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;
82 unsigned long long max_size;
85 * Hmm, should we make sure that ->io_size <= ->real_file_size?
87 max_size = f->io_size;
88 if (max_size > f->real_file_size)
89 max_size = f->real_file_size;
91 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
99 * Return the next free block in the map.
101 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir, unsigned long long *b)
104 unsigned long long min_bs = td->o.rw_min_bs;
107 i = f->last_free_lookup;
108 *b = (i * BLOCKS_PER_MAP);
109 while ((*b) * min_bs < f->real_file_size) {
110 if (f->file_map[i] != -1UL) {
111 *b += fio_ffz(f->file_map[i]);
112 if (*b > last_block(td, f, ddir))
114 f->last_free_lookup = i;
118 *b += BLOCKS_PER_MAP;
122 dprint(FD_IO, "failed finding a free block\n");
126 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
129 unsigned long long r;
133 r = os_random_long(&td->random_state);
134 dprint(FD_RANDOM, "off rand %llu\n", r);
135 *b = (last_block(td, f, ddir) - 1)
136 * (r / ((unsigned long long) RAND_MAX + 1.0));
139 * if we are not maintaining a random map, we are done.
141 if (!file_randommap(td, f))
145 * calculate map offset and check if it's free
147 if (random_map_free(f, *b))
150 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
155 * we get here, if we didn't suceed in looking up a block. generate
156 * a random start offset into the filemap, and find the first free
161 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
162 if (!get_next_free_block(td, f, ddir, b))
165 r = os_random_long(&td->random_state);
169 * that didn't work either, try exhaustive search from the start
171 f->last_free_lookup = 0;
172 return get_next_free_block(td, f, ddir, b);
176 * For random io, generate a random new block and see if it's used. Repeat
177 * until we find a free one. For sequential io, just return the end of
178 * the last io issued.
180 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
182 struct fio_file *f = io_u->file;
183 unsigned long long b;
184 enum fio_ddir ddir = io_u->ddir;
186 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
187 td->ddir_nr = td->o.ddir_nr;
189 if (get_next_rand_offset(td, f, ddir, &b))
192 if (f->last_pos >= f->real_file_size) {
193 if (!td_random(td) ||
194 get_next_rand_offset(td, f, ddir, &b))
197 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
200 io_u->offset = b * td->o.min_bs[ddir];
201 if (io_u->offset >= f->io_size) {
202 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
203 io_u->offset, f->io_size);
207 io_u->offset += f->file_offset;
208 if (io_u->offset >= f->real_file_size) {
209 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
210 io_u->offset, f->real_file_size);
217 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
219 const int ddir = io_u->ddir;
220 unsigned int buflen = buflen; /* silence dumb gcc warning */
223 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
224 buflen = td->o.min_bs[ddir];
226 r = os_random_long(&td->bsrange_state);
227 if (!td->o.bssplit_nr) {
228 buflen = (unsigned int)
229 (1 + (double) (td->o.max_bs[ddir] - 1)
230 * r / (RAND_MAX + 1.0));
235 for (i = 0; i < td->o.bssplit_nr; i++) {
236 struct bssplit *bsp = &td->o.bssplit[i];
240 if (r <= ((LONG_MAX / 100L) * perc))
244 if (!td->o.bs_unaligned) {
245 buflen = (buflen + td->o.min_bs[ddir] - 1)
246 & ~(td->o.min_bs[ddir] - 1);
250 if (io_u->offset + buflen > io_u->file->real_file_size) {
251 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
252 td->o.min_bs[ddir], ddir);
253 buflen = td->o.min_bs[ddir];
259 static void set_rwmix_bytes(struct thread_data *td)
264 * we do time or byte based switch. this is needed because
265 * buffered writes may issue a lot quicker than they complete,
266 * whereas reads do not.
268 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
269 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
272 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
277 r = os_random_long(&td->rwmix_state);
278 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
279 if (v <= td->o.rwmix[DDIR_READ])
286 * Return the data direction for the next io_u. If the job is a
287 * mixed read/write workload, check the rwmix cycle and switch if
290 static enum fio_ddir get_rw_ddir(struct thread_data *td)
294 * Check if it's time to seed a new data direction.
296 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
297 unsigned long long max_bytes;
301 * Put a top limit on how many bytes we do for
302 * one data direction, to avoid overflowing the
305 ddir = get_rand_ddir(td);
306 max_bytes = td->this_io_bytes[ddir];
308 (td->o.size * td->o.rwmix[ddir] / 100)) {
309 if (!td->rw_end_set[ddir])
310 td->rw_end_set[ddir] = 1;
315 if (ddir != td->rwmix_ddir)
318 td->rwmix_ddir = ddir;
320 return td->rwmix_ddir;
321 } else if (td_read(td))
327 static void put_file_log(struct thread_data *td, struct fio_file *f)
329 int ret = put_file(td, f);
332 td_verror(td, ret, "file close");
335 void put_io_u(struct thread_data *td, struct io_u *io_u)
337 assert((io_u->flags & IO_U_F_FREE) == 0);
338 io_u->flags |= IO_U_F_FREE;
341 put_file_log(td, io_u->file);
344 list_del(&io_u->list);
345 list_add(&io_u->list, &td->io_u_freelist);
349 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
351 struct io_u *__io_u = *io_u;
353 __io_u->flags |= IO_U_F_FREE;
354 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
355 td->io_issues[__io_u->ddir]--;
357 __io_u->flags &= ~IO_U_F_FLIGHT;
359 list_del(&__io_u->list);
360 list_add_tail(&__io_u->list, &td->io_u_requeues);
365 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
367 if (td->io_ops->flags & FIO_NOIO)
371 * see if it's time to sync
373 if (td->o.fsync_blocks &&
374 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
375 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
376 io_u->ddir = DDIR_SYNC;
380 io_u->ddir = get_rw_ddir(td);
383 * See if it's time to switch to a new zone
385 if (td->zone_bytes >= td->o.zone_size) {
387 io_u->file->last_pos += td->o.zone_skip;
388 td->io_skip_bytes += td->o.zone_skip;
392 * No log, let the seq/rand engine retrieve the next buflen and
395 if (get_next_offset(td, io_u)) {
396 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
400 io_u->buflen = get_next_buflen(td, io_u);
402 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
406 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
407 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
408 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
409 io_u->buflen, io_u->file->real_file_size);
414 * mark entry before potentially trimming io_u
416 if (td_random(td) && file_randommap(td, io_u->file))
417 mark_random_map(td, io_u);
420 * If using a write iolog, store this entry.
423 dprint_io_u(io_u, "fill_io_u");
424 td->zone_bytes += io_u->buflen;
429 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
458 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
460 __io_u_mark_map(td->ts.io_u_submit, nr);
461 td->ts.total_submit++;
464 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
466 __io_u_mark_map(td->ts.io_u_complete, nr);
467 td->ts.total_complete++;
470 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
474 switch (td->cur_depth) {
496 td->ts.io_u_map[index] += nr;
499 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
536 assert(index < FIO_IO_U_LAT_U_NR);
537 td->ts.io_u_lat_u[index]++;
540 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
581 assert(index < FIO_IO_U_LAT_M_NR);
582 td->ts.io_u_lat_m[index]++;
585 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
588 io_u_mark_lat_usec(td, usec);
590 io_u_mark_lat_msec(td, usec / 1000);
594 * Get next file to service by choosing one at random
596 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
603 long r = os_random_long(&td->next_file_state);
605 fno = (unsigned int) ((double) td->o.nr_files
606 * (r / (RAND_MAX + 1.0)));
608 if (f->flags & FIO_FILE_DONE)
611 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
612 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
619 * Get next file to service by doing round robin between all available ones
621 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
624 unsigned int old_next_file = td->next_file;
628 f = td->files[td->next_file];
631 if (td->next_file >= td->o.nr_files)
634 if (f->flags & FIO_FILE_DONE) {
639 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
643 } while (td->next_file != old_next_file);
645 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
649 static struct fio_file *get_next_file(struct thread_data *td)
653 assert(td->o.nr_files <= td->files_index);
655 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
656 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
657 " nr_files=%d\n", td->nr_open_files,
663 f = td->file_service_file;
664 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
667 if (td->o.file_service_type == FIO_FSERVICE_RR)
668 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
670 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
672 td->file_service_file = f;
673 td->file_service_left = td->file_service_nr - 1;
675 dprint(FD_FILE, "get_next_file: %p\n", f);
679 static struct fio_file *find_next_new_file(struct thread_data *td)
683 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
686 if (td->o.file_service_type == FIO_FSERVICE_RR)
687 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
689 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
694 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
699 f = get_next_file(td);
707 if (!fill_io_u(td, io_u))
711 * optimization to prevent close/open of the same file. This
712 * way we preserve queueing etc.
714 if (td->o.nr_files == 1 && td->o.time_based) {
721 * td_io_close() does a put_file() as well, so no need to
725 td_io_close_file(td, f);
726 f->flags |= FIO_FILE_DONE;
730 * probably not the right place to do this, but see
731 * if we need to open a new file
733 if (td->nr_open_files < td->o.open_files &&
734 td->o.open_files != td->o.nr_files) {
735 f = find_next_new_file(td);
737 if (!f || td_io_open_file(td, f))
748 struct io_u *__get_io_u(struct thread_data *td)
750 struct io_u *io_u = NULL;
752 if (!list_empty(&td->io_u_requeues))
753 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
754 else if (!queue_full(td)) {
755 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
764 assert(io_u->flags & IO_U_F_FREE);
765 io_u->flags &= ~IO_U_F_FREE;
768 list_del(&io_u->list);
769 list_add(&io_u->list, &td->io_u_busylist);
777 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
778 * etc. The returned io_u is fully ready to be prepped and submitted.
780 struct io_u *get_io_u(struct thread_data *td)
785 io_u = __get_io_u(td);
787 dprint(FD_IO, "__get_io_u failed\n");
792 * from a requeue, io_u already setup
798 * If using an iolog, grab next piece if any available.
800 if (td->o.read_iolog_file) {
801 if (read_iolog_get(td, io_u))
803 } else if (set_io_u_file(td, io_u)) {
804 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
809 assert(f->flags & FIO_FILE_OPEN);
811 if (io_u->ddir != DDIR_SYNC) {
812 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
813 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
817 f->last_pos = io_u->offset + io_u->buflen;
819 if (td->o.verify != VERIFY_NONE)
820 populate_verify_io_u(td, io_u);
821 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
822 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
826 * Set io data pointers.
828 io_u->endpos = io_u->offset + io_u->buflen;
829 io_u->xfer_buf = io_u->buf;
830 io_u->xfer_buflen = io_u->buflen;
833 if (!td_io_prep(td, io_u)) {
834 fio_gettime(&io_u->start_time, NULL);
838 dprint(FD_IO, "get_io_u failed\n");
843 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
845 const char *msg[] = { "read", "write", "sync" };
847 log_err("fio: io_u error");
850 log_err(" on file %s", io_u->file->file_name);
852 log_err(": %s\n", strerror(io_u->error));
854 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
855 io_u->offset, io_u->xfer_buflen);
858 td_verror(td, io_u->error, "io_u error");
861 static void io_completed(struct thread_data *td, struct io_u *io_u,
862 struct io_completion_data *icd)
866 dprint_io_u(io_u, "io complete");
868 assert(io_u->flags & IO_U_F_FLIGHT);
869 io_u->flags &= ~IO_U_F_FLIGHT;
871 if (io_u->ddir == DDIR_SYNC) {
872 td->last_was_sync = 1;
876 td->last_was_sync = 0;
879 unsigned int bytes = io_u->buflen - io_u->resid;
880 const enum fio_ddir idx = io_u->ddir;
883 td->io_blocks[idx]++;
884 td->io_bytes[idx] += bytes;
885 td->this_io_bytes[idx] += bytes;
887 usec = utime_since(&io_u->issue_time, &icd->time);
889 add_clat_sample(td, idx, usec);
890 add_bw_sample(td, idx, &icd->time);
891 io_u_mark_latency(td, usec);
893 if (td_write(td) && idx == DDIR_WRITE &&
895 td->o.verify != VERIFY_NONE)
896 log_io_piece(td, io_u);
898 icd->bytes_done[idx] += bytes;
901 ret = io_u->end_io(td, io_u);
902 if (ret && !icd->error)
906 icd->error = io_u->error;
907 io_u_log_error(td, io_u);
911 static void init_icd(struct io_completion_data *icd, int nr)
913 fio_gettime(&icd->time, NULL);
918 icd->bytes_done[0] = icd->bytes_done[1] = 0;
921 static void ios_completed(struct thread_data *td,
922 struct io_completion_data *icd)
927 for (i = 0; i < icd->nr; i++) {
928 io_u = td->io_ops->event(td, i);
930 io_completed(td, io_u, icd);
936 * Complete a single io_u for the sync engines.
938 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
940 struct io_completion_data icd;
943 io_completed(td, io_u, &icd);
947 return icd.bytes_done[0] + icd.bytes_done[1];
949 td_verror(td, icd.error, "io_u_sync_complete");
954 * Called to complete min_events number of io for the async engines.
956 long io_u_queued_complete(struct thread_data *td, int min_events)
958 struct io_completion_data icd;
959 struct timespec *tvp = NULL;
961 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
963 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
968 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
970 td_verror(td, -ret, "td_io_getevents");
976 ios_completed(td, &icd);
978 return icd.bytes_done[0] + icd.bytes_done[1];
980 td_verror(td, icd.error, "io_u_queued_complete");
985 * Call when io_u is really queued, to update the submission latency.
987 void io_u_queued(struct thread_data *td, struct io_u *io_u)
989 unsigned long slat_time;
991 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
992 add_slat_sample(td, io_u->ddir, slat_time);
996 * "randomly" fill the buffer contents
998 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1001 long *ptr = io_u->buf;
1003 if (!td->o.zero_buffers) {
1004 while ((void *) ptr - io_u->buf < max_bs) {
1005 *ptr = rand() * GOLDEN_RATIO_PRIME;
1009 memset(ptr, 0, max_bs);
1012 #ifdef FIO_USE_TIMEOUT
1013 void io_u_set_timeout(struct thread_data *td)
1015 assert(td->cur_depth);
1017 td->timer.it_interval.tv_sec = 0;
1018 td->timer.it_interval.tv_usec = 0;
1019 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1020 td->timer.it_value.tv_usec = 0;
1021 setitimer(ITIMER_REAL, &td->timer, NULL);
1022 fio_gettime(&td->timeout_end, NULL);
1025 static void io_u_dump(struct io_u *io_u)
1027 unsigned long t_start = mtime_since_now(&io_u->start_time);
1028 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1030 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
1031 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1032 io_u->xfer_buf, io_u->buflen,
1035 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1038 void io_u_set_timeout(struct thread_data fio_unused *td)
1043 #ifdef FIO_USE_TIMEOUT
1044 static void io_u_timeout_handler(int fio_unused sig)
1046 struct thread_data *td, *__td;
1047 pid_t pid = getpid();
1048 struct list_head *entry;
1052 log_err("fio: io_u timeout\n");
1055 * TLS would be nice...
1058 for_each_td(__td, i) {
1059 if (__td->pid == pid) {
1066 log_err("fio: io_u timeout, can't find job\n");
1070 if (!td->cur_depth) {
1071 log_err("fio: timeout without pending work?\n");
1075 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1077 list_for_each(entry, &td->io_u_busylist) {
1078 io_u = list_entry(entry, struct io_u, list);
1083 td_verror(td, ETIMEDOUT, "io_u timeout");
1088 void io_u_init_timeout(void)
1090 #ifdef FIO_USE_TIMEOUT
1091 signal(SIGALRM, io_u_timeout_handler);