12 * Change this define to play with the timeout handling
14 #undef FIO_USE_TIMEOUT
16 struct io_completion_data {
19 int error; /* output */
20 unsigned long bytes_done[2]; /* output */
21 struct timeval time; /* output */
25 * The ->file_map[] contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 static int random_map_free(struct thread_data *td, struct fio_file *f,
29 unsigned long long block)
31 unsigned int idx = RAND_MAP_IDX(td, f, block);
32 unsigned int bit = RAND_MAP_BIT(td, f, block);
34 return (f->file_map[idx] & (1UL << bit)) == 0;
38 * Mark a given offset as used in the map.
40 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
42 unsigned int min_bs = td->rw_min_bs;
43 struct fio_file *f = io_u->file;
44 unsigned long long block;
46 unsigned int nr_blocks;
48 block = io_u->offset / (unsigned long long) min_bs;
50 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
52 while (blocks < nr_blocks) {
53 unsigned int idx, bit;
55 if (!random_map_free(td, f, block))
58 idx = RAND_MAP_IDX(td, f, block);
59 bit = RAND_MAP_BIT(td, f, block);
61 fio_assert(td, idx < f->num_maps);
63 f->file_map[idx] |= (1UL << bit);
68 if ((blocks * min_bs) < io_u->buflen)
69 io_u->buflen = blocks * min_bs;
73 * Return the next free block in the map.
75 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
76 unsigned long long *b)
80 i = f->last_free_lookup;
81 *b = (i * BLOCKS_PER_MAP);
82 while ((*b) * td->rw_min_bs < f->real_file_size) {
83 if (f->file_map[i] != -1UL) {
84 *b += ffz(f->file_map[i]);
85 f->last_free_lookup = i;
97 * For random io, generate a random new block and see if it's used. Repeat
98 * until we find a free one. For sequential io, just return the end of
101 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
103 struct fio_file *f = io_u->file;
104 const int ddir = io_u->ddir;
105 unsigned long long b, rb;
109 unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
116 r = os_random_long(&td->random_state);
117 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
120 rb = b + (f->file_offset / td->min_bs[ddir]);
122 } while (!random_map_free(td, f, rb) && loops);
125 * if we failed to retrieve a truly random offset within
126 * the loops assigned, see if there are free ones left at all
128 if (!loops && get_next_free_block(td, f, &b))
131 b = f->last_pos / td->min_bs[ddir];
133 io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
134 if (io_u->offset >= f->real_file_size)
140 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
142 struct fio_file *f = io_u->file;
143 const int ddir = io_u->ddir;
147 if (td->min_bs[ddir] == td->max_bs[ddir])
148 buflen = td->min_bs[ddir];
150 r = os_random_long(&td->bsrange_state);
151 buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
152 if (!td->bs_unaligned)
153 buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
156 while (buflen + io_u->offset > f->real_file_size) {
157 if (buflen == td->min_bs[ddir])
160 buflen = td->min_bs[ddir];
167 * Return the data direction for the next io_u. If the job is a
168 * mixed read/write workload, check the rwmix cycle and switch if
171 static enum fio_ddir get_rw_ddir(struct thread_data *td)
175 unsigned long elapsed;
177 fio_gettime(&now, NULL);
178 elapsed = mtime_since_now(&td->rwmix_switch);
181 * Check if it's time to seed a new data direction.
183 if (elapsed >= td->rwmixcycle) {
187 r = os_random_long(&td->rwmix_state);
188 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
189 if (v < td->rwmixread)
190 td->rwmix_ddir = DDIR_READ;
192 td->rwmix_ddir = DDIR_WRITE;
193 memcpy(&td->rwmix_switch, &now, sizeof(now));
195 return td->rwmix_ddir;
196 } else if (td_read(td))
202 void put_io_u(struct thread_data *td, struct io_u *io_u)
204 assert((io_u->flags & IO_U_F_FREE) == 0);
205 io_u->flags |= IO_U_F_FREE;
208 list_del(&io_u->list);
209 list_add(&io_u->list, &td->io_u_freelist);
213 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
215 struct io_u *__io_u = *io_u;
217 __io_u->flags |= IO_U_F_FREE;
218 __io_u->flags &= ~IO_U_F_FLIGHT;
220 list_del(&__io_u->list);
221 list_add_tail(&__io_u->list, &td->io_u_requeues);
226 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
229 * If using an iolog, grab next piece if any available.
232 return read_iolog_get(td, io_u);
235 * see if it's time to sync
237 if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
238 && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
239 io_u->ddir = DDIR_SYNC;
243 io_u->ddir = get_rw_ddir(td);
246 * No log, let the seq/rand engine retrieve the next buflen and
249 if (get_next_offset(td, io_u))
252 io_u->buflen = get_next_buflen(td, io_u);
257 * mark entry before potentially trimming io_u
259 if (!td->read_iolog && td_random(td) && !td->norandommap)
260 mark_random_map(td, io_u);
263 * If using a write iolog, store this entry.
265 if (td->write_iolog_file)
266 write_iolog_put(td, io_u);
271 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
275 if (io_u->ddir == DDIR_SYNC)
278 switch (td->cur_depth) {
295 td->ts.io_u_map[index]++;
296 td->ts.total_io_u[io_u->ddir]++;
299 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
330 td->ts.io_u_lat[index]++;
334 * Get next file to service by choosing one at random
336 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
343 long r = os_random_long(&td->next_file_state);
345 fno = (unsigned int) ((double) td->nr_files * (r / (RAND_MAX + 1.0)));
348 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
354 * Get next file to service by doing round robin between all available ones
356 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
359 unsigned int old_next_file = td->next_file;
363 f = &td->files[td->next_file];
366 if (td->next_file >= td->nr_files)
369 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
373 } while (td->next_file != old_next_file);
378 static struct fio_file *get_next_file(struct thread_data *td)
382 assert(td->nr_files <= td->files_index);
384 if (!td->nr_open_files)
387 f = td->file_service_file;
388 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
391 if (td->file_service_type == FIO_FSERVICE_RR)
392 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
394 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
396 td->file_service_file = f;
397 td->file_service_left = td->file_service_nr - 1;
401 static struct fio_file *find_next_new_file(struct thread_data *td)
405 if (td->file_service_type == FIO_FSERVICE_RR)
406 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
408 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
413 struct io_u *__get_io_u(struct thread_data *td)
415 struct io_u *io_u = NULL;
417 if (!list_empty(&td->io_u_requeues))
418 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
419 else if (!queue_full(td)) {
420 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
429 assert(io_u->flags & IO_U_F_FREE);
430 io_u->flags &= ~IO_U_F_FREE;
433 list_del(&io_u->list);
434 list_add(&io_u->list, &td->io_u_busylist);
442 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
443 * etc. The returned io_u is fully ready to be prepped and submitted.
445 struct io_u *get_io_u(struct thread_data *td)
451 io_u = __get_io_u(td);
456 * from a requeue, io_u already setup
462 f = get_next_file(td);
471 if (!fill_io_u(td, io_u))
475 * No more to do for this file, close it
478 td_io_close_file(td, f);
481 * probably not the right place to do this, but see
482 * if we need to open a new file
484 if (td->nr_open_files < td->nr_files &&
485 td->open_files != td->nr_files) {
486 f = find_next_new_file(td);
488 if (!f || (ret = td_io_open_file(td, f))) {
496 if (td->zone_bytes >= td->zone_size) {
498 f->last_pos += td->zone_skip;
501 if (io_u->buflen + io_u->offset > f->real_file_size) {
502 if (td->io_ops->flags & FIO_RAWIO) {
507 io_u->buflen = f->real_file_size - io_u->offset;
510 if (io_u->ddir != DDIR_SYNC) {
516 f->last_pos = io_u->offset + io_u->buflen;
518 if (td->verify != VERIFY_NONE)
519 populate_verify_io_u(td, io_u);
523 * Set io data pointers.
526 io_u->xfer_buf = io_u->buf;
527 io_u->xfer_buflen = io_u->buflen;
529 if (td_io_prep(td, io_u)) {
534 fio_gettime(&io_u->start_time, NULL);
538 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
540 const char *msg[] = { "read", "write", "sync" };
542 log_err("fio: io_u error");
545 log_err(" on file %s", io_u->file->file_name);
547 log_err(": %s\n", strerror(io_u->error));
549 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
552 td_verror(td, io_u->error, "io_u error");
555 static void io_completed(struct thread_data *td, struct io_u *io_u,
556 struct io_completion_data *icd)
560 assert(io_u->flags & IO_U_F_FLIGHT);
561 io_u->flags &= ~IO_U_F_FLIGHT;
563 put_file(td, io_u->file);
565 if (io_u->ddir == DDIR_SYNC) {
566 td->last_was_sync = 1;
570 td->last_was_sync = 0;
573 unsigned int bytes = io_u->buflen - io_u->resid;
574 const enum fio_ddir idx = io_u->ddir;
577 td->io_blocks[idx]++;
578 td->io_bytes[idx] += bytes;
579 td->zone_bytes += bytes;
580 td->this_io_bytes[idx] += bytes;
582 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
584 msec = mtime_since(&io_u->issue_time, &icd->time);
586 add_clat_sample(td, idx, msec);
587 add_bw_sample(td, idx, &icd->time);
588 io_u_mark_latency(td, msec);
590 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
591 log_io_piece(td, io_u);
593 icd->bytes_done[idx] += bytes;
596 ret = io_u->end_io(io_u);
597 if (ret && !icd->error)
601 icd->error = io_u->error;
602 io_u_log_error(td, io_u);
606 static void init_icd(struct io_completion_data *icd, int nr)
608 fio_gettime(&icd->time, NULL);
613 icd->bytes_done[0] = icd->bytes_done[1] = 0;
616 static void ios_completed(struct thread_data *td,
617 struct io_completion_data *icd)
622 for (i = 0; i < icd->nr; i++) {
623 io_u = td->io_ops->event(td, i);
625 io_completed(td, io_u, icd);
631 * Complete a single io_u for the sync engines.
633 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
635 struct io_completion_data icd;
638 io_completed(td, io_u, &icd);
642 return icd.bytes_done[0] + icd.bytes_done[1];
644 td_verror(td, icd.error, "io_u_sync_complete");
649 * Called to complete min_events number of io for the async engines.
651 long io_u_queued_complete(struct thread_data *td, int min_events)
653 struct io_completion_data icd;
654 struct timespec *tvp = NULL;
658 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
663 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
665 td_verror(td, -ret, "td_io_getevents");
671 ios_completed(td, &icd);
673 return icd.bytes_done[0] + icd.bytes_done[1];
675 td_verror(td, icd.error, "io_u_queued_complete");
680 * Call when io_u is really queued, to update the submission latency.
682 void io_u_queued(struct thread_data *td, struct io_u *io_u)
684 unsigned long slat_time;
686 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
687 add_slat_sample(td, io_u->ddir, slat_time);
690 #ifdef FIO_USE_TIMEOUT
691 void io_u_set_timeout(struct thread_data *td)
693 assert(td->cur_depth);
695 td->timer.it_interval.tv_sec = 0;
696 td->timer.it_interval.tv_usec = 0;
697 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
698 td->timer.it_value.tv_usec = 0;
699 setitimer(ITIMER_REAL, &td->timer, NULL);
700 fio_gettime(&td->timeout_end, NULL);
703 static void io_u_dump(struct io_u *io_u)
705 unsigned long t_start = mtime_since_now(&io_u->start_time);
706 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
708 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
709 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
710 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
713 void io_u_set_timeout(struct thread_data fio_unused *td)
718 #ifdef FIO_USE_TIMEOUT
719 static void io_u_timeout_handler(int fio_unused sig)
721 struct thread_data *td, *__td;
722 pid_t pid = getpid();
723 struct list_head *entry;
727 log_err("fio: io_u timeout\n");
730 * TLS would be nice...
733 for_each_td(__td, i) {
734 if (__td->pid == pid) {
741 log_err("fio: io_u timeout, can't find job\n");
745 if (!td->cur_depth) {
746 log_err("fio: timeout without pending work?\n");
750 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
752 list_for_each(entry, &td->io_u_busylist) {
753 io_u = list_entry(entry, struct io_u, list);
758 td_verror(td, ETIMEDOUT, "io_u timeout");
763 void io_u_init_timeout(void)
765 #ifdef FIO_USE_TIMEOUT
766 signal(SIGALRM, io_u_timeout_handler);