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 void write_iolog_put(struct thread_data *td, struct io_u *io_u)
621 fprintf(td->iolog_f, "%d,%llu,%u\n", io_u->ddir, io_u->offset, io_u->buflen);
624 static int read_iolog_get(struct thread_data *td, struct io_u *io_u)
626 struct io_piece *ipo;
628 if (!list_empty(&td->io_log_list)) {
629 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
630 list_del(&ipo->list);
631 io_u->offset = ipo->offset;
632 io_u->buflen = ipo->len;
633 io_u->ddir = ipo->ddir;
641 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
644 * If using an iolog, grab next piece if any available.
647 return read_iolog_get(td, io_u);
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);
659 * If using a write iolog, store this entry.
662 write_iolog_put(td, io_u);
671 #define queue_full(td) (list_empty(&(td)->io_u_freelist))
673 struct io_u *__get_io_u(struct thread_data *td)
680 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
683 list_del(&io_u->list);
684 list_add(&io_u->list, &td->io_u_busylist);
689 static struct io_u *get_io_u(struct thread_data *td)
693 io_u = __get_io_u(td);
697 if (td->zone_bytes >= td->zone_size) {
699 td->last_pos += td->zone_skip;
702 if (fill_io_u(td, io_u)) {
707 if (io_u->buflen + io_u->offset > td->real_file_size)
708 io_u->buflen = td->real_file_size - io_u->offset;
715 if (!td->read_iolog && !td->sequential)
716 mark_random_map(td, io_u);
718 td->last_pos += io_u->buflen;
720 if (td->verify != VERIFY_NONE)
721 populate_io_u(td, io_u);
723 if (td_io_prep(td, io_u)) {
728 gettimeofday(&io_u->start_time, NULL);
732 static inline void td_set_runstate(struct thread_data *td, int runstate)
734 td->old_runstate = td->runstate;
735 td->runstate = runstate;
738 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
740 struct io_piece *ipo;
742 if (list_empty(&td->io_hist_list))
745 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
746 list_del(&ipo->list);
748 io_u->offset = ipo->offset;
749 io_u->buflen = ipo->len;
750 io_u->ddir = DDIR_READ;
755 static void prune_io_piece_log(struct thread_data *td)
757 struct io_piece *ipo;
759 while (!list_empty(&td->io_hist_list)) {
760 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
762 list_del(&ipo->list);
768 * log a succesful write, so we can unwind the log for verify
770 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
772 struct io_piece *ipo = malloc(sizeof(struct io_piece));
773 struct list_head *entry;
775 INIT_LIST_HEAD(&ipo->list);
776 ipo->offset = io_u->offset;
777 ipo->len = io_u->buflen;
780 * for random io where the writes extend the file, it will typically
781 * be laid out with the block scattered as written. it's faster to
782 * read them in in that order again, so don't sort
784 if (td->sequential || !td->overwrite) {
785 list_add_tail(&ipo->list, &td->io_hist_list);
790 * for random io, sort the list so verify will run faster
792 entry = &td->io_hist_list;
793 while ((entry = entry->prev) != &td->io_hist_list) {
794 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
796 if (__ipo->offset < ipo->offset)
800 list_add(&ipo->list, entry);
803 static void write_iolog_close(struct thread_data *td)
810 static int init_iolog(struct thread_data *td)
812 unsigned long long offset;
816 int rw, i, reads, writes;
818 if (!td->read_iolog && !td->write_iolog)
822 f = fopen(td->iolog_file, "r");
824 f = fopen(td->iolog_file, "w");
827 perror("fopen iolog");
828 printf("file %s, %d/%d\n", td->iolog_file, td->read_iolog, td->write_iolog);
833 * That's it for writing, setup a log buffer and we're done.
835 if (td->write_iolog) {
837 td->iolog_buf = malloc(8192);
838 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
843 * Read in the read iolog and store it, reuse the infrastructure
844 * for doing verifications.
847 reads = writes = i = 0;
848 while ((p = fgets(str, 4096, f)) != NULL) {
849 struct io_piece *ipo;
851 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
852 fprintf(stderr, "bad iolog: %s\n", p);
857 else if (rw == DDIR_WRITE)
860 fprintf(stderr, "bad ddir: %d\n", rw);
864 ipo = malloc(sizeof(*ipo));
865 INIT_LIST_HEAD(&ipo->list);
866 ipo->offset = offset;
868 if (bytes > td->max_bs)
871 list_add_tail(&ipo->list, &td->io_log_list);
881 if (reads && !writes)
882 td->ddir = DDIR_READ;
883 else if (!reads && writes)
884 td->ddir = DDIR_READ;
891 static int sync_td(struct thread_data *td)
894 return td->io_sync(td);
899 static int io_u_getevents(struct thread_data *td, int min, int max,
902 return td->io_getevents(td, min, max, t);
905 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
907 gettimeofday(&io_u->issue_time, NULL);
909 return td->io_queue(td, io_u);
912 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
914 static void io_completed(struct thread_data *td, struct io_u *io_u,
915 struct io_completion_data *icd)
920 gettimeofday(&e, NULL);
923 unsigned int bytes = io_u->buflen - io_u->resid;
924 const int idx = io_u->ddir;
926 td->io_blocks[idx]++;
927 td->io_bytes[idx] += bytes;
928 td->zone_bytes += bytes;
929 td->this_io_bytes[idx] += bytes;
931 msec = mtime_since(&io_u->issue_time, &e);
933 add_clat_sample(td, idx, msec);
934 add_bw_sample(td, idx);
936 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
937 log_io_piece(td, io_u);
939 icd->bytes_done[idx] += bytes;
941 icd->error = io_u->error;
944 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
950 icd->bytes_done[0] = icd->bytes_done[1] = 0;
952 for (i = 0; i < icd->nr; i++) {
953 io_u = td->io_event(td, i);
955 io_completed(td, io_u, icd);
960 static void cleanup_pending_aio(struct thread_data *td)
962 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
963 struct list_head *entry, *n;
964 struct io_completion_data icd;
969 * get immediately available events, if any
971 r = io_u_getevents(td, 0, td->cur_depth, &ts);
974 ios_completed(td, &icd);
978 * now cancel remaining active events
981 list_for_each_safe(entry, n, &td->io_u_busylist) {
982 io_u = list_entry(entry, struct io_u, list);
984 r = td->io_cancel(td, io_u);
991 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
994 ios_completed(td, &icd);
999 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
1001 struct io_u *v_io_u = *io_u;
1005 ret = verify_io_u(v_io_u);
1006 put_io_u(td, v_io_u);
1013 static void do_verify(struct thread_data *td)
1016 struct io_u *io_u, *v_io_u = NULL;
1017 struct io_completion_data icd;
1020 td_set_runstate(td, TD_VERIFYING);
1026 gettimeofday(&t, NULL);
1027 if (runtime_exceeded(td, &t))
1030 io_u = __get_io_u(td);
1034 if (get_next_verify(td, io_u)) {
1039 if (td_io_prep(td, io_u)) {
1044 ret = io_u_queue(td, io_u);
1052 * we have one pending to verify, do that while
1053 * we are doing io on the next one
1055 if (do_io_u_verify(td, &v_io_u))
1058 ret = io_u_getevents(td, 1, 1, NULL);
1065 v_io_u = td->io_event(td, 0);
1068 io_completed(td, v_io_u, &icd);
1071 td_verror(td, icd.error);
1072 put_io_u(td, v_io_u);
1078 * if we can't submit more io, we need to verify now
1080 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1085 do_io_u_verify(td, &v_io_u);
1088 cleanup_pending_aio(td);
1090 td_set_runstate(td, TD_RUNNING);
1093 static void do_io(struct thread_data *td)
1095 struct io_completion_data icd;
1096 struct timeval s, e;
1099 while (td->this_io_bytes[td->ddir] < td->io_size) {
1100 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1101 struct timespec *timeout;
1102 int ret, min_evts = 0;
1108 io_u = get_io_u(td);
1112 memcpy(&s, &io_u->start_time, sizeof(s));
1114 ret = io_u_queue(td, io_u);
1121 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1123 if (td->cur_depth < td->iodepth) {
1131 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1139 ios_completed(td, &icd);
1141 td_verror(td, icd.error);
1146 * the rate is batched for now, it should work for batches
1147 * of completions except the very first one which may look
1150 gettimeofday(&e, NULL);
1151 usec = utime_since(&s, &e);
1153 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1155 if (check_min_rate(td, &e)) {
1156 td_verror(td, ENOMEM);
1160 if (runtime_exceeded(td, &e))
1164 usec_sleep(td, td->thinktime);
1166 if (should_fsync(td) && td->fsync_blocks &&
1167 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1172 cleanup_pending_aio(td);
1174 if (should_fsync(td) && td->end_fsync)
1178 static void cleanup_io(struct thread_data *td)
1184 static int init_io(struct thread_data *td)
1186 if (td->io_engine == FIO_SYNCIO)
1187 return fio_syncio_init(td);
1188 else if (td->io_engine == FIO_MMAPIO)
1189 return fio_mmapio_init(td);
1190 else if (td->io_engine == FIO_LIBAIO)
1191 return fio_libaio_init(td);
1192 else if (td->io_engine == FIO_POSIXAIO)
1193 return fio_posixaio_init(td);
1194 else if (td->io_engine == FIO_SGIO)
1195 return fio_sgio_init(td);
1196 else if (td->io_engine == FIO_SPLICEIO)
1197 return fio_spliceio_init(td);
1199 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1204 static void cleanup_io_u(struct thread_data *td)
1206 struct list_head *entry, *n;
1209 list_for_each_safe(entry, n, &td->io_u_freelist) {
1210 io_u = list_entry(entry, struct io_u, list);
1212 list_del(&io_u->list);
1216 if (td->mem_type == MEM_MALLOC)
1217 free(td->orig_buffer);
1218 else if (td->mem_type == MEM_SHM) {
1219 struct shmid_ds sbuf;
1221 shmdt(td->orig_buffer);
1222 shmctl(td->shm_id, IPC_RMID, &sbuf);
1223 } else if (td->mem_type == MEM_MMAP)
1224 munmap(td->orig_buffer, td->orig_buffer_size);
1226 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1228 td->orig_buffer = NULL;
1231 static int init_io_u(struct thread_data *td)
1237 if (td->io_engine & FIO_SYNCIO)
1240 max_units = td->iodepth;
1242 td->orig_buffer_size = td->max_bs * max_units + MASK;
1244 if (td->mem_type == MEM_MALLOC)
1245 td->orig_buffer = malloc(td->orig_buffer_size);
1246 else if (td->mem_type == MEM_SHM) {
1247 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1248 if (td->shm_id < 0) {
1249 td_verror(td, errno);
1254 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1255 if (td->orig_buffer == (void *) -1) {
1256 td_verror(td, errno);
1258 td->orig_buffer = NULL;
1261 } else if (td->mem_type == MEM_MMAP) {
1262 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1263 if (td->orig_buffer == MAP_FAILED) {
1264 td_verror(td, errno);
1266 td->orig_buffer = NULL;
1271 p = ALIGN(td->orig_buffer);
1272 for (i = 0; i < max_units; i++) {
1273 io_u = malloc(sizeof(*io_u));
1274 memset(io_u, 0, sizeof(*io_u));
1275 INIT_LIST_HEAD(&io_u->list);
1277 io_u->buf = p + td->max_bs * i;
1279 list_add(&io_u->list, &td->io_u_freelist);
1285 static int create_file(struct thread_data *td, unsigned long long size,
1288 unsigned long long left;
1294 * unless specifically asked for overwrite, let normal io extend it
1296 if (td_write(td) && !td->overwrite)
1300 fprintf(stderr, "Need size for create\n");
1301 td_verror(td, EINVAL);
1306 oflags = O_CREAT | O_TRUNC;
1307 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1310 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1313 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1315 td_verror(td, errno);
1319 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1320 td_verror(td, errno);
1324 td->io_size = td->file_size;
1325 b = malloc(td->max_bs);
1326 memset(b, 0, td->max_bs);
1329 while (left && !td->terminate) {
1334 r = write(td->fd, b, bs);
1336 if (r == (int) bs) {
1341 td_verror(td, errno);
1350 unlink(td->file_name);
1351 else if (td->create_fsync)
1360 static int file_size(struct thread_data *td)
1364 if (fstat(td->fd, &st) == -1) {
1365 td_verror(td, errno);
1369 td->real_file_size = st.st_size;
1371 if (!td->file_size || td->file_size > td->real_file_size)
1372 td->file_size = td->real_file_size;
1374 td->file_size -= td->file_offset;
1378 static int bdev_size(struct thread_data *td)
1380 unsigned long long bytes;
1383 r = blockdev_size(td->fd, &bytes);
1389 td->real_file_size = bytes;
1392 * no extend possibilities, so limit size to device size if too large
1394 if (!td->file_size || td->file_size > td->real_file_size)
1395 td->file_size = td->real_file_size;
1397 td->file_size -= td->file_offset;
1401 static int get_file_size(struct thread_data *td)
1405 if (td->filetype == FIO_TYPE_FILE)
1406 ret = file_size(td);
1407 else if (td->filetype == FIO_TYPE_BD)
1408 ret = bdev_size(td);
1410 td->real_file_size = -1;
1415 if (td->file_offset > td->real_file_size) {
1416 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1420 td->io_size = td->file_size;
1421 if (td->io_size == 0) {
1422 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1423 td_verror(td, EINVAL);
1428 td->zone_size = td->io_size;
1430 td->total_io_size = td->io_size * td->loops;
1434 static int setup_file_mmap(struct thread_data *td)
1439 flags = PROT_READ | PROT_WRITE;
1440 else if (td_write(td)) {
1443 if (td->verify != VERIFY_NONE)
1448 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1449 if (td->mmap == MAP_FAILED) {
1451 td_verror(td, errno);
1455 if (td->invalidate_cache) {
1456 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1457 td_verror(td, errno);
1462 if (td->sequential) {
1463 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1464 td_verror(td, errno);
1468 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1469 td_verror(td, errno);
1477 static int setup_file_plain(struct thread_data *td)
1479 if (td->invalidate_cache) {
1480 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1481 td_verror(td, errno);
1486 if (td->sequential) {
1487 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1488 td_verror(td, errno);
1492 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1493 td_verror(td, errno);
1501 static int setup_file(struct thread_data *td)
1506 if (stat(td->file_name, &st) == -1) {
1507 if (errno != ENOENT) {
1508 td_verror(td, errno);
1511 if (!td->create_file) {
1512 td_verror(td, ENOENT);
1515 if (create_file(td, td->file_size, 0))
1517 } else if (td->filetype == FIO_TYPE_FILE) {
1518 if (st.st_size < (off_t) td->file_size) {
1519 if (create_file(td, td->file_size - st.st_size, 1))
1527 if (td_write(td) || td_rw(td)) {
1528 if (td->filetype == FIO_TYPE_FILE) {
1539 td->fd = open(td->file_name, flags, 0600);
1541 if (td->filetype == FIO_TYPE_CHAR)
1546 td->fd = open(td->file_name, flags);
1550 td_verror(td, errno);
1554 if (get_file_size(td))
1557 if (td->io_engine != FIO_MMAPIO)
1558 return setup_file_plain(td);
1560 return setup_file_mmap(td);
1563 static int check_dev_match(dev_t dev, char *path)
1565 unsigned int major, minor;
1569 f = fopen(path, "r");
1571 perror("open path");
1575 p = fgets(line, sizeof(line), f);
1581 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1586 if (((major << 8) | minor) == dev) {
1595 static int find_block_dir(dev_t dev, char *path)
1606 while ((dir = readdir(D)) != NULL) {
1607 char full_path[256];
1609 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1611 if (!strcmp(dir->d_name, "device"))
1614 sprintf(full_path, "%s/%s", path, dir->d_name);
1616 if (!strcmp(dir->d_name, "dev")) {
1617 if (!check_dev_match(dev, full_path)) {
1623 if (stat(full_path, &st) == -1) {
1628 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1631 found = find_block_dir(dev, full_path);
1633 strcpy(path, full_path);
1642 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1649 f = fopen(du->path, "r");
1653 p = fgets(line, sizeof(line), f);
1659 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) {
1668 static void update_io_tick_disk(struct disk_util *du)
1670 struct disk_util_stat __dus, *dus, *ldus;
1673 if (get_io_ticks(du, &__dus))
1677 ldus = &du->last_dus;
1679 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1680 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1681 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1682 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1683 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1684 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1685 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1686 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1687 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1688 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1690 gettimeofday(&t, NULL);
1691 du->msec += mtime_since(&du->time, &t);
1692 memcpy(&du->time, &t, sizeof(t));
1693 memcpy(ldus, &__dus, sizeof(__dus));
1696 static void update_io_ticks(void)
1698 struct list_head *entry;
1699 struct disk_util *du;
1701 list_for_each(entry, &disk_list) {
1702 du = list_entry(entry, struct disk_util, list);
1703 update_io_tick_disk(du);
1707 static int disk_util_exists(dev_t dev)
1709 struct list_head *entry;
1710 struct disk_util *du;
1712 list_for_each(entry, &disk_list) {
1713 du = list_entry(entry, struct disk_util, list);
1722 static void disk_util_add(dev_t dev, char *path)
1724 struct disk_util *du = malloc(sizeof(*du));
1726 memset(du, 0, sizeof(*du));
1727 INIT_LIST_HEAD(&du->list);
1728 sprintf(du->path, "%s/stat", path);
1729 du->name = strdup(basename(path));
1732 gettimeofday(&du->time, NULL);
1733 get_io_ticks(du, &du->last_dus);
1735 list_add_tail(&du->list, &disk_list);
1738 static void init_disk_util(struct thread_data *td)
1741 char foo[256], tmp[256];
1745 if (!td->do_disk_util)
1748 if (!stat(td->file_name, &st)) {
1749 if (S_ISBLK(st.st_mode))
1755 * must be a file, open "." in that path
1757 strcpy(foo, td->file_name);
1760 perror("disk util stat");
1767 if (disk_util_exists(dev))
1770 sprintf(foo, "/sys/block");
1771 if (!find_block_dir(dev, foo))
1775 * If there's a ../queue/ directory there, we are inside a partition.
1776 * Check if that is the case and jump back. For loop/md/dm etc we
1777 * are already in the right spot.
1779 sprintf(tmp, "%s/../queue", foo);
1780 if (!stat(tmp, &st)) {
1782 sprintf(tmp, "%s/queue", p);
1783 if (stat(tmp, &st)) {
1784 fprintf(stderr, "unknown sysfs layout\n");
1787 sprintf(foo, "%s", p);
1790 disk_util_add(dev, foo);
1793 static void disk_util_timer_arm(void)
1795 itimer.it_value.tv_sec = 0;
1796 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1797 setitimer(ITIMER_REAL, &itimer, NULL);
1800 static void clear_io_state(struct thread_data *td)
1802 if (td->io_engine == FIO_SYNCIO)
1803 lseek(td->fd, SEEK_SET, 0);
1806 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1807 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1811 memset(td->file_map, 0, td->num_maps * sizeof(long));
1814 static void update_rusage_stat(struct thread_data *td)
1816 if (!(td->runtime[0] + td->runtime[1]))
1819 getrusage(RUSAGE_SELF, &td->ru_end);
1821 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1822 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1823 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1826 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1829 static void *thread_main(void *data)
1831 struct thread_data *td = data;
1833 if (!td->use_thread)
1838 INIT_LIST_HEAD(&td->io_u_freelist);
1839 INIT_LIST_HEAD(&td->io_u_busylist);
1840 INIT_LIST_HEAD(&td->io_hist_list);
1841 INIT_LIST_HEAD(&td->io_log_list);
1846 if (fio_setaffinity(td) == -1) {
1847 td_verror(td, errno);
1858 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1859 td_verror(td, errno);
1864 if (nice(td->nice) < 0) {
1865 td_verror(td, errno);
1869 if (init_random_state(td))
1872 td_set_runstate(td, TD_INITIALIZED);
1873 sem_post(&startup_sem);
1874 sem_wait(&td->mutex);
1876 if (!td->create_serialize && setup_file(td))
1879 gettimeofday(&td->epoch, NULL);
1881 while (td->loops--) {
1882 getrusage(RUSAGE_SELF, &td->ru_start);
1883 gettimeofday(&td->start, NULL);
1884 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1887 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1890 prune_io_piece_log(td);
1894 td->runtime[td->ddir] += mtime_since_now(&td->start);
1895 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1896 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1898 update_rusage_stat(td);
1900 if (td->error || td->terminate)
1903 if (td->verify == VERIFY_NONE)
1907 gettimeofday(&td->start, NULL);
1911 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1913 if (td->error || td->terminate)
1918 finish_log(td, td->bw_log, "bw");
1920 finish_log(td, td->slat_log, "slat");
1922 finish_log(td, td->clat_log, "clat");
1923 if (td->write_iolog)
1924 write_iolog_close(td);
1926 if (exitall_on_terminate)
1927 terminate_threads(td->groupid);
1935 munmap(td->mmap, td->file_size);
1938 td_set_runstate(td, TD_EXITED);
1943 static void *fork_main(int shmid, int offset)
1945 struct thread_data *td;
1948 data = shmat(shmid, NULL, 0);
1949 if (data == (void *) -1) {
1954 td = data + offset * sizeof(struct thread_data);
1960 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1961 double *mean, double *dev)
1965 if (is->samples == 0)
1971 n = (double) is->samples;
1972 *mean = (double) is->val / n;
1973 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1974 if (!(*min + *max) && !(*mean + *dev))
1980 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1983 char *ddir_str[] = { "read ", "write" };
1984 unsigned long min, max;
1985 unsigned long long bw;
1988 if (!td->runtime[ddir])
1991 bw = td->io_bytes[ddir] / td->runtime[ddir];
1992 printf(" %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1994 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1995 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1997 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1998 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
2000 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
2003 p_of_agg = mean * 100 / (double) rs->agg[ddir];
2004 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);
2008 static void show_thread_status(struct thread_data *td,
2009 struct group_run_stats *rs)
2011 double usr_cpu, sys_cpu;
2013 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
2016 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
2018 show_ddir_status(td, rs, td->ddir);
2019 if (td->io_bytes[td->ddir ^ 1])
2020 show_ddir_status(td, rs, td->ddir ^ 1);
2022 if (td->runtime[0] + td->runtime[1]) {
2023 double runt = td->runtime[0] + td->runtime[1];
2025 usr_cpu = (double) td->usr_time * 100 / runt;
2026 sys_cpu = (double) td->sys_time * 100 / runt;
2032 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
2035 static void check_str_update(struct thread_data *td)
2037 char c = run_str[td->thread_number - 1];
2039 if (td->runstate == td->old_runstate)
2042 switch (td->runstate) {
2055 } else if (td_read(td)) {
2073 case TD_INITIALIZED:
2076 case TD_NOT_CREATED:
2080 printf("state %d\n", td->runstate);
2083 run_str[td->thread_number - 1] = c;
2084 td->old_runstate = td->runstate;
2087 static void eta_to_str(char *str, int eta_sec)
2089 unsigned int d, h, m, s;
2090 static int always_d, always_h;
2102 if (d || always_d) {
2104 str += sprintf(str, "%02dd:", d);
2106 if (h || always_h) {
2108 str += sprintf(str, "%02dh:", h);
2111 str += sprintf(str, "%02dm:", m);
2112 str += sprintf(str, "%02ds", s);
2115 static int thread_eta(struct thread_data *td, unsigned long elapsed)
2117 unsigned long long bytes_total, bytes_done;
2118 unsigned int eta_sec = 0;
2120 bytes_total = td->total_io_size;
2123 * if writing, bytes_total will be twice the size. If mixing,
2124 * assume a 50/50 split and thus bytes_total will be 50% larger.
2128 bytes_total = bytes_total * 3 / 2;
2132 if (td->zone_size && td->zone_skip)
2133 bytes_total /= (td->zone_skip / td->zone_size);
2135 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2138 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2139 perc = (double) bytes_done / (double) bytes_total;
2143 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2145 if (td->timeout && eta_sec > (td->timeout - elapsed))
2146 eta_sec = td->timeout - elapsed;
2147 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2148 || td->runstate == TD_INITIALIZED) {
2149 int t_eta = 0, r_eta = 0;
2152 * We can only guess - assume it'll run the full timeout
2153 * if given, otherwise assume it'll run at the specified rate.
2156 t_eta = td->timeout + td->start_delay - elapsed;
2158 r_eta = (bytes_total / 1024) / td->rate;
2159 r_eta += td->start_delay - elapsed;
2163 eta_sec = min(r_eta, t_eta);
2172 * thread is already done
2180 static void print_thread_status(void)
2182 unsigned long elapsed = time_since_now(&genesis);
2183 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
2187 eta_secs = malloc(thread_number * sizeof(int));
2188 memset(eta_secs, 0, thread_number * sizeof(int));
2190 nr_running = t_rate = m_rate = 0;
2191 for (i = 0; i < thread_number; i++) {
2192 struct thread_data *td = &threads[i];
2194 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2197 m_rate += td->ratemin;
2201 eta_secs[i] = thread_eta(td, elapsed);
2203 eta_secs[i] = INT_MAX;
2205 check_str_update(td);
2208 if (exitall_on_terminate)
2213 for (i = 0; i < thread_number; i++) {
2214 if (exitall_on_terminate) {
2215 if (eta_secs[i] < eta_sec)
2216 eta_sec = eta_secs[i];
2218 if (eta_secs[i] > eta_sec)
2219 eta_sec = eta_secs[i];
2223 if (eta_sec != INT_MAX && elapsed) {
2224 perc = (double) elapsed / (double) (elapsed + eta_sec);
2225 eta_to_str(eta_str, eta_sec);
2228 printf("Threads now running (%d)", nr_running);
2229 if (m_rate || t_rate)
2230 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2231 if (eta_sec != INT_MAX) {
2233 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2240 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2245 * reap exited threads (TD_EXITED -> TD_REAPED)
2247 for (i = 0; i < thread_number; i++) {
2248 struct thread_data *td = &threads[i];
2250 if (td->runstate != TD_EXITED)
2253 td_set_runstate(td, TD_REAPED);
2255 if (td->use_thread) {
2258 if (pthread_join(td->thread, (void *) &ret))
2259 perror("thread_join");
2261 waitpid(td->pid, NULL, 0);
2264 (*m_rate) -= td->ratemin;
2265 (*t_rate) -= td->rate;
2269 static void fio_unpin_memory(void *pinned)
2272 if (munlock(pinned, mlock_size) < 0)
2274 munmap(pinned, mlock_size);
2278 static void *fio_pin_memory(void)
2280 long pagesize, pages;
2287 * Don't allow mlock of more than real_mem-128MB
2289 pagesize = sysconf(_SC_PAGESIZE);
2290 pages = sysconf(_SC_PHYS_PAGES);
2291 if (pages != -1 && pagesize != -1) {
2292 unsigned long long real_mem = pages * pagesize;
2294 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2295 mlock_size = real_mem - 128 * 1024 * 1024;
2296 printf("fio: limiting mlocked memory to %lluMiB\n",
2301 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2303 perror("malloc locked mem");
2306 if (mlock(ptr, mlock_size) < 0) {
2307 munmap(ptr, mlock_size);
2315 static void run_threads(void)
2317 struct thread_data *td;
2318 unsigned long spent;
2319 int i, todo, nr_running, m_rate, t_rate, nr_started;
2322 mlocked_mem = fio_pin_memory();
2324 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2327 signal(SIGINT, sig_handler);
2328 signal(SIGALRM, sig_handler);
2330 todo = thread_number;
2333 m_rate = t_rate = 0;
2335 for (i = 0; i < thread_number; i++) {
2338 run_str[td->thread_number - 1] = 'P';
2342 if (!td->create_serialize)
2346 * do file setup here so it happens sequentially,
2347 * we don't want X number of threads getting their
2348 * client data interspersed on disk
2350 if (setup_file(td)) {
2351 td_set_runstate(td, TD_REAPED);
2356 gettimeofday(&genesis, NULL);
2359 struct thread_data *map[MAX_JOBS];
2360 struct timeval this_start;
2361 int this_jobs = 0, left;
2364 * create threads (TD_NOT_CREATED -> TD_CREATED)
2366 for (i = 0; i < thread_number; i++) {
2369 if (td->runstate != TD_NOT_CREATED)
2373 * never got a chance to start, killed by other
2374 * thread for some reason
2376 if (td->terminate) {
2381 if (td->start_delay) {
2382 spent = mtime_since_now(&genesis);
2384 if (td->start_delay * 1000 > spent)
2388 if (td->stonewall && (nr_started || nr_running))
2392 * Set state to created. Thread will transition
2393 * to TD_INITIALIZED when it's done setting up.
2395 td_set_runstate(td, TD_CREATED);
2396 map[this_jobs++] = td;
2397 sem_init(&startup_sem, 0, 1);
2400 if (td->use_thread) {
2401 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2402 perror("thread_create");
2407 sem_wait(&startup_sem);
2409 fork_main(shm_id, i);
2416 * Wait for the started threads to transition to
2419 printf("fio: Waiting for threads to initialize...\n");
2420 gettimeofday(&this_start, NULL);
2423 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2428 for (i = 0; i < this_jobs; i++) {
2432 if (td->runstate == TD_INITIALIZED) {
2435 } else if (td->runstate >= TD_EXITED) {
2439 nr_running++; /* work-around... */
2445 fprintf(stderr, "fio: %d jobs failed to start\n", left);
2446 for (i = 0; i < this_jobs; i++) {
2450 kill(td->pid, SIGTERM);
2456 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2458 printf("fio: Go for launch\n");
2459 for (i = 0; i < thread_number; i++) {
2462 if (td->runstate != TD_INITIALIZED)
2465 td_set_runstate(td, TD_RUNNING);
2468 m_rate += td->ratemin;
2471 sem_post(&td->mutex);
2474 reap_threads(&nr_running, &t_rate, &m_rate);
2480 while (nr_running) {
2481 reap_threads(&nr_running, &t_rate, &m_rate);
2486 fio_unpin_memory(mlocked_mem);
2489 static void show_group_stats(struct group_run_stats *rs, int id)
2491 printf("\nRun status group %d (all jobs):\n", id);
2493 if (rs->max_run[DDIR_READ])
2494 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]);
2495 if (rs->max_run[DDIR_WRITE])
2496 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]);
2499 static void show_disk_util(void)
2501 struct disk_util_stat *dus;
2502 struct list_head *entry;
2503 struct disk_util *du;
2506 printf("\nDisk stats (read/write):\n");
2508 list_for_each(entry, &disk_list) {
2509 du = list_entry(entry, struct disk_util, list);
2512 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2516 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);
2520 static void show_run_stats(void)
2522 struct group_run_stats *runstats, *rs;
2523 struct thread_data *td;
2526 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2528 for (i = 0; i < groupid + 1; i++) {
2531 memset(rs, 0, sizeof(*rs));
2532 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2533 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2536 for (i = 0; i < thread_number; i++) {
2537 unsigned long long rbw, wbw;
2542 printf("Client%d: %s\n", td->thread_number, td->verror);
2546 rs = &runstats[td->groupid];
2548 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2549 rs->min_run[0] = td->runtime[0];
2550 if (td->runtime[0] > rs->max_run[0])
2551 rs->max_run[0] = td->runtime[0];
2552 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2553 rs->min_run[1] = td->runtime[1];
2554 if (td->runtime[1] > rs->max_run[1])
2555 rs->max_run[1] = td->runtime[1];
2559 rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
2561 wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
2563 if (rbw < rs->min_bw[0])
2564 rs->min_bw[0] = rbw;
2565 if (wbw < rs->min_bw[1])
2566 rs->min_bw[1] = wbw;
2567 if (rbw > rs->max_bw[0])
2568 rs->max_bw[0] = rbw;
2569 if (wbw > rs->max_bw[1])
2570 rs->max_bw[1] = wbw;
2572 rs->io_mb[0] += td->io_bytes[0] >> 20;
2573 rs->io_mb[1] += td->io_bytes[1] >> 20;
2576 for (i = 0; i < groupid + 1; i++) {
2580 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2582 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2586 * don't overwrite last signal output
2590 for (i = 0; i < thread_number; i++) {
2592 rs = &runstats[td->groupid];
2594 show_thread_status(td, rs);
2597 for (i = 0; i < groupid + 1; i++)
2598 show_group_stats(&runstats[i], i);
2603 int main(int argc, char *argv[])
2605 if (parse_options(argc, argv))
2608 if (!thread_number) {
2609 printf("Nothing to do\n");
2613 disk_util_timer_arm();