2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
33 #include <sys/types.h>
38 #include <sys/ioctl.h>
46 #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
49 int thread_number = 0;
50 static char run_str[MAX_JOBS + 1];
52 static LIST_HEAD(disk_list);
53 static struct itimerval itimer;
54 static struct timeval genesis;
56 static void update_io_ticks(void);
57 static void disk_util_timer_arm(void);
58 static void print_thread_status(void);
60 extern unsigned long long mlock_size;
74 #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
76 static sem_t startup_sem;
78 #define TERMINATE_ALL (-1)
80 static void terminate_threads(int group_id)
84 for (i = 0; i < thread_number; i++) {
85 struct thread_data *td = &threads[i];
87 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
94 static void sig_handler(int sig)
99 disk_util_timer_arm();
100 print_thread_status();
103 printf("\nfio: terminating on signal\n");
105 terminate_threads(TERMINATE_ALL);
110 static unsigned long utime_since(struct timeval *s, struct timeval *e)
114 sec = e->tv_sec - s->tv_sec;
115 usec = e->tv_usec - s->tv_usec;
116 if (sec > 0 && usec < 0) {
121 sec *= (double) 1000000;
126 static unsigned long utime_since_now(struct timeval *s)
130 gettimeofday(&t, NULL);
131 return utime_since(s, &t);
134 static unsigned long mtime_since(struct timeval *s, struct timeval *e)
138 sec = e->tv_sec - s->tv_sec;
139 usec = e->tv_usec - s->tv_usec;
140 if (sec > 0 && usec < 0) {
145 sec *= (double) 1000;
146 usec /= (double) 1000;
151 static unsigned long mtime_since_now(struct timeval *s)
155 gettimeofday(&t, NULL);
156 return mtime_since(s, &t);
159 static inline unsigned long msec_now(struct timeval *s)
161 return s->tv_sec * 1000 + s->tv_usec / 1000;
164 static unsigned long time_since_now(struct timeval *s)
166 return mtime_since_now(s) / 1000;
169 static int random_map_free(struct thread_data *td, unsigned long long block)
171 unsigned int idx = RAND_MAP_IDX(td, block);
172 unsigned int bit = RAND_MAP_BIT(td, block);
174 return (td->file_map[idx] & (1UL << bit)) == 0;
177 static int get_next_free_block(struct thread_data *td, unsigned long long *b)
183 while ((*b) * td->min_bs < td->io_size) {
184 if (td->file_map[i] != -1UL) {
185 *b += ffz(td->file_map[i]);
189 *b += BLOCKS_PER_MAP;
196 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
198 unsigned long block = io_u->offset / td->min_bs;
199 unsigned int blocks = 0;
201 while (blocks < (io_u->buflen / td->min_bs)) {
202 unsigned int idx, bit;
204 if (!random_map_free(td, block))
207 idx = RAND_MAP_IDX(td, block);
208 bit = RAND_MAP_BIT(td, block);
210 assert(idx < td->num_maps);
212 td->file_map[idx] |= (1UL << bit);
217 if ((blocks * td->min_bs) < io_u->buflen)
218 io_u->buflen = blocks * td->min_bs;
221 static inline void add_stat_sample(struct io_stat *is, unsigned long val)
223 if (val > is->max_val)
225 if (val < is->min_val)
229 is->val_sq += val * val;
233 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
234 unsigned long val, int ddir)
236 if (iolog->nr_samples == iolog->max_samples) {
237 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
239 iolog->log = realloc(iolog->log, new_size);
240 iolog->max_samples <<= 1;
243 iolog->log[iolog->nr_samples].val = val;
244 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
245 iolog->log[iolog->nr_samples].ddir = ddir;
249 static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
251 add_stat_sample(&td->clat_stat[ddir], msec);
254 add_log_sample(td, td->clat_log, msec, ddir);
257 static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
259 add_stat_sample(&td->slat_stat[ddir], msec);
262 add_log_sample(td, td->slat_log, msec, ddir);
265 static void add_bw_sample(struct thread_data *td, int ddir)
267 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
270 if (spent < td->bw_avg_time)
273 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
274 add_stat_sample(&td->bw_stat[ddir], rate);
277 add_log_sample(td, td->bw_log, rate, ddir);
279 gettimeofday(&td->stat_sample_time[ddir], NULL);
280 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
283 static int get_next_offset(struct thread_data *td, unsigned long long *offset)
285 unsigned long long b, rb;
288 if (!td->sequential) {
289 unsigned long max_blocks = td->io_size / td->min_bs;
293 lrand48_r(&td->random_state, &r);
294 b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
295 rb = b + (td->file_offset / td->min_bs);
297 } while (!random_map_free(td, rb) && loops);
300 if (get_next_free_block(td, &b))
304 b = td->last_pos / td->min_bs;
306 *offset = (b * td->min_bs) + td->file_offset;
307 if (*offset > td->real_file_size)
313 static unsigned int get_next_buflen(struct thread_data *td)
318 if (td->min_bs == td->max_bs)
321 lrand48_r(&td->bsrange_state, &r);
322 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
323 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
326 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
327 buflen = td->io_size - td->this_io_bytes[td->ddir];
333 * busy looping version for the last few usec
335 static void __usec_sleep(unsigned int usec)
337 struct timeval start;
339 gettimeofday(&start, NULL);
340 while (utime_since_now(&start) < usec)
344 static void usec_sleep(struct thread_data *td, unsigned long usec)
346 struct timespec req, rem;
348 req.tv_sec = usec / 1000000;
349 req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
357 rem.tv_sec = rem.tv_nsec = 0;
358 if (nanosleep(&req, &rem) < 0)
361 if ((rem.tv_sec + rem.tv_nsec) == 0)
364 req.tv_nsec = rem.tv_nsec;
365 req.tv_sec = rem.tv_sec;
367 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
368 } while (!td->terminate);
371 static void rate_throttle(struct thread_data *td, unsigned long time_spent,
374 unsigned long usec_cycle;
379 usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
381 if (time_spent < usec_cycle) {
382 unsigned long s = usec_cycle - time_spent;
384 td->rate_pending_usleep += s;
385 if (td->rate_pending_usleep >= 100000) {
386 usec_sleep(td, td->rate_pending_usleep);
387 td->rate_pending_usleep = 0;
390 long overtime = time_spent - usec_cycle;
392 td->rate_pending_usleep -= overtime;
396 static int check_min_rate(struct thread_data *td, struct timeval *now)
403 * allow a 2 second settle period in the beginning
405 if (mtime_since(&td->start, now) < 2000)
409 * if rate blocks is set, sample is running
411 if (td->rate_bytes) {
412 spent = mtime_since(&td->lastrate, now);
413 if (spent < td->ratecycle)
416 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
417 if (rate < td->ratemin) {
418 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
420 terminate_threads(td->groupid);
425 td->rate_bytes = td->this_io_bytes[ddir];
426 memcpy(&td->lastrate, now, sizeof(*now));
430 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
434 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
440 static void fill_random_bytes(struct thread_data *td,
441 unsigned char *p, unsigned int len)
447 drand48_r(&td->verify_state, &r);
450 * lrand48_r seems to be broken and only fill the bottom
451 * 32-bits, even on 64-bit archs with 64-bit longs
464 static void hexdump(void *buffer, int len)
466 unsigned char *p = buffer;
469 for (i = 0; i < len; i++)
470 printf("%02x", p[i]);
474 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
476 unsigned char *p = (unsigned char *) io_u->buf;
481 c = crc32(p, hdr->len - sizeof(*hdr));
482 ret = c != hdr->crc32;
485 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
486 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
492 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
494 unsigned char *p = (unsigned char *) io_u->buf;
495 struct md5_ctx md5_ctx;
498 memset(&md5_ctx, 0, sizeof(md5_ctx));
500 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
502 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
504 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
505 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
506 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
512 static int verify_io_u(struct io_u *io_u)
514 struct verify_header *hdr = (struct verify_header *) io_u->buf;
517 if (hdr->fio_magic != FIO_HDR_MAGIC)
520 if (hdr->verify_type == VERIFY_MD5)
521 ret = verify_io_u_md5(hdr, io_u);
522 else if (hdr->verify_type == VERIFY_CRC32)
523 ret = verify_io_u_crc32(hdr, io_u);
525 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
532 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
534 hdr->crc32 = crc32(p, len);
537 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
539 struct md5_ctx md5_ctx;
541 memset(&md5_ctx, 0, sizeof(md5_ctx));
542 md5_update(&md5_ctx, p, len);
543 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
546 unsigned int hweight32(unsigned int w)
548 unsigned int res = w - ((w >> 1) & 0x55555555);
550 res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
551 res = (res + (res >> 4)) & 0x0F0F0F0F;
552 res = res + (res >> 8);
554 return (res + (res >> 16)) & 0x000000FF;
557 unsigned long hweight64(unsigned long long w)
560 return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w);
561 #elif __WORDSIZE == 64
562 unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul);
564 v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul);
565 v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful;
569 return (v + (v >> 32)) & 0x00000000000000FFul;
571 #error __WORDSIZE not defined
575 static int get_rw_ddir(struct thread_data *td)
578 * perhaps cheasy, but use the hamming weight of the position
579 * as a randomizer for data direction.
582 return hweight64(td->last_pos) & 1;
583 else if (td_read(td))
590 * fill body of io_u->buf with random data and add a header with the
591 * (eg) sha1sum of that data.
593 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
595 unsigned char *p = (unsigned char *) io_u->buf;
596 struct verify_header hdr;
598 hdr.fio_magic = FIO_HDR_MAGIC;
599 hdr.len = io_u->buflen;
601 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
603 if (td->verify == VERIFY_MD5) {
604 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
605 hdr.verify_type = VERIFY_MD5;
607 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
608 hdr.verify_type = VERIFY_CRC32;
611 memcpy(io_u->buf, &hdr, sizeof(hdr));
614 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
616 if (td->io_prep && td->io_prep(td, io_u))
622 void put_io_u(struct thread_data *td, struct io_u *io_u)
624 list_del(&io_u->list);
625 list_add(&io_u->list, &td->io_u_freelist);
629 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
632 * If using an iolog, grab next piece if any available.
635 struct io_piece *ipo;
637 if (list_empty(&td->io_log_list))
640 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
641 list_del(&ipo->list);
642 io_u->offset = ipo->offset;
643 io_u->buflen = ipo->len;
644 io_u->ddir = ipo->ddir;
650 * No log, let the seq/rand engine retrieve the next position.
652 if (!get_next_offset(td, &io_u->offset)) {
653 io_u->buflen = get_next_buflen(td);
656 io_u->ddir = get_rw_ddir(td);
664 #define queue_full(td) (list_empty(&(td)->io_u_freelist))
666 struct io_u *__get_io_u(struct thread_data *td)
673 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
676 list_del(&io_u->list);
677 list_add(&io_u->list, &td->io_u_busylist);
682 static struct io_u *get_io_u(struct thread_data *td)
686 io_u = __get_io_u(td);
690 if (td->zone_bytes >= td->zone_size) {
692 td->last_pos += td->zone_skip;
695 if (fill_io_u(td, io_u)) {
700 if (io_u->buflen + io_u->offset > td->real_file_size)
701 io_u->buflen = td->real_file_size - io_u->offset;
708 if (!td->iolog && !td->sequential)
709 mark_random_map(td, io_u);
711 td->last_pos += io_u->buflen;
713 if (td->verify != VERIFY_NONE)
714 populate_io_u(td, io_u);
716 if (td_io_prep(td, io_u)) {
721 gettimeofday(&io_u->start_time, NULL);
725 static inline void td_set_runstate(struct thread_data *td, int runstate)
727 td->old_runstate = td->runstate;
728 td->runstate = runstate;
731 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
733 struct io_piece *ipo;
735 if (list_empty(&td->io_hist_list))
738 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
739 list_del(&ipo->list);
741 io_u->offset = ipo->offset;
742 io_u->buflen = ipo->len;
743 io_u->ddir = DDIR_READ;
748 static void prune_io_piece_log(struct thread_data *td)
750 struct io_piece *ipo;
752 while (!list_empty(&td->io_hist_list)) {
753 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
755 list_del(&ipo->list);
761 * log a succesful write, so we can unwind the log for verify
763 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
765 struct io_piece *ipo = malloc(sizeof(struct io_piece));
766 struct list_head *entry;
768 INIT_LIST_HEAD(&ipo->list);
769 ipo->offset = io_u->offset;
770 ipo->len = io_u->buflen;
773 * for random io where the writes extend the file, it will typically
774 * be laid out with the block scattered as written. it's faster to
775 * read them in in that order again, so don't sort
777 if (td->sequential || !td->overwrite) {
778 list_add_tail(&ipo->list, &td->io_hist_list);
783 * for random io, sort the list so verify will run faster
785 entry = &td->io_hist_list;
786 while ((entry = entry->prev) != &td->io_hist_list) {
787 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
789 if (__ipo->offset < ipo->offset)
793 list_add(&ipo->list, entry);
796 static int init_iolog(struct thread_data *td)
798 unsigned long long offset;
802 int rw, i, reads, writes;
807 f = fopen(td->iolog_file, "r");
809 perror("fopen iolog");
814 reads = writes = i = 0;
815 while ((p = fgets(str, 4096, f)) != NULL) {
816 struct io_piece *ipo;
818 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
819 fprintf(stderr, "bad iolog: %s\n", p);
824 else if (rw == DDIR_WRITE)
827 fprintf(stderr, "bad ddir: %d\n", rw);
831 ipo = malloc(sizeof(*ipo));
832 INIT_LIST_HEAD(&ipo->list);
833 ipo->offset = offset;
835 if (bytes > td->max_bs)
838 list_add_tail(&ipo->list, &td->io_log_list);
848 if (reads && !writes)
849 td->ddir = DDIR_READ;
850 else if (!reads && writes)
851 td->ddir = DDIR_READ;
858 static int sync_td(struct thread_data *td)
861 return td->io_sync(td);
866 static int io_u_getevents(struct thread_data *td, int min, int max,
869 return td->io_getevents(td, min, max, t);
872 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
874 gettimeofday(&io_u->issue_time, NULL);
876 return td->io_queue(td, io_u);
879 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
881 static void io_completed(struct thread_data *td, struct io_u *io_u,
882 struct io_completion_data *icd)
887 gettimeofday(&e, NULL);
890 unsigned int bytes = io_u->buflen - io_u->resid;
891 const int idx = io_u->ddir;
893 td->io_blocks[idx]++;
894 td->io_bytes[idx] += bytes;
895 td->zone_bytes += bytes;
896 td->this_io_bytes[idx] += bytes;
898 msec = mtime_since(&io_u->issue_time, &e);
900 add_clat_sample(td, idx, msec);
901 add_bw_sample(td, idx);
903 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
904 log_io_piece(td, io_u);
906 icd->bytes_done[idx] += bytes;
908 icd->error = io_u->error;
911 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
917 icd->bytes_done[0] = icd->bytes_done[1] = 0;
919 for (i = 0; i < icd->nr; i++) {
920 io_u = td->io_event(td, i);
922 io_completed(td, io_u, icd);
927 static void cleanup_pending_aio(struct thread_data *td)
929 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
930 struct list_head *entry, *n;
931 struct io_completion_data icd;
936 * get immediately available events, if any
938 r = io_u_getevents(td, 0, td->cur_depth, &ts);
941 ios_completed(td, &icd);
945 * now cancel remaining active events
948 list_for_each_safe(entry, n, &td->io_u_busylist) {
949 io_u = list_entry(entry, struct io_u, list);
951 r = td->io_cancel(td, io_u);
958 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
961 ios_completed(td, &icd);
966 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
968 struct io_u *v_io_u = *io_u;
972 ret = verify_io_u(v_io_u);
973 put_io_u(td, v_io_u);
980 static void do_verify(struct thread_data *td)
983 struct io_u *io_u, *v_io_u = NULL;
984 struct io_completion_data icd;
987 td_set_runstate(td, TD_VERIFYING);
993 gettimeofday(&t, NULL);
994 if (runtime_exceeded(td, &t))
997 io_u = __get_io_u(td);
1001 if (get_next_verify(td, io_u)) {
1006 if (td_io_prep(td, io_u)) {
1011 ret = io_u_queue(td, io_u);
1019 * we have one pending to verify, do that while
1020 * we are doing io on the next one
1022 if (do_io_u_verify(td, &v_io_u))
1025 ret = io_u_getevents(td, 1, 1, NULL);
1032 v_io_u = td->io_event(td, 0);
1035 io_completed(td, v_io_u, &icd);
1038 td_verror(td, icd.error);
1039 put_io_u(td, v_io_u);
1045 * if we can't submit more io, we need to verify now
1047 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1052 do_io_u_verify(td, &v_io_u);
1055 cleanup_pending_aio(td);
1057 td_set_runstate(td, TD_RUNNING);
1060 static void do_io(struct thread_data *td)
1062 struct io_completion_data icd;
1063 struct timeval s, e;
1066 while (td->this_io_bytes[td->ddir] < td->io_size) {
1067 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1068 struct timespec *timeout;
1069 int ret, min_evts = 0;
1075 io_u = get_io_u(td);
1079 memcpy(&s, &io_u->start_time, sizeof(s));
1081 ret = io_u_queue(td, io_u);
1088 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1090 if (td->cur_depth < td->iodepth) {
1098 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1106 ios_completed(td, &icd);
1108 td_verror(td, icd.error);
1113 * the rate is batched for now, it should work for batches
1114 * of completions except the very first one which may look
1117 gettimeofday(&e, NULL);
1118 usec = utime_since(&s, &e);
1120 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1122 if (check_min_rate(td, &e)) {
1123 td_verror(td, ENOMEM);
1127 if (runtime_exceeded(td, &e))
1131 usec_sleep(td, td->thinktime);
1133 if (should_fsync(td) && td->fsync_blocks &&
1134 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1139 cleanup_pending_aio(td);
1141 if (should_fsync(td) && td->end_fsync)
1145 static void cleanup_io(struct thread_data *td)
1151 static int init_io(struct thread_data *td)
1153 if (td->io_engine == FIO_SYNCIO)
1154 return fio_syncio_init(td);
1155 else if (td->io_engine == FIO_MMAPIO)
1156 return fio_mmapio_init(td);
1157 else if (td->io_engine == FIO_LIBAIO)
1158 return fio_libaio_init(td);
1159 else if (td->io_engine == FIO_POSIXAIO)
1160 return fio_posixaio_init(td);
1161 else if (td->io_engine == FIO_SGIO)
1162 return fio_sgio_init(td);
1163 else if (td->io_engine == FIO_SPLICEIO)
1164 return fio_spliceio_init(td);
1166 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1171 static void cleanup_io_u(struct thread_data *td)
1173 struct list_head *entry, *n;
1176 list_for_each_safe(entry, n, &td->io_u_freelist) {
1177 io_u = list_entry(entry, struct io_u, list);
1179 list_del(&io_u->list);
1183 if (td->mem_type == MEM_MALLOC)
1184 free(td->orig_buffer);
1185 else if (td->mem_type == MEM_SHM) {
1186 struct shmid_ds sbuf;
1188 shmdt(td->orig_buffer);
1189 shmctl(td->shm_id, IPC_RMID, &sbuf);
1190 } else if (td->mem_type == MEM_MMAP)
1191 munmap(td->orig_buffer, td->orig_buffer_size);
1193 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1195 td->orig_buffer = NULL;
1198 static int init_io_u(struct thread_data *td)
1204 if (td->io_engine & FIO_SYNCIO)
1207 max_units = td->iodepth;
1209 td->orig_buffer_size = td->max_bs * max_units + MASK;
1211 if (td->mem_type == MEM_MALLOC)
1212 td->orig_buffer = malloc(td->orig_buffer_size);
1213 else if (td->mem_type == MEM_SHM) {
1214 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1215 if (td->shm_id < 0) {
1216 td_verror(td, errno);
1221 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1222 if (td->orig_buffer == (void *) -1) {
1223 td_verror(td, errno);
1225 td->orig_buffer = NULL;
1228 } else if (td->mem_type == MEM_MMAP) {
1229 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1230 if (td->orig_buffer == MAP_FAILED) {
1231 td_verror(td, errno);
1233 td->orig_buffer = NULL;
1238 p = ALIGN(td->orig_buffer);
1239 for (i = 0; i < max_units; i++) {
1240 io_u = malloc(sizeof(*io_u));
1241 memset(io_u, 0, sizeof(*io_u));
1242 INIT_LIST_HEAD(&io_u->list);
1244 io_u->buf = p + td->max_bs * i;
1246 list_add(&io_u->list, &td->io_u_freelist);
1252 static int create_file(struct thread_data *td, unsigned long long size,
1255 unsigned long long left;
1261 * unless specifically asked for overwrite, let normal io extend it
1263 if (td_write(td) && !td->overwrite)
1267 fprintf(stderr, "Need size for create\n");
1268 td_verror(td, EINVAL);
1273 oflags = O_CREAT | O_TRUNC;
1274 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1277 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1280 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1282 td_verror(td, errno);
1286 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1287 td_verror(td, errno);
1291 td->io_size = td->file_size;
1292 b = malloc(td->max_bs);
1293 memset(b, 0, td->max_bs);
1296 while (left && !td->terminate) {
1301 r = write(td->fd, b, bs);
1303 if (r == (int) bs) {
1308 td_verror(td, errno);
1317 unlink(td->file_name);
1318 else if (td->create_fsync)
1327 static int file_size(struct thread_data *td)
1331 if (fstat(td->fd, &st) == -1) {
1332 td_verror(td, errno);
1336 td->real_file_size = st.st_size;
1338 if (!td->file_size || td->file_size > td->real_file_size)
1339 td->file_size = td->real_file_size;
1341 td->file_size -= td->file_offset;
1345 static int bdev_size(struct thread_data *td)
1347 unsigned long long bytes;
1350 r = blockdev_size(td->fd, &bytes);
1356 td->real_file_size = bytes;
1359 * no extend possibilities, so limit size to device size if too large
1361 if (!td->file_size || td->file_size > td->real_file_size)
1362 td->file_size = td->real_file_size;
1364 td->file_size -= td->file_offset;
1368 static int get_file_size(struct thread_data *td)
1372 if (td->filetype == FIO_TYPE_FILE)
1373 ret = file_size(td);
1374 else if (td->filetype == FIO_TYPE_BD)
1375 ret = bdev_size(td);
1377 td->real_file_size = -1;
1382 if (td->file_offset > td->real_file_size) {
1383 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1387 td->io_size = td->file_size;
1388 if (td->io_size == 0) {
1389 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1390 td_verror(td, EINVAL);
1395 td->zone_size = td->io_size;
1397 td->total_io_size = td->io_size * td->loops;
1401 static int setup_file_mmap(struct thread_data *td)
1406 flags = PROT_READ | PROT_WRITE;
1407 else if (td_write(td)) {
1410 if (td->verify != VERIFY_NONE)
1415 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1416 if (td->mmap == MAP_FAILED) {
1418 td_verror(td, errno);
1422 if (td->invalidate_cache) {
1423 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1424 td_verror(td, errno);
1429 if (td->sequential) {
1430 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1431 td_verror(td, errno);
1435 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1436 td_verror(td, errno);
1444 static int setup_file_plain(struct thread_data *td)
1446 if (td->invalidate_cache) {
1447 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1448 td_verror(td, errno);
1453 if (td->sequential) {
1454 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1455 td_verror(td, errno);
1459 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1460 td_verror(td, errno);
1468 static int setup_file(struct thread_data *td)
1473 if (stat(td->file_name, &st) == -1) {
1474 if (errno != ENOENT) {
1475 td_verror(td, errno);
1478 if (!td->create_file) {
1479 td_verror(td, ENOENT);
1482 if (create_file(td, td->file_size, 0))
1484 } else if (td->filetype == FIO_TYPE_FILE) {
1485 if (st.st_size < (off_t) td->file_size) {
1486 if (create_file(td, td->file_size - st.st_size, 1))
1494 if (td_write(td) || td_rw(td)) {
1495 if (td->filetype == FIO_TYPE_FILE) {
1506 td->fd = open(td->file_name, flags, 0600);
1508 if (td->filetype == FIO_TYPE_CHAR)
1513 td->fd = open(td->file_name, flags);
1517 td_verror(td, errno);
1521 if (get_file_size(td))
1524 if (td->io_engine != FIO_MMAPIO)
1525 return setup_file_plain(td);
1527 return setup_file_mmap(td);
1530 static int check_dev_match(dev_t dev, char *path)
1532 unsigned int major, minor;
1536 f = fopen(path, "r");
1538 perror("open path");
1542 p = fgets(line, sizeof(line), f);
1548 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1553 if (((major << 8) | minor) == dev) {
1562 static int find_block_dir(dev_t dev, char *path)
1573 while ((dir = readdir(D)) != NULL) {
1574 char full_path[256];
1576 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1578 if (!strcmp(dir->d_name, "device"))
1581 sprintf(full_path, "%s/%s", path, dir->d_name);
1583 if (!strcmp(dir->d_name, "dev")) {
1584 if (!check_dev_match(dev, full_path)) {
1590 if (stat(full_path, &st) == -1) {
1595 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1598 found = find_block_dir(dev, full_path);
1600 strcpy(path, full_path);
1609 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1616 f = fopen(du->path, "r");
1620 p = fgets(line, sizeof(line), f);
1626 if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
1635 static void update_io_tick_disk(struct disk_util *du)
1637 struct disk_util_stat __dus, *dus, *ldus;
1640 if (get_io_ticks(du, &__dus))
1644 ldus = &du->last_dus;
1646 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1647 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1648 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1649 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1650 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1651 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1652 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1653 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1654 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1655 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1657 gettimeofday(&t, NULL);
1658 du->msec += mtime_since(&du->time, &t);
1659 memcpy(&du->time, &t, sizeof(t));
1660 memcpy(ldus, &__dus, sizeof(__dus));
1663 static void update_io_ticks(void)
1665 struct list_head *entry;
1666 struct disk_util *du;
1668 list_for_each(entry, &disk_list) {
1669 du = list_entry(entry, struct disk_util, list);
1670 update_io_tick_disk(du);
1674 static int disk_util_exists(dev_t dev)
1676 struct list_head *entry;
1677 struct disk_util *du;
1679 list_for_each(entry, &disk_list) {
1680 du = list_entry(entry, struct disk_util, list);
1689 static void disk_util_add(dev_t dev, char *path)
1691 struct disk_util *du = malloc(sizeof(*du));
1693 memset(du, 0, sizeof(*du));
1694 INIT_LIST_HEAD(&du->list);
1695 sprintf(du->path, "%s/stat", path);
1696 du->name = strdup(basename(path));
1699 gettimeofday(&du->time, NULL);
1700 get_io_ticks(du, &du->last_dus);
1702 list_add_tail(&du->list, &disk_list);
1705 static void init_disk_util(struct thread_data *td)
1708 char foo[256], tmp[256];
1712 if (!td->do_disk_util)
1715 if (!stat(td->file_name, &st)) {
1716 if (S_ISBLK(st.st_mode))
1722 * must be a file, open "." in that path
1724 strcpy(foo, td->file_name);
1727 perror("disk util stat");
1734 if (disk_util_exists(dev))
1737 sprintf(foo, "/sys/block");
1738 if (!find_block_dir(dev, foo))
1742 * for md/dm, there's no queue dir. we already have the right place
1744 sprintf(tmp, "%s/stat", foo);
1745 if (stat(tmp, &st)) {
1747 * if this is inside a partition dir, jump back to parent
1749 sprintf(tmp, "%s/queue", foo);
1750 if (stat(tmp, &st)) {
1752 sprintf(tmp, "%s/queue", p);
1753 if (stat(tmp, &st)) {
1754 fprintf(stderr, "unknown sysfs layout\n");
1757 sprintf(foo, "%s", p);
1761 disk_util_add(dev, foo);
1764 static void disk_util_timer_arm(void)
1766 itimer.it_value.tv_sec = 0;
1767 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1768 setitimer(ITIMER_REAL, &itimer, NULL);
1771 static void clear_io_state(struct thread_data *td)
1773 if (td->io_engine == FIO_SYNCIO)
1774 lseek(td->fd, SEEK_SET, 0);
1777 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1778 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1782 memset(td->file_map, 0, td->num_maps * sizeof(long));
1785 static void update_rusage_stat(struct thread_data *td)
1787 if (!(td->runtime[0] + td->runtime[1]))
1790 getrusage(RUSAGE_SELF, &td->ru_end);
1792 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1793 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1794 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1797 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1800 static void *thread_main(void *data)
1802 struct thread_data *td = data;
1805 if (!td->use_thread)
1810 INIT_LIST_HEAD(&td->io_u_freelist);
1811 INIT_LIST_HEAD(&td->io_u_busylist);
1812 INIT_LIST_HEAD(&td->io_hist_list);
1813 INIT_LIST_HEAD(&td->io_log_list);
1818 if (fio_setaffinity(td) == -1) {
1819 td_verror(td, errno);
1830 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1831 td_verror(td, errno);
1836 sem_post(&startup_sem);
1837 sem_wait(&td->mutex);
1839 if (!td->create_serialize && setup_file(td))
1842 if (init_random_state(td))
1845 gettimeofday(&td->epoch, NULL);
1847 while (td->loops--) {
1848 getrusage(RUSAGE_SELF, &td->ru_start);
1849 gettimeofday(&td->start, NULL);
1850 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1853 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1856 prune_io_piece_log(td);
1860 td->runtime[td->ddir] += mtime_since_now(&td->start);
1861 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1862 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1864 update_rusage_stat(td);
1866 if (td->error || td->terminate)
1869 if (td->verify == VERIFY_NONE)
1873 gettimeofday(&td->start, NULL);
1877 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1879 if (td->error || td->terminate)
1886 finish_log(td, td->bw_log, "bw");
1888 finish_log(td, td->slat_log, "slat");
1890 finish_log(td, td->clat_log, "clat");
1892 if (exitall_on_terminate)
1893 terminate_threads(td->groupid);
1901 munmap(td->mmap, td->file_size);
1905 sem_post(&startup_sem);
1906 sem_wait(&td->mutex);
1908 td_set_runstate(td, TD_EXITED);
1913 static void *fork_main(int shmid, int offset)
1915 struct thread_data *td;
1918 data = shmat(shmid, NULL, 0);
1919 if (data == (void *) -1) {
1924 td = data + offset * sizeof(struct thread_data);
1930 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1931 double *mean, double *dev)
1935 if (is->samples == 0)
1941 n = (double) is->samples;
1942 *mean = (double) is->val / n;
1943 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1944 if (!(*min + *max) && !(*mean + *dev))
1950 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1953 char *ddir_str[] = { "read ", "write" };
1954 unsigned long min, max, bw;
1957 if (!td->runtime[ddir])
1960 bw = td->io_bytes[ddir] / td->runtime[ddir];
1961 printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1963 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1964 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1966 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1967 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1969 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1972 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1973 printf(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
1977 static void show_thread_status(struct thread_data *td,
1978 struct group_run_stats *rs)
1980 double usr_cpu, sys_cpu;
1982 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1985 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1987 show_ddir_status(td, rs, td->ddir);
1988 if (td->io_bytes[td->ddir ^ 1])
1989 show_ddir_status(td, rs, td->ddir ^ 1);
1991 if (td->runtime[0] + td->runtime[1]) {
1992 double runt = td->runtime[0] + td->runtime[1];
1994 usr_cpu = (double) td->usr_time * 100 / runt;
1995 sys_cpu = (double) td->sys_time * 100 / runt;
2001 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
2004 static void check_str_update(struct thread_data *td)
2006 char c = run_str[td->thread_number - 1];
2008 if (td->runstate == td->old_runstate)
2011 switch (td->runstate) {
2024 } else if (td_read(td)) {
2042 case TD_NOT_CREATED:
2046 printf("state %d\n", td->runstate);
2049 run_str[td->thread_number - 1] = c;
2050 td->old_runstate = td->runstate;
2053 static void eta_to_str(char *str, int eta_sec)
2055 unsigned int d, h, m, s;
2056 static int always_d, always_h;
2068 if (d || always_d) {
2070 str += sprintf(str, "%02dd:", d);
2072 if (h || always_h) {
2074 str += sprintf(str, "%02dh:", h);
2077 str += sprintf(str, "%02dm:", m);
2078 str += sprintf(str, "%02ds", s);
2081 static int thread_eta(struct thread_data *td, unsigned long elapsed)
2083 unsigned long long bytes_total, bytes_done;
2084 unsigned int eta_sec = 0;
2086 bytes_total = td->total_io_size;
2089 * if writing, bytes_total will be twice the size. If mixing,
2090 * assume a 50/50 split and thus bytes_total will be 50% larger.
2094 bytes_total = bytes_total * 3 / 2;
2098 if (td->zone_size && td->zone_skip)
2099 bytes_total /= (td->zone_skip / td->zone_size);
2101 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2104 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2105 perc = (double) bytes_done / (double) bytes_total;
2109 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2111 if (td->timeout && eta_sec > (td->timeout - elapsed))
2112 eta_sec = td->timeout - elapsed;
2113 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
2114 int t_eta = 0, r_eta = 0;
2117 * We can only guess - assume it'll run the full timeout
2118 * if given, otherwise assume it'll run at the specified rate.
2121 t_eta = td->timeout + td->start_delay - elapsed;
2123 r_eta = (bytes_total / 1024) / td->rate;
2124 r_eta += td->start_delay - elapsed;
2128 eta_sec = min(r_eta, t_eta);
2137 * thread is already done
2145 static void print_thread_status(void)
2147 unsigned long elapsed = time_since_now(&genesis);
2148 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
2152 eta_secs = malloc(thread_number * sizeof(int));
2153 memset(eta_secs, 0, thread_number * sizeof(int));
2155 nr_running = t_rate = m_rate = 0;
2156 for (i = 0; i < thread_number; i++) {
2157 struct thread_data *td = &threads[i];
2159 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2162 m_rate += td->ratemin;
2166 eta_secs[i] = thread_eta(td, elapsed);
2168 eta_secs[i] = INT_MAX;
2170 check_str_update(td);
2173 if (exitall_on_terminate)
2178 for (i = 0; i < thread_number; i++) {
2179 if (exitall_on_terminate) {
2180 if (eta_secs[i] < eta_sec)
2181 eta_sec = eta_secs[i];
2183 if (eta_secs[i] > eta_sec)
2184 eta_sec = eta_secs[i];
2188 if (eta_sec != INT_MAX && elapsed) {
2189 perc = (double) elapsed / (double) (elapsed + eta_sec);
2190 eta_to_str(eta_str, eta_sec);
2193 printf("Threads now running (%d)", nr_running);
2194 if (m_rate || t_rate)
2195 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2196 if (eta_sec != INT_MAX) {
2198 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2205 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2210 * reap exited threads (TD_EXITED -> TD_REAPED)
2212 for (i = 0; i < thread_number; i++) {
2213 struct thread_data *td = &threads[i];
2215 if (td->runstate != TD_EXITED)
2218 td_set_runstate(td, TD_REAPED);
2220 if (td->use_thread) {
2223 if (pthread_join(td->thread, (void *) &ret))
2224 perror("thread_join");
2226 waitpid(td->pid, NULL, 0);
2229 (*m_rate) -= td->ratemin;
2230 (*t_rate) -= td->rate;
2234 static void run_threads(void)
2236 struct thread_data *td;
2237 unsigned long spent;
2238 int i, todo, nr_running, m_rate, t_rate, nr_started;
2239 void *mlocked_mem = NULL;
2241 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2243 signal(SIGINT, sig_handler);
2244 signal(SIGALRM, sig_handler);
2247 mlocked_mem = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2249 perror("mmap locked mem");
2252 if (mlock(mlocked_mem, mlock_size) < 0) {
2253 munmap(mlocked_mem, mlock_size);
2260 todo = thread_number;
2263 m_rate = t_rate = 0;
2265 for (i = 0; i < thread_number; i++) {
2268 run_str[td->thread_number - 1] = 'P';
2272 if (!td->create_serialize)
2276 * do file setup here so it happens sequentially,
2277 * we don't want X number of threads getting their
2278 * client data interspersed on disk
2280 if (setup_file(td)) {
2281 td_set_runstate(td, TD_REAPED);
2286 gettimeofday(&genesis, NULL);
2290 * create threads (TD_NOT_CREATED -> TD_CREATED)
2292 for (i = 0; i < thread_number; i++) {
2295 if (td->runstate != TD_NOT_CREATED)
2299 * never got a chance to start, killed by other
2300 * thread for some reason
2302 if (td->terminate) {
2307 if (td->start_delay) {
2308 spent = mtime_since_now(&genesis);
2310 if (td->start_delay * 1000 > spent)
2314 if (td->stonewall && (nr_started || nr_running))
2317 td_set_runstate(td, TD_CREATED);
2318 sem_init(&startup_sem, 0, 1);
2322 if (td->use_thread) {
2323 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2324 perror("thread_create");
2329 sem_wait(&startup_sem);
2331 fork_main(shm_id, i);
2338 * start created threads (TD_CREATED -> TD_RUNNING)
2340 for (i = 0; i < thread_number; i++) {
2343 if (td->runstate != TD_CREATED)
2346 td_set_runstate(td, TD_RUNNING);
2349 m_rate += td->ratemin;
2351 sem_post(&td->mutex);
2354 reap_threads(&nr_running, &t_rate, &m_rate);
2360 while (nr_running) {
2361 reap_threads(&nr_running, &t_rate, &m_rate);
2368 if (munlock(mlocked_mem, mlock_size) < 0)
2370 munmap(mlocked_mem, mlock_size);
2374 static void show_group_stats(struct group_run_stats *rs, int id)
2376 printf("\nRun status group %d (all jobs):\n", id);
2378 if (rs->max_run[DDIR_READ])
2379 printf(" READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
2380 if (rs->max_run[DDIR_WRITE])
2381 printf(" WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
2384 static void show_disk_util(void)
2386 struct disk_util_stat *dus;
2387 struct list_head *entry;
2388 struct disk_util *du;
2391 printf("\nDisk stats (read/write):\n");
2393 list_for_each(entry, &disk_list) {
2394 du = list_entry(entry, struct disk_util, list);
2397 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2401 printf(" %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
2405 static void show_run_stats(void)
2407 struct group_run_stats *runstats, *rs;
2408 struct thread_data *td;
2411 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2413 for (i = 0; i < groupid + 1; i++) {
2416 memset(rs, 0, sizeof(*rs));
2417 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2418 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2421 for (i = 0; i < thread_number; i++) {
2422 unsigned long rbw, wbw;
2427 printf("Client%d: %s\n", td->thread_number, td->verror);
2431 rs = &runstats[td->groupid];
2433 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2434 rs->min_run[0] = td->runtime[0];
2435 if (td->runtime[0] > rs->max_run[0])
2436 rs->max_run[0] = td->runtime[0];
2437 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2438 rs->min_run[1] = td->runtime[1];
2439 if (td->runtime[1] > rs->max_run[1])
2440 rs->max_run[1] = td->runtime[1];
2444 rbw = td->io_bytes[0] / td->runtime[0];
2446 wbw = td->io_bytes[1] / td->runtime[1];
2448 if (rbw < rs->min_bw[0])
2449 rs->min_bw[0] = rbw;
2450 if (wbw < rs->min_bw[1])
2451 rs->min_bw[1] = wbw;
2452 if (rbw > rs->max_bw[0])
2453 rs->max_bw[0] = rbw;
2454 if (wbw > rs->max_bw[1])
2455 rs->max_bw[1] = wbw;
2457 rs->io_mb[0] += td->io_bytes[0] >> 20;
2458 rs->io_mb[1] += td->io_bytes[1] >> 20;
2461 for (i = 0; i < groupid + 1; i++) {
2465 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2467 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2471 * don't overwrite last signal output
2475 for (i = 0; i < thread_number; i++) {
2477 rs = &runstats[td->groupid];
2479 show_thread_status(td, rs);
2482 for (i = 0; i < groupid + 1; i++)
2483 show_group_stats(&runstats[i], i);
2488 int main(int argc, char *argv[])
2490 if (parse_options(argc, argv))
2493 if (!thread_number) {
2494 printf("Nothing to do\n");
2498 disk_util_timer_arm();