13 #include "lib/axmap.h"
18 struct io_completion_data {
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timespec time; /* output */
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
30 static bool random_map_free(struct fio_file *f, const uint64_t block)
32 return !axmap_isset(f->io_axmap, block);
36 * Mark a given offset as used in the map.
38 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40 unsigned int min_bs = td->o.min_bs[io_u->ddir];
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
55 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
61 assert(ddir_rw(ddir));
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 * -> not for now since there is code assuming it could go either.
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
72 max_size = td->o.zone_range;
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
85 struct flist_head list;
89 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
90 enum fio_ddir ddir, uint64_t *b,
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
160 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
162 enum fio_ddir ddir, uint64_t *b)
164 struct zone_split_index *zsi;
165 uint64_t lastb, send, stotal;
169 lastb = last_block(td, f, ddir);
173 if (!td->o.zone_split_nr[ddir]) {
175 return __get_next_rand_offset(td, f, ddir, b, lastb);
179 * Generate a value, v, between 1 and 100, both inclusive
181 v = rand32_between(&td->zone_state, 1, 100);
184 * Find our generated table. 'send' is the end block of this zone,
185 * 'stotal' is our start offset.
187 zsi = &td->zone_state_index[ddir][v - 1];
188 stotal = zsi->size_prev / td->o.ba[ddir];
189 send = zsi->size / td->o.ba[ddir];
192 * Should never happen
196 log_err("fio: bug in zoned generation\n");
200 } else if (send > lastb) {
202 * This happens if the user specifies ranges that exceed
203 * the file/device size. We can't handle that gracefully,
206 log_err("fio: zoned_abs sizes exceed file size\n");
211 * Generate index from 0..send-stotal
213 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
220 static int __get_next_rand_offset_zoned(struct thread_data *td,
221 struct fio_file *f, enum fio_ddir ddir,
224 unsigned int v, send, stotal;
225 uint64_t offset, lastb;
227 struct zone_split_index *zsi;
229 lastb = last_block(td, f, ddir);
233 if (!td->o.zone_split_nr[ddir]) {
235 return __get_next_rand_offset(td, f, ddir, b, lastb);
239 * Generate a value, v, between 1 and 100, both inclusive
241 v = rand32_between(&td->zone_state, 1, 100);
243 zsi = &td->zone_state_index[ddir][v - 1];
244 stotal = zsi->size_perc_prev;
245 send = zsi->size_perc;
248 * Should never happen
252 log_err("fio: bug in zoned generation\n");
259 * 'send' is some percentage below or equal to 100 that
260 * marks the end of the current IO range. 'stotal' marks
261 * the start, in percent.
264 offset = stotal * lastb / 100ULL;
268 lastb = lastb * (send - stotal) / 100ULL;
271 * Generate index from 0..send-of-lastb
273 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
277 * Add our start offset, if any
285 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
287 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
288 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
290 return r1->off - r2->off;
293 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
294 enum fio_ddir ddir, uint64_t *b)
296 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
299 lastb = last_block(td, f, ddir);
303 return __get_next_rand_offset(td, f, ddir, b, lastb);
304 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
305 return __get_next_rand_offset_zipf(td, f, ddir, b);
306 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
307 return __get_next_rand_offset_pareto(td, f, ddir, b);
308 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
309 return __get_next_rand_offset_gauss(td, f, ddir, b);
310 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
311 return __get_next_rand_offset_zoned(td, f, ddir, b);
312 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
313 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
315 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
320 * Sort the reads for a verify phase in batches of verifysort_nr, if
323 static inline bool should_sort_io(struct thread_data *td)
325 if (!td->o.verifysort_nr || !td->o.do_verify)
329 if (td->runstate != TD_VERIFYING)
331 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
332 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
338 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
342 if (td->o.perc_rand[ddir] == 100)
345 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
347 return v <= td->o.perc_rand[ddir];
350 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
351 enum fio_ddir ddir, uint64_t *b)
356 if (!should_sort_io(td))
357 return get_off_from_method(td, f, ddir, b);
359 if (!flist_empty(&td->next_rand_list)) {
361 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
368 for (i = 0; i < td->o.verifysort_nr; i++) {
369 r = malloc(sizeof(*r));
371 ret = get_off_from_method(td, f, ddir, &r->off);
377 flist_add(&r->list, &td->next_rand_list);
383 assert(!flist_empty(&td->next_rand_list));
384 flist_sort(NULL, &td->next_rand_list, flist_cmp);
388 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
390 struct thread_options *o = &td->o;
392 if (o->invalidate_cache && !o->odirect) {
395 ret = file_invalidate_cache(td, f);
399 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
400 enum fio_ddir ddir, uint64_t *b)
402 if (!get_next_rand_offset(td, f, ddir, b))
405 if (td->o.time_based ||
406 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
407 fio_file_reset(td, f);
408 if (!get_next_rand_offset(td, f, ddir, b))
410 loop_cache_invalidate(td, f);
413 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
414 f->file_name, (unsigned long long) f->last_pos[ddir],
415 (unsigned long long) f->real_file_size);
419 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
420 enum fio_ddir ddir, uint64_t *offset)
422 struct thread_options *o = &td->o;
424 assert(ddir_rw(ddir));
427 * If we reach the end for a time based run, reset us back to 0
428 * and invalidate the cache, if we need to.
430 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
432 f->last_pos[ddir] = f->file_offset;
433 loop_cache_invalidate(td, f);
436 if (f->last_pos[ddir] < f->real_file_size) {
439 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
440 if (f->real_file_size > f->io_size)
441 f->last_pos[ddir] = f->io_size;
443 f->last_pos[ddir] = f->real_file_size;
446 pos = f->last_pos[ddir] - f->file_offset;
447 if (pos && o->ddir_seq_add) {
448 pos += o->ddir_seq_add;
451 * If we reach beyond the end of the file
452 * with holed IO, wrap around to the
453 * beginning again. If we're doing backwards IO,
456 if (pos >= f->real_file_size) {
457 if (o->ddir_seq_add > 0)
458 pos = f->file_offset;
460 if (f->real_file_size > f->io_size)
463 pos = f->real_file_size;
465 pos += o->ddir_seq_add;
477 static int get_next_block(struct thread_data *td, struct io_u *io_u,
478 enum fio_ddir ddir, int rw_seq,
479 unsigned int *is_random)
481 struct fio_file *f = io_u->file;
485 assert(ddir_rw(ddir));
491 if (should_do_random(td, ddir)) {
492 ret = get_next_rand_block(td, f, ddir, &b);
496 io_u_set(td, io_u, IO_U_F_BUSY_OK);
497 ret = get_next_seq_offset(td, f, ddir, &offset);
499 ret = get_next_rand_block(td, f, ddir, &b);
503 ret = get_next_seq_offset(td, f, ddir, &offset);
506 io_u_set(td, io_u, IO_U_F_BUSY_OK);
509 if (td->o.rw_seq == RW_SEQ_SEQ) {
510 ret = get_next_seq_offset(td, f, ddir, &offset);
512 ret = get_next_rand_block(td, f, ddir, &b);
515 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
516 if (f->last_start[ddir] != -1ULL)
517 offset = f->last_start[ddir] - f->file_offset;
522 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
529 io_u->offset = offset;
531 io_u->offset = b * td->o.ba[ddir];
533 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
542 * For random io, generate a random new block and see if it's used. Repeat
543 * until we find a free one. For sequential io, just return the end of
544 * the last io issued.
546 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
547 unsigned int *is_random)
549 struct fio_file *f = io_u->file;
550 enum fio_ddir ddir = io_u->ddir;
553 assert(ddir_rw(ddir));
555 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
557 td->ddir_seq_nr = td->o.ddir_seq_nr;
560 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
563 if (io_u->offset >= f->io_size) {
564 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
565 (unsigned long long) io_u->offset,
566 (unsigned long long) f->io_size);
570 io_u->offset += f->file_offset;
571 if (io_u->offset >= f->real_file_size) {
572 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
573 (unsigned long long) io_u->offset,
574 (unsigned long long) f->real_file_size);
581 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
582 unsigned int *is_random)
584 if (td->flags & TD_F_PROFILE_OPS) {
585 struct prof_io_ops *ops = &td->prof_io_ops;
587 if (ops->fill_io_u_off)
588 return ops->fill_io_u_off(td, io_u, is_random);
591 return __get_next_offset(td, io_u, is_random);
594 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
597 struct fio_file *f = io_u->file;
599 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
602 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
603 unsigned int is_random)
605 int ddir = io_u->ddir;
606 unsigned int buflen = 0;
607 unsigned int minbs, maxbs;
608 uint64_t frand_max, r;
611 assert(ddir_rw(ddir));
613 if (td->o.bs_is_seq_rand)
614 ddir = is_random ? DDIR_WRITE: DDIR_READ;
616 minbs = td->o.min_bs[ddir];
617 maxbs = td->o.max_bs[ddir];
623 * If we can't satisfy the min block size from here, then fail
625 if (!io_u_fits(td, io_u, minbs))
628 frand_max = rand_max(&td->bsrange_state[ddir]);
630 r = __rand(&td->bsrange_state[ddir]);
632 if (!td->o.bssplit_nr[ddir]) {
633 buflen = 1 + (unsigned int) ((double) maxbs *
634 (r / (frand_max + 1.0)));
641 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
642 struct bssplit *bsp = &td->o.bssplit[ddir][i];
648 if ((r / perc <= frand_max / 100ULL) &&
649 io_u_fits(td, io_u, buflen))
654 power_2 = is_power_of_2(minbs);
655 if (!td->o.bs_unaligned && power_2)
656 buflen &= ~(minbs - 1);
657 else if (!td->o.bs_unaligned && !power_2)
658 buflen -= buflen % minbs;
659 } while (!io_u_fits(td, io_u, buflen));
664 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
665 unsigned int is_random)
667 if (td->flags & TD_F_PROFILE_OPS) {
668 struct prof_io_ops *ops = &td->prof_io_ops;
670 if (ops->fill_io_u_size)
671 return ops->fill_io_u_size(td, io_u, is_random);
674 return __get_next_buflen(td, io_u, is_random);
677 static void set_rwmix_bytes(struct thread_data *td)
682 * we do time or byte based switch. this is needed because
683 * buffered writes may issue a lot quicker than they complete,
684 * whereas reads do not.
686 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
687 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
690 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
694 v = rand32_between(&td->rwmix_state, 1, 100);
696 if (v <= td->o.rwmix[DDIR_READ])
702 int io_u_quiesce(struct thread_data *td)
707 * We are going to sleep, ensure that we flush anything pending as
708 * not to skew our latency numbers.
710 * Changed to only monitor 'in flight' requests here instead of the
711 * td->cur_depth, b/c td->cur_depth does not accurately represent
712 * io's that have been actually submitted to an async engine,
713 * and cur_depth is meaningless for sync engines.
715 if (td->io_u_queued || td->cur_depth) {
718 ret = td_io_commit(td);
721 while (td->io_u_in_flight) {
724 ret = io_u_queued_complete(td, 1);
729 if (td->flags & TD_F_REGROW_LOGS)
735 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
737 enum fio_ddir odir = ddir ^ 1;
741 assert(ddir_rw(ddir));
742 now = utime_since_now(&td->start);
745 * if rate_next_io_time is in the past, need to catch up to rate
747 if (td->rate_next_io_time[ddir] <= now)
751 * We are ahead of rate in this direction. See if we
754 if (td_rw(td) && td->o.rwmix[odir]) {
756 * Other direction is behind rate, switch
758 if (td->rate_next_io_time[odir] <= now)
762 * Both directions are ahead of rate. sleep the min,
763 * switch if necessary
765 if (td->rate_next_io_time[ddir] <=
766 td->rate_next_io_time[odir]) {
767 usec = td->rate_next_io_time[ddir] - now;
769 usec = td->rate_next_io_time[odir] - now;
773 usec = td->rate_next_io_time[ddir] - now;
775 if (td->o.io_submit_mode == IO_MODE_INLINE)
778 usec_sleep(td, usec);
783 * Return the data direction for the next io_u. If the job is a
784 * mixed read/write workload, check the rwmix cycle and switch if
787 static enum fio_ddir get_rw_ddir(struct thread_data *td)
792 * See if it's time to fsync/fdatasync/sync_file_range first,
793 * and if not then move on to check regular I/Os.
795 if (should_fsync(td)) {
796 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
797 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
800 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
801 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
802 return DDIR_DATASYNC;
804 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
805 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
806 return DDIR_SYNC_FILE_RANGE;
811 * Check if it's time to seed a new data direction.
813 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
815 * Put a top limit on how many bytes we do for
816 * one data direction, to avoid overflowing the
819 ddir = get_rand_ddir(td);
821 if (ddir != td->rwmix_ddir)
824 td->rwmix_ddir = ddir;
826 ddir = td->rwmix_ddir;
827 } else if (td_read(td))
829 else if (td_write(td))
831 else if (td_trim(td))
836 td->rwmix_ddir = rate_ddir(td, ddir);
837 return td->rwmix_ddir;
840 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
842 enum fio_ddir ddir = get_rw_ddir(td);
844 if (td_trimwrite(td)) {
845 struct fio_file *f = io_u->file;
846 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
852 io_u->ddir = io_u->acct_ddir = ddir;
854 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
855 td->o.barrier_blocks &&
856 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
857 td->io_issues[DDIR_WRITE])
858 io_u_set(td, io_u, IO_U_F_BARRIER);
861 void put_file_log(struct thread_data *td, struct fio_file *f)
863 unsigned int ret = put_file(td, f);
866 td_verror(td, ret, "file close");
869 void put_io_u(struct thread_data *td, struct io_u *io_u)
876 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
877 put_file_log(td, io_u->file);
880 io_u_set(td, io_u, IO_U_F_FREE);
882 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
884 assert(!(td->flags & TD_F_CHILD));
886 io_u_qpush(&td->io_u_freelist, io_u);
888 td_io_u_free_notify(td);
891 void clear_io_u(struct thread_data *td, struct io_u *io_u)
893 io_u_clear(td, io_u, IO_U_F_FLIGHT);
897 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
899 struct io_u *__io_u = *io_u;
900 enum fio_ddir ddir = acct_ddir(__io_u);
902 dprint(FD_IO, "requeue %p\n", __io_u);
909 io_u_set(td, __io_u, IO_U_F_FREE);
910 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
911 td->io_issues[ddir]--;
913 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
914 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
916 assert(!(td->flags & TD_F_CHILD));
919 io_u_rpush(&td->io_u_requeues, __io_u);
921 td_io_u_free_notify(td);
925 static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
927 struct fio_file *f = io_u->file;
930 * See if it's time to switch to a new zone
932 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
934 f->file_offset += td->o.zone_range + td->o.zone_skip;
937 * Wrap from the beginning, if we exceed the file size
939 if (f->file_offset >= f->real_file_size)
940 f->file_offset = f->real_file_size - f->file_offset;
941 f->last_pos[io_u->ddir] = f->file_offset;
942 td->io_skip_bytes += td->o.zone_skip;
946 * If zone_size > zone_range, then maintain the same zone until
947 * zone_bytes >= zone_size.
949 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
950 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
951 f->file_offset, f->last_pos[io_u->ddir]);
952 f->last_pos[io_u->ddir] = f->file_offset;
956 * For random: if 'norandommap' is not set and zone_size > zone_range,
957 * map needs to be reset as it's done with zone_range everytime.
959 if ((td->zone_bytes % td->o.zone_range) == 0) {
960 fio_file_reset(td, f);
964 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
966 unsigned int is_random;
968 if (td_ioengine_flagged(td, FIO_NOIO))
971 set_rw_ddir(td, io_u);
974 * fsync() or fdatasync() or trim etc, we are done
976 if (!ddir_rw(io_u->ddir))
980 * When file is zoned zone_range is always positive
982 if (td->o.zone_range) {
983 __fill_io_u_zone(td, io_u);
987 * No log, let the seq/rand engine retrieve the next buflen and
990 if (get_next_offset(td, io_u, &is_random)) {
991 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
995 io_u->buflen = get_next_buflen(td, io_u, is_random);
997 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
1001 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
1002 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
1004 (unsigned long long) io_u->offset, io_u->buflen,
1005 (unsigned long long) io_u->file->real_file_size);
1010 * mark entry before potentially trimming io_u
1012 if (td_random(td) && file_randommap(td, io_u->file))
1013 mark_random_map(td, io_u);
1016 dprint_io_u(io_u, "fill");
1017 td->zone_bytes += io_u->buflen;
1021 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
1050 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1052 __io_u_mark_map(td->ts.io_u_submit, nr);
1053 td->ts.total_submit++;
1056 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1058 __io_u_mark_map(td->ts.io_u_complete, nr);
1059 td->ts.total_complete++;
1062 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1066 switch (td->cur_depth) {
1088 td->ts.io_u_map[idx] += nr;
1091 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1095 assert(nsec < 1000);
1128 assert(idx < FIO_IO_U_LAT_N_NR);
1129 td->ts.io_u_lat_n[idx]++;
1132 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1136 assert(usec < 1000 && usec >= 1);
1169 assert(idx < FIO_IO_U_LAT_U_NR);
1170 td->ts.io_u_lat_u[idx]++;
1173 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1216 assert(idx < FIO_IO_U_LAT_M_NR);
1217 td->ts.io_u_lat_m[idx]++;
1220 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1223 io_u_mark_lat_nsec(td, nsec);
1224 else if (nsec < 1000000)
1225 io_u_mark_lat_usec(td, nsec / 1000);
1227 io_u_mark_lat_msec(td, nsec / 1000000);
1230 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1232 unsigned long fileno;
1234 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1235 uint64_t frand_max = rand_max(&td->next_file_state);
1238 r = __rand(&td->next_file_state);
1239 return (unsigned int) ((double) td->o.nr_files
1240 * (r / (frand_max + 1.0)));
1243 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1244 fileno = zipf_next(&td->next_file_zipf);
1245 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1246 fileno = pareto_next(&td->next_file_zipf);
1247 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1248 fileno = gauss_next(&td->next_file_gauss);
1250 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1255 return fileno >> FIO_FSERVICE_SHIFT;
1259 * Get next file to service by choosing one at random
1261 static struct fio_file *get_next_file_rand(struct thread_data *td,
1262 enum fio_file_flags goodf,
1263 enum fio_file_flags badf)
1271 fno = __get_next_fileno_rand(td);
1274 if (fio_file_done(f))
1277 if (!fio_file_open(f)) {
1280 if (td->nr_open_files >= td->o.open_files)
1281 return ERR_PTR(-EBUSY);
1283 err = td_io_open_file(td, f);
1289 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1290 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1294 td_io_close_file(td, f);
1299 * Get next file to service by doing round robin between all available ones
1301 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1304 unsigned int old_next_file = td->next_file;
1310 f = td->files[td->next_file];
1313 if (td->next_file >= td->o.nr_files)
1316 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1317 if (fio_file_done(f)) {
1322 if (!fio_file_open(f)) {
1325 if (td->nr_open_files >= td->o.open_files)
1326 return ERR_PTR(-EBUSY);
1328 err = td_io_open_file(td, f);
1330 dprint(FD_FILE, "error %d on open of %s\n",
1338 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1340 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1344 td_io_close_file(td, f);
1347 } while (td->next_file != old_next_file);
1349 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1353 static struct fio_file *__get_next_file(struct thread_data *td)
1357 assert(td->o.nr_files <= td->files_index);
1359 if (td->nr_done_files >= td->o.nr_files) {
1360 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1361 " nr_files=%d\n", td->nr_open_files,
1367 f = td->file_service_file;
1368 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1369 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1371 if (td->file_service_left--)
1375 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1376 td->o.file_service_type == FIO_FSERVICE_SEQ)
1377 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1379 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1384 td->file_service_file = f;
1385 td->file_service_left = td->file_service_nr - 1;
1388 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1390 dprint(FD_FILE, "get_next_file: NULL\n");
1394 static struct fio_file *get_next_file(struct thread_data *td)
1396 if (td->flags & TD_F_PROFILE_OPS) {
1397 struct prof_io_ops *ops = &td->prof_io_ops;
1399 if (ops->get_next_file)
1400 return ops->get_next_file(td);
1403 return __get_next_file(td);
1406 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1411 f = get_next_file(td);
1412 if (IS_ERR_OR_NULL(f))
1418 if (!fill_io_u(td, io_u))
1421 put_file_log(td, f);
1422 td_io_close_file(td, f);
1424 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1425 fio_file_reset(td, f);
1427 fio_file_set_done(f);
1428 td->nr_done_files++;
1429 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1430 td->nr_done_files, td->o.nr_files);
1437 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1438 unsigned long long tnsec, unsigned long long max_nsec)
1441 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1442 td_verror(td, ETIMEDOUT, "max latency exceeded");
1443 icd->error = ETIMEDOUT;
1446 static void lat_new_cycle(struct thread_data *td)
1448 fio_gettime(&td->latency_ts, NULL);
1449 td->latency_ios = ddir_rw_sum(td->io_blocks);
1450 td->latency_failed = 0;
1454 * We had an IO outside the latency target. Reduce the queue depth. If we
1455 * are at QD=1, then it's time to give up.
1457 static bool __lat_target_failed(struct thread_data *td)
1459 if (td->latency_qd == 1)
1462 td->latency_qd_high = td->latency_qd;
1464 if (td->latency_qd == td->latency_qd_low)
1465 td->latency_qd_low--;
1467 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1469 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1472 * When we ramp QD down, quiesce existing IO to prevent
1473 * a storm of ramp downs due to pending higher depth.
1480 static bool lat_target_failed(struct thread_data *td)
1482 if (td->o.latency_percentile.u.f == 100.0)
1483 return __lat_target_failed(td);
1485 td->latency_failed++;
1489 void lat_target_init(struct thread_data *td)
1491 td->latency_end_run = 0;
1493 if (td->o.latency_target) {
1494 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1495 fio_gettime(&td->latency_ts, NULL);
1497 td->latency_qd_high = td->o.iodepth;
1498 td->latency_qd_low = 1;
1499 td->latency_ios = ddir_rw_sum(td->io_blocks);
1501 td->latency_qd = td->o.iodepth;
1504 void lat_target_reset(struct thread_data *td)
1506 if (!td->latency_end_run)
1507 lat_target_init(td);
1510 static void lat_target_success(struct thread_data *td)
1512 const unsigned int qd = td->latency_qd;
1513 struct thread_options *o = &td->o;
1515 td->latency_qd_low = td->latency_qd;
1518 * If we haven't failed yet, we double up to a failing value instead
1519 * of bisecting from highest possible queue depth. If we have set
1520 * a limit other than td->o.iodepth, bisect between that.
1522 if (td->latency_qd_high != o->iodepth)
1523 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1525 td->latency_qd *= 2;
1527 if (td->latency_qd > o->iodepth)
1528 td->latency_qd = o->iodepth;
1530 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1533 * Same as last one, we are done. Let it run a latency cycle, so
1534 * we get only the results from the targeted depth.
1536 if (td->latency_qd == qd) {
1537 if (td->latency_end_run) {
1538 dprint(FD_RATE, "We are done\n");
1541 dprint(FD_RATE, "Quiesce and final run\n");
1543 td->latency_end_run = 1;
1544 reset_all_stats(td);
1553 * Check if we can bump the queue depth
1555 void lat_target_check(struct thread_data *td)
1557 uint64_t usec_window;
1561 usec_window = utime_since_now(&td->latency_ts);
1562 if (usec_window < td->o.latency_window)
1565 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1566 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1567 success_ios *= 100.0;
1569 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1571 if (success_ios >= td->o.latency_percentile.u.f)
1572 lat_target_success(td);
1574 __lat_target_failed(td);
1578 * If latency target is enabled, we might be ramping up or down and not
1579 * using the full queue depth available.
1581 bool queue_full(const struct thread_data *td)
1583 const int qempty = io_u_qempty(&td->io_u_freelist);
1587 if (!td->o.latency_target)
1590 return td->cur_depth >= td->latency_qd;
1593 struct io_u *__get_io_u(struct thread_data *td)
1595 struct io_u *io_u = NULL;
1603 if (!io_u_rempty(&td->io_u_requeues))
1604 io_u = io_u_rpop(&td->io_u_requeues);
1605 else if (!queue_full(td)) {
1606 io_u = io_u_qpop(&td->io_u_freelist);
1611 io_u->end_io = NULL;
1615 assert(io_u->flags & IO_U_F_FREE);
1616 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1617 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1621 io_u->acct_ddir = -1;
1623 assert(!(td->flags & TD_F_CHILD));
1624 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1626 } else if (td_async_processing(td)) {
1628 * We ran out, wait for async verify threads to finish and
1631 assert(!(td->flags & TD_F_CHILD));
1632 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1640 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1642 if (!(td->flags & TD_F_TRIM_BACKLOG))
1644 if (!td->trim_entries)
1647 if (td->trim_batch) {
1649 if (get_next_trim(td, io_u))
1651 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1652 td->last_ddir != DDIR_READ) {
1653 td->trim_batch = td->o.trim_batch;
1654 if (!td->trim_batch)
1655 td->trim_batch = td->o.trim_backlog;
1656 if (get_next_trim(td, io_u))
1663 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1665 if (!(td->flags & TD_F_VER_BACKLOG))
1668 if (td->io_hist_len) {
1671 if (td->verify_batch)
1673 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1674 td->last_ddir != DDIR_READ) {
1675 td->verify_batch = td->o.verify_batch;
1676 if (!td->verify_batch)
1677 td->verify_batch = td->o.verify_backlog;
1681 if (get_verify && !get_next_verify(td, io_u)) {
1691 * Fill offset and start time into the buffer content, to prevent too
1692 * easy compressible data for simple de-dupe attempts. Do this for every
1693 * 512b block in the range, since that should be the smallest block size
1694 * we can expect from a device.
1696 static void small_content_scramble(struct io_u *io_u)
1698 unsigned int i, nr_blocks = io_u->buflen >> 9;
1699 unsigned int offset;
1700 uint64_t boffset, *iptr;
1707 boffset = io_u->offset;
1709 if (io_u->buf_filled_len)
1710 io_u->buf_filled_len = 0;
1713 * Generate random index between 0..7. We do chunks of 512b, if
1714 * we assume a cacheline is 64 bytes, then we have 8 of those.
1715 * Scramble content within the blocks in the same cacheline to
1718 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1720 for (i = 0; i < nr_blocks; i++) {
1722 * Fill offset into start of cacheline, time into end
1725 iptr = (void *) p + (offset << 6);
1728 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1729 iptr[0] = io_u->start_time.tv_sec;
1730 iptr[1] = io_u->start_time.tv_nsec;
1738 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1739 * etc. The returned io_u is fully ready to be prepped and submitted.
1741 struct io_u *get_io_u(struct thread_data *td)
1745 int do_scramble = 0;
1748 io_u = __get_io_u(td);
1750 dprint(FD_IO, "__get_io_u failed\n");
1754 if (check_get_verify(td, io_u))
1756 if (check_get_trim(td, io_u))
1760 * from a requeue, io_u already setup
1766 * If using an iolog, grab next piece if any available.
1768 if (td->flags & TD_F_READ_IOLOG) {
1769 if (read_iolog_get(td, io_u))
1771 } else if (set_io_u_file(td, io_u)) {
1773 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1779 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1783 assert(fio_file_open(f));
1785 if (ddir_rw(io_u->ddir)) {
1786 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1787 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1791 f->last_start[io_u->ddir] = io_u->offset;
1792 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1794 if (io_u->ddir == DDIR_WRITE) {
1795 if (td->flags & TD_F_REFILL_BUFFERS) {
1796 io_u_fill_buffer(td, io_u,
1797 td->o.min_bs[DDIR_WRITE],
1799 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1800 !(td->flags & TD_F_COMPRESS))
1802 if (td->flags & TD_F_VER_NONE) {
1803 populate_verify_io_u(td, io_u);
1806 } else if (io_u->ddir == DDIR_READ) {
1808 * Reset the buf_filled parameters so next time if the
1809 * buffer is used for writes it is refilled.
1811 io_u->buf_filled_len = 0;
1816 * Set io data pointers.
1818 io_u->xfer_buf = io_u->buf;
1819 io_u->xfer_buflen = io_u->buflen;
1823 if (!td_io_prep(td, io_u)) {
1824 if (!td->o.disable_lat)
1825 fio_gettime(&io_u->start_time, NULL);
1828 small_content_scramble(io_u);
1833 dprint(FD_IO, "get_io_u failed\n");
1835 return ERR_PTR(ret);
1838 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1840 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1842 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1845 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1846 io_u->file ? " on file " : "",
1847 io_u->file ? io_u->file->file_name : "",
1848 strerror(io_u->error),
1849 io_ddir_name(io_u->ddir),
1850 io_u->offset, io_u->xfer_buflen);
1852 if (td->io_ops->errdetails) {
1853 char *err = td->io_ops->errdetails(io_u);
1855 log_err("fio: %s\n", err);
1860 td_verror(td, io_u->error, "io_u error");
1863 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1865 __io_u_log_error(td, io_u);
1867 __io_u_log_error(td->parent, io_u);
1870 static inline bool gtod_reduce(struct thread_data *td)
1872 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1873 || td->o.gtod_reduce;
1876 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1877 struct io_completion_data *icd,
1878 const enum fio_ddir idx, unsigned int bytes)
1880 const int no_reduce = !gtod_reduce(td);
1881 unsigned long long llnsec = 0;
1886 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1890 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1892 if (!td->o.disable_lat) {
1893 unsigned long long tnsec;
1895 tnsec = ntime_since(&io_u->start_time, &icd->time);
1896 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1898 if (td->flags & TD_F_PROFILE_OPS) {
1899 struct prof_io_ops *ops = &td->prof_io_ops;
1902 icd->error = ops->io_u_lat(td, tnsec);
1905 if (td->o.max_latency && tnsec > td->o.max_latency)
1906 lat_fatal(td, icd, tnsec, td->o.max_latency);
1907 if (td->o.latency_target && tnsec > td->o.latency_target) {
1908 if (lat_target_failed(td))
1909 lat_fatal(td, icd, tnsec, td->o.latency_target);
1914 if (!td->o.disable_clat) {
1915 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1916 io_u_mark_latency(td, llnsec);
1919 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1920 add_bw_sample(td, io_u, bytes, llnsec);
1922 if (no_reduce && per_unit_log(td->iops_log))
1923 add_iops_sample(td, io_u, bytes);
1924 } else if (ddir_sync(idx) && !td->o.disable_clat)
1925 add_sync_clat_sample(&td->ts, llnsec);
1927 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1928 uint32_t *info = io_u_block_info(td, io_u);
1929 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1930 if (io_u->ddir == DDIR_TRIM) {
1931 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1932 BLOCK_INFO_TRIMS(*info) + 1);
1933 } else if (io_u->ddir == DDIR_WRITE) {
1934 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1941 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1942 uint64_t offset, unsigned int bytes)
1949 if (f->first_write == -1ULL || offset < f->first_write)
1950 f->first_write = offset;
1951 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1952 f->last_write = offset + bytes;
1954 if (!f->last_write_comp)
1957 idx = f->last_write_idx++;
1958 f->last_write_comp[idx] = offset;
1959 if (f->last_write_idx == td->o.iodepth)
1960 f->last_write_idx = 0;
1963 static bool should_account(struct thread_data *td)
1965 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1966 td->runstate == TD_VERIFYING);
1969 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1970 struct io_completion_data *icd)
1972 struct io_u *io_u = *io_u_ptr;
1973 enum fio_ddir ddir = io_u->ddir;
1974 struct fio_file *f = io_u->file;
1976 dprint_io_u(io_u, "complete");
1978 assert(io_u->flags & IO_U_F_FLIGHT);
1979 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1982 * Mark IO ok to verify
1986 * Remove errored entry from the verification list
1989 unlog_io_piece(td, io_u);
1991 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1996 if (ddir_sync(ddir)) {
1997 td->last_was_sync = true;
1999 f->first_write = -1ULL;
2000 f->last_write = -1ULL;
2002 if (should_account(td))
2003 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
2007 td->last_was_sync = false;
2008 td->last_ddir = ddir;
2010 if (!io_u->error && ddir_rw(ddir)) {
2011 unsigned int bytes = io_u->buflen - io_u->resid;
2014 td->io_blocks[ddir]++;
2015 td->io_bytes[ddir] += bytes;
2017 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2018 td->this_io_blocks[ddir]++;
2019 td->this_io_bytes[ddir] += bytes;
2022 if (ddir == DDIR_WRITE)
2023 file_log_write_comp(td, f, io_u->offset, bytes);
2025 if (should_account(td))
2026 account_io_completion(td, io_u, icd, ddir, bytes);
2028 icd->bytes_done[ddir] += bytes;
2031 ret = io_u->end_io(td, io_u_ptr);
2033 if (ret && !icd->error)
2036 } else if (io_u->error) {
2037 icd->error = io_u->error;
2038 io_u_log_error(td, io_u);
2041 enum error_type_bit eb = td_error_type(ddir, icd->error);
2043 if (!td_non_fatal_error(td, eb, icd->error))
2047 * If there is a non_fatal error, then add to the error count
2048 * and clear all the errors.
2050 update_error_count(td, icd->error);
2058 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2063 if (!gtod_reduce(td))
2064 fio_gettime(&icd->time, NULL);
2069 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2070 icd->bytes_done[ddir] = 0;
2073 static void ios_completed(struct thread_data *td,
2074 struct io_completion_data *icd)
2079 for (i = 0; i < icd->nr; i++) {
2080 io_u = td->io_ops->event(td, i);
2082 io_completed(td, &io_u, icd);
2090 * Complete a single io_u for the sync engines.
2092 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2094 struct io_completion_data icd;
2097 init_icd(td, &icd, 1);
2098 io_completed(td, &io_u, &icd);
2104 td_verror(td, icd.error, "io_u_sync_complete");
2108 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2109 td->bytes_done[ddir] += icd.bytes_done[ddir];
2115 * Called to complete min_events number of io for the async engines.
2117 int io_u_queued_complete(struct thread_data *td, int min_evts)
2119 struct io_completion_data icd;
2120 struct timespec *tvp = NULL;
2122 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2124 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2128 else if (min_evts > td->cur_depth)
2129 min_evts = td->cur_depth;
2131 /* No worries, td_io_getevents fixes min and max if they are
2132 * set incorrectly */
2133 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2135 td_verror(td, -ret, "td_io_getevents");
2140 init_icd(td, &icd, ret);
2141 ios_completed(td, &icd);
2143 td_verror(td, icd.error, "io_u_queued_complete");
2147 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2148 td->bytes_done[ddir] += icd.bytes_done[ddir];
2154 * Call when io_u is really queued, to update the submission latency.
2156 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2158 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2159 unsigned long slat_time;
2161 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2166 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2172 * See if we should reuse the last seed, if dedupe is enabled
2174 static struct frand_state *get_buf_state(struct thread_data *td)
2178 if (!td->o.dedupe_percentage)
2179 return &td->buf_state;
2180 else if (td->o.dedupe_percentage == 100) {
2181 frand_copy(&td->buf_state_prev, &td->buf_state);
2182 return &td->buf_state;
2185 v = rand32_between(&td->dedupe_state, 1, 100);
2187 if (v <= td->o.dedupe_percentage)
2188 return &td->buf_state_prev;
2190 return &td->buf_state;
2193 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2195 if (td->o.dedupe_percentage == 100)
2196 frand_copy(rs, &td->buf_state_prev);
2197 else if (rs == &td->buf_state)
2198 frand_copy(&td->buf_state_prev, rs);
2201 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2202 unsigned int max_bs)
2204 struct thread_options *o = &td->o;
2206 if (o->mem_type == MEM_CUDA_MALLOC)
2209 if (o->compress_percentage || o->dedupe_percentage) {
2210 unsigned int perc = td->o.compress_percentage;
2211 struct frand_state *rs;
2212 unsigned int left = max_bs;
2213 unsigned int this_write;
2216 rs = get_buf_state(td);
2218 min_write = min(min_write, left);
2221 this_write = min_not_zero(min_write,
2222 td->o.compress_chunk);
2224 fill_random_buf_percentage(rs, buf, perc,
2225 this_write, this_write,
2227 o->buffer_pattern_bytes);
2229 fill_random_buf(rs, buf, min_write);
2230 this_write = min_write;
2235 save_buf_state(td, rs);
2237 } else if (o->buffer_pattern_bytes)
2238 fill_buffer_pattern(td, buf, max_bs);
2239 else if (o->zero_buffers)
2240 memset(buf, 0, max_bs);
2242 fill_random_buf(get_buf_state(td), buf, max_bs);
2246 * "randomly" fill the buffer contents
2248 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2249 unsigned int min_write, unsigned int max_bs)
2251 io_u->buf_filled_len = 0;
2252 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2255 static int do_sync_file_range(const struct thread_data *td,
2258 off64_t offset, nbytes;
2260 offset = f->first_write;
2261 nbytes = f->last_write - f->first_write;
2266 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2269 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2273 if (io_u->ddir == DDIR_SYNC) {
2274 ret = fsync(io_u->file->fd);
2275 } else if (io_u->ddir == DDIR_DATASYNC) {
2276 #ifdef CONFIG_FDATASYNC
2277 ret = fdatasync(io_u->file->fd);
2279 ret = io_u->xfer_buflen;
2280 io_u->error = EINVAL;
2282 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2283 ret = do_sync_file_range(td, io_u->file);
2285 ret = io_u->xfer_buflen;
2286 io_u->error = EINVAL;
2290 io_u->error = errno;
2295 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2297 #ifndef FIO_HAVE_TRIM
2298 io_u->error = EINVAL;
2301 struct fio_file *f = io_u->file;
2304 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2306 return io_u->xfer_buflen;