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;
55 if (!td->o.ddir_nr && !random_map_free(td, f, block))
58 idx = RAND_MAP_IDX(td, f, block);
59 bit = RAND_MAP_BIT(td, f, block);
61 fio_assert(td, idx < f->num_maps);
63 f->file_map[idx] |= (1UL << bit);
68 if ((blocks * min_bs) < io_u->buflen)
69 io_u->buflen = blocks * min_bs;
73 * Return the next free block in the map.
75 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
76 unsigned long long *b)
80 i = f->last_free_lookup;
81 *b = (i * BLOCKS_PER_MAP);
82 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
83 if (f->file_map[i] != -1UL) {
84 *b += ffz(f->file_map[i]);
85 f->last_free_lookup = i;
97 * For random io, generate a random new block and see if it's used. Repeat
98 * until we find a free one. For sequential io, just return the end of
101 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
103 struct fio_file *f = io_u->file;
104 const int ddir = io_u->ddir;
105 unsigned long long b, rb;
109 unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir];
114 td->ddir_nr = td->o.ddir_nr;
115 else if (--td->ddir_nr) {
116 b = f->last_pos / td->o.min_bs[ddir];
119 td->ddir_nr = td->o.ddir_nr;
123 r = os_random_long(&td->random_state);
127 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
128 if (td->o.norandommap)
130 rb = b + (f->file_offset / td->o.min_bs[ddir]);
132 } while (!random_map_free(td, f, rb) && loops);
135 * if we failed to retrieve a truly random offset within
136 * the loops assigned, see if there are free ones left at all
138 if (!loops && get_next_free_block(td, f, &b))
141 b = f->last_pos / td->o.min_bs[ddir];
144 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
145 if (io_u->offset >= f->real_file_size)
151 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
153 struct fio_file *f = io_u->file;
154 const int ddir = io_u->ddir;
158 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
159 buflen = td->o.min_bs[ddir];
161 r = os_random_long(&td->bsrange_state);
162 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
163 if (!td->o.bs_unaligned)
164 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
167 while (buflen + io_u->offset > f->real_file_size) {
168 if (buflen == td->o.min_bs[ddir]) {
169 if (!td->o.odirect) {
170 assert(io_u->offset <= f->real_file_size);
171 buflen = f->real_file_size - io_u->offset;
177 buflen = td->o.min_bs[ddir];
183 static void set_rwmix_bytes(struct thread_data *td)
185 unsigned long long rbytes;
189 * we do time or byte based switch. this is needed because
190 * buffered writes may issue a lot quicker than they complete,
191 * whereas reads do not.
193 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
194 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
196 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
199 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
204 r = os_random_long(&td->rwmix_state);
205 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
206 if (v < td->o.rwmix[DDIR_READ])
213 * Return the data direction for the next io_u. If the job is a
214 * mixed read/write workload, check the rwmix cycle and switch if
217 static enum fio_ddir get_rw_ddir(struct thread_data *td)
221 unsigned long elapsed;
224 fio_gettime(&now, NULL);
225 elapsed = mtime_since_now(&td->rwmix_switch);
228 * if this is the first cycle, make it shorter
230 cycle = td->o.rwmixcycle;
231 if (!td->rwmix_bytes)
235 * Check if it's time to seed a new data direction.
237 if (elapsed >= cycle ||
238 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
239 unsigned long long max_bytes;
243 * Put a top limit on how many bytes we do for
244 * one data direction, to avoid overflowing the
247 ddir = get_rand_ddir(td);
248 max_bytes = td->this_io_bytes[ddir];
249 if (max_bytes >= (td->io_size * td->o.rwmix[ddir] / 100)) {
250 if (!td->rw_end_set[ddir]) {
251 td->rw_end_set[ddir] = 1;
252 memcpy(&td->rw_end[ddir], &now, sizeof(now));
257 if (ddir != td->rwmix_ddir)
260 td->rwmix_ddir = ddir;
261 memcpy(&td->rwmix_switch, &now, sizeof(now));
263 return td->rwmix_ddir;
264 } else if (td_read(td))
270 void put_io_u(struct thread_data *td, struct io_u *io_u)
272 assert((io_u->flags & IO_U_F_FREE) == 0);
273 io_u->flags |= IO_U_F_FREE;
276 list_del(&io_u->list);
277 list_add(&io_u->list, &td->io_u_freelist);
281 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
283 struct io_u *__io_u = *io_u;
285 __io_u->flags |= IO_U_F_FREE;
286 __io_u->flags &= ~IO_U_F_FLIGHT;
288 list_del(&__io_u->list);
289 list_add_tail(&__io_u->list, &td->io_u_requeues);
294 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
297 * If using an iolog, grab next piece if any available.
299 if (td->o.read_iolog)
300 return read_iolog_get(td, io_u);
303 * see if it's time to sync
305 if (td->o.fsync_blocks &&
306 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
307 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
308 io_u->ddir = DDIR_SYNC;
312 io_u->ddir = get_rw_ddir(td);
315 * No log, let the seq/rand engine retrieve the next buflen and
318 if (get_next_offset(td, io_u))
321 io_u->buflen = get_next_buflen(td, io_u);
326 * mark entry before potentially trimming io_u
328 if (!td->o.read_iolog && td_random(td) && !td->o.norandommap)
329 mark_random_map(td, io_u);
332 * If using a write iolog, store this entry.
334 if (td->o.write_iolog_file)
335 write_iolog_put(td, io_u);
340 void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
344 if (io_u->ddir == DDIR_SYNC)
347 switch (td->cur_depth) {
364 td->ts.io_u_map[index]++;
365 td->ts.total_io_u[io_u->ddir]++;
368 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
399 td->ts.io_u_lat[index]++;
403 * Get next file to service by choosing one at random
405 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
412 long r = os_random_long(&td->next_file_state);
414 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
417 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
423 * Get next file to service by doing round robin between all available ones
425 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
428 unsigned int old_next_file = td->next_file;
432 f = &td->files[td->next_file];
435 if (td->next_file >= td->o.nr_files)
438 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
442 } while (td->next_file != old_next_file);
447 static struct fio_file *get_next_file(struct thread_data *td)
451 assert(td->o.nr_files <= td->files_index);
453 if (!td->nr_open_files)
456 f = td->file_service_file;
457 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
460 if (td->o.file_service_type == FIO_FSERVICE_RR)
461 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
463 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
465 td->file_service_file = f;
466 td->file_service_left = td->file_service_nr - 1;
470 static struct fio_file *find_next_new_file(struct thread_data *td)
474 if (td->o.file_service_type == FIO_FSERVICE_RR)
475 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
477 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
482 struct io_u *__get_io_u(struct thread_data *td)
484 struct io_u *io_u = NULL;
486 if (!list_empty(&td->io_u_requeues))
487 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
488 else if (!queue_full(td)) {
489 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
498 assert(io_u->flags & IO_U_F_FREE);
499 io_u->flags &= ~IO_U_F_FREE;
502 list_del(&io_u->list);
503 list_add(&io_u->list, &td->io_u_busylist);
511 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
512 * etc. The returned io_u is fully ready to be prepped and submitted.
514 struct io_u *get_io_u(struct thread_data *td)
520 io_u = __get_io_u(td);
525 * from a requeue, io_u already setup
531 f = get_next_file(td);
540 if (!fill_io_u(td, io_u))
544 * No more to do for this file, close it
547 td_io_close_file(td, f);
550 * probably not the right place to do this, but see
551 * if we need to open a new file
553 if (td->nr_open_files < td->o.open_files &&
554 td->o.open_files != td->o.nr_files) {
555 f = find_next_new_file(td);
557 if (!f || (ret = td_io_open_file(td, f))) {
565 if (td->zone_bytes >= td->o.zone_size) {
567 f->last_pos += td->o.zone_skip;
570 if (io_u->buflen + io_u->offset > f->real_file_size) {
571 if (td->io_ops->flags & FIO_RAWIO) {
576 io_u->buflen = f->real_file_size - io_u->offset;
579 if (io_u->ddir != DDIR_SYNC) {
585 f->last_pos = io_u->offset + io_u->buflen;
587 if (td->o.verify != VERIFY_NONE)
588 populate_verify_io_u(td, io_u);
592 * Set io data pointers.
595 io_u->xfer_buf = io_u->buf;
596 io_u->xfer_buflen = io_u->buflen;
598 if (td_io_prep(td, io_u)) {
603 fio_gettime(&io_u->start_time, NULL);
607 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
609 const char *msg[] = { "read", "write", "sync" };
611 log_err("fio: io_u error");
614 log_err(" on file %s", io_u->file->file_name);
616 log_err(": %s\n", strerror(io_u->error));
618 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
621 td_verror(td, io_u->error, "io_u error");
624 static void io_completed(struct thread_data *td, struct io_u *io_u,
625 struct io_completion_data *icd)
629 assert(io_u->flags & IO_U_F_FLIGHT);
630 io_u->flags &= ~IO_U_F_FLIGHT;
632 put_file(td, io_u->file);
634 if (io_u->ddir == DDIR_SYNC) {
635 td->last_was_sync = 1;
639 td->last_was_sync = 0;
642 unsigned int bytes = io_u->buflen - io_u->resid;
643 const enum fio_ddir idx = io_u->ddir;
646 td->io_blocks[idx]++;
647 td->io_bytes[idx] += bytes;
648 td->zone_bytes += bytes;
649 td->this_io_bytes[idx] += bytes;
651 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
653 msec = mtime_since(&io_u->issue_time, &icd->time);
655 add_clat_sample(td, idx, msec);
656 add_bw_sample(td, idx, &icd->time);
657 io_u_mark_latency(td, msec);
659 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
660 log_io_piece(td, io_u);
662 icd->bytes_done[idx] += bytes;
665 ret = io_u->end_io(io_u);
666 if (ret && !icd->error)
670 icd->error = io_u->error;
671 io_u_log_error(td, io_u);
675 static void init_icd(struct io_completion_data *icd, int nr)
677 fio_gettime(&icd->time, NULL);
682 icd->bytes_done[0] = icd->bytes_done[1] = 0;
685 static void ios_completed(struct thread_data *td,
686 struct io_completion_data *icd)
691 for (i = 0; i < icd->nr; i++) {
692 io_u = td->io_ops->event(td, i);
694 io_completed(td, io_u, icd);
700 * Complete a single io_u for the sync engines.
702 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
704 struct io_completion_data icd;
707 io_completed(td, io_u, &icd);
711 return icd.bytes_done[0] + icd.bytes_done[1];
713 td_verror(td, icd.error, "io_u_sync_complete");
718 * Called to complete min_events number of io for the async engines.
720 long io_u_queued_complete(struct thread_data *td, int min_events)
722 struct io_completion_data icd;
723 struct timespec *tvp = NULL;
725 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
730 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
732 td_verror(td, -ret, "td_io_getevents");
738 ios_completed(td, &icd);
740 return icd.bytes_done[0] + icd.bytes_done[1];
742 td_verror(td, icd.error, "io_u_queued_complete");
747 * Call when io_u is really queued, to update the submission latency.
749 void io_u_queued(struct thread_data *td, struct io_u *io_u)
751 unsigned long slat_time;
753 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
754 add_slat_sample(td, io_u->ddir, slat_time);
757 #ifdef FIO_USE_TIMEOUT
758 void io_u_set_timeout(struct thread_data *td)
760 assert(td->cur_depth);
762 td->timer.it_interval.tv_sec = 0;
763 td->timer.it_interval.tv_usec = 0;
764 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
765 td->timer.it_value.tv_usec = 0;
766 setitimer(ITIMER_REAL, &td->timer, NULL);
767 fio_gettime(&td->timeout_end, NULL);
770 static void io_u_dump(struct io_u *io_u)
772 unsigned long t_start = mtime_since_now(&io_u->start_time);
773 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
775 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
776 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);
777 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
780 void io_u_set_timeout(struct thread_data fio_unused *td)
785 #ifdef FIO_USE_TIMEOUT
786 static void io_u_timeout_handler(int fio_unused sig)
788 struct thread_data *td, *__td;
789 pid_t pid = getpid();
790 struct list_head *entry;
794 log_err("fio: io_u timeout\n");
797 * TLS would be nice...
800 for_each_td(__td, i) {
801 if (__td->pid == pid) {
808 log_err("fio: io_u timeout, can't find job\n");
812 if (!td->cur_depth) {
813 log_err("fio: timeout without pending work?\n");
817 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
819 list_for_each(entry, &td->io_u_busylist) {
820 io_u = list_entry(entry, struct io_u, list);
825 td_verror(td, ETIMEDOUT, "io_u timeout");
830 void io_u_init_timeout(void)
832 #ifdef FIO_USE_TIMEOUT
833 signal(SIGALRM, io_u_timeout_handler);