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(struct thread_data *td,
161 struct fio_file *f, enum fio_ddir ddir,
164 unsigned int v, send, stotal;
165 uint64_t offset, lastb;
167 struct zone_split_index *zsi;
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);
183 zsi = &td->zone_state_index[ddir][v - 1];
184 stotal = zsi->size_perc_prev;
185 send = zsi->size_perc;
188 * Should never happen
192 log_err("fio: bug in zoned generation\n");
199 * 'send' is some percentage below or equal to 100 that
200 * marks the end of the current IO range. 'stotal' marks
201 * the start, in percent.
204 offset = stotal * lastb / 100ULL;
208 lastb = lastb * (send - stotal) / 100ULL;
211 * Generate index from 0..send-of-lastb
213 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
217 * Add our start offset, if any
225 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
227 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
228 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
230 return r1->off - r2->off;
233 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
234 enum fio_ddir ddir, uint64_t *b)
236 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
239 lastb = last_block(td, f, ddir);
243 return __get_next_rand_offset(td, f, ddir, b, lastb);
244 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
245 return __get_next_rand_offset_zipf(td, f, ddir, b);
246 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
247 return __get_next_rand_offset_pareto(td, f, ddir, b);
248 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
249 return __get_next_rand_offset_gauss(td, f, ddir, b);
250 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
251 return __get_next_rand_offset_zoned(td, f, ddir, b);
253 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
258 * Sort the reads for a verify phase in batches of verifysort_nr, if
261 static inline bool should_sort_io(struct thread_data *td)
263 if (!td->o.verifysort_nr || !td->o.do_verify)
267 if (td->runstate != TD_VERIFYING)
269 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
270 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
276 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
280 if (td->o.perc_rand[ddir] == 100)
283 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
285 return v <= td->o.perc_rand[ddir];
288 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
289 enum fio_ddir ddir, uint64_t *b)
294 if (!should_sort_io(td))
295 return get_off_from_method(td, f, ddir, b);
297 if (!flist_empty(&td->next_rand_list)) {
299 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
306 for (i = 0; i < td->o.verifysort_nr; i++) {
307 r = malloc(sizeof(*r));
309 ret = get_off_from_method(td, f, ddir, &r->off);
315 flist_add(&r->list, &td->next_rand_list);
321 assert(!flist_empty(&td->next_rand_list));
322 flist_sort(NULL, &td->next_rand_list, flist_cmp);
326 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
328 struct thread_options *o = &td->o;
330 if (o->invalidate_cache && !o->odirect) {
333 ret = file_invalidate_cache(td, f);
337 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
338 enum fio_ddir ddir, uint64_t *b)
340 if (!get_next_rand_offset(td, f, ddir, b))
343 if (td->o.time_based ||
344 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
345 fio_file_reset(td, f);
346 if (!get_next_rand_offset(td, f, ddir, b))
348 loop_cache_invalidate(td, f);
351 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
352 f->file_name, (unsigned long long) f->last_pos[ddir],
353 (unsigned long long) f->real_file_size);
357 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
358 enum fio_ddir ddir, uint64_t *offset)
360 struct thread_options *o = &td->o;
362 assert(ddir_rw(ddir));
365 * If we reach the end for a time based run, reset us back to 0
366 * and invalidate the cache, if we need to.
368 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
370 f->last_pos[ddir] = f->file_offset;
371 loop_cache_invalidate(td, f);
374 if (f->last_pos[ddir] < f->real_file_size) {
377 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
378 if (f->real_file_size > f->io_size)
379 f->last_pos[ddir] = f->io_size;
381 f->last_pos[ddir] = f->real_file_size;
384 pos = f->last_pos[ddir] - f->file_offset;
385 if (pos && o->ddir_seq_add) {
386 pos += o->ddir_seq_add;
389 * If we reach beyond the end of the file
390 * with holed IO, wrap around to the
391 * beginning again. If we're doing backwards IO,
394 if (pos >= f->real_file_size) {
395 if (o->ddir_seq_add > 0)
396 pos = f->file_offset;
398 if (f->real_file_size > f->io_size)
401 pos = f->real_file_size;
403 pos += o->ddir_seq_add;
415 static int get_next_block(struct thread_data *td, struct io_u *io_u,
416 enum fio_ddir ddir, int rw_seq,
417 unsigned int *is_random)
419 struct fio_file *f = io_u->file;
423 assert(ddir_rw(ddir));
429 if (should_do_random(td, ddir)) {
430 ret = get_next_rand_block(td, f, ddir, &b);
434 io_u_set(td, io_u, IO_U_F_BUSY_OK);
435 ret = get_next_seq_offset(td, f, ddir, &offset);
437 ret = get_next_rand_block(td, f, ddir, &b);
441 ret = get_next_seq_offset(td, f, ddir, &offset);
444 io_u_set(td, io_u, IO_U_F_BUSY_OK);
447 if (td->o.rw_seq == RW_SEQ_SEQ) {
448 ret = get_next_seq_offset(td, f, ddir, &offset);
450 ret = get_next_rand_block(td, f, ddir, &b);
453 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
454 if (f->last_start[ddir] != -1ULL)
455 offset = f->last_start[ddir] - f->file_offset;
460 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
467 io_u->offset = offset;
469 io_u->offset = b * td->o.ba[ddir];
471 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
480 * For random io, generate a random new block and see if it's used. Repeat
481 * until we find a free one. For sequential io, just return the end of
482 * the last io issued.
484 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
485 unsigned int *is_random)
487 struct fio_file *f = io_u->file;
488 enum fio_ddir ddir = io_u->ddir;
491 assert(ddir_rw(ddir));
493 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
495 td->ddir_seq_nr = td->o.ddir_seq_nr;
498 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
501 if (io_u->offset >= f->io_size) {
502 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
503 (unsigned long long) io_u->offset,
504 (unsigned long long) f->io_size);
508 io_u->offset += f->file_offset;
509 if (io_u->offset >= f->real_file_size) {
510 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
511 (unsigned long long) io_u->offset,
512 (unsigned long long) f->real_file_size);
519 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
520 unsigned int *is_random)
522 if (td->flags & TD_F_PROFILE_OPS) {
523 struct prof_io_ops *ops = &td->prof_io_ops;
525 if (ops->fill_io_u_off)
526 return ops->fill_io_u_off(td, io_u, is_random);
529 return __get_next_offset(td, io_u, is_random);
532 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
535 struct fio_file *f = io_u->file;
537 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
540 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
541 unsigned int is_random)
543 int ddir = io_u->ddir;
544 unsigned int buflen = 0;
545 unsigned int minbs, maxbs;
546 uint64_t frand_max, r;
549 assert(ddir_rw(ddir));
551 if (td->o.bs_is_seq_rand)
552 ddir = is_random ? DDIR_WRITE: DDIR_READ;
554 minbs = td->o.min_bs[ddir];
555 maxbs = td->o.max_bs[ddir];
561 * If we can't satisfy the min block size from here, then fail
563 if (!io_u_fits(td, io_u, minbs))
566 frand_max = rand_max(&td->bsrange_state[ddir]);
568 r = __rand(&td->bsrange_state[ddir]);
570 if (!td->o.bssplit_nr[ddir]) {
571 buflen = 1 + (unsigned int) ((double) maxbs *
572 (r / (frand_max + 1.0)));
579 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
580 struct bssplit *bsp = &td->o.bssplit[ddir][i];
586 if ((r / perc <= frand_max / 100ULL) &&
587 io_u_fits(td, io_u, buflen))
592 power_2 = is_power_of_2(minbs);
593 if (!td->o.bs_unaligned && power_2)
594 buflen &= ~(minbs - 1);
595 else if (!td->o.bs_unaligned && !power_2)
596 buflen -= buflen % minbs;
597 } while (!io_u_fits(td, io_u, buflen));
602 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
603 unsigned int is_random)
605 if (td->flags & TD_F_PROFILE_OPS) {
606 struct prof_io_ops *ops = &td->prof_io_ops;
608 if (ops->fill_io_u_size)
609 return ops->fill_io_u_size(td, io_u, is_random);
612 return __get_next_buflen(td, io_u, is_random);
615 static void set_rwmix_bytes(struct thread_data *td)
620 * we do time or byte based switch. this is needed because
621 * buffered writes may issue a lot quicker than they complete,
622 * whereas reads do not.
624 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
625 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
628 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
632 v = rand32_between(&td->rwmix_state, 1, 100);
634 if (v <= td->o.rwmix[DDIR_READ])
640 int io_u_quiesce(struct thread_data *td)
645 * We are going to sleep, ensure that we flush anything pending as
646 * not to skew our latency numbers.
648 * Changed to only monitor 'in flight' requests here instead of the
649 * td->cur_depth, b/c td->cur_depth does not accurately represent
650 * io's that have been actually submitted to an async engine,
651 * and cur_depth is meaningless for sync engines.
653 if (td->io_u_queued || td->cur_depth) {
656 ret = td_io_commit(td);
659 while (td->io_u_in_flight) {
662 ret = io_u_queued_complete(td, 1);
667 if (td->flags & TD_F_REGROW_LOGS)
673 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
675 enum fio_ddir odir = ddir ^ 1;
679 assert(ddir_rw(ddir));
680 now = utime_since_now(&td->start);
683 * if rate_next_io_time is in the past, need to catch up to rate
685 if (td->rate_next_io_time[ddir] <= now)
689 * We are ahead of rate in this direction. See if we
692 if (td_rw(td) && td->o.rwmix[odir]) {
694 * Other direction is behind rate, switch
696 if (td->rate_next_io_time[odir] <= now)
700 * Both directions are ahead of rate. sleep the min
701 * switch if necissary
703 if (td->rate_next_io_time[ddir] <=
704 td->rate_next_io_time[odir]) {
705 usec = td->rate_next_io_time[ddir] - now;
707 usec = td->rate_next_io_time[odir] - now;
711 usec = td->rate_next_io_time[ddir] - now;
713 if (td->o.io_submit_mode == IO_MODE_INLINE)
716 usec = usec_sleep(td, usec);
722 * Return the data direction for the next io_u. If the job is a
723 * mixed read/write workload, check the rwmix cycle and switch if
726 static enum fio_ddir get_rw_ddir(struct thread_data *td)
731 * See if it's time to fsync/fdatasync/sync_file_range first,
732 * and if not then move on to check regular I/Os.
734 if (should_fsync(td)) {
735 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
736 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
739 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
740 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
741 return DDIR_DATASYNC;
743 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
744 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
745 return DDIR_SYNC_FILE_RANGE;
750 * Check if it's time to seed a new data direction.
752 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
754 * Put a top limit on how many bytes we do for
755 * one data direction, to avoid overflowing the
758 ddir = get_rand_ddir(td);
760 if (ddir != td->rwmix_ddir)
763 td->rwmix_ddir = ddir;
765 ddir = td->rwmix_ddir;
766 } else if (td_read(td))
768 else if (td_write(td))
770 else if (td_trim(td))
775 td->rwmix_ddir = rate_ddir(td, ddir);
776 return td->rwmix_ddir;
779 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
781 enum fio_ddir ddir = get_rw_ddir(td);
783 if (td_trimwrite(td)) {
784 struct fio_file *f = io_u->file;
785 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
791 io_u->ddir = io_u->acct_ddir = ddir;
793 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
794 td->o.barrier_blocks &&
795 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
796 td->io_issues[DDIR_WRITE])
797 io_u_set(td, io_u, IO_U_F_BARRIER);
800 void put_file_log(struct thread_data *td, struct fio_file *f)
802 unsigned int ret = put_file(td, f);
805 td_verror(td, ret, "file close");
808 void put_io_u(struct thread_data *td, struct io_u *io_u)
815 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
816 put_file_log(td, io_u->file);
819 io_u_set(td, io_u, IO_U_F_FREE);
821 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
823 assert(!(td->flags & TD_F_CHILD));
825 io_u_qpush(&td->io_u_freelist, io_u);
827 td_io_u_free_notify(td);
830 void clear_io_u(struct thread_data *td, struct io_u *io_u)
832 io_u_clear(td, io_u, IO_U_F_FLIGHT);
836 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
838 struct io_u *__io_u = *io_u;
839 enum fio_ddir ddir = acct_ddir(__io_u);
841 dprint(FD_IO, "requeue %p\n", __io_u);
848 io_u_set(td, __io_u, IO_U_F_FREE);
849 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
850 td->io_issues[ddir]--;
852 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
853 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
855 assert(!(td->flags & TD_F_CHILD));
858 io_u_rpush(&td->io_u_requeues, __io_u);
860 td_io_u_free_notify(td);
864 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
866 unsigned int is_random;
868 if (td_ioengine_flagged(td, FIO_NOIO))
871 set_rw_ddir(td, io_u);
874 * fsync() or fdatasync() or trim etc, we are done
876 if (!ddir_rw(io_u->ddir))
880 * See if it's time to switch to a new zone
882 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
883 struct fio_file *f = io_u->file;
886 f->file_offset += td->o.zone_range + td->o.zone_skip;
889 * Wrap from the beginning, if we exceed the file size
891 if (f->file_offset >= f->real_file_size)
892 f->file_offset = f->real_file_size - f->file_offset;
893 f->last_pos[io_u->ddir] = f->file_offset;
894 td->io_skip_bytes += td->o.zone_skip;
898 * No log, let the seq/rand engine retrieve the next buflen and
901 if (get_next_offset(td, io_u, &is_random)) {
902 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
906 io_u->buflen = get_next_buflen(td, io_u, is_random);
908 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
912 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
913 dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n",
915 dprint(FD_IO, " offset=%llu/buflen=%lu > %llu\n",
916 (unsigned long long) io_u->offset, io_u->buflen,
917 (unsigned long long) io_u->file->real_file_size);
922 * mark entry before potentially trimming io_u
924 if (td_random(td) && file_randommap(td, io_u->file))
925 mark_random_map(td, io_u);
928 dprint_io_u(io_u, "fill_io_u");
929 td->zone_bytes += io_u->buflen;
933 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
962 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
964 __io_u_mark_map(td->ts.io_u_submit, nr);
965 td->ts.total_submit++;
968 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
970 __io_u_mark_map(td->ts.io_u_complete, nr);
971 td->ts.total_complete++;
974 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
978 switch (td->cur_depth) {
1000 td->ts.io_u_map[idx] += nr;
1003 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1007 assert(nsec < 1000);
1040 assert(idx < FIO_IO_U_LAT_N_NR);
1041 td->ts.io_u_lat_n[idx]++;
1044 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1048 assert(usec < 1000 && usec >= 1);
1081 assert(idx < FIO_IO_U_LAT_U_NR);
1082 td->ts.io_u_lat_u[idx]++;
1085 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1128 assert(idx < FIO_IO_U_LAT_M_NR);
1129 td->ts.io_u_lat_m[idx]++;
1132 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1135 io_u_mark_lat_nsec(td, nsec);
1136 else if (nsec < 1000000)
1137 io_u_mark_lat_usec(td, nsec / 1000);
1139 io_u_mark_lat_msec(td, nsec / 1000000);
1142 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1144 unsigned long fileno;
1146 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1147 uint64_t frand_max = rand_max(&td->next_file_state);
1150 r = __rand(&td->next_file_state);
1151 return (unsigned int) ((double) td->o.nr_files
1152 * (r / (frand_max + 1.0)));
1155 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1156 fileno = zipf_next(&td->next_file_zipf);
1157 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1158 fileno = pareto_next(&td->next_file_zipf);
1159 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1160 fileno = gauss_next(&td->next_file_gauss);
1162 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1167 return fileno >> FIO_FSERVICE_SHIFT;
1171 * Get next file to service by choosing one at random
1173 static struct fio_file *get_next_file_rand(struct thread_data *td,
1174 enum fio_file_flags goodf,
1175 enum fio_file_flags badf)
1183 fno = __get_next_fileno_rand(td);
1186 if (fio_file_done(f))
1189 if (!fio_file_open(f)) {
1192 if (td->nr_open_files >= td->o.open_files)
1193 return ERR_PTR(-EBUSY);
1195 err = td_io_open_file(td, f);
1201 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1202 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1206 td_io_close_file(td, f);
1211 * Get next file to service by doing round robin between all available ones
1213 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1216 unsigned int old_next_file = td->next_file;
1222 f = td->files[td->next_file];
1225 if (td->next_file >= td->o.nr_files)
1228 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1229 if (fio_file_done(f)) {
1234 if (!fio_file_open(f)) {
1237 if (td->nr_open_files >= td->o.open_files)
1238 return ERR_PTR(-EBUSY);
1240 err = td_io_open_file(td, f);
1242 dprint(FD_FILE, "error %d on open of %s\n",
1250 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1252 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1256 td_io_close_file(td, f);
1259 } while (td->next_file != old_next_file);
1261 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1265 static struct fio_file *__get_next_file(struct thread_data *td)
1269 assert(td->o.nr_files <= td->files_index);
1271 if (td->nr_done_files >= td->o.nr_files) {
1272 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1273 " nr_files=%d\n", td->nr_open_files,
1279 f = td->file_service_file;
1280 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1281 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1283 if (td->file_service_left--)
1287 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1288 td->o.file_service_type == FIO_FSERVICE_SEQ)
1289 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1291 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1296 td->file_service_file = f;
1297 td->file_service_left = td->file_service_nr - 1;
1300 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1302 dprint(FD_FILE, "get_next_file: NULL\n");
1306 static struct fio_file *get_next_file(struct thread_data *td)
1308 if (td->flags & TD_F_PROFILE_OPS) {
1309 struct prof_io_ops *ops = &td->prof_io_ops;
1311 if (ops->get_next_file)
1312 return ops->get_next_file(td);
1315 return __get_next_file(td);
1318 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1323 f = get_next_file(td);
1324 if (IS_ERR_OR_NULL(f))
1330 if (!fill_io_u(td, io_u))
1333 put_file_log(td, f);
1334 td_io_close_file(td, f);
1336 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1337 fio_file_reset(td, f);
1339 fio_file_set_done(f);
1340 td->nr_done_files++;
1341 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1342 td->nr_done_files, td->o.nr_files);
1349 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1350 unsigned long tusec, unsigned long max_usec)
1353 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1354 td_verror(td, ETIMEDOUT, "max latency exceeded");
1355 icd->error = ETIMEDOUT;
1358 static void lat_new_cycle(struct thread_data *td)
1360 fio_gettime(&td->latency_ts, NULL);
1361 td->latency_ios = ddir_rw_sum(td->io_blocks);
1362 td->latency_failed = 0;
1366 * We had an IO outside the latency target. Reduce the queue depth. If we
1367 * are at QD=1, then it's time to give up.
1369 static bool __lat_target_failed(struct thread_data *td)
1371 if (td->latency_qd == 1)
1374 td->latency_qd_high = td->latency_qd;
1376 if (td->latency_qd == td->latency_qd_low)
1377 td->latency_qd_low--;
1379 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1381 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1384 * When we ramp QD down, quiesce existing IO to prevent
1385 * a storm of ramp downs due to pending higher depth.
1392 static bool lat_target_failed(struct thread_data *td)
1394 if (td->o.latency_percentile.u.f == 100.0)
1395 return __lat_target_failed(td);
1397 td->latency_failed++;
1401 void lat_target_init(struct thread_data *td)
1403 td->latency_end_run = 0;
1405 if (td->o.latency_target) {
1406 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1407 fio_gettime(&td->latency_ts, NULL);
1409 td->latency_qd_high = td->o.iodepth;
1410 td->latency_qd_low = 1;
1411 td->latency_ios = ddir_rw_sum(td->io_blocks);
1413 td->latency_qd = td->o.iodepth;
1416 void lat_target_reset(struct thread_data *td)
1418 if (!td->latency_end_run)
1419 lat_target_init(td);
1422 static void lat_target_success(struct thread_data *td)
1424 const unsigned int qd = td->latency_qd;
1425 struct thread_options *o = &td->o;
1427 td->latency_qd_low = td->latency_qd;
1430 * If we haven't failed yet, we double up to a failing value instead
1431 * of bisecting from highest possible queue depth. If we have set
1432 * a limit other than td->o.iodepth, bisect between that.
1434 if (td->latency_qd_high != o->iodepth)
1435 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1437 td->latency_qd *= 2;
1439 if (td->latency_qd > o->iodepth)
1440 td->latency_qd = o->iodepth;
1442 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1445 * Same as last one, we are done. Let it run a latency cycle, so
1446 * we get only the results from the targeted depth.
1448 if (td->latency_qd == qd) {
1449 if (td->latency_end_run) {
1450 dprint(FD_RATE, "We are done\n");
1453 dprint(FD_RATE, "Quiesce and final run\n");
1455 td->latency_end_run = 1;
1456 reset_all_stats(td);
1465 * Check if we can bump the queue depth
1467 void lat_target_check(struct thread_data *td)
1469 uint64_t usec_window;
1473 usec_window = utime_since_now(&td->latency_ts);
1474 if (usec_window < td->o.latency_window)
1477 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1478 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1479 success_ios *= 100.0;
1481 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1483 if (success_ios >= td->o.latency_percentile.u.f)
1484 lat_target_success(td);
1486 __lat_target_failed(td);
1490 * If latency target is enabled, we might be ramping up or down and not
1491 * using the full queue depth available.
1493 bool queue_full(const struct thread_data *td)
1495 const int qempty = io_u_qempty(&td->io_u_freelist);
1499 if (!td->o.latency_target)
1502 return td->cur_depth >= td->latency_qd;
1505 struct io_u *__get_io_u(struct thread_data *td)
1507 struct io_u *io_u = NULL;
1515 if (!io_u_rempty(&td->io_u_requeues))
1516 io_u = io_u_rpop(&td->io_u_requeues);
1517 else if (!queue_full(td)) {
1518 io_u = io_u_qpop(&td->io_u_freelist);
1523 io_u->end_io = NULL;
1527 assert(io_u->flags & IO_U_F_FREE);
1528 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1529 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1533 io_u->acct_ddir = -1;
1535 assert(!(td->flags & TD_F_CHILD));
1536 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1538 } else if (td_async_processing(td)) {
1540 * We ran out, wait for async verify threads to finish and
1543 assert(!(td->flags & TD_F_CHILD));
1544 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1552 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1554 if (!(td->flags & TD_F_TRIM_BACKLOG))
1557 if (td->trim_entries) {
1560 if (td->trim_batch) {
1563 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1564 td->last_ddir != DDIR_READ) {
1565 td->trim_batch = td->o.trim_batch;
1566 if (!td->trim_batch)
1567 td->trim_batch = td->o.trim_backlog;
1571 if (get_trim && get_next_trim(td, io_u))
1578 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1580 if (!(td->flags & TD_F_VER_BACKLOG))
1583 if (td->io_hist_len) {
1586 if (td->verify_batch)
1588 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1589 td->last_ddir != DDIR_READ) {
1590 td->verify_batch = td->o.verify_batch;
1591 if (!td->verify_batch)
1592 td->verify_batch = td->o.verify_backlog;
1596 if (get_verify && !get_next_verify(td, io_u)) {
1606 * Fill offset and start time into the buffer content, to prevent too
1607 * easy compressible data for simple de-dupe attempts. Do this for every
1608 * 512b block in the range, since that should be the smallest block size
1609 * we can expect from a device.
1611 static void small_content_scramble(struct io_u *io_u)
1613 unsigned int i, nr_blocks = io_u->buflen / 512;
1615 unsigned int offset;
1622 boffset = io_u->offset;
1623 io_u->buf_filled_len = 0;
1625 for (i = 0; i < nr_blocks; i++) {
1627 * Fill the byte offset into a "random" start offset of
1628 * the buffer, given by the product of the usec time
1629 * and the actual offset.
1631 offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511;
1632 offset &= ~(sizeof(uint64_t) - 1);
1633 if (offset >= 512 - sizeof(uint64_t))
1634 offset -= sizeof(uint64_t);
1635 memcpy(p + offset, &boffset, sizeof(boffset));
1637 end = p + 512 - sizeof(io_u->start_time);
1638 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1645 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1646 * etc. The returned io_u is fully ready to be prepped and submitted.
1648 struct io_u *get_io_u(struct thread_data *td)
1652 int do_scramble = 0;
1655 io_u = __get_io_u(td);
1657 dprint(FD_IO, "__get_io_u failed\n");
1661 if (check_get_verify(td, io_u))
1663 if (check_get_trim(td, io_u))
1667 * from a requeue, io_u already setup
1673 * If using an iolog, grab next piece if any available.
1675 if (td->flags & TD_F_READ_IOLOG) {
1676 if (read_iolog_get(td, io_u))
1678 } else if (set_io_u_file(td, io_u)) {
1680 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1686 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1690 assert(fio_file_open(f));
1692 if (ddir_rw(io_u->ddir)) {
1693 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1694 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1698 f->last_start[io_u->ddir] = io_u->offset;
1699 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1701 if (io_u->ddir == DDIR_WRITE) {
1702 if (td->flags & TD_F_REFILL_BUFFERS) {
1703 io_u_fill_buffer(td, io_u,
1704 td->o.min_bs[DDIR_WRITE],
1706 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1707 !(td->flags & TD_F_COMPRESS))
1709 if (td->flags & TD_F_VER_NONE) {
1710 populate_verify_io_u(td, io_u);
1713 } else if (io_u->ddir == DDIR_READ) {
1715 * Reset the buf_filled parameters so next time if the
1716 * buffer is used for writes it is refilled.
1718 io_u->buf_filled_len = 0;
1723 * Set io data pointers.
1725 io_u->xfer_buf = io_u->buf;
1726 io_u->xfer_buflen = io_u->buflen;
1730 if (!td_io_prep(td, io_u)) {
1731 if (!td->o.disable_lat)
1732 fio_gettime(&io_u->start_time, NULL);
1735 small_content_scramble(io_u);
1740 dprint(FD_IO, "get_io_u failed\n");
1742 return ERR_PTR(ret);
1745 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1747 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1749 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1752 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1753 io_u->file ? " on file " : "",
1754 io_u->file ? io_u->file->file_name : "",
1755 strerror(io_u->error),
1756 io_ddir_name(io_u->ddir),
1757 io_u->offset, io_u->xfer_buflen);
1759 if (td->io_ops->errdetails) {
1760 char *err = td->io_ops->errdetails(io_u);
1762 log_err("fio: %s\n", err);
1767 td_verror(td, io_u->error, "io_u error");
1770 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1772 __io_u_log_error(td, io_u);
1774 __io_u_log_error(td->parent, io_u);
1777 static inline bool gtod_reduce(struct thread_data *td)
1779 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1780 || td->o.gtod_reduce;
1783 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1784 struct io_completion_data *icd,
1785 const enum fio_ddir idx, unsigned int bytes)
1787 const int no_reduce = !gtod_reduce(td);
1788 unsigned long long llnsec = 0;
1793 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1797 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1799 if (!td->o.disable_lat) {
1800 unsigned long long tnsec;
1802 tnsec = ntime_since(&io_u->start_time, &icd->time);
1803 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1805 if (td->flags & TD_F_PROFILE_OPS) {
1806 struct prof_io_ops *ops = &td->prof_io_ops;
1809 icd->error = ops->io_u_lat(td, tnsec/1000);
1812 if (td->o.max_latency && tnsec/1000 > td->o.max_latency)
1813 lat_fatal(td, icd, tnsec/1000, td->o.max_latency);
1814 if (td->o.latency_target && tnsec/1000 > td->o.latency_target) {
1815 if (lat_target_failed(td))
1816 lat_fatal(td, icd, tnsec/1000, td->o.latency_target);
1821 if (!td->o.disable_clat) {
1822 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1823 io_u_mark_latency(td, llnsec);
1826 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1827 add_bw_sample(td, io_u, bytes, llnsec);
1829 if (no_reduce && per_unit_log(td->iops_log))
1830 add_iops_sample(td, io_u, bytes);
1833 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1834 uint32_t *info = io_u_block_info(td, io_u);
1835 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1836 if (io_u->ddir == DDIR_TRIM) {
1837 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1838 BLOCK_INFO_TRIMS(*info) + 1);
1839 } else if (io_u->ddir == DDIR_WRITE) {
1840 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1847 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1848 uint64_t offset, unsigned int bytes)
1855 if (f->first_write == -1ULL || offset < f->first_write)
1856 f->first_write = offset;
1857 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1858 f->last_write = offset + bytes;
1860 if (!f->last_write_comp)
1863 idx = f->last_write_idx++;
1864 f->last_write_comp[idx] = offset;
1865 if (f->last_write_idx == td->o.iodepth)
1866 f->last_write_idx = 0;
1869 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1870 struct io_completion_data *icd)
1872 struct io_u *io_u = *io_u_ptr;
1873 enum fio_ddir ddir = io_u->ddir;
1874 struct fio_file *f = io_u->file;
1876 dprint_io_u(io_u, "io complete");
1878 assert(io_u->flags & IO_U_F_FLIGHT);
1879 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1882 * Mark IO ok to verify
1886 * Remove errored entry from the verification list
1889 unlog_io_piece(td, io_u);
1891 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1896 if (ddir_sync(ddir)) {
1897 td->last_was_sync = 1;
1899 f->first_write = -1ULL;
1900 f->last_write = -1ULL;
1905 td->last_was_sync = 0;
1906 td->last_ddir = ddir;
1908 if (!io_u->error && ddir_rw(ddir)) {
1909 unsigned int bytes = io_u->buflen - io_u->resid;
1912 td->io_blocks[ddir]++;
1913 td->this_io_blocks[ddir]++;
1914 td->io_bytes[ddir] += bytes;
1916 if (!(io_u->flags & IO_U_F_VER_LIST))
1917 td->this_io_bytes[ddir] += bytes;
1919 if (ddir == DDIR_WRITE)
1920 file_log_write_comp(td, f, io_u->offset, bytes);
1922 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1923 td->runstate == TD_VERIFYING))
1924 account_io_completion(td, io_u, icd, ddir, bytes);
1926 icd->bytes_done[ddir] += bytes;
1929 ret = io_u->end_io(td, io_u_ptr);
1931 if (ret && !icd->error)
1934 } else if (io_u->error) {
1935 icd->error = io_u->error;
1936 io_u_log_error(td, io_u);
1939 enum error_type_bit eb = td_error_type(ddir, icd->error);
1941 if (!td_non_fatal_error(td, eb, icd->error))
1945 * If there is a non_fatal error, then add to the error count
1946 * and clear all the errors.
1948 update_error_count(td, icd->error);
1956 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1961 if (!gtod_reduce(td))
1962 fio_gettime(&icd->time, NULL);
1967 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1968 icd->bytes_done[ddir] = 0;
1971 static void ios_completed(struct thread_data *td,
1972 struct io_completion_data *icd)
1977 for (i = 0; i < icd->nr; i++) {
1978 io_u = td->io_ops->event(td, i);
1980 io_completed(td, &io_u, icd);
1988 * Complete a single io_u for the sync engines.
1990 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1992 struct io_completion_data icd;
1995 init_icd(td, &icd, 1);
1996 io_completed(td, &io_u, &icd);
2002 td_verror(td, icd.error, "io_u_sync_complete");
2006 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2007 td->bytes_done[ddir] += icd.bytes_done[ddir];
2013 * Called to complete min_events number of io for the async engines.
2015 int io_u_queued_complete(struct thread_data *td, int min_evts)
2017 struct io_completion_data icd;
2018 struct timespec *tvp = NULL;
2020 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2022 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2026 else if (min_evts > td->cur_depth)
2027 min_evts = td->cur_depth;
2029 /* No worries, td_io_getevents fixes min and max if they are
2030 * set incorrectly */
2031 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2033 td_verror(td, -ret, "td_io_getevents");
2038 init_icd(td, &icd, ret);
2039 ios_completed(td, &icd);
2041 td_verror(td, icd.error, "io_u_queued_complete");
2045 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2046 td->bytes_done[ddir] += icd.bytes_done[ddir];
2052 * Call when io_u is really queued, to update the submission latency.
2054 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2056 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2057 unsigned long slat_time;
2059 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2064 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2070 * See if we should reuse the last seed, if dedupe is enabled
2072 static struct frand_state *get_buf_state(struct thread_data *td)
2076 if (!td->o.dedupe_percentage)
2077 return &td->buf_state;
2078 else if (td->o.dedupe_percentage == 100) {
2079 frand_copy(&td->buf_state_prev, &td->buf_state);
2080 return &td->buf_state;
2083 v = rand32_between(&td->dedupe_state, 1, 100);
2085 if (v <= td->o.dedupe_percentage)
2086 return &td->buf_state_prev;
2088 return &td->buf_state;
2091 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2093 if (td->o.dedupe_percentage == 100)
2094 frand_copy(rs, &td->buf_state_prev);
2095 else if (rs == &td->buf_state)
2096 frand_copy(&td->buf_state_prev, rs);
2099 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2100 unsigned int max_bs)
2102 struct thread_options *o = &td->o;
2104 if (o->mem_type == MEM_CUDA_MALLOC)
2107 if (o->compress_percentage || o->dedupe_percentage) {
2108 unsigned int perc = td->o.compress_percentage;
2109 struct frand_state *rs;
2110 unsigned int left = max_bs;
2111 unsigned int this_write;
2114 rs = get_buf_state(td);
2116 min_write = min(min_write, left);
2119 this_write = min_not_zero(min_write,
2120 td->o.compress_chunk);
2122 fill_random_buf_percentage(rs, buf, perc,
2123 this_write, this_write,
2125 o->buffer_pattern_bytes);
2127 fill_random_buf(rs, buf, min_write);
2128 this_write = min_write;
2133 save_buf_state(td, rs);
2135 } else if (o->buffer_pattern_bytes)
2136 fill_buffer_pattern(td, buf, max_bs);
2137 else if (o->zero_buffers)
2138 memset(buf, 0, max_bs);
2140 fill_random_buf(get_buf_state(td), buf, max_bs);
2144 * "randomly" fill the buffer contents
2146 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2147 unsigned int min_write, unsigned int max_bs)
2149 io_u->buf_filled_len = 0;
2150 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2153 static int do_sync_file_range(const struct thread_data *td,
2156 off64_t offset, nbytes;
2158 offset = f->first_write;
2159 nbytes = f->last_write - f->first_write;
2164 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2167 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2171 if (io_u->ddir == DDIR_SYNC) {
2172 ret = fsync(io_u->file->fd);
2173 } else if (io_u->ddir == DDIR_DATASYNC) {
2174 #ifdef CONFIG_FDATASYNC
2175 ret = fdatasync(io_u->file->fd);
2177 ret = io_u->xfer_buflen;
2178 io_u->error = EINVAL;
2180 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2181 ret = do_sync_file_range(td, io_u->file);
2183 ret = io_u->xfer_buflen;
2184 io_u->error = EINVAL;
2188 io_u->error = errno;
2193 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2195 #ifndef FIO_HAVE_TRIM
2196 io_u->error = EINVAL;
2199 struct fio_file *f = io_u->file;
2202 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2204 return io_u->xfer_buflen;