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;
56 * If we have a mixed random workload, we may
57 * encounter blocks we already did IO to.
59 if (!td->o.ddir_nr && !random_map_free(td, f, block))
62 idx = RAND_MAP_IDX(td, f, block);
63 bit = RAND_MAP_BIT(td, f, block);
65 fio_assert(td, idx < f->num_maps);
67 f->file_map[idx] |= (1UL << bit);
72 if ((blocks * min_bs) < io_u->buflen)
73 io_u->buflen = blocks * min_bs;
77 * Return the next free block in the map.
79 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
80 unsigned long long *b)
84 i = f->last_free_lookup;
85 *b = (i * BLOCKS_PER_MAP);
86 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
87 if (f->file_map[i] != -1UL) {
88 *b += ffz(f->file_map[i]);
89 f->last_free_lookup = i;
100 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
101 int ddir, unsigned long long *b)
103 unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
104 unsigned long long r, rb;
108 r = os_random_long(&td->random_state);
112 *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
113 if (td->o.norandommap)
115 rb = *b + (f->file_offset / td->o.min_bs[ddir]);
117 } while (!random_map_free(td, f, rb) && loops);
120 * if we failed to retrieve a truly random offset within
121 * the loops assigned, see if there are free ones left at all
123 if (!loops && get_next_free_block(td, f, b))
130 * For random io, generate a random new block and see if it's used. Repeat
131 * until we find a free one. For sequential io, just return the end of
132 * the last io issued.
134 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
136 struct fio_file *f = io_u->file;
137 const int ddir = io_u->ddir;
138 unsigned long long b;
140 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
141 td->ddir_nr = td->o.ddir_nr;
143 if (get_next_rand_offset(td, f, ddir, &b))
146 if (f->last_pos >= f->real_file_size)
149 b = f->last_pos / td->o.min_bs[ddir];
152 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
153 if (io_u->offset >= f->real_file_size)
159 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
161 const int ddir = io_u->ddir;
165 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
166 buflen = td->o.min_bs[ddir];
168 r = os_random_long(&td->bsrange_state);
169 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
170 if (!td->o.bs_unaligned)
171 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
177 static void set_rwmix_bytes(struct thread_data *td)
179 unsigned long long rbytes;
183 * we do time or byte based switch. this is needed because
184 * buffered writes may issue a lot quicker than they complete,
185 * whereas reads do not.
187 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
188 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
190 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
193 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
198 r = os_random_long(&td->rwmix_state);
199 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
200 if (v < td->o.rwmix[DDIR_READ])
207 * Return the data direction for the next io_u. If the job is a
208 * mixed read/write workload, check the rwmix cycle and switch if
211 static enum fio_ddir get_rw_ddir(struct thread_data *td)
215 unsigned long elapsed;
218 fio_gettime(&now, NULL);
219 elapsed = mtime_since_now(&td->rwmix_switch);
222 * if this is the first cycle, make it shorter
224 cycle = td->o.rwmixcycle;
225 if (!td->rwmix_bytes)
229 * Check if it's time to seed a new data direction.
231 if (elapsed >= cycle ||
232 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
233 unsigned long long max_bytes;
237 * Put a top limit on how many bytes we do for
238 * one data direction, to avoid overflowing the
241 ddir = get_rand_ddir(td);
242 max_bytes = td->this_io_bytes[ddir];
243 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
244 if (!td->rw_end_set[ddir]) {
245 td->rw_end_set[ddir] = 1;
246 memcpy(&td->rw_end[ddir], &now, sizeof(now));
251 if (ddir != td->rwmix_ddir)
254 td->rwmix_ddir = ddir;
255 memcpy(&td->rwmix_switch, &now, sizeof(now));
257 return td->rwmix_ddir;
258 } else if (td_read(td))
264 void put_io_u(struct thread_data *td, struct io_u *io_u)
266 assert((io_u->flags & IO_U_F_FREE) == 0);
267 io_u->flags |= IO_U_F_FREE;
270 list_del(&io_u->list);
271 list_add(&io_u->list, &td->io_u_freelist);
275 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
277 struct io_u *__io_u = *io_u;
279 __io_u->flags |= IO_U_F_FREE;
280 __io_u->flags &= ~IO_U_F_FLIGHT;
282 list_del(&__io_u->list);
283 list_add_tail(&__io_u->list, &td->io_u_requeues);
288 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
291 * If using an iolog, grab next piece if any available.
293 if (td->o.read_iolog)
294 return read_iolog_get(td, io_u);
297 * see if it's time to sync
299 if (td->o.fsync_blocks &&
300 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
301 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
302 io_u->ddir = DDIR_SYNC;
306 io_u->ddir = get_rw_ddir(td);
309 * No log, let the seq/rand engine retrieve the next buflen and
312 if (get_next_offset(td, io_u))
315 io_u->buflen = get_next_buflen(td, io_u);
320 * mark entry before potentially trimming io_u
322 if (td_random(td) && !td->o.norandommap)
323 mark_random_map(td, io_u);
326 * If using a write iolog, store this entry.
329 if (td->o.write_iolog_file)
330 write_iolog_put(td, io_u);
335 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
339 if (io_u->ddir == DDIR_SYNC)
342 switch (td->cur_depth) {
359 td->ts.io_u_map[index]++;
360 td->ts.total_io_u[io_u->ddir]++;
363 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
394 td->ts.io_u_lat[index]++;
398 * Get next file to service by choosing one at random
400 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
407 long r = os_random_long(&td->next_file_state);
409 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
411 if (f->flags & FIO_FILE_DONE)
414 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
420 * Get next file to service by doing round robin between all available ones
422 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
425 unsigned int old_next_file = td->next_file;
429 f = &td->files[td->next_file];
432 if (td->next_file >= td->o.nr_files)
435 if (f->flags & FIO_FILE_DONE) {
440 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
444 } while (td->next_file != old_next_file);
449 static struct fio_file *get_next_file(struct thread_data *td)
453 assert(td->o.nr_files <= td->files_index);
455 if (!td->nr_open_files)
458 f = td->file_service_file;
459 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
462 if (td->o.file_service_type == FIO_FSERVICE_RR)
463 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
465 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
467 td->file_service_file = f;
468 td->file_service_left = td->file_service_nr - 1;
472 static struct fio_file *find_next_new_file(struct thread_data *td)
476 if (td->o.file_service_type == FIO_FSERVICE_RR)
477 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
479 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
484 struct io_u *__get_io_u(struct thread_data *td)
486 struct io_u *io_u = NULL;
488 if (!list_empty(&td->io_u_requeues))
489 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
490 else if (!queue_full(td)) {
491 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
500 assert(io_u->flags & IO_U_F_FREE);
501 io_u->flags &= ~IO_U_F_FREE;
504 list_del(&io_u->list);
505 list_add(&io_u->list, &td->io_u_busylist);
513 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
514 * etc. The returned io_u is fully ready to be prepped and submitted.
516 struct io_u *get_io_u(struct thread_data *td)
522 io_u = __get_io_u(td);
527 * from a requeue, io_u already setup
533 f = get_next_file(td);
542 if (!fill_io_u(td, io_u))
546 * No more to do for this file, close it
549 td_io_close_file(td, f);
550 f->flags |= FIO_FILE_DONE;
553 * probably not the right place to do this, but see
554 * if we need to open a new file
556 if (td->nr_open_files < td->o.open_files &&
557 td->o.open_files != td->o.nr_files) {
558 f = find_next_new_file(td);
560 if (!f || (ret = td_io_open_file(td, f))) {
568 if (td->zone_bytes >= td->o.zone_size) {
570 f->last_pos += td->o.zone_skip;
573 if (io_u->ddir != DDIR_SYNC) {
579 f->last_pos = io_u->offset + io_u->buflen;
581 if (td->o.verify != VERIFY_NONE)
582 populate_verify_io_u(td, io_u);
586 * Set io data pointers.
589 io_u->xfer_buf = io_u->buf;
590 io_u->xfer_buflen = io_u->buflen;
592 if (td_io_prep(td, io_u)) {
597 fio_gettime(&io_u->start_time, NULL);
601 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
603 const char *msg[] = { "read", "write", "sync" };
605 log_err("fio: io_u error");
608 log_err(" on file %s", io_u->file->file_name);
610 log_err(": %s\n", strerror(io_u->error));
612 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
615 td_verror(td, io_u->error, "io_u error");
618 static void io_completed(struct thread_data *td, struct io_u *io_u,
619 struct io_completion_data *icd)
623 assert(io_u->flags & IO_U_F_FLIGHT);
624 io_u->flags &= ~IO_U_F_FLIGHT;
626 put_file(td, io_u->file);
628 if (io_u->ddir == DDIR_SYNC) {
629 td->last_was_sync = 1;
633 td->last_was_sync = 0;
636 unsigned int bytes = io_u->buflen - io_u->resid;
637 const enum fio_ddir idx = io_u->ddir;
640 td->io_blocks[idx]++;
641 td->io_bytes[idx] += bytes;
642 td->zone_bytes += bytes;
643 td->this_io_bytes[idx] += bytes;
645 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
647 msec = mtime_since(&io_u->issue_time, &icd->time);
649 add_clat_sample(td, idx, msec);
650 add_bw_sample(td, idx, &icd->time);
651 io_u_mark_latency(td, msec);
653 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
654 td->o.verify != VERIFY_NONE)
655 log_io_piece(td, io_u);
657 icd->bytes_done[idx] += bytes;
660 ret = io_u->end_io(td, io_u);
661 if (ret && !icd->error)
665 icd->error = io_u->error;
666 io_u_log_error(td, io_u);
670 static void init_icd(struct io_completion_data *icd, int nr)
672 fio_gettime(&icd->time, NULL);
677 icd->bytes_done[0] = icd->bytes_done[1] = 0;
680 static void ios_completed(struct thread_data *td,
681 struct io_completion_data *icd)
686 for (i = 0; i < icd->nr; i++) {
687 io_u = td->io_ops->event(td, i);
689 io_completed(td, io_u, icd);
695 * Complete a single io_u for the sync engines.
697 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
699 struct io_completion_data icd;
702 io_completed(td, io_u, &icd);
706 return icd.bytes_done[0] + icd.bytes_done[1];
708 td_verror(td, icd.error, "io_u_sync_complete");
713 * Called to complete min_events number of io for the async engines.
715 long io_u_queued_complete(struct thread_data *td, int min_events)
717 struct io_completion_data icd;
718 struct timespec *tvp = NULL;
720 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
725 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
727 td_verror(td, -ret, "td_io_getevents");
733 ios_completed(td, &icd);
735 return icd.bytes_done[0] + icd.bytes_done[1];
737 td_verror(td, icd.error, "io_u_queued_complete");
742 * Call when io_u is really queued, to update the submission latency.
744 void io_u_queued(struct thread_data *td, struct io_u *io_u)
746 unsigned long slat_time;
748 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
749 add_slat_sample(td, io_u->ddir, slat_time);
752 #ifdef FIO_USE_TIMEOUT
753 void io_u_set_timeout(struct thread_data *td)
755 assert(td->cur_depth);
757 td->timer.it_interval.tv_sec = 0;
758 td->timer.it_interval.tv_usec = 0;
759 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
760 td->timer.it_value.tv_usec = 0;
761 setitimer(ITIMER_REAL, &td->timer, NULL);
762 fio_gettime(&td->timeout_end, NULL);
765 static void io_u_dump(struct io_u *io_u)
767 unsigned long t_start = mtime_since_now(&io_u->start_time);
768 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
770 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
771 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);
772 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
775 void io_u_set_timeout(struct thread_data fio_unused *td)
780 #ifdef FIO_USE_TIMEOUT
781 static void io_u_timeout_handler(int fio_unused sig)
783 struct thread_data *td, *__td;
784 pid_t pid = getpid();
785 struct list_head *entry;
789 log_err("fio: io_u timeout\n");
792 * TLS would be nice...
795 for_each_td(__td, i) {
796 if (__td->pid == pid) {
803 log_err("fio: io_u timeout, can't find job\n");
807 if (!td->cur_depth) {
808 log_err("fio: timeout without pending work?\n");
812 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
814 list_for_each(entry, &td->io_u_busylist) {
815 io_u = list_entry(entry, struct io_u, list);
820 td_verror(td, ETIMEDOUT, "io_u timeout");
825 void io_u_init_timeout(void)
827 #ifdef FIO_USE_TIMEOUT
828 signal(SIGALRM, io_u_timeout_handler);