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 void mark_random_map(struct thread_data *td, struct io_u *io_u)
36 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
37 struct fio_file *f = io_u->file;
38 unsigned long long nr_blocks;
41 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
42 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44 if (!(io_u->flags & IO_U_F_BUSY_OK))
45 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47 if ((nr_blocks * min_bs) < io_u->buflen)
48 io_u->buflen = nr_blocks * min_bs;
51 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
57 assert(ddir_rw(ddir));
60 * Hmm, should we make sure that ->io_size <= ->real_file_size?
61 * -> not for now since there is code assuming it could go either.
63 max_size = f->io_size;
64 if (max_size > f->real_file_size)
65 max_size = f->real_file_size;
67 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
68 max_size = td->o.zone_range;
70 if (td->o.min_bs[ddir] > td->o.ba[ddir])
71 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
73 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
80 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
81 enum fio_ddir ddir, uint64_t *b,
86 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
87 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
89 r = __rand(&td->random_state);
91 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
93 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
97 assert(fio_file_lfsr(f));
99 if (lfsr_next(&f->lfsr, &off))
106 * if we are not maintaining a random map, we are done.
108 if (!file_randommap(td, f))
112 * calculate map offset and check if it's free
114 if (random_map_free(f, *b))
117 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
118 (unsigned long long) *b);
120 *b = axmap_next_free(f->io_axmap, *b);
121 if (*b == (uint64_t) -1ULL)
127 static int __get_next_rand_offset_zipf(struct thread_data *td,
128 struct fio_file *f, enum fio_ddir ddir,
131 *b = zipf_next(&f->zipf);
135 static int __get_next_rand_offset_pareto(struct thread_data *td,
136 struct fio_file *f, enum fio_ddir ddir,
139 *b = pareto_next(&f->zipf);
143 static int __get_next_rand_offset_gauss(struct thread_data *td,
144 struct fio_file *f, enum fio_ddir ddir,
147 *b = gauss_next(&f->gauss);
151 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
153 enum fio_ddir ddir, uint64_t *b)
155 struct zone_split_index *zsi;
156 uint64_t lastb, send, stotal;
159 lastb = last_block(td, f, ddir);
163 if (!td->o.zone_split_nr[ddir]) {
165 return __get_next_rand_offset(td, f, ddir, b, lastb);
169 * Generate a value, v, between 1 and 100, both inclusive
171 v = rand_between(&td->zone_state, 1, 100);
174 * Find our generated table. 'send' is the end block of this zone,
175 * 'stotal' is our start offset.
177 zsi = &td->zone_state_index[ddir][v - 1];
178 stotal = zsi->size_prev / td->o.ba[ddir];
179 send = zsi->size / td->o.ba[ddir];
182 * Should never happen
185 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
186 log_err("fio: bug in zoned generation\n");
188 } else if (send > lastb) {
190 * This happens if the user specifies ranges that exceed
191 * the file/device size. We can't handle that gracefully,
194 log_err("fio: zoned_abs sizes exceed file size\n");
199 * Generate index from 0..send-stotal
201 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
208 static int __get_next_rand_offset_zoned(struct thread_data *td,
209 struct fio_file *f, enum fio_ddir ddir,
212 unsigned int v, send, stotal;
213 uint64_t offset, lastb;
214 struct zone_split_index *zsi;
216 lastb = last_block(td, f, ddir);
220 if (!td->o.zone_split_nr[ddir]) {
222 return __get_next_rand_offset(td, f, ddir, b, lastb);
226 * Generate a value, v, between 1 and 100, both inclusive
228 v = rand_between(&td->zone_state, 1, 100);
230 zsi = &td->zone_state_index[ddir][v - 1];
231 stotal = zsi->size_perc_prev;
232 send = zsi->size_perc;
235 * Should never happen
238 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
239 log_err("fio: bug in zoned generation\n");
244 * 'send' is some percentage below or equal to 100 that
245 * marks the end of the current IO range. 'stotal' marks
246 * the start, in percent.
249 offset = stotal * lastb / 100ULL;
253 lastb = lastb * (send - stotal) / 100ULL;
256 * Generate index from 0..send-of-lastb
258 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
262 * Add our start offset, if any
270 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
271 enum fio_ddir ddir, uint64_t *b)
273 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
276 lastb = last_block(td, f, ddir);
280 return __get_next_rand_offset(td, f, ddir, b, lastb);
281 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
282 return __get_next_rand_offset_zipf(td, f, ddir, b);
283 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
284 return __get_next_rand_offset_pareto(td, f, ddir, b);
285 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
286 return __get_next_rand_offset_gauss(td, f, ddir, b);
287 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
288 return __get_next_rand_offset_zoned(td, f, ddir, b);
289 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
290 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
292 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
296 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
300 if (td->o.perc_rand[ddir] == 100)
303 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
305 return v <= td->o.perc_rand[ddir];
308 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
310 struct thread_options *o = &td->o;
312 if (o->invalidate_cache && !o->odirect) {
315 ret = file_invalidate_cache(td, f);
319 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
320 enum fio_ddir ddir, uint64_t *b)
322 if (!get_next_rand_offset(td, f, ddir, b))
325 if (td->o.time_based ||
326 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
327 fio_file_reset(td, f);
328 loop_cache_invalidate(td, f);
329 if (!get_next_rand_offset(td, f, ddir, b))
333 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
334 f->file_name, (unsigned long long) f->last_pos[ddir],
335 (unsigned long long) f->real_file_size);
339 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
340 enum fio_ddir ddir, uint64_t *offset)
342 struct thread_options *o = &td->o;
344 assert(ddir_rw(ddir));
347 * If we reach the end for a time based run, reset us back to 0
348 * and invalidate the cache, if we need to.
350 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
352 f->last_pos[ddir] = f->file_offset;
353 loop_cache_invalidate(td, f);
356 if (f->last_pos[ddir] < f->real_file_size) {
360 * Only rewind if we already hit the end
362 if (f->last_pos[ddir] == f->file_offset &&
363 f->file_offset && o->ddir_seq_add < 0) {
364 if (f->real_file_size > f->io_size)
365 f->last_pos[ddir] = f->io_size;
367 f->last_pos[ddir] = f->real_file_size;
370 pos = f->last_pos[ddir] - f->file_offset;
371 if (pos && o->ddir_seq_add) {
372 pos += o->ddir_seq_add;
375 * If we reach beyond the end of the file
376 * with holed IO, wrap around to the
377 * beginning again. If we're doing backwards IO,
380 if (pos >= f->real_file_size) {
381 if (o->ddir_seq_add > 0)
382 pos = f->file_offset;
384 if (f->real_file_size > f->io_size)
387 pos = f->real_file_size;
389 pos += o->ddir_seq_add;
401 static int get_next_block(struct thread_data *td, struct io_u *io_u,
402 enum fio_ddir ddir, int rw_seq,
405 struct fio_file *f = io_u->file;
409 assert(ddir_rw(ddir));
415 if (should_do_random(td, ddir)) {
416 ret = get_next_rand_block(td, f, ddir, &b);
420 io_u_set(td, io_u, IO_U_F_BUSY_OK);
421 ret = get_next_seq_offset(td, f, ddir, &offset);
423 ret = get_next_rand_block(td, f, ddir, &b);
427 ret = get_next_seq_offset(td, f, ddir, &offset);
430 io_u_set(td, io_u, IO_U_F_BUSY_OK);
433 if (td->o.rw_seq == RW_SEQ_SEQ) {
434 ret = get_next_seq_offset(td, f, ddir, &offset);
436 ret = get_next_rand_block(td, f, ddir, &b);
439 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
440 if (f->last_start[ddir] != -1ULL)
441 offset = f->last_start[ddir] - f->file_offset;
446 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
453 io_u->offset = offset;
455 io_u->offset = b * td->o.ba[ddir];
457 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
466 * For random io, generate a random new block and see if it's used. Repeat
467 * until we find a free one. For sequential io, just return the end of
468 * the last io issued.
470 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
473 struct fio_file *f = io_u->file;
474 enum fio_ddir ddir = io_u->ddir;
477 assert(ddir_rw(ddir));
479 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
481 td->ddir_seq_nr = td->o.ddir_seq_nr;
484 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
487 if (io_u->offset >= f->io_size) {
488 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
489 (unsigned long long) io_u->offset,
490 (unsigned long long) f->io_size);
494 io_u->offset += f->file_offset;
495 if (io_u->offset >= f->real_file_size) {
496 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
497 (unsigned long long) io_u->offset,
498 (unsigned long long) f->real_file_size);
505 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
506 unsigned long long buflen)
508 struct fio_file *f = io_u->file;
510 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
513 static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
516 int ddir = io_u->ddir;
517 unsigned long long buflen = 0;
518 unsigned long long minbs, maxbs;
519 uint64_t frand_max, r;
522 assert(ddir_rw(ddir));
524 if (td->o.bs_is_seq_rand)
525 ddir = is_random ? DDIR_WRITE : DDIR_READ;
527 minbs = td->o.min_bs[ddir];
528 maxbs = td->o.max_bs[ddir];
534 * If we can't satisfy the min block size from here, then fail
536 if (!io_u_fits(td, io_u, minbs))
539 frand_max = rand_max(&td->bsrange_state[ddir]);
541 r = __rand(&td->bsrange_state[ddir]);
543 if (!td->o.bssplit_nr[ddir]) {
544 buflen = minbs + (unsigned long long) ((double) maxbs *
545 (r / (frand_max + 1.0)));
550 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
551 struct bssplit *bsp = &td->o.bssplit[ddir][i];
557 if ((r / perc <= frand_max / 100ULL) &&
558 io_u_fits(td, io_u, buflen))
563 power_2 = is_power_of_2(minbs);
564 if (!td->o.bs_unaligned && power_2)
565 buflen &= ~(minbs - 1);
566 else if (!td->o.bs_unaligned && !power_2)
567 buflen -= buflen % minbs;
568 } while (!io_u_fits(td, io_u, buflen));
573 static void set_rwmix_bytes(struct thread_data *td)
578 * we do time or byte based switch. this is needed because
579 * buffered writes may issue a lot quicker than they complete,
580 * whereas reads do not.
582 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
583 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
586 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
590 v = rand_between(&td->rwmix_state, 1, 100);
592 if (v <= td->o.rwmix[DDIR_READ])
598 int io_u_quiesce(struct thread_data *td)
603 * We are going to sleep, ensure that we flush anything pending as
604 * not to skew our latency numbers.
606 * Changed to only monitor 'in flight' requests here instead of the
607 * td->cur_depth, b/c td->cur_depth does not accurately represent
608 * io's that have been actually submitted to an async engine,
609 * and cur_depth is meaningless for sync engines.
611 if (td->io_u_queued || td->cur_depth)
614 while (td->io_u_in_flight) {
617 ret = io_u_queued_complete(td, 1);
622 if (td->flags & TD_F_REGROW_LOGS)
628 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
630 enum fio_ddir odir = ddir ^ 1;
634 assert(ddir_rw(ddir));
635 now = utime_since_now(&td->start);
638 * if rate_next_io_time is in the past, need to catch up to rate
640 if (td->rate_next_io_time[ddir] <= now)
644 * We are ahead of rate in this direction. See if we
647 if (td_rw(td) && td->o.rwmix[odir]) {
649 * Other direction is behind rate, switch
651 if (td->rate_next_io_time[odir] <= now)
655 * Both directions are ahead of rate. sleep the min,
656 * switch if necessary
658 if (td->rate_next_io_time[ddir] <=
659 td->rate_next_io_time[odir]) {
660 usec = td->rate_next_io_time[ddir] - now;
662 usec = td->rate_next_io_time[odir] - now;
666 usec = td->rate_next_io_time[ddir] - now;
668 if (td->o.io_submit_mode == IO_MODE_INLINE)
671 usec_sleep(td, usec);
676 * Return the data direction for the next io_u. If the job is a
677 * mixed read/write workload, check the rwmix cycle and switch if
680 static enum fio_ddir get_rw_ddir(struct thread_data *td)
685 * See if it's time to fsync/fdatasync/sync_file_range first,
686 * and if not then move on to check regular I/Os.
688 if (should_fsync(td)) {
689 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
690 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
693 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
694 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
695 return DDIR_DATASYNC;
697 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
698 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
699 return DDIR_SYNC_FILE_RANGE;
704 * Check if it's time to seed a new data direction.
706 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
708 * Put a top limit on how many bytes we do for
709 * one data direction, to avoid overflowing the
712 ddir = get_rand_ddir(td);
714 if (ddir != td->rwmix_ddir)
717 td->rwmix_ddir = ddir;
719 ddir = td->rwmix_ddir;
720 } else if (td_read(td))
722 else if (td_write(td))
724 else if (td_trim(td))
729 td->rwmix_ddir = rate_ddir(td, ddir);
730 return td->rwmix_ddir;
733 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
735 enum fio_ddir ddir = get_rw_ddir(td);
737 if (td_trimwrite(td)) {
738 struct fio_file *f = io_u->file;
739 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
745 io_u->ddir = io_u->acct_ddir = ddir;
747 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
748 td->o.barrier_blocks &&
749 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
750 td->io_issues[DDIR_WRITE])
751 io_u_set(td, io_u, IO_U_F_BARRIER);
754 void put_file_log(struct thread_data *td, struct fio_file *f)
756 unsigned int ret = put_file(td, f);
759 td_verror(td, ret, "file close");
762 void put_io_u(struct thread_data *td, struct io_u *io_u)
769 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
770 put_file_log(td, io_u->file);
773 io_u_set(td, io_u, IO_U_F_FREE);
775 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
777 assert(!(td->flags & TD_F_CHILD));
779 io_u_qpush(&td->io_u_freelist, io_u);
780 td_io_u_free_notify(td);
784 void clear_io_u(struct thread_data *td, struct io_u *io_u)
786 io_u_clear(td, io_u, IO_U_F_FLIGHT);
790 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
792 struct io_u *__io_u = *io_u;
793 enum fio_ddir ddir = acct_ddir(__io_u);
795 dprint(FD_IO, "requeue %p\n", __io_u);
802 io_u_set(td, __io_u, IO_U_F_FREE);
803 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
804 td->io_issues[ddir]--;
806 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
807 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
809 assert(!(td->flags & TD_F_CHILD));
812 io_u_rpush(&td->io_u_requeues, __io_u);
813 td_io_u_free_notify(td);
818 static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
820 struct fio_file *f = io_u->file;
822 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
823 assert(td->o.zone_size);
824 assert(td->o.zone_range);
827 * See if it's time to switch to a new zone
829 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
831 f->file_offset += td->o.zone_range + td->o.zone_skip;
834 * Wrap from the beginning, if we exceed the file size
836 if (f->file_offset >= f->real_file_size)
837 f->file_offset = get_start_offset(td, f);
839 f->last_pos[io_u->ddir] = f->file_offset;
840 td->io_skip_bytes += td->o.zone_skip;
844 * If zone_size > zone_range, then maintain the same zone until
845 * zone_bytes >= zone_size.
847 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
848 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
849 f->file_offset, f->last_pos[io_u->ddir]);
850 f->last_pos[io_u->ddir] = f->file_offset;
854 * For random: if 'norandommap' is not set and zone_size > zone_range,
855 * map needs to be reset as it's done with zone_range everytime.
857 if ((td->zone_bytes % td->o.zone_range) == 0)
858 fio_file_reset(td, f);
861 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
865 if (td_ioengine_flagged(td, FIO_NOIO))
868 set_rw_ddir(td, io_u);
871 * fsync() or fdatasync() or trim etc, we are done
873 if (!ddir_rw(io_u->ddir))
876 if (td->o.zone_mode == ZONE_MODE_STRIDED)
877 setup_strided_zone_mode(td, io_u);
880 * No log, let the seq/rand engine retrieve the next buflen and
883 if (get_next_offset(td, io_u, &is_random)) {
884 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
888 io_u->buflen = get_next_buflen(td, io_u, is_random);
890 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
894 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
895 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
897 (unsigned long long) io_u->offset, io_u->buflen,
898 (unsigned long long) io_u->file->real_file_size);
903 * mark entry before potentially trimming io_u
905 if (td_random(td) && file_randommap(td, io_u->file))
906 mark_random_map(td, io_u);
909 dprint_io_u(io_u, "fill");
910 td->zone_bytes += io_u->buflen;
914 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
943 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
945 __io_u_mark_map(td->ts.io_u_submit, nr);
946 td->ts.total_submit++;
949 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
951 __io_u_mark_map(td->ts.io_u_complete, nr);
952 td->ts.total_complete++;
955 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
959 switch (td->cur_depth) {
981 td->ts.io_u_map[idx] += nr;
984 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1021 assert(idx < FIO_IO_U_LAT_N_NR);
1022 td->ts.io_u_lat_n[idx]++;
1025 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1029 assert(usec < 1000 && usec >= 1);
1062 assert(idx < FIO_IO_U_LAT_U_NR);
1063 td->ts.io_u_lat_u[idx]++;
1066 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1109 assert(idx < FIO_IO_U_LAT_M_NR);
1110 td->ts.io_u_lat_m[idx]++;
1113 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1116 io_u_mark_lat_nsec(td, nsec);
1117 else if (nsec < 1000000)
1118 io_u_mark_lat_usec(td, nsec / 1000);
1120 io_u_mark_lat_msec(td, nsec / 1000000);
1123 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1125 unsigned long fileno;
1127 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1128 uint64_t frand_max = rand_max(&td->next_file_state);
1131 r = __rand(&td->next_file_state);
1132 return (unsigned int) ((double) td->o.nr_files
1133 * (r / (frand_max + 1.0)));
1136 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1137 fileno = zipf_next(&td->next_file_zipf);
1138 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1139 fileno = pareto_next(&td->next_file_zipf);
1140 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1141 fileno = gauss_next(&td->next_file_gauss);
1143 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1148 return fileno >> FIO_FSERVICE_SHIFT;
1152 * Get next file to service by choosing one at random
1154 static struct fio_file *get_next_file_rand(struct thread_data *td,
1155 enum fio_file_flags goodf,
1156 enum fio_file_flags badf)
1164 fno = __get_next_fileno_rand(td);
1167 if (fio_file_done(f))
1170 if (!fio_file_open(f)) {
1173 if (td->nr_open_files >= td->o.open_files)
1174 return ERR_PTR(-EBUSY);
1176 err = td_io_open_file(td, f);
1182 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1183 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1187 td_io_close_file(td, f);
1192 * Get next file to service by doing round robin between all available ones
1194 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1197 unsigned int old_next_file = td->next_file;
1203 f = td->files[td->next_file];
1206 if (td->next_file >= td->o.nr_files)
1209 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1210 if (fio_file_done(f)) {
1215 if (!fio_file_open(f)) {
1218 if (td->nr_open_files >= td->o.open_files)
1219 return ERR_PTR(-EBUSY);
1221 err = td_io_open_file(td, f);
1223 dprint(FD_FILE, "error %d on open of %s\n",
1231 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1233 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1237 td_io_close_file(td, f);
1240 } while (td->next_file != old_next_file);
1242 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1246 static struct fio_file *__get_next_file(struct thread_data *td)
1250 assert(td->o.nr_files <= td->files_index);
1252 if (td->nr_done_files >= td->o.nr_files) {
1253 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1254 " nr_files=%d\n", td->nr_open_files,
1260 f = td->file_service_file;
1261 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1262 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1264 if (td->file_service_left--)
1268 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1269 td->o.file_service_type == FIO_FSERVICE_SEQ)
1270 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1272 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1277 td->file_service_file = f;
1278 td->file_service_left = td->file_service_nr - 1;
1281 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1283 dprint(FD_FILE, "get_next_file: NULL\n");
1287 static struct fio_file *get_next_file(struct thread_data *td)
1289 return __get_next_file(td);
1292 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1297 f = get_next_file(td);
1298 if (IS_ERR_OR_NULL(f))
1304 if (!fill_io_u(td, io_u))
1307 put_file_log(td, f);
1308 td_io_close_file(td, f);
1310 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1311 fio_file_reset(td, f);
1313 fio_file_set_done(f);
1314 td->nr_done_files++;
1315 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1316 td->nr_done_files, td->o.nr_files);
1323 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1324 unsigned long long tnsec, unsigned long long max_nsec)
1327 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1328 td_verror(td, ETIMEDOUT, "max latency exceeded");
1329 icd->error = ETIMEDOUT;
1332 static void lat_new_cycle(struct thread_data *td)
1334 fio_gettime(&td->latency_ts, NULL);
1335 td->latency_ios = ddir_rw_sum(td->io_blocks);
1336 td->latency_failed = 0;
1340 * We had an IO outside the latency target. Reduce the queue depth. If we
1341 * are at QD=1, then it's time to give up.
1343 static bool __lat_target_failed(struct thread_data *td)
1345 if (td->latency_qd == 1)
1348 td->latency_qd_high = td->latency_qd;
1350 if (td->latency_qd == td->latency_qd_low)
1351 td->latency_qd_low--;
1353 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1355 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1358 * When we ramp QD down, quiesce existing IO to prevent
1359 * a storm of ramp downs due to pending higher depth.
1366 static bool lat_target_failed(struct thread_data *td)
1368 if (td->o.latency_percentile.u.f == 100.0)
1369 return __lat_target_failed(td);
1371 td->latency_failed++;
1375 void lat_target_init(struct thread_data *td)
1377 td->latency_end_run = 0;
1379 if (td->o.latency_target) {
1380 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1381 fio_gettime(&td->latency_ts, NULL);
1383 td->latency_qd_high = td->o.iodepth;
1384 td->latency_qd_low = 1;
1385 td->latency_ios = ddir_rw_sum(td->io_blocks);
1387 td->latency_qd = td->o.iodepth;
1390 void lat_target_reset(struct thread_data *td)
1392 if (!td->latency_end_run)
1393 lat_target_init(td);
1396 static void lat_target_success(struct thread_data *td)
1398 const unsigned int qd = td->latency_qd;
1399 struct thread_options *o = &td->o;
1401 td->latency_qd_low = td->latency_qd;
1404 * If we haven't failed yet, we double up to a failing value instead
1405 * of bisecting from highest possible queue depth. If we have set
1406 * a limit other than td->o.iodepth, bisect between that.
1408 if (td->latency_qd_high != o->iodepth)
1409 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1411 td->latency_qd *= 2;
1413 if (td->latency_qd > o->iodepth)
1414 td->latency_qd = o->iodepth;
1416 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1419 * Same as last one, we are done. Let it run a latency cycle, so
1420 * we get only the results from the targeted depth.
1422 if (td->latency_qd == qd) {
1423 if (td->latency_end_run) {
1424 dprint(FD_RATE, "We are done\n");
1427 dprint(FD_RATE, "Quiesce and final run\n");
1429 td->latency_end_run = 1;
1430 reset_all_stats(td);
1439 * Check if we can bump the queue depth
1441 void lat_target_check(struct thread_data *td)
1443 uint64_t usec_window;
1447 usec_window = utime_since_now(&td->latency_ts);
1448 if (usec_window < td->o.latency_window)
1451 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1452 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1453 success_ios *= 100.0;
1455 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1457 if (success_ios >= td->o.latency_percentile.u.f)
1458 lat_target_success(td);
1460 __lat_target_failed(td);
1464 * If latency target is enabled, we might be ramping up or down and not
1465 * using the full queue depth available.
1467 bool queue_full(const struct thread_data *td)
1469 const int qempty = io_u_qempty(&td->io_u_freelist);
1473 if (!td->o.latency_target)
1476 return td->cur_depth >= td->latency_qd;
1479 struct io_u *__get_io_u(struct thread_data *td)
1481 struct io_u *io_u = NULL;
1490 if (!io_u_rempty(&td->io_u_requeues))
1491 io_u = io_u_rpop(&td->io_u_requeues);
1492 else if (!queue_full(td)) {
1493 io_u = io_u_qpop(&td->io_u_freelist);
1498 io_u->end_io = NULL;
1502 assert(io_u->flags & IO_U_F_FREE);
1503 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1504 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1508 io_u->acct_ddir = -1;
1510 assert(!(td->flags & TD_F_CHILD));
1511 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1513 } else if (td_async_processing(td)) {
1515 * We ran out, wait for async verify threads to finish and
1518 assert(!(td->flags & TD_F_CHILD));
1519 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1528 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1530 if (!(td->flags & TD_F_TRIM_BACKLOG))
1532 if (!td->trim_entries)
1535 if (td->trim_batch) {
1537 if (get_next_trim(td, io_u))
1539 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1540 td->last_ddir != DDIR_READ) {
1541 td->trim_batch = td->o.trim_batch;
1542 if (!td->trim_batch)
1543 td->trim_batch = td->o.trim_backlog;
1544 if (get_next_trim(td, io_u))
1551 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1553 if (!(td->flags & TD_F_VER_BACKLOG))
1556 if (td->io_hist_len) {
1559 if (td->verify_batch)
1561 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1562 td->last_ddir != DDIR_READ) {
1563 td->verify_batch = td->o.verify_batch;
1564 if (!td->verify_batch)
1565 td->verify_batch = td->o.verify_backlog;
1569 if (get_verify && !get_next_verify(td, io_u)) {
1579 * Fill offset and start time into the buffer content, to prevent too
1580 * easy compressible data for simple de-dupe attempts. Do this for every
1581 * 512b block in the range, since that should be the smallest block size
1582 * we can expect from a device.
1584 static void small_content_scramble(struct io_u *io_u)
1586 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1587 unsigned int offset;
1588 uint64_t boffset, *iptr;
1595 boffset = io_u->offset;
1597 if (io_u->buf_filled_len)
1598 io_u->buf_filled_len = 0;
1601 * Generate random index between 0..7. We do chunks of 512b, if
1602 * we assume a cacheline is 64 bytes, then we have 8 of those.
1603 * Scramble content within the blocks in the same cacheline to
1606 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1608 for (i = 0; i < nr_blocks; i++) {
1610 * Fill offset into start of cacheline, time into end
1613 iptr = (void *) p + (offset << 6);
1616 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1617 iptr[0] = io_u->start_time.tv_sec;
1618 iptr[1] = io_u->start_time.tv_nsec;
1626 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1627 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1629 struct io_u *get_io_u(struct thread_data *td)
1633 int do_scramble = 0;
1636 io_u = __get_io_u(td);
1638 dprint(FD_IO, "__get_io_u failed\n");
1642 if (check_get_verify(td, io_u))
1644 if (check_get_trim(td, io_u))
1648 * from a requeue, io_u already setup
1654 * If using an iolog, grab next piece if any available.
1656 if (td->flags & TD_F_READ_IOLOG) {
1657 if (read_iolog_get(td, io_u))
1659 } else if (set_io_u_file(td, io_u)) {
1661 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1667 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1671 assert(fio_file_open(f));
1673 if (ddir_rw(io_u->ddir)) {
1674 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1675 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1679 f->last_start[io_u->ddir] = io_u->offset;
1680 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1682 if (io_u->ddir == DDIR_WRITE) {
1683 if (td->flags & TD_F_REFILL_BUFFERS) {
1684 io_u_fill_buffer(td, io_u,
1685 td->o.min_bs[DDIR_WRITE],
1687 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1688 !(td->flags & TD_F_COMPRESS) &&
1689 !(td->flags & TD_F_DO_VERIFY))
1691 } else if (io_u->ddir == DDIR_READ) {
1693 * Reset the buf_filled parameters so next time if the
1694 * buffer is used for writes it is refilled.
1696 io_u->buf_filled_len = 0;
1701 * Set io data pointers.
1703 io_u->xfer_buf = io_u->buf;
1704 io_u->xfer_buflen = io_u->buflen;
1708 if (!td_io_prep(td, io_u)) {
1709 if (!td->o.disable_lat)
1710 fio_gettime(&io_u->start_time, NULL);
1713 small_content_scramble(io_u);
1718 dprint(FD_IO, "get_io_u failed\n");
1720 return ERR_PTR(ret);
1723 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1725 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1727 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1730 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1731 io_u->file ? " on file " : "",
1732 io_u->file ? io_u->file->file_name : "",
1733 strerror(io_u->error),
1734 io_ddir_name(io_u->ddir),
1735 io_u->offset, io_u->xfer_buflen);
1737 if (td->io_ops->errdetails) {
1738 char *err = td->io_ops->errdetails(io_u);
1740 log_err("fio: %s\n", err);
1745 td_verror(td, io_u->error, "io_u error");
1748 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1750 __io_u_log_error(td, io_u);
1752 __io_u_log_error(td->parent, io_u);
1755 static inline bool gtod_reduce(struct thread_data *td)
1757 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1758 || td->o.gtod_reduce;
1761 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1762 struct io_completion_data *icd,
1763 const enum fio_ddir idx, unsigned int bytes)
1765 const int no_reduce = !gtod_reduce(td);
1766 unsigned long long llnsec = 0;
1771 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1775 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1777 if (!td->o.disable_lat) {
1778 unsigned long long tnsec;
1780 tnsec = ntime_since(&io_u->start_time, &icd->time);
1781 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1783 if (td->flags & TD_F_PROFILE_OPS) {
1784 struct prof_io_ops *ops = &td->prof_io_ops;
1787 icd->error = ops->io_u_lat(td, tnsec);
1790 if (td->o.max_latency && tnsec > td->o.max_latency)
1791 lat_fatal(td, icd, tnsec, td->o.max_latency);
1792 if (td->o.latency_target && tnsec > td->o.latency_target) {
1793 if (lat_target_failed(td))
1794 lat_fatal(td, icd, tnsec, td->o.latency_target);
1799 if (!td->o.disable_clat) {
1800 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1801 io_u_mark_latency(td, llnsec);
1804 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1805 add_bw_sample(td, io_u, bytes, llnsec);
1807 if (no_reduce && per_unit_log(td->iops_log))
1808 add_iops_sample(td, io_u, bytes);
1809 } else if (ddir_sync(idx) && !td->o.disable_clat)
1810 add_sync_clat_sample(&td->ts, llnsec);
1812 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1813 uint32_t *info = io_u_block_info(td, io_u);
1814 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1815 if (io_u->ddir == DDIR_TRIM) {
1816 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1817 BLOCK_INFO_TRIMS(*info) + 1);
1818 } else if (io_u->ddir == DDIR_WRITE) {
1819 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1826 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1827 uint64_t offset, unsigned int bytes)
1834 if (f->first_write == -1ULL || offset < f->first_write)
1835 f->first_write = offset;
1836 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1837 f->last_write = offset + bytes;
1839 if (!f->last_write_comp)
1842 idx = f->last_write_idx++;
1843 f->last_write_comp[idx] = offset;
1844 if (f->last_write_idx == td->o.iodepth)
1845 f->last_write_idx = 0;
1848 static bool should_account(struct thread_data *td)
1850 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1851 td->runstate == TD_VERIFYING);
1854 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1855 struct io_completion_data *icd)
1857 struct io_u *io_u = *io_u_ptr;
1858 enum fio_ddir ddir = io_u->ddir;
1859 struct fio_file *f = io_u->file;
1861 dprint_io_u(io_u, "complete");
1863 assert(io_u->flags & IO_U_F_FLIGHT);
1864 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1867 * Mark IO ok to verify
1871 * Remove errored entry from the verification list
1874 unlog_io_piece(td, io_u);
1876 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1881 if (ddir_sync(ddir)) {
1882 td->last_was_sync = true;
1884 f->first_write = -1ULL;
1885 f->last_write = -1ULL;
1887 if (should_account(td))
1888 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1892 td->last_was_sync = false;
1893 td->last_ddir = ddir;
1895 if (!io_u->error && ddir_rw(ddir)) {
1896 unsigned long long bytes = io_u->buflen - io_u->resid;
1899 td->io_blocks[ddir]++;
1900 td->io_bytes[ddir] += bytes;
1902 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1903 td->this_io_blocks[ddir]++;
1904 td->this_io_bytes[ddir] += bytes;
1907 if (ddir == DDIR_WRITE)
1908 file_log_write_comp(td, f, io_u->offset, bytes);
1910 if (should_account(td))
1911 account_io_completion(td, io_u, icd, ddir, bytes);
1913 icd->bytes_done[ddir] += bytes;
1916 ret = io_u->end_io(td, io_u_ptr);
1918 if (ret && !icd->error)
1921 } else if (io_u->error) {
1922 icd->error = io_u->error;
1923 io_u_log_error(td, io_u);
1926 enum error_type_bit eb = td_error_type(ddir, icd->error);
1928 if (!td_non_fatal_error(td, eb, icd->error))
1932 * If there is a non_fatal error, then add to the error count
1933 * and clear all the errors.
1935 update_error_count(td, icd->error);
1943 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1948 if (!gtod_reduce(td))
1949 fio_gettime(&icd->time, NULL);
1954 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1955 icd->bytes_done[ddir] = 0;
1958 static void ios_completed(struct thread_data *td,
1959 struct io_completion_data *icd)
1964 for (i = 0; i < icd->nr; i++) {
1965 io_u = td->io_ops->event(td, i);
1967 io_completed(td, &io_u, icd);
1975 * Complete a single io_u for the sync engines.
1977 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1979 struct io_completion_data icd;
1982 init_icd(td, &icd, 1);
1983 io_completed(td, &io_u, &icd);
1989 td_verror(td, icd.error, "io_u_sync_complete");
1993 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1994 td->bytes_done[ddir] += icd.bytes_done[ddir];
2000 * Called to complete min_events number of io for the async engines.
2002 int io_u_queued_complete(struct thread_data *td, int min_evts)
2004 struct io_completion_data icd;
2005 struct timespec *tvp = NULL;
2007 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2009 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2013 else if (min_evts > td->cur_depth)
2014 min_evts = td->cur_depth;
2016 /* No worries, td_io_getevents fixes min and max if they are
2017 * set incorrectly */
2018 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2020 td_verror(td, -ret, "td_io_getevents");
2025 init_icd(td, &icd, ret);
2026 ios_completed(td, &icd);
2028 td_verror(td, icd.error, "io_u_queued_complete");
2032 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2033 td->bytes_done[ddir] += icd.bytes_done[ddir];
2039 * Call when io_u is really queued, to update the submission latency.
2041 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2043 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2044 unsigned long slat_time;
2046 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2051 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2057 * See if we should reuse the last seed, if dedupe is enabled
2059 static struct frand_state *get_buf_state(struct thread_data *td)
2063 if (!td->o.dedupe_percentage)
2064 return &td->buf_state;
2065 else if (td->o.dedupe_percentage == 100) {
2066 frand_copy(&td->buf_state_prev, &td->buf_state);
2067 return &td->buf_state;
2070 v = rand_between(&td->dedupe_state, 1, 100);
2072 if (v <= td->o.dedupe_percentage)
2073 return &td->buf_state_prev;
2075 return &td->buf_state;
2078 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2080 if (td->o.dedupe_percentage == 100)
2081 frand_copy(rs, &td->buf_state_prev);
2082 else if (rs == &td->buf_state)
2083 frand_copy(&td->buf_state_prev, rs);
2086 void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2087 unsigned long long max_bs)
2089 struct thread_options *o = &td->o;
2091 if (o->mem_type == MEM_CUDA_MALLOC)
2094 if (o->compress_percentage || o->dedupe_percentage) {
2095 unsigned int perc = td->o.compress_percentage;
2096 struct frand_state *rs;
2097 unsigned long long left = max_bs;
2098 unsigned long long this_write;
2101 rs = get_buf_state(td);
2103 min_write = min(min_write, left);
2106 this_write = min_not_zero(min_write,
2107 (unsigned long long) td->o.compress_chunk);
2109 fill_random_buf_percentage(rs, buf, perc,
2110 this_write, this_write,
2112 o->buffer_pattern_bytes);
2114 fill_random_buf(rs, buf, min_write);
2115 this_write = min_write;
2120 save_buf_state(td, rs);
2122 } else if (o->buffer_pattern_bytes)
2123 fill_buffer_pattern(td, buf, max_bs);
2124 else if (o->zero_buffers)
2125 memset(buf, 0, max_bs);
2127 fill_random_buf(get_buf_state(td), buf, max_bs);
2131 * "randomly" fill the buffer contents
2133 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2134 unsigned long long min_write, unsigned long long max_bs)
2136 io_u->buf_filled_len = 0;
2137 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2140 static int do_sync_file_range(const struct thread_data *td,
2143 off64_t offset, nbytes;
2145 offset = f->first_write;
2146 nbytes = f->last_write - f->first_write;
2151 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2154 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2158 if (io_u->ddir == DDIR_SYNC) {
2159 ret = fsync(io_u->file->fd);
2160 } else if (io_u->ddir == DDIR_DATASYNC) {
2161 #ifdef CONFIG_FDATASYNC
2162 ret = fdatasync(io_u->file->fd);
2164 ret = io_u->xfer_buflen;
2165 io_u->error = EINVAL;
2167 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2168 ret = do_sync_file_range(td, io_u->file);
2170 ret = io_u->xfer_buflen;
2171 io_u->error = EINVAL;
2175 io_u->error = errno;
2180 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2182 #ifndef FIO_HAVE_TRIM
2183 io_u->error = EINVAL;
2186 struct fio_file *f = io_u->file;
2189 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2191 return io_u->xfer_buflen;