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
34 #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 struct timeval genesis;
54 static void print_thread_status(void);
56 extern unsigned long long mlock_size;
71 #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
73 static sem_t startup_sem;
75 #define TERMINATE_ALL (-1)
76 #define JOB_START_TIMEOUT (5 * 1000)
78 static void terminate_threads(int group_id)
82 for (i = 0; i < thread_number; i++) {
83 struct thread_data *td = &threads[i];
85 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
92 static void sig_handler(int sig)
97 disk_util_timer_arm();
98 print_thread_status();
101 printf("\nfio: terminating on signal\n");
103 terminate_threads(TERMINATE_ALL);
108 static int random_map_free(struct thread_data *td, unsigned long long block)
110 unsigned int idx = RAND_MAP_IDX(td, block);
111 unsigned int bit = RAND_MAP_BIT(td, block);
113 return (td->file_map[idx] & (1UL << bit)) == 0;
116 static int get_next_free_block(struct thread_data *td, unsigned long long *b)
122 while ((*b) * td->min_bs < td->io_size) {
123 if (td->file_map[i] != -1UL) {
124 *b += ffz(td->file_map[i]);
128 *b += BLOCKS_PER_MAP;
135 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
137 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
138 unsigned int blocks = 0;
140 while (blocks < (io_u->buflen / td->min_bs)) {
141 unsigned int idx, bit;
143 if (!random_map_free(td, block))
146 idx = RAND_MAP_IDX(td, block);
147 bit = RAND_MAP_BIT(td, block);
149 assert(idx < td->num_maps);
151 td->file_map[idx] |= (1UL << bit);
156 if ((blocks * td->min_bs) < io_u->buflen)
157 io_u->buflen = blocks * td->min_bs;
160 static int get_next_offset(struct thread_data *td, unsigned long long *offset)
162 unsigned long long b, rb;
165 if (!td->sequential) {
166 unsigned long long max_blocks = td->io_size / td->min_bs;
170 lrand48_r(&td->random_state, &r);
171 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
172 rb = b + (td->file_offset / td->min_bs);
174 } while (!random_map_free(td, rb) && loops);
177 if (get_next_free_block(td, &b))
181 b = td->last_pos / td->min_bs;
183 *offset = (b * td->min_bs) + td->file_offset;
184 if (*offset > td->real_file_size)
190 static unsigned int get_next_buflen(struct thread_data *td)
195 if (td->min_bs == td->max_bs)
198 lrand48_r(&td->bsrange_state, &r);
199 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
200 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
203 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
204 buflen = td->io_size - td->this_io_bytes[td->ddir];
209 static int check_min_rate(struct thread_data *td, struct timeval *now)
216 * allow a 2 second settle period in the beginning
218 if (mtime_since(&td->start, now) < 2000)
222 * if rate blocks is set, sample is running
224 if (td->rate_bytes) {
225 spent = mtime_since(&td->lastrate, now);
226 if (spent < td->ratecycle)
229 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
230 if (rate < td->ratemin) {
231 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
233 terminate_threads(td->groupid);
238 td->rate_bytes = td->this_io_bytes[ddir];
239 memcpy(&td->lastrate, now, sizeof(*now));
243 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
247 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
253 static void fill_random_bytes(struct thread_data *td,
254 unsigned char *p, unsigned int len)
260 drand48_r(&td->verify_state, &r);
263 * lrand48_r seems to be broken and only fill the bottom
264 * 32-bits, even on 64-bit archs with 64-bit longs
277 static void hexdump(void *buffer, int len)
279 unsigned char *p = buffer;
282 for (i = 0; i < len; i++)
283 printf("%02x", p[i]);
287 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
289 unsigned char *p = (unsigned char *) io_u->buf;
294 c = crc32(p, hdr->len - sizeof(*hdr));
295 ret = c != hdr->crc32;
298 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
299 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
305 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
307 unsigned char *p = (unsigned char *) io_u->buf;
308 struct md5_ctx md5_ctx;
311 memset(&md5_ctx, 0, sizeof(md5_ctx));
313 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
315 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
317 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
318 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
319 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
325 static int verify_io_u(struct io_u *io_u)
327 struct verify_header *hdr = (struct verify_header *) io_u->buf;
330 if (hdr->fio_magic != FIO_HDR_MAGIC)
333 if (hdr->verify_type == VERIFY_MD5)
334 ret = verify_io_u_md5(hdr, io_u);
335 else if (hdr->verify_type == VERIFY_CRC32)
336 ret = verify_io_u_crc32(hdr, io_u);
338 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
345 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
347 hdr->crc32 = crc32(p, len);
350 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
352 struct md5_ctx md5_ctx;
354 memset(&md5_ctx, 0, sizeof(md5_ctx));
355 md5_update(&md5_ctx, p, len);
356 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
359 static int get_rw_ddir(struct thread_data *td)
363 unsigned long elapsed;
365 gettimeofday(&now, NULL);
366 elapsed = mtime_since_now(&td->rwmix_switch);
369 * Check if it's time to seed a new data direction.
371 if (elapsed >= td->rwmixcycle) {
375 lrand48_r(&td->random_state, &r);
376 v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
377 if (v < td->rwmixread)
378 td->rwmix_ddir = DDIR_READ;
380 td->rwmix_ddir = DDIR_WRITE;
381 memcpy(&td->rwmix_switch, &now, sizeof(now));
383 return td->rwmix_ddir;
384 } else if (td_read(td))
391 * fill body of io_u->buf with random data and add a header with the
392 * (eg) sha1sum of that data.
394 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
396 unsigned char *p = (unsigned char *) io_u->buf;
397 struct verify_header hdr;
399 hdr.fio_magic = FIO_HDR_MAGIC;
400 hdr.len = io_u->buflen;
402 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
404 if (td->verify == VERIFY_MD5) {
405 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
406 hdr.verify_type = VERIFY_MD5;
408 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
409 hdr.verify_type = VERIFY_CRC32;
412 memcpy(io_u->buf, &hdr, sizeof(hdr));
415 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
417 if (td->io_prep && td->io_prep(td, io_u))
423 void put_io_u(struct thread_data *td, struct io_u *io_u)
425 list_del(&io_u->list);
426 list_add(&io_u->list, &td->io_u_freelist);
430 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
433 * If using an iolog, grab next piece if any available.
436 return read_iolog_get(td, io_u);
439 * No log, let the seq/rand engine retrieve the next position.
441 if (!get_next_offset(td, &io_u->offset)) {
442 io_u->buflen = get_next_buflen(td);
445 io_u->ddir = get_rw_ddir(td);
448 * If using a write iolog, store this entry.
451 write_iolog_put(td, io_u);
460 #define queue_full(td) (list_empty(&(td)->io_u_freelist))
462 struct io_u *__get_io_u(struct thread_data *td)
469 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
472 list_del(&io_u->list);
473 list_add(&io_u->list, &td->io_u_busylist);
478 static struct io_u *get_io_u(struct thread_data *td)
482 io_u = __get_io_u(td);
486 if (td->zone_bytes >= td->zone_size) {
488 td->last_pos += td->zone_skip;
491 if (fill_io_u(td, io_u)) {
496 if (io_u->buflen + io_u->offset > td->real_file_size)
497 io_u->buflen = td->real_file_size - io_u->offset;
504 if (!td->read_iolog && !td->sequential)
505 mark_random_map(td, io_u);
507 td->last_pos += io_u->buflen;
509 if (td->verify != VERIFY_NONE)
510 populate_io_u(td, io_u);
512 if (td_io_prep(td, io_u)) {
517 gettimeofday(&io_u->start_time, NULL);
521 static inline void td_set_runstate(struct thread_data *td, int runstate)
523 td->old_runstate = td->runstate;
524 td->runstate = runstate;
527 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
529 struct io_piece *ipo;
531 if (list_empty(&td->io_hist_list))
534 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
535 list_del(&ipo->list);
537 io_u->offset = ipo->offset;
538 io_u->buflen = ipo->len;
539 io_u->ddir = DDIR_READ;
544 static int sync_td(struct thread_data *td)
547 return td->io_sync(td);
552 static int io_u_getevents(struct thread_data *td, int min, int max,
555 return td->io_getevents(td, min, max, t);
558 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
560 gettimeofday(&io_u->issue_time, NULL);
562 return td->io_queue(td, io_u);
565 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
567 static void io_completed(struct thread_data *td, struct io_u *io_u,
568 struct io_completion_data *icd)
573 gettimeofday(&e, NULL);
576 unsigned int bytes = io_u->buflen - io_u->resid;
577 const int idx = io_u->ddir;
579 td->io_blocks[idx]++;
580 td->io_bytes[idx] += bytes;
581 td->zone_bytes += bytes;
582 td->this_io_bytes[idx] += bytes;
584 msec = mtime_since(&io_u->issue_time, &e);
586 add_clat_sample(td, idx, msec);
587 add_bw_sample(td, idx);
589 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
590 log_io_piece(td, io_u);
592 icd->bytes_done[idx] += bytes;
594 icd->error = io_u->error;
597 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
603 icd->bytes_done[0] = icd->bytes_done[1] = 0;
605 for (i = 0; i < icd->nr; i++) {
606 io_u = td->io_event(td, i);
608 io_completed(td, io_u, icd);
613 static void cleanup_pending_aio(struct thread_data *td)
615 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
616 struct list_head *entry, *n;
617 struct io_completion_data icd;
622 * get immediately available events, if any
624 r = io_u_getevents(td, 0, td->cur_depth, &ts);
627 ios_completed(td, &icd);
631 * now cancel remaining active events
634 list_for_each_safe(entry, n, &td->io_u_busylist) {
635 io_u = list_entry(entry, struct io_u, list);
637 r = td->io_cancel(td, io_u);
644 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
647 ios_completed(td, &icd);
652 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
654 struct io_u *v_io_u = *io_u;
658 ret = verify_io_u(v_io_u);
659 put_io_u(td, v_io_u);
666 static void do_verify(struct thread_data *td)
669 struct io_u *io_u, *v_io_u = NULL;
670 struct io_completion_data icd;
673 td_set_runstate(td, TD_VERIFYING);
679 gettimeofday(&t, NULL);
680 if (runtime_exceeded(td, &t))
683 io_u = __get_io_u(td);
687 if (get_next_verify(td, io_u)) {
692 if (td_io_prep(td, io_u)) {
697 ret = io_u_queue(td, io_u);
705 * we have one pending to verify, do that while
706 * we are doing io on the next one
708 if (do_io_u_verify(td, &v_io_u))
711 ret = io_u_getevents(td, 1, 1, NULL);
718 v_io_u = td->io_event(td, 0);
721 io_completed(td, v_io_u, &icd);
724 td_verror(td, icd.error);
725 put_io_u(td, v_io_u);
731 * if we can't submit more io, we need to verify now
733 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
738 do_io_u_verify(td, &v_io_u);
741 cleanup_pending_aio(td);
743 td_set_runstate(td, TD_RUNNING);
746 static void do_io(struct thread_data *td)
748 struct io_completion_data icd;
752 while (td->this_io_bytes[td->ddir] < td->io_size) {
753 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
754 struct timespec *timeout;
755 int ret, min_evts = 0;
765 memcpy(&s, &io_u->start_time, sizeof(s));
767 ret = io_u_queue(td, io_u);
774 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
776 if (td->cur_depth < td->iodepth) {
784 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
792 ios_completed(td, &icd);
794 td_verror(td, icd.error);
799 * the rate is batched for now, it should work for batches
800 * of completions except the very first one which may look
803 gettimeofday(&e, NULL);
804 usec = utime_since(&s, &e);
806 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
808 if (check_min_rate(td, &e)) {
809 td_verror(td, ENOMEM);
813 if (runtime_exceeded(td, &e))
817 usec_sleep(td, td->thinktime);
819 if (should_fsync(td) && td->fsync_blocks &&
820 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
825 cleanup_pending_aio(td);
827 if (should_fsync(td) && td->end_fsync)
831 static void cleanup_io(struct thread_data *td)
837 static int init_io(struct thread_data *td)
839 if (td->io_engine == FIO_SYNCIO)
840 return fio_syncio_init(td);
841 else if (td->io_engine == FIO_MMAPIO)
842 return fio_mmapio_init(td);
843 else if (td->io_engine == FIO_LIBAIO)
844 return fio_libaio_init(td);
845 else if (td->io_engine == FIO_POSIXAIO)
846 return fio_posixaio_init(td);
847 else if (td->io_engine == FIO_SGIO)
848 return fio_sgio_init(td);
849 else if (td->io_engine == FIO_SPLICEIO)
850 return fio_spliceio_init(td);
852 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
857 static void cleanup_io_u(struct thread_data *td)
859 struct list_head *entry, *n;
862 list_for_each_safe(entry, n, &td->io_u_freelist) {
863 io_u = list_entry(entry, struct io_u, list);
865 list_del(&io_u->list);
869 if (td->mem_type == MEM_MALLOC)
870 free(td->orig_buffer);
871 else if (td->mem_type == MEM_SHM) {
872 struct shmid_ds sbuf;
874 shmdt(td->orig_buffer);
875 shmctl(td->shm_id, IPC_RMID, &sbuf);
876 } else if (td->mem_type == MEM_MMAP)
877 munmap(td->orig_buffer, td->orig_buffer_size);
879 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
881 td->orig_buffer = NULL;
884 static int init_io_u(struct thread_data *td)
890 if (td->io_engine & FIO_SYNCIO)
893 max_units = td->iodepth;
895 td->orig_buffer_size = td->max_bs * max_units + MASK;
897 if (td->mem_type == MEM_MALLOC)
898 td->orig_buffer = malloc(td->orig_buffer_size);
899 else if (td->mem_type == MEM_SHM) {
900 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
901 if (td->shm_id < 0) {
902 td_verror(td, errno);
907 td->orig_buffer = shmat(td->shm_id, NULL, 0);
908 if (td->orig_buffer == (void *) -1) {
909 td_verror(td, errno);
911 td->orig_buffer = NULL;
914 } else if (td->mem_type == MEM_MMAP) {
915 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
916 if (td->orig_buffer == MAP_FAILED) {
917 td_verror(td, errno);
919 td->orig_buffer = NULL;
924 p = ALIGN(td->orig_buffer);
925 for (i = 0; i < max_units; i++) {
926 io_u = malloc(sizeof(*io_u));
927 memset(io_u, 0, sizeof(*io_u));
928 INIT_LIST_HEAD(&io_u->list);
930 io_u->buf = p + td->max_bs * i;
932 list_add(&io_u->list, &td->io_u_freelist);
938 static void cleanup_allocs(struct thread_data *td)
943 free(td->iolog_file);
945 free(td->exec_prerun);
946 if (td->exec_postrun)
947 free(td->exec_postrun);
949 free(td->ioscheduler);
951 free(td->sysfs_root);
954 static int create_file(struct thread_data *td, unsigned long long size,
957 unsigned long long left;
963 * unless specifically asked for overwrite, let normal io extend it
965 if (td_write(td) && !td->overwrite)
969 fprintf(stderr, "Need size for create\n");
970 td_verror(td, EINVAL);
975 oflags = O_CREAT | O_TRUNC;
976 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
979 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
982 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
984 td_verror(td, errno);
988 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
989 td_verror(td, errno);
993 td->io_size = td->file_size;
994 b = malloc(td->max_bs);
995 memset(b, 0, td->max_bs);
998 while (left && !td->terminate) {
1003 r = write(td->fd, b, bs);
1005 if (r == (int) bs) {
1010 td_verror(td, errno);
1019 unlink(td->file_name);
1020 else if (td->create_fsync)
1029 static int file_size(struct thread_data *td)
1033 if (fstat(td->fd, &st) == -1) {
1034 td_verror(td, errno);
1038 td->real_file_size = st.st_size;
1040 if (!td->file_size || td->file_size > td->real_file_size)
1041 td->file_size = td->real_file_size;
1043 td->file_size -= td->file_offset;
1047 static int bdev_size(struct thread_data *td)
1049 unsigned long long bytes;
1052 r = blockdev_size(td->fd, &bytes);
1058 td->real_file_size = bytes;
1061 * no extend possibilities, so limit size to device size if too large
1063 if (!td->file_size || td->file_size > td->real_file_size)
1064 td->file_size = td->real_file_size;
1066 td->file_size -= td->file_offset;
1070 static int get_file_size(struct thread_data *td)
1074 if (td->filetype == FIO_TYPE_FILE)
1075 ret = file_size(td);
1076 else if (td->filetype == FIO_TYPE_BD)
1077 ret = bdev_size(td);
1079 td->real_file_size = -1;
1084 if (td->file_offset > td->real_file_size) {
1085 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1089 td->io_size = td->file_size;
1090 if (td->io_size == 0) {
1091 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1092 td_verror(td, EINVAL);
1097 td->zone_size = td->io_size;
1099 td->total_io_size = td->io_size * td->loops;
1103 static int setup_file_mmap(struct thread_data *td)
1108 flags = PROT_READ | PROT_WRITE;
1109 else if (td_write(td)) {
1112 if (td->verify != VERIFY_NONE)
1117 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1118 if (td->mmap == MAP_FAILED) {
1120 td_verror(td, errno);
1124 if (td->invalidate_cache) {
1125 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1126 td_verror(td, errno);
1131 if (td->sequential) {
1132 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1133 td_verror(td, errno);
1137 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1138 td_verror(td, errno);
1146 static int setup_file_plain(struct thread_data *td)
1148 if (td->invalidate_cache) {
1149 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1150 td_verror(td, errno);
1155 if (td->sequential) {
1156 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1157 td_verror(td, errno);
1161 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1162 td_verror(td, errno);
1170 static int setup_file(struct thread_data *td)
1175 if (stat(td->file_name, &st) == -1) {
1176 if (errno != ENOENT) {
1177 td_verror(td, errno);
1180 if (!td->create_file) {
1181 td_verror(td, ENOENT);
1184 if (create_file(td, td->file_size, 0))
1186 } else if (td->filetype == FIO_TYPE_FILE) {
1187 if (st.st_size < (off_t) td->file_size) {
1188 if (create_file(td, td->file_size - st.st_size, 1))
1196 if (td_write(td) || td_rw(td)) {
1197 if (td->filetype == FIO_TYPE_FILE) {
1208 td->fd = open(td->file_name, flags, 0600);
1210 if (td->filetype == FIO_TYPE_CHAR)
1215 td->fd = open(td->file_name, flags);
1219 td_verror(td, errno);
1223 if (get_file_size(td))
1226 if (td->io_engine != FIO_MMAPIO)
1227 return setup_file_plain(td);
1229 return setup_file_mmap(td);
1232 static int switch_ioscheduler(struct thread_data *td)
1234 char tmp[256], tmp2[128];
1238 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1240 f = fopen(tmp, "r+");
1242 td_verror(td, errno);
1249 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1250 if (ferror(f) || ret != 1) {
1251 td_verror(td, errno);
1259 * Read back and check that the selected scheduler is now the default.
1261 ret = fread(tmp, 1, sizeof(tmp), f);
1262 if (ferror(f) || ret < 0) {
1263 td_verror(td, errno);
1268 sprintf(tmp2, "[%s]", td->ioscheduler);
1269 if (!strstr(tmp, tmp2)) {
1270 fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
1271 td_verror(td, EINVAL);
1280 static void clear_io_state(struct thread_data *td)
1282 if (td->io_engine == FIO_SYNCIO)
1283 lseek(td->fd, SEEK_SET, 0);
1286 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1287 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1291 memset(td->file_map, 0, td->num_maps * sizeof(long));
1294 static void *thread_main(void *data)
1296 struct thread_data *td = data;
1298 if (!td->use_thread)
1303 INIT_LIST_HEAD(&td->io_u_freelist);
1304 INIT_LIST_HEAD(&td->io_u_busylist);
1305 INIT_LIST_HEAD(&td->io_hist_list);
1306 INIT_LIST_HEAD(&td->io_log_list);
1311 if (fio_setaffinity(td) == -1) {
1312 td_verror(td, errno);
1323 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1324 td_verror(td, errno);
1329 if (nice(td->nice) < 0) {
1330 td_verror(td, errno);
1334 if (init_random_state(td))
1337 if (td->ioscheduler && switch_ioscheduler(td))
1340 td_set_runstate(td, TD_INITIALIZED);
1341 sem_post(&startup_sem);
1342 sem_wait(&td->mutex);
1344 if (!td->create_serialize && setup_file(td))
1347 gettimeofday(&td->epoch, NULL);
1349 if (td->exec_prerun)
1350 system(td->exec_prerun);
1352 while (td->loops--) {
1353 getrusage(RUSAGE_SELF, &td->ru_start);
1354 gettimeofday(&td->start, NULL);
1355 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1358 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1361 prune_io_piece_log(td);
1365 td->runtime[td->ddir] += mtime_since_now(&td->start);
1366 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1367 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1369 update_rusage_stat(td);
1371 if (td->error || td->terminate)
1374 if (td->verify == VERIFY_NONE)
1378 gettimeofday(&td->start, NULL);
1382 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1384 if (td->error || td->terminate)
1389 finish_log(td, td->bw_log, "bw");
1391 finish_log(td, td->slat_log, "slat");
1393 finish_log(td, td->clat_log, "clat");
1394 if (td->write_iolog)
1395 write_iolog_close(td);
1396 if (td->exec_postrun)
1397 system(td->exec_postrun);
1399 if (exitall_on_terminate)
1400 terminate_threads(td->groupid);
1408 munmap(td->mmap, td->file_size);
1412 td_set_runstate(td, TD_EXITED);
1417 static void *fork_main(int shmid, int offset)
1419 struct thread_data *td;
1422 data = shmat(shmid, NULL, 0);
1423 if (data == (void *) -1) {
1428 td = data + offset * sizeof(struct thread_data);
1434 static void check_str_update(struct thread_data *td)
1436 char c = run_str[td->thread_number - 1];
1438 if (td->runstate == td->old_runstate)
1441 switch (td->runstate) {
1454 } else if (td_read(td)) {
1472 case TD_INITIALIZED:
1475 case TD_NOT_CREATED:
1479 printf("state %d\n", td->runstate);
1482 run_str[td->thread_number - 1] = c;
1483 td->old_runstate = td->runstate;
1486 static void eta_to_str(char *str, int eta_sec)
1488 unsigned int d, h, m, s;
1489 static int always_d, always_h;
1501 if (d || always_d) {
1503 str += sprintf(str, "%02dd:", d);
1505 if (h || always_h) {
1507 str += sprintf(str, "%02dh:", h);
1510 str += sprintf(str, "%02dm:", m);
1511 str += sprintf(str, "%02ds", s);
1514 static int thread_eta(struct thread_data *td, unsigned long elapsed)
1516 unsigned long long bytes_total, bytes_done;
1517 unsigned int eta_sec = 0;
1519 bytes_total = td->total_io_size;
1522 * if writing, bytes_total will be twice the size. If mixing,
1523 * assume a 50/50 split and thus bytes_total will be 50% larger.
1527 bytes_total = bytes_total * 3 / 2;
1531 if (td->zone_size && td->zone_skip)
1532 bytes_total /= (td->zone_skip / td->zone_size);
1534 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1537 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1538 perc = (double) bytes_done / (double) bytes_total;
1542 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1544 if (td->timeout && eta_sec > (td->timeout - elapsed))
1545 eta_sec = td->timeout - elapsed;
1546 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1547 || td->runstate == TD_INITIALIZED) {
1548 int t_eta = 0, r_eta = 0;
1551 * We can only guess - assume it'll run the full timeout
1552 * if given, otherwise assume it'll run at the specified rate.
1555 t_eta = td->timeout + td->start_delay - elapsed;
1557 r_eta = (bytes_total / 1024) / td->rate;
1558 r_eta += td->start_delay - elapsed;
1562 eta_sec = min(r_eta, t_eta);
1571 * thread is already done
1579 static void print_thread_status(void)
1581 unsigned long elapsed = time_since_now(&genesis);
1582 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
1586 eta_secs = malloc(thread_number * sizeof(int));
1587 memset(eta_secs, 0, thread_number * sizeof(int));
1589 nr_running = t_rate = m_rate = 0;
1590 for (i = 0; i < thread_number; i++) {
1591 struct thread_data *td = &threads[i];
1593 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1596 m_rate += td->ratemin;
1600 eta_secs[i] = thread_eta(td, elapsed);
1602 eta_secs[i] = INT_MAX;
1604 check_str_update(td);
1607 if (exitall_on_terminate)
1612 for (i = 0; i < thread_number; i++) {
1613 if (exitall_on_terminate) {
1614 if (eta_secs[i] < eta_sec)
1615 eta_sec = eta_secs[i];
1617 if (eta_secs[i] > eta_sec)
1618 eta_sec = eta_secs[i];
1622 if (eta_sec != INT_MAX && elapsed) {
1623 perc = (double) elapsed / (double) (elapsed + eta_sec);
1624 eta_to_str(eta_str, eta_sec);
1627 printf("Threads now running (%d)", nr_running);
1628 if (m_rate || t_rate)
1629 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1630 if (eta_sec != INT_MAX) {
1632 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1639 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1644 * reap exited threads (TD_EXITED -> TD_REAPED)
1646 for (i = 0; i < thread_number; i++) {
1647 struct thread_data *td = &threads[i];
1649 if (td->runstate != TD_EXITED)
1652 td_set_runstate(td, TD_REAPED);
1654 if (td->use_thread) {
1657 if (pthread_join(td->thread, (void *) &ret))
1658 perror("thread_join");
1660 waitpid(td->pid, NULL, 0);
1663 (*m_rate) -= td->ratemin;
1664 (*t_rate) -= td->rate;
1668 static void fio_unpin_memory(void *pinned)
1671 if (munlock(pinned, mlock_size) < 0)
1673 munmap(pinned, mlock_size);
1677 static void *fio_pin_memory(void)
1679 long pagesize, pages;
1686 * Don't allow mlock of more than real_mem-128MB
1688 pagesize = sysconf(_SC_PAGESIZE);
1689 pages = sysconf(_SC_PHYS_PAGES);
1690 if (pages != -1 && pagesize != -1) {
1691 unsigned long long real_mem = pages * pagesize;
1693 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
1694 mlock_size = real_mem - 128 * 1024 * 1024;
1695 printf("fio: limiting mlocked memory to %lluMiB\n",
1700 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1702 perror("malloc locked mem");
1705 if (mlock(ptr, mlock_size) < 0) {
1706 munmap(ptr, mlock_size);
1714 static void run_threads(void)
1716 struct thread_data *td;
1717 unsigned long spent;
1718 int i, todo, nr_running, m_rate, t_rate, nr_started;
1721 mlocked_mem = fio_pin_memory();
1723 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1726 signal(SIGINT, sig_handler);
1727 signal(SIGALRM, sig_handler);
1729 todo = thread_number;
1732 m_rate = t_rate = 0;
1734 for (i = 0; i < thread_number; i++) {
1737 run_str[td->thread_number - 1] = 'P';
1741 if (!td->create_serialize)
1745 * do file setup here so it happens sequentially,
1746 * we don't want X number of threads getting their
1747 * client data interspersed on disk
1749 if (setup_file(td)) {
1750 td_set_runstate(td, TD_REAPED);
1755 gettimeofday(&genesis, NULL);
1758 struct thread_data *map[MAX_JOBS];
1759 struct timeval this_start;
1760 int this_jobs = 0, left;
1763 * create threads (TD_NOT_CREATED -> TD_CREATED)
1765 for (i = 0; i < thread_number; i++) {
1768 if (td->runstate != TD_NOT_CREATED)
1772 * never got a chance to start, killed by other
1773 * thread for some reason
1775 if (td->terminate) {
1780 if (td->start_delay) {
1781 spent = mtime_since_now(&genesis);
1783 if (td->start_delay * 1000 > spent)
1787 if (td->stonewall && (nr_started || nr_running))
1791 * Set state to created. Thread will transition
1792 * to TD_INITIALIZED when it's done setting up.
1794 td_set_runstate(td, TD_CREATED);
1795 map[this_jobs++] = td;
1796 sem_init(&startup_sem, 0, 1);
1799 if (td->use_thread) {
1800 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1801 perror("thread_create");
1806 sem_wait(&startup_sem);
1808 fork_main(shm_id, i);
1815 * Wait for the started threads to transition to
1818 printf("fio: Waiting for threads to initialize...\n");
1819 gettimeofday(&this_start, NULL);
1822 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1827 for (i = 0; i < this_jobs; i++) {
1831 if (td->runstate == TD_INITIALIZED) {
1834 } else if (td->runstate >= TD_EXITED) {
1838 nr_running++; /* work-around... */
1844 fprintf(stderr, "fio: %d jobs failed to start\n", left);
1845 for (i = 0; i < this_jobs; i++) {
1849 kill(td->pid, SIGTERM);
1855 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1857 printf("fio: Go for launch\n");
1858 for (i = 0; i < thread_number; i++) {
1861 if (td->runstate != TD_INITIALIZED)
1864 td_set_runstate(td, TD_RUNNING);
1867 m_rate += td->ratemin;
1870 sem_post(&td->mutex);
1873 reap_threads(&nr_running, &t_rate, &m_rate);
1879 while (nr_running) {
1880 reap_threads(&nr_running, &t_rate, &m_rate);
1885 fio_unpin_memory(mlocked_mem);
1888 int main(int argc, char *argv[])
1890 if (parse_options(argc, argv))
1893 if (!thread_number) {
1894 printf("Nothing to do\n");
1898 disk_util_timer_arm();