14 struct io_completion_data {
17 int error; /* output */
18 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
19 struct timespec time; /* output */
23 * The ->io_axmap contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 static bool random_map_free(struct fio_file *f, const uint64_t block)
28 return !axmap_isset(f->io_axmap, block);
32 * Mark a given offset as used in the map.
34 static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
35 uint64_t offset, uint64_t buflen)
37 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
38 struct fio_file *f = io_u->file;
39 unsigned long long nr_blocks;
42 block = (offset - f->file_offset) / (uint64_t) min_bs;
43 nr_blocks = (buflen + min_bs - 1) / min_bs;
45 if (!(io_u->flags & IO_U_F_BUSY_OK))
46 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48 if ((nr_blocks * min_bs) < buflen)
49 buflen = nr_blocks * min_bs;
54 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
60 assert(ddir_rw(ddir));
63 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 * -> not for now since there is code assuming it could go either.
66 max_size = f->io_size;
67 if (max_size > f->real_file_size)
68 max_size = f->real_file_size;
70 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
71 max_size = td->o.zone_range;
73 if (td->o.min_bs[ddir] > td->o.ba[ddir])
74 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
83 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b,
89 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
90 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
92 r = __rand(&td->random_state);
94 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
96 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
100 assert(fio_file_lfsr(f));
102 if (lfsr_next(&f->lfsr, &off))
109 * if we are not maintaining a random map, we are done.
111 if (!file_randommap(td, f))
115 * calculate map offset and check if it's free
117 if (random_map_free(f, *b))
120 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
121 (unsigned long long) *b);
123 *b = axmap_next_free(f->io_axmap, *b);
124 if (*b == (uint64_t) -1ULL)
130 static int __get_next_rand_offset_zipf(struct thread_data *td,
131 struct fio_file *f, enum fio_ddir ddir,
134 *b = zipf_next(&f->zipf);
138 static int __get_next_rand_offset_pareto(struct thread_data *td,
139 struct fio_file *f, enum fio_ddir ddir,
142 *b = pareto_next(&f->zipf);
146 static int __get_next_rand_offset_gauss(struct thread_data *td,
147 struct fio_file *f, enum fio_ddir ddir,
150 *b = gauss_next(&f->gauss);
154 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
156 enum fio_ddir ddir, uint64_t *b)
158 struct zone_split_index *zsi;
159 uint64_t lastb, send, stotal;
162 lastb = last_block(td, f, ddir);
166 if (!td->o.zone_split_nr[ddir]) {
168 return __get_next_rand_offset(td, f, ddir, b, lastb);
172 * Generate a value, v, between 1 and 100, both inclusive
174 v = rand_between(&td->zone_state, 1, 100);
177 * Find our generated table. 'send' is the end block of this zone,
178 * 'stotal' is our start offset.
180 zsi = &td->zone_state_index[ddir][v - 1];
181 stotal = zsi->size_prev / td->o.ba[ddir];
182 send = zsi->size / td->o.ba[ddir];
185 * Should never happen
188 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
189 log_err("fio: bug in zoned generation\n");
191 } else if (send > lastb) {
193 * This happens if the user specifies ranges that exceed
194 * the file/device size. We can't handle that gracefully,
197 log_err("fio: zoned_abs sizes exceed file size\n");
202 * Generate index from 0..send-stotal
204 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
211 static int __get_next_rand_offset_zoned(struct thread_data *td,
212 struct fio_file *f, enum fio_ddir ddir,
215 unsigned int v, send, stotal;
216 uint64_t offset, lastb;
217 struct zone_split_index *zsi;
219 lastb = last_block(td, f, ddir);
223 if (!td->o.zone_split_nr[ddir]) {
225 return __get_next_rand_offset(td, f, ddir, b, lastb);
229 * Generate a value, v, between 1 and 100, both inclusive
231 v = rand_between(&td->zone_state, 1, 100);
233 zsi = &td->zone_state_index[ddir][v - 1];
234 stotal = zsi->size_perc_prev;
235 send = zsi->size_perc;
238 * Should never happen
241 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
242 log_err("fio: bug in zoned generation\n");
247 * 'send' is some percentage below or equal to 100 that
248 * marks the end of the current IO range. 'stotal' marks
249 * the start, in percent.
252 offset = stotal * lastb / 100ULL;
256 lastb = lastb * (send - stotal) / 100ULL;
259 * Generate index from 0..send-of-lastb
261 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
265 * Add our start offset, if any
273 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
274 enum fio_ddir ddir, uint64_t *b)
276 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
279 lastb = last_block(td, f, ddir);
283 return __get_next_rand_offset(td, f, ddir, b, lastb);
284 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
285 return __get_next_rand_offset_zipf(td, f, ddir, b);
286 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
287 return __get_next_rand_offset_pareto(td, f, ddir, b);
288 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
289 return __get_next_rand_offset_gauss(td, f, ddir, b);
290 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
291 return __get_next_rand_offset_zoned(td, f, ddir, b);
292 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
293 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
295 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
299 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
303 if (td->o.perc_rand[ddir] == 100)
306 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
308 return v <= td->o.perc_rand[ddir];
311 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
313 struct thread_options *o = &td->o;
315 if (o->invalidate_cache && !o->odirect) {
318 ret = file_invalidate_cache(td, f);
322 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
323 enum fio_ddir ddir, uint64_t *b)
325 if (!get_next_rand_offset(td, f, ddir, b))
328 if (td->o.time_based ||
329 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
330 fio_file_reset(td, f);
331 loop_cache_invalidate(td, f);
332 if (!get_next_rand_offset(td, f, ddir, b))
336 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
337 f->file_name, (unsigned long long) f->last_pos[ddir],
338 (unsigned long long) f->real_file_size);
342 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
343 enum fio_ddir ddir, uint64_t *offset)
345 struct thread_options *o = &td->o;
347 assert(ddir_rw(ddir));
350 * If we reach the end for a time based run, reset us back to 0
351 * and invalidate the cache, if we need to.
353 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
355 f->last_pos[ddir] = f->file_offset;
356 loop_cache_invalidate(td, f);
359 if (f->last_pos[ddir] < f->real_file_size) {
363 * Only rewind if we already hit the end
365 if (f->last_pos[ddir] == f->file_offset &&
366 f->file_offset && o->ddir_seq_add < 0) {
367 if (f->real_file_size > f->io_size)
368 f->last_pos[ddir] = f->io_size;
370 f->last_pos[ddir] = f->real_file_size;
373 pos = f->last_pos[ddir] - f->file_offset;
374 if (pos && o->ddir_seq_add) {
375 pos += o->ddir_seq_add;
378 * If we reach beyond the end of the file
379 * with holed IO, wrap around to the
380 * beginning again. If we're doing backwards IO,
383 if (pos >= f->real_file_size) {
384 if (o->ddir_seq_add > 0)
385 pos = f->file_offset;
387 if (f->real_file_size > f->io_size)
390 pos = f->real_file_size;
392 pos += o->ddir_seq_add;
404 static int get_next_block(struct thread_data *td, struct io_u *io_u,
405 enum fio_ddir ddir, int rw_seq,
408 struct fio_file *f = io_u->file;
412 assert(ddir_rw(ddir));
418 if (should_do_random(td, ddir)) {
419 ret = get_next_rand_block(td, f, ddir, &b);
423 io_u_set(td, io_u, IO_U_F_BUSY_OK);
424 ret = get_next_seq_offset(td, f, ddir, &offset);
426 ret = get_next_rand_block(td, f, ddir, &b);
430 ret = get_next_seq_offset(td, f, ddir, &offset);
433 io_u_set(td, io_u, IO_U_F_BUSY_OK);
436 if (td->o.rw_seq == RW_SEQ_SEQ) {
437 ret = get_next_seq_offset(td, f, ddir, &offset);
439 ret = get_next_rand_block(td, f, ddir, &b);
442 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
443 if (f->last_start[ddir] != -1ULL)
444 offset = f->last_start[ddir] - f->file_offset;
449 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
456 io_u->offset = offset;
458 io_u->offset = b * td->o.ba[ddir];
460 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
469 * For random io, generate a random new block and see if it's used. Repeat
470 * until we find a free one. For sequential io, just return the end of
471 * the last io issued.
473 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
476 struct fio_file *f = io_u->file;
477 enum fio_ddir ddir = io_u->ddir;
480 assert(ddir_rw(ddir));
482 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
484 td->ddir_seq_nr = td->o.ddir_seq_nr;
487 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
490 if (io_u->offset >= f->io_size) {
491 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
492 (unsigned long long) io_u->offset,
493 (unsigned long long) f->io_size);
497 io_u->offset += f->file_offset;
498 if (io_u->offset >= f->real_file_size) {
499 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
500 (unsigned long long) io_u->offset,
501 (unsigned long long) f->real_file_size);
508 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
509 unsigned long long buflen)
511 struct fio_file *f = io_u->file;
513 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
516 static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
519 int ddir = io_u->ddir;
520 unsigned long long buflen = 0;
521 unsigned long long minbs, maxbs;
522 uint64_t frand_max, r;
525 assert(ddir_rw(ddir));
527 if (td->o.bs_is_seq_rand)
528 ddir = is_random ? DDIR_WRITE : DDIR_READ;
530 minbs = td->o.min_bs[ddir];
531 maxbs = td->o.max_bs[ddir];
537 * If we can't satisfy the min block size from here, then fail
539 if (!io_u_fits(td, io_u, minbs))
542 frand_max = rand_max(&td->bsrange_state[ddir]);
544 r = __rand(&td->bsrange_state[ddir]);
546 if (!td->o.bssplit_nr[ddir]) {
547 buflen = minbs + (unsigned long long) ((double) maxbs *
548 (r / (frand_max + 1.0)));
553 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
554 struct bssplit *bsp = &td->o.bssplit[ddir][i];
560 if ((r / perc <= frand_max / 100ULL) &&
561 io_u_fits(td, io_u, buflen))
566 power_2 = is_power_of_2(minbs);
567 if (!td->o.bs_unaligned && power_2)
568 buflen &= ~(minbs - 1);
569 else if (!td->o.bs_unaligned && !power_2)
570 buflen -= buflen % minbs;
571 } while (!io_u_fits(td, io_u, buflen));
576 static void set_rwmix_bytes(struct thread_data *td)
581 * we do time or byte based switch. this is needed because
582 * buffered writes may issue a lot quicker than they complete,
583 * whereas reads do not.
585 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
586 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
589 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
593 v = rand_between(&td->rwmix_state, 1, 100);
595 if (v <= td->o.rwmix[DDIR_READ])
601 int io_u_quiesce(struct thread_data *td)
606 * We are going to sleep, ensure that we flush anything pending as
607 * not to skew our latency numbers.
609 * Changed to only monitor 'in flight' requests here instead of the
610 * td->cur_depth, b/c td->cur_depth does not accurately represent
611 * io's that have been actually submitted to an async engine,
612 * and cur_depth is meaningless for sync engines.
614 if (td->io_u_queued || td->cur_depth)
617 while (td->io_u_in_flight) {
620 ret = io_u_queued_complete(td, 1);
625 if (td->flags & TD_F_REGROW_LOGS)
631 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
633 enum fio_ddir odir = ddir ^ 1;
637 assert(ddir_rw(ddir));
638 now = utime_since_now(&td->start);
641 * if rate_next_io_time is in the past, need to catch up to rate
643 if (td->rate_next_io_time[ddir] <= now)
647 * We are ahead of rate in this direction. See if we
650 if (td_rw(td) && td->o.rwmix[odir]) {
652 * Other direction is behind rate, switch
654 if (td->rate_next_io_time[odir] <= now)
658 * Both directions are ahead of rate. sleep the min,
659 * switch if necessary
661 if (td->rate_next_io_time[ddir] <=
662 td->rate_next_io_time[odir]) {
663 usec = td->rate_next_io_time[ddir] - now;
665 usec = td->rate_next_io_time[odir] - now;
669 usec = td->rate_next_io_time[ddir] - now;
671 if (td->o.io_submit_mode == IO_MODE_INLINE)
674 usec_sleep(td, usec);
679 * Return the data direction for the next io_u. If the job is a
680 * mixed read/write workload, check the rwmix cycle and switch if
683 static enum fio_ddir get_rw_ddir(struct thread_data *td)
688 * See if it's time to fsync/fdatasync/sync_file_range first,
689 * and if not then move on to check regular I/Os.
691 if (should_fsync(td)) {
692 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
693 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
696 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
697 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
698 return DDIR_DATASYNC;
700 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
701 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
702 return DDIR_SYNC_FILE_RANGE;
707 * Check if it's time to seed a new data direction.
709 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
711 * Put a top limit on how many bytes we do for
712 * one data direction, to avoid overflowing the
715 ddir = get_rand_ddir(td);
717 if (ddir != td->rwmix_ddir)
720 td->rwmix_ddir = ddir;
722 ddir = td->rwmix_ddir;
723 } else if (td_read(td))
725 else if (td_write(td))
727 else if (td_trim(td))
732 td->rwmix_ddir = rate_ddir(td, ddir);
733 return td->rwmix_ddir;
736 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
738 enum fio_ddir ddir = get_rw_ddir(td);
740 if (td_trimwrite(td)) {
741 struct fio_file *f = io_u->file;
742 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
748 io_u->ddir = io_u->acct_ddir = ddir;
750 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
751 td->o.barrier_blocks &&
752 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
753 td->io_issues[DDIR_WRITE])
754 io_u_set(td, io_u, IO_U_F_BARRIER);
757 void put_file_log(struct thread_data *td, struct fio_file *f)
759 unsigned int ret = put_file(td, f);
762 td_verror(td, ret, "file close");
765 void put_io_u(struct thread_data *td, struct io_u *io_u)
767 if (io_u->post_submit) {
768 io_u->post_submit(io_u, io_u->error == 0);
769 io_u->post_submit = NULL;
777 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
778 put_file_log(td, io_u->file);
781 io_u_set(td, io_u, IO_U_F_FREE);
783 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
785 assert(!(td->flags & TD_F_CHILD));
787 io_u_qpush(&td->io_u_freelist, io_u);
788 td_io_u_free_notify(td);
792 void clear_io_u(struct thread_data *td, struct io_u *io_u)
794 io_u_clear(td, io_u, IO_U_F_FLIGHT);
798 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
800 struct io_u *__io_u = *io_u;
801 enum fio_ddir ddir = acct_ddir(__io_u);
803 dprint(FD_IO, "requeue %p\n", __io_u);
810 io_u_set(td, __io_u, IO_U_F_FREE);
811 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
812 td->io_issues[ddir]--;
814 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
815 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
817 assert(!(td->flags & TD_F_CHILD));
820 io_u_rpush(&td->io_u_requeues, __io_u);
821 td_io_u_free_notify(td);
826 static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
828 struct fio_file *f = io_u->file;
830 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
831 assert(td->o.zone_size);
832 assert(td->o.zone_range);
835 * See if it's time to switch to a new zone
837 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
839 f->file_offset += td->o.zone_range + td->o.zone_skip;
842 * Wrap from the beginning, if we exceed the file size
844 if (f->file_offset >= f->real_file_size)
845 f->file_offset = get_start_offset(td, f);
847 f->last_pos[io_u->ddir] = f->file_offset;
848 td->io_skip_bytes += td->o.zone_skip;
852 * If zone_size > zone_range, then maintain the same zone until
853 * zone_bytes >= zone_size.
855 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
856 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
857 f->file_offset, f->last_pos[io_u->ddir]);
858 f->last_pos[io_u->ddir] = f->file_offset;
862 * For random: if 'norandommap' is not set and zone_size > zone_range,
863 * map needs to be reset as it's done with zone_range everytime.
865 if ((td->zone_bytes % td->o.zone_range) == 0)
866 fio_file_reset(td, f);
869 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
873 if (td_ioengine_flagged(td, FIO_NOIO))
876 set_rw_ddir(td, io_u);
879 * fsync() or fdatasync() or trim etc, we are done
881 if (!ddir_rw(io_u->ddir))
884 if (td->o.zone_mode == ZONE_MODE_STRIDED)
885 setup_strided_zone_mode(td, io_u);
888 * No log, let the seq/rand engine retrieve the next buflen and
891 if (get_next_offset(td, io_u, &is_random)) {
892 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
896 io_u->buflen = get_next_buflen(td, io_u, is_random);
898 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
902 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
903 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
905 (unsigned long long) io_u->offset, io_u->buflen,
906 (unsigned long long) io_u->file->real_file_size);
911 * mark entry before potentially trimming io_u
913 if (td_random(td) && file_randommap(td, io_u->file))
914 io_u->buflen = mark_random_map(td, io_u, io_u->offset,
918 dprint_io_u(io_u, "fill");
919 td->zone_bytes += io_u->buflen;
923 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
952 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
954 __io_u_mark_map(td->ts.io_u_submit, nr);
955 td->ts.total_submit++;
958 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
960 __io_u_mark_map(td->ts.io_u_complete, nr);
961 td->ts.total_complete++;
964 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
968 switch (td->cur_depth) {
990 td->ts.io_u_map[idx] += nr;
993 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1030 assert(idx < FIO_IO_U_LAT_N_NR);
1031 td->ts.io_u_lat_n[idx]++;
1034 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1038 assert(usec < 1000 && usec >= 1);
1071 assert(idx < FIO_IO_U_LAT_U_NR);
1072 td->ts.io_u_lat_u[idx]++;
1075 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1118 assert(idx < FIO_IO_U_LAT_M_NR);
1119 td->ts.io_u_lat_m[idx]++;
1122 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1125 io_u_mark_lat_nsec(td, nsec);
1126 else if (nsec < 1000000)
1127 io_u_mark_lat_usec(td, nsec / 1000);
1129 io_u_mark_lat_msec(td, nsec / 1000000);
1132 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1134 unsigned long fileno;
1136 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1137 uint64_t frand_max = rand_max(&td->next_file_state);
1140 r = __rand(&td->next_file_state);
1141 return (unsigned int) ((double) td->o.nr_files
1142 * (r / (frand_max + 1.0)));
1145 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1146 fileno = zipf_next(&td->next_file_zipf);
1147 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1148 fileno = pareto_next(&td->next_file_zipf);
1149 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1150 fileno = gauss_next(&td->next_file_gauss);
1152 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1157 return fileno >> FIO_FSERVICE_SHIFT;
1161 * Get next file to service by choosing one at random
1163 static struct fio_file *get_next_file_rand(struct thread_data *td,
1164 enum fio_file_flags goodf,
1165 enum fio_file_flags badf)
1173 fno = __get_next_fileno_rand(td);
1176 if (fio_file_done(f))
1179 if (!fio_file_open(f)) {
1182 if (td->nr_open_files >= td->o.open_files)
1183 return ERR_PTR(-EBUSY);
1185 err = td_io_open_file(td, f);
1191 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1192 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1196 td_io_close_file(td, f);
1201 * Get next file to service by doing round robin between all available ones
1203 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1206 unsigned int old_next_file = td->next_file;
1212 f = td->files[td->next_file];
1215 if (td->next_file >= td->o.nr_files)
1218 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1219 if (fio_file_done(f)) {
1224 if (!fio_file_open(f)) {
1227 if (td->nr_open_files >= td->o.open_files)
1228 return ERR_PTR(-EBUSY);
1230 err = td_io_open_file(td, f);
1232 dprint(FD_FILE, "error %d on open of %s\n",
1240 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1242 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1246 td_io_close_file(td, f);
1249 } while (td->next_file != old_next_file);
1251 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1255 static struct fio_file *__get_next_file(struct thread_data *td)
1259 assert(td->o.nr_files <= td->files_index);
1261 if (td->nr_done_files >= td->o.nr_files) {
1262 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1263 " nr_files=%d\n", td->nr_open_files,
1269 f = td->file_service_file;
1270 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1271 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1273 if (td->file_service_left--)
1277 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1278 td->o.file_service_type == FIO_FSERVICE_SEQ)
1279 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1281 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1286 td->file_service_file = f;
1287 td->file_service_left = td->file_service_nr - 1;
1290 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1292 dprint(FD_FILE, "get_next_file: NULL\n");
1296 static struct fio_file *get_next_file(struct thread_data *td)
1298 return __get_next_file(td);
1301 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1306 f = get_next_file(td);
1307 if (IS_ERR_OR_NULL(f))
1313 if (!fill_io_u(td, io_u))
1316 if (io_u->post_submit) {
1317 io_u->post_submit(io_u, false);
1318 io_u->post_submit = NULL;
1321 put_file_log(td, f);
1322 td_io_close_file(td, f);
1324 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1325 fio_file_reset(td, f);
1327 fio_file_set_done(f);
1328 td->nr_done_files++;
1329 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1330 td->nr_done_files, td->o.nr_files);
1337 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1338 unsigned long long tnsec, unsigned long long max_nsec)
1341 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1342 td_verror(td, ETIMEDOUT, "max latency exceeded");
1343 icd->error = ETIMEDOUT;
1346 static void lat_new_cycle(struct thread_data *td)
1348 fio_gettime(&td->latency_ts, NULL);
1349 td->latency_ios = ddir_rw_sum(td->io_blocks);
1350 td->latency_failed = 0;
1354 * We had an IO outside the latency target. Reduce the queue depth. If we
1355 * are at QD=1, then it's time to give up.
1357 static bool __lat_target_failed(struct thread_data *td)
1359 if (td->latency_qd == 1)
1362 td->latency_qd_high = td->latency_qd;
1364 if (td->latency_qd == td->latency_qd_low)
1365 td->latency_qd_low--;
1367 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1369 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1372 * When we ramp QD down, quiesce existing IO to prevent
1373 * a storm of ramp downs due to pending higher depth.
1380 static bool lat_target_failed(struct thread_data *td)
1382 if (td->o.latency_percentile.u.f == 100.0)
1383 return __lat_target_failed(td);
1385 td->latency_failed++;
1389 void lat_target_init(struct thread_data *td)
1391 td->latency_end_run = 0;
1393 if (td->o.latency_target) {
1394 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1395 fio_gettime(&td->latency_ts, NULL);
1397 td->latency_qd_high = td->o.iodepth;
1398 td->latency_qd_low = 1;
1399 td->latency_ios = ddir_rw_sum(td->io_blocks);
1401 td->latency_qd = td->o.iodepth;
1404 void lat_target_reset(struct thread_data *td)
1406 if (!td->latency_end_run)
1407 lat_target_init(td);
1410 static void lat_target_success(struct thread_data *td)
1412 const unsigned int qd = td->latency_qd;
1413 struct thread_options *o = &td->o;
1415 td->latency_qd_low = td->latency_qd;
1418 * If we haven't failed yet, we double up to a failing value instead
1419 * of bisecting from highest possible queue depth. If we have set
1420 * a limit other than td->o.iodepth, bisect between that.
1422 if (td->latency_qd_high != o->iodepth)
1423 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1425 td->latency_qd *= 2;
1427 if (td->latency_qd > o->iodepth)
1428 td->latency_qd = o->iodepth;
1430 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1433 * Same as last one, we are done. Let it run a latency cycle, so
1434 * we get only the results from the targeted depth.
1436 if (td->latency_qd == qd) {
1437 if (td->latency_end_run) {
1438 dprint(FD_RATE, "We are done\n");
1441 dprint(FD_RATE, "Quiesce and final run\n");
1443 td->latency_end_run = 1;
1444 reset_all_stats(td);
1453 * Check if we can bump the queue depth
1455 void lat_target_check(struct thread_data *td)
1457 uint64_t usec_window;
1461 usec_window = utime_since_now(&td->latency_ts);
1462 if (usec_window < td->o.latency_window)
1465 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1466 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1467 success_ios *= 100.0;
1469 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1471 if (success_ios >= td->o.latency_percentile.u.f)
1472 lat_target_success(td);
1474 __lat_target_failed(td);
1478 * If latency target is enabled, we might be ramping up or down and not
1479 * using the full queue depth available.
1481 bool queue_full(const struct thread_data *td)
1483 const int qempty = io_u_qempty(&td->io_u_freelist);
1487 if (!td->o.latency_target)
1490 return td->cur_depth >= td->latency_qd;
1493 struct io_u *__get_io_u(struct thread_data *td)
1495 struct io_u *io_u = NULL;
1504 if (!io_u_rempty(&td->io_u_requeues))
1505 io_u = io_u_rpop(&td->io_u_requeues);
1506 else if (!queue_full(td)) {
1507 io_u = io_u_qpop(&td->io_u_freelist);
1512 io_u->end_io = NULL;
1516 assert(io_u->flags & IO_U_F_FREE);
1517 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1518 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1522 io_u->acct_ddir = -1;
1524 assert(!(td->flags & TD_F_CHILD));
1525 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1527 } else if (td_async_processing(td)) {
1529 * We ran out, wait for async verify threads to finish and
1532 assert(!(td->flags & TD_F_CHILD));
1533 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1542 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1544 if (!(td->flags & TD_F_TRIM_BACKLOG))
1546 if (!td->trim_entries)
1549 if (td->trim_batch) {
1551 if (get_next_trim(td, io_u))
1553 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1554 td->last_ddir != DDIR_READ) {
1555 td->trim_batch = td->o.trim_batch;
1556 if (!td->trim_batch)
1557 td->trim_batch = td->o.trim_backlog;
1558 if (get_next_trim(td, io_u))
1565 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1567 if (!(td->flags & TD_F_VER_BACKLOG))
1570 if (td->io_hist_len) {
1573 if (td->verify_batch)
1575 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1576 td->last_ddir != DDIR_READ) {
1577 td->verify_batch = td->o.verify_batch;
1578 if (!td->verify_batch)
1579 td->verify_batch = td->o.verify_backlog;
1583 if (get_verify && !get_next_verify(td, io_u)) {
1593 * Fill offset and start time into the buffer content, to prevent too
1594 * easy compressible data for simple de-dupe attempts. Do this for every
1595 * 512b block in the range, since that should be the smallest block size
1596 * we can expect from a device.
1598 static void small_content_scramble(struct io_u *io_u)
1600 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1601 unsigned int offset;
1602 uint64_t boffset, *iptr;
1609 boffset = io_u->offset;
1611 if (io_u->buf_filled_len)
1612 io_u->buf_filled_len = 0;
1615 * Generate random index between 0..7. We do chunks of 512b, if
1616 * we assume a cacheline is 64 bytes, then we have 8 of those.
1617 * Scramble content within the blocks in the same cacheline to
1620 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1622 for (i = 0; i < nr_blocks; i++) {
1624 * Fill offset into start of cacheline, time into end
1627 iptr = (void *) p + (offset << 6);
1630 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1631 iptr[0] = io_u->start_time.tv_sec;
1632 iptr[1] = io_u->start_time.tv_nsec;
1640 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1641 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1643 struct io_u *get_io_u(struct thread_data *td)
1647 int do_scramble = 0;
1650 io_u = __get_io_u(td);
1652 dprint(FD_IO, "__get_io_u failed\n");
1656 if (check_get_verify(td, io_u))
1658 if (check_get_trim(td, io_u))
1662 * from a requeue, io_u already setup
1668 * If using an iolog, grab next piece if any available.
1670 if (td->flags & TD_F_READ_IOLOG) {
1671 if (read_iolog_get(td, io_u))
1673 } else if (set_io_u_file(td, io_u)) {
1675 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1681 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1685 assert(fio_file_open(f));
1687 if (ddir_rw(io_u->ddir)) {
1688 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1689 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1693 f->last_start[io_u->ddir] = io_u->offset;
1694 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1696 if (io_u->ddir == DDIR_WRITE) {
1697 if (td->flags & TD_F_REFILL_BUFFERS) {
1698 io_u_fill_buffer(td, io_u,
1699 td->o.min_bs[DDIR_WRITE],
1701 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1702 !(td->flags & TD_F_COMPRESS) &&
1703 !(td->flags & TD_F_DO_VERIFY))
1705 } else if (io_u->ddir == DDIR_READ) {
1707 * Reset the buf_filled parameters so next time if the
1708 * buffer is used for writes it is refilled.
1710 io_u->buf_filled_len = 0;
1715 * Set io data pointers.
1717 io_u->xfer_buf = io_u->buf;
1718 io_u->xfer_buflen = io_u->buflen;
1722 if (!td_io_prep(td, io_u)) {
1723 if (!td->o.disable_lat)
1724 fio_gettime(&io_u->start_time, NULL);
1727 small_content_scramble(io_u);
1732 dprint(FD_IO, "get_io_u failed\n");
1734 return ERR_PTR(ret);
1737 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1739 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1741 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1744 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1745 io_u->file ? " on file " : "",
1746 io_u->file ? io_u->file->file_name : "",
1747 strerror(io_u->error),
1748 io_ddir_name(io_u->ddir),
1749 io_u->offset, io_u->xfer_buflen);
1751 if (td->io_ops->errdetails) {
1752 char *err = td->io_ops->errdetails(io_u);
1754 log_err("fio: %s\n", err);
1759 td_verror(td, io_u->error, "io_u error");
1762 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1764 __io_u_log_error(td, io_u);
1766 __io_u_log_error(td->parent, io_u);
1769 static inline bool gtod_reduce(struct thread_data *td)
1771 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1772 || td->o.gtod_reduce;
1775 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1776 struct io_completion_data *icd,
1777 const enum fio_ddir idx, unsigned int bytes)
1779 const int no_reduce = !gtod_reduce(td);
1780 unsigned long long llnsec = 0;
1785 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1789 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1791 if (!td->o.disable_lat) {
1792 unsigned long long tnsec;
1794 tnsec = ntime_since(&io_u->start_time, &icd->time);
1795 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1797 if (td->flags & TD_F_PROFILE_OPS) {
1798 struct prof_io_ops *ops = &td->prof_io_ops;
1801 icd->error = ops->io_u_lat(td, tnsec);
1804 if (td->o.max_latency && tnsec > td->o.max_latency)
1805 lat_fatal(td, icd, tnsec, td->o.max_latency);
1806 if (td->o.latency_target && tnsec > td->o.latency_target) {
1807 if (lat_target_failed(td))
1808 lat_fatal(td, icd, tnsec, td->o.latency_target);
1813 if (!td->o.disable_clat) {
1814 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1815 io_u_mark_latency(td, llnsec);
1818 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1819 add_bw_sample(td, io_u, bytes, llnsec);
1821 if (no_reduce && per_unit_log(td->iops_log))
1822 add_iops_sample(td, io_u, bytes);
1823 } else if (ddir_sync(idx) && !td->o.disable_clat)
1824 add_sync_clat_sample(&td->ts, llnsec);
1826 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1827 uint32_t *info = io_u_block_info(td, io_u);
1828 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1829 if (io_u->ddir == DDIR_TRIM) {
1830 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1831 BLOCK_INFO_TRIMS(*info) + 1);
1832 } else if (io_u->ddir == DDIR_WRITE) {
1833 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1840 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1841 uint64_t offset, unsigned int bytes)
1848 if (f->first_write == -1ULL || offset < f->first_write)
1849 f->first_write = offset;
1850 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1851 f->last_write = offset + bytes;
1853 if (!f->last_write_comp)
1856 idx = f->last_write_idx++;
1857 f->last_write_comp[idx] = offset;
1858 if (f->last_write_idx == td->o.iodepth)
1859 f->last_write_idx = 0;
1862 static bool should_account(struct thread_data *td)
1864 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1865 td->runstate == TD_VERIFYING);
1868 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1869 struct io_completion_data *icd)
1871 struct io_u *io_u = *io_u_ptr;
1872 enum fio_ddir ddir = io_u->ddir;
1873 struct fio_file *f = io_u->file;
1875 dprint_io_u(io_u, "complete");
1877 assert(io_u->flags & IO_U_F_FLIGHT);
1878 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1881 * Mark IO ok to verify
1885 * Remove errored entry from the verification list
1888 unlog_io_piece(td, io_u);
1890 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1895 if (ddir_sync(ddir)) {
1896 td->last_was_sync = true;
1898 f->first_write = -1ULL;
1899 f->last_write = -1ULL;
1901 if (should_account(td))
1902 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1906 td->last_was_sync = false;
1907 td->last_ddir = ddir;
1909 if (!io_u->error && ddir_rw(ddir)) {
1910 unsigned long long bytes = io_u->buflen - io_u->resid;
1913 td->io_blocks[ddir]++;
1914 td->io_bytes[ddir] += bytes;
1916 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1917 td->this_io_blocks[ddir]++;
1918 td->this_io_bytes[ddir] += bytes;
1921 if (ddir == DDIR_WRITE)
1922 file_log_write_comp(td, f, io_u->offset, bytes);
1924 if (should_account(td))
1925 account_io_completion(td, io_u, icd, ddir, bytes);
1927 icd->bytes_done[ddir] += bytes;
1930 ret = io_u->end_io(td, io_u_ptr);
1932 if (ret && !icd->error)
1935 } else if (io_u->error) {
1936 icd->error = io_u->error;
1937 io_u_log_error(td, io_u);
1940 enum error_type_bit eb = td_error_type(ddir, icd->error);
1942 if (!td_non_fatal_error(td, eb, icd->error))
1946 * If there is a non_fatal error, then add to the error count
1947 * and clear all the errors.
1949 update_error_count(td, icd->error);
1957 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1962 if (!gtod_reduce(td))
1963 fio_gettime(&icd->time, NULL);
1968 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1969 icd->bytes_done[ddir] = 0;
1972 static void ios_completed(struct thread_data *td,
1973 struct io_completion_data *icd)
1978 for (i = 0; i < icd->nr; i++) {
1979 io_u = td->io_ops->event(td, i);
1981 io_completed(td, &io_u, icd);
1989 * Complete a single io_u for the sync engines.
1991 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1993 struct io_completion_data icd;
1996 init_icd(td, &icd, 1);
1997 io_completed(td, &io_u, &icd);
2003 td_verror(td, icd.error, "io_u_sync_complete");
2007 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2008 td->bytes_done[ddir] += icd.bytes_done[ddir];
2014 * Called to complete min_events number of io for the async engines.
2016 int io_u_queued_complete(struct thread_data *td, int min_evts)
2018 struct io_completion_data icd;
2019 struct timespec *tvp = NULL;
2021 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2023 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2027 else if (min_evts > td->cur_depth)
2028 min_evts = td->cur_depth;
2030 /* No worries, td_io_getevents fixes min and max if they are
2031 * set incorrectly */
2032 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2034 td_verror(td, -ret, "td_io_getevents");
2039 init_icd(td, &icd, ret);
2040 ios_completed(td, &icd);
2042 td_verror(td, icd.error, "io_u_queued_complete");
2046 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2047 td->bytes_done[ddir] += icd.bytes_done[ddir];
2053 * Call when io_u is really queued, to update the submission latency.
2055 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2057 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2058 unsigned long slat_time;
2060 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2065 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2071 * See if we should reuse the last seed, if dedupe is enabled
2073 static struct frand_state *get_buf_state(struct thread_data *td)
2077 if (!td->o.dedupe_percentage)
2078 return &td->buf_state;
2079 else if (td->o.dedupe_percentage == 100) {
2080 frand_copy(&td->buf_state_prev, &td->buf_state);
2081 return &td->buf_state;
2084 v = rand_between(&td->dedupe_state, 1, 100);
2086 if (v <= td->o.dedupe_percentage)
2087 return &td->buf_state_prev;
2089 return &td->buf_state;
2092 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2094 if (td->o.dedupe_percentage == 100)
2095 frand_copy(rs, &td->buf_state_prev);
2096 else if (rs == &td->buf_state)
2097 frand_copy(&td->buf_state_prev, rs);
2100 void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2101 unsigned long long max_bs)
2103 struct thread_options *o = &td->o;
2105 if (o->mem_type == MEM_CUDA_MALLOC)
2108 if (o->compress_percentage || o->dedupe_percentage) {
2109 unsigned int perc = td->o.compress_percentage;
2110 struct frand_state *rs;
2111 unsigned long long left = max_bs;
2112 unsigned long long this_write;
2115 rs = get_buf_state(td);
2117 min_write = min(min_write, left);
2120 this_write = min_not_zero(min_write,
2121 (unsigned long long) td->o.compress_chunk);
2123 fill_random_buf_percentage(rs, buf, perc,
2124 this_write, this_write,
2126 o->buffer_pattern_bytes);
2128 fill_random_buf(rs, buf, min_write);
2129 this_write = min_write;
2134 save_buf_state(td, rs);
2136 } else if (o->buffer_pattern_bytes)
2137 fill_buffer_pattern(td, buf, max_bs);
2138 else if (o->zero_buffers)
2139 memset(buf, 0, max_bs);
2141 fill_random_buf(get_buf_state(td), buf, max_bs);
2145 * "randomly" fill the buffer contents
2147 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2148 unsigned long long min_write, unsigned long long max_bs)
2150 io_u->buf_filled_len = 0;
2151 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2154 static int do_sync_file_range(const struct thread_data *td,
2157 off64_t offset, nbytes;
2159 offset = f->first_write;
2160 nbytes = f->last_write - f->first_write;
2165 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2168 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2172 if (io_u->ddir == DDIR_SYNC) {
2173 ret = fsync(io_u->file->fd);
2174 } else if (io_u->ddir == DDIR_DATASYNC) {
2175 #ifdef CONFIG_FDATASYNC
2176 ret = fdatasync(io_u->file->fd);
2178 ret = io_u->xfer_buflen;
2179 io_u->error = EINVAL;
2181 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2182 ret = do_sync_file_range(td, io_u->file);
2184 ret = io_u->xfer_buflen;
2185 io_u->error = EINVAL;
2189 io_u->error = errno;
2194 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2196 #ifndef FIO_HAVE_TRIM
2197 io_u->error = EINVAL;
2200 struct fio_file *f = io_u->file;
2203 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2205 return io_u->xfer_buflen;