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;
75 static inline unsigned long long last_block(struct thread_data *td,
78 unsigned long long max_blocks;
80 max_blocks = f->io_size / td->o.rw_min_bs;
84 return max_blocks - 1;
88 * Return the next free block in the map.
90 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
91 unsigned long long *b)
95 i = f->last_free_lookup;
96 *b = (i * BLOCKS_PER_MAP);
97 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
98 if (f->file_map[i] != -1UL) {
99 *b += fio_ffz(f->file_map[i]);
100 if (*b > last_block(td, f))
102 f->last_free_lookup = i;
106 *b += BLOCKS_PER_MAP;
110 dprint(FD_IO, "failed finding a free block\n");
114 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
115 unsigned long long *b)
117 unsigned long long r, rb;
121 r = os_random_long(&td->random_state);
122 *b = last_block(td, f);
125 * if we are not maintaining a random map, we are done.
127 if (td->o.norandommap)
131 * calculate map offset and chec if it's free
134 if (random_map_free(td, f, rb))
140 * we get here, if we didn't suceed in looking up a block. generate
141 * a random start offset into the filemap, and find the first free
146 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
147 if (!get_next_free_block(td, f, b))
150 r = os_random_long(&td->random_state);
154 * that didn't work either, try exhaustive search from the start
156 f->last_free_lookup = 0;
157 return get_next_free_block(td, f, b);
161 * For random io, generate a random new block and see if it's used. Repeat
162 * until we find a free one. For sequential io, just return the end of
163 * the last io issued.
165 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
167 struct fio_file *f = io_u->file;
168 unsigned long long b;
170 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
171 td->ddir_nr = td->o.ddir_nr;
173 if (get_next_rand_offset(td, f, &b))
176 if (f->last_pos >= f->real_file_size) {
177 if (!td_random(td) || get_next_rand_offset(td, f, &b))
180 b = (f->last_pos - f->file_offset) / td->o.rw_min_bs;
183 io_u->offset = (b * td->o.rw_min_bs) + f->file_offset;
184 if (io_u->offset >= f->real_file_size) {
185 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
186 io_u->offset, f->real_file_size);
193 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
195 const int ddir = io_u->ddir;
199 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
200 buflen = td->o.min_bs[ddir];
202 r = os_random_long(&td->bsrange_state);
203 if (!td->o.bssplit_nr)
204 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
209 for (i = 0; i < td->o.bssplit_nr; i++) {
210 struct bssplit *bsp = &td->o.bssplit[i];
214 if (r <= ((LONG_MAX / 100L) * perc))
218 if (!td->o.bs_unaligned)
219 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
222 if (io_u->offset + buflen > io_u->file->real_file_size)
223 buflen = td->o.min_bs[ddir];
228 static void set_rwmix_bytes(struct thread_data *td)
230 unsigned long long rbytes;
234 * we do time or byte based switch. this is needed because
235 * buffered writes may issue a lot quicker than they complete,
236 * whereas reads do not.
238 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
239 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
241 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
244 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
249 r = os_random_long(&td->rwmix_state);
250 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
251 if (v < td->o.rwmix[DDIR_READ])
258 * Return the data direction for the next io_u. If the job is a
259 * mixed read/write workload, check the rwmix cycle and switch if
262 static enum fio_ddir get_rw_ddir(struct thread_data *td)
266 unsigned long elapsed;
269 fio_gettime(&now, NULL);
270 elapsed = mtime_since_now(&td->rwmix_switch);
273 * if this is the first cycle, make it shorter
275 cycle = td->o.rwmixcycle;
276 if (!td->rwmix_bytes)
280 * Check if it's time to seed a new data direction.
282 if (elapsed >= cycle ||
283 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
284 unsigned long long max_bytes;
288 * Put a top limit on how many bytes we do for
289 * one data direction, to avoid overflowing the
292 ddir = get_rand_ddir(td);
293 max_bytes = td->this_io_bytes[ddir];
294 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
295 if (!td->rw_end_set[ddir]) {
296 td->rw_end_set[ddir] = 1;
297 memcpy(&td->rw_end[ddir], &now, sizeof(now));
302 if (ddir != td->rwmix_ddir)
305 td->rwmix_ddir = ddir;
306 memcpy(&td->rwmix_switch, &now, sizeof(now));
308 return td->rwmix_ddir;
309 } else if (td_read(td))
315 void put_io_u(struct thread_data *td, struct io_u *io_u)
317 assert((io_u->flags & IO_U_F_FREE) == 0);
318 io_u->flags |= IO_U_F_FREE;
321 put_file(td, io_u->file);
324 list_del(&io_u->list);
325 list_add(&io_u->list, &td->io_u_freelist);
329 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
331 struct io_u *__io_u = *io_u;
333 __io_u->flags |= IO_U_F_FREE;
334 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
335 td->io_issues[__io_u->ddir]--;
337 __io_u->flags &= ~IO_U_F_FLIGHT;
339 list_del(&__io_u->list);
340 list_add_tail(&__io_u->list, &td->io_u_requeues);
345 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
347 if (td->io_ops->flags & FIO_NOIO)
351 * see if it's time to sync
353 if (td->o.fsync_blocks &&
354 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
355 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
356 io_u->ddir = DDIR_SYNC;
360 io_u->ddir = get_rw_ddir(td);
363 * See if it's time to switch to a new zone
365 if (td->zone_bytes >= td->o.zone_size) {
367 io_u->file->last_pos += td->o.zone_skip;
368 td->io_skip_bytes += td->o.zone_skip;
372 * No log, let the seq/rand engine retrieve the next buflen and
375 if (get_next_offset(td, io_u)) {
376 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
380 io_u->buflen = get_next_buflen(td, io_u);
382 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
386 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
387 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
392 * mark entry before potentially trimming io_u
394 if (td_random(td) && !td->o.norandommap)
395 mark_random_map(td, io_u);
398 * If using a write iolog, store this entry.
401 dprint_io_u(io_u, "fill_io_u");
402 td->zone_bytes += io_u->buflen;
407 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
411 if (io_u->ddir == DDIR_SYNC)
414 switch (td->cur_depth) {
436 td->ts.io_u_map[index]++;
437 td->ts.total_io_u[io_u->ddir]++;
440 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
477 assert(index < FIO_IO_U_LAT_U_NR);
478 td->ts.io_u_lat_u[index]++;
481 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
522 assert(index < FIO_IO_U_LAT_M_NR);
523 td->ts.io_u_lat_m[index]++;
526 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
529 io_u_mark_lat_usec(td, usec);
531 io_u_mark_lat_msec(td, usec / 1000);
535 * Get next file to service by choosing one at random
537 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
544 long r = os_random_long(&td->next_file_state);
546 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
548 if (f->flags & FIO_FILE_DONE)
551 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
552 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
559 * Get next file to service by doing round robin between all available ones
561 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
564 unsigned int old_next_file = td->next_file;
568 f = &td->files[td->next_file];
571 if (td->next_file >= td->o.nr_files)
574 if (f->flags & FIO_FILE_DONE) {
579 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
583 } while (td->next_file != old_next_file);
585 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
589 static struct fio_file *get_next_file(struct thread_data *td)
593 assert(td->o.nr_files <= td->files_index);
595 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
596 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files);
600 f = td->file_service_file;
601 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
604 if (td->o.file_service_type == FIO_FSERVICE_RR)
605 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
607 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
609 td->file_service_file = f;
610 td->file_service_left = td->file_service_nr - 1;
612 dprint(FD_FILE, "get_next_file: %p\n", f);
616 static struct fio_file *find_next_new_file(struct thread_data *td)
620 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
623 if (td->o.file_service_type == FIO_FSERVICE_RR)
624 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
626 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
631 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
636 f = get_next_file(td);
644 if (!fill_io_u(td, io_u))
648 * td_io_close() does a put_file() as well, so no need to
652 td_io_close_file(td, f);
653 f->flags |= FIO_FILE_DONE;
657 * probably not the right place to do this, but see
658 * if we need to open a new file
660 if (td->nr_open_files < td->o.open_files &&
661 td->o.open_files != td->o.nr_files) {
662 f = find_next_new_file(td);
664 if (!f || td_io_open_file(td, f))
675 struct io_u *__get_io_u(struct thread_data *td)
677 struct io_u *io_u = NULL;
679 if (!list_empty(&td->io_u_requeues))
680 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
681 else if (!queue_full(td)) {
682 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
691 assert(io_u->flags & IO_U_F_FREE);
692 io_u->flags &= ~IO_U_F_FREE;
695 list_del(&io_u->list);
696 list_add(&io_u->list, &td->io_u_busylist);
704 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
705 * etc. The returned io_u is fully ready to be prepped and submitted.
707 struct io_u *get_io_u(struct thread_data *td)
712 io_u = __get_io_u(td);
714 dprint(FD_IO, "__get_io_u failed\n");
719 * from a requeue, io_u already setup
725 * If using an iolog, grab next piece if any available.
727 if (td->o.read_iolog_file) {
728 if (read_iolog_get(td, io_u))
730 } else if (set_io_u_file(td, io_u)) {
731 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
736 assert(f->flags & FIO_FILE_OPEN);
738 if (io_u->ddir != DDIR_SYNC) {
739 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
740 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
744 f->last_pos = io_u->offset + io_u->buflen;
746 if (td->o.verify != VERIFY_NONE)
747 populate_verify_io_u(td, io_u);
751 * Set io data pointers.
753 io_u->endpos = io_u->offset + io_u->buflen;
755 io_u->xfer_buf = io_u->buf;
756 io_u->xfer_buflen = io_u->buflen;
758 if (!td_io_prep(td, io_u)) {
759 fio_gettime(&io_u->start_time, NULL);
763 dprint(FD_IO, "get_io_u failed\n");
768 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
770 const char *msg[] = { "read", "write", "sync" };
772 log_err("fio: io_u error");
775 log_err(" on file %s", io_u->file->file_name);
777 log_err(": %s\n", strerror(io_u->error));
779 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
782 td_verror(td, io_u->error, "io_u error");
785 static void io_completed(struct thread_data *td, struct io_u *io_u,
786 struct io_completion_data *icd)
790 dprint_io_u(io_u, "io complete");
792 assert(io_u->flags & IO_U_F_FLIGHT);
793 io_u->flags &= ~IO_U_F_FLIGHT;
795 if (io_u->ddir == DDIR_SYNC) {
796 td->last_was_sync = 1;
800 td->last_was_sync = 0;
803 unsigned int bytes = io_u->buflen - io_u->resid;
804 const enum fio_ddir idx = io_u->ddir;
807 td->io_blocks[idx]++;
808 td->io_bytes[idx] += bytes;
809 td->this_io_bytes[idx] += bytes;
811 io_u->file->last_completed_pos = io_u->endpos;
813 usec = utime_since(&io_u->issue_time, &icd->time);
815 add_clat_sample(td, idx, usec);
816 add_bw_sample(td, idx, &icd->time);
817 io_u_mark_latency(td, usec);
819 if (td_write(td) && idx == DDIR_WRITE &&
821 td->o.verify != VERIFY_NONE)
822 log_io_piece(td, io_u);
824 icd->bytes_done[idx] += bytes;
827 ret = io_u->end_io(td, io_u);
828 if (ret && !icd->error)
832 icd->error = io_u->error;
833 io_u_log_error(td, io_u);
837 static void init_icd(struct io_completion_data *icd, int nr)
839 fio_gettime(&icd->time, NULL);
844 icd->bytes_done[0] = icd->bytes_done[1] = 0;
847 static void ios_completed(struct thread_data *td,
848 struct io_completion_data *icd)
853 for (i = 0; i < icd->nr; i++) {
854 io_u = td->io_ops->event(td, i);
856 io_completed(td, io_u, icd);
862 * Complete a single io_u for the sync engines.
864 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
866 struct io_completion_data icd;
869 io_completed(td, io_u, &icd);
873 return icd.bytes_done[0] + icd.bytes_done[1];
875 td_verror(td, icd.error, "io_u_sync_complete");
880 * Called to complete min_events number of io for the async engines.
882 long io_u_queued_complete(struct thread_data *td, int min_events)
884 struct io_completion_data icd;
885 struct timespec *tvp = NULL;
887 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
889 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
894 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
896 td_verror(td, -ret, "td_io_getevents");
902 ios_completed(td, &icd);
904 return icd.bytes_done[0] + icd.bytes_done[1];
906 td_verror(td, icd.error, "io_u_queued_complete");
911 * Call when io_u is really queued, to update the submission latency.
913 void io_u_queued(struct thread_data *td, struct io_u *io_u)
915 unsigned long slat_time;
917 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
918 add_slat_sample(td, io_u->ddir, slat_time);
921 #ifdef FIO_USE_TIMEOUT
922 void io_u_set_timeout(struct thread_data *td)
924 assert(td->cur_depth);
926 td->timer.it_interval.tv_sec = 0;
927 td->timer.it_interval.tv_usec = 0;
928 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
929 td->timer.it_value.tv_usec = 0;
930 setitimer(ITIMER_REAL, &td->timer, NULL);
931 fio_gettime(&td->timeout_end, NULL);
934 static void io_u_dump(struct io_u *io_u)
936 unsigned long t_start = mtime_since_now(&io_u->start_time);
937 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
939 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
940 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);
941 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
944 void io_u_set_timeout(struct thread_data fio_unused *td)
949 #ifdef FIO_USE_TIMEOUT
950 static void io_u_timeout_handler(int fio_unused sig)
952 struct thread_data *td, *__td;
953 pid_t pid = getpid();
954 struct list_head *entry;
958 log_err("fio: io_u timeout\n");
961 * TLS would be nice...
964 for_each_td(__td, i) {
965 if (__td->pid == pid) {
972 log_err("fio: io_u timeout, can't find job\n");
976 if (!td->cur_depth) {
977 log_err("fio: timeout without pending work?\n");
981 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
983 list_for_each(entry, &td->io_u_busylist) {
984 io_u = list_entry(entry, struct io_u, list);
989 td_verror(td, ETIMEDOUT, "io_u timeout");
994 void io_u_init_timeout(void)
996 #ifdef FIO_USE_TIMEOUT
997 signal(SIGALRM, io_u_timeout_handler);