1 /*
2  * Copyright (c) 2004 SuSE, Inc.  All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11  *
12  * Further, this software is distributed without any warranty that it is
13  * free of the rightful claim of any third person regarding infringement
14  * or the like.  Any license provided herein, whether implied or
15  * otherwise, applies only to this software file.  Patent licenses, if
16  * any, provided herein do not apply to combinations of this program with
17  * other software, or any other product whatsoever.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, write the Free Software Foundation, Inc., 59
21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22  *
23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24  * Mountain View, CA  94043, or:
25  *
26  *
27  * aio-stress
28  *
29  * will open or create each file on the command line, and start a series
30  * of aio to it.
31  *
32  * aio is done in a rotating loop.  first file1 gets 8 requests, then
33  * file2, then file3 etc.  As each file finishes writing, it is switched
34  * to reads
35  *
36  * io buffers are aligned in case you want to do raw io
37  *
38  * compile with gcc -Wall -laio -lpthread -o aio-stress aio-stress.c
39  *
40  * run aio-stress -h to see the options
41  *
42  * Please mail Chris Mason (mason@suse.com) with bug reports or patches
43  */
44 #define _FILE_OFFSET_BITS 64
45 #define PROG_VERSION "0.21"
46 #define NEW_GETEVENTS
47 
48 #include <stdio.h>
49 #include <errno.h>
50 #include <assert.h>
51 #include <stdlib.h>
52 
53 #include <sys/types.h>
54 #include <sys/stat.h>
55 #include <fcntl.h>
56 #include <unistd.h>
57 #include <sys/time.h>
58 #include <libaio.h>
59 #include <sys/ipc.h>
60 #include <sys/shm.h>
61 #include <sys/mman.h>
62 #include <string.h>
63 #include <pthread.h>
64 
65 #define IO_FREE 0
66 #define IO_PENDING 1
67 #define RUN_FOREVER -1
68 
69 #ifndef O_DIRECT
70 #define O_DIRECT         040000 /* direct disk access hint */
71 #endif
72 
73 enum {
74     WRITE,
75     READ,
76     RWRITE,
77     RREAD,
78     LAST_STAGE,
79 };
80 
81 #define USE_MALLOC 0
82 #define USE_SHM 1
83 #define USE_SHMFS 2
84 
85 /*
86  * various globals, these are effectively read only by the time the threads
87  * are started
88  */
89 long stages = 0;
90 unsigned long page_size_mask;
91 int o_direct = 0;
92 int o_sync = 0;
93 int latency_stats = 0;
94 int completion_latency_stats = 0;
95 int io_iter = 8;
96 int iterations = RUN_FOREVER;
97 int max_io_submit = 0;
98 long rec_len = 64 * 1024;
99 int depth = 64;
100 int num_threads = 1;
101 int num_contexts = 1;
102 off_t context_offset = 2 * 1024 * 1024;
103 int fsync_stages = 1;
104 int use_shm = 0;
105 int shm_id;
106 char *unaligned_buffer = NULL;
107 char *aligned_buffer = NULL;
108 int padded_reclen = 0;
109 int stonewall = 1;
110 int verify = 0;
111 char *verify_buf = NULL;
112 int unlink_files = 0;
113 
114 struct io_unit;
115 struct thread_info;
116 
117 /* pthread mutexes and other globals for keeping the threads in sync */
118 pthread_cond_t stage_cond = PTHREAD_COND_INITIALIZER;
119 pthread_mutex_t stage_mutex = PTHREAD_MUTEX_INITIALIZER;
120 int threads_ending = 0;
121 int threads_starting = 0;
122 struct timeval global_stage_start_time;
123 struct thread_info *global_thread_info;
124 
125 /*
126  * latencies during io_submit are measured, these are the
127  * granularities for deviations
128  */
129 #define DEVIATIONS 6
130 int deviations[DEVIATIONS] = { 100, 250, 500, 1000, 5000, 10000 };
131 struct io_latency {
132     double max;
133     double min;
134     double total_io;
135     double total_lat;
136     double deviations[DEVIATIONS];
137 };
138 
139 /* container for a series of operations to a file */
140 struct io_oper {
141     /* already open file descriptor, valid for whatever operation you want */
142     int fd;
143 
144     /* starting byte of the operation */
145     off_t start;
146 
147     /* ending byte of the operation */
148     off_t end;
149 
150     /* size of the read/write buffer */
151     int reclen;
152 
153     /* max number of pending requests before a wait is triggered */
154     int depth;
155 
156     /* current number of pending requests */
157     int num_pending;
158 
159     /* last error, zero if there were none */
160     int last_err;
161 
162     /* total number of errors hit. */
163     int num_err;
164 
165     /* read,write, random, etc */
166     int rw;
167 
168     /* number of ios that will get sent to aio */
169     int total_ios;
170 
171     /* number of ios we've already sent */
172     int started_ios;
173 
174     /* last offset used in an io operation */
175     off_t last_offset;
176 
177     /* stonewalled = 1 when we got cut off before submitting all our ios */
178     int stonewalled;
179 
180     /* list management */
181     struct io_oper *next;
182     struct io_oper *prev;
183 
184     struct timeval start_time;
185 
186     char *file_name;
187 };
188 
189 /* a single io, and all the tracking needed for it */
190 struct io_unit {
191     /* note, iocb must go first! */
192     struct iocb iocb;
193 
194     /* pointer to parent io operation struct */
195     struct io_oper *io_oper;
196 
197     /* aligned buffer */
198     char *buf;
199 
200     /* size of the aligned buffer (record size) */
201     int buf_size;
202 
203     /* state of this io unit (free, pending, done) */
204     int busy;
205 
206     /* result of last operation */
207     long res;
208 
209     struct io_unit *next;
210 
211     struct timeval io_start_time;		/* time of io_submit */
212 };
213 
214 struct thread_info {
215     io_context_t io_ctx;
216     pthread_t tid;
217 
218     /* allocated array of io_unit structs */
219     struct io_unit *ios;
220 
221     /* list of io units available for io */
222     struct io_unit *free_ious;
223 
224     /* number of io units in the ios array */
225     int num_global_ios;
226 
227     /* number of io units in flight */
228     int num_global_pending;
229 
230     /* preallocated array of iocb pointers, only used in run_active */
231     struct iocb **iocbs;
232 
233     /* preallocated array of events */
234     struct io_event *events;
235 
236     /* size of the events array */
237     int num_global_events;
238 
239     /* latency stats for io_submit */
240     struct io_latency io_submit_latency;
241 
242     /* list of operations still in progress, and of those finished */
243     struct io_oper *active_opers;
244     struct io_oper *finished_opers;
245 
246     /* number of files this thread is doing io on */
247     int num_files;
248 
249     /* how much io this thread did in the last stage */
250     double stage_mb_trans;
251 
252     /* latency completion stats i/o time from io_submit until io_getevents */
253     struct io_latency io_completion_latency;
254 };
255 
256 /*
257  * return seconds between start_tv and stop_tv in double precision
258  */
time_since(struct timeval * start_tv,struct timeval * stop_tv)259 static double time_since(struct timeval *start_tv, struct timeval *stop_tv)
260 {
261     double sec, usec;
262     double ret;
263     sec = stop_tv->tv_sec - start_tv->tv_sec;
264     usec = stop_tv->tv_usec - start_tv->tv_usec;
265     if (sec > 0 && usec < 0) {
266         sec--;
267 	usec += 1000000;
268     }
269     ret = sec + usec / (double)1000000;
270     if (ret < 0)
271         ret = 0;
272     return ret;
273 }
274 
275 /*
276  * return seconds between start_tv and now in double precision
277  */
time_since_now(struct timeval * start_tv)278 static double time_since_now(struct timeval *start_tv)
279 {
280     struct timeval stop_time;
281     gettimeofday(&stop_time, NULL);
282     return time_since(start_tv, &stop_time);
283 }
284 
285 /*
286  * Add latency info to latency struct
287  */
calc_latency(struct timeval * start_tv,struct timeval * stop_tv,struct io_latency * lat)288 static void calc_latency(struct timeval *start_tv, struct timeval *stop_tv,
289 			struct io_latency *lat)
290 {
291     double delta;
292     int i;
293     delta = time_since(start_tv, stop_tv);
294     delta = delta * 1000;
295 
296     if (delta > lat->max)
297     	lat->max = delta;
298     if (!lat->min || delta < lat->min)
299     	lat->min = delta;
300     lat->total_io++;
301     lat->total_lat += delta;
302     for (i = 0 ; i < DEVIATIONS ; i++) {
303         if (delta < deviations[i]) {
304 	    lat->deviations[i]++;
305 	    break;
306 	}
307     }
308 }
309 
oper_list_add(struct io_oper * oper,struct io_oper ** list)310 static void oper_list_add(struct io_oper *oper, struct io_oper **list)
311 {
312     if (!*list) {
313         *list = oper;
314 	oper->prev = oper->next = oper;
315 	return;
316     }
317     oper->prev = (*list)->prev;
318     oper->next = *list;
319     (*list)->prev->next = oper;
320     (*list)->prev = oper;
321     return;
322 }
323 
oper_list_del(struct io_oper * oper,struct io_oper ** list)324 static void oper_list_del(struct io_oper *oper, struct io_oper **list)
325 {
326     if ((*list)->next == (*list)->prev && *list == (*list)->next) {
327         *list = NULL;
328 	return;
329     }
330     oper->prev->next = oper->next;
331     oper->next->prev = oper->prev;
332     if (*list == oper)
333         *list = oper->next;
334 }
335 
336 /* worker func to check error fields in the io unit */
check_finished_io(struct io_unit * io)337 static int check_finished_io(struct io_unit *io) {
338     int i;
339     if (io->res != io->buf_size) {
340 
341   		 struct stat s;
342   		 fstat(io->io_oper->fd, &s);
343 
344   		 /*
345   		  * If file size is large enough for the read, then this short
346   		  * read is an error.
347   		  */
348   		 if ((io->io_oper->rw == READ || io->io_oper->rw == RREAD) &&
349   		     s.st_size > (io->iocb.u.c.offset + io->res)) {
350 
351   		 		 fprintf(stderr, "io err %lu (%s) op %d, off %Lu size %d\n",
352   		 		 		 io->res, strerror(-io->res), io->iocb.aio_lio_opcode,
353   		 		 		 io->iocb.u.c.offset, io->buf_size);
354   		 		 io->io_oper->last_err = io->res;
355   		 		 io->io_oper->num_err++;
356   		 		 return -1;
357   		 }
358     }
359     if (verify && io->io_oper->rw == READ) {
360         if (memcmp(io->buf, verify_buf, io->io_oper->reclen)) {
361 	    fprintf(stderr, "verify error, file %s offset %Lu contents (offset:bad:good):\n",
362 	            io->io_oper->file_name, io->iocb.u.c.offset);
363 
364 	    for (i = 0 ; i < io->io_oper->reclen ; i++) {
365 	        if (io->buf[i] != verify_buf[i]) {
366 		    fprintf(stderr, "%d:%c:%c ", i, io->buf[i], verify_buf[i]);
367 		}
368 	    }
369 	    fprintf(stderr, "\n");
370 	}
371 
372     }
373     return 0;
374 }
375 
376 /* worker func to check the busy bits and get an io unit ready for use */
grab_iou(struct io_unit * io,struct io_oper * oper)377 static int grab_iou(struct io_unit *io, struct io_oper *oper) {
378     if (io->busy == IO_PENDING)
379         return -1;
380 
381     io->busy = IO_PENDING;
382     io->res = 0;
383     io->io_oper = oper;
384     return 0;
385 }
386 
stage_name(int rw)387 char *stage_name(int rw) {
388     switch(rw) {
389     case WRITE:
390         return "write";
391     case READ:
392         return "read";
393     case RWRITE:
394         return "random write";
395     case RREAD:
396         return "random read";
397     }
398     return "unknown";
399 }
400 
oper_mb_trans(struct io_oper * oper)401 static inline double oper_mb_trans(struct io_oper *oper) {
402     return ((double)oper->started_ios * (double)oper->reclen) /
403                 (double)(1024 * 1024);
404 }
405 
print_time(struct io_oper * oper)406 static void print_time(struct io_oper *oper) {
407     double runtime;
408     double tput;
409     double mb;
410 
411     runtime = time_since_now(&oper->start_time);
412     mb = oper_mb_trans(oper);
413     tput = mb / runtime;
414     fprintf(stderr, "%s on %s (%.2f MB/s) %.2f MB in %.2fs\n",
415 	    stage_name(oper->rw), oper->file_name, tput, mb, runtime);
416 }
417 
print_lat(char * str,struct io_latency * lat)418 static void print_lat(char *str, struct io_latency *lat) {
419     double avg = lat->total_lat / lat->total_io;
420     int i;
421     double total_counted = 0;
422     fprintf(stderr, "%s min %.2f avg %.2f max %.2f\n\t",
423             str, lat->min, avg, lat->max);
424 
425     for (i = 0 ; i < DEVIATIONS ; i++) {
426 	fprintf(stderr, " %.0f < %d", lat->deviations[i], deviations[i]);
427 	total_counted += lat->deviations[i];
428     }
429     if (total_counted && lat->total_io - total_counted)
430         fprintf(stderr, " < %.0f", lat->total_io - total_counted);
431     fprintf(stderr, "\n");
432     memset(lat, 0, sizeof(*lat));
433 }
434 
print_latency(struct thread_info * t)435 static void print_latency(struct thread_info *t)
436 {
437     struct io_latency *lat = &t->io_submit_latency;
438     print_lat("latency", lat);
439 }
440 
print_completion_latency(struct thread_info * t)441 static void print_completion_latency(struct thread_info *t)
442 {
443     struct io_latency *lat = &t->io_completion_latency;
444     print_lat("completion latency", lat);
445 }
446 
447 /*
448  * updates the fields in the io operation struct that belongs to this
449  * io unit, and make the io unit reusable again
450  */
finish_io(struct thread_info * t,struct io_unit * io,long result,struct timeval * tv_now)451 void finish_io(struct thread_info *t, struct io_unit *io, long result,
452 		struct timeval *tv_now) {
453     struct io_oper *oper = io->io_oper;
454 
455     calc_latency(&io->io_start_time, tv_now, &t->io_completion_latency);
456     io->res = result;
457     io->busy = IO_FREE;
458     io->next = t->free_ious;
459     t->free_ious = io;
460     oper->num_pending--;
461     t->num_global_pending--;
462     check_finished_io(io);
463     if (oper->num_pending == 0 &&
464        (oper->started_ios == oper->total_ios || oper->stonewalled))
465     {
466         print_time(oper);
467     }
468 }
469 
read_some_events(struct thread_info * t)470 int read_some_events(struct thread_info *t) {
471     struct io_unit *event_io;
472     struct io_event *event;
473     int nr;
474     int i;
475     int min_nr = io_iter;
476     struct timeval stop_time;
477 
478     if (t->num_global_pending < io_iter)
479         min_nr = t->num_global_pending;
480 
481 #ifdef NEW_GETEVENTS
482     nr = io_getevents(t->io_ctx, min_nr, t->num_global_events, t->events,NULL);
483 #else
484     nr = io_getevents(t->io_ctx, t->num_global_events, t->events, NULL);
485 #endif
486     if (nr <= 0)
487         return nr;
488 
489     gettimeofday(&stop_time, NULL);
490     for (i = 0 ; i < nr ; i++) {
491 	event = t->events + i;
492 	event_io = (struct io_unit *)((unsigned long)event->obj);
493 	finish_io(t, event_io, event->res, &stop_time);
494     }
495     return nr;
496 }
497 
498 /*
499  * finds a free io unit, waiting for pending requests if required.  returns
500  * null if none could be found
501  */
find_iou(struct thread_info * t,struct io_oper * oper)502 static struct io_unit *find_iou(struct thread_info *t, struct io_oper *oper)
503 {
504     struct io_unit *event_io;
505     int nr;
506 
507 retry:
508     if (t->free_ious) {
509         event_io = t->free_ious;
510 	t->free_ious = t->free_ious->next;
511 	if (grab_iou(event_io, oper)) {
512 	    fprintf(stderr, "io unit on free list but not free\n");
513 	    abort();
514 	}
515 	return event_io;
516     }
517     nr = read_some_events(t);
518     if (nr > 0)
519     	goto retry;
520     else
521     	fprintf(stderr, "no free ious after read_some_events\n");
522     return NULL;
523 }
524 
525 /*
526  * wait for all pending requests for this io operation to finish
527  */
io_oper_wait(struct thread_info * t,struct io_oper * oper)528 static int io_oper_wait(struct thread_info *t, struct io_oper *oper) {
529     struct io_event event;
530     struct io_unit *event_io;
531 
532     if (oper == NULL) {
533         return 0;
534     }
535 
536     if (oper->num_pending == 0)
537         goto done;
538 
539     /* this func is not speed sensitive, no need to go wild reading
540      * more than one event at a time
541      */
542 #ifdef NEW_GETEVENTS
543     while(io_getevents(t->io_ctx, 1, 1, &event, NULL) > 0) {
544 #else
545     while(io_getevents(t->io_ctx, 1, &event, NULL) > 0) {
546 #endif
547 	struct timeval tv_now;
548         event_io = (struct io_unit *)((unsigned long)event.obj);
549 
550 	gettimeofday(&tv_now, NULL);
551 	finish_io(t, event_io, event.res, &tv_now);
552 
553 	if (oper->num_pending == 0)
554 	    break;
555     }
556 done:
557     if (oper->num_err) {
558         fprintf(stderr, "%u errors on oper, last %u\n",
559 	        oper->num_err, oper->last_err);
560     }
561     return 0;
562 }
563 
564 off_t random_byte_offset(struct io_oper *oper) {
565     off_t num;
566     off_t rand_byte = oper->start;
567     off_t range;
568     off_t offset = 1;
569 
570     range = (oper->end - oper->start) / (1024 * 1024);
571     if ((page_size_mask+1) > (1024 * 1024))
572         offset = (page_size_mask+1) / (1024 * 1024);
573     if (range < offset)
574         range = 0;
575     else
576         range -= offset;
577 
578     /* find a random mb offset */
579     num = 1 + (int)((double)range * rand() / (RAND_MAX + 1.0 ));
580     rand_byte += num * 1024 * 1024;
581 
582     /* find a random byte offset */
583     num = 1 + (int)((double)(1024 * 1024) * rand() / (RAND_MAX + 1.0));
584 
585     /* page align */
586     num = (num + page_size_mask) & ~page_size_mask;
587     rand_byte += num;
588 
589     if (rand_byte + oper->reclen > oper->end) {
590 	rand_byte -= oper->reclen;
591     }
592     return rand_byte;
593 }
594 
595 /*
596  * build an aio iocb for an operation, based on oper->rw and the
597  * last offset used.  This finds the struct io_unit that will be attached
598  * to the iocb, and things are ready for submission to aio after this
599  * is called.
600  *
601  * returns null on error
602  */
603 static struct io_unit *build_iocb(struct thread_info *t, struct io_oper *oper)
604 {
605     struct io_unit *io;
606     off_t rand_byte;
607 
608     io = find_iou(t, oper);
609     if (!io) {
610         fprintf(stderr, "unable to find io unit\n");
611 	return NULL;
612     }
613 
614     switch(oper->rw) {
615     case WRITE:
616         io_prep_pwrite(&io->iocb,oper->fd, io->buf, oper->reclen,
617 	               oper->last_offset);
618 	oper->last_offset += oper->reclen;
619 	break;
620     case READ:
621         io_prep_pread(&io->iocb,oper->fd, io->buf, oper->reclen,
622 	              oper->last_offset);
623 	oper->last_offset += oper->reclen;
624 	break;
625     case RREAD:
626 	rand_byte = random_byte_offset(oper);
627 	oper->last_offset = rand_byte;
628         io_prep_pread(&io->iocb,oper->fd, io->buf, oper->reclen,
629 	              rand_byte);
630         break;
631     case RWRITE:
632 	rand_byte = random_byte_offset(oper);
633 	oper->last_offset = rand_byte;
634         io_prep_pwrite(&io->iocb,oper->fd, io->buf, oper->reclen,
635 	              rand_byte);
636 
637         break;
638     }
639 
640     return io;
641 }
642 
643 /*
644  * wait for any pending requests, and then free all ram associated with
645  * an operation.  returns the last error the operation hit (zero means none)
646  */
647 static int
648 finish_oper(struct thread_info *t, struct io_oper *oper)
649 {
650     unsigned long last_err;
651 
652     io_oper_wait(t, oper);
653     last_err = oper->last_err;
654     if (oper->num_pending > 0) {
655         fprintf(stderr, "oper num_pending is %d\n", oper->num_pending);
656     }
657     close(oper->fd);
658     free(oper);
659     return last_err;
660 }
661 
662 /*
663  * allocates an io operation and fills in all the fields.  returns
664  * null on error
665  */
666 static struct io_oper *
667 create_oper(int fd, int rw, off_t start, off_t end, int reclen, int depth,
668             int iter, char *file_name)
669 {
670     struct io_oper *oper;
671 
672     oper = malloc (sizeof(*oper));
673     if (!oper) {
674 	fprintf(stderr, "unable to allocate io oper\n");
675 	return NULL;
676     }
677     memset(oper, 0, sizeof(*oper));
678 
679     oper->depth = depth;
680     oper->start = start;
681     oper->end = end;
682     oper->last_offset = oper->start;
683     oper->fd = fd;
684     oper->reclen = reclen;
685     oper->rw = rw;
686     oper->total_ios = (oper->end - oper->start) / oper->reclen;
687     oper->file_name = file_name;
688 
689     return oper;
690 }
691 
692 /*
693  * does setup on num_ios worth of iocbs, but does not actually
694  * start any io
695  */
696 int build_oper(struct thread_info *t, struct io_oper *oper, int num_ios,
697                struct iocb **my_iocbs)
698 {
699     int i;
700     struct io_unit *io;
701 
702     if (oper->started_ios == 0)
703 	gettimeofday(&oper->start_time, NULL);
704 
705     if (num_ios == 0)
706         num_ios = oper->total_ios;
707 
708     if ((oper->started_ios + num_ios) > oper->total_ios)
709         num_ios = oper->total_ios - oper->started_ios;
710 
711     for( i = 0 ; i < num_ios ; i++) {
712 	io = build_iocb(t, oper);
713 	if (!io) {
714 	    return -1;
715 	}
716 	my_iocbs[i] = &io->iocb;
717     }
718     return num_ios;
719 }
720 
721 /*
722  * runs through the iocbs in the array provided and updates
723  * counters in the associated oper struct
724  */
725 static void update_iou_counters(struct iocb **my_iocbs, int nr,
726 	struct timeval *tv_now)
727 {
728     struct io_unit *io;
729     int i;
730     for (i = 0 ; i < nr ; i++) {
731 	io = (struct io_unit *)(my_iocbs[i]);
732 	io->io_oper->num_pending++;
733 	io->io_oper->started_ios++;
734 	io->io_start_time = *tv_now;	/* set time of io_submit */
735     }
736 }
737 
738 /* starts some io for a given file, returns zero if all went well */
739 int run_built(struct thread_info *t, int num_ios, struct iocb **my_iocbs)
740 {
741     int ret;
742     struct timeval start_time;
743     struct timeval stop_time;
744 
745 resubmit:
746     gettimeofday(&start_time, NULL);
747     ret = io_submit(t->io_ctx, num_ios, my_iocbs);
748     gettimeofday(&stop_time, NULL);
749     calc_latency(&start_time, &stop_time, &t->io_submit_latency);
750 
751     if (ret != num_ios) {
752 	/* some ios got through */
753 	if (ret > 0) {
754 	    update_iou_counters(my_iocbs, ret, &stop_time);
755 	    my_iocbs += ret;
756 	    t->num_global_pending += ret;
757 	    num_ios -= ret;
758 	}
759 	/*
760 	 * we've used all the requests allocated in aio_init, wait and
761 	 * retry
762 	 */
763 	if (ret > 0 || ret == -EAGAIN) {
764 	    int old_ret = ret;
765 	    if ((ret = read_some_events(t) > 0)) {
766 		goto resubmit;
767 	    } else {
768 	    	fprintf(stderr, "ret was %d and now is %d\n", ret, old_ret);
769 		abort();
770 	    }
771 	}
772 
773 	fprintf(stderr, "ret %d (%s) on io_submit\n", ret, strerror(-ret));
774 	return -1;
775     }
776     update_iou_counters(my_iocbs, ret, &stop_time);
777     t->num_global_pending += ret;
778     return 0;
779 }
780 
781 /*
782  * changes oper->rw to the next in a command sequence, or returns zero
783  * to say this operation is really, completely done for
784  */
785 static int restart_oper(struct io_oper *oper) {
786     int new_rw  = 0;
787     if (oper->last_err)
788         return 0;
789 
790     /* this switch falls through */
791     switch(oper->rw) {
792     case WRITE:
793 	if (stages & (1 << READ))
794 	    new_rw = READ;
795     case READ:
796 	if (!new_rw && stages & (1 << RWRITE))
797 	    new_rw = RWRITE;
798     case RWRITE:
799 	if (!new_rw && stages & (1 << RREAD))
800 	    new_rw = RREAD;
801     }
802 
803     if (new_rw) {
804 	oper->started_ios = 0;
805 	oper->last_offset = oper->start;
806 	oper->stonewalled = 0;
807 
808 	/*
809 	 * we're restarting an operation with pending requests, so the
810 	 * timing info won't be printed by finish_io.  Printing it here
811 	 */
812 	if (oper->num_pending)
813 	    print_time(oper);
814 
815 	oper->rw = new_rw;
816 	return 1;
817     }
818     return 0;
819 }
820 
821 static int oper_runnable(struct io_oper *oper) {
822     struct stat buf;
823     int ret;
824 
825     /* first context is always runnable, if started_ios > 0, no need to
826      * redo the calculations
827      */
828     if (oper->started_ios || oper->start == 0)
829         return 1;
830     /*
831      * only the sequential phases force delays in starting */
832     if (oper->rw >= RWRITE)
833         return 1;
834     ret = fstat(oper->fd, &buf);
835     if (ret < 0) {
836         perror("fstat");
837 	exit(1);
838     }
839     if (S_ISREG(buf.st_mode) && buf.st_size < oper->start)
840         return 0;
841     return 1;
842 }
843 
844 /*
845  * runs through all the io operations on the active list, and starts
846  * a chunk of io on each.  If any io operations are completely finished,
847  * it either switches them to the next stage or puts them on the
848  * finished list.
849  *
850  * this function stops after max_io_submit iocbs are sent down the
851  * pipe, even if it has not yet touched all the operations on the
852  * active list.  Any operations that have finished are moved onto
853  * the finished_opers list.
854  */
855 static int run_active_list(struct thread_info *t,
856 			 int io_iter,
857 			 int max_io_submit)
858 {
859     struct io_oper *oper;
860     struct io_oper *built_opers = NULL;
861     struct iocb **my_iocbs = t->iocbs;
862     int ret = 0;
863     int num_built = 0;
864 
865     oper = t->active_opers;
866     while(oper) {
867 	if (!oper_runnable(oper)) {
868 	    oper = oper->next;
869 	    if (oper == t->active_opers)
870 	        break;
871 	    continue;
872 	}
873 	ret = build_oper(t, oper, io_iter, my_iocbs);
874 	if (ret >= 0) {
875 	    my_iocbs += ret;
876 	    num_built += ret;
877 	    oper_list_del(oper, &t->active_opers);
878 	    oper_list_add(oper, &built_opers);
879 	    oper = t->active_opers;
880 	    if (num_built + io_iter > max_io_submit)
881 	        break;
882 	} else
883 	    break;
884     }
885     if (num_built) {
886 	ret = run_built(t, num_built, t->iocbs);
887 	if (ret < 0) {
888 	    fprintf(stderr, "error %d on run_built\n", ret);
889 	    exit(1);
890 	}
891 	while(built_opers) {
892 	    oper = built_opers;
893 	    oper_list_del(oper, &built_opers);
894 	    oper_list_add(oper, &t->active_opers);
895 	    if (oper->started_ios == oper->total_ios) {
896 		oper_list_del(oper, &t->active_opers);
897 		oper_list_add(oper, &t->finished_opers);
898 	    }
899 	}
900     }
901     return 0;
902 }
903 
904 void drop_shm() {
905     int ret;
906     struct shmid_ds ds;
907     if (use_shm != USE_SHM)
908         return;
909 
910     ret = shmctl(shm_id, IPC_RMID, &ds);
911     if (ret) {
912         perror("shmctl IPC_RMID");
913     }
914 }
915 
916 void aio_setup(io_context_t *io_ctx, int n)
917 {
918     int res = io_queue_init(n, io_ctx);
919     if (res != 0) {
920 	fprintf(stderr, "io_queue_setup(%d) returned %d (%s)\n",
921 		n, res, strerror(-res));
922 	exit(3);
923     }
924 }
925 
926 /*
927  * allocate io operation and event arrays for a given thread
928  */
929 int setup_ious(struct thread_info *t,
930               int num_files, int depth,
931 	      int reclen, int max_io_submit) {
932     int i;
933     size_t bytes = num_files * depth * sizeof(*t->ios);
934 
935     t->ios = malloc(bytes);
936     if (!t->ios) {
937 	fprintf(stderr, "unable to allocate io units\n");
938 	return -1;
939     }
940     memset(t->ios, 0, bytes);
941 
942     for (i = 0 ; i < depth * num_files; i++) {
943 	t->ios[i].buf = aligned_buffer;
944 	aligned_buffer += padded_reclen;
945 	t->ios[i].buf_size = reclen;
946 	if (verify)
947 	    memset(t->ios[i].buf, 'b', reclen);
948 	else
949 	    memset(t->ios[i].buf, 0, reclen);
950 	t->ios[i].next = t->free_ious;
951 	t->free_ious = t->ios + i;
952     }
953     if (verify) {
954         verify_buf = aligned_buffer;
955         memset(verify_buf, 'b', reclen);
956     }
957 
958     t->iocbs = malloc(sizeof(struct iocb *) * max_io_submit);
959     if (!t->iocbs) {
960         fprintf(stderr, "unable to allocate iocbs\n");
961 	goto free_buffers;
962     }
963 
964     memset(t->iocbs, 0, max_io_submit * sizeof(struct iocb *));
965 
966     t->events = malloc(sizeof(struct io_event) * depth * num_files);
967     if (!t->events) {
968         fprintf(stderr, "unable to allocate ram for events\n");
969 	goto free_buffers;
970     }
971     memset(t->events, 0, num_files * sizeof(struct io_event)*depth);
972 
973     t->num_global_ios = num_files * depth;
974     t->num_global_events = t->num_global_ios;
975     return 0;
976 
977 free_buffers:
978     if (t->ios)
979         free(t->ios);
980     if (t->iocbs)
981         free(t->iocbs);
982     if (t->events)
983         free(t->events);
984     return -1;
985 }
986 
987 /*
988  * The buffers used for file data are allocated as a single big
989  * malloc, and then each thread and operation takes a piece and uses
990  * that for file data.  This lets us do a large shm or bigpages alloc
991  * and without trying to find a special place in each thread to map the
992  * buffers to
993  */
994 int setup_shared_mem(int num_threads, int num_files, int depth,
995                      int reclen, int max_io_submit)
996 {
997     char *p = NULL;
998     size_t total_ram;
999 
1000     padded_reclen = (reclen + page_size_mask) / (page_size_mask+1);
1001     padded_reclen = padded_reclen * (page_size_mask+1);
1002     total_ram = num_files * depth * padded_reclen + num_threads;
1003     if (verify)
1004     	total_ram += padded_reclen;
1005 
1006     if (use_shm == USE_MALLOC) {
1007 	p = malloc(total_ram + page_size_mask);
1008     } else if (use_shm == USE_SHM) {
1009         shm_id = shmget(IPC_PRIVATE, total_ram, IPC_CREAT | 0700);
1010 	if (shm_id < 0) {
1011 	    perror("shmget");
1012 	    drop_shm();
1013 	    goto free_buffers;
1014 	}
1015 	p = shmat(shm_id, (char *)0x50000000, 0);
1016         if ((long)p == -1) {
1017 	    perror("shmat");
1018 	    goto free_buffers;
1019 	}
1020 	/* won't really be dropped until we shmdt */
1021 	drop_shm();
1022     } else if (use_shm == USE_SHMFS) {
1023         char mmap_name[16]; /* /dev/shm/ + null + XXXXXX */
1024 	int fd;
1025 
1026 	strcpy(mmap_name, "/dev/shm/XXXXXX");
1027 	fd = mkstemp(mmap_name);
1028         if (fd < 0) {
1029 	    perror("mkstemp");
1030 	    goto free_buffers;
1031 	}
1032 	unlink(mmap_name);
1033 	ftruncate(fd, total_ram);
1034 	shm_id = fd;
1035 	p = mmap((char *)0x50000000, total_ram,
1036 	         PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1037 
1038         if (p == MAP_FAILED) {
1039 	    perror("mmap");
1040 	    goto free_buffers;
1041 	}
1042     }
1043     if (!p) {
1044         fprintf(stderr, "unable to allocate buffers\n");
1045 	goto free_buffers;
1046     }
1047     unaligned_buffer = p;
1048     p = (char*)((intptr_t) (p + page_size_mask) & ~page_size_mask);
1049     aligned_buffer = p;
1050     return 0;
1051 
1052 free_buffers:
1053     drop_shm();
1054     if (unaligned_buffer)
1055         free(unaligned_buffer);
1056     return -1;
1057 }
1058 
1059 /*
1060  * runs through all the thread_info structs and calculates a combined
1061  * throughput
1062  */
1063 void global_thread_throughput(struct thread_info *t, char *this_stage) {
1064     int i;
1065     double runtime = time_since_now(&global_stage_start_time);
1066     double total_mb = 0;
1067     double min_trans = 0;
1068 
1069     for (i = 0 ; i < num_threads ; i++) {
1070         total_mb += global_thread_info[i].stage_mb_trans;
1071 	if (!min_trans || t->stage_mb_trans < min_trans)
1072 	    min_trans = t->stage_mb_trans;
1073     }
1074     if (total_mb) {
1075 	fprintf(stderr, "%s throughput (%.2f MB/s) ", this_stage,
1076 	        total_mb / runtime);
1077 	fprintf(stderr, "%.2f MB in %.2fs", total_mb, runtime);
1078         if (stonewall)
1079 	    fprintf(stderr, " min transfer %.2fMB", min_trans);
1080         fprintf(stderr, "\n");
1081     }
1082 }
1083 
1084 
1085 /* this is the meat of the state machine.  There is a list of
1086  * active operations structs, and as each one finishes the required
1087  * io it is moved to a list of finished operations.  Once they have
1088  * all finished whatever stage they were in, they are given the chance
1089  * to restart and pick a different stage (read/write/random read etc)
1090  *
1091  * various timings are printed in between the stages, along with
1092  * thread synchronization if there are more than one threads.
1093  */
1094 int worker(struct thread_info *t)
1095 {
1096     struct io_oper *oper;
1097     char *this_stage = NULL;
1098     struct timeval stage_time;
1099     int status = 0;
1100     int iteration = 0;
1101     int cnt;
1102 
1103     aio_setup(&t->io_ctx, 512);
1104 
1105 restart:
1106     if (num_threads > 1) {
1107         pthread_mutex_lock(&stage_mutex);
1108 	threads_starting++;
1109 	if (threads_starting == num_threads) {
1110 	    threads_ending = 0;
1111 	    gettimeofday(&global_stage_start_time, NULL);
1112 	    pthread_cond_broadcast(&stage_cond);
1113 	}
1114 	while (threads_starting != num_threads)
1115 	    pthread_cond_wait(&stage_cond, &stage_mutex);
1116         pthread_mutex_unlock(&stage_mutex);
1117     }
1118     if (t->active_opers) {
1119         this_stage = stage_name(t->active_opers->rw);
1120 	gettimeofday(&stage_time, NULL);
1121 	t->stage_mb_trans = 0;
1122     }
1123 
1124     cnt = 0;
1125     /* first we send everything through aio */
1126     while(t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
1127 	if (stonewall && threads_ending) {
1128 	    oper = t->active_opers;
1129 	    oper->stonewalled = 1;
1130 	    oper_list_del(oper, &t->active_opers);
1131 	    oper_list_add(oper, &t->finished_opers);
1132 	} else {
1133 	    run_active_list(t, io_iter,  max_io_submit);
1134         }
1135 	cnt++;
1136     }
1137     if (latency_stats)
1138         print_latency(t);
1139 
1140     if (completion_latency_stats)
1141 	print_completion_latency(t);
1142 
1143     /* then we wait for all the operations to finish */
1144     oper = t->finished_opers;
1145     do {
1146 	if (!oper)
1147 		break;
1148 	io_oper_wait(t, oper);
1149 	oper = oper->next;
1150     } while(oper != t->finished_opers);
1151 
1152     /* then we do an fsync to get the timing for any future operations
1153      * right, and check to see if any of these need to get restarted
1154      */
1155     oper = t->finished_opers;
1156     while(oper) {
1157 	if (fsync_stages)
1158             fsync(oper->fd);
1159 	t->stage_mb_trans += oper_mb_trans(oper);
1160 	if (restart_oper(oper)) {
1161 	    oper_list_del(oper, &t->finished_opers);
1162 	    oper_list_add(oper, &t->active_opers);
1163 	    oper = t->finished_opers;
1164 	    continue;
1165 	}
1166 	oper = oper->next;
1167 	if (oper == t->finished_opers)
1168 	    break;
1169     }
1170 
1171     if (t->stage_mb_trans && t->num_files > 0) {
1172         double seconds = time_since_now(&stage_time);
1173 	fprintf(stderr, "thread %d %s totals (%.2f MB/s) %.2f MB in %.2fs\n",
1174 	        t - global_thread_info, this_stage, t->stage_mb_trans/seconds,
1175 		t->stage_mb_trans, seconds);
1176     }
1177 
1178     if (num_threads > 1) {
1179 	pthread_mutex_lock(&stage_mutex);
1180 	threads_ending++;
1181 	if (threads_ending == num_threads) {
1182 	    threads_starting = 0;
1183 	    pthread_cond_broadcast(&stage_cond);
1184 	    global_thread_throughput(t, this_stage);
1185 	}
1186 	while(threads_ending != num_threads)
1187 	    pthread_cond_wait(&stage_cond, &stage_mutex);
1188 	pthread_mutex_unlock(&stage_mutex);
1189     }
1190 
1191     /* someone got restarted, go back to the beginning */
1192     if (t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
1193 	iteration++;
1194         goto restart;
1195     }
1196 
1197     /* finally, free all the ram */
1198     while(t->finished_opers) {
1199 	oper = t->finished_opers;
1200 	oper_list_del(oper, &t->finished_opers);
1201 	status = finish_oper(t, oper);
1202     }
1203 
1204     if (t->num_global_pending) {
1205         fprintf(stderr, "global num pending is %d\n", t->num_global_pending);
1206     }
1207     io_queue_release(t->io_ctx);
1208 
1209     return status;
1210 }
1211 
1212 typedef void * (*start_routine)(void *);
1213 int run_workers(struct thread_info *t, int num_threads)
1214 {
1215     int ret;
1216     int thread_ret;
1217     int i;
1218 
1219     for(i = 0 ; i < num_threads ; i++) {
1220         ret = pthread_create(&t[i].tid, NULL, (start_routine)worker, t + i);
1221 	if (ret) {
1222 	    perror("pthread_create");
1223 	    exit(1);
1224 	}
1225     }
1226     for(i = 0 ; i < num_threads ; i++) {
1227         ret = pthread_join(t[i].tid, (void *)&thread_ret);
1228         if (ret) {
1229 	    perror("pthread_join");
1230 	    exit(1);
1231 	}
1232     }
1233     return 0;
1234 }
1235 
1236 off_t parse_size(char *size_arg, off_t mult) {
1237     char c;
1238     int num;
1239     off_t ret;
1240     c = size_arg[strlen(size_arg) - 1];
1241     if (c > '9') {
1242         size_arg[strlen(size_arg) - 1] = '\0';
1243     }
1244     num = atoi(size_arg);
1245     switch(c) {
1246     case 'g':
1247     case 'G':
1248         mult = 1024 * 1024 * 1024;
1249 	break;
1250     case 'm':
1251     case 'M':
1252         mult = 1024 * 1024;
1253 	break;
1254     case 'k':
1255     case 'K':
1256         mult = 1024;
1257 	break;
1258     case 'b':
1259     case 'B':
1260         mult = 1;
1261 	break;
1262     }
1263     ret = mult * num;
1264     return ret;
1265 }
1266 
1267 void print_usage(void) {
1268     printf("usage: aio-stress [-s size] [-r size] [-a size] [-d num] [-b num]\n");
1269     printf("                  [-i num] [-t num] [-c num] [-C size] [-nxhOS ]\n");
1270     printf("                  file1 [file2 ...]\n");
1271     printf("\t-a size in KB at which to align buffers\n");
1272     printf("\t-b max number of iocbs to give io_submit at once\n");
1273     printf("\t-c number of io contexts per file\n");
1274     printf("\t-C offset between contexts, default 2MB\n");
1275     printf("\t-s size in MB of the test file(s), default 1024MB\n");
1276     printf("\t-r record size in KB used for each io, default 64KB\n");
1277     printf("\t-d number of pending aio requests for each file, default 64\n");
1278     printf("\t-i number of ios per file sent before switching\n\t   to the next file, default 8\n");
1279     printf("\t-I total number of ayncs IOs the program will run, default is run until Cntl-C\n");
1280     printf("\t-O Use O_DIRECT (not available in 2.4 kernels),\n");
1281     printf("\t-S Use O_SYNC for writes\n");
1282     printf("\t-o add an operation to the list: write=0, read=1,\n");
1283     printf("\t   random write=2, random read=3.\n");
1284     printf("\t   repeat -o to specify multiple ops: -o 0 -o 1 etc.\n");
1285     printf("\t-m shm use ipc shared memory for io buffers instead of malloc\n");
1286     printf("\t-m shmfs mmap a file in /dev/shm for io buffers\n");
1287     printf("\t-n no fsyncs between write stage and read stage\n");
1288     printf("\t-l print io_submit latencies after each stage\n");
1289     printf("\t-L print io completion latencies after each stage\n");
1290     printf("\t-t number of threads to run\n");
1291     printf("\t-u unlink files after completion\n");
1292     printf("\t-v verification of bytes written\n");
1293     printf("\t-x turn off thread stonewalling\n");
1294     printf("\t-h this message\n");
1295     printf("\n\t   the size options (-a -s and -r) allow modifiers -s 400{k,m,g}\n");
1296     printf("\t   translate to 400KB, 400MB and 400GB\n");
1297     printf("version %s\n", PROG_VERSION);
1298 }
1299 
1300 int main(int ac, char **av)
1301 {
1302     int rwfd;
1303     int i;
1304     int j;
1305     int c;
1306 
1307     off_t file_size = 1 * 1024 * 1024 * 1024;
1308     int first_stage = WRITE;
1309     struct io_oper *oper;
1310     int status = 0;
1311     int num_files = 0;
1312     int open_fds = 0;
1313     struct thread_info *t;
1314 
1315     page_size_mask = getpagesize() - 1;
1316 
1317     while(1) {
1318 	c = getopt(ac, av, "a:b:c:C:m:s:r:d:i:I:o:t:lLnhOSxvu");
1319 	if  (c < 0)
1320 	    break;
1321 
1322         switch(c) {
1323 	case 'a':
1324 	    page_size_mask = parse_size(optarg, 1024);
1325 	    page_size_mask--;
1326 	    break;
1327 	case 'c':
1328 	    num_contexts = atoi(optarg);
1329 	    break;
1330 	case 'C':
1331 	    context_offset = parse_size(optarg, 1024 * 1024);
1332 	case 'b':
1333 	    max_io_submit = atoi(optarg);
1334 	    break;
1335 	case 's':
1336 	    file_size = parse_size(optarg, 1024 * 1024);
1337 	    break;
1338 	case 'd':
1339 	    depth = atoi(optarg);
1340 	    break;
1341 	case 'r':
1342 	    rec_len = parse_size(optarg, 1024);
1343 	    break;
1344 	case 'i':
1345 	    io_iter = atoi(optarg);
1346 	    break;
1347         case 'I':
1348           iterations = atoi(optarg);
1349         break;
1350 	case 'n':
1351 	    fsync_stages = 0;
1352 	    break;
1353 	case 'l':
1354 	    latency_stats = 1;
1355 	    break;
1356 	case 'L':
1357 	    completion_latency_stats = 1;
1358 	    break;
1359 	case 'm':
1360 	    if (!strcmp(optarg, "shm")) {
1361 		fprintf(stderr, "using ipc shm\n");
1362 	        use_shm = USE_SHM;
1363 	    } else if (!strcmp(optarg, "shmfs")) {
1364 	        fprintf(stderr, "using /dev/shm for buffers\n");
1365 		use_shm = USE_SHMFS;
1366 	    }
1367 	    break;
1368 	case 'o':
1369 	    i = atoi(optarg);
1370 	    stages |= 1 << i;
1371 	    fprintf(stderr, "adding stage %s\n", stage_name(i));
1372 	    break;
1373 	case 'O':
1374 	    o_direct = O_DIRECT;
1375 	    break;
1376 	case 'S':
1377 	    o_sync = O_SYNC;
1378 	    break;
1379 	case 't':
1380 	    num_threads = atoi(optarg);
1381 	    break;
1382 	case 'x':
1383 	    stonewall = 0;
1384 	    break;
1385 	case 'u':
1386 	    unlink_files = 1;
1387 	    break;
1388 	case 'v':
1389 	    verify = 1;
1390 	    break;
1391 	case 'h':
1392 	default:
1393 	    print_usage();
1394 	    exit(1);
1395 	}
1396     }
1397 
1398     /*
1399      * make sure we don't try to submit more ios than we have allocated
1400      * memory for
1401      */
1402     if (depth < io_iter) {
1403 	io_iter = depth;
1404         fprintf(stderr, "dropping io_iter to %d\n", io_iter);
1405     }
1406 
1407     if (optind >= ac) {
1408 	print_usage();
1409 	exit(1);
1410     }
1411 
1412     num_files = ac - optind;
1413 
1414     if (num_threads > (num_files * num_contexts)) {
1415         num_threads = num_files * num_contexts;
1416 	fprintf(stderr, "dropping thread count to the number of contexts %d\n",
1417 	        num_threads);
1418     }
1419 
1420     t = malloc(num_threads * sizeof(*t));
1421     if (!t) {
1422         perror("malloc");
1423 	exit(1);
1424     }
1425     global_thread_info = t;
1426 
1427     /* by default, allow a huge number of iocbs to be sent towards
1428      * io_submit
1429      */
1430     if (!max_io_submit)
1431         max_io_submit = num_files * io_iter * num_contexts;
1432 
1433     /*
1434      * make sure we don't try to submit more ios than max_io_submit allows
1435      */
1436     if (max_io_submit < io_iter) {
1437         io_iter = max_io_submit;
1438 	fprintf(stderr, "dropping io_iter to %d\n", io_iter);
1439     }
1440 
1441     if (!stages) {
1442         stages = (1 << WRITE) | (1 << READ) | (1 << RREAD) | (1 << RWRITE);
1443     } else {
1444         for (i = 0 ; i < LAST_STAGE; i++) {
1445 	    if (stages & (1 << i)) {
1446 	        first_stage = i;
1447 		fprintf(stderr, "starting with %s\n", stage_name(i));
1448 		break;
1449 	    }
1450 	}
1451     }
1452 
1453     if (file_size < num_contexts * context_offset) {
1454         fprintf(stderr, "file size %Lu too small for %d contexts\n",
1455 	        file_size, num_contexts);
1456 	exit(1);
1457     }
1458 
1459     fprintf(stderr, "file size %LuMB, record size %luKB, depth %d, ios per iteration %d\n", file_size / (1024 * 1024), rec_len / 1024, depth, io_iter);
1460     fprintf(stderr, "max io_submit %d, buffer alignment set to %luKB\n",
1461             max_io_submit, (page_size_mask + 1)/1024);
1462     fprintf(stderr, "threads %d files %d contexts %d context offset %LuMB verification %s\n",
1463             num_threads, num_files, num_contexts,
1464 	    context_offset / (1024 * 1024), verify ? "on" : "off");
1465     /* open all the files and do any required setup for them */
1466     for (i = optind ; i < ac ; i++) {
1467 	int thread_index;
1468 	for (j = 0 ; j < num_contexts ; j++) {
1469 	    thread_index = open_fds % num_threads;
1470 	    open_fds++;
1471 
1472 	    rwfd = open(av[i], O_CREAT | O_RDWR | o_direct | o_sync, 0600);
1473 	    assert(rwfd != -1);
1474 
1475 	    oper = create_oper(rwfd, first_stage, j * context_offset,
1476 	                       file_size - j * context_offset, rec_len,
1477 			       depth, io_iter, av[i]);
1478 	    if (!oper) {
1479 		fprintf(stderr, "error in create_oper\n");
1480 		exit(-1);
1481 	    }
1482 	    oper_list_add(oper, &t[thread_index].active_opers);
1483 	    t[thread_index].num_files++;
1484 	}
1485     }
1486     if (setup_shared_mem(num_threads, num_files * num_contexts,
1487                          depth, rec_len, max_io_submit))
1488     {
1489         exit(1);
1490     }
1491     for (i = 0 ; i < num_threads ; i++) {
1492 	if (setup_ious(&t[i], t[i].num_files, depth, rec_len, max_io_submit))
1493 		exit(1);
1494     }
1495     if (num_threads > 1){
1496         printf("Running multi thread version num_threads:%d\n", num_threads);
1497         run_workers(t, num_threads);
1498     } else {
1499         printf("Running single thread version \n");
1500 	status = worker(t);
1501     }
1502     if (unlink_files) {
1503 	for (i = optind ; i < ac ; i++) {
1504 	    printf("Cleaning up file %s \n", av[i]);
1505 	    unlink(av[i]);
1506 	}
1507     }
1508 
1509     if (status) {
1510 	exit(1);
1511     }
1512     return status;
1513 }
1514 
1515