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]) {
224 * if using direct/raw io, we may not be able to
225 * shrink the size. so just fail it.
227 if (td->io_ops->flags & FIO_RAWIO)
230 buflen = td->io_size - td->this_io_bytes[td->ddir];
237 * Check if we are above the minimum rate given.
239 static int check_min_rate(struct thread_data *td, struct timeval *now)
246 * allow a 2 second settle period in the beginning
248 if (mtime_since(&td->start, now) < 2000)
252 * if rate blocks is set, sample is running
254 if (td->rate_bytes) {
255 spent = mtime_since(&td->lastrate, now);
256 if (spent < td->ratecycle)
259 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
260 if (rate < td->ratemin) {
261 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
263 terminate_threads(td->groupid);
268 td->rate_bytes = td->this_io_bytes[ddir];
269 memcpy(&td->lastrate, now, sizeof(*now));
273 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
277 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
283 static void fill_random_bytes(struct thread_data *td,
284 unsigned char *p, unsigned int len)
290 r = os_random_double(&td->verify_state);
293 * lrand48_r seems to be broken and only fill the bottom
294 * 32-bits, even on 64-bit archs with 64-bit longs
307 static void hexdump(void *buffer, int len)
309 unsigned char *p = buffer;
312 for (i = 0; i < len; i++)
313 fprintf(f_out, "%02x", p[i]);
314 fprintf(f_out, "\n");
317 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
319 unsigned char *p = (unsigned char *) io_u->buf;
323 c = crc32(p, hdr->len - sizeof(*hdr));
325 if (c != hdr->crc32) {
326 log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
327 log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c);
334 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
336 unsigned char *p = (unsigned char *) io_u->buf;
337 struct md5_ctx md5_ctx;
339 memset(&md5_ctx, 0, sizeof(md5_ctx));
341 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
343 if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
344 log_err("md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
345 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
346 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
353 static int verify_io_u(struct io_u *io_u)
355 struct verify_header *hdr = (struct verify_header *) io_u->buf;
358 if (hdr->fio_magic != FIO_HDR_MAGIC)
361 if (hdr->verify_type == VERIFY_MD5)
362 ret = verify_io_u_md5(hdr, io_u);
363 else if (hdr->verify_type == VERIFY_CRC32)
364 ret = verify_io_u_crc32(hdr, io_u);
366 log_err("Bad verify type %d\n", hdr->verify_type);
373 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
375 hdr->crc32 = crc32(p, len);
378 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
380 struct md5_ctx md5_ctx;
382 memset(&md5_ctx, 0, sizeof(md5_ctx));
383 md5_update(&md5_ctx, p, len);
384 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
388 * Return the data direction for the next io_u. If the job is a
389 * mixed read/write workload, check the rwmix cycle and switch if
392 static int get_rw_ddir(struct thread_data *td)
396 unsigned long elapsed;
398 gettimeofday(&now, NULL);
399 elapsed = mtime_since_now(&td->rwmix_switch);
402 * Check if it's time to seed a new data direction.
404 if (elapsed >= td->rwmixcycle) {
408 r = os_random_long(&td->rwmix_state);
409 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
410 if (v < td->rwmixread)
411 td->rwmix_ddir = DDIR_READ;
413 td->rwmix_ddir = DDIR_WRITE;
414 memcpy(&td->rwmix_switch, &now, sizeof(now));
416 return td->rwmix_ddir;
417 } else if (td_read(td))
424 * fill body of io_u->buf with random data and add a header with the
425 * crc32 or md5 sum of that data.
427 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
429 unsigned char *p = (unsigned char *) io_u->buf;
430 struct verify_header hdr;
432 hdr.fio_magic = FIO_HDR_MAGIC;
433 hdr.len = io_u->buflen;
435 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
437 if (td->verify == VERIFY_MD5) {
438 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
439 hdr.verify_type = VERIFY_MD5;
441 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
442 hdr.verify_type = VERIFY_CRC32;
445 memcpy(io_u->buf, &hdr, sizeof(hdr));
448 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
450 if (td->io_ops->prep && td->io_ops->prep(td, io_u))
456 void put_io_u(struct thread_data *td, struct io_u *io_u)
459 list_del(&io_u->list);
460 list_add(&io_u->list, &td->io_u_freelist);
464 static int fill_io_u(struct thread_data *td, struct fio_file *f,
468 * If using an iolog, grab next piece if any available.
471 return read_iolog_get(td, io_u);
474 * No log, let the seq/rand engine retrieve the next position.
476 if (!get_next_offset(td, f, &io_u->offset)) {
477 io_u->buflen = get_next_buflen(td);
480 io_u->ddir = get_rw_ddir(td);
483 * If using a write iolog, store this entry.
486 write_iolog_put(td, io_u);
496 #define queue_full(td) list_empty(&(td)->io_u_freelist)
498 struct io_u *__get_io_u(struct thread_data *td)
500 struct io_u *io_u = NULL;
502 if (!queue_full(td)) {
503 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
507 list_del(&io_u->list);
508 list_add(&io_u->list, &td->io_u_busylist);
516 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
517 * etc. The returned io_u is fully ready to be prepped and submitted.
519 static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
523 io_u = __get_io_u(td);
527 if (td->zone_bytes >= td->zone_size) {
529 f->last_pos += td->zone_skip;
532 if (fill_io_u(td, f, io_u)) {
537 if (io_u->buflen + io_u->offset > f->file_size) {
538 if (td->io_ops->flags & FIO_RAWIO) {
543 io_u->buflen = f->file_size - io_u->offset;
551 if (!td->read_iolog && !td->sequential)
552 mark_random_map(td, f, io_u);
554 f->last_pos += io_u->buflen;
556 if (td->verify != VERIFY_NONE)
557 populate_io_u(td, io_u);
559 if (td_io_prep(td, io_u)) {
564 gettimeofday(&io_u->start_time, NULL);
568 static inline void td_set_runstate(struct thread_data *td, int runstate)
570 td->runstate = runstate;
573 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
575 struct io_piece *ipo;
577 if (!list_empty(&td->io_hist_list)) {
578 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
580 list_del(&ipo->list);
582 io_u->offset = ipo->offset;
583 io_u->buflen = ipo->len;
584 io_u->ddir = DDIR_READ;
592 static struct fio_file *get_next_file(struct thread_data *td)
594 int old_next_file = td->next_file;
598 f = &td->files[td->next_file];
601 if (td->next_file >= td->nr_files)
608 } while (td->next_file != old_next_file);
613 static int td_io_sync(struct thread_data *td, struct fio_file *f)
615 if (td->io_ops->sync)
616 return td->io_ops->sync(td, f);
621 static int io_u_getevents(struct thread_data *td, int min, int max,
624 return td->io_ops->getevents(td, min, max, t);
627 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
629 gettimeofday(&io_u->issue_time, NULL);
631 return td->io_ops->queue(td, io_u);
634 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
636 static void io_completed(struct thread_data *td, struct io_u *io_u,
637 struct io_completion_data *icd)
642 gettimeofday(&e, NULL);
645 unsigned int bytes = io_u->buflen - io_u->resid;
646 const int idx = io_u->ddir;
648 td->io_blocks[idx]++;
649 td->io_bytes[idx] += bytes;
650 td->zone_bytes += bytes;
651 td->this_io_bytes[idx] += bytes;
653 msec = mtime_since(&io_u->issue_time, &e);
655 add_clat_sample(td, idx, msec);
656 add_bw_sample(td, idx);
658 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
659 log_io_piece(td, io_u);
661 icd->bytes_done[idx] += bytes;
663 icd->error = io_u->error;
666 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
672 icd->bytes_done[0] = icd->bytes_done[1] = 0;
674 for (i = 0; i < icd->nr; i++) {
675 io_u = td->io_ops->event(td, i);
677 io_completed(td, io_u, icd);
683 * When job exits, we can cancel the in-flight IO if we are using async
684 * io. Attempt to do so.
686 static void cleanup_pending_aio(struct thread_data *td)
688 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
689 struct list_head *entry, *n;
690 struct io_completion_data icd;
695 * get immediately available events, if any
697 r = io_u_getevents(td, 0, td->cur_depth, &ts);
700 ios_completed(td, &icd);
704 * now cancel remaining active events
706 if (td->io_ops->cancel) {
707 list_for_each_safe(entry, n, &td->io_u_busylist) {
708 io_u = list_entry(entry, struct io_u, list);
710 r = td->io_ops->cancel(td, io_u);
717 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
720 ios_completed(td, &icd);
725 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
727 struct io_u *v_io_u = *io_u;
731 ret = verify_io_u(v_io_u);
732 put_io_u(td, v_io_u);
740 * The main verify engine. Runs over the writes we previusly submitted,
741 * reads the blocks back in, and checks the crc/md5 of the data.
743 static void do_verify(struct thread_data *td)
746 struct io_u *io_u, *v_io_u = NULL;
747 struct io_completion_data icd;
751 td_set_runstate(td, TD_VERIFYING);
757 gettimeofday(&t, NULL);
758 if (runtime_exceeded(td, &t))
761 io_u = __get_io_u(td);
765 if (get_next_verify(td, io_u)) {
770 f = get_next_file(td);
776 if (td_io_prep(td, io_u)) {
781 ret = io_u_queue(td, io_u);
789 * we have one pending to verify, do that while
790 * we are doing io on the next one
792 if (do_io_u_verify(td, &v_io_u))
795 ret = io_u_getevents(td, 1, 1, NULL);
802 v_io_u = td->io_ops->event(td, 0);
805 io_completed(td, v_io_u, &icd);
808 td_verror(td, icd.error);
809 put_io_u(td, v_io_u);
815 * if we can't submit more io, we need to verify now
817 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
822 do_io_u_verify(td, &v_io_u);
825 cleanup_pending_aio(td);
827 td_set_runstate(td, TD_RUNNING);
831 * Not really an io thread, all it does is burn CPU cycles in the specified
834 static void do_cpuio(struct thread_data *td)
837 int split = 100 / td->cpuload;
840 while (!td->terminate) {
841 gettimeofday(&e, NULL);
843 if (runtime_exceeded(td, &e))
849 usec_sleep(td, 10000);
856 * Main IO worker function. It retrieves io_u's to process and queues
857 * and reaps them, checking for rate and errors along the way.
859 static void do_io(struct thread_data *td)
861 struct io_completion_data icd;
867 td_set_runstate(td, TD_RUNNING);
869 while (td->this_io_bytes[td->ddir] < td->io_size) {
870 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
871 struct timespec *timeout;
872 int ret, min_evts = 0;
878 f = get_next_file(td);
882 io_u = get_io_u(td, f);
886 memcpy(&s, &io_u->start_time, sizeof(s));
888 ret = io_u_queue(td, io_u);
895 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
897 if (td->cur_depth < td->iodepth) {
905 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
913 ios_completed(td, &icd);
915 td_verror(td, icd.error);
920 * the rate is batched for now, it should work for batches
921 * of completions except the very first one which may look
924 gettimeofday(&e, NULL);
925 usec = utime_since(&s, &e);
927 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
929 if (check_min_rate(td, &e)) {
930 td_verror(td, ENOMEM);
934 if (runtime_exceeded(td, &e))
938 usec_sleep(td, td->thinktime);
940 if (should_fsync(td) && td->fsync_blocks &&
941 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
946 cleanup_pending_aio(td);
948 if (should_fsync(td) && td->end_fsync) {
949 td_set_runstate(td, TD_FSYNCING);
950 for_each_file(td, f, i)
955 static int init_io(struct thread_data *td)
957 if (td->io_ops->init)
958 return td->io_ops->init(td);
963 static void cleanup_io_u(struct thread_data *td)
965 struct list_head *entry, *n;
968 list_for_each_safe(entry, n, &td->io_u_freelist) {
969 io_u = list_entry(entry, struct io_u, list);
971 list_del(&io_u->list);
975 if (td->mem_type == MEM_MALLOC)
976 free(td->orig_buffer);
977 else if (td->mem_type == MEM_SHM) {
978 struct shmid_ds sbuf;
980 shmdt(td->orig_buffer);
981 shmctl(td->shm_id, IPC_RMID, &sbuf);
982 } else if (td->mem_type == MEM_MMAP)
983 munmap(td->orig_buffer, td->orig_buffer_size);
985 log_err("Bad memory type %d\n", td->mem_type);
987 td->orig_buffer = NULL;
990 static int init_io_u(struct thread_data *td)
996 if (td->io_ops->flags & FIO_CPUIO)
999 if (td->io_ops->flags & FIO_SYNCIO)
1002 max_units = td->iodepth;
1004 td->orig_buffer_size = td->max_bs * max_units + MASK;
1006 if (td->mem_type == MEM_MALLOC)
1007 td->orig_buffer = malloc(td->orig_buffer_size);
1008 else if (td->mem_type == MEM_SHM) {
1009 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1010 if (td->shm_id < 0) {
1011 td_verror(td, errno);
1016 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1017 if (td->orig_buffer == (void *) -1) {
1018 td_verror(td, errno);
1020 td->orig_buffer = NULL;
1023 } else if (td->mem_type == MEM_MMAP) {
1024 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1025 if (td->orig_buffer == MAP_FAILED) {
1026 td_verror(td, errno);
1028 td->orig_buffer = NULL;
1033 p = ALIGN(td->orig_buffer);
1034 for (i = 0; i < max_units; i++) {
1035 io_u = malloc(sizeof(*io_u));
1036 memset(io_u, 0, sizeof(*io_u));
1037 INIT_LIST_HEAD(&io_u->list);
1039 io_u->buf = p + td->max_bs * i;
1041 list_add(&io_u->list, &td->io_u_freelist);
1047 static int switch_ioscheduler(struct thread_data *td)
1049 char tmp[256], tmp2[128];
1053 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1055 f = fopen(tmp, "r+");
1057 td_verror(td, errno);
1064 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1065 if (ferror(f) || ret != 1) {
1066 td_verror(td, errno);
1074 * Read back and check that the selected scheduler is now the default.
1076 ret = fread(tmp, 1, sizeof(tmp), f);
1077 if (ferror(f) || ret < 0) {
1078 td_verror(td, errno);
1083 sprintf(tmp2, "[%s]", td->ioscheduler);
1084 if (!strstr(tmp, tmp2)) {
1085 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
1086 td_verror(td, EINVAL);
1095 static void clear_io_state(struct thread_data *td)
1100 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1101 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1104 for_each_file(td, f, i) {
1106 if (td->io_ops->flags & FIO_SYNCIO)
1107 lseek(f->fd, SEEK_SET, 0);
1110 memset(f->file_map, 0, f->num_maps * sizeof(long));
1115 * Entry point for the thread based jobs. The process based jobs end up
1116 * here as well, after a little setup.
1118 static void *thread_main(void *data)
1120 struct thread_data *td = data;
1122 if (!td->use_thread)
1127 INIT_LIST_HEAD(&td->io_u_freelist);
1128 INIT_LIST_HEAD(&td->io_u_busylist);
1129 INIT_LIST_HEAD(&td->io_hist_list);
1130 INIT_LIST_HEAD(&td->io_log_list);
1135 if (fio_setaffinity(td) == -1) {
1136 td_verror(td, errno);
1147 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1148 td_verror(td, errno);
1153 if (nice(td->nice) == -1) {
1154 td_verror(td, errno);
1158 if (init_random_state(td))
1161 if (td->ioscheduler && switch_ioscheduler(td))
1164 td_set_runstate(td, TD_INITIALIZED);
1165 fio_sem_up(&startup_sem);
1166 fio_sem_down(&td->mutex);
1168 if (!td->create_serialize && setup_files(td))
1171 gettimeofday(&td->epoch, NULL);
1173 if (td->exec_prerun)
1174 system(td->exec_prerun);
1176 while (td->loops--) {
1177 getrusage(RUSAGE_SELF, &td->ru_start);
1178 gettimeofday(&td->start, NULL);
1179 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1182 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1185 prune_io_piece_log(td);
1187 if (td->io_ops->flags & FIO_CPUIO)
1192 td->runtime[td->ddir] += mtime_since_now(&td->start);
1193 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1194 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1196 update_rusage_stat(td);
1198 if (td->error || td->terminate)
1201 if (td->verify == VERIFY_NONE)
1205 gettimeofday(&td->start, NULL);
1209 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1211 if (td->error || td->terminate)
1216 finish_log(td, td->bw_log, "bw");
1218 finish_log(td, td->slat_log, "slat");
1220 finish_log(td, td->clat_log, "clat");
1221 if (td->write_iolog)
1222 write_iolog_close(td);
1223 if (td->exec_postrun)
1224 system(td->exec_postrun);
1226 if (exitall_on_terminate)
1227 terminate_threads(td->groupid);
1233 td_set_runstate(td, TD_EXITED);
1239 * We cannot pass the td data into a forked process, so attach the td and
1240 * pass it to the thread worker.
1242 static void *fork_main(int shmid, int offset)
1244 struct thread_data *td;
1247 data = shmat(shmid, NULL, 0);
1248 if (data == (void *) -1) {
1253 td = data + offset * sizeof(struct thread_data);
1260 * Sets the status of the 'td' in the printed status map.
1262 static void check_str_update(struct thread_data *td)
1264 char c = run_str[td->thread_number - 1];
1266 switch (td->runstate) {
1279 } else if (td_read(td)) {
1300 case TD_INITIALIZED:
1303 case TD_NOT_CREATED:
1307 log_err("state %d\n", td->runstate);
1310 run_str[td->thread_number - 1] = c;
1314 * Convert seconds to a printable string.
1316 static void eta_to_str(char *str, int eta_sec)
1318 unsigned int d, h, m, s;
1319 static int always_d, always_h;
1331 if (d || always_d) {
1333 str += sprintf(str, "%02dd:", d);
1335 if (h || always_h) {
1337 str += sprintf(str, "%02dh:", h);
1340 str += sprintf(str, "%02dm:", m);
1341 str += sprintf(str, "%02ds", s);
1345 * Best effort calculation of the estimated pending runtime of a job.
1347 static int thread_eta(struct thread_data *td, unsigned long elapsed)
1349 unsigned long long bytes_total, bytes_done;
1350 unsigned int eta_sec = 0;
1352 bytes_total = td->total_io_size;
1355 * if writing, bytes_total will be twice the size. If mixing,
1356 * assume a 50/50 split and thus bytes_total will be 50% larger.
1360 bytes_total = bytes_total * 3 / 2;
1364 if (td->zone_size && td->zone_skip)
1365 bytes_total /= (td->zone_skip / td->zone_size);
1367 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1370 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1371 perc = (double) bytes_done / (double) bytes_total;
1375 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1377 if (td->timeout && eta_sec > (td->timeout - elapsed))
1378 eta_sec = td->timeout - elapsed;
1379 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1380 || td->runstate == TD_INITIALIZED) {
1381 int t_eta = 0, r_eta = 0;
1384 * We can only guess - assume it'll run the full timeout
1385 * if given, otherwise assume it'll run at the specified rate.
1388 t_eta = td->timeout + td->start_delay - elapsed;
1390 r_eta = (bytes_total / 1024) / td->rate;
1391 r_eta += td->start_delay - elapsed;
1395 eta_sec = min(r_eta, t_eta);
1404 * thread is already done or waiting for fsync
1413 * Print status of the jobs we know about. This includes rate estimates,
1414 * ETA, thread state, etc.
1416 static void print_thread_status(void)
1418 unsigned long elapsed = time_since_now(&genesis);
1419 int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec;
1423 if (temp_stall_ts || terse_output)
1426 eta_secs = malloc(thread_number * sizeof(int));
1427 memset(eta_secs, 0, thread_number * sizeof(int));
1429 nr_pending = nr_running = t_rate = m_rate = 0;
1430 for (i = 0; i < thread_number; i++) {
1431 struct thread_data *td = &threads[i];
1433 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING||
1434 td->runstate == TD_FSYNCING) {
1437 m_rate += td->ratemin;
1438 } else if (td->runstate < TD_RUNNING)
1442 eta_secs[i] = thread_eta(td, elapsed);
1444 eta_secs[i] = INT_MAX;
1446 check_str_update(td);
1449 if (exitall_on_terminate)
1454 for (i = 0; i < thread_number; i++) {
1455 if (exitall_on_terminate) {
1456 if (eta_secs[i] < eta_sec)
1457 eta_sec = eta_secs[i];
1459 if (eta_secs[i] > eta_sec)
1460 eta_sec = eta_secs[i];
1464 if (eta_sec != INT_MAX && elapsed) {
1465 perc = (double) elapsed / (double) (elapsed + eta_sec);
1466 eta_to_str(eta_str, eta_sec);
1469 if (!nr_running && !nr_pending)
1472 printf("Threads running: %d", nr_running);
1473 if (m_rate || t_rate)
1474 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1475 if (eta_sec != INT_MAX && nr_running) {
1477 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1485 * Run over the job map and reap the threads that have exited, if any.
1487 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1492 * reap exited threads (TD_EXITED -> TD_REAPED)
1494 for (i = 0, cputhreads = 0; i < thread_number; i++) {
1495 struct thread_data *td = &threads[i];
1497 if (td->io_ops->flags & FIO_CPUIO)
1500 if (td->runstate != TD_EXITED)
1503 td_set_runstate(td, TD_REAPED);
1505 if (td->use_thread) {
1508 if (pthread_join(td->thread, (void *) &ret))
1509 perror("thread_join");
1511 waitpid(td->pid, NULL, 0);
1514 (*m_rate) -= td->ratemin;
1515 (*t_rate) -= td->rate;
1518 if (*nr_running == cputhreads)
1519 terminate_threads(TERMINATE_ALL);
1522 static void fio_unpin_memory(void *pinned)
1525 if (munlock(pinned, mlock_size) < 0)
1527 munmap(pinned, mlock_size);
1531 static void *fio_pin_memory(void)
1533 unsigned long long phys_mem;
1540 * Don't allow mlock of more than real_mem-128MB
1542 phys_mem = os_phys_mem();
1544 if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
1545 mlock_size = phys_mem - 128 * 1024 * 1024;
1546 fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
1550 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1552 perror("malloc locked mem");
1555 if (mlock(ptr, mlock_size) < 0) {
1556 munmap(ptr, mlock_size);
1565 * Main function for kicking off and reaping jobs, as needed.
1567 static void run_threads(void)
1569 struct thread_data *td;
1570 unsigned long spent;
1571 int i, todo, nr_running, m_rate, t_rate, nr_started;
1574 mlocked_mem = fio_pin_memory();
1576 if (!terse_output) {
1577 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1581 signal(SIGINT, sig_handler);
1582 signal(SIGALRM, sig_handler);
1584 todo = thread_number;
1587 m_rate = t_rate = 0;
1589 for (i = 0; i < thread_number; i++) {
1592 run_str[td->thread_number - 1] = 'P';
1596 if (!td->create_serialize)
1600 * do file setup here so it happens sequentially,
1601 * we don't want X number of threads getting their
1602 * client data interspersed on disk
1604 if (setup_files(td)) {
1605 td_set_runstate(td, TD_REAPED);
1610 gettimeofday(&genesis, NULL);
1613 struct thread_data *map[MAX_JOBS];
1614 struct timeval this_start;
1615 int this_jobs = 0, left;
1618 * create threads (TD_NOT_CREATED -> TD_CREATED)
1620 for (i = 0; i < thread_number; i++) {
1623 if (td->runstate != TD_NOT_CREATED)
1627 * never got a chance to start, killed by other
1628 * thread for some reason
1630 if (td->terminate) {
1635 if (td->start_delay) {
1636 spent = mtime_since_now(&genesis);
1638 if (td->start_delay * 1000 > spent)
1642 if (td->stonewall && (nr_started || nr_running))
1646 * Set state to created. Thread will transition
1647 * to TD_INITIALIZED when it's done setting up.
1649 td_set_runstate(td, TD_CREATED);
1650 map[this_jobs++] = td;
1651 fio_sem_init(&startup_sem, 1);
1654 if (td->use_thread) {
1655 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1656 perror("thread_create");
1661 fio_sem_down(&startup_sem);
1663 fork_main(shm_id, i);
1670 * Wait for the started threads to transition to
1673 gettimeofday(&this_start, NULL);
1676 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1681 for (i = 0; i < this_jobs; i++) {
1685 if (td->runstate == TD_INITIALIZED) {
1688 } else if (td->runstate >= TD_EXITED) {
1692 nr_running++; /* work-around... */
1698 log_err("fio: %d jobs failed to start\n", left);
1699 for (i = 0; i < this_jobs; i++) {
1703 kill(td->pid, SIGTERM);
1709 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1711 for (i = 0; i < thread_number; i++) {
1714 if (td->runstate != TD_INITIALIZED)
1717 td_set_runstate(td, TD_RUNNING);
1720 m_rate += td->ratemin;
1723 fio_sem_up(&td->mutex);
1726 reap_threads(&nr_running, &t_rate, &m_rate);
1732 while (nr_running) {
1733 reap_threads(&nr_running, &t_rate, &m_rate);
1738 fio_unpin_memory(mlocked_mem);
1741 int main(int argc, char *argv[])
1743 if (parse_options(argc, argv))
1746 if (!thread_number) {
1747 log_err("Nothing to do\n");
1751 disk_util_timer_arm();