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->o.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->o.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->o.min_bs[ddir];
113 r = os_random_long(&td->random_state);
117 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
118 if (td->o.norandommap)
120 rb = b + (f->file_offset / td->o.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->o.min_bs[ddir];
133 io_u->offset = (b * td->o.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->o.min_bs[ddir] == td->o.max_bs[ddir])
148 buflen = td->o.min_bs[ddir];
150 r = os_random_long(&td->bsrange_state);
151 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
152 if (!td->o.bs_unaligned)
153 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
156 while (buflen + io_u->offset > f->real_file_size) {
157 if (buflen == td->o.min_bs[ddir]) {
158 if (!td->o.odirect) {
159 assert(io_u->offset <= f->real_file_size);
160 buflen = f->real_file_size - io_u->offset;
166 buflen = td->o.min_bs[ddir];
172 static void set_rwmix_bytes(struct thread_data *td)
174 unsigned long long rbytes;
178 * we do time or byte based switch. this is needed because
179 * buffered writes may issue a lot quicker than they complete,
180 * whereas reads do not.
182 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
183 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
185 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
188 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
193 r = os_random_long(&td->rwmix_state);
194 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
195 if (v < td->o.rwmix[DDIR_READ])
202 * Return the data direction for the next io_u. If the job is a
203 * mixed read/write workload, check the rwmix cycle and switch if
206 static enum fio_ddir get_rw_ddir(struct thread_data *td)
210 unsigned long elapsed;
213 fio_gettime(&now, NULL);
214 elapsed = mtime_since_now(&td->rwmix_switch);
217 * if this is the first cycle, make it shorter
219 cycle = td->o.rwmixcycle;
220 if (!td->rwmix_bytes)
224 * Check if it's time to seed a new data direction.
226 if (elapsed >= cycle ||
227 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
228 unsigned long long max_bytes;
232 * Put a top limit on how many bytes we do for
233 * one data direction, to avoid overflowing the
236 ddir = get_rand_ddir(td);
237 max_bytes = td->this_io_bytes[ddir];
238 if (max_bytes >= (td->io_size * td->o.rwmix[ddir] / 100)) {
239 if (!td->rw_end_set[ddir]) {
240 td->rw_end_set[ddir] = 1;
241 memcpy(&td->rw_end[ddir], &now, sizeof(now));
246 if (ddir != td->rwmix_ddir)
249 td->rwmix_ddir = ddir;
250 memcpy(&td->rwmix_switch, &now, sizeof(now));
252 return td->rwmix_ddir;
253 } else if (td_read(td))
259 void put_io_u(struct thread_data *td, struct io_u *io_u)
261 assert((io_u->flags & IO_U_F_FREE) == 0);
262 io_u->flags |= IO_U_F_FREE;
265 list_del(&io_u->list);
266 list_add(&io_u->list, &td->io_u_freelist);
270 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
272 struct io_u *__io_u = *io_u;
274 __io_u->flags |= IO_U_F_FREE;
275 __io_u->flags &= ~IO_U_F_FLIGHT;
277 list_del(&__io_u->list);
278 list_add_tail(&__io_u->list, &td->io_u_requeues);
283 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
286 * If using an iolog, grab next piece if any available.
288 if (td->o.read_iolog)
289 return read_iolog_get(td, io_u);
292 * see if it's time to sync
294 if (td->o.fsync_blocks &&
295 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
296 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
297 io_u->ddir = DDIR_SYNC;
301 io_u->ddir = get_rw_ddir(td);
304 * No log, let the seq/rand engine retrieve the next buflen and
307 if (get_next_offset(td, io_u))
310 io_u->buflen = get_next_buflen(td, io_u);
315 * mark entry before potentially trimming io_u
317 if (!td->o.read_iolog && td_random(td) && !td->o.norandommap)
318 mark_random_map(td, io_u);
321 * If using a write iolog, store this entry.
323 if (td->o.write_iolog_file)
324 write_iolog_put(td, io_u);
329 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
333 if (io_u->ddir == DDIR_SYNC)
336 switch (td->cur_depth) {
353 td->ts.io_u_map[index]++;
354 td->ts.total_io_u[io_u->ddir]++;
357 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
388 td->ts.io_u_lat[index]++;
392 * Get next file to service by choosing one at random
394 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
401 long r = os_random_long(&td->next_file_state);
403 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
406 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
412 * Get next file to service by doing round robin between all available ones
414 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
417 unsigned int old_next_file = td->next_file;
421 f = &td->files[td->next_file];
424 if (td->next_file >= td->o.nr_files)
427 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
431 } while (td->next_file != old_next_file);
436 static struct fio_file *get_next_file(struct thread_data *td)
440 assert(td->o.nr_files <= td->files_index);
442 if (!td->nr_open_files)
445 f = td->file_service_file;
446 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
449 if (td->o.file_service_type == FIO_FSERVICE_RR)
450 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
452 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
454 td->file_service_file = f;
455 td->file_service_left = td->file_service_nr - 1;
459 static struct fio_file *find_next_new_file(struct thread_data *td)
463 if (td->o.file_service_type == FIO_FSERVICE_RR)
464 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
466 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
471 struct io_u *__get_io_u(struct thread_data *td)
473 struct io_u *io_u = NULL;
475 if (!list_empty(&td->io_u_requeues))
476 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
477 else if (!queue_full(td)) {
478 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
487 assert(io_u->flags & IO_U_F_FREE);
488 io_u->flags &= ~IO_U_F_FREE;
491 list_del(&io_u->list);
492 list_add(&io_u->list, &td->io_u_busylist);
500 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
501 * etc. The returned io_u is fully ready to be prepped and submitted.
503 struct io_u *get_io_u(struct thread_data *td)
509 io_u = __get_io_u(td);
514 * from a requeue, io_u already setup
520 f = get_next_file(td);
529 if (!fill_io_u(td, io_u))
533 * No more to do for this file, close it
536 td_io_close_file(td, f);
539 * probably not the right place to do this, but see
540 * if we need to open a new file
542 if (td->nr_open_files < td->o.open_files &&
543 td->o.open_files != td->o.nr_files) {
544 f = find_next_new_file(td);
546 if (!f || (ret = td_io_open_file(td, f))) {
554 if (td->zone_bytes >= td->o.zone_size) {
556 f->last_pos += td->o.zone_skip;
559 if (io_u->buflen + io_u->offset > f->real_file_size) {
560 if (td->io_ops->flags & FIO_RAWIO) {
565 io_u->buflen = f->real_file_size - io_u->offset;
568 if (io_u->ddir != DDIR_SYNC) {
574 f->last_pos = io_u->offset + io_u->buflen;
576 if (td->o.verify != VERIFY_NONE)
577 populate_verify_io_u(td, io_u);
581 * Set io data pointers.
584 io_u->xfer_buf = io_u->buf;
585 io_u->xfer_buflen = io_u->buflen;
587 if (td_io_prep(td, io_u)) {
592 fio_gettime(&io_u->start_time, NULL);
596 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
598 const char *msg[] = { "read", "write", "sync" };
600 log_err("fio: io_u error");
603 log_err(" on file %s", io_u->file->file_name);
605 log_err(": %s\n", strerror(io_u->error));
607 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
610 td_verror(td, io_u->error, "io_u error");
613 static void io_completed(struct thread_data *td, struct io_u *io_u,
614 struct io_completion_data *icd)
618 assert(io_u->flags & IO_U_F_FLIGHT);
619 io_u->flags &= ~IO_U_F_FLIGHT;
621 put_file(td, io_u->file);
623 if (io_u->ddir == DDIR_SYNC) {
624 td->last_was_sync = 1;
628 td->last_was_sync = 0;
631 unsigned int bytes = io_u->buflen - io_u->resid;
632 const enum fio_ddir idx = io_u->ddir;
635 td->io_blocks[idx]++;
636 td->io_bytes[idx] += bytes;
637 td->zone_bytes += bytes;
638 td->this_io_bytes[idx] += bytes;
640 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
642 msec = mtime_since(&io_u->issue_time, &icd->time);
644 add_clat_sample(td, idx, msec);
645 add_bw_sample(td, idx, &icd->time);
646 io_u_mark_latency(td, msec);
648 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
649 log_io_piece(td, io_u);
651 icd->bytes_done[idx] += bytes;
654 ret = io_u->end_io(io_u);
655 if (ret && !icd->error)
659 icd->error = io_u->error;
660 io_u_log_error(td, io_u);
664 static void init_icd(struct io_completion_data *icd, int nr)
666 fio_gettime(&icd->time, NULL);
671 icd->bytes_done[0] = icd->bytes_done[1] = 0;
674 static void ios_completed(struct thread_data *td,
675 struct io_completion_data *icd)
680 for (i = 0; i < icd->nr; i++) {
681 io_u = td->io_ops->event(td, i);
683 io_completed(td, io_u, icd);
689 * Complete a single io_u for the sync engines.
691 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
693 struct io_completion_data icd;
696 io_completed(td, io_u, &icd);
700 return icd.bytes_done[0] + icd.bytes_done[1];
702 td_verror(td, icd.error, "io_u_sync_complete");
707 * Called to complete min_events number of io for the async engines.
709 long io_u_queued_complete(struct thread_data *td, int min_events)
711 struct io_completion_data icd;
712 struct timespec *tvp = NULL;
714 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
719 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
721 td_verror(td, -ret, "td_io_getevents");
727 ios_completed(td, &icd);
729 return icd.bytes_done[0] + icd.bytes_done[1];
731 td_verror(td, icd.error, "io_u_queued_complete");
736 * Call when io_u is really queued, to update the submission latency.
738 void io_u_queued(struct thread_data *td, struct io_u *io_u)
740 unsigned long slat_time;
742 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
743 add_slat_sample(td, io_u->ddir, slat_time);
746 #ifdef FIO_USE_TIMEOUT
747 void io_u_set_timeout(struct thread_data *td)
749 assert(td->cur_depth);
751 td->timer.it_interval.tv_sec = 0;
752 td->timer.it_interval.tv_usec = 0;
753 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
754 td->timer.it_value.tv_usec = 0;
755 setitimer(ITIMER_REAL, &td->timer, NULL);
756 fio_gettime(&td->timeout_end, NULL);
759 static void io_u_dump(struct io_u *io_u)
761 unsigned long t_start = mtime_since_now(&io_u->start_time);
762 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
764 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
765 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);
766 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
769 void io_u_set_timeout(struct thread_data fio_unused *td)
774 #ifdef FIO_USE_TIMEOUT
775 static void io_u_timeout_handler(int fio_unused sig)
777 struct thread_data *td, *__td;
778 pid_t pid = getpid();
779 struct list_head *entry;
783 log_err("fio: io_u timeout\n");
786 * TLS would be nice...
789 for_each_td(__td, i) {
790 if (__td->pid == pid) {
797 log_err("fio: io_u timeout, can't find job\n");
801 if (!td->cur_depth) {
802 log_err("fio: timeout without pending work?\n");
806 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
808 list_for_each(entry, &td->io_u_busylist) {
809 io_u = list_entry(entry, struct io_u, list);
814 td_verror(td, ETIMEDOUT, "io_u timeout");
819 void io_u_init_timeout(void)
821 #ifdef FIO_USE_TIMEOUT
822 signal(SIGALRM, io_u_timeout_handler);