2ee7dcaf81a8c49f800fee31cd58677506896fc4
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 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 <signal.h>
28 #include <time.h>
29 #include <locale.h>
30 #include <assert.h>
31 #include <sys/stat.h>
32 #include <sys/wait.h>
33 #include <sys/ipc.h>
34 #include <sys/shm.h>
35 #include <sys/mman.h>
36
37 #include "fio.h"
38 #include "hash.h"
39
40 unsigned long page_mask;
41 unsigned long page_size;
42 #define ALIGN(buf)      \
43         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
44
45 int groupid = 0;
46 int thread_number = 0;
47 int nr_process = 0;
48 int nr_thread = 0;
49 int shm_id = 0;
50 int temp_stall_ts;
51
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
55
56 struct io_log *agg_io_log[2];
57
58 #define TERMINATE_ALL           (-1)
59 #define JOB_START_TIMEOUT       (5 * 1000)
60
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
62 {
63         dprint(FD_PROCESS, "%d: runstate %d -> %d\n", td->pid, td->runstate,
64                                                         runstate);
65         td->runstate = runstate;
66 }
67
68 static void terminate_threads(int group_id)
69 {
70         struct thread_data *td;
71         int i;
72
73         for_each_td(td, i) {
74                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
75                         dprint(FD_PROCESS, "setting terminate on %d\n",td->pid);
76                         /*
77                          * if the thread is running, just let it exit
78                          */
79                         if (td->runstate < TD_RUNNING)
80                                 kill(td->pid, SIGQUIT);
81                         td->terminate = 1;
82                         td->o.start_delay = 0;
83                 }
84         }
85 }
86
87 static void sig_handler(int sig)
88 {
89         switch (sig) {
90                 case SIGALRM:
91                         update_io_ticks();
92                         disk_util_timer_arm();
93                         print_thread_status();
94                         break;
95                 default:
96                         printf("\nfio: terminating on signal %d\n", sig);
97                         fflush(stdout);
98                         terminate_threads(TERMINATE_ALL);
99                         break;
100         }
101 }
102
103 /*
104  * Check if we are above the minimum rate given.
105  */
106 static int check_min_rate(struct thread_data *td, struct timeval *now)
107 {
108         unsigned long long bytes = 0;
109         unsigned long iops = 0;
110         unsigned long spent;
111         unsigned long rate;
112
113         /*
114          * No minimum rate set, always ok
115          */
116         if (!td->o.ratemin && !td->o.rate_iops_min)
117                 return 0;
118
119         /*
120          * allow a 2 second settle period in the beginning
121          */
122         if (mtime_since(&td->start, now) < 2000)
123                 return 0;
124
125         if (td_read(td)) {
126                 iops += td->io_blocks[DDIR_READ];
127                 bytes += td->this_io_bytes[DDIR_READ];
128         }
129         if (td_write(td)) {
130                 iops += td->io_blocks[DDIR_WRITE];
131                 bytes += td->this_io_bytes[DDIR_WRITE];
132         }
133
134         /*
135          * if rate blocks is set, sample is running
136          */
137         if (td->rate_bytes || td->rate_blocks) {
138                 spent = mtime_since(&td->lastrate, now);
139                 if (spent < td->o.ratecycle)
140                         return 0;
141
142                 if (td->o.rate) {
143                         /*
144                          * check bandwidth specified rate
145                          */
146                         if (bytes < td->rate_bytes) {
147                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
148                                 return 1;
149                         } else {
150                                 rate = (bytes - td->rate_bytes) / spent;
151                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
152                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
153                                         return 1;
154                                 }
155                         }
156                 } else {
157                         /*
158                          * checks iops specified rate
159                          */
160                         if (iops < td->o.rate_iops) {
161                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
162                                 return 1;
163                         } else {
164                                 rate = (iops - td->rate_blocks) / spent;
165                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
166                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
167                                 }
168                         }
169                 }
170         }
171
172         td->rate_bytes = bytes;
173         td->rate_blocks = iops;
174         memcpy(&td->lastrate, now, sizeof(*now));
175         return 0;
176 }
177
178 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
179 {
180         if (!td->o.timeout)
181                 return 0;
182         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
183                 return 1;
184
185         return 0;
186 }
187
188 /*
189  * When job exits, we can cancel the in-flight IO if we are using async
190  * io. Attempt to do so.
191  */
192 static void cleanup_pending_aio(struct thread_data *td)
193 {
194         struct list_head *entry, *n;
195         struct io_u *io_u;
196         int r;
197
198         /*
199          * get immediately available events, if any
200          */
201         r = io_u_queued_complete(td, 0);
202         if (r < 0)
203                 return;
204
205         /*
206          * now cancel remaining active events
207          */
208         if (td->io_ops->cancel) {
209                 list_for_each_safe(entry, n, &td->io_u_busylist) {
210                         io_u = list_entry(entry, struct io_u, list);
211
212                         /*
213                          * if the io_u isn't in flight, then that generally
214                          * means someone leaked an io_u. complain but fix
215                          * it up, so we don't stall here.
216                          */
217                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
218                                 log_err("fio: non-busy IO on busy list\n");
219                                 put_io_u(td, io_u);
220                         } else {
221                                 r = td->io_ops->cancel(td, io_u);
222                                 if (!r)
223                                         put_io_u(td, io_u);
224                         }
225                 }
226         }
227
228         if (td->cur_depth)
229                 r = io_u_queued_complete(td, td->cur_depth);
230 }
231
232 /*
233  * Helper to handle the final sync of a file. Works just like the normal
234  * io path, just does everything sync.
235  */
236 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
237 {
238         struct io_u *io_u = __get_io_u(td);
239         int ret;
240
241         if (!io_u)
242                 return 1;
243
244         io_u->ddir = DDIR_SYNC;
245         io_u->file = f;
246
247         if (td_io_prep(td, io_u)) {
248                 put_io_u(td, io_u);
249                 return 1;
250         }
251
252 requeue:
253         ret = td_io_queue(td, io_u);
254         if (ret < 0) {
255                 td_verror(td, io_u->error, "td_io_queue");
256                 put_io_u(td, io_u);
257                 return 1;
258         } else if (ret == FIO_Q_QUEUED) {
259                 if (io_u_queued_complete(td, 1) < 0)
260                         return 1;
261         } else if (ret == FIO_Q_COMPLETED) {
262                 if (io_u->error) {
263                         td_verror(td, io_u->error, "td_io_queue");
264                         return 1;
265                 }
266
267                 if (io_u_sync_complete(td, io_u) < 0)
268                         return 1;
269         } else if (ret == FIO_Q_BUSY) {
270                 if (td_io_commit(td))
271                         return 1;
272                 goto requeue;
273         }
274
275         return 0;
276 }
277
278 /*
279  * The main verify engine. Runs over the writes we previously submitted,
280  * reads the blocks back in, and checks the crc/md5 of the data.
281  */
282 static void do_verify(struct thread_data *td)
283 {
284         struct fio_file *f;
285         struct io_u *io_u;
286         int ret, min_events;
287         unsigned int i;
288
289         /*
290          * sync io first and invalidate cache, to make sure we really
291          * read from disk.
292          */
293         for_each_file(td, f, i) {
294                 if (!(f->flags & FIO_FILE_OPEN))
295                         continue;
296                 if (fio_io_sync(td, f))
297                         break;
298                 if (file_invalidate_cache(td, f))
299                         break;
300         }
301
302         if (td->error)
303                 return;
304
305         td_set_runstate(td, TD_VERIFYING);
306
307         io_u = NULL;
308         while (!td->terminate) {
309                 int ret2;
310
311                 io_u = __get_io_u(td);
312                 if (!io_u)
313                         break;
314
315                 if (runtime_exceeded(td, &io_u->start_time)) {
316                         put_io_u(td, io_u);
317                         td->terminate = 1;
318                         break;
319                 }
320
321                 if (get_next_verify(td, io_u)) {
322                         put_io_u(td, io_u);
323                         break;
324                 }
325
326                 if (td_io_prep(td, io_u)) {
327                         put_io_u(td, io_u);
328                         break;
329                 }
330
331                 io_u->end_io = verify_io_u;
332
333                 ret = td_io_queue(td, io_u);
334                 switch (ret) {
335                 case FIO_Q_COMPLETED:
336                         if (io_u->error)
337                                 ret = -io_u->error;
338                         else if (io_u->resid) {
339                                 int bytes = io_u->xfer_buflen - io_u->resid;
340                                 struct fio_file *f = io_u->file;
341
342                                 /*
343                                  * zero read, fail
344                                  */
345                                 if (!bytes) {
346                                         td_verror(td, ENODATA, "full resid");
347                                         put_io_u(td, io_u);
348                                         break;
349                                 }
350
351                                 io_u->xfer_buflen = io_u->resid;
352                                 io_u->xfer_buf += bytes;
353                                 io_u->offset += bytes;
354                                 f->last_completed_pos = io_u->offset;
355
356                                 td->ts.short_io_u[io_u->ddir]++;
357
358                                 if (io_u->offset == f->real_file_size)
359                                         goto sync_done;
360
361                                 requeue_io_u(td, &io_u);
362                         } else {
363 sync_done:
364                                 ret = io_u_sync_complete(td, io_u);
365                                 if (ret < 0)
366                                         break;
367                         }
368                         continue;
369                 case FIO_Q_QUEUED:
370                         break;
371                 case FIO_Q_BUSY:
372                         requeue_io_u(td, &io_u);
373                         ret2 = td_io_commit(td);
374                         if (ret2 < 0)
375                                 ret = ret2;
376                         break;
377                 default:
378                         assert(ret < 0);
379                         td_verror(td, -ret, "td_io_queue");
380                         break;
381                 }
382
383                 if (ret < 0 || td->error)
384                         break;
385
386                 /*
387                  * if we can queue more, do so. but check if there are
388                  * completed io_u's first.
389                  */
390                 min_events = 0;
391                 if (queue_full(td) || ret == FIO_Q_BUSY) {
392                         min_events = 1;
393
394                         if (td->cur_depth > td->o.iodepth_low)
395                                 min_events = td->cur_depth - td->o.iodepth_low;
396                 }
397
398                 /*
399                  * Reap required number of io units, if any, and do the
400                  * verification on them through the callback handler
401                  */
402                 if (io_u_queued_complete(td, min_events) < 0)
403                         break;
404         }
405
406         if (!td->error) {
407                 min_events = td->cur_depth;
408
409                 if (min_events)
410                         ret = io_u_queued_complete(td, min_events);
411         } else
412                 cleanup_pending_aio(td);
413
414         td_set_runstate(td, TD_RUNNING);
415 }
416
417 /*
418  * Main IO worker function. It retrieves io_u's to process and queues
419  * and reaps them, checking for rate and errors along the way.
420  */
421 static void do_io(struct thread_data *td)
422 {
423         struct timeval s;
424         unsigned long usec;
425         unsigned int i;
426         int ret = 0;
427
428         td_set_runstate(td, TD_RUNNING);
429
430         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
431                 struct timeval comp_time;
432                 long bytes_done = 0;
433                 int min_evts = 0;
434                 struct io_u *io_u;
435                 int ret2;
436
437                 if (td->terminate)
438                         break;
439
440                 io_u = get_io_u(td);
441                 if (!io_u)
442                         break;
443
444                 memcpy(&s, &io_u->start_time, sizeof(s));
445
446                 if (runtime_exceeded(td, &s)) {
447                         put_io_u(td, io_u);
448                         td->terminate = 1;
449                         break;
450                 }
451
452                 /*
453                  * Add verification end_io handler, if asked to verify
454                  * a previously written file.
455                  */
456                 if (td->o.verify != VERIFY_NONE)
457                         io_u->end_io = verify_io_u;
458
459                 ret = td_io_queue(td, io_u);
460                 switch (ret) {
461                 case FIO_Q_COMPLETED:
462                         if (io_u->error)
463                                 ret = -io_u->error;
464                         else if (io_u->resid) {
465                                 int bytes = io_u->xfer_buflen - io_u->resid;
466                                 struct fio_file *f = io_u->file;
467
468                                 /*
469                                  * zero read, fail
470                                  */
471                                 if (!bytes) {
472                                         td_verror(td, ENODATA, "full resid");
473                                         put_io_u(td, io_u);
474                                         break;
475                                 }
476
477                                 io_u->xfer_buflen = io_u->resid;
478                                 io_u->xfer_buf += bytes;
479                                 io_u->offset += bytes;
480                                 f->last_completed_pos = io_u->offset;
481
482                                 td->ts.short_io_u[io_u->ddir]++;
483
484                                 if (io_u->offset == f->real_file_size)
485                                         goto sync_done;
486
487                                 requeue_io_u(td, &io_u);
488                         } else {
489 sync_done:
490                                 fio_gettime(&comp_time, NULL);
491                                 bytes_done = io_u_sync_complete(td, io_u);
492                                 if (bytes_done < 0)
493                                         ret = bytes_done;
494                         }
495                         break;
496                 case FIO_Q_QUEUED:
497                         /*
498                          * if the engine doesn't have a commit hook,
499                          * the io_u is really queued. if it does have such
500                          * a hook, it has to call io_u_queued() itself.
501                          */
502                         if (td->io_ops->commit == NULL)
503                                 io_u_queued(td, io_u);
504                         break;
505                 case FIO_Q_BUSY:
506                         requeue_io_u(td, &io_u);
507                         ret2 = td_io_commit(td);
508                         if (ret2 < 0)
509                                 ret = ret2;
510                         break;
511                 default:
512                         assert(ret < 0);
513                         put_io_u(td, io_u);
514                         break;
515                 }
516
517                 if (ret < 0 || td->error)
518                         break;
519
520                 /*
521                  * See if we need to complete some commands
522                  */
523                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
524                         min_evts = 0;
525                         if (queue_full(td) || ret == FIO_Q_BUSY) {
526                                 min_evts = 1;
527
528                                 if (td->cur_depth > td->o.iodepth_low)
529                                         min_evts = td->cur_depth - td->o.iodepth_low;
530                         }
531
532                         fio_gettime(&comp_time, NULL);
533                         bytes_done = io_u_queued_complete(td, min_evts);
534                         if (bytes_done < 0)
535                                 break;
536                 }
537
538                 if (!bytes_done)
539                         continue;
540
541                 /*
542                  * the rate is batched for now, it should work for batches
543                  * of completions except the very first one which may look
544                  * a little bursty
545                  */
546                 usec = utime_since(&s, &comp_time);
547
548                 rate_throttle(td, usec, bytes_done);
549
550                 if (check_min_rate(td, &comp_time)) {
551                         if (exitall_on_terminate)
552                                 terminate_threads(td->groupid);
553                         td_verror(td, ENODATA, "check_min_rate");
554                         break;
555                 }
556
557                 if (td->o.thinktime) {
558                         unsigned long long b;
559
560                         b = td->io_blocks[0] + td->io_blocks[1];
561                         if (!(b % td->o.thinktime_blocks)) {
562                                 int left;
563
564                                 if (td->o.thinktime_spin)
565                                         __usec_sleep(td->o.thinktime_spin);
566
567                                 left = td->o.thinktime - td->o.thinktime_spin;
568                                 if (left)
569                                         usec_sleep(td, left);
570                         }
571                 }
572         }
573
574         /*
575          * stop job if we failed doing any IO
576          */
577         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
578                 td->done = 1;
579
580         if (td->o.fill_device && td->error == ENOSPC) {
581                 td->error = 0;
582                 td->terminate = 1;
583         }
584         if (!td->error) {
585                 struct fio_file *f;
586
587                 i = td->cur_depth;
588                 if (i)
589                         ret = io_u_queued_complete(td, i);
590
591                 if (should_fsync(td) && td->o.end_fsync) {
592                         td_set_runstate(td, TD_FSYNCING);
593
594                         for_each_file(td, f, i) {
595                                 if (!(f->flags & FIO_FILE_OPEN))
596                                         continue;
597                                 fio_io_sync(td, f);
598                         }
599                 }
600         } else
601                 cleanup_pending_aio(td);
602 }
603
604 static void cleanup_io_u(struct thread_data *td)
605 {
606         struct list_head *entry, *n;
607         struct io_u *io_u;
608
609         list_for_each_safe(entry, n, &td->io_u_freelist) {
610                 io_u = list_entry(entry, struct io_u, list);
611
612                 list_del(&io_u->list);
613                 free(io_u);
614         }
615
616         free_io_mem(td);
617 }
618
619 /*
620  * "randomly" fill the buffer contents
621  */
622 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
623 {
624         long *ptr = io_u->buf;
625
626         if (!td->o.zero_buffers) {
627                 while ((void *) ptr - io_u->buf < max_bs) {
628                         *ptr = rand() * GOLDEN_RATIO_PRIME;
629                         ptr++;
630                 }
631         } else
632                 memset(ptr, 0, max_bs);
633 }
634
635 static int init_io_u(struct thread_data *td)
636 {
637         struct io_u *io_u;
638         unsigned int max_bs;
639         int i, max_units;
640         char *p;
641
642         if (td->io_ops->flags & FIO_SYNCIO)
643                 max_units = 1;
644         else
645                 max_units = td->o.iodepth;
646
647         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
648         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
649
650         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
651                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
652
653         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
654                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
655                 return 1;
656         }
657
658         if (allocate_io_mem(td))
659                 return 1;
660
661         if (td->o.odirect)
662                 p = ALIGN(td->orig_buffer);
663         else
664                 p = td->orig_buffer;
665
666         for (i = 0; i < max_units; i++) {
667                 if (td->terminate)
668                         return 1;
669                 io_u = malloc(sizeof(*io_u));
670                 memset(io_u, 0, sizeof(*io_u));
671                 INIT_LIST_HEAD(&io_u->list);
672
673                 if (!(td->io_ops->flags & FIO_NOIO)) {
674                         io_u->buf = p + max_bs * i;
675
676                         if (td_write(td))
677                                 fill_io_buf(td, io_u, max_bs);
678                 }
679
680                 io_u->index = i;
681                 io_u->flags = IO_U_F_FREE;
682                 list_add(&io_u->list, &td->io_u_freelist);
683         }
684
685         io_u_init_timeout();
686
687         return 0;
688 }
689
690 static int switch_ioscheduler(struct thread_data *td)
691 {
692         char tmp[256], tmp2[128];
693         FILE *f;
694         int ret;
695
696         if (td->io_ops->flags & FIO_DISKLESSIO)
697                 return 0;
698
699         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
700
701         f = fopen(tmp, "r+");
702         if (!f) {
703                 if (errno == ENOENT) {
704                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
705                         return 0;
706                 }
707                 td_verror(td, errno, "fopen iosched");
708                 return 1;
709         }
710
711         /*
712          * Set io scheduler.
713          */
714         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
715         if (ferror(f) || ret != 1) {
716                 td_verror(td, errno, "fwrite");
717                 fclose(f);
718                 return 1;
719         }
720
721         rewind(f);
722
723         /*
724          * Read back and check that the selected scheduler is now the default.
725          */
726         ret = fread(tmp, 1, sizeof(tmp), f);
727         if (ferror(f) || ret < 0) {
728                 td_verror(td, errno, "fread");
729                 fclose(f);
730                 return 1;
731         }
732
733         sprintf(tmp2, "[%s]", td->o.ioscheduler);
734         if (!strstr(tmp, tmp2)) {
735                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
736                 td_verror(td, EINVAL, "iosched_switch");
737                 fclose(f);
738                 return 1;
739         }
740
741         fclose(f);
742         return 0;
743 }
744
745 static int keep_running(struct thread_data *td)
746 {
747         unsigned long long io_done;
748
749         if (td->done)
750                 return 0;
751         if (td->o.time_based)
752                 return 1;
753         if (td->o.loops) {
754                 td->o.loops--;
755                 return 1;
756         }
757
758         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
759         if (io_done < td->o.size)
760                 return 1;
761
762         return 0;
763 }
764
765 static int clear_io_state(struct thread_data *td)
766 {
767         struct fio_file *f;
768         unsigned int i;
769         int ret;
770
771         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
772         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
773         td->zone_bytes = 0;
774         td->rate_bytes = 0;
775         td->rate_blocks = 0;
776         td->rw_end_set[0] = td->rw_end_set[1] = 0;
777
778         td->last_was_sync = 0;
779
780         if (td->o.time_based)
781                 td->nr_done_files = 0;
782
783         for_each_file(td, f, i)
784                 td_io_close_file(td, f);
785
786         ret = 0;
787         for_each_file(td, f, i) {
788                 f->flags &= ~FIO_FILE_DONE;
789                 ret = td_io_open_file(td, f);
790                 if (ret)
791                         break;
792         }
793
794         return ret;
795 }
796
797 /*
798  * Entry point for the thread based jobs. The process based jobs end up
799  * here as well, after a little setup.
800  */
801 static void *thread_main(void *data)
802 {
803         unsigned long long runtime[2], elapsed;
804         struct thread_data *td = data;
805         int clear_state;
806
807         if (!td->o.use_thread)
808                 setsid();
809
810         td->pid = getpid();
811
812         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
813
814         INIT_LIST_HEAD(&td->io_u_freelist);
815         INIT_LIST_HEAD(&td->io_u_busylist);
816         INIT_LIST_HEAD(&td->io_u_requeues);
817         INIT_LIST_HEAD(&td->io_log_list);
818         INIT_LIST_HEAD(&td->io_hist_list);
819         td->io_hist_tree = RB_ROOT;
820
821         td_set_runstate(td, TD_INITIALIZED);
822         fio_sem_up(startup_sem);
823         fio_sem_down(td->mutex);
824
825         /*
826          * the ->mutex semaphore is now no longer used, close it to avoid
827          * eating a file descriptor
828          */
829         fio_sem_remove(td->mutex);
830
831         /*
832          * May alter parameters that init_io_u() will use, so we need to
833          * do this first.
834          */
835         if (init_iolog(td))
836                 goto err;
837
838         if (init_io_u(td))
839                 goto err;
840
841         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
842                 td_verror(td, errno, "cpu_set_affinity");
843                 goto err;
844         }
845
846         if (td->ioprio_set) {
847                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
848                         td_verror(td, errno, "ioprio_set");
849                         goto err;
850                 }
851         }
852
853         if (nice(td->o.nice) == -1) {
854                 td_verror(td, errno, "nice");
855                 goto err;
856         }
857
858         if (td->o.ioscheduler && switch_ioscheduler(td))
859                 goto err;
860
861         if (!td->o.create_serialize && setup_files(td))
862                 goto err;
863
864         if (td_io_init(td))
865                 goto err;
866
867         if (open_files(td))
868                 goto err;
869
870         if (init_random_map(td))
871                 goto err;
872
873         if (td->o.exec_prerun) {
874                 if (system(td->o.exec_prerun) < 0)
875                         goto err;
876         }
877
878         fio_gettime(&td->epoch, NULL);
879         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
880         getrusage(RUSAGE_SELF, &td->ts.ru_start);
881
882         runtime[0] = runtime[1] = 0;
883         clear_state = 0;
884         while (keep_running(td)) {
885                 fio_gettime(&td->start, NULL);
886                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
887
888                 if (td->o.ratemin)
889                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
890
891                 if (clear_state && clear_io_state(td))
892                         break;
893
894                 prune_io_piece_log(td);
895
896                 do_io(td);
897
898                 clear_state = 1;
899
900                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
901                         if (td->rw_end_set[DDIR_READ])
902                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
903                         else
904                                 elapsed = utime_since_now(&td->start);
905
906                         runtime[DDIR_READ] += elapsed;
907                 }
908                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
909                         if (td->rw_end_set[DDIR_WRITE])
910                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
911                         else
912                                 elapsed = utime_since_now(&td->start);
913
914                         runtime[DDIR_WRITE] += elapsed;
915                 }
916                 
917                 if (td->error || td->terminate)
918                         break;
919
920                 if (!td->o.do_verify ||
921                     td->o.verify == VERIFY_NONE ||
922                     (td->io_ops->flags & FIO_UNIDIR))
923                         continue;
924
925                 if (clear_io_state(td))
926                         break;
927
928                 fio_gettime(&td->start, NULL);
929
930                 do_verify(td);
931
932                 runtime[DDIR_READ] += utime_since_now(&td->start);
933
934                 if (td->error || td->terminate)
935                         break;
936         }
937
938         update_rusage_stat(td);
939         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
940         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
941         td->ts.total_run_time = mtime_since_now(&td->epoch);
942         td->ts.io_bytes[0] = td->io_bytes[0];
943         td->ts.io_bytes[1] = td->io_bytes[1];
944
945         if (td->ts.bw_log)
946                 finish_log(td, td->ts.bw_log, "bw");
947         if (td->ts.slat_log)
948                 finish_log(td, td->ts.slat_log, "slat");
949         if (td->ts.clat_log)
950                 finish_log(td, td->ts.clat_log, "clat");
951         if (td->o.exec_postrun) {
952                 if (system(td->o.exec_postrun) < 0)
953                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
954         }
955
956         if (exitall_on_terminate)
957                 terminate_threads(td->groupid);
958
959 err:
960         if (td->error)
961                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
962         close_files(td);
963         close_ioengine(td);
964         cleanup_io_u(td);
965
966         /*
967          * do this very late, it will log file closing as well
968          */
969         if (td->o.write_iolog_file)
970                 write_iolog_close(td);
971
972         options_mem_free(td);
973         td_set_runstate(td, TD_EXITED);
974         return (void *) (unsigned long) td->error;
975 }
976
977 /*
978  * We cannot pass the td data into a forked process, so attach the td and
979  * pass it to the thread worker.
980  */
981 static int fork_main(int shmid, int offset)
982 {
983         struct thread_data *td;
984         void *data, *ret;
985
986         data = shmat(shmid, NULL, 0);
987         if (data == (void *) -1) {
988                 int __err = errno;
989
990                 perror("shmat");
991                 return __err;
992         }
993
994         td = data + offset * sizeof(struct thread_data);
995         ret = thread_main(td);
996         shmdt(data);
997         return (int) (unsigned long) ret;
998 }
999
1000 /*
1001  * Run over the job map and reap the threads that have exited, if any.
1002  */
1003 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1004 {
1005         struct thread_data *td;
1006         int i, cputhreads, realthreads, pending, status, ret;
1007
1008         /*
1009          * reap exited threads (TD_EXITED -> TD_REAPED)
1010          */
1011         realthreads = pending = cputhreads = 0;
1012         for_each_td(td, i) {
1013                 int flags = 0;
1014
1015                 /*
1016                  * ->io_ops is NULL for a thread that has closed its
1017                  * io engine
1018                  */
1019                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1020                         cputhreads++;
1021                 else
1022                         realthreads++;
1023
1024                 if (!td->pid || td->runstate == TD_REAPED)
1025                         continue;
1026                 if (td->o.use_thread) {
1027                         if (td->runstate == TD_EXITED) {
1028                                 td_set_runstate(td, TD_REAPED);
1029                                 goto reaped;
1030                         }
1031                         continue;
1032                 }
1033
1034                 flags = WNOHANG;
1035                 if (td->runstate == TD_EXITED)
1036                         flags = 0;
1037
1038                 /*
1039                  * check if someone quit or got killed in an unusual way
1040                  */
1041                 ret = waitpid(td->pid, &status, flags);
1042                 if (ret < 0) {
1043                         if (errno == ECHILD) {
1044                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1045                                 td_set_runstate(td, TD_REAPED);
1046                                 goto reaped;
1047                         }
1048                         perror("waitpid");
1049                 } else if (ret == td->pid) {
1050                         if (WIFSIGNALED(status)) {
1051                                 int sig = WTERMSIG(status);
1052
1053                                 if (sig != SIGQUIT)
1054                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1055                                 td_set_runstate(td, TD_REAPED);
1056                                 goto reaped;
1057                         }
1058                         if (WIFEXITED(status)) {
1059                                 if (WEXITSTATUS(status) && !td->error)
1060                                         td->error = WEXITSTATUS(status);
1061
1062                                 td_set_runstate(td, TD_REAPED);
1063                                 goto reaped;
1064                         }
1065                 }
1066
1067                 /*
1068                  * thread is not dead, continue
1069                  */
1070                 pending++;
1071                 continue;
1072 reaped:
1073                 if (td->o.use_thread) {
1074                         long ret;
1075
1076                         dprint(FD_PROCESS, "joining tread %d\n", td->pid);
1077                         if (pthread_join(td->thread, (void *) &ret)) {
1078                                 dprint(FD_PROCESS, "join failed %ld\n", ret);
1079                                 perror("pthread_join");
1080                         }
1081                 }
1082
1083                 (*nr_running)--;
1084                 (*m_rate) -= td->o.ratemin;
1085                 (*t_rate) -= td->o.rate;
1086                 pending--;
1087
1088                 if (td->error)
1089                         exit_value++;
1090         }
1091
1092         if (*nr_running == cputhreads && !pending && realthreads)
1093                 terminate_threads(TERMINATE_ALL);
1094 }
1095
1096 /*
1097  * Main function for kicking off and reaping jobs, as needed.
1098  */
1099 static void run_threads(void)
1100 {
1101         struct thread_data *td;
1102         unsigned long spent;
1103         int i, todo, nr_running, m_rate, t_rate, nr_started;
1104
1105         if (fio_pin_memory())
1106                 return;
1107
1108         if (!terse_output) {
1109                 printf("Starting ");
1110                 if (nr_thread)
1111                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1112                 if (nr_process) {
1113                         if (nr_thread)
1114                                 printf(" and ");
1115                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1116                 }
1117                 printf("\n");
1118                 fflush(stdout);
1119         }
1120
1121         signal(SIGINT, sig_handler);
1122         signal(SIGALRM, sig_handler);
1123
1124         todo = thread_number;
1125         nr_running = 0;
1126         nr_started = 0;
1127         m_rate = t_rate = 0;
1128
1129         for_each_td(td, i) {
1130                 print_status_init(td->thread_number - 1);
1131
1132                 if (!td->o.create_serialize) {
1133                         init_disk_util(td);
1134                         continue;
1135                 }
1136
1137                 /*
1138                  * do file setup here so it happens sequentially,
1139                  * we don't want X number of threads getting their
1140                  * client data interspersed on disk
1141                  */
1142                 if (setup_files(td)) {
1143                         exit_value++;
1144                         if (td->error)
1145                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1146                         td_set_runstate(td, TD_REAPED);
1147                         todo--;
1148                 }
1149
1150                 init_disk_util(td);
1151         }
1152
1153         set_genesis_time();
1154
1155         while (todo) {
1156                 struct thread_data *map[MAX_JOBS];
1157                 struct timeval this_start;
1158                 int this_jobs = 0, left;
1159
1160                 /*
1161                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1162                  */
1163                 for_each_td(td, i) {
1164                         if (td->runstate != TD_NOT_CREATED)
1165                                 continue;
1166
1167                         /*
1168                          * never got a chance to start, killed by other
1169                          * thread for some reason
1170                          */
1171                         if (td->terminate) {
1172                                 todo--;
1173                                 continue;
1174                         }
1175
1176                         if (td->o.start_delay) {
1177                                 spent = mtime_since_genesis();
1178
1179                                 if (td->o.start_delay * 1000 > spent)
1180                                         continue;
1181                         }
1182
1183                         if (td->o.stonewall && (nr_started || nr_running))
1184                                 break;
1185
1186                         /*
1187                          * Set state to created. Thread will transition
1188                          * to TD_INITIALIZED when it's done setting up.
1189                          */
1190                         td_set_runstate(td, TD_CREATED);
1191                         map[this_jobs++] = td;
1192                         nr_started++;
1193
1194                         if (td->o.use_thread) {
1195                                 dprint(FD_PROCESS, "will pthread_create\n");
1196                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1197                                         perror("thread_create");
1198                                         nr_started--;
1199                                         break;
1200                                 }
1201                         } else {
1202                                 dprint(FD_PROCESS, "will fork\n");
1203                                 if (!fork()) {
1204                                         int ret = fork_main(shm_id, i);
1205
1206                                         exit(ret);
1207                                 }
1208                         }
1209                         fio_sem_down(startup_sem);
1210                 }
1211
1212                 /*
1213                  * Wait for the started threads to transition to
1214                  * TD_INITIALIZED.
1215                  */
1216                 fio_gettime(&this_start, NULL);
1217                 left = this_jobs;
1218                 while (left && !fio_abort) {
1219                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1220                                 break;
1221
1222                         usleep(100000);
1223
1224                         for (i = 0; i < this_jobs; i++) {
1225                                 td = map[i];
1226                                 if (!td)
1227                                         continue;
1228                                 if (td->runstate == TD_INITIALIZED) {
1229                                         map[i] = NULL;
1230                                         left--;
1231                                 } else if (td->runstate >= TD_EXITED) {
1232                                         map[i] = NULL;
1233                                         left--;
1234                                         todo--;
1235                                         nr_running++; /* work-around... */
1236                                 }
1237                         }
1238                 }
1239
1240                 if (left) {
1241                         log_err("fio: %d jobs failed to start\n", left);
1242                         for (i = 0; i < this_jobs; i++) {
1243                                 td = map[i];
1244                                 if (!td)
1245                                         continue;
1246                                 kill(td->pid, SIGTERM);
1247                         }
1248                         break;
1249                 }
1250
1251                 /*
1252                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1253                  */
1254                 for_each_td(td, i) {
1255                         if (td->runstate != TD_INITIALIZED)
1256                                 continue;
1257
1258                         td_set_runstate(td, TD_RUNNING);
1259                         nr_running++;
1260                         nr_started--;
1261                         m_rate += td->o.ratemin;
1262                         t_rate += td->o.rate;
1263                         todo--;
1264                         fio_sem_up(td->mutex);
1265                 }
1266
1267                 reap_threads(&nr_running, &t_rate, &m_rate);
1268
1269                 if (todo)
1270                         usleep(100000);
1271         }
1272
1273         while (nr_running) {
1274                 reap_threads(&nr_running, &t_rate, &m_rate);
1275                 usleep(10000);
1276         }
1277
1278         update_io_ticks();
1279         fio_unpin_memory();
1280 }
1281
1282 int main(int argc, char *argv[])
1283 {
1284         long ps;
1285
1286         /*
1287          * We need locale for number printing, if it isn't set then just
1288          * go with the US format.
1289          */
1290         if (!getenv("LC_NUMERIC"))
1291                 setlocale(LC_NUMERIC, "en_US");
1292
1293         if (parse_options(argc, argv))
1294                 return 1;
1295
1296         if (!thread_number)
1297                 return 0;
1298
1299         ps = sysconf(_SC_PAGESIZE);
1300         if (ps < 0) {
1301                 log_err("Failed to get page size\n");
1302                 return 1;
1303         }
1304
1305         page_size = ps;
1306         page_mask = ps - 1;
1307
1308         if (write_bw_log) {
1309                 setup_log(&agg_io_log[DDIR_READ]);
1310                 setup_log(&agg_io_log[DDIR_WRITE]);
1311         }
1312
1313         startup_sem = fio_sem_init(0);
1314
1315         set_genesis_time();
1316
1317         disk_util_timer_arm();
1318
1319         run_threads();
1320
1321         if (!fio_abort) {
1322                 show_run_stats();
1323                 if (write_bw_log) {
1324                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1325                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1326                 }
1327         }
1328
1329         fio_sem_remove(startup_sem);
1330         return exit_value;
1331 }