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,
79 unsigned long long max_blocks;
81 max_blocks = f->io_size / td->o.min_bs[ddir];
85 return max_blocks - 1;
89 * Return the next free block in the map.
91 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
92 enum fio_ddir ddir, unsigned long long *b)
96 i = f->last_free_lookup;
97 *b = (i * BLOCKS_PER_MAP);
98 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
99 if (f->file_map[i] != -1UL) {
100 *b += fio_ffz(f->file_map[i]);
101 if (*b > last_block(td, f, ddir))
103 f->last_free_lookup = i;
107 *b += BLOCKS_PER_MAP;
111 dprint(FD_IO, "failed finding a free block\n");
115 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
116 enum fio_ddir ddir, unsigned long long *b)
118 unsigned long long r, rb;
122 r = os_random_long(&td->random_state);
123 *b = last_block(td, f, ddir);
126 * if we are not maintaining a random map, we are done.
128 if (td->o.norandommap)
132 * calculate map offset and chec if it's free
135 if (random_map_free(td, f, rb))
141 * we get here, if we didn't suceed in looking up a block. generate
142 * a random start offset into the filemap, and find the first free
147 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
148 if (!get_next_free_block(td, f, ddir, b))
151 r = os_random_long(&td->random_state);
155 * that didn't work either, try exhaustive search from the start
157 f->last_free_lookup = 0;
158 return get_next_free_block(td, f, ddir, b);
162 * For random io, generate a random new block and see if it's used. Repeat
163 * until we find a free one. For sequential io, just return the end of
164 * the last io issued.
166 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
168 struct fio_file *f = io_u->file;
169 unsigned long long b;
170 enum fio_ddir ddir = io_u->ddir;
172 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
173 td->ddir_nr = td->o.ddir_nr;
175 if (get_next_rand_offset(td, f, ddir, &b))
178 if (f->last_pos >= f->real_file_size) {
179 if (!td_random(td) ||
180 get_next_rand_offset(td, f, ddir, &b))
183 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
186 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
187 if (io_u->offset >= f->real_file_size) {
188 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
189 io_u->offset, f->real_file_size);
196 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
198 const int ddir = io_u->ddir;
202 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
203 buflen = td->o.min_bs[ddir];
205 r = os_random_long(&td->bsrange_state);
206 if (!td->o.bssplit_nr)
207 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
212 for (i = 0; i < td->o.bssplit_nr; i++) {
213 struct bssplit *bsp = &td->o.bssplit[i];
217 if (r <= ((LONG_MAX / 100L) * perc))
221 if (!td->o.bs_unaligned)
222 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
225 if (io_u->offset + buflen > io_u->file->real_file_size) {
226 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
227 td->o.min_bs[ddir], ddir);
228 buflen = td->o.min_bs[ddir];
234 static void set_rwmix_bytes(struct thread_data *td)
236 unsigned long long rbytes;
240 * we do time or byte based switch. this is needed because
241 * buffered writes may issue a lot quicker than they complete,
242 * whereas reads do not.
244 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
245 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
247 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
250 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
255 r = os_random_long(&td->rwmix_state);
256 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
257 if (v < td->o.rwmix[DDIR_READ])
264 * Return the data direction for the next io_u. If the job is a
265 * mixed read/write workload, check the rwmix cycle and switch if
268 static enum fio_ddir get_rw_ddir(struct thread_data *td)
272 unsigned long elapsed;
275 fio_gettime(&now, NULL);
276 elapsed = mtime_since_now(&td->rwmix_switch);
279 * if this is the first cycle, make it shorter
281 cycle = td->o.rwmixcycle;
282 if (!td->rwmix_bytes)
286 * Check if it's time to seed a new data direction.
288 if (elapsed >= cycle ||
289 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
290 unsigned long long max_bytes;
294 * Put a top limit on how many bytes we do for
295 * one data direction, to avoid overflowing the
298 ddir = get_rand_ddir(td);
299 max_bytes = td->this_io_bytes[ddir];
300 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
301 if (!td->rw_end_set[ddir]) {
302 td->rw_end_set[ddir] = 1;
303 memcpy(&td->rw_end[ddir], &now, sizeof(now));
308 if (ddir != td->rwmix_ddir)
311 td->rwmix_ddir = ddir;
312 memcpy(&td->rwmix_switch, &now, sizeof(now));
314 return td->rwmix_ddir;
315 } else if (td_read(td))
321 void put_io_u(struct thread_data *td, struct io_u *io_u)
323 assert((io_u->flags & IO_U_F_FREE) == 0);
324 io_u->flags |= IO_U_F_FREE;
327 put_file(td, io_u->file);
330 list_del(&io_u->list);
331 list_add(&io_u->list, &td->io_u_freelist);
335 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
337 struct io_u *__io_u = *io_u;
339 __io_u->flags |= IO_U_F_FREE;
340 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
341 td->io_issues[__io_u->ddir]--;
343 __io_u->flags &= ~IO_U_F_FLIGHT;
345 list_del(&__io_u->list);
346 list_add_tail(&__io_u->list, &td->io_u_requeues);
351 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
353 if (td->io_ops->flags & FIO_NOIO)
357 * see if it's time to sync
359 if (td->o.fsync_blocks &&
360 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
361 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
362 io_u->ddir = DDIR_SYNC;
366 io_u->ddir = get_rw_ddir(td);
369 * See if it's time to switch to a new zone
371 if (td->zone_bytes >= td->o.zone_size) {
373 io_u->file->last_pos += td->o.zone_skip;
374 td->io_skip_bytes += td->o.zone_skip;
378 * No log, let the seq/rand engine retrieve the next buflen and
381 if (get_next_offset(td, io_u)) {
382 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
386 io_u->buflen = get_next_buflen(td, io_u);
388 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
392 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
393 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
394 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
395 io_u->buflen, io_u->file->real_file_size);
400 * mark entry before potentially trimming io_u
402 if (td_random(td) && !td->o.norandommap)
403 mark_random_map(td, io_u);
406 * If using a write iolog, store this entry.
409 dprint_io_u(io_u, "fill_io_u");
410 td->zone_bytes += io_u->buflen;
415 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
419 if (io_u->ddir == DDIR_SYNC)
422 switch (td->cur_depth) {
444 td->ts.io_u_map[index]++;
445 td->ts.total_io_u[io_u->ddir]++;
448 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
485 assert(index < FIO_IO_U_LAT_U_NR);
486 td->ts.io_u_lat_u[index]++;
489 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
530 assert(index < FIO_IO_U_LAT_M_NR);
531 td->ts.io_u_lat_m[index]++;
534 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
537 io_u_mark_lat_usec(td, usec);
539 io_u_mark_lat_msec(td, usec / 1000);
543 * Get next file to service by choosing one at random
545 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
552 long r = os_random_long(&td->next_file_state);
554 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
556 if (f->flags & FIO_FILE_DONE)
559 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
560 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
567 * Get next file to service by doing round robin between all available ones
569 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
572 unsigned int old_next_file = td->next_file;
576 f = &td->files[td->next_file];
579 if (td->next_file >= td->o.nr_files)
582 if (f->flags & FIO_FILE_DONE) {
587 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
591 } while (td->next_file != old_next_file);
593 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
597 static struct fio_file *get_next_file(struct thread_data *td)
601 assert(td->o.nr_files <= td->files_index);
603 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
604 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);
608 f = td->file_service_file;
609 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
612 if (td->o.file_service_type == FIO_FSERVICE_RR)
613 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
615 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
617 td->file_service_file = f;
618 td->file_service_left = td->file_service_nr - 1;
620 dprint(FD_FILE, "get_next_file: %p\n", f);
624 static struct fio_file *find_next_new_file(struct thread_data *td)
628 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
631 if (td->o.file_service_type == FIO_FSERVICE_RR)
632 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
634 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
639 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
644 f = get_next_file(td);
652 if (!fill_io_u(td, io_u))
656 * td_io_close() does a put_file() as well, so no need to
660 td_io_close_file(td, f);
661 f->flags |= FIO_FILE_DONE;
665 * probably not the right place to do this, but see
666 * if we need to open a new file
668 if (td->nr_open_files < td->o.open_files &&
669 td->o.open_files != td->o.nr_files) {
670 f = find_next_new_file(td);
672 if (!f || td_io_open_file(td, f))
683 struct io_u *__get_io_u(struct thread_data *td)
685 struct io_u *io_u = NULL;
687 if (!list_empty(&td->io_u_requeues))
688 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
689 else if (!queue_full(td)) {
690 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
699 assert(io_u->flags & IO_U_F_FREE);
700 io_u->flags &= ~IO_U_F_FREE;
703 list_del(&io_u->list);
704 list_add(&io_u->list, &td->io_u_busylist);
712 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
713 * etc. The returned io_u is fully ready to be prepped and submitted.
715 struct io_u *get_io_u(struct thread_data *td)
720 io_u = __get_io_u(td);
722 dprint(FD_IO, "__get_io_u failed\n");
727 * from a requeue, io_u already setup
733 * If using an iolog, grab next piece if any available.
735 if (td->o.read_iolog_file) {
736 if (read_iolog_get(td, io_u))
738 } else if (set_io_u_file(td, io_u)) {
739 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
744 assert(f->flags & FIO_FILE_OPEN);
746 if (io_u->ddir != DDIR_SYNC) {
747 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
748 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
752 f->last_pos = io_u->offset + io_u->buflen;
754 if (td->o.verify != VERIFY_NONE)
755 populate_verify_io_u(td, io_u);
759 * Set io data pointers.
761 io_u->endpos = io_u->offset + io_u->buflen;
763 io_u->xfer_buf = io_u->buf;
764 io_u->xfer_buflen = io_u->buflen;
766 if (!td_io_prep(td, io_u)) {
767 fio_gettime(&io_u->start_time, NULL);
771 dprint(FD_IO, "get_io_u failed\n");
776 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
778 const char *msg[] = { "read", "write", "sync" };
780 log_err("fio: io_u error");
783 log_err(" on file %s", io_u->file->file_name);
785 log_err(": %s\n", strerror(io_u->error));
787 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
790 td_verror(td, io_u->error, "io_u error");
793 static void io_completed(struct thread_data *td, struct io_u *io_u,
794 struct io_completion_data *icd)
798 dprint_io_u(io_u, "io complete");
800 assert(io_u->flags & IO_U_F_FLIGHT);
801 io_u->flags &= ~IO_U_F_FLIGHT;
803 if (io_u->ddir == DDIR_SYNC) {
804 td->last_was_sync = 1;
808 td->last_was_sync = 0;
811 unsigned int bytes = io_u->buflen - io_u->resid;
812 const enum fio_ddir idx = io_u->ddir;
815 td->io_blocks[idx]++;
816 td->io_bytes[idx] += bytes;
817 td->this_io_bytes[idx] += bytes;
819 io_u->file->last_completed_pos = io_u->endpos;
821 usec = utime_since(&io_u->issue_time, &icd->time);
823 add_clat_sample(td, idx, usec);
824 add_bw_sample(td, idx, &icd->time);
825 io_u_mark_latency(td, usec);
827 if (td_write(td) && idx == DDIR_WRITE &&
829 td->o.verify != VERIFY_NONE)
830 log_io_piece(td, io_u);
832 icd->bytes_done[idx] += bytes;
835 ret = io_u->end_io(td, io_u);
836 if (ret && !icd->error)
840 icd->error = io_u->error;
841 io_u_log_error(td, io_u);
845 static void init_icd(struct io_completion_data *icd, int nr)
847 fio_gettime(&icd->time, NULL);
852 icd->bytes_done[0] = icd->bytes_done[1] = 0;
855 static void ios_completed(struct thread_data *td,
856 struct io_completion_data *icd)
861 for (i = 0; i < icd->nr; i++) {
862 io_u = td->io_ops->event(td, i);
864 io_completed(td, io_u, icd);
870 * Complete a single io_u for the sync engines.
872 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
874 struct io_completion_data icd;
877 io_completed(td, io_u, &icd);
881 return icd.bytes_done[0] + icd.bytes_done[1];
883 td_verror(td, icd.error, "io_u_sync_complete");
888 * Called to complete min_events number of io for the async engines.
890 long io_u_queued_complete(struct thread_data *td, int min_events)
892 struct io_completion_data icd;
893 struct timespec *tvp = NULL;
895 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
897 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
902 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
904 td_verror(td, -ret, "td_io_getevents");
910 ios_completed(td, &icd);
912 return icd.bytes_done[0] + icd.bytes_done[1];
914 td_verror(td, icd.error, "io_u_queued_complete");
919 * Call when io_u is really queued, to update the submission latency.
921 void io_u_queued(struct thread_data *td, struct io_u *io_u)
923 unsigned long slat_time;
925 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
926 add_slat_sample(td, io_u->ddir, slat_time);
929 #ifdef FIO_USE_TIMEOUT
930 void io_u_set_timeout(struct thread_data *td)
932 assert(td->cur_depth);
934 td->timer.it_interval.tv_sec = 0;
935 td->timer.it_interval.tv_usec = 0;
936 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
937 td->timer.it_value.tv_usec = 0;
938 setitimer(ITIMER_REAL, &td->timer, NULL);
939 fio_gettime(&td->timeout_end, NULL);
942 static void io_u_dump(struct io_u *io_u)
944 unsigned long t_start = mtime_since_now(&io_u->start_time);
945 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
947 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
948 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);
949 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
952 void io_u_set_timeout(struct thread_data fio_unused *td)
957 #ifdef FIO_USE_TIMEOUT
958 static void io_u_timeout_handler(int fio_unused sig)
960 struct thread_data *td, *__td;
961 pid_t pid = getpid();
962 struct list_head *entry;
966 log_err("fio: io_u timeout\n");
969 * TLS would be nice...
972 for_each_td(__td, i) {
973 if (__td->pid == pid) {
980 log_err("fio: io_u timeout, can't find job\n");
984 if (!td->cur_depth) {
985 log_err("fio: timeout without pending work?\n");
989 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
991 list_for_each(entry, &td->io_u_busylist) {
992 io_u = list_entry(entry, struct io_u, list);
997 td_verror(td, ETIMEDOUT, "io_u timeout");
1002 void io_u_init_timeout(void)
1004 #ifdef FIO_USE_TIMEOUT
1005 signal(SIGALRM, io_u_timeout_handler);