11 * Change this define to play with the timeout handling
13 #undef FIO_USE_TIMEOUT
15 struct io_completion_data {
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 static int random_map_free(struct thread_data *td, struct fio_file *f,
28 unsigned long long block)
30 unsigned int idx = RAND_MAP_IDX(td, f, block);
31 unsigned int bit = RAND_MAP_BIT(td, f, block);
33 return (f->file_map[idx] & (1UL << bit)) == 0;
37 * Mark a given offset as used in the map.
39 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41 unsigned int min_bs = td->o.rw_min_bs;
42 struct fio_file *f = io_u->file;
43 unsigned long long block;
45 unsigned int nr_blocks;
47 block = io_u->offset / (unsigned long long) min_bs;
49 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
51 while (blocks < nr_blocks) {
52 unsigned int idx, bit;
55 * If we have a mixed random workload, we may
56 * encounter blocks we already did IO to.
58 if (!td->o.ddir_nr == 1 && !random_map_free(td, f, block))
61 idx = RAND_MAP_IDX(td, f, block);
62 bit = RAND_MAP_BIT(td, f, block);
64 fio_assert(td, idx < f->num_maps);
66 f->file_map[idx] |= (1UL << bit);
71 if ((blocks * min_bs) < io_u->buflen)
72 io_u->buflen = blocks * min_bs;
76 * Return the next free block in the map.
78 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
79 unsigned long long *b)
83 i = f->last_free_lookup;
84 *b = (i * BLOCKS_PER_MAP);
85 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
86 if (f->file_map[i] != -1UL) {
87 *b += fio_ffz(f->file_map[i]);
88 f->last_free_lookup = i;
99 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
100 int ddir, unsigned long long *b)
102 unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
103 unsigned long long r, rb;
107 r = os_random_long(&td->random_state);
111 *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
113 * if we are not maintaining a random map, we are done.
115 if (td->o.norandommap)
119 * calculate map offset and chec if it's free
121 rb = *b + (f->file_offset / td->o.min_bs[ddir]);
122 if (random_map_free(td, f, rb))
128 * we get here, if we didn't suceed in looking up a block. generate
129 * a random start offset into the filemap, and find the first free
134 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
135 if (!get_next_free_block(td, f, b))
138 r = os_random_long(&td->random_state);
142 * that didn't work either, try exhaustive search from the start
144 f->last_free_lookup = 0;
145 return get_next_free_block(td, f, b);
149 * For random io, generate a random new block and see if it's used. Repeat
150 * until we find a free one. For sequential io, just return the end of
151 * the last io issued.
153 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
155 struct fio_file *f = io_u->file;
156 const int ddir = io_u->ddir;
157 unsigned long long b;
159 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
160 td->ddir_nr = td->o.ddir_nr;
162 if (get_next_rand_offset(td, f, ddir, &b))
165 if (f->last_pos >= f->real_file_size) {
166 if (!td_random(td) || get_next_rand_offset(td, f, ddir, &b))
169 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
172 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
173 if (io_u->offset >= f->real_file_size)
179 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
181 const int ddir = io_u->ddir;
185 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
186 buflen = td->o.min_bs[ddir];
188 r = os_random_long(&td->bsrange_state);
189 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
190 if (!td->o.bs_unaligned)
191 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
194 if (io_u->offset + buflen > io_u->file->real_file_size)
195 buflen = td->o.min_bs[ddir];
200 static void set_rwmix_bytes(struct thread_data *td)
202 unsigned long long rbytes;
206 * we do time or byte based switch. this is needed because
207 * buffered writes may issue a lot quicker than they complete,
208 * whereas reads do not.
210 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
211 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
213 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
216 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
221 r = os_random_long(&td->rwmix_state);
222 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
223 if (v < td->o.rwmix[DDIR_READ])
230 * Return the data direction for the next io_u. If the job is a
231 * mixed read/write workload, check the rwmix cycle and switch if
234 static enum fio_ddir get_rw_ddir(struct thread_data *td)
238 unsigned long elapsed;
241 fio_gettime(&now, NULL);
242 elapsed = mtime_since_now(&td->rwmix_switch);
245 * if this is the first cycle, make it shorter
247 cycle = td->o.rwmixcycle;
248 if (!td->rwmix_bytes)
252 * Check if it's time to seed a new data direction.
254 if (elapsed >= cycle ||
255 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
256 unsigned long long max_bytes;
260 * Put a top limit on how many bytes we do for
261 * one data direction, to avoid overflowing the
264 ddir = get_rand_ddir(td);
265 max_bytes = td->this_io_bytes[ddir];
266 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
267 if (!td->rw_end_set[ddir]) {
268 td->rw_end_set[ddir] = 1;
269 memcpy(&td->rw_end[ddir], &now, sizeof(now));
274 if (ddir != td->rwmix_ddir)
277 td->rwmix_ddir = ddir;
278 memcpy(&td->rwmix_switch, &now, sizeof(now));
280 return td->rwmix_ddir;
281 } else if (td_read(td))
287 void put_io_u(struct thread_data *td, struct io_u *io_u)
289 assert((io_u->flags & IO_U_F_FREE) == 0);
290 io_u->flags |= IO_U_F_FREE;
293 put_file(td, io_u->file);
296 list_del(&io_u->list);
297 list_add(&io_u->list, &td->io_u_freelist);
301 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
303 struct io_u *__io_u = *io_u;
305 __io_u->flags |= IO_U_F_FREE;
306 __io_u->flags &= ~IO_U_F_FLIGHT;
308 list_del(&__io_u->list);
309 list_add_tail(&__io_u->list, &td->io_u_requeues);
314 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
317 * see if it's time to sync
319 if (td->o.fsync_blocks &&
320 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
321 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
322 io_u->ddir = DDIR_SYNC;
326 io_u->ddir = get_rw_ddir(td);
329 * See if it's time to switch to a new zone
331 if (td->zone_bytes >= td->o.zone_size) {
333 io_u->file->last_pos += td->o.zone_skip;
334 td->io_skip_bytes += td->o.zone_skip;
338 * No log, let the seq/rand engine retrieve the next buflen and
341 if (get_next_offset(td, io_u))
344 io_u->buflen = get_next_buflen(td, io_u);
348 if (io_u->offset + io_u->buflen > io_u->file->real_file_size)
352 * mark entry before potentially trimming io_u
354 if (td_random(td) && !td->o.norandommap)
355 mark_random_map(td, io_u);
358 * If using a write iolog, store this entry.
365 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
369 if (io_u->ddir == DDIR_SYNC)
372 switch (td->cur_depth) {
394 td->ts.io_u_map[index]++;
395 td->ts.total_io_u[io_u->ddir]++;
398 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
435 assert(index < FIO_IO_U_LAT_U_NR);
436 td->ts.io_u_lat_u[index]++;
439 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
480 assert(index < FIO_IO_U_LAT_M_NR);
481 td->ts.io_u_lat_m[index]++;
484 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
487 io_u_mark_lat_usec(td, usec);
489 io_u_mark_lat_msec(td, usec / 1000);
493 * Get next file to service by choosing one at random
495 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
502 long r = os_random_long(&td->next_file_state);
504 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
506 if (f->flags & FIO_FILE_DONE)
509 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
515 * Get next file to service by doing round robin between all available ones
517 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
520 unsigned int old_next_file = td->next_file;
524 f = &td->files[td->next_file];
527 if (td->next_file >= td->o.nr_files)
530 if (f->flags & FIO_FILE_DONE) {
535 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
539 } while (td->next_file != old_next_file);
544 static struct fio_file *get_next_file(struct thread_data *td)
548 assert(td->o.nr_files <= td->files_index);
550 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
553 f = td->file_service_file;
554 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
557 if (td->o.file_service_type == FIO_FSERVICE_RR)
558 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
560 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
562 td->file_service_file = f;
563 td->file_service_left = td->file_service_nr - 1;
567 static struct fio_file *find_next_new_file(struct thread_data *td)
571 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
574 if (td->o.file_service_type == FIO_FSERVICE_RR)
575 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
577 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
582 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
587 f = get_next_file(td);
595 if (!fill_io_u(td, io_u))
599 * td_io_close() does a put_file() as well, so no need to
603 td_io_close_file(td, f);
604 f->flags |= FIO_FILE_DONE;
608 * probably not the right place to do this, but see
609 * if we need to open a new file
611 if (td->nr_open_files < td->o.open_files &&
612 td->o.open_files != td->o.nr_files) {
613 f = find_next_new_file(td);
615 if (!f || td_io_open_file(td, f))
626 struct io_u *__get_io_u(struct thread_data *td)
628 struct io_u *io_u = NULL;
630 if (!list_empty(&td->io_u_requeues))
631 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
632 else if (!queue_full(td)) {
633 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
642 assert(io_u->flags & IO_U_F_FREE);
643 io_u->flags &= ~IO_U_F_FREE;
646 list_del(&io_u->list);
647 list_add(&io_u->list, &td->io_u_busylist);
655 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
656 * etc. The returned io_u is fully ready to be prepped and submitted.
658 struct io_u *get_io_u(struct thread_data *td)
663 io_u = __get_io_u(td);
668 * from a requeue, io_u already setup
674 * If using an iolog, grab next piece if any available.
676 if (td->o.read_iolog_file) {
677 if (read_iolog_get(td, io_u))
679 } else if (set_io_u_file(td, io_u))
683 assert(f->flags & FIO_FILE_OPEN);
685 if (io_u->ddir != DDIR_SYNC) {
689 f->last_pos = io_u->offset + io_u->buflen;
691 if (td->o.verify != VERIFY_NONE)
692 populate_verify_io_u(td, io_u);
696 * Set io data pointers.
698 io_u->endpos = io_u->offset + io_u->buflen;
700 io_u->xfer_buf = io_u->buf;
701 io_u->xfer_buflen = io_u->buflen;
703 if (!td_io_prep(td, io_u)) {
704 fio_gettime(&io_u->start_time, NULL);
712 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
714 const char *msg[] = { "read", "write", "sync" };
716 log_err("fio: io_u error");
719 log_err(" on file %s", io_u->file->file_name);
721 log_err(": %s\n", strerror(io_u->error));
723 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
726 td_verror(td, io_u->error, "io_u error");
729 static void io_completed(struct thread_data *td, struct io_u *io_u,
730 struct io_completion_data *icd)
734 assert(io_u->flags & IO_U_F_FLIGHT);
735 io_u->flags &= ~IO_U_F_FLIGHT;
737 if (io_u->ddir == DDIR_SYNC) {
738 td->last_was_sync = 1;
742 td->last_was_sync = 0;
745 unsigned int bytes = io_u->buflen - io_u->resid;
746 const enum fio_ddir idx = io_u->ddir;
749 td->io_blocks[idx]++;
750 td->io_bytes[idx] += bytes;
751 td->zone_bytes += bytes;
752 td->this_io_bytes[idx] += bytes;
754 io_u->file->last_completed_pos = io_u->endpos;
756 usec = utime_since(&io_u->issue_time, &icd->time);
758 add_clat_sample(td, idx, usec);
759 add_bw_sample(td, idx, &icd->time);
760 io_u_mark_latency(td, usec);
762 if (td_write(td) && idx == DDIR_WRITE &&
763 td->o.verify != VERIFY_NONE)
764 log_io_piece(td, io_u);
766 icd->bytes_done[idx] += bytes;
769 ret = io_u->end_io(td, io_u);
770 if (ret && !icd->error)
774 icd->error = io_u->error;
775 io_u_log_error(td, io_u);
779 static void init_icd(struct io_completion_data *icd, int nr)
781 fio_gettime(&icd->time, NULL);
786 icd->bytes_done[0] = icd->bytes_done[1] = 0;
789 static void ios_completed(struct thread_data *td,
790 struct io_completion_data *icd)
795 for (i = 0; i < icd->nr; i++) {
796 io_u = td->io_ops->event(td, i);
798 io_completed(td, io_u, icd);
804 * Complete a single io_u for the sync engines.
806 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
808 struct io_completion_data icd;
811 io_completed(td, io_u, &icd);
815 return icd.bytes_done[0] + icd.bytes_done[1];
817 td_verror(td, icd.error, "io_u_sync_complete");
822 * Called to complete min_events number of io for the async engines.
824 long io_u_queued_complete(struct thread_data *td, int min_events)
826 struct io_completion_data icd;
827 struct timespec *tvp = NULL;
829 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
834 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
836 td_verror(td, -ret, "td_io_getevents");
842 ios_completed(td, &icd);
844 return icd.bytes_done[0] + icd.bytes_done[1];
846 td_verror(td, icd.error, "io_u_queued_complete");
851 * Call when io_u is really queued, to update the submission latency.
853 void io_u_queued(struct thread_data *td, struct io_u *io_u)
855 unsigned long slat_time;
857 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
858 add_slat_sample(td, io_u->ddir, slat_time);
861 #ifdef FIO_USE_TIMEOUT
862 void io_u_set_timeout(struct thread_data *td)
864 assert(td->cur_depth);
866 td->timer.it_interval.tv_sec = 0;
867 td->timer.it_interval.tv_usec = 0;
868 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
869 td->timer.it_value.tv_usec = 0;
870 setitimer(ITIMER_REAL, &td->timer, NULL);
871 fio_gettime(&td->timeout_end, NULL);
874 static void io_u_dump(struct io_u *io_u)
876 unsigned long t_start = mtime_since_now(&io_u->start_time);
877 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
879 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
880 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);
881 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
884 void io_u_set_timeout(struct thread_data fio_unused *td)
889 #ifdef FIO_USE_TIMEOUT
890 static void io_u_timeout_handler(int fio_unused sig)
892 struct thread_data *td, *__td;
893 pid_t pid = getpid();
894 struct list_head *entry;
898 log_err("fio: io_u timeout\n");
901 * TLS would be nice...
904 for_each_td(__td, i) {
905 if (__td->pid == pid) {
912 log_err("fio: io_u timeout, can't find job\n");
916 if (!td->cur_depth) {
917 log_err("fio: timeout without pending work?\n");
921 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
923 list_for_each(entry, &td->io_u_busylist) {
924 io_u = list_entry(entry, struct io_u, list);
929 td_verror(td, ETIMEDOUT, "io_u timeout");
934 void io_u_init_timeout(void)
936 #ifdef FIO_USE_TIMEOUT
937 signal(SIGALRM, io_u_timeout_handler);