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 fio_file *f, const unsigned long long block)
30 unsigned int idx = RAND_MAP_IDX(f, block);
31 unsigned int bit = RAND_MAP_BIT(f, block);
33 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
35 return (f->file_map[idx] & (1UL << bit)) == 0;
39 * Mark a given offset as used in the map.
41 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
43 unsigned int min_bs = td->o.rw_min_bs;
44 struct fio_file *f = io_u->file;
45 unsigned long long block;
47 unsigned int nr_blocks;
49 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
51 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
53 while (blocks < nr_blocks) {
54 unsigned int idx, bit;
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
60 if ((td->o.ddir_nr == 1) && !random_map_free(f, block))
63 idx = RAND_MAP_IDX(f, block);
64 bit = RAND_MAP_BIT(f, block);
66 fio_assert(td, idx < f->num_maps);
68 f->file_map[idx] |= (1UL << bit);
73 if ((blocks * min_bs) < io_u->buflen)
74 io_u->buflen = blocks * min_bs;
77 static inline unsigned long long last_block(struct thread_data *td,
81 unsigned long long max_blocks;
83 max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
91 * Return the next free block in the map.
93 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
94 enum fio_ddir ddir, unsigned long long *b)
96 unsigned long long min_bs = td->o.rw_min_bs;
99 i = f->last_free_lookup;
100 *b = (i * BLOCKS_PER_MAP);
101 while ((*b) * min_bs < f->real_file_size) {
102 if (f->file_map[i] != -1UL) {
103 *b += fio_ffz(f->file_map[i]);
104 if (*b > last_block(td, f, ddir))
106 f->last_free_lookup = i;
110 *b += BLOCKS_PER_MAP;
114 dprint(FD_IO, "failed finding a free block\n");
118 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
119 enum fio_ddir ddir, unsigned long long *b)
121 unsigned long long r;
125 r = os_random_long(&td->random_state);
126 dprint(FD_RANDOM, "off rand %llu\n", r);
127 *b = (last_block(td, f, ddir) - 1)
128 * (r / ((unsigned long long) RAND_MAX + 1.0));
131 * if we are not maintaining a random map, we are done.
133 if (!file_randommap(td, f))
137 * calculate map offset and check if it's free
139 if (random_map_free(f, *b))
142 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
147 * we get here, if we didn't suceed in looking up a block. generate
148 * a random start offset into the filemap, and find the first free
153 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
154 if (!get_next_free_block(td, f, ddir, b))
157 r = os_random_long(&td->random_state);
161 * that didn't work either, try exhaustive search from the start
163 f->last_free_lookup = 0;
164 return get_next_free_block(td, f, ddir, b);
168 * For random io, generate a random new block and see if it's used. Repeat
169 * until we find a free one. For sequential io, just return the end of
170 * the last io issued.
172 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
174 struct fio_file *f = io_u->file;
175 unsigned long long b;
176 enum fio_ddir ddir = io_u->ddir;
178 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
179 td->ddir_nr = td->o.ddir_nr;
181 if (get_next_rand_offset(td, f, ddir, &b))
184 if (f->last_pos >= f->real_file_size) {
185 if (!td_random(td) ||
186 get_next_rand_offset(td, f, ddir, &b))
189 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
192 io_u->offset = b * td->o.min_bs[ddir];
193 if (io_u->offset >= f->io_size) {
194 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
195 io_u->offset, f->io_size);
199 io_u->offset += f->file_offset;
200 if (io_u->offset >= f->real_file_size) {
201 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
202 io_u->offset, f->real_file_size);
209 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
211 const int ddir = io_u->ddir;
215 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
216 buflen = td->o.min_bs[ddir];
218 r = os_random_long(&td->bsrange_state);
219 if (!td->o.bssplit_nr) {
220 buflen = (unsigned int)
221 (1 + (double) (td->o.max_bs[ddir] - 1)
222 * r / (RAND_MAX + 1.0));
227 for (i = 0; i < td->o.bssplit_nr; i++) {
228 struct bssplit *bsp = &td->o.bssplit[i];
232 if (r <= ((LONG_MAX / 100L) * perc))
236 if (!td->o.bs_unaligned) {
237 buflen = (buflen + td->o.min_bs[ddir] - 1)
238 & ~(td->o.min_bs[ddir] - 1);
242 if (io_u->offset + buflen > io_u->file->real_file_size) {
243 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
244 td->o.min_bs[ddir], ddir);
245 buflen = td->o.min_bs[ddir];
251 static void set_rwmix_bytes(struct thread_data *td)
253 unsigned long issues;
257 * we do time or byte based switch. this is needed because
258 * buffered writes may issue a lot quicker than they complete,
259 * whereas reads do not.
261 issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
262 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
264 td->rwmix_issues = td->io_issues[td->rwmix_ddir]
265 + (issues * ((100 - diff)) / diff);
268 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
273 r = os_random_long(&td->rwmix_state);
274 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
275 if (v < td->o.rwmix[DDIR_READ])
282 * Return the data direction for the next io_u. If the job is a
283 * mixed read/write workload, check the rwmix cycle and switch if
286 static enum fio_ddir get_rw_ddir(struct thread_data *td)
290 * Check if it's time to seed a new data direction.
292 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
293 unsigned long long max_bytes;
297 * Put a top limit on how many bytes we do for
298 * one data direction, to avoid overflowing the
301 ddir = get_rand_ddir(td);
302 max_bytes = td->this_io_bytes[ddir];
304 (td->o.size * td->o.rwmix[ddir] / 100)) {
305 if (!td->rw_end_set[ddir])
306 td->rw_end_set[ddir] = 1;
311 if (ddir != td->rwmix_ddir)
314 td->rwmix_ddir = ddir;
316 return td->rwmix_ddir;
317 } else if (td_read(td))
323 static void put_file_log(struct thread_data *td, struct fio_file *f)
325 int ret = put_file(td, f);
328 td_verror(td, ret, "file close");
331 void put_io_u(struct thread_data *td, struct io_u *io_u)
333 assert((io_u->flags & IO_U_F_FREE) == 0);
334 io_u->flags |= IO_U_F_FREE;
337 put_file_log(td, io_u->file);
340 list_del(&io_u->list);
341 list_add(&io_u->list, &td->io_u_freelist);
345 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
347 struct io_u *__io_u = *io_u;
349 __io_u->flags |= IO_U_F_FREE;
350 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
351 td->io_issues[__io_u->ddir]--;
353 __io_u->flags &= ~IO_U_F_FLIGHT;
355 list_del(&__io_u->list);
356 list_add_tail(&__io_u->list, &td->io_u_requeues);
361 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
363 if (td->io_ops->flags & FIO_NOIO)
367 * see if it's time to sync
369 if (td->o.fsync_blocks &&
370 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
371 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
372 io_u->ddir = DDIR_SYNC;
376 io_u->ddir = get_rw_ddir(td);
379 * See if it's time to switch to a new zone
381 if (td->zone_bytes >= td->o.zone_size) {
383 io_u->file->last_pos += td->o.zone_skip;
384 td->io_skip_bytes += td->o.zone_skip;
388 * No log, let the seq/rand engine retrieve the next buflen and
391 if (get_next_offset(td, io_u)) {
392 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
396 io_u->buflen = get_next_buflen(td, io_u);
398 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
402 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
403 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
404 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
405 io_u->buflen, io_u->file->real_file_size);
410 * mark entry before potentially trimming io_u
412 if (td_random(td) && file_randommap(td, io_u->file))
413 mark_random_map(td, io_u);
416 * If using a write iolog, store this entry.
419 dprint_io_u(io_u, "fill_io_u");
420 td->zone_bytes += io_u->buflen;
425 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
429 switch (td->cur_depth) {
451 td->ts.io_u_map[index] += nr;
454 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
491 assert(index < FIO_IO_U_LAT_U_NR);
492 td->ts.io_u_lat_u[index]++;
495 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
536 assert(index < FIO_IO_U_LAT_M_NR);
537 td->ts.io_u_lat_m[index]++;
540 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
543 io_u_mark_lat_usec(td, usec);
545 io_u_mark_lat_msec(td, usec / 1000);
549 * Get next file to service by choosing one at random
551 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
558 long r = os_random_long(&td->next_file_state);
560 fno = (unsigned int) ((double) td->o.nr_files
561 * (r / (RAND_MAX + 1.0)));
563 if (f->flags & FIO_FILE_DONE)
566 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
567 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
574 * Get next file to service by doing round robin between all available ones
576 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
579 unsigned int old_next_file = td->next_file;
583 f = td->files[td->next_file];
586 if (td->next_file >= td->o.nr_files)
589 if (f->flags & FIO_FILE_DONE) {
594 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
598 } while (td->next_file != old_next_file);
600 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
604 static struct fio_file *get_next_file(struct thread_data *td)
608 assert(td->o.nr_files <= td->files_index);
610 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
611 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
612 " nr_files=%d\n", td->nr_open_files,
618 f = td->file_service_file;
619 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
622 if (td->o.file_service_type == FIO_FSERVICE_RR)
623 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
625 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
627 td->file_service_file = f;
628 td->file_service_left = td->file_service_nr - 1;
630 dprint(FD_FILE, "get_next_file: %p\n", f);
634 static struct fio_file *find_next_new_file(struct thread_data *td)
638 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
641 if (td->o.file_service_type == FIO_FSERVICE_RR)
642 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
644 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
649 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
654 f = get_next_file(td);
662 if (!fill_io_u(td, io_u))
666 * optimization to prevent close/open of the same file. This
667 * way we preserve queueing etc.
669 if (td->o.nr_files == 1 && td->o.time_based) {
676 * td_io_close() does a put_file() as well, so no need to
680 td_io_close_file(td, f);
681 f->flags |= FIO_FILE_DONE;
685 * probably not the right place to do this, but see
686 * if we need to open a new file
688 if (td->nr_open_files < td->o.open_files &&
689 td->o.open_files != td->o.nr_files) {
690 f = find_next_new_file(td);
692 if (!f || td_io_open_file(td, f))
703 struct io_u *__get_io_u(struct thread_data *td)
705 struct io_u *io_u = NULL;
707 if (!list_empty(&td->io_u_requeues))
708 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
709 else if (!queue_full(td)) {
710 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
719 assert(io_u->flags & IO_U_F_FREE);
720 io_u->flags &= ~IO_U_F_FREE;
723 list_del(&io_u->list);
724 list_add(&io_u->list, &td->io_u_busylist);
732 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
733 * etc. The returned io_u is fully ready to be prepped and submitted.
735 struct io_u *get_io_u(struct thread_data *td)
740 io_u = __get_io_u(td);
742 dprint(FD_IO, "__get_io_u failed\n");
747 * from a requeue, io_u already setup
753 * If using an iolog, grab next piece if any available.
755 if (td->o.read_iolog_file) {
756 if (read_iolog_get(td, io_u))
758 } else if (set_io_u_file(td, io_u)) {
759 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
764 assert(f->flags & FIO_FILE_OPEN);
766 if (io_u->ddir != DDIR_SYNC) {
767 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
768 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
772 f->last_pos = io_u->offset + io_u->buflen;
774 if (td->o.verify != VERIFY_NONE)
775 populate_verify_io_u(td, io_u);
779 * Set io data pointers.
781 io_u->endpos = io_u->offset + io_u->buflen;
782 io_u->xfer_buf = io_u->buf;
783 io_u->xfer_buflen = io_u->buflen;
785 if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
786 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
788 if (!td_io_prep(td, io_u)) {
789 fio_gettime(&io_u->start_time, NULL);
793 dprint(FD_IO, "get_io_u failed\n");
798 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
800 const char *msg[] = { "read", "write", "sync" };
802 log_err("fio: io_u error");
805 log_err(" on file %s", io_u->file->file_name);
807 log_err(": %s\n", strerror(io_u->error));
809 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
810 io_u->offset, io_u->xfer_buflen);
813 td_verror(td, io_u->error, "io_u error");
816 static void io_completed(struct thread_data *td, struct io_u *io_u,
817 struct io_completion_data *icd)
821 dprint_io_u(io_u, "io complete");
823 assert(io_u->flags & IO_U_F_FLIGHT);
824 io_u->flags &= ~IO_U_F_FLIGHT;
826 if (io_u->ddir == DDIR_SYNC) {
827 td->last_was_sync = 1;
831 td->last_was_sync = 0;
834 unsigned int bytes = io_u->buflen - io_u->resid;
835 const enum fio_ddir idx = io_u->ddir;
838 td->io_blocks[idx]++;
839 td->io_bytes[idx] += bytes;
840 td->this_io_bytes[idx] += bytes;
842 usec = utime_since(&io_u->issue_time, &icd->time);
844 add_clat_sample(td, idx, usec);
845 add_bw_sample(td, idx, &icd->time);
846 io_u_mark_latency(td, usec);
848 if (td_write(td) && idx == DDIR_WRITE &&
850 td->o.verify != VERIFY_NONE)
851 log_io_piece(td, io_u);
853 icd->bytes_done[idx] += bytes;
856 ret = io_u->end_io(td, io_u);
857 if (ret && !icd->error)
861 icd->error = io_u->error;
862 io_u_log_error(td, io_u);
866 static void init_icd(struct io_completion_data *icd, int nr)
868 fio_gettime(&icd->time, NULL);
873 icd->bytes_done[0] = icd->bytes_done[1] = 0;
876 static void ios_completed(struct thread_data *td,
877 struct io_completion_data *icd)
882 for (i = 0; i < icd->nr; i++) {
883 io_u = td->io_ops->event(td, i);
885 io_completed(td, io_u, icd);
891 * Complete a single io_u for the sync engines.
893 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
895 struct io_completion_data icd;
898 io_completed(td, io_u, &icd);
902 return icd.bytes_done[0] + icd.bytes_done[1];
904 td_verror(td, icd.error, "io_u_sync_complete");
909 * Called to complete min_events number of io for the async engines.
911 long io_u_queued_complete(struct thread_data *td, int min_events)
913 struct io_completion_data icd;
914 struct timespec *tvp = NULL;
916 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
918 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
923 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
925 td_verror(td, -ret, "td_io_getevents");
931 ios_completed(td, &icd);
933 return icd.bytes_done[0] + icd.bytes_done[1];
935 td_verror(td, icd.error, "io_u_queued_complete");
940 * Call when io_u is really queued, to update the submission latency.
942 void io_u_queued(struct thread_data *td, struct io_u *io_u)
944 unsigned long slat_time;
946 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
947 add_slat_sample(td, io_u->ddir, slat_time);
951 * "randomly" fill the buffer contents
953 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
956 long *ptr = io_u->buf;
958 if (!td->o.zero_buffers) {
959 while ((void *) ptr - io_u->buf < max_bs) {
960 *ptr = rand() * GOLDEN_RATIO_PRIME;
964 memset(ptr, 0, max_bs);
967 #ifdef FIO_USE_TIMEOUT
968 void io_u_set_timeout(struct thread_data *td)
970 assert(td->cur_depth);
972 td->timer.it_interval.tv_sec = 0;
973 td->timer.it_interval.tv_usec = 0;
974 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
975 td->timer.it_value.tv_usec = 0;
976 setitimer(ITIMER_REAL, &td->timer, NULL);
977 fio_gettime(&td->timeout_end, NULL);
980 static void io_u_dump(struct io_u *io_u)
982 unsigned long t_start = mtime_since_now(&io_u->start_time);
983 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
985 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
986 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
987 io_u->xfer_buf, io_u->buflen,
990 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
993 void io_u_set_timeout(struct thread_data fio_unused *td)
998 #ifdef FIO_USE_TIMEOUT
999 static void io_u_timeout_handler(int fio_unused sig)
1001 struct thread_data *td, *__td;
1002 pid_t pid = getpid();
1003 struct list_head *entry;
1007 log_err("fio: io_u timeout\n");
1010 * TLS would be nice...
1013 for_each_td(__td, i) {
1014 if (__td->pid == pid) {
1021 log_err("fio: io_u timeout, can't find job\n");
1025 if (!td->cur_depth) {
1026 log_err("fio: timeout without pending work?\n");
1030 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1032 list_for_each(entry, &td->io_u_busylist) {
1033 io_u = list_entry(entry, struct io_u, list);
1038 td_verror(td, ETIMEDOUT, "io_u timeout");
1043 void io_u_init_timeout(void)
1045 #ifdef FIO_USE_TIMEOUT
1046 signal(SIGALRM, io_u_timeout_handler);