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 fio_file *f, const unsigned long long block)
29 unsigned int idx = RAND_MAP_IDX(f, block);
30 unsigned int bit = RAND_MAP_BIT(f, block);
32 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
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 - f->file_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;
56 * If we have a mixed random workload, we may
57 * encounter blocks we already did IO to.
59 if ((td->o.ddir_nr == 1) && !random_map_free(f, block))
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
65 fio_assert(td, idx < f->num_maps);
67 f->file_map[idx] |= (1UL << bit);
72 if ((blocks * min_bs) < io_u->buflen)
73 io_u->buflen = blocks * min_bs;
76 static inline unsigned long long last_block(struct thread_data *td,
80 unsigned long long max_blocks;
82 max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
90 * Return the next free block in the map.
92 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
93 enum fio_ddir ddir, unsigned long long *b)
95 unsigned long long min_bs = td->o.rw_min_bs;
98 i = f->last_free_lookup;
99 *b = (i * BLOCKS_PER_MAP);
100 while ((*b) * min_bs < f->real_file_size) {
101 if (f->file_map[i] != -1UL) {
102 *b += fio_ffz(f->file_map[i]);
103 if (*b > last_block(td, f, ddir))
105 f->last_free_lookup = i;
109 *b += BLOCKS_PER_MAP;
113 dprint(FD_IO, "failed finding a free block\n");
117 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
118 enum fio_ddir ddir, unsigned long long *b)
120 unsigned long long r;
124 r = os_random_long(&td->random_state);
125 dprint(FD_RANDOM, "off rand %llu\n", r);
126 *b = (last_block(td, f, ddir) - 1)
127 * (r / ((unsigned long long) RAND_MAX + 1.0));
130 * if we are not maintaining a random map, we are done.
132 if (!file_randommap(td, f))
136 * calculate map offset and check if it's free
138 if (random_map_free(f, *b))
141 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
146 * we get here, if we didn't suceed in looking up a block. generate
147 * a random start offset into the filemap, and find the first free
152 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
153 if (!get_next_free_block(td, f, ddir, b))
156 r = os_random_long(&td->random_state);
160 * that didn't work either, try exhaustive search from the start
162 f->last_free_lookup = 0;
163 return get_next_free_block(td, f, ddir, b);
167 * For random io, generate a random new block and see if it's used. Repeat
168 * until we find a free one. For sequential io, just return the end of
169 * the last io issued.
171 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
173 struct fio_file *f = io_u->file;
174 unsigned long long b;
175 enum fio_ddir ddir = io_u->ddir;
177 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
178 td->ddir_nr = td->o.ddir_nr;
180 if (get_next_rand_offset(td, f, ddir, &b))
183 if (f->last_pos >= f->real_file_size) {
184 if (!td_random(td) ||
185 get_next_rand_offset(td, f, ddir, &b))
188 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
191 io_u->offset = b * td->o.min_bs[ddir];
192 if (io_u->offset >= f->io_size) {
193 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
194 io_u->offset, f->io_size);
198 io_u->offset += f->file_offset;
199 if (io_u->offset >= f->real_file_size) {
200 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
201 io_u->offset, f->real_file_size);
208 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
210 const int ddir = io_u->ddir;
214 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
215 buflen = td->o.min_bs[ddir];
217 r = os_random_long(&td->bsrange_state);
218 if (!td->o.bssplit_nr) {
219 buflen = (unsigned int)
220 (1 + (double) (td->o.max_bs[ddir] - 1)
221 * r / (RAND_MAX + 1.0));
226 for (i = 0; i < td->o.bssplit_nr; i++) {
227 struct bssplit *bsp = &td->o.bssplit[i];
231 if (r <= ((LONG_MAX / 100L) * perc))
235 if (!td->o.bs_unaligned) {
236 buflen = (buflen + td->o.min_bs[ddir] - 1)
237 & ~(td->o.min_bs[ddir] - 1);
241 if (io_u->offset + buflen > io_u->file->real_file_size) {
242 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
243 td->o.min_bs[ddir], ddir);
244 buflen = td->o.min_bs[ddir];
250 static void set_rwmix_bytes(struct thread_data *td)
252 unsigned long issues;
256 * we do time or byte based switch. this is needed because
257 * buffered writes may issue a lot quicker than they complete,
258 * whereas reads do not.
260 issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
261 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
263 td->rwmix_issues = td->io_issues[td->rwmix_ddir]
264 + (issues * ((100 - diff)) / diff);
267 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
272 r = os_random_long(&td->rwmix_state);
273 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
274 if (v < td->o.rwmix[DDIR_READ])
281 * Return the data direction for the next io_u. If the job is a
282 * mixed read/write workload, check the rwmix cycle and switch if
285 static enum fio_ddir get_rw_ddir(struct thread_data *td)
289 * Check if it's time to seed a new data direction.
291 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
292 unsigned long long max_bytes;
296 * Put a top limit on how many bytes we do for
297 * one data direction, to avoid overflowing the
300 ddir = get_rand_ddir(td);
301 max_bytes = td->this_io_bytes[ddir];
303 (td->o.size * td->o.rwmix[ddir] / 100)) {
304 if (!td->rw_end_set[ddir])
305 td->rw_end_set[ddir] = 1;
310 if (ddir != td->rwmix_ddir)
313 td->rwmix_ddir = ddir;
315 return td->rwmix_ddir;
316 } else if (td_read(td))
322 void put_io_u(struct thread_data *td, struct io_u *io_u)
324 assert((io_u->flags & IO_U_F_FREE) == 0);
325 io_u->flags |= IO_U_F_FREE;
328 int ret = put_file(td, io_u->file);
331 td_verror(td, ret, "file close");
335 list_del(&io_u->list);
336 list_add(&io_u->list, &td->io_u_freelist);
340 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
342 struct io_u *__io_u = *io_u;
344 __io_u->flags |= IO_U_F_FREE;
345 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
346 td->io_issues[__io_u->ddir]--;
348 __io_u->flags &= ~IO_U_F_FLIGHT;
350 list_del(&__io_u->list);
351 list_add_tail(&__io_u->list, &td->io_u_requeues);
356 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
358 if (td->io_ops->flags & FIO_NOIO)
362 * see if it's time to sync
364 if (td->o.fsync_blocks &&
365 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
366 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
367 io_u->ddir = DDIR_SYNC;
371 io_u->ddir = get_rw_ddir(td);
374 * See if it's time to switch to a new zone
376 if (td->zone_bytes >= td->o.zone_size) {
378 io_u->file->last_pos += td->o.zone_skip;
379 td->io_skip_bytes += td->o.zone_skip;
383 * No log, let the seq/rand engine retrieve the next buflen and
386 if (get_next_offset(td, io_u)) {
387 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
391 io_u->buflen = get_next_buflen(td, io_u);
393 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
397 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
398 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
399 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
400 io_u->buflen, io_u->file->real_file_size);
405 * mark entry before potentially trimming io_u
407 if (td_random(td) && file_randommap(td, io_u->file))
408 mark_random_map(td, io_u);
411 * If using a write iolog, store this entry.
414 dprint_io_u(io_u, "fill_io_u");
415 td->zone_bytes += io_u->buflen;
420 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
424 switch (td->cur_depth) {
446 td->ts.io_u_map[index] += nr;
449 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
486 assert(index < FIO_IO_U_LAT_U_NR);
487 td->ts.io_u_lat_u[index]++;
490 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
531 assert(index < FIO_IO_U_LAT_M_NR);
532 td->ts.io_u_lat_m[index]++;
535 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
538 io_u_mark_lat_usec(td, usec);
540 io_u_mark_lat_msec(td, usec / 1000);
544 * Get next file to service by choosing one at random
546 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
553 long r = os_random_long(&td->next_file_state);
555 fno = (unsigned int) ((double) td->o.nr_files
556 * (r / (RAND_MAX + 1.0)));
558 if (f->flags & FIO_FILE_DONE)
561 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
562 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
569 * Get next file to service by doing round robin between all available ones
571 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
574 unsigned int old_next_file = td->next_file;
578 f = td->files[td->next_file];
581 if (td->next_file >= td->o.nr_files)
584 if (f->flags & FIO_FILE_DONE) {
589 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
593 } while (td->next_file != old_next_file);
595 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
599 static struct fio_file *get_next_file(struct thread_data *td)
603 assert(td->o.nr_files <= td->files_index);
605 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
606 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
607 " nr_files=%d\n", td->nr_open_files,
613 f = td->file_service_file;
614 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
617 if (td->o.file_service_type == FIO_FSERVICE_RR)
618 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
620 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
622 td->file_service_file = f;
623 td->file_service_left = td->file_service_nr - 1;
625 dprint(FD_FILE, "get_next_file: %p\n", f);
629 static struct fio_file *find_next_new_file(struct thread_data *td)
633 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
636 if (td->o.file_service_type == FIO_FSERVICE_RR)
637 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
639 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
644 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
649 f = get_next_file(td);
657 if (!fill_io_u(td, io_u))
661 * optimization to prevent close/open of the same file. This
662 * way we preserve queueing etc.
664 if (td->o.nr_files == 1 && td->o.time_based) {
671 * td_io_close() does a put_file() as well, so no need to
675 td_io_close_file(td, f);
676 f->flags |= FIO_FILE_DONE;
680 * probably not the right place to do this, but see
681 * if we need to open a new file
683 if (td->nr_open_files < td->o.open_files &&
684 td->o.open_files != td->o.nr_files) {
685 f = find_next_new_file(td);
687 if (!f || td_io_open_file(td, f))
698 struct io_u *__get_io_u(struct thread_data *td)
700 struct io_u *io_u = NULL;
702 if (!list_empty(&td->io_u_requeues))
703 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
704 else if (!queue_full(td)) {
705 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
714 assert(io_u->flags & IO_U_F_FREE);
715 io_u->flags &= ~IO_U_F_FREE;
718 list_del(&io_u->list);
719 list_add(&io_u->list, &td->io_u_busylist);
727 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
728 * etc. The returned io_u is fully ready to be prepped and submitted.
730 struct io_u *get_io_u(struct thread_data *td)
735 io_u = __get_io_u(td);
737 dprint(FD_IO, "__get_io_u failed\n");
742 * from a requeue, io_u already setup
748 * If using an iolog, grab next piece if any available.
750 if (td->o.read_iolog_file) {
751 if (read_iolog_get(td, io_u))
753 } else if (set_io_u_file(td, io_u)) {
754 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
759 assert(f->flags & FIO_FILE_OPEN);
761 if (io_u->ddir != DDIR_SYNC) {
762 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
763 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
767 f->last_pos = io_u->offset + io_u->buflen;
769 if (td->o.verify != VERIFY_NONE)
770 populate_verify_io_u(td, io_u);
774 * Set io data pointers.
776 io_u->endpos = io_u->offset + io_u->buflen;
777 io_u->xfer_buf = io_u->buf;
778 io_u->xfer_buflen = io_u->buflen;
780 if (!td_io_prep(td, io_u)) {
781 fio_gettime(&io_u->start_time, NULL);
785 dprint(FD_IO, "get_io_u failed\n");
790 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
792 const char *msg[] = { "read", "write", "sync" };
794 log_err("fio: io_u error");
797 log_err(" on file %s", io_u->file->file_name);
799 log_err(": %s\n", strerror(io_u->error));
801 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
802 io_u->offset, io_u->xfer_buflen);
805 td_verror(td, io_u->error, "io_u error");
808 static void io_completed(struct thread_data *td, struct io_u *io_u,
809 struct io_completion_data *icd)
813 dprint_io_u(io_u, "io complete");
815 assert(io_u->flags & IO_U_F_FLIGHT);
816 io_u->flags &= ~IO_U_F_FLIGHT;
818 if (io_u->ddir == DDIR_SYNC) {
819 td->last_was_sync = 1;
823 td->last_was_sync = 0;
826 unsigned int bytes = io_u->buflen - io_u->resid;
827 const enum fio_ddir idx = io_u->ddir;
830 td->io_blocks[idx]++;
831 td->io_bytes[idx] += bytes;
832 td->this_io_bytes[idx] += bytes;
834 usec = utime_since(&io_u->issue_time, &icd->time);
836 add_clat_sample(td, idx, usec);
837 add_bw_sample(td, idx, &icd->time);
838 io_u_mark_latency(td, usec);
840 if (td_write(td) && idx == DDIR_WRITE &&
842 td->o.verify != VERIFY_NONE)
843 log_io_piece(td, io_u);
845 icd->bytes_done[idx] += bytes;
848 ret = io_u->end_io(td, io_u);
849 if (ret && !icd->error)
853 icd->error = io_u->error;
854 io_u_log_error(td, io_u);
858 static void init_icd(struct io_completion_data *icd, int nr)
860 fio_gettime(&icd->time, NULL);
865 icd->bytes_done[0] = icd->bytes_done[1] = 0;
868 static void ios_completed(struct thread_data *td,
869 struct io_completion_data *icd)
874 for (i = 0; i < icd->nr; i++) {
875 io_u = td->io_ops->event(td, i);
877 io_completed(td, io_u, icd);
883 * Complete a single io_u for the sync engines.
885 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
887 struct io_completion_data icd;
890 io_completed(td, io_u, &icd);
894 return icd.bytes_done[0] + icd.bytes_done[1];
896 td_verror(td, icd.error, "io_u_sync_complete");
901 * Called to complete min_events number of io for the async engines.
903 long io_u_queued_complete(struct thread_data *td, int min_events)
905 struct io_completion_data icd;
906 struct timespec *tvp = NULL;
908 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
910 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
915 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
917 td_verror(td, -ret, "td_io_getevents");
923 ios_completed(td, &icd);
925 return icd.bytes_done[0] + icd.bytes_done[1];
927 td_verror(td, icd.error, "io_u_queued_complete");
932 * Call when io_u is really queued, to update the submission latency.
934 void io_u_queued(struct thread_data *td, struct io_u *io_u)
936 unsigned long slat_time;
938 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
939 add_slat_sample(td, io_u->ddir, slat_time);
942 #ifdef FIO_USE_TIMEOUT
943 void io_u_set_timeout(struct thread_data *td)
945 assert(td->cur_depth);
947 td->timer.it_interval.tv_sec = 0;
948 td->timer.it_interval.tv_usec = 0;
949 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
950 td->timer.it_value.tv_usec = 0;
951 setitimer(ITIMER_REAL, &td->timer, NULL);
952 fio_gettime(&td->timeout_end, NULL);
955 static void io_u_dump(struct io_u *io_u)
957 unsigned long t_start = mtime_since_now(&io_u->start_time);
958 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
960 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
961 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
962 io_u->xfer_buf, io_u->buflen,
965 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
968 void io_u_set_timeout(struct thread_data fio_unused *td)
973 #ifdef FIO_USE_TIMEOUT
974 static void io_u_timeout_handler(int fio_unused sig)
976 struct thread_data *td, *__td;
977 pid_t pid = getpid();
978 struct list_head *entry;
982 log_err("fio: io_u timeout\n");
985 * TLS would be nice...
988 for_each_td(__td, i) {
989 if (__td->pid == pid) {
996 log_err("fio: io_u timeout, can't find job\n");
1000 if (!td->cur_depth) {
1001 log_err("fio: timeout without pending work?\n");
1005 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1007 list_for_each(entry, &td->io_u_busylist) {
1008 io_u = list_entry(entry, struct io_u, list);
1013 td_verror(td, ETIMEDOUT, "io_u timeout");
1018 void io_u_init_timeout(void)
1020 #ifdef FIO_USE_TIMEOUT
1021 signal(SIGALRM, io_u_timeout_handler);