         pthread_cond_signal(&helper_cond);
2143         pthread_join(helper_thread, &ret);
2144 }
2145
2146 static void free_disk_util(void)
2147 {
2148         disk_util_prune_entries();
2149
2150         pthread_cond_destroy(&helper_cond);
2151 }
2152
2153 static void *helper_thread_main(void *data)
2154 {
2155         int ret = 0;
2156
2157         fio_mutex_up(startup_mutex);
2158
2159         while (!ret) {
2160                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2161                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2162                 struct timespec ts;
2163                 struct timeval tv;
2164
2165                 gettimeofday(&tv, NULL);
2166                 ts.tv_sec = tv.tv_sec + sec;
2167                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2168
2169                 if (ts.tv_nsec >= 1000000000ULL) {
2170                         ts.tv_nsec -= 1000000000ULL;
2171                         ts.tv_sec++;
2172                 }
2173
2174                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2175
2176                 ret = update_io_ticks();
2177
2178                 if (helper_do_stat) {
2179                         helper_do_stat = 0;
2180                         __show_running_run_stats();
2181                 }
2182
2183                 if (!is_backend)
2184                         print_thread_status();
2185         }
2186
2187         return NULL;
2188 }
2189
2190 static int create_helper_thread(void)
2191 {
2192         int ret;
2193
2194         setup_disk_util();
2195
2196         pthread_cond_init(&helper_cond, NULL);
2197         pthread_mutex_init(&helper_lock, NULL);
2198
2199         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2200         if (ret) {
2201                 log_err("Can't create helper thread: %s\n", strerror(ret));
2202                 return 1;
2203         }
2204
2205         dprint(FD_MUTEX, "wait on startup_mutex\n");
2206         fio_mutex_down(startup_mutex);
2207         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2208         return 0;
2209 }
2210
2211 int fio_backend(void)
2212 {
2213         struct thread_data *td;
2214         int i;
2215
2216         if (exec_profile) {
2217                 if (load_profile(exec_profile))
2218                         return 1;
2219                 free(exec_profile);
2220                 exec_profile = NULL;
2221         }
2222         if (!thread_number)
2223                 return 0;
2224
2225         if (write_bw_log) {
2226                 struct log_params p = {
2227                         .log_type = IO_LOG_TYPE_BW,
2228                 };
2229
2230                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2231                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2232                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2233         }
2234
2235         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2236         if (startup_mutex == NULL)
2237                 return 1;
2238
2239         set_genesis_time();
2240         stat_init();
2241         create_helper_thread();
2242
2243         cgroup_list = smalloc(sizeof(*cgroup_list));
2244         INIT_FLIST_HEAD(cgroup_list);
2245
2246         run_threads();
2247
2248         wait_for_helper_thread_exit();
2249
2250         if (!fio_abort) {
2251                 __show_run_stats();
2252                 if (write_bw_log) {
2253                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2254                                 struct io_log *log = agg_io_log[i];
2255
2256                                 flush_log(log);
2257                                 free_log(log);
2258                         }
2259                 }
2260         }
2261
2262         for_each_td(td, i) {
2263                 fio_options_free(td);
2264                 if (td->rusage_sem) {
2265                         fio_mutex_remove(td->rusage_sem);
2266                         td->rusage_sem = NULL;
2267                 }
2268         }
2269
2270         free_disk_util();
2271         cgroup_kill(cgroup_list);
2272         sfree(cgroup_list);
2273         sfree(cgroup_mnt);
2274
2275         fio_mutex_remove(startup_mutex);
2276         stat_exit();
2277         return exit_value;
2278 }