2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
32 #include <sys/ioctl.h>
40 #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
43 int thread_number = 0;
44 static char run_str[MAX_JOBS + 1];
46 static struct timeval genesis;
48 char *fio_inst_prefix = _INST_PREFIX;
50 static void print_thread_status(void);
52 extern unsigned long long mlock_size;
55 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
56 * will never back again. It may cycle between running/verififying/fsyncing.
57 * Once the thread reaches TD_EXITED, it is just waiting for the core to
71 #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
73 static volatile int 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);
109 * The ->file_map[] contains a map of blocks we have or have not done io
110 * to yet. Used to make sure we cover the entire range in a fair fashion.
112 static int random_map_free(struct thread_data *td, struct fio_file *f,
113 unsigned long long block)
115 unsigned int idx = RAND_MAP_IDX(td, f, block);
116 unsigned int bit = RAND_MAP_BIT(td, f, block);
118 return (f->file_map[idx] & (1UL << bit)) == 0;
122 * Return the next free block in the map.
124 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
125 unsigned long long *b)
131 while ((*b) * td->min_bs < f->file_size) {
132 if (f->file_map[i] != -1UL) {
133 *b += ffz(f->file_map[i]);
137 *b += BLOCKS_PER_MAP;
145 * Mark a given offset as used in the map.
147 static void mark_random_map(struct thread_data *td, struct fio_file *f,
150 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
151 unsigned int blocks = 0;
153 while (blocks < (io_u->buflen / td->min_bs)) {
154 unsigned int idx, bit;
156 if (!random_map_free(td, f, block))
159 idx = RAND_MAP_IDX(td, f, block);
160 bit = RAND_MAP_BIT(td, f, block);
162 assert(idx < f->num_maps);
164 f->file_map[idx] |= (1UL << bit);
169 if ((blocks * td->min_bs) < io_u->buflen)
170 io_u->buflen = blocks * td->min_bs;
174 * For random io, generate a random new block and see if it's used. Repeat
175 * until we find a free one. For sequential io, just return the end of
176 * the last io issued.
178 static int get_next_offset(struct thread_data *td, struct fio_file *f,
179 unsigned long long *offset)
181 unsigned long long b, rb;
184 if (!td->sequential) {
185 unsigned long long max_blocks = td->io_size / td->min_bs;
189 r = os_random_long(&td->random_state);
190 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
191 rb = b + (f->file_offset / td->min_bs);
193 } while (!random_map_free(td, f, rb) && loops);
196 if (get_next_free_block(td, f, &b))
200 b = f->last_pos / td->min_bs;
202 *offset = (b * td->min_bs) + f->file_offset;
203 if (*offset > f->file_size)
209 static unsigned int get_next_buflen(struct thread_data *td)
214 if (td->min_bs == td->max_bs)
217 r = os_random_long(&td->bsrange_state);
218 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
219 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
222 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
223 buflen = td->io_size - td->this_io_bytes[td->ddir];
229 * Check if we are above the minimum rate given.
231 static int check_min_rate(struct thread_data *td, struct timeval *now)
238 * allow a 2 second settle period in the beginning
240 if (mtime_since(&td->start, now) < 2000)
244 * if rate blocks is set, sample is running
246 if (td->rate_bytes) {
247 spent = mtime_since(&td->lastrate, now);
248 if (spent < td->ratecycle)
251 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
252 if (rate < td->ratemin) {
253 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
255 terminate_threads(td->groupid);
260 td->rate_bytes = td->this_io_bytes[ddir];
261 memcpy(&td->lastrate, now, sizeof(*now));
265 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
269 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
275 static void fill_random_bytes(struct thread_data *td,
276 unsigned char *p, unsigned int len)
282 r = os_random_double(&td->verify_state);
285 * lrand48_r seems to be broken and only fill the bottom
286 * 32-bits, even on 64-bit archs with 64-bit longs
299 static void hexdump(void *buffer, int len)
301 unsigned char *p = buffer;
304 for (i = 0; i < len; i++)
305 fprintf(f_out, "%02x", p[i]);
306 fprintf(f_out, "\n");
309 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
311 unsigned char *p = (unsigned char *) io_u->buf;
315 c = crc32(p, hdr->len - sizeof(*hdr));
317 if (c != hdr->crc32) {
318 log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
319 log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c);
326 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
328 unsigned char *p = (unsigned char *) io_u->buf;
329 struct md5_ctx md5_ctx;
331 memset(&md5_ctx, 0, sizeof(md5_ctx));
333 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
335 if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
336 log_err("md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
337 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
338 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
345 static int verify_io_u(struct io_u *io_u)
347 struct verify_header *hdr = (struct verify_header *) io_u->buf;
350 if (hdr->fio_magic != FIO_HDR_MAGIC)
353 if (hdr->verify_type == VERIFY_MD5)
354 ret = verify_io_u_md5(hdr, io_u);
355 else if (hdr->verify_type == VERIFY_CRC32)
356 ret = verify_io_u_crc32(hdr, io_u);
358 log_err("Bad verify type %d\n", hdr->verify_type);
365 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
367 hdr->crc32 = crc32(p, len);
370 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
372 struct md5_ctx md5_ctx;
374 memset(&md5_ctx, 0, sizeof(md5_ctx));
375 md5_update(&md5_ctx, p, len);
376 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
380 * Return the data direction for the next io_u. If the job is a
381 * mixed read/write workload, check the rwmix cycle and switch if
384 static int get_rw_ddir(struct thread_data *td)
388 unsigned long elapsed;
390 gettimeofday(&now, NULL);
391 elapsed = mtime_since_now(&td->rwmix_switch);
394 * Check if it's time to seed a new data direction.
396 if (elapsed >= td->rwmixcycle) {
400 r = os_random_long(&td->rwmix_state);
401 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
402 if (v < td->rwmixread)
403 td->rwmix_ddir = DDIR_READ;
405 td->rwmix_ddir = DDIR_WRITE;
406 memcpy(&td->rwmix_switch, &now, sizeof(now));
408 return td->rwmix_ddir;
409 } else if (td_read(td))
416 * fill body of io_u->buf with random data and add a header with the
417 * crc32 or md5 sum of that data.
419 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
421 unsigned char *p = (unsigned char *) io_u->buf;
422 struct verify_header hdr;
424 hdr.fio_magic = FIO_HDR_MAGIC;
425 hdr.len = io_u->buflen;
427 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
429 if (td->verify == VERIFY_MD5) {
430 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
431 hdr.verify_type = VERIFY_MD5;
433 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
434 hdr.verify_type = VERIFY_CRC32;
437 memcpy(io_u->buf, &hdr, sizeof(hdr));
440 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
442 if (td->io_ops->prep && td->io_ops->prep(td, io_u))
448 void put_io_u(struct thread_data *td, struct io_u *io_u)
451 list_del(&io_u->list);
452 list_add(&io_u->list, &td->io_u_freelist);
456 static int fill_io_u(struct thread_data *td, struct fio_file *f,
460 * If using an iolog, grab next piece if any available.
463 return read_iolog_get(td, io_u);
466 * No log, let the seq/rand engine retrieve the next position.
468 if (!get_next_offset(td, f, &io_u->offset)) {
469 io_u->buflen = get_next_buflen(td);
472 io_u->ddir = get_rw_ddir(td);
475 * If using a write iolog, store this entry.
478 write_iolog_put(td, io_u);
488 #define queue_full(td) list_empty(&(td)->io_u_freelist)
490 struct io_u *__get_io_u(struct thread_data *td)
492 struct io_u *io_u = NULL;
494 if (!queue_full(td)) {
495 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
499 list_del(&io_u->list);
500 list_add(&io_u->list, &td->io_u_busylist);
508 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
509 * etc. The returned io_u is fully ready to be prepped and submitted.
511 static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
515 io_u = __get_io_u(td);
519 if (td->zone_bytes >= td->zone_size) {
521 f->last_pos += td->zone_skip;
524 if (fill_io_u(td, f, io_u)) {
529 if (io_u->buflen + io_u->offset > f->file_size)
530 io_u->buflen = f->file_size - io_u->offset;
537 if (!td->read_iolog && !td->sequential)
538 mark_random_map(td, f, io_u);
540 f->last_pos += io_u->buflen;
542 if (td->verify != VERIFY_NONE)
543 populate_io_u(td, io_u);
545 if (td_io_prep(td, io_u)) {
550 gettimeofday(&io_u->start_time, NULL);
554 static inline void td_set_runstate(struct thread_data *td, int runstate)
556 td->runstate = runstate;
559 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
561 struct io_piece *ipo;
563 if (!list_empty(&td->io_hist_list)) {
564 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
566 list_del(&ipo->list);
568 io_u->offset = ipo->offset;
569 io_u->buflen = ipo->len;
570 io_u->ddir = DDIR_READ;
578 static struct fio_file *get_next_file(struct thread_data *td)
580 struct fio_file *f = &td->files[td->next_file];
583 if (td->next_file >= td->nr_files)
589 static int td_io_sync(struct thread_data *td, struct fio_file *f)
591 if (td->io_ops->sync)
592 return td->io_ops->sync(td, f);
597 static int io_u_getevents(struct thread_data *td, int min, int max,
600 return td->io_ops->getevents(td, min, max, t);
603 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
605 gettimeofday(&io_u->issue_time, NULL);
607 return td->io_ops->queue(td, io_u);
610 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
612 static void io_completed(struct thread_data *td, struct io_u *io_u,
613 struct io_completion_data *icd)
618 gettimeofday(&e, NULL);
621 unsigned int bytes = io_u->buflen - io_u->resid;
622 const int idx = io_u->ddir;
624 td->io_blocks[idx]++;
625 td->io_bytes[idx] += bytes;
626 td->zone_bytes += bytes;
627 td->this_io_bytes[idx] += bytes;
629 msec = mtime_since(&io_u->issue_time, &e);
631 add_clat_sample(td, idx, msec);
632 add_bw_sample(td, idx);
634 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
635 log_io_piece(td, io_u);
637 icd->bytes_done[idx] += bytes;
639 icd->error = io_u->error;
642 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
648 icd->bytes_done[0] = icd->bytes_done[1] = 0;
650 for (i = 0; i < icd->nr; i++) {
651 io_u = td->io_ops->event(td, i);
653 io_completed(td, io_u, icd);
659 * When job exits, we can cancel the in-flight IO if we are using async
660 * io. Attempt to do so.
662 static void cleanup_pending_aio(struct thread_data *td)
664 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
665 struct list_head *entry, *n;
666 struct io_completion_data icd;
671 * get immediately available events, if any
673 r = io_u_getevents(td, 0, td->cur_depth, &ts);
676 ios_completed(td, &icd);
680 * now cancel remaining active events
682 if (td->io_ops->cancel) {
683 list_for_each_safe(entry, n, &td->io_u_busylist) {
684 io_u = list_entry(entry, struct io_u, list);
686 r = td->io_ops->cancel(td, io_u);
693 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
696 ios_completed(td, &icd);
701 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
703 struct io_u *v_io_u = *io_u;
707 ret = verify_io_u(v_io_u);
708 put_io_u(td, v_io_u);
716 * The main verify engine. Runs over the writes we previusly submitted,
717 * reads the blocks back in, and checks the crc/md5 of the data.
719 static void do_verify(struct thread_data *td)
722 struct io_u *io_u, *v_io_u = NULL;
723 struct io_completion_data icd;
727 td_set_runstate(td, TD_VERIFYING);
733 gettimeofday(&t, NULL);
734 if (runtime_exceeded(td, &t))
737 io_u = __get_io_u(td);
741 if (get_next_verify(td, io_u)) {
746 f = get_next_file(td);
752 if (td_io_prep(td, io_u)) {
757 ret = io_u_queue(td, io_u);
765 * we have one pending to verify, do that while
766 * we are doing io on the next one
768 if (do_io_u_verify(td, &v_io_u))
771 ret = io_u_getevents(td, 1, 1, NULL);
778 v_io_u = td->io_ops->event(td, 0);
781 io_completed(td, v_io_u, &icd);
784 td_verror(td, icd.error);
785 put_io_u(td, v_io_u);
791 * if we can't submit more io, we need to verify now
793 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
798 do_io_u_verify(td, &v_io_u);
801 cleanup_pending_aio(td);
803 td_set_runstate(td, TD_RUNNING);
807 * Not really an io thread, all it does is burn CPU cycles in the specified
810 static void do_cpuio(struct thread_data *td)
813 int split = 100 / td->cpuload;
816 while (!td->terminate) {
817 gettimeofday(&e, NULL);
819 if (runtime_exceeded(td, &e))
825 usec_sleep(td, 10000);
832 * Main IO worker function. It retrieves io_u's to process and queues
833 * and reaps them, checking for rate and errors along the way.
835 static void do_io(struct thread_data *td)
837 struct io_completion_data icd;
843 td_set_runstate(td, TD_RUNNING);
845 while (td->this_io_bytes[td->ddir] < td->io_size) {
846 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
847 struct timespec *timeout;
848 int ret, min_evts = 0;
854 f = get_next_file(td);
858 io_u = get_io_u(td, f);
862 memcpy(&s, &io_u->start_time, sizeof(s));
864 ret = io_u_queue(td, io_u);
871 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
873 if (td->cur_depth < td->iodepth) {
881 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
889 ios_completed(td, &icd);
891 td_verror(td, icd.error);
896 * the rate is batched for now, it should work for batches
897 * of completions except the very first one which may look
900 gettimeofday(&e, NULL);
901 usec = utime_since(&s, &e);
903 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
905 if (check_min_rate(td, &e)) {
906 td_verror(td, ENOMEM);
910 if (runtime_exceeded(td, &e))
914 usec_sleep(td, td->thinktime);
916 if (should_fsync(td) && td->fsync_blocks &&
917 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
922 cleanup_pending_aio(td);
924 if (should_fsync(td) && td->end_fsync) {
925 td_set_runstate(td, TD_FSYNCING);
926 for_each_file(td, f, i)
931 static int init_io(struct thread_data *td)
933 if (td->io_ops->init)
934 return td->io_ops->init(td);
939 static void cleanup_io_u(struct thread_data *td)
941 struct list_head *entry, *n;
944 list_for_each_safe(entry, n, &td->io_u_freelist) {
945 io_u = list_entry(entry, struct io_u, list);
947 list_del(&io_u->list);
951 if (td->mem_type == MEM_MALLOC)
952 free(td->orig_buffer);
953 else if (td->mem_type == MEM_SHM) {
954 struct shmid_ds sbuf;
956 shmdt(td->orig_buffer);
957 shmctl(td->shm_id, IPC_RMID, &sbuf);
958 } else if (td->mem_type == MEM_MMAP)
959 munmap(td->orig_buffer, td->orig_buffer_size);
961 log_err("Bad memory type %d\n", td->mem_type);
963 td->orig_buffer = NULL;
966 static int init_io_u(struct thread_data *td)
972 if (td->io_ops->flags & FIO_CPUIO)
975 if (td->io_ops->flags & FIO_SYNCIO)
978 max_units = td->iodepth;
980 td->orig_buffer_size = td->max_bs * max_units + MASK;
982 if (td->mem_type == MEM_MALLOC)
983 td->orig_buffer = malloc(td->orig_buffer_size);
984 else if (td->mem_type == MEM_SHM) {
985 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
986 if (td->shm_id < 0) {
987 td_verror(td, errno);
992 td->orig_buffer = shmat(td->shm_id, NULL, 0);
993 if (td->orig_buffer == (void *) -1) {
994 td_verror(td, errno);
996 td->orig_buffer = NULL;
999 } else if (td->mem_type == MEM_MMAP) {
1000 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1001 if (td->orig_buffer == MAP_FAILED) {
1002 td_verror(td, errno);
1004 td->orig_buffer = NULL;
1009 p = ALIGN(td->orig_buffer);
1010 for (i = 0; i < max_units; i++) {
1011 io_u = malloc(sizeof(*io_u));
1012 memset(io_u, 0, sizeof(*io_u));
1013 INIT_LIST_HEAD(&io_u->list);
1015 io_u->buf = p + td->max_bs * i;
1017 list_add(&io_u->list, &td->io_u_freelist);
1023 static int switch_ioscheduler(struct thread_data *td)
1025 char tmp[256], tmp2[128];
1029 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1031 f = fopen(tmp, "r+");
1033 td_verror(td, errno);
1040 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1041 if (ferror(f) || ret != 1) {
1042 td_verror(td, errno);
1050 * Read back and check that the selected scheduler is now the default.
1052 ret = fread(tmp, 1, sizeof(tmp), f);
1053 if (ferror(f) || ret < 0) {
1054 td_verror(td, errno);
1059 sprintf(tmp2, "[%s]", td->ioscheduler);
1060 if (!strstr(tmp, tmp2)) {
1061 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
1062 td_verror(td, EINVAL);
1071 static void clear_io_state(struct thread_data *td)
1076 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1077 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1080 for_each_file(td, f, i) {
1082 if (td->io_ops->flags & FIO_SYNCIO)
1083 lseek(f->fd, SEEK_SET, 0);
1086 memset(f->file_map, 0, f->num_maps * sizeof(long));
1091 * Entry point for the thread based jobs. The process based jobs end up
1092 * here as well, after a little setup.
1094 static void *thread_main(void *data)
1096 struct thread_data *td = data;
1098 if (!td->use_thread)
1103 INIT_LIST_HEAD(&td->io_u_freelist);
1104 INIT_LIST_HEAD(&td->io_u_busylist);
1105 INIT_LIST_HEAD(&td->io_hist_list);
1106 INIT_LIST_HEAD(&td->io_log_list);
1111 if (fio_setaffinity(td) == -1) {
1112 td_verror(td, errno);
1123 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1124 td_verror(td, errno);
1129 if (nice(td->nice) == -1) {
1130 td_verror(td, errno);
1134 if (init_random_state(td))
1137 if (td->ioscheduler && switch_ioscheduler(td))
1140 td_set_runstate(td, TD_INITIALIZED);
1141 fio_sem_up(&startup_sem);
1142 fio_sem_down(&td->mutex);
1144 if (!td->create_serialize && setup_files(td))
1147 gettimeofday(&td->epoch, NULL);
1149 if (td->exec_prerun)
1150 system(td->exec_prerun);
1152 while (td->loops--) {
1153 getrusage(RUSAGE_SELF, &td->ru_start);
1154 gettimeofday(&td->start, NULL);
1155 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1158 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1161 prune_io_piece_log(td);
1163 if (td->io_ops->flags & FIO_CPUIO)
1168 td->runtime[td->ddir] += mtime_since_now(&td->start);
1169 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1170 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1172 update_rusage_stat(td);
1174 if (td->error || td->terminate)
1177 if (td->verify == VERIFY_NONE)
1181 gettimeofday(&td->start, NULL);
1185 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1187 if (td->error || td->terminate)
1192 finish_log(td, td->bw_log, "bw");
1194 finish_log(td, td->slat_log, "slat");
1196 finish_log(td, td->clat_log, "clat");
1197 if (td->write_iolog)
1198 write_iolog_close(td);
1199 if (td->exec_postrun)
1200 system(td->exec_postrun);
1202 if (exitall_on_terminate)
1203 terminate_threads(td->groupid);
1209 td_set_runstate(td, TD_EXITED);
1215 * We cannot pass the td data into a forked process, so attach the td and
1216 * pass it to the thread worker.
1218 static void *fork_main(int shmid, int offset)
1220 struct thread_data *td;
1223 data = shmat(shmid, NULL, 0);
1224 if (data == (void *) -1) {
1229 td = data + offset * sizeof(struct thread_data);
1236 * Sets the status of the 'td' in the printed status map.
1238 static void check_str_update(struct thread_data *td)
1240 char c = run_str[td->thread_number - 1];
1242 switch (td->runstate) {
1255 } else if (td_read(td)) {
1276 case TD_INITIALIZED:
1279 case TD_NOT_CREATED:
1283 log_err("state %d\n", td->runstate);
1286 run_str[td->thread_number - 1] = c;
1290 * Convert seconds to a printable string.
1292 static void eta_to_str(char *str, int eta_sec)
1294 unsigned int d, h, m, s;
1295 static int always_d, always_h;
1307 if (d || always_d) {
1309 str += sprintf(str, "%02dd:", d);
1311 if (h || always_h) {
1313 str += sprintf(str, "%02dh:", h);
1316 str += sprintf(str, "%02dm:", m);
1317 str += sprintf(str, "%02ds", s);
1321 * Best effort calculation of the estimated pending runtime of a job.
1323 static int thread_eta(struct thread_data *td, unsigned long elapsed)
1325 unsigned long long bytes_total, bytes_done;
1326 unsigned int eta_sec = 0;
1328 bytes_total = td->total_io_size;
1331 * if writing, bytes_total will be twice the size. If mixing,
1332 * assume a 50/50 split and thus bytes_total will be 50% larger.
1336 bytes_total = bytes_total * 3 / 2;
1340 if (td->zone_size && td->zone_skip)
1341 bytes_total /= (td->zone_skip / td->zone_size);
1343 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1346 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1347 perc = (double) bytes_done / (double) bytes_total;
1351 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1353 if (td->timeout && eta_sec > (td->timeout - elapsed))
1354 eta_sec = td->timeout - elapsed;
1355 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1356 || td->runstate == TD_INITIALIZED) {
1357 int t_eta = 0, r_eta = 0;
1360 * We can only guess - assume it'll run the full timeout
1361 * if given, otherwise assume it'll run at the specified rate.
1364 t_eta = td->timeout + td->start_delay - elapsed;
1366 r_eta = (bytes_total / 1024) / td->rate;
1367 r_eta += td->start_delay - elapsed;
1371 eta_sec = min(r_eta, t_eta);
1380 * thread is already done or waiting for fsync
1389 * Print status of the jobs we know about. This includes rate estimates,
1390 * ETA, thread state, etc.
1392 static void print_thread_status(void)
1394 unsigned long elapsed = time_since_now(&genesis);
1395 int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec;
1399 if (temp_stall_ts || terse_output)
1402 eta_secs = malloc(thread_number * sizeof(int));
1403 memset(eta_secs, 0, thread_number * sizeof(int));
1405 nr_pending = nr_running = t_rate = m_rate = 0;
1406 for (i = 0; i < thread_number; i++) {
1407 struct thread_data *td = &threads[i];
1409 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING||
1410 td->runstate == TD_FSYNCING) {
1413 m_rate += td->ratemin;
1414 } else if (td->runstate < TD_RUNNING)
1418 eta_secs[i] = thread_eta(td, elapsed);
1420 eta_secs[i] = INT_MAX;
1422 check_str_update(td);
1425 if (exitall_on_terminate)
1430 for (i = 0; i < thread_number; i++) {
1431 if (exitall_on_terminate) {
1432 if (eta_secs[i] < eta_sec)
1433 eta_sec = eta_secs[i];
1435 if (eta_secs[i] > eta_sec)
1436 eta_sec = eta_secs[i];
1440 if (eta_sec != INT_MAX && elapsed) {
1441 perc = (double) elapsed / (double) (elapsed + eta_sec);
1442 eta_to_str(eta_str, eta_sec);
1445 if (!nr_running && !nr_pending)
1448 printf("Threads running: %d", nr_running);
1449 if (m_rate || t_rate)
1450 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1451 if (eta_sec != INT_MAX && nr_running) {
1453 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1461 * Run over the job map and reap the threads that have exited, if any.
1463 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1468 * reap exited threads (TD_EXITED -> TD_REAPED)
1470 for (i = 0, cputhreads = 0; i < thread_number; i++) {
1471 struct thread_data *td = &threads[i];
1473 if (td->io_ops->flags & FIO_CPUIO)
1476 if (td->runstate != TD_EXITED)
1479 td_set_runstate(td, TD_REAPED);
1481 if (td->use_thread) {
1484 if (pthread_join(td->thread, (void *) &ret))
1485 perror("thread_join");
1487 waitpid(td->pid, NULL, 0);
1490 (*m_rate) -= td->ratemin;
1491 (*t_rate) -= td->rate;
1494 if (*nr_running == cputhreads)
1495 terminate_threads(TERMINATE_ALL);
1498 static void fio_unpin_memory(void *pinned)
1501 if (munlock(pinned, mlock_size) < 0)
1503 munmap(pinned, mlock_size);
1507 static void *fio_pin_memory(void)
1509 unsigned long long phys_mem;
1516 * Don't allow mlock of more than real_mem-128MB
1518 phys_mem = os_phys_mem();
1520 if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
1521 mlock_size = phys_mem - 128 * 1024 * 1024;
1522 fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
1526 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1528 perror("malloc locked mem");
1531 if (mlock(ptr, mlock_size) < 0) {
1532 munmap(ptr, mlock_size);
1541 * Main function for kicking off and reaping jobs, as needed.
1543 static void run_threads(void)
1545 struct thread_data *td;
1546 unsigned long spent;
1547 int i, todo, nr_running, m_rate, t_rate, nr_started;
1550 mlocked_mem = fio_pin_memory();
1552 if (!terse_output) {
1553 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1557 signal(SIGINT, sig_handler);
1558 signal(SIGALRM, sig_handler);
1560 todo = thread_number;
1563 m_rate = t_rate = 0;
1565 for (i = 0; i < thread_number; i++) {
1568 run_str[td->thread_number - 1] = 'P';
1572 if (!td->create_serialize)
1576 * do file setup here so it happens sequentially,
1577 * we don't want X number of threads getting their
1578 * client data interspersed on disk
1580 if (setup_files(td)) {
1581 td_set_runstate(td, TD_REAPED);
1586 gettimeofday(&genesis, NULL);
1589 struct thread_data *map[MAX_JOBS];
1590 struct timeval this_start;
1591 int this_jobs = 0, left;
1594 * create threads (TD_NOT_CREATED -> TD_CREATED)
1596 for (i = 0; i < thread_number; i++) {
1599 if (td->runstate != TD_NOT_CREATED)
1603 * never got a chance to start, killed by other
1604 * thread for some reason
1606 if (td->terminate) {
1611 if (td->start_delay) {
1612 spent = mtime_since_now(&genesis);
1614 if (td->start_delay * 1000 > spent)
1618 if (td->stonewall && (nr_started || nr_running))
1622 * Set state to created. Thread will transition
1623 * to TD_INITIALIZED when it's done setting up.
1625 td_set_runstate(td, TD_CREATED);
1626 map[this_jobs++] = td;
1627 fio_sem_init(&startup_sem, 1);
1630 if (td->use_thread) {
1631 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1632 perror("thread_create");
1637 fio_sem_down(&startup_sem);
1639 fork_main(shm_id, i);
1646 * Wait for the started threads to transition to
1649 gettimeofday(&this_start, NULL);
1652 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1657 for (i = 0; i < this_jobs; i++) {
1661 if (td->runstate == TD_INITIALIZED) {
1664 } else if (td->runstate >= TD_EXITED) {
1668 nr_running++; /* work-around... */
1674 log_err("fio: %d jobs failed to start\n", left);
1675 for (i = 0; i < this_jobs; i++) {
1679 kill(td->pid, SIGTERM);
1685 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1687 for (i = 0; i < thread_number; i++) {
1690 if (td->runstate != TD_INITIALIZED)
1693 td_set_runstate(td, TD_RUNNING);
1696 m_rate += td->ratemin;
1699 fio_sem_up(&td->mutex);
1702 reap_threads(&nr_running, &t_rate, &m_rate);
1708 while (nr_running) {
1709 reap_threads(&nr_running, &t_rate, &m_rate);
1714 fio_unpin_memory(mlocked_mem);
1717 int main(int argc, char *argv[])
1719 if (parse_options(argc, argv))
1722 if (!thread_number) {
1723 log_err("Nothing to do\n");
1727 disk_util_timer_arm();