Add options to have fio latency profile a device
[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
56 static pthread_t disk_util_thread;
57 static struct fio_mutex *disk_thread_mutex;
58 static struct fio_mutex *startup_mutex;
59 static struct fio_mutex *writeout_mutex;
60 static struct flist_head *cgroup_list;
61 static char *cgroup_mnt;
62 static int exit_value;
63 static volatile int fio_abort;
64 static unsigned int nr_process = 0;
65 static unsigned int nr_thread = 0;
66
67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
68
69 int groupid = 0;
70 unsigned int thread_number = 0;
71 unsigned int stat_number = 0;
72 int shm_id = 0;
73 int temp_stall_ts;
74 unsigned long done_secs = 0;
75 volatile int disk_util_exit = 0;
76
77 #define PAGE_ALIGN(buf) \
78         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
79
80 #define JOB_START_TIMEOUT       (5 * 1000)
81
82 static 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
97 static void sig_show_status(int sig)
98 {
99         show_running_run_stats();
100 }
101
102 static 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
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
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
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  */
140 static 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
220 static int check_min_rate(struct thread_data *td, struct timeval *now,
221                           uint64_t *bytes_done)
222 {
223         int ret = 0;
224
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);
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  */
239 static void cleanup_pending_aio(struct thread_data *td)
240 {
241         int r;
242
243         /*
244          * get immediately available events, if any
245          */
246         r = io_u_queued_complete(td, 0, NULL);
247         if (r < 0)
248                 return;
249
250         /*
251          * now cancel remaining active events
252          */
253         if (td->io_ops->cancel) {
254                 struct io_u *io_u;
255                 int i;
256
257                 io_u_qiter(&td->io_u_all, io_u, i) {
258                         if (io_u->flags & IO_U_F_FLIGHT) {
259                                 r = td->io_ops->cancel(td, io_u);
260                                 if (!r)
261                                         put_io_u(td, io_u);
262                         }
263                 }
264         }
265
266         if (td->cur_depth)
267                 r = io_u_queued_complete(td, td->cur_depth, NULL);
268 }
269
270 /*
271  * Helper to handle the final sync of a file. Works just like the normal
272  * io path, just does everything sync.
273  */
274 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
275 {
276         struct io_u *io_u = __get_io_u(td);
277         int ret;
278
279         if (!io_u)
280                 return 1;
281
282         io_u->ddir = DDIR_SYNC;
283         io_u->file = f;
284
285         if (td_io_prep(td, io_u)) {
286                 put_io_u(td, io_u);
287                 return 1;
288         }
289
290 requeue:
291         ret = td_io_queue(td, io_u);
292         if (ret < 0) {
293                 td_verror(td, io_u->error, "td_io_queue");
294                 put_io_u(td, io_u);
295                 return 1;
296         } else if (ret == FIO_Q_QUEUED) {
297                 if (io_u_queued_complete(td, 1, NULL) < 0)
298                         return 1;
299         } else if (ret == FIO_Q_COMPLETED) {
300                 if (io_u->error) {
301                         td_verror(td, io_u->error, "td_io_queue");
302                         return 1;
303                 }
304
305                 if (io_u_sync_complete(td, io_u, NULL) < 0)
306                         return 1;
307         } else if (ret == FIO_Q_BUSY) {
308                 if (td_io_commit(td))
309                         return 1;
310                 goto requeue;
311         }
312
313         return 0;
314 }
315
316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
317 {
318         int ret;
319
320         if (fio_file_open(f))
321                 return fio_io_sync(td, f);
322
323         if (td_io_open_file(td, f))
324                 return 1;
325
326         ret = fio_io_sync(td, f);
327         td_io_close_file(td, f);
328         return ret;
329 }
330
331 static inline void __update_tv_cache(struct thread_data *td)
332 {
333         fio_gettime(&td->tv_cache, NULL);
334 }
335
336 static inline void update_tv_cache(struct thread_data *td)
337 {
338         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
339                 __update_tv_cache(td);
340 }
341
342 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
343 {
344         if (in_ramp_time(td))
345                 return 0;
346         if (!td->o.timeout)
347                 return 0;
348         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
349                 return 1;
350
351         return 0;
352 }
353
354 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
355                                int *retptr)
356 {
357         int ret = *retptr;
358
359         if (ret < 0 || td->error) {
360                 int err = td->error;
361                 enum error_type_bit eb;
362
363                 if (ret < 0)
364                         err = -ret;
365
366                 eb = td_error_type(ddir, err);
367                 if (!(td->o.continue_on_error & (1 << eb)))
368                         return 1;
369
370                 if (td_non_fatal_error(td, eb, err)) {
371                         /*
372                          * Continue with the I/Os in case of
373                          * a non fatal error.
374                          */
375                         update_error_count(td, err);
376                         td_clear_error(td);
377                         *retptr = 0;
378                         return 0;
379                 } else if (td->o.fill_device && err == ENOSPC) {
380                         /*
381                          * We expect to hit this error if
382                          * fill_device option is set.
383                          */
384                         td_clear_error(td);
385                         td->terminate = 1;
386                         return 1;
387                 } else {
388                         /*
389                          * Stop the I/O in case of a fatal
390                          * error.
391                          */
392                         update_error_count(td, err);
393                         return 1;
394                 }
395         }
396
397         return 0;
398 }
399
400 static void check_update_rusage(struct thread_data *td)
401 {
402         if (td->update_rusage) {
403                 td->update_rusage = 0;
404                 update_rusage_stat(td);
405                 fio_mutex_up(td->rusage_sem);
406         }
407 }
408
409 /*
410  * The main verify engine. Runs over the writes we previously submitted,
411  * reads the blocks back in, and checks the crc/md5 of the data.
412  */
413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
414 {
415         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
416         struct fio_file *f;
417         struct io_u *io_u;
418         int ret, min_events;
419         unsigned int i;
420
421         dprint(FD_VERIFY, "starting loop\n");
422
423         /*
424          * sync io first and invalidate cache, to make sure we really
425          * read from disk.
426          */
427         for_each_file(td, f, i) {
428                 if (!fio_file_open(f))
429                         continue;
430                 if (fio_io_sync(td, f))
431                         break;
432                 if (file_invalidate_cache(td, f))
433                         break;
434         }
435
436         check_update_rusage(td);
437
438         if (td->error)
439                 return;
440
441         td_set_runstate(td, TD_VERIFYING);
442
443         io_u = NULL;
444         while (!td->terminate) {
445                 enum fio_ddir ddir;
446                 int ret2, full;
447
448                 update_tv_cache(td);
449                 check_update_rusage(td);
450
451                 if (runtime_exceeded(td, &td->tv_cache)) {
452                         __update_tv_cache(td);
453                         if (runtime_exceeded(td, &td->tv_cache)) {
454                                 td->terminate = 1;
455                                 break;
456                         }
457                 }
458
459                 if (flow_threshold_exceeded(td))
460                         continue;
461
462                 if (!td->o.experimental_verify) {
463                         io_u = __get_io_u(td);
464                         if (!io_u)
465                                 break;
466
467                         if (get_next_verify(td, io_u)) {
468                                 put_io_u(td, io_u);
469                                 break;
470                         }
471
472                         if (td_io_prep(td, io_u)) {
473                                 put_io_u(td, io_u);
474                                 break;
475                         }
476                 } else {
477                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
478                                 break;
479
480                         while ((io_u = get_io_u(td)) != NULL) {
481                                 /*
482                                  * We are only interested in the places where
483                                  * we wrote or trimmed IOs. Turn those into
484                                  * reads for verification purposes.
485                                  */
486                                 if (io_u->ddir == DDIR_READ) {
487                                         /*
488                                          * Pretend we issued it for rwmix
489                                          * accounting
490                                          */
491                                         td->io_issues[DDIR_READ]++;
492                                         put_io_u(td, io_u);
493                                         continue;
494                                 } else if (io_u->ddir == DDIR_TRIM) {
495                                         io_u->ddir = DDIR_READ;
496                                         io_u->flags |= IO_U_F_TRIMMED;
497                                         break;
498                                 } else if (io_u->ddir == DDIR_WRITE) {
499                                         io_u->ddir = DDIR_READ;
500                                         break;
501                                 } else {
502                                         put_io_u(td, io_u);
503                                         continue;
504                                 }
505                         }
506
507                         if (!io_u)
508                                 break;
509                 }
510
511                 if (td->o.verify_async)
512                         io_u->end_io = verify_io_u_async;
513                 else
514                         io_u->end_io = verify_io_u;
515
516                 ddir = io_u->ddir;
517
518                 ret = td_io_queue(td, io_u);
519                 switch (ret) {
520                 case FIO_Q_COMPLETED:
521                         if (io_u->error) {
522                                 ret = -io_u->error;
523                                 clear_io_u(td, io_u);
524                         } else if (io_u->resid) {
525                                 int bytes = io_u->xfer_buflen - io_u->resid;
526
527                                 /*
528                                  * zero read, fail
529                                  */
530                                 if (!bytes) {
531                                         td_verror(td, EIO, "full resid");
532                                         put_io_u(td, io_u);
533                                         break;
534                                 }
535
536                                 io_u->xfer_buflen = io_u->resid;
537                                 io_u->xfer_buf += bytes;
538                                 io_u->offset += bytes;
539
540                                 if (ddir_rw(io_u->ddir))
541                                         td->ts.short_io_u[io_u->ddir]++;
542
543                                 f = io_u->file;
544                                 if (io_u->offset == f->real_file_size)
545                                         goto sync_done;
546
547                                 requeue_io_u(td, &io_u);
548                         } else {
549 sync_done:
550                                 ret = io_u_sync_complete(td, io_u, bytes_done);
551                                 if (ret < 0)
552                                         break;
553                         }
554                         continue;
555                 case FIO_Q_QUEUED:
556                         break;
557                 case FIO_Q_BUSY:
558                         requeue_io_u(td, &io_u);
559                         ret2 = td_io_commit(td);
560                         if (ret2 < 0)
561                                 ret = ret2;
562                         break;
563                 default:
564                         assert(ret < 0);
565                         td_verror(td, -ret, "td_io_queue");
566                         break;
567                 }
568
569                 if (break_on_this_error(td, ddir, &ret))
570                         break;
571
572                 /*
573                  * if we can queue more, do so. but check if there are
574                  * completed io_u's first. Note that we can get BUSY even
575                  * without IO queued, if the system is resource starved.
576                  */
577                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
578                 if (full || !td->o.iodepth_batch_complete) {
579                         min_events = min(td->o.iodepth_batch_complete,
580                                          td->cur_depth);
581                         /*
582                          * if the queue is full, we MUST reap at least 1 event
583                          */
584                         if (full && !min_events)
585                                 min_events = 1;
586
587                         do {
588                                 /*
589                                  * Reap required number of io units, if any,
590                                  * and do the verification on them through
591                                  * the callback handler
592                                  */
593                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
594                                         ret = -1;
595                                         break;
596                                 }
597                         } while (full && (td->cur_depth > td->o.iodepth_low));
598                 }
599                 if (ret < 0)
600                         break;
601         }
602
603         check_update_rusage(td);
604
605         if (!td->error) {
606                 min_events = td->cur_depth;
607
608                 if (min_events)
609                         ret = io_u_queued_complete(td, min_events, NULL);
610         } else
611                 cleanup_pending_aio(td);
612
613         td_set_runstate(td, TD_RUNNING);
614
615         dprint(FD_VERIFY, "exiting loop\n");
616 }
617
618 static int io_bytes_exceeded(struct thread_data *td)
619 {
620         unsigned long long bytes;
621
622         if (td_rw(td))
623                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
624         else if (td_write(td))
625                 bytes = td->this_io_bytes[DDIR_WRITE];
626         else if (td_read(td))
627                 bytes = td->this_io_bytes[DDIR_READ];
628         else
629                 bytes = td->this_io_bytes[DDIR_TRIM];
630
631         return bytes >= td->o.size;
632 }
633
634 /*
635  * Main IO worker function. It retrieves io_u's to process and queues
636  * and reaps them, checking for rate and errors along the way.
637  *
638  * Returns number of bytes written and trimmed.
639  */
640 static uint64_t do_io(struct thread_data *td)
641 {
642         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
643         unsigned int i;
644         int ret = 0;
645         uint64_t bytes_issued = 0;
646
647         if (in_ramp_time(td))
648                 td_set_runstate(td, TD_RAMP);
649         else
650                 td_set_runstate(td, TD_RUNNING);
651
652         lat_target_init(td);
653
654         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
655                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
656                 td->o.time_based) {
657                 struct timeval comp_time;
658                 int min_evts = 0;
659                 struct io_u *io_u;
660                 int ret2, full;
661                 enum fio_ddir ddir;
662
663                 check_update_rusage(td);
664
665                 if (td->terminate || td->done)
666                         break;
667
668                 update_tv_cache(td);
669
670                 if (runtime_exceeded(td, &td->tv_cache)) {
671                         __update_tv_cache(td);
672                         if (runtime_exceeded(td, &td->tv_cache)) {
673                                 td->terminate = 1;
674                                 break;
675                         }
676                 }
677
678                 if (flow_threshold_exceeded(td))
679                         continue;
680
681                 if (bytes_issued >= (uint64_t) td->o.size)
682                         break;
683
684                 io_u = get_io_u(td);
685                 if (!io_u) {
686                         if (td->o.latency_target)
687                                 goto reap;
688                         break;
689                 }
690
691                 ddir = io_u->ddir;
692
693                 /*
694                  * Add verification end_io handler if:
695                  *      - Asked to verify (!td_rw(td))
696                  *      - Or the io_u is from our verify list (mixed write/ver)
697                  */
698                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
699                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
700                         if (td->o.verify_async)
701                                 io_u->end_io = verify_io_u_async;
702                         else
703                                 io_u->end_io = verify_io_u;
704                         td_set_runstate(td, TD_VERIFYING);
705                 } else if (in_ramp_time(td))
706                         td_set_runstate(td, TD_RAMP);
707                 else
708                         td_set_runstate(td, TD_RUNNING);
709
710                 ret = td_io_queue(td, io_u);
711                 switch (ret) {
712                 case FIO_Q_COMPLETED:
713                         if (io_u->error) {
714                                 ret = -io_u->error;
715                                 clear_io_u(td, io_u);
716                         } else if (io_u->resid) {
717                                 int bytes = io_u->xfer_buflen - io_u->resid;
718                                 struct fio_file *f = io_u->file;
719
720                                 bytes_issued += bytes;
721                                 /*
722                                  * zero read, fail
723                                  */
724                                 if (!bytes) {
725                                         td_verror(td, EIO, "full resid");
726                                         put_io_u(td, io_u);
727                                         break;
728                                 }
729
730                                 io_u->xfer_buflen = io_u->resid;
731                                 io_u->xfer_buf += bytes;
732                                 io_u->offset += bytes;
733
734                                 if (ddir_rw(io_u->ddir))
735                                         td->ts.short_io_u[io_u->ddir]++;
736
737                                 if (io_u->offset == f->real_file_size)
738                                         goto sync_done;
739
740                                 requeue_io_u(td, &io_u);
741                         } else {
742 sync_done:
743                                 if (__should_check_rate(td, DDIR_READ) ||
744                                     __should_check_rate(td, DDIR_WRITE) ||
745                                     __should_check_rate(td, DDIR_TRIM))
746                                         fio_gettime(&comp_time, NULL);
747
748                                 ret = io_u_sync_complete(td, io_u, bytes_done);
749                                 if (ret < 0)
750                                         break;
751                                 bytes_issued += io_u->xfer_buflen;
752                         }
753                         break;
754                 case FIO_Q_QUEUED:
755                         /*
756                          * if the engine doesn't have a commit hook,
757                          * the io_u is really queued. if it does have such
758                          * a hook, it has to call io_u_queued() itself.
759                          */
760                         if (td->io_ops->commit == NULL)
761                                 io_u_queued(td, io_u);
762                         bytes_issued += io_u->xfer_buflen;
763                         break;
764                 case FIO_Q_BUSY:
765                         requeue_io_u(td, &io_u);
766                         ret2 = td_io_commit(td);
767                         if (ret2 < 0)
768                                 ret = ret2;
769                         break;
770                 default:
771                         assert(ret < 0);
772                         put_io_u(td, io_u);
773                         break;
774                 }
775
776                 if (break_on_this_error(td, ddir, &ret))
777                         break;
778
779                 /*
780                  * See if we need to complete some commands. Note that we
781                  * can get BUSY even without IO queued, if the system is
782                  * resource starved.
783                  */
784 reap:
785                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
786                 if (full || !td->o.iodepth_batch_complete) {
787                         min_evts = min(td->o.iodepth_batch_complete,
788                                         td->cur_depth);
789                         /*
790                          * if the queue is full, we MUST reap at least 1 event
791                          */
792                         if (full && !min_evts)
793                                 min_evts = 1;
794
795                         if (__should_check_rate(td, DDIR_READ) ||
796                             __should_check_rate(td, DDIR_WRITE) ||
797                             __should_check_rate(td, DDIR_TRIM))
798                                 fio_gettime(&comp_time, NULL);
799
800                         do {
801                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
802                                 if (ret < 0)
803                                         break;
804
805                         } while (full && (td->cur_depth > td->o.iodepth_low));
806                 }
807
808                 if (ret < 0)
809                         break;
810                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
811                         continue;
812
813                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
814                         if (check_min_rate(td, &comp_time, bytes_done)) {
815                                 if (exitall_on_terminate)
816                                         fio_terminate_threads(td->groupid);
817                                 td_verror(td, EIO, "check_min_rate");
818                                 break;
819                         }
820                 }
821                 if (!in_ramp_time(td) && td->o.latency_target)
822                         lat_target_check(td);
823
824                 if (td->o.thinktime) {
825                         unsigned long long b;
826
827                         b = ddir_rw_sum(td->io_blocks);
828                         if (!(b % td->o.thinktime_blocks)) {
829                                 int left;
830
831                                 io_u_quiesce(td);
832
833                                 if (td->o.thinktime_spin)
834                                         usec_spin(td->o.thinktime_spin);
835
836                                 left = td->o.thinktime - td->o.thinktime_spin;
837                                 if (left)
838                                         usec_sleep(td, left);
839                         }
840                 }
841         }
842
843         check_update_rusage(td);
844
845         if (td->trim_entries)
846                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
847
848         if (td->o.fill_device && td->error == ENOSPC) {
849                 td->error = 0;
850                 td->terminate = 1;
851         }
852         if (!td->error) {
853                 struct fio_file *f;
854
855                 i = td->cur_depth;
856                 if (i) {
857                         ret = io_u_queued_complete(td, i, bytes_done);
858                         if (td->o.fill_device && td->error == ENOSPC)
859                                 td->error = 0;
860                 }
861
862                 if (should_fsync(td) && td->o.end_fsync) {
863                         td_set_runstate(td, TD_FSYNCING);
864
865                         for_each_file(td, f, i) {
866                                 if (!fio_file_fsync(td, f))
867                                         continue;
868
869                                 log_err("fio: end_fsync failed for file %s\n",
870                                                                 f->file_name);
871                         }
872                 }
873         } else
874                 cleanup_pending_aio(td);
875
876         /*
877          * stop job if we failed doing any IO
878          */
879         if (!ddir_rw_sum(td->this_io_bytes))
880                 td->done = 1;
881
882         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
883 }
884
885 static void cleanup_io_u(struct thread_data *td)
886 {
887         struct io_u *io_u;
888
889         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
890
891                 if (td->io_ops->io_u_free)
892                         td->io_ops->io_u_free(td, io_u);
893
894                 fio_memfree(io_u, sizeof(*io_u));
895         }
896
897         free_io_mem(td);
898
899         io_u_rexit(&td->io_u_requeues);
900         io_u_qexit(&td->io_u_freelist);
901         io_u_qexit(&td->io_u_all);
902 }
903
904 static int init_io_u(struct thread_data *td)
905 {
906         struct io_u *io_u;
907         unsigned int max_bs, min_write;
908         int cl_align, i, max_units;
909         int data_xfer = 1, err;
910         char *p;
911
912         max_units = td->o.iodepth;
913         max_bs = td_max_bs(td);
914         min_write = td->o.min_bs[DDIR_WRITE];
915         td->orig_buffer_size = (unsigned long long) max_bs
916                                         * (unsigned long long) max_units;
917
918         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
919                 data_xfer = 0;
920
921         err = 0;
922         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
923         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
924         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
925
926         if (err) {
927                 log_err("fio: failed setting up IO queues\n");
928                 return 1;
929         }
930
931         /*
932          * if we may later need to do address alignment, then add any
933          * possible adjustment here so that we don't cause a buffer
934          * overflow later. this adjustment may be too much if we get
935          * lucky and the allocator gives us an aligned address.
936          */
937         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
938             (td->io_ops->flags & FIO_RAWIO))
939                 td->orig_buffer_size += page_mask + td->o.mem_align;
940
941         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
942                 unsigned long bs;
943
944                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
945                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
946         }
947
948         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
949                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
950                 return 1;
951         }
952
953         if (data_xfer && allocate_io_mem(td))
954                 return 1;
955
956         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
957             (td->io_ops->flags & FIO_RAWIO))
958                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
959         else
960                 p = td->orig_buffer;
961
962         cl_align = os_cache_line_size();
963
964         for (i = 0; i < max_units; i++) {
965                 void *ptr;
966
967                 if (td->terminate)
968                         return 1;
969
970                 ptr = fio_memalign(cl_align, sizeof(*io_u));
971                 if (!ptr) {
972                         log_err("fio: unable to allocate aligned memory\n");
973                         break;
974                 }
975
976                 io_u = ptr;
977                 memset(io_u, 0, sizeof(*io_u));
978                 INIT_FLIST_HEAD(&io_u->verify_list);
979                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
980
981                 if (data_xfer) {
982                         io_u->buf = p;
983                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
984
985                         if (td_write(td))
986                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
987                         if (td_write(td) && td->o.verify_pattern_bytes) {
988                                 /*
989                                  * Fill the buffer with the pattern if we are
990                                  * going to be doing writes.
991                                  */
992                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
993                         }
994                 }
995
996                 io_u->index = i;
997                 io_u->flags = IO_U_F_FREE;
998                 io_u_qpush(&td->io_u_freelist, io_u);
999
1000                 /*
1001                  * io_u never leaves this stack, used for iteration of all
1002                  * io_u buffers.
1003                  */
1004                 io_u_qpush(&td->io_u_all, io_u);
1005
1006                 if (td->io_ops->io_u_init) {
1007                         int ret = td->io_ops->io_u_init(td, io_u);
1008
1009                         if (ret) {
1010                                 log_err("fio: failed to init engine data: %d\n", ret);
1011                                 return 1;
1012                         }
1013                 }
1014
1015                 p += max_bs;
1016         }
1017
1018         return 0;
1019 }
1020
1021 static int switch_ioscheduler(struct thread_data *td)
1022 {
1023         char tmp[256], tmp2[128];
1024         FILE *f;
1025         int ret;
1026
1027         if (td->io_ops->flags & FIO_DISKLESSIO)
1028                 return 0;
1029
1030         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1031
1032         f = fopen(tmp, "r+");
1033         if (!f) {
1034                 if (errno == ENOENT) {
1035                         log_err("fio: os or kernel doesn't support IO scheduler"
1036                                 " switching\n");
1037                         return 0;
1038                 }
1039                 td_verror(td, errno, "fopen iosched");
1040                 return 1;
1041         }
1042
1043         /*
1044          * Set io scheduler.
1045          */
1046         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1047         if (ferror(f) || ret != 1) {
1048                 td_verror(td, errno, "fwrite");
1049                 fclose(f);
1050                 return 1;
1051         }
1052
1053         rewind(f);
1054
1055         /*
1056          * Read back and check that the selected scheduler is now the default.
1057          */
1058         ret = fread(tmp, 1, sizeof(tmp), f);
1059         if (ferror(f) || ret < 0) {
1060                 td_verror(td, errno, "fread");
1061                 fclose(f);
1062                 return 1;
1063         }
1064
1065         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1066         if (!strstr(tmp, tmp2)) {
1067                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1068                 td_verror(td, EINVAL, "iosched_switch");
1069                 fclose(f);
1070                 return 1;
1071         }
1072
1073         fclose(f);
1074         return 0;
1075 }
1076
1077 static int keep_running(struct thread_data *td)
1078 {
1079         if (td->done)
1080                 return 0;
1081         if (td->o.time_based)
1082                 return 1;
1083         if (td->o.loops) {
1084                 td->o.loops--;
1085                 return 1;
1086         }
1087
1088         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1089                 uint64_t diff;
1090
1091                 /*
1092                  * If the difference is less than the minimum IO size, we
1093                  * are done.
1094                  */
1095                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1096                 if (diff < td_max_bs(td))
1097                         return 0;
1098
1099                 return 1;
1100         }
1101
1102         return 0;
1103 }
1104
1105 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1106 {
1107         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1108         char *str;
1109
1110         str = malloc(newlen);
1111         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1112
1113         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1114         ret = system(str);
1115         if (ret == -1)
1116                 log_err("fio: exec of cmd <%s> failed\n", str);
1117
1118         free(str);
1119         return ret;
1120 }
1121
1122 /*
1123  * Entry point for the thread based jobs. The process based jobs end up
1124  * here as well, after a little setup.
1125  */
1126 static void *thread_main(void *data)
1127 {
1128         unsigned long long elapsed;
1129         struct thread_data *td = data;
1130         struct thread_options *o = &td->o;
1131         pthread_condattr_t attr;
1132         int clear_state;
1133         int ret;
1134
1135         if (!o->use_thread) {
1136                 setsid();
1137                 td->pid = getpid();
1138         } else
1139                 td->pid = gettid();
1140
1141         /*
1142          * fio_time_init() may not have been called yet if running as a server
1143          */
1144         fio_time_init();
1145
1146         fio_local_clock_init(o->use_thread);
1147
1148         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1149
1150         if (is_backend)
1151                 fio_server_send_start(td);
1152
1153         INIT_FLIST_HEAD(&td->io_log_list);
1154         INIT_FLIST_HEAD(&td->io_hist_list);
1155         INIT_FLIST_HEAD(&td->verify_list);
1156         INIT_FLIST_HEAD(&td->trim_list);
1157         INIT_FLIST_HEAD(&td->next_rand_list);
1158         pthread_mutex_init(&td->io_u_lock, NULL);
1159         td->io_hist_tree = RB_ROOT;
1160
1161         pthread_condattr_init(&attr);
1162         pthread_cond_init(&td->verify_cond, &attr);
1163         pthread_cond_init(&td->free_cond, &attr);
1164
1165         td_set_runstate(td, TD_INITIALIZED);
1166         dprint(FD_MUTEX, "up startup_mutex\n");
1167         fio_mutex_up(startup_mutex);
1168         dprint(FD_MUTEX, "wait on td->mutex\n");
1169         fio_mutex_down(td->mutex);
1170         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1171
1172         /*
1173          * the ->mutex mutex is now no longer used, close it to avoid
1174          * eating a file descriptor
1175          */
1176         fio_mutex_remove(td->mutex);
1177         td->mutex = NULL;
1178
1179         /*
1180          * A new gid requires privilege, so we need to do this before setting
1181          * the uid.
1182          */
1183         if (o->gid != -1U && setgid(o->gid)) {
1184                 td_verror(td, errno, "setgid");
1185                 goto err;
1186         }
1187         if (o->uid != -1U && setuid(o->uid)) {
1188                 td_verror(td, errno, "setuid");
1189                 goto err;
1190         }
1191
1192         /*
1193          * If we have a gettimeofday() thread, make sure we exclude that
1194          * thread from this job
1195          */
1196         if (o->gtod_cpu)
1197                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1198
1199         /*
1200          * Set affinity first, in case it has an impact on the memory
1201          * allocations.
1202          */
1203         if (o->cpumask_set) {
1204                 ret = fio_setaffinity(td->pid, o->cpumask);
1205                 if (ret == -1) {
1206                         td_verror(td, errno, "cpu_set_affinity");
1207                         goto err;
1208                 }
1209         }
1210
1211 #ifdef CONFIG_LIBNUMA
1212         /* numa node setup */
1213         if (o->numa_cpumask_set || o->numa_memmask_set) {
1214                 int ret;
1215
1216                 if (numa_available() < 0) {
1217                         td_verror(td, errno, "Does not support NUMA API\n");
1218                         goto err;
1219                 }
1220
1221                 if (o->numa_cpumask_set) {
1222                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1223                         if (ret == -1) {
1224                                 td_verror(td, errno, \
1225                                         "numa_run_on_node_mask failed\n");
1226                                 goto err;
1227                         }
1228                 }
1229
1230                 if (o->numa_memmask_set) {
1231
1232                         switch (o->numa_mem_mode) {
1233                         case MPOL_INTERLEAVE:
1234                                 numa_set_interleave_mask(o->numa_memnodesmask);
1235                                 break;
1236                         case MPOL_BIND:
1237                                 numa_set_membind(o->numa_memnodesmask);
1238                                 break;
1239                         case MPOL_LOCAL:
1240                                 numa_set_localalloc();
1241                                 break;
1242                         case MPOL_PREFERRED:
1243                                 numa_set_preferred(o->numa_mem_prefer_node);
1244                                 break;
1245                         case MPOL_DEFAULT:
1246                         default:
1247                                 break;
1248                         }
1249
1250                 }
1251         }
1252 #endif
1253
1254         if (fio_pin_memory(td))
1255                 goto err;
1256
1257         /*
1258          * May alter parameters that init_io_u() will use, so we need to
1259          * do this first.
1260          */
1261         if (init_iolog(td))
1262                 goto err;
1263
1264         if (init_io_u(td))
1265                 goto err;
1266
1267         if (o->verify_async && verify_async_init(td))
1268                 goto err;
1269
1270         if (o->ioprio) {
1271                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1272                 if (ret == -1) {
1273                         td_verror(td, errno, "ioprio_set");
1274                         goto err;
1275                 }
1276         }
1277
1278         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1279                 goto err;
1280
1281         errno = 0;
1282         if (nice(o->nice) == -1 && errno != 0) {
1283                 td_verror(td, errno, "nice");
1284                 goto err;
1285         }
1286
1287         if (o->ioscheduler && switch_ioscheduler(td))
1288                 goto err;
1289
1290         if (!o->create_serialize && setup_files(td))
1291                 goto err;
1292
1293         if (td_io_init(td))
1294                 goto err;
1295
1296         if (init_random_map(td))
1297                 goto err;
1298
1299         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1300                 goto err;
1301
1302         if (o->pre_read) {
1303                 if (pre_read_files(td) < 0)
1304                         goto err;
1305         }
1306
1307         fio_verify_init(td);
1308
1309         fio_gettime(&td->epoch, NULL);
1310         fio_getrusage(&td->ru_start);
1311         clear_state = 0;
1312         while (keep_running(td)) {
1313                 uint64_t verify_bytes;
1314
1315                 fio_gettime(&td->start, NULL);
1316                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1317                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1318                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1319
1320                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1321                                 o->ratemin[DDIR_TRIM]) {
1322                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1323                                                 sizeof(td->bw_sample_time));
1324                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1325                                                 sizeof(td->bw_sample_time));
1326                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1327                                                 sizeof(td->bw_sample_time));
1328                 }
1329
1330                 if (clear_state)
1331                         clear_io_state(td);
1332
1333                 prune_io_piece_log(td);
1334
1335                 verify_bytes = do_io(td);
1336
1337                 clear_state = 1;
1338
1339                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1340                         elapsed = utime_since_now(&td->start);
1341                         td->ts.runtime[DDIR_READ] += elapsed;
1342                 }
1343                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1344                         elapsed = utime_since_now(&td->start);
1345                         td->ts.runtime[DDIR_WRITE] += elapsed;
1346                 }
1347                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1348                         elapsed = utime_since_now(&td->start);
1349                         td->ts.runtime[DDIR_TRIM] += elapsed;
1350                 }
1351
1352                 if (td->error || td->terminate)
1353                         break;
1354
1355                 if (!o->do_verify ||
1356                     o->verify == VERIFY_NONE ||
1357                     (td->io_ops->flags & FIO_UNIDIR))
1358                         continue;
1359
1360                 clear_io_state(td);
1361
1362                 fio_gettime(&td->start, NULL);
1363
1364                 do_verify(td, verify_bytes);
1365
1366                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1367
1368                 if (td->error || td->terminate)
1369                         break;
1370         }
1371
1372         update_rusage_stat(td);
1373         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1374         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1375         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1376         td->ts.total_run_time = mtime_since_now(&td->epoch);
1377         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1378         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1379         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1380
1381         fio_unpin_memory(td);
1382
1383         fio_mutex_down(writeout_mutex);
1384         if (td->bw_log) {
1385                 if (o->bw_log_file) {
1386                         finish_log_named(td, td->bw_log,
1387                                                 o->bw_log_file, "bw");
1388                 } else
1389                         finish_log(td, td->bw_log, "bw");
1390         }
1391         if (td->lat_log) {
1392                 if (o->lat_log_file) {
1393                         finish_log_named(td, td->lat_log,
1394                                                 o->lat_log_file, "lat");
1395                 } else
1396                         finish_log(td, td->lat_log, "lat");
1397         }
1398         if (td->slat_log) {
1399                 if (o->lat_log_file) {
1400                         finish_log_named(td, td->slat_log,
1401                                                 o->lat_log_file, "slat");
1402                 } else
1403                         finish_log(td, td->slat_log, "slat");
1404         }
1405         if (td->clat_log) {
1406                 if (o->lat_log_file) {
1407                         finish_log_named(td, td->clat_log,
1408                                                 o->lat_log_file, "clat");
1409                 } else
1410                         finish_log(td, td->clat_log, "clat");
1411         }
1412         if (td->iops_log) {
1413                 if (o->iops_log_file) {
1414                         finish_log_named(td, td->iops_log,
1415                                                 o->iops_log_file, "iops");
1416                 } else
1417                         finish_log(td, td->iops_log, "iops");
1418         }
1419
1420         fio_mutex_up(writeout_mutex);
1421         if (o->exec_postrun)
1422                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1423
1424         if (exitall_on_terminate)
1425                 fio_terminate_threads(td->groupid);
1426
1427 err:
1428         if (td->error)
1429                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1430                                                         td->verror);
1431
1432         if (o->verify_async)
1433                 verify_async_exit(td);
1434
1435         close_and_free_files(td);
1436         cleanup_io_u(td);
1437         close_ioengine(td);
1438         cgroup_shutdown(td, &cgroup_mnt);
1439
1440         if (o->cpumask_set) {
1441                 int ret = fio_cpuset_exit(&o->cpumask);
1442
1443                 td_verror(td, ret, "fio_cpuset_exit");
1444         }
1445
1446         /*
1447          * do this very late, it will log file closing as well
1448          */
1449         if (o->write_iolog_file)
1450                 write_iolog_close(td);
1451
1452         fio_mutex_remove(td->rusage_sem);
1453         td->rusage_sem = NULL;
1454
1455         td_set_runstate(td, TD_EXITED);
1456         return (void *) (uintptr_t) td->error;
1457 }
1458
1459
1460 /*
1461  * We cannot pass the td data into a forked process, so attach the td and
1462  * pass it to the thread worker.
1463  */
1464 static int fork_main(int shmid, int offset)
1465 {
1466         struct thread_data *td;
1467         void *data, *ret;
1468
1469 #ifndef __hpux
1470         data = shmat(shmid, NULL, 0);
1471         if (data == (void *) -1) {
1472                 int __err = errno;
1473
1474                 perror("shmat");
1475                 return __err;
1476         }
1477 #else
1478         /*
1479          * HP-UX inherits shm mappings?
1480          */
1481         data = threads;
1482 #endif
1483
1484         td = data + offset * sizeof(struct thread_data);
1485         ret = thread_main(td);
1486         shmdt(data);
1487         return (int) (uintptr_t) ret;
1488 }
1489
1490 /*
1491  * Run over the job map and reap the threads that have exited, if any.
1492  */
1493 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1494                          unsigned int *m_rate)
1495 {
1496         struct thread_data *td;
1497         unsigned int cputhreads, realthreads, pending;
1498         int i, status, ret;
1499
1500         /*
1501          * reap exited threads (TD_EXITED -> TD_REAPED)
1502          */
1503         realthreads = pending = cputhreads = 0;
1504         for_each_td(td, i) {
1505                 int flags = 0;
1506
1507                 /*
1508                  * ->io_ops is NULL for a thread that has closed its
1509                  * io engine
1510                  */
1511                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1512                         cputhreads++;
1513                 else
1514                         realthreads++;
1515
1516                 if (!td->pid) {
1517                         pending++;
1518                         continue;
1519                 }
1520                 if (td->runstate == TD_REAPED)
1521                         continue;
1522                 if (td->o.use_thread) {
1523                         if (td->runstate == TD_EXITED) {
1524                                 td_set_runstate(td, TD_REAPED);
1525                                 goto reaped;
1526                         }
1527                         continue;
1528                 }
1529
1530                 flags = WNOHANG;
1531                 if (td->runstate == TD_EXITED)
1532                         flags = 0;
1533
1534                 /*
1535                  * check if someone quit or got killed in an unusual way
1536                  */
1537                 ret = waitpid(td->pid, &status, flags);
1538                 if (ret < 0) {
1539                         if (errno == ECHILD) {
1540                                 log_err("fio: pid=%d disappeared %d\n",
1541                                                 (int) td->pid, td->runstate);
1542                                 td->sig = ECHILD;
1543                                 td_set_runstate(td, TD_REAPED);
1544                                 goto reaped;
1545                         }
1546                         perror("waitpid");
1547                 } else if (ret == td->pid) {
1548                         if (WIFSIGNALED(status)) {
1549                                 int sig = WTERMSIG(status);
1550
1551                                 if (sig != SIGTERM && sig != SIGUSR2)
1552                                         log_err("fio: pid=%d, got signal=%d\n",
1553                                                         (int) td->pid, sig);
1554                                 td->sig = sig;
1555                                 td_set_runstate(td, TD_REAPED);
1556                                 goto reaped;
1557                         }
1558                         if (WIFEXITED(status)) {
1559                                 if (WEXITSTATUS(status) && !td->error)
1560                                         td->error = WEXITSTATUS(status);
1561
1562                                 td_set_runstate(td, TD_REAPED);
1563                                 goto reaped;
1564                         }
1565                 }
1566
1567                 /*
1568                  * thread is not dead, continue
1569                  */
1570                 pending++;
1571                 continue;
1572 reaped:
1573                 (*nr_running)--;
1574                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1575                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1576                 if (!td->pid)
1577                         pending--;
1578
1579                 if (td->error)
1580                         exit_value++;
1581
1582                 done_secs += mtime_since_now(&td->epoch) / 1000;
1583                 profile_td_exit(td);
1584         }
1585
1586         if (*nr_running == cputhreads && !pending && realthreads)
1587                 fio_terminate_threads(TERMINATE_ALL);
1588 }
1589
1590 static void do_usleep(unsigned int usecs)
1591 {
1592         check_for_running_stats();
1593         usleep(usecs);
1594 }
1595
1596 /*
1597  * Main function for kicking off and reaping jobs, as needed.
1598  */
1599 static void run_threads(void)
1600 {
1601         struct thread_data *td;
1602         unsigned long spent;
1603         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1604
1605         if (fio_gtod_offload && fio_start_gtod_thread())
1606                 return;
1607
1608         fio_idle_prof_init();
1609
1610         set_sig_handlers();
1611
1612         nr_thread = nr_process = 0;
1613         for_each_td(td, i) {
1614                 if (td->o.use_thread)
1615                         nr_thread++;
1616                 else
1617                         nr_process++;
1618         }
1619
1620         if (output_format == FIO_OUTPUT_NORMAL) {
1621                 log_info("Starting ");
1622                 if (nr_thread)
1623                         log_info("%d thread%s", nr_thread,
1624                                                 nr_thread > 1 ? "s" : "");
1625                 if (nr_process) {
1626                         if (nr_thread)
1627                                 log_info(" and ");
1628                         log_info("%d process%s", nr_process,
1629                                                 nr_process > 1 ? "es" : "");
1630                 }
1631                 log_info("\n");
1632                 fflush(stdout);
1633         }
1634
1635         todo = thread_number;
1636         nr_running = 0;
1637         nr_started = 0;
1638         m_rate = t_rate = 0;
1639
1640         for_each_td(td, i) {
1641                 print_status_init(td->thread_number - 1);
1642
1643                 if (!td->o.create_serialize)
1644                         continue;
1645
1646                 /*
1647                  * do file setup here so it happens sequentially,
1648                  * we don't want X number of threads getting their
1649                  * client data interspersed on disk
1650                  */
1651                 if (setup_files(td)) {
1652                         exit_value++;
1653                         if (td->error)
1654                                 log_err("fio: pid=%d, err=%d/%s\n",
1655                                         (int) td->pid, td->error, td->verror);
1656                         td_set_runstate(td, TD_REAPED);
1657                         todo--;
1658                 } else {
1659                         struct fio_file *f;
1660                         unsigned int j;
1661
1662                         /*
1663                          * for sharing to work, each job must always open
1664                          * its own files. so close them, if we opened them
1665                          * for creation
1666                          */
1667                         for_each_file(td, f, j) {
1668                                 if (fio_file_open(f))
1669                                         td_io_close_file(td, f);
1670                         }
1671                 }
1672         }
1673
1674         /* start idle threads before io threads start to run */
1675         fio_idle_prof_start();
1676
1677         set_genesis_time();
1678
1679         while (todo) {
1680                 struct thread_data *map[REAL_MAX_JOBS];
1681                 struct timeval this_start;
1682                 int this_jobs = 0, left;
1683
1684                 /*
1685                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1686                  */
1687                 for_each_td(td, i) {
1688                         if (td->runstate != TD_NOT_CREATED)
1689                                 continue;
1690
1691                         /*
1692                          * never got a chance to start, killed by other
1693                          * thread for some reason
1694                          */
1695                         if (td->terminate) {
1696                                 todo--;
1697                                 continue;
1698                         }
1699
1700                         if (td->o.start_delay) {
1701                                 spent = mtime_since_genesis();
1702
1703                                 if (td->o.start_delay * 1000 > spent)
1704                                         continue;
1705                         }
1706
1707                         if (td->o.stonewall && (nr_started || nr_running)) {
1708                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1709                                                         td->o.name);
1710                                 break;
1711                         }
1712
1713                         init_disk_util(td);
1714
1715                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1716                         td->update_rusage = 0;
1717
1718                         /*
1719                          * Set state to created. Thread will transition
1720                          * to TD_INITIALIZED when it's done setting up.
1721                          */
1722                         td_set_runstate(td, TD_CREATED);
1723                         map[this_jobs++] = td;
1724                         nr_started++;
1725
1726                         if (td->o.use_thread) {
1727                                 int ret;
1728
1729                                 dprint(FD_PROCESS, "will pthread_create\n");
1730                                 ret = pthread_create(&td->thread, NULL,
1731                                                         thread_main, td);
1732                                 if (ret) {
1733                                         log_err("pthread_create: %s\n",
1734                                                         strerror(ret));
1735                                         nr_started--;
1736                                         break;
1737                                 }
1738                                 ret = pthread_detach(td->thread);
1739                                 if (ret)
1740                                         log_err("pthread_detach: %s",
1741                                                         strerror(ret));
1742                         } else {
1743                                 pid_t pid;
1744                                 dprint(FD_PROCESS, "will fork\n");
1745                                 pid = fork();
1746                                 if (!pid) {
1747                                         int ret = fork_main(shm_id, i);
1748
1749                                         _exit(ret);
1750                                 } else if (i == fio_debug_jobno)
1751                                         *fio_debug_jobp = pid;
1752                         }
1753                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1754                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1755                                 log_err("fio: job startup hung? exiting.\n");
1756                                 fio_terminate_threads(TERMINATE_ALL);
1757                                 fio_abort = 1;
1758                                 nr_started--;
1759                                 break;
1760                         }
1761                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1762                 }
1763
1764                 /*
1765                  * Wait for the started threads to transition to
1766                  * TD_INITIALIZED.
1767                  */
1768                 fio_gettime(&this_start, NULL);
1769                 left = this_jobs;
1770                 while (left && !fio_abort) {
1771                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1772                                 break;
1773
1774                         do_usleep(100000);
1775
1776                         for (i = 0; i < this_jobs; i++) {
1777                                 td = map[i];
1778                                 if (!td)
1779                                         continue;
1780                                 if (td->runstate == TD_INITIALIZED) {
1781                                         map[i] = NULL;
1782                                         left--;
1783                                 } else if (td->runstate >= TD_EXITED) {
1784                                         map[i] = NULL;
1785                                         left--;
1786                                         todo--;
1787                                         nr_running++; /* work-around... */
1788                                 }
1789                         }
1790                 }
1791
1792                 if (left) {
1793                         log_err("fio: %d job%s failed to start\n", left,
1794                                         left > 1 ? "s" : "");
1795                         for (i = 0; i < this_jobs; i++) {
1796                                 td = map[i];
1797                                 if (!td)
1798                                         continue;
1799                                 kill(td->pid, SIGTERM);
1800                         }
1801                         break;
1802                 }
1803
1804                 /*
1805                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1806                  */
1807                 for_each_td(td, i) {
1808                         if (td->runstate != TD_INITIALIZED)
1809                                 continue;
1810
1811                         if (in_ramp_time(td))
1812                                 td_set_runstate(td, TD_RAMP);
1813                         else
1814                                 td_set_runstate(td, TD_RUNNING);
1815                         nr_running++;
1816                         nr_started--;
1817                         m_rate += ddir_rw_sum(td->o.ratemin);
1818                         t_rate += ddir_rw_sum(td->o.rate);
1819                         todo--;
1820                         fio_mutex_up(td->mutex);
1821                 }
1822
1823                 reap_threads(&nr_running, &t_rate, &m_rate);
1824
1825                 if (todo)
1826                         do_usleep(100000);
1827         }
1828
1829         while (nr_running) {
1830                 reap_threads(&nr_running, &t_rate, &m_rate);
1831                 do_usleep(10000);
1832         }
1833
1834         fio_idle_prof_stop();
1835
1836         update_io_ticks();
1837 }
1838
1839 void wait_for_disk_thread_exit(void)
1840 {
1841         fio_mutex_down(disk_thread_mutex);
1842 }
1843
1844 static void free_disk_util(void)
1845 {
1846         disk_util_start_exit();
1847         wait_for_disk_thread_exit();
1848         disk_util_prune_entries();
1849 }
1850
1851 static void *disk_thread_main(void *data)
1852 {
1853         int ret = 0;
1854
1855         fio_mutex_up(startup_mutex);
1856
1857         while (threads && !ret) {
1858                 usleep(DISK_UTIL_MSEC * 1000);
1859                 if (!threads)
1860                         break;
1861                 ret = update_io_ticks();
1862
1863                 if (!is_backend)
1864                         print_thread_status();
1865         }
1866
1867         fio_mutex_up(disk_thread_mutex);
1868         return NULL;
1869 }
1870
1871 static int create_disk_util_thread(void)
1872 {
1873         int ret;
1874
1875         setup_disk_util();
1876
1877         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1878
1879         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1880         if (ret) {
1881                 fio_mutex_remove(disk_thread_mutex);
1882                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1883                 return 1;
1884         }
1885
1886         ret = pthread_detach(disk_util_thread);
1887         if (ret) {
1888                 fio_mutex_remove(disk_thread_mutex);
1889                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1890                 return 1;
1891         }
1892
1893         dprint(FD_MUTEX, "wait on startup_mutex\n");
1894         fio_mutex_down(startup_mutex);
1895         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1896         return 0;
1897 }
1898
1899 int fio_backend(void)
1900 {
1901         struct thread_data *td;
1902         int i;
1903
1904         if (exec_profile) {
1905                 if (load_profile(exec_profile))
1906                         return 1;
1907                 free(exec_profile);
1908                 exec_profile = NULL;
1909         }
1910         if (!thread_number)
1911                 return 0;
1912
1913         if (write_bw_log) {
1914                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1915                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1916                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1917         }
1918
1919         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1920         if (startup_mutex == NULL)
1921                 return 1;
1922         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1923         if (writeout_mutex == NULL)
1924                 return 1;
1925
1926         set_genesis_time();
1927         stat_init();
1928         create_disk_util_thread();
1929
1930         cgroup_list = smalloc(sizeof(*cgroup_list));
1931         INIT_FLIST_HEAD(cgroup_list);
1932
1933         run_threads();
1934
1935         if (!fio_abort) {
1936                 show_run_stats();
1937                 if (write_bw_log) {
1938                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1939                         __finish_log(agg_io_log[DDIR_WRITE],
1940                                         "agg-write_bw.log");
1941                         __finish_log(agg_io_log[DDIR_TRIM],
1942                                         "agg-write_bw.log");
1943                 }
1944         }
1945
1946         for_each_td(td, i)
1947                 fio_options_free(td);
1948
1949         free_disk_util();
1950         cgroup_kill(cgroup_list);
1951         sfree(cgroup_list);
1952         sfree(cgroup_mnt);
1953
1954         fio_mutex_remove(startup_mutex);
1955         fio_mutex_remove(writeout_mutex);
1956         fio_mutex_remove(disk_thread_mutex);
1957         stat_exit();
1958         return exit_value;
1959 }