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 int min_bs = td->o.min_bs[io_u->ddir];
37 struct fio_file *f = io_u->file;
38 unsigned int 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;
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,
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 int get_next_buflen(struct thread_data *td, struct io_u *io_u,
516 int ddir = io_u->ddir;
517 unsigned int buflen = 0;
518 unsigned int 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 = 1 + (unsigned int) ((double) maxbs *
545 (r / (frand_max + 1.0)));
552 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
553 struct bssplit *bsp = &td->o.bssplit[ddir][i];
559 if ((r / perc <= frand_max / 100ULL) &&
560 io_u_fits(td, io_u, buflen))
565 power_2 = is_power_of_2(minbs);
566 if (!td->o.bs_unaligned && power_2)
567 buflen &= ~(minbs - 1);
568 else if (!td->o.bs_unaligned && !power_2)
569 buflen -= buflen % minbs;
570 } while (!io_u_fits(td, io_u, buflen));
575 static void set_rwmix_bytes(struct thread_data *td)
580 * we do time or byte based switch. this is needed because
581 * buffered writes may issue a lot quicker than they complete,
582 * whereas reads do not.
584 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
585 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
588 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
592 v = rand_between(&td->rwmix_state, 1, 100);
594 if (v <= td->o.rwmix[DDIR_READ])
600 int io_u_quiesce(struct thread_data *td)
605 * We are going to sleep, ensure that we flush anything pending as
606 * not to skew our latency numbers.
608 * Changed to only monitor 'in flight' requests here instead of the
609 * td->cur_depth, b/c td->cur_depth does not accurately represent
610 * io's that have been actually submitted to an async engine,
611 * and cur_depth is meaningless for sync engines.
613 if (td->io_u_queued || td->cur_depth)
616 while (td->io_u_in_flight) {
619 ret = io_u_queued_complete(td, 1);
624 if (td->flags & TD_F_REGROW_LOGS)
630 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
632 enum fio_ddir odir = ddir ^ 1;
636 assert(ddir_rw(ddir));
637 now = utime_since_now(&td->start);
640 * if rate_next_io_time is in the past, need to catch up to rate
642 if (td->rate_next_io_time[ddir] <= now)
646 * We are ahead of rate in this direction. See if we
649 if (td_rw(td) && td->o.rwmix[odir]) {
651 * Other direction is behind rate, switch
653 if (td->rate_next_io_time[odir] <= now)
657 * Both directions are ahead of rate. sleep the min,
658 * switch if necessary
660 if (td->rate_next_io_time[ddir] <=
661 td->rate_next_io_time[odir]) {
662 usec = td->rate_next_io_time[ddir] - now;
664 usec = td->rate_next_io_time[odir] - now;
668 usec = td->rate_next_io_time[ddir] - now;
670 if (td->o.io_submit_mode == IO_MODE_INLINE)
673 usec_sleep(td, usec);
678 * Return the data direction for the next io_u. If the job is a
679 * mixed read/write workload, check the rwmix cycle and switch if
682 static enum fio_ddir get_rw_ddir(struct thread_data *td)
687 * See if it's time to fsync/fdatasync/sync_file_range first,
688 * and if not then move on to check regular I/Os.
690 if (should_fsync(td)) {
691 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
692 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
695 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
696 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
697 return DDIR_DATASYNC;
699 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
700 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
701 return DDIR_SYNC_FILE_RANGE;
706 * Check if it's time to seed a new data direction.
708 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
710 * Put a top limit on how many bytes we do for
711 * one data direction, to avoid overflowing the
714 ddir = get_rand_ddir(td);
716 if (ddir != td->rwmix_ddir)
719 td->rwmix_ddir = ddir;
721 ddir = td->rwmix_ddir;
722 } else if (td_read(td))
724 else if (td_write(td))
726 else if (td_trim(td))
731 td->rwmix_ddir = rate_ddir(td, ddir);
732 return td->rwmix_ddir;
735 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
737 enum fio_ddir ddir = get_rw_ddir(td);
739 if (td_trimwrite(td)) {
740 struct fio_file *f = io_u->file;
741 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
747 io_u->ddir = io_u->acct_ddir = ddir;
749 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
750 td->o.barrier_blocks &&
751 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
752 td->io_issues[DDIR_WRITE])
753 io_u_set(td, io_u, IO_U_F_BARRIER);
756 void put_file_log(struct thread_data *td, struct fio_file *f)
758 unsigned int ret = put_file(td, f);
761 td_verror(td, ret, "file close");
764 void put_io_u(struct thread_data *td, struct io_u *io_u)
771 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
772 put_file_log(td, io_u->file);
775 io_u_set(td, io_u, IO_U_F_FREE);
777 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
779 assert(!(td->flags & TD_F_CHILD));
781 io_u_qpush(&td->io_u_freelist, io_u);
782 td_io_u_free_notify(td);
786 void clear_io_u(struct thread_data *td, struct io_u *io_u)
788 io_u_clear(td, io_u, IO_U_F_FLIGHT);
792 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
794 struct io_u *__io_u = *io_u;
795 enum fio_ddir ddir = acct_ddir(__io_u);
797 dprint(FD_IO, "requeue %p\n", __io_u);
804 io_u_set(td, __io_u, IO_U_F_FREE);
805 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
806 td->io_issues[ddir]--;
808 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
809 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
811 assert(!(td->flags & TD_F_CHILD));
814 io_u_rpush(&td->io_u_requeues, __io_u);
815 td_io_u_free_notify(td);
820 static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
822 struct fio_file *f = io_u->file;
825 * See if it's time to switch to a new zone
827 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
829 f->file_offset += td->o.zone_range + td->o.zone_skip;
832 * Wrap from the beginning, if we exceed the file size
834 if (f->file_offset >= f->real_file_size)
835 f->file_offset = get_start_offset(td, f);
837 f->last_pos[io_u->ddir] = f->file_offset;
838 td->io_skip_bytes += td->o.zone_skip;
842 * If zone_size > zone_range, then maintain the same zone until
843 * zone_bytes >= zone_size.
845 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
846 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
847 f->file_offset, f->last_pos[io_u->ddir]);
848 f->last_pos[io_u->ddir] = f->file_offset;
852 * For random: if 'norandommap' is not set and zone_size > zone_range,
853 * map needs to be reset as it's done with zone_range everytime.
855 if ((td->zone_bytes % td->o.zone_range) == 0)
856 fio_file_reset(td, f);
859 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
863 if (td_ioengine_flagged(td, FIO_NOIO))
866 set_rw_ddir(td, io_u);
869 * fsync() or fdatasync() or trim etc, we are done
871 if (!ddir_rw(io_u->ddir))
875 * When file is zoned zone_range is always positive
877 if (td->o.zone_range)
878 __fill_io_u_zone(td, io_u);
881 * No log, let the seq/rand engine retrieve the next buflen and
884 if (get_next_offset(td, io_u, &is_random)) {
885 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
889 io_u->buflen = get_next_buflen(td, io_u, is_random);
891 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
895 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
896 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
898 (unsigned long long) io_u->offset, io_u->buflen,
899 (unsigned long long) io_u->file->real_file_size);
904 * mark entry before potentially trimming io_u
906 if (td_random(td) && file_randommap(td, io_u->file))
907 mark_random_map(td, io_u);
910 dprint_io_u(io_u, "fill");
911 td->zone_bytes += io_u->buflen;
915 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
944 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
946 __io_u_mark_map(td->ts.io_u_submit, nr);
947 td->ts.total_submit++;
950 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
952 __io_u_mark_map(td->ts.io_u_complete, nr);
953 td->ts.total_complete++;
956 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
960 switch (td->cur_depth) {
982 td->ts.io_u_map[idx] += nr;
985 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1022 assert(idx < FIO_IO_U_LAT_N_NR);
1023 td->ts.io_u_lat_n[idx]++;
1026 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1030 assert(usec < 1000 && usec >= 1);
1063 assert(idx < FIO_IO_U_LAT_U_NR);
1064 td->ts.io_u_lat_u[idx]++;
1067 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1110 assert(idx < FIO_IO_U_LAT_M_NR);
1111 td->ts.io_u_lat_m[idx]++;
1114 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1117 io_u_mark_lat_nsec(td, nsec);
1118 else if (nsec < 1000000)
1119 io_u_mark_lat_usec(td, nsec / 1000);
1121 io_u_mark_lat_msec(td, nsec / 1000000);
1124 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1126 unsigned long fileno;
1128 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1129 uint64_t frand_max = rand_max(&td->next_file_state);
1132 r = __rand(&td->next_file_state);
1133 return (unsigned int) ((double) td->o.nr_files
1134 * (r / (frand_max + 1.0)));
1137 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1138 fileno = zipf_next(&td->next_file_zipf);
1139 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1140 fileno = pareto_next(&td->next_file_zipf);
1141 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1142 fileno = gauss_next(&td->next_file_gauss);
1144 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1149 return fileno >> FIO_FSERVICE_SHIFT;
1153 * Get next file to service by choosing one at random
1155 static struct fio_file *get_next_file_rand(struct thread_data *td,
1156 enum fio_file_flags goodf,
1157 enum fio_file_flags badf)
1165 fno = __get_next_fileno_rand(td);
1168 if (fio_file_done(f))
1171 if (!fio_file_open(f)) {
1174 if (td->nr_open_files >= td->o.open_files)
1175 return ERR_PTR(-EBUSY);
1177 err = td_io_open_file(td, f);
1183 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1184 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1188 td_io_close_file(td, f);
1193 * Get next file to service by doing round robin between all available ones
1195 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1198 unsigned int old_next_file = td->next_file;
1204 f = td->files[td->next_file];
1207 if (td->next_file >= td->o.nr_files)
1210 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1211 if (fio_file_done(f)) {
1216 if (!fio_file_open(f)) {
1219 if (td->nr_open_files >= td->o.open_files)
1220 return ERR_PTR(-EBUSY);
1222 err = td_io_open_file(td, f);
1224 dprint(FD_FILE, "error %d on open of %s\n",
1232 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1234 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1238 td_io_close_file(td, f);
1241 } while (td->next_file != old_next_file);
1243 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1247 static struct fio_file *__get_next_file(struct thread_data *td)
1251 assert(td->o.nr_files <= td->files_index);
1253 if (td->nr_done_files >= td->o.nr_files) {
1254 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1255 " nr_files=%d\n", td->nr_open_files,
1261 f = td->file_service_file;
1262 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1263 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1265 if (td->file_service_left--)
1269 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1270 td->o.file_service_type == FIO_FSERVICE_SEQ)
1271 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1273 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1278 td->file_service_file = f;
1279 td->file_service_left = td->file_service_nr - 1;
1282 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1284 dprint(FD_FILE, "get_next_file: NULL\n");
1288 static struct fio_file *get_next_file(struct thread_data *td)
1290 return __get_next_file(td);
1293 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1298 f = get_next_file(td);
1299 if (IS_ERR_OR_NULL(f))
1305 if (!fill_io_u(td, io_u))
1308 put_file_log(td, f);
1309 td_io_close_file(td, f);
1311 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1312 fio_file_reset(td, f);
1314 fio_file_set_done(f);
1315 td->nr_done_files++;
1316 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1317 td->nr_done_files, td->o.nr_files);
1324 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1325 unsigned long long tnsec, unsigned long long max_nsec)
1328 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1329 td_verror(td, ETIMEDOUT, "max latency exceeded");
1330 icd->error = ETIMEDOUT;
1333 static void lat_new_cycle(struct thread_data *td)
1335 fio_gettime(&td->latency_ts, NULL);
1336 td->latency_ios = ddir_rw_sum(td->io_blocks);
1337 td->latency_failed = 0;
1341 * We had an IO outside the latency target. Reduce the queue depth. If we
1342 * are at QD=1, then it's time to give up.
1344 static bool __lat_target_failed(struct thread_data *td)
1346 if (td->latency_qd == 1)
1349 td->latency_qd_high = td->latency_qd;
1351 if (td->latency_qd == td->latency_qd_low)
1352 td->latency_qd_low--;
1354 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1356 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1359 * When we ramp QD down, quiesce existing IO to prevent
1360 * a storm of ramp downs due to pending higher depth.
1367 static bool lat_target_failed(struct thread_data *td)
1369 if (td->o.latency_percentile.u.f == 100.0)
1370 return __lat_target_failed(td);
1372 td->latency_failed++;
1376 void lat_target_init(struct thread_data *td)
1378 td->latency_end_run = 0;
1380 if (td->o.latency_target) {
1381 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1382 fio_gettime(&td->latency_ts, NULL);
1384 td->latency_qd_high = td->o.iodepth;
1385 td->latency_qd_low = 1;
1386 td->latency_ios = ddir_rw_sum(td->io_blocks);
1388 td->latency_qd = td->o.iodepth;
1391 void lat_target_reset(struct thread_data *td)
1393 if (!td->latency_end_run)
1394 lat_target_init(td);
1397 static void lat_target_success(struct thread_data *td)
1399 const unsigned int qd = td->latency_qd;
1400 struct thread_options *o = &td->o;
1402 td->latency_qd_low = td->latency_qd;
1405 * If we haven't failed yet, we double up to a failing value instead
1406 * of bisecting from highest possible queue depth. If we have set
1407 * a limit other than td->o.iodepth, bisect between that.
1409 if (td->latency_qd_high != o->iodepth)
1410 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1412 td->latency_qd *= 2;
1414 if (td->latency_qd > o->iodepth)
1415 td->latency_qd = o->iodepth;
1417 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1420 * Same as last one, we are done. Let it run a latency cycle, so
1421 * we get only the results from the targeted depth.
1423 if (td->latency_qd == qd) {
1424 if (td->latency_end_run) {
1425 dprint(FD_RATE, "We are done\n");
1428 dprint(FD_RATE, "Quiesce and final run\n");
1430 td->latency_end_run = 1;
1431 reset_all_stats(td);
1440 * Check if we can bump the queue depth
1442 void lat_target_check(struct thread_data *td)
1444 uint64_t usec_window;
1448 usec_window = utime_since_now(&td->latency_ts);
1449 if (usec_window < td->o.latency_window)
1452 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1453 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1454 success_ios *= 100.0;
1456 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1458 if (success_ios >= td->o.latency_percentile.u.f)
1459 lat_target_success(td);
1461 __lat_target_failed(td);
1465 * If latency target is enabled, we might be ramping up or down and not
1466 * using the full queue depth available.
1468 bool queue_full(const struct thread_data *td)
1470 const int qempty = io_u_qempty(&td->io_u_freelist);
1474 if (!td->o.latency_target)
1477 return td->cur_depth >= td->latency_qd;
1480 struct io_u *__get_io_u(struct thread_data *td)
1482 struct io_u *io_u = NULL;
1491 if (!io_u_rempty(&td->io_u_requeues))
1492 io_u = io_u_rpop(&td->io_u_requeues);
1493 else if (!queue_full(td)) {
1494 io_u = io_u_qpop(&td->io_u_freelist);
1499 io_u->end_io = NULL;
1503 assert(io_u->flags & IO_U_F_FREE);
1504 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1505 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1509 io_u->acct_ddir = -1;
1511 assert(!(td->flags & TD_F_CHILD));
1512 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1514 } else if (td_async_processing(td)) {
1516 * We ran out, wait for async verify threads to finish and
1519 assert(!(td->flags & TD_F_CHILD));
1520 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1529 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1531 if (!(td->flags & TD_F_TRIM_BACKLOG))
1533 if (!td->trim_entries)
1536 if (td->trim_batch) {
1538 if (get_next_trim(td, io_u))
1540 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1541 td->last_ddir != DDIR_READ) {
1542 td->trim_batch = td->o.trim_batch;
1543 if (!td->trim_batch)
1544 td->trim_batch = td->o.trim_backlog;
1545 if (get_next_trim(td, io_u))
1552 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1554 if (!(td->flags & TD_F_VER_BACKLOG))
1557 if (td->io_hist_len) {
1560 if (td->verify_batch)
1562 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1563 td->last_ddir != DDIR_READ) {
1564 td->verify_batch = td->o.verify_batch;
1565 if (!td->verify_batch)
1566 td->verify_batch = td->o.verify_backlog;
1570 if (get_verify && !get_next_verify(td, io_u)) {
1580 * Fill offset and start time into the buffer content, to prevent too
1581 * easy compressible data for simple de-dupe attempts. Do this for every
1582 * 512b block in the range, since that should be the smallest block size
1583 * we can expect from a device.
1585 static void small_content_scramble(struct io_u *io_u)
1587 unsigned int i, nr_blocks = io_u->buflen >> 9;
1588 unsigned int offset;
1589 uint64_t boffset, *iptr;
1596 boffset = io_u->offset;
1598 if (io_u->buf_filled_len)
1599 io_u->buf_filled_len = 0;
1602 * Generate random index between 0..7. We do chunks of 512b, if
1603 * we assume a cacheline is 64 bytes, then we have 8 of those.
1604 * Scramble content within the blocks in the same cacheline to
1607 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1609 for (i = 0; i < nr_blocks; i++) {
1611 * Fill offset into start of cacheline, time into end
1614 iptr = (void *) p + (offset << 6);
1617 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1618 iptr[0] = io_u->start_time.tv_sec;
1619 iptr[1] = io_u->start_time.tv_nsec;
1627 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1628 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1630 struct io_u *get_io_u(struct thread_data *td)
1634 int do_scramble = 0;
1637 io_u = __get_io_u(td);
1639 dprint(FD_IO, "__get_io_u failed\n");
1643 if (check_get_verify(td, io_u))
1645 if (check_get_trim(td, io_u))
1649 * from a requeue, io_u already setup
1655 * If using an iolog, grab next piece if any available.
1657 if (td->flags & TD_F_READ_IOLOG) {
1658 if (read_iolog_get(td, io_u))
1660 } else if (set_io_u_file(td, io_u)) {
1662 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1668 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1672 assert(fio_file_open(f));
1674 if (ddir_rw(io_u->ddir)) {
1675 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1676 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1680 f->last_start[io_u->ddir] = io_u->offset;
1681 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1683 if (io_u->ddir == DDIR_WRITE) {
1684 if (td->flags & TD_F_REFILL_BUFFERS) {
1685 io_u_fill_buffer(td, io_u,
1686 td->o.min_bs[DDIR_WRITE],
1688 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1689 !(td->flags & TD_F_COMPRESS) &&
1690 !(td->flags & TD_F_DO_VERIFY))
1692 } else if (io_u->ddir == DDIR_READ) {
1694 * Reset the buf_filled parameters so next time if the
1695 * buffer is used for writes it is refilled.
1697 io_u->buf_filled_len = 0;
1702 * Set io data pointers.
1704 io_u->xfer_buf = io_u->buf;
1705 io_u->xfer_buflen = io_u->buflen;
1709 if (!td_io_prep(td, io_u)) {
1710 if (!td->o.disable_lat)
1711 fio_gettime(&io_u->start_time, NULL);
1714 small_content_scramble(io_u);
1719 dprint(FD_IO, "get_io_u failed\n");
1721 return ERR_PTR(ret);
1724 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1726 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1728 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1731 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1732 io_u->file ? " on file " : "",
1733 io_u->file ? io_u->file->file_name : "",
1734 strerror(io_u->error),
1735 io_ddir_name(io_u->ddir),
1736 io_u->offset, io_u->xfer_buflen);
1738 if (td->io_ops->errdetails) {
1739 char *err = td->io_ops->errdetails(io_u);
1741 log_err("fio: %s\n", err);
1746 td_verror(td, io_u->error, "io_u error");
1749 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1751 __io_u_log_error(td, io_u);
1753 __io_u_log_error(td->parent, io_u);
1756 static inline bool gtod_reduce(struct thread_data *td)
1758 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1759 || td->o.gtod_reduce;
1762 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1763 struct io_completion_data *icd,
1764 const enum fio_ddir idx, unsigned int bytes)
1766 const int no_reduce = !gtod_reduce(td);
1767 unsigned long long llnsec = 0;
1772 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1776 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1778 if (!td->o.disable_lat) {
1779 unsigned long long tnsec;
1781 tnsec = ntime_since(&io_u->start_time, &icd->time);
1782 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1784 if (td->flags & TD_F_PROFILE_OPS) {
1785 struct prof_io_ops *ops = &td->prof_io_ops;
1788 icd->error = ops->io_u_lat(td, tnsec);
1791 if (td->o.max_latency && tnsec > td->o.max_latency)
1792 lat_fatal(td, icd, tnsec, td->o.max_latency);
1793 if (td->o.latency_target && tnsec > td->o.latency_target) {
1794 if (lat_target_failed(td))
1795 lat_fatal(td, icd, tnsec, td->o.latency_target);
1800 if (!td->o.disable_clat) {
1801 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1802 io_u_mark_latency(td, llnsec);
1805 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1806 add_bw_sample(td, io_u, bytes, llnsec);
1808 if (no_reduce && per_unit_log(td->iops_log))
1809 add_iops_sample(td, io_u, bytes);
1810 } else if (ddir_sync(idx) && !td->o.disable_clat)
1811 add_sync_clat_sample(&td->ts, llnsec);
1813 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1814 uint32_t *info = io_u_block_info(td, io_u);
1815 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1816 if (io_u->ddir == DDIR_TRIM) {
1817 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1818 BLOCK_INFO_TRIMS(*info) + 1);
1819 } else if (io_u->ddir == DDIR_WRITE) {
1820 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1827 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1828 uint64_t offset, unsigned int bytes)
1835 if (f->first_write == -1ULL || offset < f->first_write)
1836 f->first_write = offset;
1837 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1838 f->last_write = offset + bytes;
1840 if (!f->last_write_comp)
1843 idx = f->last_write_idx++;
1844 f->last_write_comp[idx] = offset;
1845 if (f->last_write_idx == td->o.iodepth)
1846 f->last_write_idx = 0;
1849 static bool should_account(struct thread_data *td)
1851 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1852 td->runstate == TD_VERIFYING);
1855 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1856 struct io_completion_data *icd)
1858 struct io_u *io_u = *io_u_ptr;
1859 enum fio_ddir ddir = io_u->ddir;
1860 struct fio_file *f = io_u->file;
1862 dprint_io_u(io_u, "complete");
1864 assert(io_u->flags & IO_U_F_FLIGHT);
1865 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1868 * Mark IO ok to verify
1872 * Remove errored entry from the verification list
1875 unlog_io_piece(td, io_u);
1877 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1882 if (ddir_sync(ddir)) {
1883 td->last_was_sync = true;
1885 f->first_write = -1ULL;
1886 f->last_write = -1ULL;
1888 if (should_account(td))
1889 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1893 td->last_was_sync = false;
1894 td->last_ddir = ddir;
1896 if (!io_u->error && ddir_rw(ddir)) {
1897 unsigned int bytes = io_u->buflen - io_u->resid;
1900 td->io_blocks[ddir]++;
1901 td->io_bytes[ddir] += bytes;
1903 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1904 td->this_io_blocks[ddir]++;
1905 td->this_io_bytes[ddir] += bytes;
1908 if (ddir == DDIR_WRITE)
1909 file_log_write_comp(td, f, io_u->offset, bytes);
1911 if (should_account(td))
1912 account_io_completion(td, io_u, icd, ddir, bytes);
1914 icd->bytes_done[ddir] += bytes;
1917 ret = io_u->end_io(td, io_u_ptr);
1919 if (ret && !icd->error)
1922 } else if (io_u->error) {
1923 icd->error = io_u->error;
1924 io_u_log_error(td, io_u);
1927 enum error_type_bit eb = td_error_type(ddir, icd->error);
1929 if (!td_non_fatal_error(td, eb, icd->error))
1933 * If there is a non_fatal error, then add to the error count
1934 * and clear all the errors.
1936 update_error_count(td, icd->error);
1944 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1949 if (!gtod_reduce(td))
1950 fio_gettime(&icd->time, NULL);
1955 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1956 icd->bytes_done[ddir] = 0;
1959 static void ios_completed(struct thread_data *td,
1960 struct io_completion_data *icd)
1965 for (i = 0; i < icd->nr; i++) {
1966 io_u = td->io_ops->event(td, i);
1968 io_completed(td, &io_u, icd);
1976 * Complete a single io_u for the sync engines.
1978 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1980 struct io_completion_data icd;
1983 init_icd(td, &icd, 1);
1984 io_completed(td, &io_u, &icd);
1990 td_verror(td, icd.error, "io_u_sync_complete");
1994 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1995 td->bytes_done[ddir] += icd.bytes_done[ddir];
2001 * Called to complete min_events number of io for the async engines.
2003 int io_u_queued_complete(struct thread_data *td, int min_evts)
2005 struct io_completion_data icd;
2006 struct timespec *tvp = NULL;
2008 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2010 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2014 else if (min_evts > td->cur_depth)
2015 min_evts = td->cur_depth;
2017 /* No worries, td_io_getevents fixes min and max if they are
2018 * set incorrectly */
2019 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2021 td_verror(td, -ret, "td_io_getevents");
2026 init_icd(td, &icd, ret);
2027 ios_completed(td, &icd);
2029 td_verror(td, icd.error, "io_u_queued_complete");
2033 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2034 td->bytes_done[ddir] += icd.bytes_done[ddir];
2040 * Call when io_u is really queued, to update the submission latency.
2042 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2044 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2045 unsigned long slat_time;
2047 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2052 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2058 * See if we should reuse the last seed, if dedupe is enabled
2060 static struct frand_state *get_buf_state(struct thread_data *td)
2064 if (!td->o.dedupe_percentage)
2065 return &td->buf_state;
2066 else if (td->o.dedupe_percentage == 100) {
2067 frand_copy(&td->buf_state_prev, &td->buf_state);
2068 return &td->buf_state;
2071 v = rand_between(&td->dedupe_state, 1, 100);
2073 if (v <= td->o.dedupe_percentage)
2074 return &td->buf_state_prev;
2076 return &td->buf_state;
2079 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2081 if (td->o.dedupe_percentage == 100)
2082 frand_copy(rs, &td->buf_state_prev);
2083 else if (rs == &td->buf_state)
2084 frand_copy(&td->buf_state_prev, rs);
2087 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2088 unsigned int max_bs)
2090 struct thread_options *o = &td->o;
2092 if (o->mem_type == MEM_CUDA_MALLOC)
2095 if (o->compress_percentage || o->dedupe_percentage) {
2096 unsigned int perc = td->o.compress_percentage;
2097 struct frand_state *rs;
2098 unsigned int left = max_bs;
2099 unsigned int this_write;
2102 rs = get_buf_state(td);
2104 min_write = min(min_write, left);
2107 this_write = min_not_zero(min_write,
2108 td->o.compress_chunk);
2110 fill_random_buf_percentage(rs, buf, perc,
2111 this_write, this_write,
2113 o->buffer_pattern_bytes);
2115 fill_random_buf(rs, buf, min_write);
2116 this_write = min_write;
2121 save_buf_state(td, rs);
2123 } else if (o->buffer_pattern_bytes)
2124 fill_buffer_pattern(td, buf, max_bs);
2125 else if (o->zero_buffers)
2126 memset(buf, 0, max_bs);
2128 fill_random_buf(get_buf_state(td), buf, max_bs);
2132 * "randomly" fill the buffer contents
2134 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2135 unsigned int min_write, unsigned int max_bs)
2137 io_u->buf_filled_len = 0;
2138 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2141 static int do_sync_file_range(const struct thread_data *td,
2144 off64_t offset, nbytes;
2146 offset = f->first_write;
2147 nbytes = f->last_write - f->first_write;
2152 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2155 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2159 if (io_u->ddir == DDIR_SYNC) {
2160 ret = fsync(io_u->file->fd);
2161 } else if (io_u->ddir == DDIR_DATASYNC) {
2162 #ifdef CONFIG_FDATASYNC
2163 ret = fdatasync(io_u->file->fd);
2165 ret = io_u->xfer_buflen;
2166 io_u->error = EINVAL;
2168 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2169 ret = do_sync_file_range(td, io_u->file);
2171 ret = io_u->xfer_buflen;
2172 io_u->error = EINVAL;
2176 io_u->error = errno;
2181 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2183 #ifndef FIO_HAVE_TRIM
2184 io_u->error = EINVAL;
2187 struct fio_file *f = io_u->file;
2190 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2192 return io_u->xfer_buflen;