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 / td->o.min_bs[ddir];
173 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
174 if (io_u->offset >= f->real_file_size)
180 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
182 const int ddir = io_u->ddir;
186 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
187 buflen = td->o.min_bs[ddir];
189 r = os_random_long(&td->bsrange_state);
190 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
191 if (!td->o.bs_unaligned)
192 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
195 if (io_u->offset + buflen > io_u->file->real_file_size)
196 buflen = td->o.min_bs[ddir];
201 static void set_rwmix_bytes(struct thread_data *td)
203 unsigned long long rbytes;
207 * we do time or byte based switch. this is needed because
208 * buffered writes may issue a lot quicker than they complete,
209 * whereas reads do not.
211 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
212 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
214 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
217 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
222 r = os_random_long(&td->rwmix_state);
223 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
224 if (v < td->o.rwmix[DDIR_READ])
231 * Return the data direction for the next io_u. If the job is a
232 * mixed read/write workload, check the rwmix cycle and switch if
235 static enum fio_ddir get_rw_ddir(struct thread_data *td)
239 unsigned long elapsed;
242 fio_gettime(&now, NULL);
243 elapsed = mtime_since_now(&td->rwmix_switch);
246 * if this is the first cycle, make it shorter
248 cycle = td->o.rwmixcycle;
249 if (!td->rwmix_bytes)
253 * Check if it's time to seed a new data direction.
255 if (elapsed >= cycle ||
256 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
257 unsigned long long max_bytes;
261 * Put a top limit on how many bytes we do for
262 * one data direction, to avoid overflowing the
265 ddir = get_rand_ddir(td);
266 max_bytes = td->this_io_bytes[ddir];
267 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
268 if (!td->rw_end_set[ddir]) {
269 td->rw_end_set[ddir] = 1;
270 memcpy(&td->rw_end[ddir], &now, sizeof(now));
275 if (ddir != td->rwmix_ddir)
278 td->rwmix_ddir = ddir;
279 memcpy(&td->rwmix_switch, &now, sizeof(now));
281 return td->rwmix_ddir;
282 } else if (td_read(td))
288 void put_io_u(struct thread_data *td, struct io_u *io_u)
290 assert((io_u->flags & IO_U_F_FREE) == 0);
291 io_u->flags |= IO_U_F_FREE;
294 put_file(td, io_u->file);
297 list_del(&io_u->list);
298 list_add(&io_u->list, &td->io_u_freelist);
302 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
304 struct io_u *__io_u = *io_u;
306 __io_u->flags |= IO_U_F_FREE;
307 __io_u->flags &= ~IO_U_F_FLIGHT;
309 list_del(&__io_u->list);
310 list_add_tail(&__io_u->list, &td->io_u_requeues);
315 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
318 * see if it's time to sync
320 if (td->o.fsync_blocks &&
321 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
322 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
323 io_u->ddir = DDIR_SYNC;
327 io_u->ddir = get_rw_ddir(td);
330 * See if it's time to switch to a new zone
332 if (td->zone_bytes >= td->o.zone_size) {
334 io_u->file->last_pos += td->o.zone_skip;
335 td->io_skip_bytes += td->o.zone_skip;
339 * No log, let the seq/rand engine retrieve the next buflen and
342 if (get_next_offset(td, io_u))
345 io_u->buflen = get_next_buflen(td, io_u);
349 if (io_u->offset + io_u->buflen > io_u->file->real_file_size)
353 * mark entry before potentially trimming io_u
355 if (td_random(td) && !td->o.norandommap)
356 mark_random_map(td, io_u);
359 * If using a write iolog, store this entry.
366 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
370 if (io_u->ddir == DDIR_SYNC)
373 switch (td->cur_depth) {
395 td->ts.io_u_map[index]++;
396 td->ts.total_io_u[io_u->ddir]++;
399 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
436 assert(index < FIO_IO_U_LAT_U_NR);
437 td->ts.io_u_lat_u[index]++;
440 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
481 assert(index < FIO_IO_U_LAT_M_NR);
482 td->ts.io_u_lat_m[index]++;
485 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
488 io_u_mark_lat_usec(td, usec);
490 io_u_mark_lat_msec(td, usec / 1000);
494 * Get next file to service by choosing one at random
496 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
503 long r = os_random_long(&td->next_file_state);
505 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
507 if (f->flags & FIO_FILE_DONE)
510 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
516 * Get next file to service by doing round robin between all available ones
518 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
521 unsigned int old_next_file = td->next_file;
525 f = &td->files[td->next_file];
528 if (td->next_file >= td->o.nr_files)
531 if (f->flags & FIO_FILE_DONE) {
536 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
540 } while (td->next_file != old_next_file);
545 static struct fio_file *get_next_file(struct thread_data *td)
549 assert(td->o.nr_files <= td->files_index);
551 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
554 f = td->file_service_file;
555 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
558 if (td->o.file_service_type == FIO_FSERVICE_RR)
559 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
561 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
563 td->file_service_file = f;
564 td->file_service_left = td->file_service_nr - 1;
568 static struct fio_file *find_next_new_file(struct thread_data *td)
572 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
575 if (td->o.file_service_type == FIO_FSERVICE_RR)
576 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
578 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
583 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
588 f = get_next_file(td);
596 if (!fill_io_u(td, io_u))
600 * td_io_close() does a put_file() as well, so no need to
604 td_io_close_file(td, f);
605 f->flags |= FIO_FILE_DONE;
609 * probably not the right place to do this, but see
610 * if we need to open a new file
612 if (td->nr_open_files < td->o.open_files &&
613 td->o.open_files != td->o.nr_files) {
614 f = find_next_new_file(td);
616 if (!f || td_io_open_file(td, f))
627 struct io_u *__get_io_u(struct thread_data *td)
629 struct io_u *io_u = NULL;
631 if (!list_empty(&td->io_u_requeues))
632 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
633 else if (!queue_full(td)) {
634 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
643 assert(io_u->flags & IO_U_F_FREE);
644 io_u->flags &= ~IO_U_F_FREE;
647 list_del(&io_u->list);
648 list_add(&io_u->list, &td->io_u_busylist);
656 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
657 * etc. The returned io_u is fully ready to be prepped and submitted.
659 struct io_u *get_io_u(struct thread_data *td)
664 io_u = __get_io_u(td);
669 * from a requeue, io_u already setup
675 * If using an iolog, grab next piece if any available.
677 if (td->o.read_iolog_file) {
678 if (read_iolog_get(td, io_u))
680 } else if (set_io_u_file(td, io_u))
684 assert(f->flags & FIO_FILE_OPEN);
686 if (io_u->ddir != DDIR_SYNC) {
690 f->last_pos = io_u->offset + io_u->buflen;
692 if (td->o.verify != VERIFY_NONE)
693 populate_verify_io_u(td, io_u);
697 * Set io data pointers.
699 io_u->endpos = io_u->offset + io_u->buflen;
701 io_u->xfer_buf = io_u->buf;
702 io_u->xfer_buflen = io_u->buflen;
704 if (!td_io_prep(td, io_u)) {
705 fio_gettime(&io_u->start_time, NULL);
713 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
715 const char *msg[] = { "read", "write", "sync" };
717 log_err("fio: io_u error");
720 log_err(" on file %s", io_u->file->file_name);
722 log_err(": %s\n", strerror(io_u->error));
724 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
727 td_verror(td, io_u->error, "io_u error");
730 static void io_completed(struct thread_data *td, struct io_u *io_u,
731 struct io_completion_data *icd)
735 assert(io_u->flags & IO_U_F_FLIGHT);
736 io_u->flags &= ~IO_U_F_FLIGHT;
738 if (io_u->ddir == DDIR_SYNC) {
739 td->last_was_sync = 1;
743 td->last_was_sync = 0;
746 unsigned int bytes = io_u->buflen - io_u->resid;
747 const enum fio_ddir idx = io_u->ddir;
750 td->io_blocks[idx]++;
751 td->io_bytes[idx] += bytes;
752 td->zone_bytes += bytes;
753 td->this_io_bytes[idx] += bytes;
755 io_u->file->last_completed_pos = io_u->endpos;
757 usec = utime_since(&io_u->issue_time, &icd->time);
759 add_clat_sample(td, idx, usec);
760 add_bw_sample(td, idx, &icd->time);
761 io_u_mark_latency(td, usec);
763 if (td_write(td) && idx == DDIR_WRITE &&
764 td->o.verify != VERIFY_NONE)
765 log_io_piece(td, io_u);
767 icd->bytes_done[idx] += bytes;
770 ret = io_u->end_io(td, io_u);
771 if (ret && !icd->error)
775 icd->error = io_u->error;
776 io_u_log_error(td, io_u);
780 static void init_icd(struct io_completion_data *icd, int nr)
782 fio_gettime(&icd->time, NULL);
787 icd->bytes_done[0] = icd->bytes_done[1] = 0;
790 static void ios_completed(struct thread_data *td,
791 struct io_completion_data *icd)
796 for (i = 0; i < icd->nr; i++) {
797 io_u = td->io_ops->event(td, i);
799 io_completed(td, io_u, icd);
805 * Complete a single io_u for the sync engines.
807 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
809 struct io_completion_data icd;
812 io_completed(td, io_u, &icd);
816 return icd.bytes_done[0] + icd.bytes_done[1];
818 td_verror(td, icd.error, "io_u_sync_complete");
823 * Called to complete min_events number of io for the async engines.
825 long io_u_queued_complete(struct thread_data *td, int min_events)
827 struct io_completion_data icd;
828 struct timespec *tvp = NULL;
830 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
835 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
837 td_verror(td, -ret, "td_io_getevents");
843 ios_completed(td, &icd);
845 return icd.bytes_done[0] + icd.bytes_done[1];
847 td_verror(td, icd.error, "io_u_queued_complete");
852 * Call when io_u is really queued, to update the submission latency.
854 void io_u_queued(struct thread_data *td, struct io_u *io_u)
856 unsigned long slat_time;
858 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
859 add_slat_sample(td, io_u->ddir, slat_time);
862 #ifdef FIO_USE_TIMEOUT
863 void io_u_set_timeout(struct thread_data *td)
865 assert(td->cur_depth);
867 td->timer.it_interval.tv_sec = 0;
868 td->timer.it_interval.tv_usec = 0;
869 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
870 td->timer.it_value.tv_usec = 0;
871 setitimer(ITIMER_REAL, &td->timer, NULL);
872 fio_gettime(&td->timeout_end, NULL);
875 static void io_u_dump(struct io_u *io_u)
877 unsigned long t_start = mtime_since_now(&io_u->start_time);
878 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
880 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
881 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);
882 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
885 void io_u_set_timeout(struct thread_data fio_unused *td)
890 #ifdef FIO_USE_TIMEOUT
891 static void io_u_timeout_handler(int fio_unused sig)
893 struct thread_data *td, *__td;
894 pid_t pid = getpid();
895 struct list_head *entry;
899 log_err("fio: io_u timeout\n");
902 * TLS would be nice...
905 for_each_td(__td, i) {
906 if (__td->pid == pid) {
913 log_err("fio: io_u timeout, can't find job\n");
917 if (!td->cur_depth) {
918 log_err("fio: timeout without pending work?\n");
922 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
924 list_for_each(entry, &td->io_u_busylist) {
925 io_u = list_entry(entry, struct io_u, list);
930 td_verror(td, ETIMEDOUT, "io_u timeout");
935 void io_u_init_timeout(void)
937 #ifdef FIO_USE_TIMEOUT
938 signal(SIGALRM, io_u_timeout_handler);