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;
75 #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
77 static sem_t startup_sem;
79 #define TERMINATE_ALL (-1)
80 #define JOB_START_TIMEOUT (5 * 1000)
82 static void terminate_threads(int group_id)
86 for (i = 0; i < thread_number; i++) {
87 struct thread_data *td = &threads[i];
89 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
96 static void sig_handler(int sig)
101 disk_util_timer_arm();
102 print_thread_status();
105 printf("\nfio: terminating on signal\n");
107 terminate_threads(TERMINATE_ALL);
112 static unsigned long utime_since(struct timeval *s, struct timeval *e)
116 sec = e->tv_sec - s->tv_sec;
117 usec = e->tv_usec - s->tv_usec;
118 if (sec > 0 && usec < 0) {
123 sec *= (double) 1000000;
128 static unsigned long utime_since_now(struct timeval *s)
132 gettimeofday(&t, NULL);
133 return utime_since(s, &t);
136 static unsigned long mtime_since(struct timeval *s, struct timeval *e)
140 sec = e->tv_sec - s->tv_sec;
141 usec = e->tv_usec - s->tv_usec;
142 if (sec > 0 && usec < 0) {
147 sec *= (double) 1000;
148 usec /= (double) 1000;
153 static unsigned long mtime_since_now(struct timeval *s)
157 gettimeofday(&t, NULL);
158 return mtime_since(s, &t);
161 static inline unsigned long msec_now(struct timeval *s)
163 return s->tv_sec * 1000 + s->tv_usec / 1000;
166 static unsigned long time_since_now(struct timeval *s)
168 return mtime_since_now(s) / 1000;
171 static int random_map_free(struct thread_data *td, unsigned long long block)
173 unsigned int idx = RAND_MAP_IDX(td, block);
174 unsigned int bit = RAND_MAP_BIT(td, block);
176 return (td->file_map[idx] & (1UL << bit)) == 0;
179 static int get_next_free_block(struct thread_data *td, unsigned long long *b)
185 while ((*b) * td->min_bs < td->io_size) {
186 if (td->file_map[i] != -1UL) {
187 *b += ffz(td->file_map[i]);
191 *b += BLOCKS_PER_MAP;
198 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
200 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
201 unsigned int blocks = 0;
203 while (blocks < (io_u->buflen / td->min_bs)) {
204 unsigned int idx, bit;
206 if (!random_map_free(td, block))
209 idx = RAND_MAP_IDX(td, block);
210 bit = RAND_MAP_BIT(td, block);
212 assert(idx < td->num_maps);
214 td->file_map[idx] |= (1UL << bit);
219 if ((blocks * td->min_bs) < io_u->buflen)
220 io_u->buflen = blocks * td->min_bs;
223 static inline void add_stat_sample(struct io_stat *is, unsigned long val)
225 if (val > is->max_val)
227 if (val < is->min_val)
231 is->val_sq += val * val;
235 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
236 unsigned long val, int ddir)
238 if (iolog->nr_samples == iolog->max_samples) {
239 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
241 iolog->log = realloc(iolog->log, new_size);
242 iolog->max_samples <<= 1;
245 iolog->log[iolog->nr_samples].val = val;
246 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
247 iolog->log[iolog->nr_samples].ddir = ddir;
251 static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
253 add_stat_sample(&td->clat_stat[ddir], msec);
256 add_log_sample(td, td->clat_log, msec, ddir);
259 static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
261 add_stat_sample(&td->slat_stat[ddir], msec);
264 add_log_sample(td, td->slat_log, msec, ddir);
267 static void add_bw_sample(struct thread_data *td, int ddir)
269 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
272 if (spent < td->bw_avg_time)
275 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
276 add_stat_sample(&td->bw_stat[ddir], rate);
279 add_log_sample(td, td->bw_log, rate, ddir);
281 gettimeofday(&td->stat_sample_time[ddir], NULL);
282 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
285 static int get_next_offset(struct thread_data *td, unsigned long long *offset)
287 unsigned long long b, rb;
290 if (!td->sequential) {
291 unsigned long long max_blocks = td->io_size / td->min_bs;
295 lrand48_r(&td->random_state, &r);
296 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
297 rb = b + (td->file_offset / td->min_bs);
299 } while (!random_map_free(td, rb) && loops);
302 if (get_next_free_block(td, &b))
306 b = td->last_pos / td->min_bs;
308 *offset = (b * td->min_bs) + td->file_offset;
309 if (*offset > td->real_file_size)
315 static unsigned int get_next_buflen(struct thread_data *td)
320 if (td->min_bs == td->max_bs)
323 lrand48_r(&td->bsrange_state, &r);
324 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
325 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
328 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
329 buflen = td->io_size - td->this_io_bytes[td->ddir];
335 * busy looping version for the last few usec
337 static void __usec_sleep(unsigned int usec)
339 struct timeval start;
341 gettimeofday(&start, NULL);
342 while (utime_since_now(&start) < usec)
346 static void usec_sleep(struct thread_data *td, unsigned long usec)
348 struct timespec req, rem;
350 req.tv_sec = usec / 1000000;
351 req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
359 rem.tv_sec = rem.tv_nsec = 0;
360 if (nanosleep(&req, &rem) < 0)
363 if ((rem.tv_sec + rem.tv_nsec) == 0)
366 req.tv_nsec = rem.tv_nsec;
367 req.tv_sec = rem.tv_sec;
369 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
370 } while (!td->terminate);
373 static void rate_throttle(struct thread_data *td, unsigned long time_spent,
376 unsigned long usec_cycle;
381 usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
383 if (time_spent < usec_cycle) {
384 unsigned long s = usec_cycle - time_spent;
386 td->rate_pending_usleep += s;
387 if (td->rate_pending_usleep >= 100000) {
388 usec_sleep(td, td->rate_pending_usleep);
389 td->rate_pending_usleep = 0;
392 long overtime = time_spent - usec_cycle;
394 td->rate_pending_usleep -= overtime;
398 static int check_min_rate(struct thread_data *td, struct timeval *now)
405 * allow a 2 second settle period in the beginning
407 if (mtime_since(&td->start, now) < 2000)
411 * if rate blocks is set, sample is running
413 if (td->rate_bytes) {
414 spent = mtime_since(&td->lastrate, now);
415 if (spent < td->ratecycle)
418 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
419 if (rate < td->ratemin) {
420 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
422 terminate_threads(td->groupid);
427 td->rate_bytes = td->this_io_bytes[ddir];
428 memcpy(&td->lastrate, now, sizeof(*now));
432 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
436 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
442 static void fill_random_bytes(struct thread_data *td,
443 unsigned char *p, unsigned int len)
449 drand48_r(&td->verify_state, &r);
452 * lrand48_r seems to be broken and only fill the bottom
453 * 32-bits, even on 64-bit archs with 64-bit longs
466 static void hexdump(void *buffer, int len)
468 unsigned char *p = buffer;
471 for (i = 0; i < len; i++)
472 printf("%02x", p[i]);
476 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
478 unsigned char *p = (unsigned char *) io_u->buf;
483 c = crc32(p, hdr->len - sizeof(*hdr));
484 ret = c != hdr->crc32;
487 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
488 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
494 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
496 unsigned char *p = (unsigned char *) io_u->buf;
497 struct md5_ctx md5_ctx;
500 memset(&md5_ctx, 0, sizeof(md5_ctx));
502 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
504 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
506 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
507 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
508 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
514 static int verify_io_u(struct io_u *io_u)
516 struct verify_header *hdr = (struct verify_header *) io_u->buf;
519 if (hdr->fio_magic != FIO_HDR_MAGIC)
522 if (hdr->verify_type == VERIFY_MD5)
523 ret = verify_io_u_md5(hdr, io_u);
524 else if (hdr->verify_type == VERIFY_CRC32)
525 ret = verify_io_u_crc32(hdr, io_u);
527 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
534 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
536 hdr->crc32 = crc32(p, len);
539 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
541 struct md5_ctx md5_ctx;
543 memset(&md5_ctx, 0, sizeof(md5_ctx));
544 md5_update(&md5_ctx, p, len);
545 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
548 static int get_rw_ddir(struct thread_data *td)
552 unsigned long elapsed;
554 gettimeofday(&now, NULL);
555 elapsed = mtime_since_now(&td->rwmix_switch);
558 * Check if it's time to seed a new data direction.
560 if (elapsed >= td->rwmixcycle) {
564 lrand48_r(&td->random_state, &r);
565 v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
566 if (v < td->rwmixread)
567 td->rwmix_ddir = DDIR_READ;
569 td->rwmix_ddir = DDIR_WRITE;
570 memcpy(&td->rwmix_switch, &now, sizeof(now));
572 return td->rwmix_ddir;
573 } else if (td_read(td))
580 * fill body of io_u->buf with random data and add a header with the
581 * (eg) sha1sum of that data.
583 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
585 unsigned char *p = (unsigned char *) io_u->buf;
586 struct verify_header hdr;
588 hdr.fio_magic = FIO_HDR_MAGIC;
589 hdr.len = io_u->buflen;
591 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
593 if (td->verify == VERIFY_MD5) {
594 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
595 hdr.verify_type = VERIFY_MD5;
597 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
598 hdr.verify_type = VERIFY_CRC32;
601 memcpy(io_u->buf, &hdr, sizeof(hdr));
604 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
606 if (td->io_prep && td->io_prep(td, io_u))
612 void put_io_u(struct thread_data *td, struct io_u *io_u)
614 list_del(&io_u->list);
615 list_add(&io_u->list, &td->io_u_freelist);
619 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
622 * If using an iolog, grab next piece if any available.
625 struct io_piece *ipo;
627 if (list_empty(&td->io_log_list))
630 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
631 list_del(&ipo->list);
632 io_u->offset = ipo->offset;
633 io_u->buflen = ipo->len;
634 io_u->ddir = ipo->ddir;
640 * No log, let the seq/rand engine retrieve the next position.
642 if (!get_next_offset(td, &io_u->offset)) {
643 io_u->buflen = get_next_buflen(td);
646 io_u->ddir = get_rw_ddir(td);
654 #define queue_full(td) (list_empty(&(td)->io_u_freelist))
656 struct io_u *__get_io_u(struct thread_data *td)
663 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
666 list_del(&io_u->list);
667 list_add(&io_u->list, &td->io_u_busylist);
672 static struct io_u *get_io_u(struct thread_data *td)
676 io_u = __get_io_u(td);
680 if (td->zone_bytes >= td->zone_size) {
682 td->last_pos += td->zone_skip;
685 if (fill_io_u(td, io_u)) {
690 if (io_u->buflen + io_u->offset > td->real_file_size)
691 io_u->buflen = td->real_file_size - io_u->offset;
698 if (!td->iolog && !td->sequential)
699 mark_random_map(td, io_u);
701 td->last_pos += io_u->buflen;
703 if (td->verify != VERIFY_NONE)
704 populate_io_u(td, io_u);
706 if (td_io_prep(td, io_u)) {
711 gettimeofday(&io_u->start_time, NULL);
715 static inline void td_set_runstate(struct thread_data *td, int runstate)
717 td->old_runstate = td->runstate;
718 td->runstate = runstate;
721 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
723 struct io_piece *ipo;
725 if (list_empty(&td->io_hist_list))
728 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
729 list_del(&ipo->list);
731 io_u->offset = ipo->offset;
732 io_u->buflen = ipo->len;
733 io_u->ddir = DDIR_READ;
738 static void prune_io_piece_log(struct thread_data *td)
740 struct io_piece *ipo;
742 while (!list_empty(&td->io_hist_list)) {
743 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
745 list_del(&ipo->list);
751 * log a succesful write, so we can unwind the log for verify
753 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
755 struct io_piece *ipo = malloc(sizeof(struct io_piece));
756 struct list_head *entry;
758 INIT_LIST_HEAD(&ipo->list);
759 ipo->offset = io_u->offset;
760 ipo->len = io_u->buflen;
763 * for random io where the writes extend the file, it will typically
764 * be laid out with the block scattered as written. it's faster to
765 * read them in in that order again, so don't sort
767 if (td->sequential || !td->overwrite) {
768 list_add_tail(&ipo->list, &td->io_hist_list);
773 * for random io, sort the list so verify will run faster
775 entry = &td->io_hist_list;
776 while ((entry = entry->prev) != &td->io_hist_list) {
777 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
779 if (__ipo->offset < ipo->offset)
783 list_add(&ipo->list, entry);
786 static int init_iolog(struct thread_data *td)
788 unsigned long long offset;
792 int rw, i, reads, writes;
797 f = fopen(td->iolog_file, "r");
799 perror("fopen iolog");
804 reads = writes = i = 0;
805 while ((p = fgets(str, 4096, f)) != NULL) {
806 struct io_piece *ipo;
808 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
809 fprintf(stderr, "bad iolog: %s\n", p);
814 else if (rw == DDIR_WRITE)
817 fprintf(stderr, "bad ddir: %d\n", rw);
821 ipo = malloc(sizeof(*ipo));
822 INIT_LIST_HEAD(&ipo->list);
823 ipo->offset = offset;
825 if (bytes > td->max_bs)
828 list_add_tail(&ipo->list, &td->io_log_list);
838 if (reads && !writes)
839 td->ddir = DDIR_READ;
840 else if (!reads && writes)
841 td->ddir = DDIR_READ;
848 static int sync_td(struct thread_data *td)
851 return td->io_sync(td);
856 static int io_u_getevents(struct thread_data *td, int min, int max,
859 return td->io_getevents(td, min, max, t);
862 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
864 gettimeofday(&io_u->issue_time, NULL);
866 return td->io_queue(td, io_u);
869 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
871 static void io_completed(struct thread_data *td, struct io_u *io_u,
872 struct io_completion_data *icd)
877 gettimeofday(&e, NULL);
880 unsigned int bytes = io_u->buflen - io_u->resid;
881 const int idx = io_u->ddir;
883 td->io_blocks[idx]++;
884 td->io_bytes[idx] += bytes;
885 td->zone_bytes += bytes;
886 td->this_io_bytes[idx] += bytes;
888 msec = mtime_since(&io_u->issue_time, &e);
890 add_clat_sample(td, idx, msec);
891 add_bw_sample(td, idx);
893 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
894 log_io_piece(td, io_u);
896 icd->bytes_done[idx] += bytes;
898 icd->error = io_u->error;
901 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
907 icd->bytes_done[0] = icd->bytes_done[1] = 0;
909 for (i = 0; i < icd->nr; i++) {
910 io_u = td->io_event(td, i);
912 io_completed(td, io_u, icd);
917 static void cleanup_pending_aio(struct thread_data *td)
919 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
920 struct list_head *entry, *n;
921 struct io_completion_data icd;
926 * get immediately available events, if any
928 r = io_u_getevents(td, 0, td->cur_depth, &ts);
931 ios_completed(td, &icd);
935 * now cancel remaining active events
938 list_for_each_safe(entry, n, &td->io_u_busylist) {
939 io_u = list_entry(entry, struct io_u, list);
941 r = td->io_cancel(td, io_u);
948 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
951 ios_completed(td, &icd);
956 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
958 struct io_u *v_io_u = *io_u;
962 ret = verify_io_u(v_io_u);
963 put_io_u(td, v_io_u);
970 static void do_verify(struct thread_data *td)
973 struct io_u *io_u, *v_io_u = NULL;
974 struct io_completion_data icd;
977 td_set_runstate(td, TD_VERIFYING);
983 gettimeofday(&t, NULL);
984 if (runtime_exceeded(td, &t))
987 io_u = __get_io_u(td);
991 if (get_next_verify(td, io_u)) {
996 if (td_io_prep(td, io_u)) {
1001 ret = io_u_queue(td, io_u);
1009 * we have one pending to verify, do that while
1010 * we are doing io on the next one
1012 if (do_io_u_verify(td, &v_io_u))
1015 ret = io_u_getevents(td, 1, 1, NULL);
1022 v_io_u = td->io_event(td, 0);
1025 io_completed(td, v_io_u, &icd);
1028 td_verror(td, icd.error);
1029 put_io_u(td, v_io_u);
1035 * if we can't submit more io, we need to verify now
1037 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1042 do_io_u_verify(td, &v_io_u);
1045 cleanup_pending_aio(td);
1047 td_set_runstate(td, TD_RUNNING);
1050 static void do_io(struct thread_data *td)
1052 struct io_completion_data icd;
1053 struct timeval s, e;
1056 while (td->this_io_bytes[td->ddir] < td->io_size) {
1057 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1058 struct timespec *timeout;
1059 int ret, min_evts = 0;
1065 io_u = get_io_u(td);
1069 memcpy(&s, &io_u->start_time, sizeof(s));
1071 ret = io_u_queue(td, io_u);
1078 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1080 if (td->cur_depth < td->iodepth) {
1088 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1096 ios_completed(td, &icd);
1098 td_verror(td, icd.error);
1103 * the rate is batched for now, it should work for batches
1104 * of completions except the very first one which may look
1107 gettimeofday(&e, NULL);
1108 usec = utime_since(&s, &e);
1110 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1112 if (check_min_rate(td, &e)) {
1113 td_verror(td, ENOMEM);
1117 if (runtime_exceeded(td, &e))
1121 usec_sleep(td, td->thinktime);
1123 if (should_fsync(td) && td->fsync_blocks &&
1124 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1129 cleanup_pending_aio(td);
1131 if (should_fsync(td) && td->end_fsync)
1135 static void cleanup_io(struct thread_data *td)
1141 static int init_io(struct thread_data *td)
1143 if (td->io_engine == FIO_SYNCIO)
1144 return fio_syncio_init(td);
1145 else if (td->io_engine == FIO_MMAPIO)
1146 return fio_mmapio_init(td);
1147 else if (td->io_engine == FIO_LIBAIO)
1148 return fio_libaio_init(td);
1149 else if (td->io_engine == FIO_POSIXAIO)
1150 return fio_posixaio_init(td);
1151 else if (td->io_engine == FIO_SGIO)
1152 return fio_sgio_init(td);
1153 else if (td->io_engine == FIO_SPLICEIO)
1154 return fio_spliceio_init(td);
1156 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1161 static void cleanup_io_u(struct thread_data *td)
1163 struct list_head *entry, *n;
1166 list_for_each_safe(entry, n, &td->io_u_freelist) {
1167 io_u = list_entry(entry, struct io_u, list);
1169 list_del(&io_u->list);
1173 if (td->mem_type == MEM_MALLOC)
1174 free(td->orig_buffer);
1175 else if (td->mem_type == MEM_SHM) {
1176 struct shmid_ds sbuf;
1178 shmdt(td->orig_buffer);
1179 shmctl(td->shm_id, IPC_RMID, &sbuf);
1180 } else if (td->mem_type == MEM_MMAP)
1181 munmap(td->orig_buffer, td->orig_buffer_size);
1183 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1185 td->orig_buffer = NULL;
1188 static int init_io_u(struct thread_data *td)
1194 if (td->io_engine & FIO_SYNCIO)
1197 max_units = td->iodepth;
1199 td->orig_buffer_size = td->max_bs * max_units + MASK;
1201 if (td->mem_type == MEM_MALLOC)
1202 td->orig_buffer = malloc(td->orig_buffer_size);
1203 else if (td->mem_type == MEM_SHM) {
1204 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1205 if (td->shm_id < 0) {
1206 td_verror(td, errno);
1211 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1212 if (td->orig_buffer == (void *) -1) {
1213 td_verror(td, errno);
1215 td->orig_buffer = NULL;
1218 } else if (td->mem_type == MEM_MMAP) {
1219 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1220 if (td->orig_buffer == MAP_FAILED) {
1221 td_verror(td, errno);
1223 td->orig_buffer = NULL;
1228 p = ALIGN(td->orig_buffer);
1229 for (i = 0; i < max_units; i++) {
1230 io_u = malloc(sizeof(*io_u));
1231 memset(io_u, 0, sizeof(*io_u));
1232 INIT_LIST_HEAD(&io_u->list);
1234 io_u->buf = p + td->max_bs * i;
1236 list_add(&io_u->list, &td->io_u_freelist);
1242 static int create_file(struct thread_data *td, unsigned long long size,
1245 unsigned long long left;
1251 * unless specifically asked for overwrite, let normal io extend it
1253 if (td_write(td) && !td->overwrite)
1257 fprintf(stderr, "Need size for create\n");
1258 td_verror(td, EINVAL);
1263 oflags = O_CREAT | O_TRUNC;
1264 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1267 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1270 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1272 td_verror(td, errno);
1276 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1277 td_verror(td, errno);
1281 td->io_size = td->file_size;
1282 b = malloc(td->max_bs);
1283 memset(b, 0, td->max_bs);
1286 while (left && !td->terminate) {
1291 r = write(td->fd, b, bs);
1293 if (r == (int) bs) {
1298 td_verror(td, errno);
1307 unlink(td->file_name);
1308 else if (td->create_fsync)
1317 static int file_size(struct thread_data *td)
1321 if (fstat(td->fd, &st) == -1) {
1322 td_verror(td, errno);
1326 td->real_file_size = st.st_size;
1328 if (!td->file_size || td->file_size > td->real_file_size)
1329 td->file_size = td->real_file_size;
1331 td->file_size -= td->file_offset;
1335 static int bdev_size(struct thread_data *td)
1337 unsigned long long bytes;
1340 r = blockdev_size(td->fd, &bytes);
1346 td->real_file_size = bytes;
1349 * no extend possibilities, so limit size to device size if too large
1351 if (!td->file_size || td->file_size > td->real_file_size)
1352 td->file_size = td->real_file_size;
1354 td->file_size -= td->file_offset;
1358 static int get_file_size(struct thread_data *td)
1362 if (td->filetype == FIO_TYPE_FILE)
1363 ret = file_size(td);
1364 else if (td->filetype == FIO_TYPE_BD)
1365 ret = bdev_size(td);
1367 td->real_file_size = -1;
1372 if (td->file_offset > td->real_file_size) {
1373 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1377 td->io_size = td->file_size;
1378 if (td->io_size == 0) {
1379 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1380 td_verror(td, EINVAL);
1385 td->zone_size = td->io_size;
1387 td->total_io_size = td->io_size * td->loops;
1391 static int setup_file_mmap(struct thread_data *td)
1396 flags = PROT_READ | PROT_WRITE;
1397 else if (td_write(td)) {
1400 if (td->verify != VERIFY_NONE)
1405 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1406 if (td->mmap == MAP_FAILED) {
1408 td_verror(td, errno);
1412 if (td->invalidate_cache) {
1413 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1414 td_verror(td, errno);
1419 if (td->sequential) {
1420 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1421 td_verror(td, errno);
1425 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1426 td_verror(td, errno);
1434 static int setup_file_plain(struct thread_data *td)
1436 if (td->invalidate_cache) {
1437 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1438 td_verror(td, errno);
1443 if (td->sequential) {
1444 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1445 td_verror(td, errno);
1449 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1450 td_verror(td, errno);
1458 static int setup_file(struct thread_data *td)
1463 if (stat(td->file_name, &st) == -1) {
1464 if (errno != ENOENT) {
1465 td_verror(td, errno);
1468 if (!td->create_file) {
1469 td_verror(td, ENOENT);
1472 if (create_file(td, td->file_size, 0))
1474 } else if (td->filetype == FIO_TYPE_FILE) {
1475 if (st.st_size < (off_t) td->file_size) {
1476 if (create_file(td, td->file_size - st.st_size, 1))
1484 if (td_write(td) || td_rw(td)) {
1485 if (td->filetype == FIO_TYPE_FILE) {
1496 td->fd = open(td->file_name, flags, 0600);
1498 if (td->filetype == FIO_TYPE_CHAR)
1503 td->fd = open(td->file_name, flags);
1507 td_verror(td, errno);
1511 if (get_file_size(td))
1514 if (td->io_engine != FIO_MMAPIO)
1515 return setup_file_plain(td);
1517 return setup_file_mmap(td);
1520 static int check_dev_match(dev_t dev, char *path)
1522 unsigned int major, minor;
1526 f = fopen(path, "r");
1528 perror("open path");
1532 p = fgets(line, sizeof(line), f);
1538 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1543 if (((major << 8) | minor) == dev) {
1552 static int find_block_dir(dev_t dev, char *path)
1563 while ((dir = readdir(D)) != NULL) {
1564 char full_path[256];
1566 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1568 if (!strcmp(dir->d_name, "device"))
1571 sprintf(full_path, "%s/%s", path, dir->d_name);
1573 if (!strcmp(dir->d_name, "dev")) {
1574 if (!check_dev_match(dev, full_path)) {
1580 if (stat(full_path, &st) == -1) {
1585 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1588 found = find_block_dir(dev, full_path);
1590 strcpy(path, full_path);
1599 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1606 f = fopen(du->path, "r");
1610 p = fgets(line, sizeof(line), f);
1616 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) {
1625 static void update_io_tick_disk(struct disk_util *du)
1627 struct disk_util_stat __dus, *dus, *ldus;
1630 if (get_io_ticks(du, &__dus))
1634 ldus = &du->last_dus;
1636 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1637 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1638 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1639 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1640 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1641 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1642 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1643 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1644 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1645 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1647 gettimeofday(&t, NULL);
1648 du->msec += mtime_since(&du->time, &t);
1649 memcpy(&du->time, &t, sizeof(t));
1650 memcpy(ldus, &__dus, sizeof(__dus));
1653 static void update_io_ticks(void)
1655 struct list_head *entry;
1656 struct disk_util *du;
1658 list_for_each(entry, &disk_list) {
1659 du = list_entry(entry, struct disk_util, list);
1660 update_io_tick_disk(du);
1664 static int disk_util_exists(dev_t dev)
1666 struct list_head *entry;
1667 struct disk_util *du;
1669 list_for_each(entry, &disk_list) {
1670 du = list_entry(entry, struct disk_util, list);
1679 static void disk_util_add(dev_t dev, char *path)
1681 struct disk_util *du = malloc(sizeof(*du));
1683 memset(du, 0, sizeof(*du));
1684 INIT_LIST_HEAD(&du->list);
1685 sprintf(du->path, "%s/stat", path);
1686 du->name = strdup(basename(path));
1689 gettimeofday(&du->time, NULL);
1690 get_io_ticks(du, &du->last_dus);
1692 list_add_tail(&du->list, &disk_list);
1695 static void init_disk_util(struct thread_data *td)
1698 char foo[256], tmp[256];
1702 if (!td->do_disk_util)
1705 if (!stat(td->file_name, &st)) {
1706 if (S_ISBLK(st.st_mode))
1712 * must be a file, open "." in that path
1714 strcpy(foo, td->file_name);
1717 perror("disk util stat");
1724 if (disk_util_exists(dev))
1727 sprintf(foo, "/sys/block");
1728 if (!find_block_dir(dev, foo))
1732 * If there's a ../queue/ directory there, we are inside a partition.
1733 * Check if that is the case and jump back. For loop/md/dm etc we
1734 * are already in the right spot.
1736 sprintf(tmp, "%s/../queue", foo);
1737 if (!stat(tmp, &st)) {
1739 sprintf(tmp, "%s/queue", p);
1740 if (stat(tmp, &st)) {
1741 fprintf(stderr, "unknown sysfs layout\n");
1744 sprintf(foo, "%s", p);
1747 disk_util_add(dev, foo);
1750 static void disk_util_timer_arm(void)
1752 itimer.it_value.tv_sec = 0;
1753 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1754 setitimer(ITIMER_REAL, &itimer, NULL);
1757 static void clear_io_state(struct thread_data *td)
1759 if (td->io_engine == FIO_SYNCIO)
1760 lseek(td->fd, SEEK_SET, 0);
1763 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1764 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1768 memset(td->file_map, 0, td->num_maps * sizeof(long));
1771 static void update_rusage_stat(struct thread_data *td)
1773 if (!(td->runtime[0] + td->runtime[1]))
1776 getrusage(RUSAGE_SELF, &td->ru_end);
1778 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1779 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1780 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1783 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1786 static void *thread_main(void *data)
1788 struct thread_data *td = data;
1790 if (!td->use_thread)
1795 INIT_LIST_HEAD(&td->io_u_freelist);
1796 INIT_LIST_HEAD(&td->io_u_busylist);
1797 INIT_LIST_HEAD(&td->io_hist_list);
1798 INIT_LIST_HEAD(&td->io_log_list);
1803 if (fio_setaffinity(td) == -1) {
1804 td_verror(td, errno);
1815 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1816 td_verror(td, errno);
1821 if (nice(td->nice) < 0) {
1822 td_verror(td, errno);
1826 if (init_random_state(td))
1829 td_set_runstate(td, TD_INITIALIZED);
1830 sem_post(&startup_sem);
1831 sem_wait(&td->mutex);
1833 if (!td->create_serialize && setup_file(td))
1836 gettimeofday(&td->epoch, NULL);
1838 while (td->loops--) {
1839 getrusage(RUSAGE_SELF, &td->ru_start);
1840 gettimeofday(&td->start, NULL);
1841 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1844 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1847 prune_io_piece_log(td);
1851 td->runtime[td->ddir] += mtime_since_now(&td->start);
1852 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1853 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1855 update_rusage_stat(td);
1857 if (td->error || td->terminate)
1860 if (td->verify == VERIFY_NONE)
1864 gettimeofday(&td->start, NULL);
1868 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1870 if (td->error || td->terminate)
1875 finish_log(td, td->bw_log, "bw");
1877 finish_log(td, td->slat_log, "slat");
1879 finish_log(td, td->clat_log, "clat");
1881 if (exitall_on_terminate)
1882 terminate_threads(td->groupid);
1890 munmap(td->mmap, td->file_size);
1893 td_set_runstate(td, TD_EXITED);
1898 static void *fork_main(int shmid, int offset)
1900 struct thread_data *td;
1903 data = shmat(shmid, NULL, 0);
1904 if (data == (void *) -1) {
1909 td = data + offset * sizeof(struct thread_data);
1915 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1916 double *mean, double *dev)
1920 if (is->samples == 0)
1926 n = (double) is->samples;
1927 *mean = (double) is->val / n;
1928 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1929 if (!(*min + *max) && !(*mean + *dev))
1935 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1938 char *ddir_str[] = { "read ", "write" };
1939 unsigned long min, max;
1940 unsigned long long bw;
1943 if (!td->runtime[ddir])
1946 bw = td->io_bytes[ddir] / td->runtime[ddir];
1947 printf(" %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1949 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1950 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1952 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1953 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1955 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1958 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1959 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);
1963 static void show_thread_status(struct thread_data *td,
1964 struct group_run_stats *rs)
1966 double usr_cpu, sys_cpu;
1968 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1971 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1973 show_ddir_status(td, rs, td->ddir);
1974 if (td->io_bytes[td->ddir ^ 1])
1975 show_ddir_status(td, rs, td->ddir ^ 1);
1977 if (td->runtime[0] + td->runtime[1]) {
1978 double runt = td->runtime[0] + td->runtime[1];
1980 usr_cpu = (double) td->usr_time * 100 / runt;
1981 sys_cpu = (double) td->sys_time * 100 / runt;
1987 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1990 static void check_str_update(struct thread_data *td)
1992 char c = run_str[td->thread_number - 1];
1994 if (td->runstate == td->old_runstate)
1997 switch (td->runstate) {
2010 } else if (td_read(td)) {
2028 case TD_INITIALIZED:
2031 case TD_NOT_CREATED:
2035 printf("state %d\n", td->runstate);
2038 run_str[td->thread_number - 1] = c;
2039 td->old_runstate = td->runstate;
2042 static void eta_to_str(char *str, int eta_sec)
2044 unsigned int d, h, m, s;
2045 static int always_d, always_h;
2057 if (d || always_d) {
2059 str += sprintf(str, "%02dd:", d);
2061 if (h || always_h) {
2063 str += sprintf(str, "%02dh:", h);
2066 str += sprintf(str, "%02dm:", m);
2067 str += sprintf(str, "%02ds", s);
2070 static int thread_eta(struct thread_data *td, unsigned long elapsed)
2072 unsigned long long bytes_total, bytes_done;
2073 unsigned int eta_sec = 0;
2075 bytes_total = td->total_io_size;
2078 * if writing, bytes_total will be twice the size. If mixing,
2079 * assume a 50/50 split and thus bytes_total will be 50% larger.
2083 bytes_total = bytes_total * 3 / 2;
2087 if (td->zone_size && td->zone_skip)
2088 bytes_total /= (td->zone_skip / td->zone_size);
2090 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2093 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2094 perc = (double) bytes_done / (double) bytes_total;
2098 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2100 if (td->timeout && eta_sec > (td->timeout - elapsed))
2101 eta_sec = td->timeout - elapsed;
2102 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2103 || td->runstate == TD_INITIALIZED) {
2104 int t_eta = 0, r_eta = 0;
2107 * We can only guess - assume it'll run the full timeout
2108 * if given, otherwise assume it'll run at the specified rate.
2111 t_eta = td->timeout + td->start_delay - elapsed;
2113 r_eta = (bytes_total / 1024) / td->rate;
2114 r_eta += td->start_delay - elapsed;
2118 eta_sec = min(r_eta, t_eta);
2127 * thread is already done
2135 static void print_thread_status(void)
2137 unsigned long elapsed = time_since_now(&genesis);
2138 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
2142 eta_secs = malloc(thread_number * sizeof(int));
2143 memset(eta_secs, 0, thread_number * sizeof(int));
2145 nr_running = t_rate = m_rate = 0;
2146 for (i = 0; i < thread_number; i++) {
2147 struct thread_data *td = &threads[i];
2149 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2152 m_rate += td->ratemin;
2156 eta_secs[i] = thread_eta(td, elapsed);
2158 eta_secs[i] = INT_MAX;
2160 check_str_update(td);
2163 if (exitall_on_terminate)
2168 for (i = 0; i < thread_number; i++) {
2169 if (exitall_on_terminate) {
2170 if (eta_secs[i] < eta_sec)
2171 eta_sec = eta_secs[i];
2173 if (eta_secs[i] > eta_sec)
2174 eta_sec = eta_secs[i];
2178 if (eta_sec != INT_MAX && elapsed) {
2179 perc = (double) elapsed / (double) (elapsed + eta_sec);
2180 eta_to_str(eta_str, eta_sec);
2183 printf("Threads now running (%d)", nr_running);
2184 if (m_rate || t_rate)
2185 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2186 if (eta_sec != INT_MAX) {
2188 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2195 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2200 * reap exited threads (TD_EXITED -> TD_REAPED)
2202 for (i = 0; i < thread_number; i++) {
2203 struct thread_data *td = &threads[i];
2205 if (td->runstate != TD_EXITED)
2208 td_set_runstate(td, TD_REAPED);
2210 if (td->use_thread) {
2213 if (pthread_join(td->thread, (void *) &ret))
2214 perror("thread_join");
2216 waitpid(td->pid, NULL, 0);
2219 (*m_rate) -= td->ratemin;
2220 (*t_rate) -= td->rate;
2224 static void fio_unpin_memory(void *pinned)
2227 if (munlock(pinned, mlock_size) < 0)
2229 munmap(pinned, mlock_size);
2233 static void *fio_pin_memory(void)
2235 long pagesize, pages;
2242 * Don't allow mlock of more than real_mem-128MB
2244 pagesize = sysconf(_SC_PAGESIZE);
2245 pages = sysconf(_SC_PHYS_PAGES);
2246 if (pages != -1 && pagesize != -1) {
2247 unsigned long long real_mem = pages * pagesize;
2249 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2250 mlock_size = real_mem - 128 * 1024 * 1024;
2251 printf("fio: limiting mlocked memory to %lluMiB\n",
2256 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2258 perror("malloc locked mem");
2261 if (mlock(ptr, mlock_size) < 0) {
2262 munmap(ptr, mlock_size);
2270 static void run_threads(void)
2272 struct thread_data *td;
2273 unsigned long spent;
2274 int i, todo, nr_running, m_rate, t_rate, nr_started;
2277 mlocked_mem = fio_pin_memory();
2279 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2282 signal(SIGINT, sig_handler);
2283 signal(SIGALRM, sig_handler);
2285 todo = thread_number;
2288 m_rate = t_rate = 0;
2290 for (i = 0; i < thread_number; i++) {
2293 run_str[td->thread_number - 1] = 'P';
2297 if (!td->create_serialize)
2301 * do file setup here so it happens sequentially,
2302 * we don't want X number of threads getting their
2303 * client data interspersed on disk
2305 if (setup_file(td)) {
2306 td_set_runstate(td, TD_REAPED);
2311 gettimeofday(&genesis, NULL);
2314 struct thread_data *map[MAX_JOBS];
2315 struct timeval this_start;
2316 int this_jobs = 0, left;
2319 * create threads (TD_NOT_CREATED -> TD_CREATED)
2321 for (i = 0; i < thread_number; i++) {
2324 if (td->runstate != TD_NOT_CREATED)
2328 * never got a chance to start, killed by other
2329 * thread for some reason
2331 if (td->terminate) {
2336 if (td->start_delay) {
2337 spent = mtime_since_now(&genesis);
2339 if (td->start_delay * 1000 > spent)
2343 if (td->stonewall && (nr_started || nr_running))
2347 * Set state to created. Thread will transition
2348 * to TD_INITIALIZED when it's done setting up.
2350 td_set_runstate(td, TD_CREATED);
2351 map[this_jobs++] = td;
2352 sem_init(&startup_sem, 0, 1);
2355 if (td->use_thread) {
2356 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2357 perror("thread_create");
2362 sem_wait(&startup_sem);
2364 fork_main(shm_id, i);
2371 * Wait for the started threads to transition to
2374 printf("fio: Waiting for threads to initialize...\n");
2375 gettimeofday(&this_start, NULL);
2378 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2383 for (i = 0; i < this_jobs; i++) {
2387 if (td->runstate == TD_INITIALIZED) {
2390 } else if (td->runstate >= TD_EXITED) {
2394 nr_running++; /* work-around... */
2400 fprintf(stderr, "fio: %d jobs failed to start\n", left);
2401 for (i = 0; i < this_jobs; i++) {
2405 kill(td->pid, SIGTERM);
2411 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2413 printf("fio: Go for launch\n");
2414 for (i = 0; i < thread_number; i++) {
2417 if (td->runstate != TD_INITIALIZED)
2420 td_set_runstate(td, TD_RUNNING);
2423 m_rate += td->ratemin;
2426 sem_post(&td->mutex);
2429 reap_threads(&nr_running, &t_rate, &m_rate);
2435 while (nr_running) {
2436 reap_threads(&nr_running, &t_rate, &m_rate);
2441 fio_unpin_memory(mlocked_mem);
2444 static void show_group_stats(struct group_run_stats *rs, int id)
2446 printf("\nRun status group %d (all jobs):\n", id);
2448 if (rs->max_run[DDIR_READ])
2449 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]);
2450 if (rs->max_run[DDIR_WRITE])
2451 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]);
2454 static void show_disk_util(void)
2456 struct disk_util_stat *dus;
2457 struct list_head *entry;
2458 struct disk_util *du;
2461 printf("\nDisk stats (read/write):\n");
2463 list_for_each(entry, &disk_list) {
2464 du = list_entry(entry, struct disk_util, list);
2467 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2471 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);
2475 static void show_run_stats(void)
2477 struct group_run_stats *runstats, *rs;
2478 struct thread_data *td;
2481 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2483 for (i = 0; i < groupid + 1; i++) {
2486 memset(rs, 0, sizeof(*rs));
2487 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2488 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2491 for (i = 0; i < thread_number; i++) {
2492 unsigned long long rbw, wbw;
2497 printf("Client%d: %s\n", td->thread_number, td->verror);
2501 rs = &runstats[td->groupid];
2503 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2504 rs->min_run[0] = td->runtime[0];
2505 if (td->runtime[0] > rs->max_run[0])
2506 rs->max_run[0] = td->runtime[0];
2507 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2508 rs->min_run[1] = td->runtime[1];
2509 if (td->runtime[1] > rs->max_run[1])
2510 rs->max_run[1] = td->runtime[1];
2514 rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
2516 wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
2518 if (rbw < rs->min_bw[0])
2519 rs->min_bw[0] = rbw;
2520 if (wbw < rs->min_bw[1])
2521 rs->min_bw[1] = wbw;
2522 if (rbw > rs->max_bw[0])
2523 rs->max_bw[0] = rbw;
2524 if (wbw > rs->max_bw[1])
2525 rs->max_bw[1] = wbw;
2527 rs->io_mb[0] += td->io_bytes[0] >> 20;
2528 rs->io_mb[1] += td->io_bytes[1] >> 20;
2531 for (i = 0; i < groupid + 1; i++) {
2535 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2537 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2541 * don't overwrite last signal output
2545 for (i = 0; i < thread_number; i++) {
2547 rs = &runstats[td->groupid];
2549 show_thread_status(td, rs);
2552 for (i = 0; i < groupid + 1; i++)
2553 show_group_stats(&runstats[i], i);
2558 int main(int argc, char *argv[])
2560 if (parse_options(argc, argv))
2563 if (!thread_number) {
2564 printf("Nothing to do\n");
2568 disk_util_timer_arm();