Replace list based free/busy/requeue list with FIFO + ring
[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
646         if (in_ramp_time(td))
647                 td_set_runstate(td, TD_RAMP);
648         else
649                 td_set_runstate(td, TD_RUNNING);
650
651         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
652                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
653                 td->o.time_based) {
654                 struct timeval comp_time;
655                 int min_evts = 0;
656                 struct io_u *io_u;
657                 int ret2, full;
658                 enum fio_ddir ddir;
659
660                 check_update_rusage(td);
661
662                 if (td->terminate || td->done)
663                         break;
664
665                 update_tv_cache(td);
666
667                 if (runtime_exceeded(td, &td->tv_cache)) {
668                         __update_tv_cache(td);
669                         if (runtime_exceeded(td, &td->tv_cache)) {
670                                 td->terminate = 1;
671                                 break;
672                         }
673                 }
674
675                 if (flow_threshold_exceeded(td))
676                         continue;
677
678                 io_u = get_io_u(td);
679                 if (!io_u)
680                         break;
681
682                 ddir = io_u->ddir;
683
684                 /*
685                  * Add verification end_io handler if:
686                  *      - Asked to verify (!td_rw(td))
687                  *      - Or the io_u is from our verify list (mixed write/ver)
688                  */
689                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
690                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
691                         if (td->o.verify_async)
692                                 io_u->end_io = verify_io_u_async;
693                         else
694                                 io_u->end_io = verify_io_u;
695                         td_set_runstate(td, TD_VERIFYING);
696                 } else if (in_ramp_time(td))
697                         td_set_runstate(td, TD_RAMP);
698                 else
699                         td_set_runstate(td, TD_RUNNING);
700
701                 ret = td_io_queue(td, io_u);
702                 switch (ret) {
703                 case FIO_Q_COMPLETED:
704                         if (io_u->error) {
705                                 ret = -io_u->error;
706                                 clear_io_u(td, io_u);
707                         } else if (io_u->resid) {
708                                 int bytes = io_u->xfer_buflen - io_u->resid;
709                                 struct fio_file *f = io_u->file;
710
711                                 /*
712                                  * zero read, fail
713                                  */
714                                 if (!bytes) {
715                                         td_verror(td, EIO, "full resid");
716                                         put_io_u(td, io_u);
717                                         break;
718                                 }
719
720                                 io_u->xfer_buflen = io_u->resid;
721                                 io_u->xfer_buf += bytes;
722                                 io_u->offset += bytes;
723
724                                 if (ddir_rw(io_u->ddir))
725                                         td->ts.short_io_u[io_u->ddir]++;
726
727                                 if (io_u->offset == f->real_file_size)
728                                         goto sync_done;
729
730                                 requeue_io_u(td, &io_u);
731                         } else {
732 sync_done:
733                                 if (__should_check_rate(td, DDIR_READ) ||
734                                     __should_check_rate(td, DDIR_WRITE) ||
735                                     __should_check_rate(td, DDIR_TRIM))
736                                         fio_gettime(&comp_time, NULL);
737
738                                 ret = io_u_sync_complete(td, io_u, bytes_done);
739                                 if (ret < 0)
740                                         break;
741                         }
742                         break;
743                 case FIO_Q_QUEUED:
744                         /*
745                          * if the engine doesn't have a commit hook,
746                          * the io_u is really queued. if it does have such
747                          * a hook, it has to call io_u_queued() itself.
748                          */
749                         if (td->io_ops->commit == NULL)
750                                 io_u_queued(td, io_u);
751                         break;
752                 case FIO_Q_BUSY:
753                         requeue_io_u(td, &io_u);
754                         ret2 = td_io_commit(td);
755                         if (ret2 < 0)
756                                 ret = ret2;
757                         break;
758                 default:
759                         assert(ret < 0);
760                         put_io_u(td, io_u);
761                         break;
762                 }
763
764                 if (break_on_this_error(td, ddir, &ret))
765                         break;
766
767                 /*
768                  * See if we need to complete some commands. Note that we
769                  * can get BUSY even without IO queued, if the system is
770                  * resource starved.
771                  */
772                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
773                 if (full || !td->o.iodepth_batch_complete) {
774                         min_evts = min(td->o.iodepth_batch_complete,
775                                         td->cur_depth);
776                         /*
777                          * if the queue is full, we MUST reap at least 1 event
778                          */
779                         if (full && !min_evts)
780                                 min_evts = 1;
781
782                         if (__should_check_rate(td, DDIR_READ) ||
783                             __should_check_rate(td, DDIR_WRITE) ||
784                             __should_check_rate(td, DDIR_TRIM))
785                                 fio_gettime(&comp_time, NULL);
786
787                         do {
788                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
789                                 if (ret < 0)
790                                         break;
791
792                         } while (full && (td->cur_depth > td->o.iodepth_low));
793                 }
794
795                 if (ret < 0)
796                         break;
797                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
798                         continue;
799
800                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
801                         if (check_min_rate(td, &comp_time, bytes_done)) {
802                                 if (exitall_on_terminate)
803                                         fio_terminate_threads(td->groupid);
804                                 td_verror(td, EIO, "check_min_rate");
805                                 break;
806                         }
807                 }
808
809                 if (td->o.thinktime) {
810                         unsigned long long b;
811
812                         b = ddir_rw_sum(td->io_blocks);
813                         if (!(b % td->o.thinktime_blocks)) {
814                                 int left;
815
816                                 io_u_quiesce(td);
817
818                                 if (td->o.thinktime_spin)
819                                         usec_spin(td->o.thinktime_spin);
820
821                                 left = td->o.thinktime - td->o.thinktime_spin;
822                                 if (left)
823                                         usec_sleep(td, left);
824                         }
825                 }
826         }
827
828         check_update_rusage(td);
829
830         if (td->trim_entries)
831                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
832
833         if (td->o.fill_device && td->error == ENOSPC) {
834                 td->error = 0;
835                 td->terminate = 1;
836         }
837         if (!td->error) {
838                 struct fio_file *f;
839
840                 i = td->cur_depth;
841                 if (i) {
842                         ret = io_u_queued_complete(td, i, bytes_done);
843                         if (td->o.fill_device && td->error == ENOSPC)
844                                 td->error = 0;
845                 }
846
847                 if (should_fsync(td) && td->o.end_fsync) {
848                         td_set_runstate(td, TD_FSYNCING);
849
850                         for_each_file(td, f, i) {
851                                 if (!fio_file_fsync(td, f))
852                                         continue;
853
854                                 log_err("fio: end_fsync failed for file %s\n",
855                                                                 f->file_name);
856                         }
857                 }
858         } else
859                 cleanup_pending_aio(td);
860
861         /*
862          * stop job if we failed doing any IO
863          */
864         if (!ddir_rw_sum(td->this_io_bytes))
865                 td->done = 1;
866
867         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
868 }
869
870 static void cleanup_io_u(struct thread_data *td)
871 {
872         struct io_u *io_u;
873
874         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
875
876                 if (td->io_ops->io_u_free)
877                         td->io_ops->io_u_free(td, io_u);
878
879                 fio_memfree(io_u, sizeof(*io_u));
880         }
881
882         free_io_mem(td);
883
884         io_u_rexit(&td->io_u_requeues);
885         io_u_qexit(&td->io_u_freelist);
886         io_u_qexit(&td->io_u_all);
887 }
888
889 static int init_io_u(struct thread_data *td)
890 {
891         struct io_u *io_u;
892         unsigned int max_bs, min_write;
893         int cl_align, i, max_units;
894         int data_xfer = 1, err;
895         char *p;
896
897         max_units = td->o.iodepth;
898         max_bs = td_max_bs(td);
899         min_write = td->o.min_bs[DDIR_WRITE];
900         td->orig_buffer_size = (unsigned long long) max_bs
901                                         * (unsigned long long) max_units;
902
903         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
904                 data_xfer = 0;
905
906         err = 0;
907         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
908         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
909         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
910
911         if (err) {
912                 log_err("fio: failed setting up IO queues\n");
913                 return 1;
914         }
915
916         /*
917          * if we may later need to do address alignment, then add any
918          * possible adjustment here so that we don't cause a buffer
919          * overflow later. this adjustment may be too much if we get
920          * lucky and the allocator gives us an aligned address.
921          */
922         if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO))
923                 td->orig_buffer_size += page_mask + td->o.mem_align;
924
925         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
926                 unsigned long bs;
927
928                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
929                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
930         }
931
932         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
933                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
934                 return 1;
935         }
936
937         if (data_xfer && allocate_io_mem(td))
938                 return 1;
939
940         if (td->o.odirect || td->o.mem_align ||
941             (td->io_ops->flags & FIO_RAWIO))
942                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
943         else
944                 p = td->orig_buffer;
945
946         cl_align = os_cache_line_size();
947
948         for (i = 0; i < max_units; i++) {
949                 void *ptr;
950
951                 if (td->terminate)
952                         return 1;
953
954                 ptr = fio_memalign(cl_align, sizeof(*io_u));
955                 if (!ptr) {
956                         log_err("fio: unable to allocate aligned memory\n");
957                         break;
958                 }
959
960                 io_u = ptr;
961                 memset(io_u, 0, sizeof(*io_u));
962                 INIT_FLIST_HEAD(&io_u->verify_list);
963                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
964
965                 if (data_xfer) {
966                         io_u->buf = p;
967                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
968
969                         if (td_write(td))
970                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
971                         if (td_write(td) && td->o.verify_pattern_bytes) {
972                                 /*
973                                  * Fill the buffer with the pattern if we are
974                                  * going to be doing writes.
975                                  */
976                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
977                         }
978                 }
979
980                 io_u->index = i;
981                 io_u->flags = IO_U_F_FREE;
982                 io_u_qpush(&td->io_u_freelist, io_u);
983
984                 /*
985                  * io_u never leaves this stack, used for iteration of all
986                  * io_u buffers.
987                  */
988                 io_u_qpush(&td->io_u_all, io_u);
989
990                 if (td->io_ops->io_u_init) {
991                         int ret = td->io_ops->io_u_init(td, io_u);
992
993                         if (ret) {
994                                 log_err("fio: failed to init engine data: %d\n", ret);
995                                 return 1;
996                         }
997                 }
998
999                 p += max_bs;
1000         }
1001
1002         return 0;
1003 }
1004
1005 static int switch_ioscheduler(struct thread_data *td)
1006 {
1007         char tmp[256], tmp2[128];
1008         FILE *f;
1009         int ret;
1010
1011         if (td->io_ops->flags & FIO_DISKLESSIO)
1012                 return 0;
1013
1014         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1015
1016         f = fopen(tmp, "r+");
1017         if (!f) {
1018                 if (errno == ENOENT) {
1019                         log_err("fio: os or kernel doesn't support IO scheduler"
1020                                 " switching\n");
1021                         return 0;
1022                 }
1023                 td_verror(td, errno, "fopen iosched");
1024                 return 1;
1025         }
1026
1027         /*
1028          * Set io scheduler.
1029          */
1030         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1031         if (ferror(f) || ret != 1) {
1032                 td_verror(td, errno, "fwrite");
1033                 fclose(f);
1034                 return 1;
1035         }
1036
1037         rewind(f);
1038
1039         /*
1040          * Read back and check that the selected scheduler is now the default.
1041          */
1042         ret = fread(tmp, 1, sizeof(tmp), f);
1043         if (ferror(f) || ret < 0) {
1044                 td_verror(td, errno, "fread");
1045                 fclose(f);
1046                 return 1;
1047         }
1048
1049         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1050         if (!strstr(tmp, tmp2)) {
1051                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1052                 td_verror(td, EINVAL, "iosched_switch");
1053                 fclose(f);
1054                 return 1;
1055         }
1056
1057         fclose(f);
1058         return 0;
1059 }
1060
1061 static int keep_running(struct thread_data *td)
1062 {
1063         if (td->done)
1064                 return 0;
1065         if (td->o.time_based)
1066                 return 1;
1067         if (td->o.loops) {
1068                 td->o.loops--;
1069                 return 1;
1070         }
1071
1072         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1073                 uint64_t diff;
1074
1075                 /*
1076                  * If the difference is less than the minimum IO size, we
1077                  * are done.
1078                  */
1079                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1080                 if (diff < td_max_bs(td))
1081                         return 0;
1082
1083                 return 1;
1084         }
1085
1086         return 0;
1087 }
1088
1089 static int exec_string(const char *string)
1090 {
1091         int ret, newlen = strlen(string) + 1 + 8;
1092         char *str;
1093
1094         str = malloc(newlen);
1095         sprintf(str, "sh -c %s", string);
1096
1097         ret = system(str);
1098         if (ret == -1)
1099                 log_err("fio: exec of cmd <%s> failed\n", str);
1100
1101         free(str);
1102         return ret;
1103 }
1104
1105 /*
1106  * Entry point for the thread based jobs. The process based jobs end up
1107  * here as well, after a little setup.
1108  */
1109 static void *thread_main(void *data)
1110 {
1111         unsigned long long elapsed;
1112         struct thread_data *td = data;
1113         struct thread_options *o = &td->o;
1114         pthread_condattr_t attr;
1115         int clear_state;
1116         int ret;
1117
1118         if (!o->use_thread) {
1119                 setsid();
1120                 td->pid = getpid();
1121         } else
1122                 td->pid = gettid();
1123
1124         fio_local_clock_init(o->use_thread);
1125
1126         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1127
1128         if (is_backend)
1129                 fio_server_send_start(td);
1130
1131         INIT_FLIST_HEAD(&td->io_log_list);
1132         INIT_FLIST_HEAD(&td->io_hist_list);
1133         INIT_FLIST_HEAD(&td->verify_list);
1134         INIT_FLIST_HEAD(&td->trim_list);
1135         INIT_FLIST_HEAD(&td->next_rand_list);
1136         pthread_mutex_init(&td->io_u_lock, NULL);
1137         td->io_hist_tree = RB_ROOT;
1138
1139         pthread_condattr_init(&attr);
1140         pthread_cond_init(&td->verify_cond, &attr);
1141         pthread_cond_init(&td->free_cond, &attr);
1142
1143         td_set_runstate(td, TD_INITIALIZED);
1144         dprint(FD_MUTEX, "up startup_mutex\n");
1145         fio_mutex_up(startup_mutex);
1146         dprint(FD_MUTEX, "wait on td->mutex\n");
1147         fio_mutex_down(td->mutex);
1148         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1149
1150         /*
1151          * the ->mutex mutex is now no longer used, close it to avoid
1152          * eating a file descriptor
1153          */
1154         fio_mutex_remove(td->mutex);
1155         td->mutex = NULL;
1156
1157         /*
1158          * A new gid requires privilege, so we need to do this before setting
1159          * the uid.
1160          */
1161         if (o->gid != -1U && setgid(o->gid)) {
1162                 td_verror(td, errno, "setgid");
1163                 goto err;
1164         }
1165         if (o->uid != -1U && setuid(o->uid)) {
1166                 td_verror(td, errno, "setuid");
1167                 goto err;
1168         }
1169
1170         /*
1171          * If we have a gettimeofday() thread, make sure we exclude that
1172          * thread from this job
1173          */
1174         if (o->gtod_cpu)
1175                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1176
1177         /*
1178          * Set affinity first, in case it has an impact on the memory
1179          * allocations.
1180          */
1181         if (o->cpumask_set) {
1182                 ret = fio_setaffinity(td->pid, o->cpumask);
1183                 if (ret == -1) {
1184                         td_verror(td, errno, "cpu_set_affinity");
1185                         goto err;
1186                 }
1187         }
1188
1189 #ifdef CONFIG_LIBNUMA
1190         /* numa node setup */
1191         if (o->numa_cpumask_set || o->numa_memmask_set) {
1192                 int ret;
1193
1194                 if (numa_available() < 0) {
1195                         td_verror(td, errno, "Does not support NUMA API\n");
1196                         goto err;
1197                 }
1198
1199                 if (o->numa_cpumask_set) {
1200                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1201                         if (ret == -1) {
1202                                 td_verror(td, errno, \
1203                                         "numa_run_on_node_mask failed\n");
1204                                 goto err;
1205                         }
1206                 }
1207
1208                 if (o->numa_memmask_set) {
1209
1210                         switch (o->numa_mem_mode) {
1211                         case MPOL_INTERLEAVE:
1212                                 numa_set_interleave_mask(o->numa_memnodesmask);
1213                                 break;
1214                         case MPOL_BIND:
1215                                 numa_set_membind(o->numa_memnodesmask);
1216                                 break;
1217                         case MPOL_LOCAL:
1218                                 numa_set_localalloc();
1219                                 break;
1220                         case MPOL_PREFERRED:
1221                                 numa_set_preferred(o->numa_mem_prefer_node);
1222                                 break;
1223                         case MPOL_DEFAULT:
1224                         default:
1225                                 break;
1226                         }
1227
1228                 }
1229         }
1230 #endif
1231
1232         if (fio_pin_memory(td))
1233                 goto err;
1234
1235         /*
1236          * May alter parameters that init_io_u() will use, so we need to
1237          * do this first.
1238          */
1239         if (init_iolog(td))
1240                 goto err;
1241
1242         if (init_io_u(td))
1243                 goto err;
1244
1245         if (o->verify_async && verify_async_init(td))
1246                 goto err;
1247
1248         if (o->ioprio) {
1249                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1250                 if (ret == -1) {
1251                         td_verror(td, errno, "ioprio_set");
1252                         goto err;
1253                 }
1254         }
1255
1256         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1257                 goto err;
1258
1259         errno = 0;
1260         if (nice(o->nice) == -1 && errno != 0) {
1261                 td_verror(td, errno, "nice");
1262                 goto err;
1263         }
1264
1265         if (o->ioscheduler && switch_ioscheduler(td))
1266                 goto err;
1267
1268         if (!o->create_serialize && setup_files(td))
1269                 goto err;
1270
1271         if (td_io_init(td))
1272                 goto err;
1273
1274         if (init_random_map(td))
1275                 goto err;
1276
1277         if (o->exec_prerun && exec_string(o->exec_prerun))
1278                 goto err;
1279
1280         if (o->pre_read) {
1281                 if (pre_read_files(td) < 0)
1282                         goto err;
1283         }
1284
1285         fio_verify_init(td);
1286
1287         fio_gettime(&td->epoch, NULL);
1288         fio_getrusage(&td->ru_start);
1289         clear_state = 0;
1290         while (keep_running(td)) {
1291                 uint64_t verify_bytes;
1292
1293                 fio_gettime(&td->start, NULL);
1294                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1295                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1296                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1297
1298                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1299                                 o->ratemin[DDIR_TRIM]) {
1300                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1301                                                 sizeof(td->bw_sample_time));
1302                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1303                                                 sizeof(td->bw_sample_time));
1304                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1305                                                 sizeof(td->bw_sample_time));
1306                 }
1307
1308                 if (clear_state)
1309                         clear_io_state(td);
1310
1311                 prune_io_piece_log(td);
1312
1313                 verify_bytes = do_io(td);
1314
1315                 clear_state = 1;
1316
1317                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1318                         elapsed = utime_since_now(&td->start);
1319                         td->ts.runtime[DDIR_READ] += elapsed;
1320                 }
1321                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1322                         elapsed = utime_since_now(&td->start);
1323                         td->ts.runtime[DDIR_WRITE] += elapsed;
1324                 }
1325                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1326                         elapsed = utime_since_now(&td->start);
1327                         td->ts.runtime[DDIR_TRIM] += elapsed;
1328                 }
1329
1330                 if (td->error || td->terminate)
1331                         break;
1332
1333                 if (!o->do_verify ||
1334                     o->verify == VERIFY_NONE ||
1335                     (td->io_ops->flags & FIO_UNIDIR))
1336                         continue;
1337
1338                 clear_io_state(td);
1339
1340                 fio_gettime(&td->start, NULL);
1341
1342                 do_verify(td, verify_bytes);
1343
1344                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1345
1346                 if (td->error || td->terminate)
1347                         break;
1348         }
1349
1350         update_rusage_stat(td);
1351         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1352         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1353         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1354         td->ts.total_run_time = mtime_since_now(&td->epoch);
1355         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1356         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1357         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1358
1359         fio_unpin_memory(td);
1360
1361         fio_mutex_down(writeout_mutex);
1362         if (td->bw_log) {
1363                 if (o->bw_log_file) {
1364                         finish_log_named(td, td->bw_log,
1365                                                 o->bw_log_file, "bw");
1366                 } else
1367                         finish_log(td, td->bw_log, "bw");
1368         }
1369         if (td->lat_log) {
1370                 if (o->lat_log_file) {
1371                         finish_log_named(td, td->lat_log,
1372                                                 o->lat_log_file, "lat");
1373                 } else
1374                         finish_log(td, td->lat_log, "lat");
1375         }
1376         if (td->slat_log) {
1377                 if (o->lat_log_file) {
1378                         finish_log_named(td, td->slat_log,
1379                                                 o->lat_log_file, "slat");
1380                 } else
1381                         finish_log(td, td->slat_log, "slat");
1382         }
1383         if (td->clat_log) {
1384                 if (o->lat_log_file) {
1385                         finish_log_named(td, td->clat_log,
1386                                                 o->lat_log_file, "clat");
1387                 } else
1388                         finish_log(td, td->clat_log, "clat");
1389         }
1390         if (td->iops_log) {
1391                 if (o->iops_log_file) {
1392                         finish_log_named(td, td->iops_log,
1393                                                 o->iops_log_file, "iops");
1394                 } else
1395                         finish_log(td, td->iops_log, "iops");
1396         }
1397
1398         fio_mutex_up(writeout_mutex);
1399         if (o->exec_postrun)
1400                 exec_string(o->exec_postrun);
1401
1402         if (exitall_on_terminate)
1403                 fio_terminate_threads(td->groupid);
1404
1405 err:
1406         if (td->error)
1407                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1408                                                         td->verror);
1409
1410         if (o->verify_async)
1411                 verify_async_exit(td);
1412
1413         close_and_free_files(td);
1414         cleanup_io_u(td);
1415         close_ioengine(td);
1416         cgroup_shutdown(td, &cgroup_mnt);
1417
1418         if (o->cpumask_set) {
1419                 int ret = fio_cpuset_exit(&o->cpumask);
1420
1421                 td_verror(td, ret, "fio_cpuset_exit");
1422         }
1423
1424         /*
1425          * do this very late, it will log file closing as well
1426          */
1427         if (o->write_iolog_file)
1428                 write_iolog_close(td);
1429
1430         fio_mutex_remove(td->rusage_sem);
1431         td->rusage_sem = NULL;
1432
1433         td_set_runstate(td, TD_EXITED);
1434         return (void *) (uintptr_t) td->error;
1435 }
1436
1437
1438 /*
1439  * We cannot pass the td data into a forked process, so attach the td and
1440  * pass it to the thread worker.
1441  */
1442 static int fork_main(int shmid, int offset)
1443 {
1444         struct thread_data *td;
1445         void *data, *ret;
1446
1447 #ifndef __hpux
1448         data = shmat(shmid, NULL, 0);
1449         if (data == (void *) -1) {
1450                 int __err = errno;
1451
1452                 perror("shmat");
1453                 return __err;
1454         }
1455 #else
1456         /*
1457          * HP-UX inherits shm mappings?
1458          */
1459         data = threads;
1460 #endif
1461
1462         td = data + offset * sizeof(struct thread_data);
1463         ret = thread_main(td);
1464         shmdt(data);
1465         return (int) (uintptr_t) ret;
1466 }
1467
1468 /*
1469  * Run over the job map and reap the threads that have exited, if any.
1470  */
1471 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1472                          unsigned int *m_rate)
1473 {
1474         struct thread_data *td;
1475         unsigned int cputhreads, realthreads, pending;
1476         int i, status, ret;
1477
1478         /*
1479          * reap exited threads (TD_EXITED -> TD_REAPED)
1480          */
1481         realthreads = pending = cputhreads = 0;
1482         for_each_td(td, i) {
1483                 int flags = 0;
1484
1485                 /*
1486                  * ->io_ops is NULL for a thread that has closed its
1487                  * io engine
1488                  */
1489                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1490                         cputhreads++;
1491                 else
1492                         realthreads++;
1493
1494                 if (!td->pid) {
1495                         pending++;
1496                         continue;
1497                 }
1498                 if (td->runstate == TD_REAPED)
1499                         continue;
1500                 if (td->o.use_thread) {
1501                         if (td->runstate == TD_EXITED) {
1502                                 td_set_runstate(td, TD_REAPED);
1503                                 goto reaped;
1504                         }
1505                         continue;
1506                 }
1507
1508                 flags = WNOHANG;
1509                 if (td->runstate == TD_EXITED)
1510                         flags = 0;
1511
1512                 /*
1513                  * check if someone quit or got killed in an unusual way
1514                  */
1515                 ret = waitpid(td->pid, &status, flags);
1516                 if (ret < 0) {
1517                         if (errno == ECHILD) {
1518                                 log_err("fio: pid=%d disappeared %d\n",
1519                                                 (int) td->pid, td->runstate);
1520                                 td->sig = ECHILD;
1521                                 td_set_runstate(td, TD_REAPED);
1522                                 goto reaped;
1523                         }
1524                         perror("waitpid");
1525                 } else if (ret == td->pid) {
1526                         if (WIFSIGNALED(status)) {
1527                                 int sig = WTERMSIG(status);
1528
1529                                 if (sig != SIGTERM && sig != SIGUSR2)
1530                                         log_err("fio: pid=%d, got signal=%d\n",
1531                                                         (int) td->pid, sig);
1532                                 td->sig = sig;
1533                                 td_set_runstate(td, TD_REAPED);
1534                                 goto reaped;
1535                         }
1536                         if (WIFEXITED(status)) {
1537                                 if (WEXITSTATUS(status) && !td->error)
1538                                         td->error = WEXITSTATUS(status);
1539
1540                                 td_set_runstate(td, TD_REAPED);
1541                                 goto reaped;
1542                         }
1543                 }
1544
1545                 /*
1546                  * thread is not dead, continue
1547                  */
1548                 pending++;
1549                 continue;
1550 reaped:
1551                 (*nr_running)--;
1552                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1553                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1554                 if (!td->pid)
1555                         pending--;
1556
1557                 if (td->error)
1558                         exit_value++;
1559
1560                 done_secs += mtime_since_now(&td->epoch) / 1000;
1561                 profile_td_exit(td);
1562         }
1563
1564         if (*nr_running == cputhreads && !pending && realthreads)
1565                 fio_terminate_threads(TERMINATE_ALL);
1566 }
1567
1568 static void do_usleep(unsigned int usecs)
1569 {
1570         check_for_running_stats();
1571         usleep(usecs);
1572 }
1573
1574 /*
1575  * Main function for kicking off and reaping jobs, as needed.
1576  */
1577 static void run_threads(void)
1578 {
1579         struct thread_data *td;
1580         unsigned long spent;
1581         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1582
1583         if (fio_gtod_offload && fio_start_gtod_thread())
1584                 return;
1585         
1586         fio_idle_prof_init();
1587
1588         set_sig_handlers();
1589
1590         nr_thread = nr_process = 0;
1591         for_each_td(td, i) {
1592                 if (td->o.use_thread)
1593                         nr_thread++;
1594                 else
1595                         nr_process++;
1596         }
1597
1598         if (output_format == FIO_OUTPUT_NORMAL) {
1599                 log_info("Starting ");
1600                 if (nr_thread)
1601                         log_info("%d thread%s", nr_thread,
1602                                                 nr_thread > 1 ? "s" : "");
1603                 if (nr_process) {
1604                         if (nr_thread)
1605                                 log_info(" and ");
1606                         log_info("%d process%s", nr_process,
1607                                                 nr_process > 1 ? "es" : "");
1608                 }
1609                 log_info("\n");
1610                 fflush(stdout);
1611         }
1612
1613         todo = thread_number;
1614         nr_running = 0;
1615         nr_started = 0;
1616         m_rate = t_rate = 0;
1617
1618         for_each_td(td, i) {
1619                 print_status_init(td->thread_number - 1);
1620
1621                 if (!td->o.create_serialize)
1622                         continue;
1623
1624                 /*
1625                  * do file setup here so it happens sequentially,
1626                  * we don't want X number of threads getting their
1627                  * client data interspersed on disk
1628                  */
1629                 if (setup_files(td)) {
1630                         exit_value++;
1631                         if (td->error)
1632                                 log_err("fio: pid=%d, err=%d/%s\n",
1633                                         (int) td->pid, td->error, td->verror);
1634                         td_set_runstate(td, TD_REAPED);
1635                         todo--;
1636                 } else {
1637                         struct fio_file *f;
1638                         unsigned int j;
1639
1640                         /*
1641                          * for sharing to work, each job must always open
1642                          * its own files. so close them, if we opened them
1643                          * for creation
1644                          */
1645                         for_each_file(td, f, j) {
1646                                 if (fio_file_open(f))
1647                                         td_io_close_file(td, f);
1648                         }
1649                 }
1650         }
1651
1652         /* start idle threads before io threads start to run */
1653         fio_idle_prof_start();
1654
1655         set_genesis_time();
1656
1657         while (todo) {
1658                 struct thread_data *map[REAL_MAX_JOBS];
1659                 struct timeval this_start;
1660                 int this_jobs = 0, left;
1661
1662                 /*
1663                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1664                  */
1665                 for_each_td(td, i) {
1666                         if (td->runstate != TD_NOT_CREATED)
1667                                 continue;
1668
1669                         /*
1670                          * never got a chance to start, killed by other
1671                          * thread for some reason
1672                          */
1673                         if (td->terminate) {
1674                                 todo--;
1675                                 continue;
1676                         }
1677
1678                         if (td->o.start_delay) {
1679                                 spent = mtime_since_genesis();
1680
1681                                 if (td->o.start_delay * 1000 > spent)
1682                                         continue;
1683                         }
1684
1685                         if (td->o.stonewall && (nr_started || nr_running)) {
1686                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1687                                                         td->o.name);
1688                                 break;
1689                         }
1690
1691                         init_disk_util(td);
1692
1693                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1694                         td->update_rusage = 0;
1695
1696                         /*
1697                          * Set state to created. Thread will transition
1698                          * to TD_INITIALIZED when it's done setting up.
1699                          */
1700                         td_set_runstate(td, TD_CREATED);
1701                         map[this_jobs++] = td;
1702                         nr_started++;
1703
1704                         if (td->o.use_thread) {
1705                                 int ret;
1706
1707                                 dprint(FD_PROCESS, "will pthread_create\n");
1708                                 ret = pthread_create(&td->thread, NULL,
1709                                                         thread_main, td);
1710                                 if (ret) {
1711                                         log_err("pthread_create: %s\n",
1712                                                         strerror(ret));
1713                                         nr_started--;
1714                                         break;
1715                                 }
1716                                 ret = pthread_detach(td->thread);
1717                                 if (ret)
1718                                         log_err("pthread_detach: %s",
1719                                                         strerror(ret));
1720                         } else {
1721                                 pid_t pid;
1722                                 dprint(FD_PROCESS, "will fork\n");
1723                                 pid = fork();
1724                                 if (!pid) {
1725                                         int ret = fork_main(shm_id, i);
1726
1727                                         _exit(ret);
1728                                 } else if (i == fio_debug_jobno)
1729                                         *fio_debug_jobp = pid;
1730                         }
1731                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1732                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1733                                 log_err("fio: job startup hung? exiting.\n");
1734                                 fio_terminate_threads(TERMINATE_ALL);
1735                                 fio_abort = 1;
1736                                 nr_started--;
1737                                 break;
1738                         }
1739                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1740                 }
1741
1742                 /*
1743                  * Wait for the started threads to transition to
1744                  * TD_INITIALIZED.
1745                  */
1746                 fio_gettime(&this_start, NULL);
1747                 left = this_jobs;
1748                 while (left && !fio_abort) {
1749                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1750                                 break;
1751
1752                         do_usleep(100000);
1753
1754                         for (i = 0; i < this_jobs; i++) {
1755                                 td = map[i];
1756                                 if (!td)
1757                                         continue;
1758                                 if (td->runstate == TD_INITIALIZED) {
1759                                         map[i] = NULL;
1760                                         left--;
1761                                 } else if (td->runstate >= TD_EXITED) {
1762                                         map[i] = NULL;
1763                                         left--;
1764                                         todo--;
1765                                         nr_running++; /* work-around... */
1766                                 }
1767                         }
1768                 }
1769
1770                 if (left) {
1771                         log_err("fio: %d job%s failed to start\n", left,
1772                                         left > 1 ? "s" : "");
1773                         for (i = 0; i < this_jobs; i++) {
1774                                 td = map[i];
1775                                 if (!td)
1776                                         continue;
1777                                 kill(td->pid, SIGTERM);
1778                         }
1779                         break;
1780                 }
1781
1782                 /*
1783                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1784                  */
1785                 for_each_td(td, i) {
1786                         if (td->runstate != TD_INITIALIZED)
1787                                 continue;
1788
1789                         if (in_ramp_time(td))
1790                                 td_set_runstate(td, TD_RAMP);
1791                         else
1792                                 td_set_runstate(td, TD_RUNNING);
1793                         nr_running++;
1794                         nr_started--;
1795                         m_rate += ddir_rw_sum(td->o.ratemin);
1796                         t_rate += ddir_rw_sum(td->o.rate);
1797                         todo--;
1798                         fio_mutex_up(td->mutex);
1799                 }
1800
1801                 reap_threads(&nr_running, &t_rate, &m_rate);
1802
1803                 if (todo)
1804                         do_usleep(100000);
1805         }
1806
1807         while (nr_running) {
1808                 reap_threads(&nr_running, &t_rate, &m_rate);
1809                 do_usleep(10000);
1810         }
1811
1812         fio_idle_prof_stop();
1813
1814         update_io_ticks();
1815 }
1816
1817 void wait_for_disk_thread_exit(void)
1818 {
1819         fio_mutex_down(disk_thread_mutex);
1820 }
1821
1822 static void free_disk_util(void)
1823 {
1824         disk_util_start_exit();
1825         wait_for_disk_thread_exit();
1826         disk_util_prune_entries();
1827 }
1828
1829 static void *disk_thread_main(void *data)
1830 {
1831         int ret = 0;
1832
1833         fio_mutex_up(startup_mutex);
1834
1835         while (threads && !ret) {
1836                 usleep(DISK_UTIL_MSEC * 1000);
1837                 if (!threads)
1838                         break;
1839                 ret = update_io_ticks();
1840
1841                 if (!is_backend)
1842                         print_thread_status();
1843         }
1844
1845         fio_mutex_up(disk_thread_mutex);
1846         return NULL;
1847 }
1848
1849 static int create_disk_util_thread(void)
1850 {
1851         int ret;
1852
1853         setup_disk_util();
1854
1855         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1856
1857         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1858         if (ret) {
1859                 fio_mutex_remove(disk_thread_mutex);
1860                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1861                 return 1;
1862         }
1863
1864         ret = pthread_detach(disk_util_thread);
1865         if (ret) {
1866                 fio_mutex_remove(disk_thread_mutex);
1867                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1868                 return 1;
1869         }
1870
1871         dprint(FD_MUTEX, "wait on startup_mutex\n");
1872         fio_mutex_down(startup_mutex);
1873         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1874         return 0;
1875 }
1876
1877 int fio_backend(void)
1878 {
1879         struct thread_data *td;
1880         int i;
1881
1882         if (exec_profile) {
1883                 if (load_profile(exec_profile))
1884                         return 1;
1885                 free(exec_profile);
1886                 exec_profile = NULL;
1887         }
1888         if (!thread_number)
1889                 return 0;
1890
1891         if (write_bw_log) {
1892                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1893                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1894                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1895         }
1896
1897         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1898         if (startup_mutex == NULL)
1899                 return 1;
1900         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1901         if (writeout_mutex == NULL)
1902                 return 1;
1903
1904         set_genesis_time();
1905         stat_init();
1906         create_disk_util_thread();
1907
1908         cgroup_list = smalloc(sizeof(*cgroup_list));
1909         INIT_FLIST_HEAD(cgroup_list);
1910
1911         run_threads();
1912
1913         if (!fio_abort) {
1914                 show_run_stats();
1915                 if (write_bw_log) {
1916                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1917                         __finish_log(agg_io_log[DDIR_WRITE],
1918                                         "agg-write_bw.log");
1919                         __finish_log(agg_io_log[DDIR_TRIM],
1920                                         "agg-write_bw.log");
1921                 }
1922         }
1923
1924         for_each_td(td, i)
1925                 fio_options_free(td);
1926
1927         free_disk_util();
1928         cgroup_kill(cgroup_list);
1929         sfree(cgroup_list);
1930         sfree(cgroup_mnt);
1931
1932         fio_mutex_remove(startup_mutex);
1933         fio_mutex_remove(writeout_mutex);
1934         fio_mutex_remove(disk_thread_mutex);
1935         stat_exit();
1936         return exit_value;
1937 }