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));
364 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
366 struct thread_options *o = &td->o;
367 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
369 if (io_size > f->last_pos[ddir])
370 f->last_pos[ddir] = 0;
372 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
374 loop_cache_invalidate(td, f);
377 if (f->last_pos[ddir] < f->real_file_size) {
380 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
381 if (f->real_file_size > f->io_size)
382 f->last_pos[ddir] = f->io_size;
384 f->last_pos[ddir] = f->real_file_size;
387 pos = f->last_pos[ddir] - f->file_offset;
388 if (pos && o->ddir_seq_add) {
389 pos += o->ddir_seq_add;
392 * If we reach beyond the end of the file
393 * with holed IO, wrap around to the
394 * beginning again. If we're doing backwards IO,
397 if (pos >= f->real_file_size) {
398 if (o->ddir_seq_add > 0)
399 pos = f->file_offset;
401 if (f->real_file_size > f->io_size)
404 pos = f->real_file_size;
406 pos += o->ddir_seq_add;
418 static int get_next_block(struct thread_data *td, struct io_u *io_u,
419 enum fio_ddir ddir, int rw_seq,
420 unsigned int *is_random)
422 struct fio_file *f = io_u->file;
426 assert(ddir_rw(ddir));
432 if (should_do_random(td, ddir)) {
433 ret = get_next_rand_block(td, f, ddir, &b);
437 io_u_set(td, io_u, IO_U_F_BUSY_OK);
438 ret = get_next_seq_offset(td, f, ddir, &offset);
440 ret = get_next_rand_block(td, f, ddir, &b);
444 ret = get_next_seq_offset(td, f, ddir, &offset);
447 io_u_set(td, io_u, IO_U_F_BUSY_OK);
450 if (td->o.rw_seq == RW_SEQ_SEQ) {
451 ret = get_next_seq_offset(td, f, ddir, &offset);
453 ret = get_next_rand_block(td, f, ddir, &b);
456 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
457 if (f->last_start[ddir] != -1ULL)
458 offset = f->last_start[ddir] - f->file_offset;
463 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
470 io_u->offset = offset;
472 io_u->offset = b * td->o.ba[ddir];
474 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
483 * For random io, generate a random new block and see if it's used. Repeat
484 * until we find a free one. For sequential io, just return the end of
485 * the last io issued.
487 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
488 unsigned int *is_random)
490 struct fio_file *f = io_u->file;
491 enum fio_ddir ddir = io_u->ddir;
494 assert(ddir_rw(ddir));
496 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
498 td->ddir_seq_nr = td->o.ddir_seq_nr;
501 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
504 if (io_u->offset >= f->io_size) {
505 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
506 (unsigned long long) io_u->offset,
507 (unsigned long long) f->io_size);
511 io_u->offset += f->file_offset;
512 if (io_u->offset >= f->real_file_size) {
513 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
514 (unsigned long long) io_u->offset,
515 (unsigned long long) f->real_file_size);
522 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
523 unsigned int *is_random)
525 if (td->flags & TD_F_PROFILE_OPS) {
526 struct prof_io_ops *ops = &td->prof_io_ops;
528 if (ops->fill_io_u_off)
529 return ops->fill_io_u_off(td, io_u, is_random);
532 return __get_next_offset(td, io_u, is_random);
535 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
538 struct fio_file *f = io_u->file;
540 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
543 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
544 unsigned int is_random)
546 int ddir = io_u->ddir;
547 unsigned int buflen = 0;
548 unsigned int minbs, maxbs;
549 uint64_t frand_max, r;
552 assert(ddir_rw(ddir));
554 if (td->o.bs_is_seq_rand)
555 ddir = is_random ? DDIR_WRITE: DDIR_READ;
557 minbs = td->o.min_bs[ddir];
558 maxbs = td->o.max_bs[ddir];
564 * If we can't satisfy the min block size from here, then fail
566 if (!io_u_fits(td, io_u, minbs))
569 frand_max = rand_max(&td->bsrange_state[ddir]);
571 r = __rand(&td->bsrange_state[ddir]);
573 if (!td->o.bssplit_nr[ddir]) {
574 buflen = 1 + (unsigned int) ((double) maxbs *
575 (r / (frand_max + 1.0)));
582 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
583 struct bssplit *bsp = &td->o.bssplit[ddir][i];
589 if ((r / perc <= frand_max / 100ULL) &&
590 io_u_fits(td, io_u, buflen))
595 power_2 = is_power_of_2(minbs);
596 if (!td->o.bs_unaligned && power_2)
597 buflen &= ~(minbs - 1);
598 else if (!td->o.bs_unaligned && !power_2)
599 buflen -= buflen % minbs;
600 } while (!io_u_fits(td, io_u, buflen));
605 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
606 unsigned int is_random)
608 if (td->flags & TD_F_PROFILE_OPS) {
609 struct prof_io_ops *ops = &td->prof_io_ops;
611 if (ops->fill_io_u_size)
612 return ops->fill_io_u_size(td, io_u, is_random);
615 return __get_next_buflen(td, io_u, is_random);
618 static void set_rwmix_bytes(struct thread_data *td)
623 * we do time or byte based switch. this is needed because
624 * buffered writes may issue a lot quicker than they complete,
625 * whereas reads do not.
627 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
628 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
631 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
635 v = rand32_between(&td->rwmix_state, 1, 100);
637 if (v <= td->o.rwmix[DDIR_READ])
643 int io_u_quiesce(struct thread_data *td)
648 * We are going to sleep, ensure that we flush anything pending as
649 * not to skew our latency numbers.
651 * Changed to only monitor 'in flight' requests here instead of the
652 * td->cur_depth, b/c td->cur_depth does not accurately represent
653 * io's that have been actually submitted to an async engine,
654 * and cur_depth is meaningless for sync engines.
656 if (td->io_u_queued || td->cur_depth) {
659 ret = td_io_commit(td);
662 while (td->io_u_in_flight) {
665 ret = io_u_queued_complete(td, 1);
670 if (td->flags & TD_F_REGROW_LOGS)
676 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
678 enum fio_ddir odir = ddir ^ 1;
682 assert(ddir_rw(ddir));
683 now = utime_since_now(&td->start);
686 * if rate_next_io_time is in the past, need to catch up to rate
688 if (td->rate_next_io_time[ddir] <= now)
692 * We are ahead of rate in this direction. See if we
695 if (td_rw(td) && td->o.rwmix[odir]) {
697 * Other direction is behind rate, switch
699 if (td->rate_next_io_time[odir] <= now)
703 * Both directions are ahead of rate. sleep the min
704 * switch if necissary
706 if (td->rate_next_io_time[ddir] <=
707 td->rate_next_io_time[odir]) {
708 usec = td->rate_next_io_time[ddir] - now;
710 usec = td->rate_next_io_time[odir] - now;
714 usec = td->rate_next_io_time[ddir] - now;
716 if (td->o.io_submit_mode == IO_MODE_INLINE)
719 usec = usec_sleep(td, usec);
725 * Return the data direction for the next io_u. If the job is a
726 * mixed read/write workload, check the rwmix cycle and switch if
729 static enum fio_ddir get_rw_ddir(struct thread_data *td)
734 * See if it's time to fsync/fdatasync/sync_file_range first,
735 * and if not then move on to check regular I/Os.
737 if (should_fsync(td)) {
738 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
739 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
742 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
743 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
744 return DDIR_DATASYNC;
746 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
747 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
748 return DDIR_SYNC_FILE_RANGE;
753 * Check if it's time to seed a new data direction.
755 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
757 * Put a top limit on how many bytes we do for
758 * one data direction, to avoid overflowing the
761 ddir = get_rand_ddir(td);
763 if (ddir != td->rwmix_ddir)
766 td->rwmix_ddir = ddir;
768 ddir = td->rwmix_ddir;
769 } else if (td_read(td))
771 else if (td_write(td))
773 else if (td_trim(td))
778 td->rwmix_ddir = rate_ddir(td, ddir);
779 return td->rwmix_ddir;
782 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
784 enum fio_ddir ddir = get_rw_ddir(td);
786 if (td_trimwrite(td)) {
787 struct fio_file *f = io_u->file;
788 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
794 io_u->ddir = io_u->acct_ddir = ddir;
796 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
797 td->o.barrier_blocks &&
798 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
799 td->io_issues[DDIR_WRITE])
800 io_u_set(td, io_u, IO_U_F_BARRIER);
803 void put_file_log(struct thread_data *td, struct fio_file *f)
805 unsigned int ret = put_file(td, f);
808 td_verror(td, ret, "file close");
811 void put_io_u(struct thread_data *td, struct io_u *io_u)
818 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
819 put_file_log(td, io_u->file);
822 io_u_set(td, io_u, IO_U_F_FREE);
824 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
826 assert(!(td->flags & TD_F_CHILD));
828 io_u_qpush(&td->io_u_freelist, io_u);
830 td_io_u_free_notify(td);
833 void clear_io_u(struct thread_data *td, struct io_u *io_u)
835 io_u_clear(td, io_u, IO_U_F_FLIGHT);
839 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
841 struct io_u *__io_u = *io_u;
842 enum fio_ddir ddir = acct_ddir(__io_u);
844 dprint(FD_IO, "requeue %p\n", __io_u);
851 io_u_set(td, __io_u, IO_U_F_FREE);
852 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
853 td->io_issues[ddir]--;
855 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
856 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
858 assert(!(td->flags & TD_F_CHILD));
861 io_u_rpush(&td->io_u_requeues, __io_u);
863 td_io_u_free_notify(td);
867 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
869 unsigned int is_random;
871 if (td_ioengine_flagged(td, FIO_NOIO))
874 set_rw_ddir(td, io_u);
877 * fsync() or fdatasync() or trim etc, we are done
879 if (!ddir_rw(io_u->ddir))
883 * See if it's time to switch to a new zone
885 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
886 struct fio_file *f = io_u->file;
889 f->file_offset += td->o.zone_range + td->o.zone_skip;
892 * Wrap from the beginning, if we exceed the file size
894 if (f->file_offset >= f->real_file_size)
895 f->file_offset = f->real_file_size - f->file_offset;
896 f->last_pos[io_u->ddir] = f->file_offset;
897 td->io_skip_bytes += td->o.zone_skip;
901 * No log, let the seq/rand engine retrieve the next buflen and
904 if (get_next_offset(td, io_u, &is_random)) {
905 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
909 io_u->buflen = get_next_buflen(td, io_u, is_random);
911 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
915 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
916 dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n",
918 dprint(FD_IO, " offset=%llu/buflen=%lu > %llu\n",
919 (unsigned long long) io_u->offset, io_u->buflen,
920 (unsigned long long) io_u->file->real_file_size);
925 * mark entry before potentially trimming io_u
927 if (td_random(td) && file_randommap(td, io_u->file))
928 mark_random_map(td, io_u);
931 dprint_io_u(io_u, "fill_io_u");
932 td->zone_bytes += io_u->buflen;
936 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
965 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
967 __io_u_mark_map(td->ts.io_u_submit, nr);
968 td->ts.total_submit++;
971 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
973 __io_u_mark_map(td->ts.io_u_complete, nr);
974 td->ts.total_complete++;
977 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
981 switch (td->cur_depth) {
1003 td->ts.io_u_map[idx] += nr;
1006 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1010 assert(nsec < 1000);
1043 assert(idx < FIO_IO_U_LAT_N_NR);
1044 td->ts.io_u_lat_n[idx]++;
1047 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1051 assert(usec < 1000 && usec >= 1);
1084 assert(idx < FIO_IO_U_LAT_U_NR);
1085 td->ts.io_u_lat_u[idx]++;
1088 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1131 assert(idx < FIO_IO_U_LAT_M_NR);
1132 td->ts.io_u_lat_m[idx]++;
1135 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1138 io_u_mark_lat_nsec(td, nsec);
1139 else if (nsec < 1000000)
1140 io_u_mark_lat_usec(td, nsec / 1000);
1142 io_u_mark_lat_msec(td, nsec / 1000000);
1145 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1147 unsigned long fileno;
1149 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1150 uint64_t frand_max = rand_max(&td->next_file_state);
1153 r = __rand(&td->next_file_state);
1154 return (unsigned int) ((double) td->o.nr_files
1155 * (r / (frand_max + 1.0)));
1158 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1159 fileno = zipf_next(&td->next_file_zipf);
1160 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1161 fileno = pareto_next(&td->next_file_zipf);
1162 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1163 fileno = gauss_next(&td->next_file_gauss);
1165 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1170 return fileno >> FIO_FSERVICE_SHIFT;
1174 * Get next file to service by choosing one at random
1176 static struct fio_file *get_next_file_rand(struct thread_data *td,
1177 enum fio_file_flags goodf,
1178 enum fio_file_flags badf)
1186 fno = __get_next_fileno_rand(td);
1189 if (fio_file_done(f))
1192 if (!fio_file_open(f)) {
1195 if (td->nr_open_files >= td->o.open_files)
1196 return ERR_PTR(-EBUSY);
1198 err = td_io_open_file(td, f);
1204 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1205 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1209 td_io_close_file(td, f);
1214 * Get next file to service by doing round robin between all available ones
1216 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1219 unsigned int old_next_file = td->next_file;
1225 f = td->files[td->next_file];
1228 if (td->next_file >= td->o.nr_files)
1231 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1232 if (fio_file_done(f)) {
1237 if (!fio_file_open(f)) {
1240 if (td->nr_open_files >= td->o.open_files)
1241 return ERR_PTR(-EBUSY);
1243 err = td_io_open_file(td, f);
1245 dprint(FD_FILE, "error %d on open of %s\n",
1253 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1255 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1259 td_io_close_file(td, f);
1262 } while (td->next_file != old_next_file);
1264 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1268 static struct fio_file *__get_next_file(struct thread_data *td)
1272 assert(td->o.nr_files <= td->files_index);
1274 if (td->nr_done_files >= td->o.nr_files) {
1275 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1276 " nr_files=%d\n", td->nr_open_files,
1282 f = td->file_service_file;
1283 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1284 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1286 if (td->file_service_left--)
1290 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1291 td->o.file_service_type == FIO_FSERVICE_SEQ)
1292 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1294 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1299 td->file_service_file = f;
1300 td->file_service_left = td->file_service_nr - 1;
1303 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1305 dprint(FD_FILE, "get_next_file: NULL\n");
1309 static struct fio_file *get_next_file(struct thread_data *td)
1311 if (td->flags & TD_F_PROFILE_OPS) {
1312 struct prof_io_ops *ops = &td->prof_io_ops;
1314 if (ops->get_next_file)
1315 return ops->get_next_file(td);
1318 return __get_next_file(td);
1321 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1326 f = get_next_file(td);
1327 if (IS_ERR_OR_NULL(f))
1333 if (!fill_io_u(td, io_u))
1336 put_file_log(td, f);
1337 td_io_close_file(td, f);
1339 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1340 fio_file_reset(td, f);
1342 fio_file_set_done(f);
1343 td->nr_done_files++;
1344 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1345 td->nr_done_files, td->o.nr_files);
1352 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1353 unsigned long tusec, unsigned long max_usec)
1356 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1357 td_verror(td, ETIMEDOUT, "max latency exceeded");
1358 icd->error = ETIMEDOUT;
1361 static void lat_new_cycle(struct thread_data *td)
1363 fio_gettime(&td->latency_ts, NULL);
1364 td->latency_ios = ddir_rw_sum(td->io_blocks);
1365 td->latency_failed = 0;
1369 * We had an IO outside the latency target. Reduce the queue depth. If we
1370 * are at QD=1, then it's time to give up.
1372 static bool __lat_target_failed(struct thread_data *td)
1374 if (td->latency_qd == 1)
1377 td->latency_qd_high = td->latency_qd;
1379 if (td->latency_qd == td->latency_qd_low)
1380 td->latency_qd_low--;
1382 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1384 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1387 * When we ramp QD down, quiesce existing IO to prevent
1388 * a storm of ramp downs due to pending higher depth.
1395 static bool lat_target_failed(struct thread_data *td)
1397 if (td->o.latency_percentile.u.f == 100.0)
1398 return __lat_target_failed(td);
1400 td->latency_failed++;
1404 void lat_target_init(struct thread_data *td)
1406 td->latency_end_run = 0;
1408 if (td->o.latency_target) {
1409 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1410 fio_gettime(&td->latency_ts, NULL);
1412 td->latency_qd_high = td->o.iodepth;
1413 td->latency_qd_low = 1;
1414 td->latency_ios = ddir_rw_sum(td->io_blocks);
1416 td->latency_qd = td->o.iodepth;
1419 void lat_target_reset(struct thread_data *td)
1421 if (!td->latency_end_run)
1422 lat_target_init(td);
1425 static void lat_target_success(struct thread_data *td)
1427 const unsigned int qd = td->latency_qd;
1428 struct thread_options *o = &td->o;
1430 td->latency_qd_low = td->latency_qd;
1433 * If we haven't failed yet, we double up to a failing value instead
1434 * of bisecting from highest possible queue depth. If we have set
1435 * a limit other than td->o.iodepth, bisect between that.
1437 if (td->latency_qd_high != o->iodepth)
1438 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1440 td->latency_qd *= 2;
1442 if (td->latency_qd > o->iodepth)
1443 td->latency_qd = o->iodepth;
1445 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1448 * Same as last one, we are done. Let it run a latency cycle, so
1449 * we get only the results from the targeted depth.
1451 if (td->latency_qd == qd) {
1452 if (td->latency_end_run) {
1453 dprint(FD_RATE, "We are done\n");
1456 dprint(FD_RATE, "Quiesce and final run\n");
1458 td->latency_end_run = 1;
1459 reset_all_stats(td);
1468 * Check if we can bump the queue depth
1470 void lat_target_check(struct thread_data *td)
1472 uint64_t usec_window;
1476 usec_window = utime_since_now(&td->latency_ts);
1477 if (usec_window < td->o.latency_window)
1480 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1481 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1482 success_ios *= 100.0;
1484 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1486 if (success_ios >= td->o.latency_percentile.u.f)
1487 lat_target_success(td);
1489 __lat_target_failed(td);
1493 * If latency target is enabled, we might be ramping up or down and not
1494 * using the full queue depth available.
1496 bool queue_full(const struct thread_data *td)
1498 const int qempty = io_u_qempty(&td->io_u_freelist);
1502 if (!td->o.latency_target)
1505 return td->cur_depth >= td->latency_qd;
1508 struct io_u *__get_io_u(struct thread_data *td)
1510 struct io_u *io_u = NULL;
1518 if (!io_u_rempty(&td->io_u_requeues))
1519 io_u = io_u_rpop(&td->io_u_requeues);
1520 else if (!queue_full(td)) {
1521 io_u = io_u_qpop(&td->io_u_freelist);
1526 io_u->end_io = NULL;
1530 assert(io_u->flags & IO_U_F_FREE);
1531 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1532 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1536 io_u->acct_ddir = -1;
1538 assert(!(td->flags & TD_F_CHILD));
1539 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1541 } else if (td_async_processing(td)) {
1543 * We ran out, wait for async verify threads to finish and
1546 assert(!(td->flags & TD_F_CHILD));
1547 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1555 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1557 if (!(td->flags & TD_F_TRIM_BACKLOG))
1560 if (td->trim_entries) {
1563 if (td->trim_batch) {
1566 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1567 td->last_ddir != DDIR_READ) {
1568 td->trim_batch = td->o.trim_batch;
1569 if (!td->trim_batch)
1570 td->trim_batch = td->o.trim_backlog;
1574 if (get_trim && get_next_trim(td, io_u))
1581 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1583 if (!(td->flags & TD_F_VER_BACKLOG))
1586 if (td->io_hist_len) {
1589 if (td->verify_batch)
1591 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1592 td->last_ddir != DDIR_READ) {
1593 td->verify_batch = td->o.verify_batch;
1594 if (!td->verify_batch)
1595 td->verify_batch = td->o.verify_backlog;
1599 if (get_verify && !get_next_verify(td, io_u)) {
1609 * Fill offset and start time into the buffer content, to prevent too
1610 * easy compressible data for simple de-dupe attempts. Do this for every
1611 * 512b block in the range, since that should be the smallest block size
1612 * we can expect from a device.
1614 static void small_content_scramble(struct io_u *io_u)
1616 unsigned int i, nr_blocks = io_u->buflen / 512;
1618 unsigned int offset;
1625 boffset = io_u->offset;
1626 io_u->buf_filled_len = 0;
1628 for (i = 0; i < nr_blocks; i++) {
1630 * Fill the byte offset into a "random" start offset of
1631 * the buffer, given by the product of the usec time
1632 * and the actual offset.
1634 offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511;
1635 offset &= ~(sizeof(uint64_t) - 1);
1636 if (offset >= 512 - sizeof(uint64_t))
1637 offset -= sizeof(uint64_t);
1638 memcpy(p + offset, &boffset, sizeof(boffset));
1640 end = p + 512 - sizeof(io_u->start_time);
1641 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1648 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1649 * etc. The returned io_u is fully ready to be prepped and submitted.
1651 struct io_u *get_io_u(struct thread_data *td)
1655 int do_scramble = 0;
1658 io_u = __get_io_u(td);
1660 dprint(FD_IO, "__get_io_u failed\n");
1664 if (check_get_verify(td, io_u))
1666 if (check_get_trim(td, io_u))
1670 * from a requeue, io_u already setup
1676 * If using an iolog, grab next piece if any available.
1678 if (td->flags & TD_F_READ_IOLOG) {
1679 if (read_iolog_get(td, io_u))
1681 } else if (set_io_u_file(td, io_u)) {
1683 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1689 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1693 assert(fio_file_open(f));
1695 if (ddir_rw(io_u->ddir)) {
1696 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1697 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1701 f->last_start[io_u->ddir] = io_u->offset;
1702 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1704 if (io_u->ddir == DDIR_WRITE) {
1705 if (td->flags & TD_F_REFILL_BUFFERS) {
1706 io_u_fill_buffer(td, io_u,
1707 td->o.min_bs[DDIR_WRITE],
1709 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1710 !(td->flags & TD_F_COMPRESS))
1712 if (td->flags & TD_F_VER_NONE) {
1713 populate_verify_io_u(td, io_u);
1716 } else if (io_u->ddir == DDIR_READ) {
1718 * Reset the buf_filled parameters so next time if the
1719 * buffer is used for writes it is refilled.
1721 io_u->buf_filled_len = 0;
1726 * Set io data pointers.
1728 io_u->xfer_buf = io_u->buf;
1729 io_u->xfer_buflen = io_u->buflen;
1733 if (!td_io_prep(td, io_u)) {
1734 if (!td->o.disable_lat)
1735 fio_gettime(&io_u->start_time, NULL);
1738 small_content_scramble(io_u);
1743 dprint(FD_IO, "get_io_u failed\n");
1745 return ERR_PTR(ret);
1748 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1750 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1752 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1755 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1756 io_u->file ? " on file " : "",
1757 io_u->file ? io_u->file->file_name : "",
1758 strerror(io_u->error),
1759 io_ddir_name(io_u->ddir),
1760 io_u->offset, io_u->xfer_buflen);
1762 if (td->io_ops->errdetails) {
1763 char *err = td->io_ops->errdetails(io_u);
1765 log_err("fio: %s\n", err);
1770 td_verror(td, io_u->error, "io_u error");
1773 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1775 __io_u_log_error(td, io_u);
1777 __io_u_log_error(td->parent, io_u);
1780 static inline bool gtod_reduce(struct thread_data *td)
1782 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1783 || td->o.gtod_reduce;
1786 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1787 struct io_completion_data *icd,
1788 const enum fio_ddir idx, unsigned int bytes)
1790 const int no_reduce = !gtod_reduce(td);
1791 unsigned long long llnsec = 0;
1796 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1800 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1802 if (!td->o.disable_lat) {
1803 unsigned long long tnsec;
1805 tnsec = ntime_since(&io_u->start_time, &icd->time);
1806 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1808 if (td->flags & TD_F_PROFILE_OPS) {
1809 struct prof_io_ops *ops = &td->prof_io_ops;
1812 icd->error = ops->io_u_lat(td, tnsec/1000);
1815 if (td->o.max_latency && tnsec/1000 > td->o.max_latency)
1816 lat_fatal(td, icd, tnsec/1000, td->o.max_latency);
1817 if (td->o.latency_target && tnsec/1000 > td->o.latency_target) {
1818 if (lat_target_failed(td))
1819 lat_fatal(td, icd, tnsec/1000, td->o.latency_target);
1824 if (!td->o.disable_clat) {
1825 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1826 io_u_mark_latency(td, llnsec);
1829 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1830 add_bw_sample(td, io_u, bytes, llnsec);
1832 if (no_reduce && per_unit_log(td->iops_log))
1833 add_iops_sample(td, io_u, bytes);
1836 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1837 uint32_t *info = io_u_block_info(td, io_u);
1838 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1839 if (io_u->ddir == DDIR_TRIM) {
1840 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1841 BLOCK_INFO_TRIMS(*info) + 1);
1842 } else if (io_u->ddir == DDIR_WRITE) {
1843 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1850 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1851 uint64_t offset, unsigned int bytes)
1858 if (f->first_write == -1ULL || offset < f->first_write)
1859 f->first_write = offset;
1860 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1861 f->last_write = offset + bytes;
1863 if (!f->last_write_comp)
1866 idx = f->last_write_idx++;
1867 f->last_write_comp[idx] = offset;
1868 if (f->last_write_idx == td->o.iodepth)
1869 f->last_write_idx = 0;
1872 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1873 struct io_completion_data *icd)
1875 struct io_u *io_u = *io_u_ptr;
1876 enum fio_ddir ddir = io_u->ddir;
1877 struct fio_file *f = io_u->file;
1879 dprint_io_u(io_u, "io complete");
1881 assert(io_u->flags & IO_U_F_FLIGHT);
1882 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1885 * Mark IO ok to verify
1889 * Remove errored entry from the verification list
1892 unlog_io_piece(td, io_u);
1894 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1899 if (ddir_sync(ddir)) {
1900 td->last_was_sync = 1;
1902 f->first_write = -1ULL;
1903 f->last_write = -1ULL;
1908 td->last_was_sync = 0;
1909 td->last_ddir = ddir;
1911 if (!io_u->error && ddir_rw(ddir)) {
1912 unsigned int bytes = io_u->buflen - io_u->resid;
1915 td->io_blocks[ddir]++;
1916 td->this_io_blocks[ddir]++;
1917 td->io_bytes[ddir] += bytes;
1919 if (!(io_u->flags & IO_U_F_VER_LIST))
1920 td->this_io_bytes[ddir] += bytes;
1922 if (ddir == DDIR_WRITE)
1923 file_log_write_comp(td, f, io_u->offset, bytes);
1925 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1926 td->runstate == TD_VERIFYING))
1927 account_io_completion(td, io_u, icd, ddir, bytes);
1929 icd->bytes_done[ddir] += bytes;
1932 ret = io_u->end_io(td, io_u_ptr);
1934 if (ret && !icd->error)
1937 } else if (io_u->error) {
1938 icd->error = io_u->error;
1939 io_u_log_error(td, io_u);
1942 enum error_type_bit eb = td_error_type(ddir, icd->error);
1944 if (!td_non_fatal_error(td, eb, icd->error))
1948 * If there is a non_fatal error, then add to the error count
1949 * and clear all the errors.
1951 update_error_count(td, icd->error);
1959 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1964 if (!gtod_reduce(td))
1965 fio_gettime(&icd->time, NULL);
1970 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1971 icd->bytes_done[ddir] = 0;
1974 static void ios_completed(struct thread_data *td,
1975 struct io_completion_data *icd)
1980 for (i = 0; i < icd->nr; i++) {
1981 io_u = td->io_ops->event(td, i);
1983 io_completed(td, &io_u, icd);
1991 * Complete a single io_u for the sync engines.
1993 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1995 struct io_completion_data icd;
1998 init_icd(td, &icd, 1);
1999 io_completed(td, &io_u, &icd);
2005 td_verror(td, icd.error, "io_u_sync_complete");
2009 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2010 td->bytes_done[ddir] += icd.bytes_done[ddir];
2016 * Called to complete min_events number of io for the async engines.
2018 int io_u_queued_complete(struct thread_data *td, int min_evts)
2020 struct io_completion_data icd;
2021 struct timespec *tvp = NULL;
2023 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2025 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2029 else if (min_evts > td->cur_depth)
2030 min_evts = td->cur_depth;
2032 /* No worries, td_io_getevents fixes min and max if they are
2033 * set incorrectly */
2034 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2036 td_verror(td, -ret, "td_io_getevents");
2041 init_icd(td, &icd, ret);
2042 ios_completed(td, &icd);
2044 td_verror(td, icd.error, "io_u_queued_complete");
2048 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2049 td->bytes_done[ddir] += icd.bytes_done[ddir];
2055 * Call when io_u is really queued, to update the submission latency.
2057 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2059 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2060 unsigned long slat_time;
2062 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2067 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2073 * See if we should reuse the last seed, if dedupe is enabled
2075 static struct frand_state *get_buf_state(struct thread_data *td)
2079 if (!td->o.dedupe_percentage)
2080 return &td->buf_state;
2081 else if (td->o.dedupe_percentage == 100) {
2082 frand_copy(&td->buf_state_prev, &td->buf_state);
2083 return &td->buf_state;
2086 v = rand32_between(&td->dedupe_state, 1, 100);
2088 if (v <= td->o.dedupe_percentage)
2089 return &td->buf_state_prev;
2091 return &td->buf_state;
2094 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2096 if (td->o.dedupe_percentage == 100)
2097 frand_copy(rs, &td->buf_state_prev);
2098 else if (rs == &td->buf_state)
2099 frand_copy(&td->buf_state_prev, rs);
2102 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2103 unsigned int max_bs)
2105 struct thread_options *o = &td->o;
2107 if (o->mem_type == MEM_CUDA_MALLOC)
2110 if (o->compress_percentage || o->dedupe_percentage) {
2111 unsigned int perc = td->o.compress_percentage;
2112 struct frand_state *rs;
2113 unsigned int left = max_bs;
2114 unsigned int this_write;
2117 rs = get_buf_state(td);
2119 min_write = min(min_write, left);
2122 this_write = min_not_zero(min_write,
2123 td->o.compress_chunk);
2125 fill_random_buf_percentage(rs, buf, perc,
2126 this_write, this_write,
2128 o->buffer_pattern_bytes);
2130 fill_random_buf(rs, buf, min_write);
2131 this_write = min_write;
2136 save_buf_state(td, rs);
2138 } else if (o->buffer_pattern_bytes)
2139 fill_buffer_pattern(td, buf, max_bs);
2140 else if (o->zero_buffers)
2141 memset(buf, 0, max_bs);
2143 fill_random_buf(get_buf_state(td), buf, max_bs);
2147 * "randomly" fill the buffer contents
2149 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2150 unsigned int min_write, unsigned int max_bs)
2152 io_u->buf_filled_len = 0;
2153 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2156 static int do_sync_file_range(const struct thread_data *td,
2159 off64_t offset, nbytes;
2161 offset = f->first_write;
2162 nbytes = f->last_write - f->first_write;
2167 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2170 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2174 if (io_u->ddir == DDIR_SYNC) {
2175 ret = fsync(io_u->file->fd);
2176 } else if (io_u->ddir == DDIR_DATASYNC) {
2177 #ifdef CONFIG_FDATASYNC
2178 ret = fdatasync(io_u->file->fd);
2180 ret = io_u->xfer_buflen;
2181 io_u->error = EINVAL;
2183 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2184 ret = do_sync_file_range(td, io_u->file);
2186 ret = io_u->xfer_buflen;
2187 io_u->error = EINVAL;
2191 io_u->error = errno;
2196 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2198 #ifndef FIO_HAVE_TRIM
2199 io_u->error = EINVAL;
2202 struct fio_file *f = io_u->file;
2205 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2207 return io_u->xfer_buflen;