13 #include "lib/axmap.h"
18 struct io_completion_data {
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timespec time; /* output */
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
30 static bool random_map_free(struct fio_file *f, const uint64_t block)
32 return !axmap_isset(f->io_axmap, block);
36 * Mark a given offset as used in the map.
38 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40 unsigned int min_bs = td->o.min_bs[io_u->ddir];
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
55 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
61 assert(ddir_rw(ddir));
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 * -> not for now since there is code assuming it could go either.
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
72 max_size = td->o.zone_range;
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
85 struct flist_head list;
89 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
90 enum fio_ddir ddir, uint64_t *b,
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
160 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
162 enum fio_ddir ddir, uint64_t *b)
164 struct zone_split_index *zsi;
165 uint64_t offset, lastb;
166 uint64_t send, stotal;
170 lastb = last_block(td, f, ddir);
174 if (!td->o.zone_split_nr[ddir]) {
176 return __get_next_rand_offset(td, f, ddir, b, lastb);
180 * Generate a value, v, between 1 and 100, both inclusive
182 v = rand32_between(&td->zone_state, 1, 100);
184 zsi = &td->zone_state_index[ddir][v - 1];
185 stotal = zsi->size_prev / td->o.ba[ddir];
186 send = zsi->size / td->o.ba[ddir];
189 * Should never happen
193 log_err("fio: bug in zoned generation\n");
197 } else if (send > lastb) {
199 * This happens if the user specifies ranges that exceed
200 * the file/device size. We can't handle that gracefully,
203 log_err("fio: zoned_abs sizes exceed file size\n");
208 * 'send' is some percentage below or equal to 100 that
209 * marks the end of the current IO range. 'stotal' marks
210 * the start, in percent.
217 lastb = send - stotal;
220 * Generate index from 0..send-of-lastb
222 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
226 * Add our start offset, if any
234 static int __get_next_rand_offset_zoned(struct thread_data *td,
235 struct fio_file *f, enum fio_ddir ddir,
238 unsigned int v, send, stotal;
239 uint64_t offset, lastb;
241 struct zone_split_index *zsi;
243 lastb = last_block(td, f, ddir);
247 if (!td->o.zone_split_nr[ddir]) {
249 return __get_next_rand_offset(td, f, ddir, b, lastb);
253 * Generate a value, v, between 1 and 100, both inclusive
255 v = rand32_between(&td->zone_state, 1, 100);
257 zsi = &td->zone_state_index[ddir][v - 1];
258 stotal = zsi->size_perc_prev;
259 send = zsi->size_perc;
262 * Should never happen
266 log_err("fio: bug in zoned generation\n");
273 * 'send' is some percentage below or equal to 100 that
274 * marks the end of the current IO range. 'stotal' marks
275 * the start, in percent.
278 offset = stotal * lastb / 100ULL;
282 lastb = lastb * (send - stotal) / 100ULL;
285 * Generate index from 0..send-of-lastb
287 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
291 * Add our start offset, if any
299 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
301 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
302 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
304 return r1->off - r2->off;
307 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
308 enum fio_ddir ddir, uint64_t *b)
310 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
313 lastb = last_block(td, f, ddir);
317 return __get_next_rand_offset(td, f, ddir, b, lastb);
318 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
319 return __get_next_rand_offset_zipf(td, f, ddir, b);
320 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
321 return __get_next_rand_offset_pareto(td, f, ddir, b);
322 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
323 return __get_next_rand_offset_gauss(td, f, ddir, b);
324 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
325 return __get_next_rand_offset_zoned(td, f, ddir, b);
326 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
327 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
329 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
334 * Sort the reads for a verify phase in batches of verifysort_nr, if
337 static inline bool should_sort_io(struct thread_data *td)
339 if (!td->o.verifysort_nr || !td->o.do_verify)
343 if (td->runstate != TD_VERIFYING)
345 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
346 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
352 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
356 if (td->o.perc_rand[ddir] == 100)
359 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
361 return v <= td->o.perc_rand[ddir];
364 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
365 enum fio_ddir ddir, uint64_t *b)
370 if (!should_sort_io(td))
371 return get_off_from_method(td, f, ddir, b);
373 if (!flist_empty(&td->next_rand_list)) {
375 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
382 for (i = 0; i < td->o.verifysort_nr; i++) {
383 r = malloc(sizeof(*r));
385 ret = get_off_from_method(td, f, ddir, &r->off);
391 flist_add(&r->list, &td->next_rand_list);
397 assert(!flist_empty(&td->next_rand_list));
398 flist_sort(NULL, &td->next_rand_list, flist_cmp);
402 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
404 struct thread_options *o = &td->o;
406 if (o->invalidate_cache && !o->odirect) {
409 ret = file_invalidate_cache(td, f);
413 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
414 enum fio_ddir ddir, uint64_t *b)
416 if (!get_next_rand_offset(td, f, ddir, b))
419 if (td->o.time_based ||
420 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
421 fio_file_reset(td, f);
422 if (!get_next_rand_offset(td, f, ddir, b))
424 loop_cache_invalidate(td, f);
427 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
428 f->file_name, (unsigned long long) f->last_pos[ddir],
429 (unsigned long long) f->real_file_size);
433 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
434 enum fio_ddir ddir, uint64_t *offset)
436 struct thread_options *o = &td->o;
438 assert(ddir_rw(ddir));
441 * If we reach the end for a time based run, reset us back to 0
442 * and invalidate the cache, if we need to.
444 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
446 f->last_pos[ddir] = f->file_offset;
447 loop_cache_invalidate(td, f);
450 if (f->last_pos[ddir] < f->real_file_size) {
453 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
454 if (f->real_file_size > f->io_size)
455 f->last_pos[ddir] = f->io_size;
457 f->last_pos[ddir] = f->real_file_size;
460 pos = f->last_pos[ddir] - f->file_offset;
461 if (pos && o->ddir_seq_add) {
462 pos += o->ddir_seq_add;
465 * If we reach beyond the end of the file
466 * with holed IO, wrap around to the
467 * beginning again. If we're doing backwards IO,
470 if (pos >= f->real_file_size) {
471 if (o->ddir_seq_add > 0)
472 pos = f->file_offset;
474 if (f->real_file_size > f->io_size)
477 pos = f->real_file_size;
479 pos += o->ddir_seq_add;
491 static int get_next_block(struct thread_data *td, struct io_u *io_u,
492 enum fio_ddir ddir, int rw_seq,
493 unsigned int *is_random)
495 struct fio_file *f = io_u->file;
499 assert(ddir_rw(ddir));
505 if (should_do_random(td, ddir)) {
506 ret = get_next_rand_block(td, f, ddir, &b);
510 io_u_set(td, io_u, IO_U_F_BUSY_OK);
511 ret = get_next_seq_offset(td, f, ddir, &offset);
513 ret = get_next_rand_block(td, f, ddir, &b);
517 ret = get_next_seq_offset(td, f, ddir, &offset);
520 io_u_set(td, io_u, IO_U_F_BUSY_OK);
523 if (td->o.rw_seq == RW_SEQ_SEQ) {
524 ret = get_next_seq_offset(td, f, ddir, &offset);
526 ret = get_next_rand_block(td, f, ddir, &b);
529 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
530 if (f->last_start[ddir] != -1ULL)
531 offset = f->last_start[ddir] - f->file_offset;
536 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
543 io_u->offset = offset;
545 io_u->offset = b * td->o.ba[ddir];
547 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
556 * For random io, generate a random new block and see if it's used. Repeat
557 * until we find a free one. For sequential io, just return the end of
558 * the last io issued.
560 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
561 unsigned int *is_random)
563 struct fio_file *f = io_u->file;
564 enum fio_ddir ddir = io_u->ddir;
567 assert(ddir_rw(ddir));
569 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
571 td->ddir_seq_nr = td->o.ddir_seq_nr;
574 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
577 if (io_u->offset >= f->io_size) {
578 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
579 (unsigned long long) io_u->offset,
580 (unsigned long long) f->io_size);
584 io_u->offset += f->file_offset;
585 if (io_u->offset >= f->real_file_size) {
586 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
587 (unsigned long long) io_u->offset,
588 (unsigned long long) f->real_file_size);
595 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
596 unsigned int *is_random)
598 if (td->flags & TD_F_PROFILE_OPS) {
599 struct prof_io_ops *ops = &td->prof_io_ops;
601 if (ops->fill_io_u_off)
602 return ops->fill_io_u_off(td, io_u, is_random);
605 return __get_next_offset(td, io_u, is_random);
608 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
611 struct fio_file *f = io_u->file;
613 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
616 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
617 unsigned int is_random)
619 int ddir = io_u->ddir;
620 unsigned int buflen = 0;
621 unsigned int minbs, maxbs;
622 uint64_t frand_max, r;
625 assert(ddir_rw(ddir));
627 if (td->o.bs_is_seq_rand)
628 ddir = is_random ? DDIR_WRITE: DDIR_READ;
630 minbs = td->o.min_bs[ddir];
631 maxbs = td->o.max_bs[ddir];
637 * If we can't satisfy the min block size from here, then fail
639 if (!io_u_fits(td, io_u, minbs))
642 frand_max = rand_max(&td->bsrange_state[ddir]);
644 r = __rand(&td->bsrange_state[ddir]);
646 if (!td->o.bssplit_nr[ddir]) {
647 buflen = 1 + (unsigned int) ((double) maxbs *
648 (r / (frand_max + 1.0)));
655 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
656 struct bssplit *bsp = &td->o.bssplit[ddir][i];
662 if ((r / perc <= frand_max / 100ULL) &&
663 io_u_fits(td, io_u, buflen))
668 power_2 = is_power_of_2(minbs);
669 if (!td->o.bs_unaligned && power_2)
670 buflen &= ~(minbs - 1);
671 else if (!td->o.bs_unaligned && !power_2)
672 buflen -= buflen % minbs;
673 } while (!io_u_fits(td, io_u, buflen));
678 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
679 unsigned int is_random)
681 if (td->flags & TD_F_PROFILE_OPS) {
682 struct prof_io_ops *ops = &td->prof_io_ops;
684 if (ops->fill_io_u_size)
685 return ops->fill_io_u_size(td, io_u, is_random);
688 return __get_next_buflen(td, io_u, is_random);
691 static void set_rwmix_bytes(struct thread_data *td)
696 * we do time or byte based switch. this is needed because
697 * buffered writes may issue a lot quicker than they complete,
698 * whereas reads do not.
700 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
701 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
704 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
708 v = rand32_between(&td->rwmix_state, 1, 100);
710 if (v <= td->o.rwmix[DDIR_READ])
716 int io_u_quiesce(struct thread_data *td)
721 * We are going to sleep, ensure that we flush anything pending as
722 * not to skew our latency numbers.
724 * Changed to only monitor 'in flight' requests here instead of the
725 * td->cur_depth, b/c td->cur_depth does not accurately represent
726 * io's that have been actually submitted to an async engine,
727 * and cur_depth is meaningless for sync engines.
729 if (td->io_u_queued || td->cur_depth) {
732 ret = td_io_commit(td);
735 while (td->io_u_in_flight) {
738 ret = io_u_queued_complete(td, 1);
743 if (td->flags & TD_F_REGROW_LOGS)
749 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
751 enum fio_ddir odir = ddir ^ 1;
755 assert(ddir_rw(ddir));
756 now = utime_since_now(&td->start);
759 * if rate_next_io_time is in the past, need to catch up to rate
761 if (td->rate_next_io_time[ddir] <= now)
765 * We are ahead of rate in this direction. See if we
768 if (td_rw(td) && td->o.rwmix[odir]) {
770 * Other direction is behind rate, switch
772 if (td->rate_next_io_time[odir] <= now)
776 * Both directions are ahead of rate. sleep the min
777 * switch if necissary
779 if (td->rate_next_io_time[ddir] <=
780 td->rate_next_io_time[odir]) {
781 usec = td->rate_next_io_time[ddir] - now;
783 usec = td->rate_next_io_time[odir] - now;
787 usec = td->rate_next_io_time[ddir] - now;
789 if (td->o.io_submit_mode == IO_MODE_INLINE)
792 usec = usec_sleep(td, usec);
798 * Return the data direction for the next io_u. If the job is a
799 * mixed read/write workload, check the rwmix cycle and switch if
802 static enum fio_ddir get_rw_ddir(struct thread_data *td)
807 * See if it's time to fsync/fdatasync/sync_file_range first,
808 * and if not then move on to check regular I/Os.
810 if (should_fsync(td)) {
811 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
812 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
815 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
816 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
817 return DDIR_DATASYNC;
819 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
820 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
821 return DDIR_SYNC_FILE_RANGE;
826 * Check if it's time to seed a new data direction.
828 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
830 * Put a top limit on how many bytes we do for
831 * one data direction, to avoid overflowing the
834 ddir = get_rand_ddir(td);
836 if (ddir != td->rwmix_ddir)
839 td->rwmix_ddir = ddir;
841 ddir = td->rwmix_ddir;
842 } else if (td_read(td))
844 else if (td_write(td))
846 else if (td_trim(td))
851 td->rwmix_ddir = rate_ddir(td, ddir);
852 return td->rwmix_ddir;
855 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
857 enum fio_ddir ddir = get_rw_ddir(td);
859 if (td_trimwrite(td)) {
860 struct fio_file *f = io_u->file;
861 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
867 io_u->ddir = io_u->acct_ddir = ddir;
869 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
870 td->o.barrier_blocks &&
871 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
872 td->io_issues[DDIR_WRITE])
873 io_u_set(td, io_u, IO_U_F_BARRIER);
876 void put_file_log(struct thread_data *td, struct fio_file *f)
878 unsigned int ret = put_file(td, f);
881 td_verror(td, ret, "file close");
884 void put_io_u(struct thread_data *td, struct io_u *io_u)
891 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
892 put_file_log(td, io_u->file);
895 io_u_set(td, io_u, IO_U_F_FREE);
897 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
899 assert(!(td->flags & TD_F_CHILD));
901 io_u_qpush(&td->io_u_freelist, io_u);
903 td_io_u_free_notify(td);
906 void clear_io_u(struct thread_data *td, struct io_u *io_u)
908 io_u_clear(td, io_u, IO_U_F_FLIGHT);
912 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
914 struct io_u *__io_u = *io_u;
915 enum fio_ddir ddir = acct_ddir(__io_u);
917 dprint(FD_IO, "requeue %p\n", __io_u);
924 io_u_set(td, __io_u, IO_U_F_FREE);
925 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
926 td->io_issues[ddir]--;
928 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
929 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
931 assert(!(td->flags & TD_F_CHILD));
934 io_u_rpush(&td->io_u_requeues, __io_u);
936 td_io_u_free_notify(td);
940 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
942 unsigned int is_random;
944 if (td_ioengine_flagged(td, FIO_NOIO))
947 set_rw_ddir(td, io_u);
950 * fsync() or fdatasync() or trim etc, we are done
952 if (!ddir_rw(io_u->ddir))
956 * See if it's time to switch to a new zone
958 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
959 struct fio_file *f = io_u->file;
962 f->file_offset += td->o.zone_range + td->o.zone_skip;
965 * Wrap from the beginning, if we exceed the file size
967 if (f->file_offset >= f->real_file_size)
968 f->file_offset = f->real_file_size - f->file_offset;
969 f->last_pos[io_u->ddir] = f->file_offset;
970 td->io_skip_bytes += td->o.zone_skip;
974 * No log, let the seq/rand engine retrieve the next buflen and
977 if (get_next_offset(td, io_u, &is_random)) {
978 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
982 io_u->buflen = get_next_buflen(td, io_u, is_random);
984 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
988 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
989 dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n",
991 dprint(FD_IO, " offset=%llu/buflen=%lu > %llu\n",
992 (unsigned long long) io_u->offset, io_u->buflen,
993 (unsigned long long) io_u->file->real_file_size);
998 * mark entry before potentially trimming io_u
1000 if (td_random(td) && file_randommap(td, io_u->file))
1001 mark_random_map(td, io_u);
1004 dprint_io_u(io_u, "fill_io_u");
1005 td->zone_bytes += io_u->buflen;
1009 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
1038 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1040 __io_u_mark_map(td->ts.io_u_submit, nr);
1041 td->ts.total_submit++;
1044 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1046 __io_u_mark_map(td->ts.io_u_complete, nr);
1047 td->ts.total_complete++;
1050 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1054 switch (td->cur_depth) {
1076 td->ts.io_u_map[idx] += nr;
1079 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1083 assert(nsec < 1000);
1116 assert(idx < FIO_IO_U_LAT_N_NR);
1117 td->ts.io_u_lat_n[idx]++;
1120 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1124 assert(usec < 1000 && usec >= 1);
1157 assert(idx < FIO_IO_U_LAT_U_NR);
1158 td->ts.io_u_lat_u[idx]++;
1161 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1204 assert(idx < FIO_IO_U_LAT_M_NR);
1205 td->ts.io_u_lat_m[idx]++;
1208 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1211 io_u_mark_lat_nsec(td, nsec);
1212 else if (nsec < 1000000)
1213 io_u_mark_lat_usec(td, nsec / 1000);
1215 io_u_mark_lat_msec(td, nsec / 1000000);
1218 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1220 unsigned long fileno;
1222 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1223 uint64_t frand_max = rand_max(&td->next_file_state);
1226 r = __rand(&td->next_file_state);
1227 return (unsigned int) ((double) td->o.nr_files
1228 * (r / (frand_max + 1.0)));
1231 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1232 fileno = zipf_next(&td->next_file_zipf);
1233 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1234 fileno = pareto_next(&td->next_file_zipf);
1235 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1236 fileno = gauss_next(&td->next_file_gauss);
1238 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1243 return fileno >> FIO_FSERVICE_SHIFT;
1247 * Get next file to service by choosing one at random
1249 static struct fio_file *get_next_file_rand(struct thread_data *td,
1250 enum fio_file_flags goodf,
1251 enum fio_file_flags badf)
1259 fno = __get_next_fileno_rand(td);
1262 if (fio_file_done(f))
1265 if (!fio_file_open(f)) {
1268 if (td->nr_open_files >= td->o.open_files)
1269 return ERR_PTR(-EBUSY);
1271 err = td_io_open_file(td, f);
1277 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1278 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1282 td_io_close_file(td, f);
1287 * Get next file to service by doing round robin between all available ones
1289 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1292 unsigned int old_next_file = td->next_file;
1298 f = td->files[td->next_file];
1301 if (td->next_file >= td->o.nr_files)
1304 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1305 if (fio_file_done(f)) {
1310 if (!fio_file_open(f)) {
1313 if (td->nr_open_files >= td->o.open_files)
1314 return ERR_PTR(-EBUSY);
1316 err = td_io_open_file(td, f);
1318 dprint(FD_FILE, "error %d on open of %s\n",
1326 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1328 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1332 td_io_close_file(td, f);
1335 } while (td->next_file != old_next_file);
1337 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1341 static struct fio_file *__get_next_file(struct thread_data *td)
1345 assert(td->o.nr_files <= td->files_index);
1347 if (td->nr_done_files >= td->o.nr_files) {
1348 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1349 " nr_files=%d\n", td->nr_open_files,
1355 f = td->file_service_file;
1356 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1357 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1359 if (td->file_service_left--)
1363 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1364 td->o.file_service_type == FIO_FSERVICE_SEQ)
1365 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1367 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1372 td->file_service_file = f;
1373 td->file_service_left = td->file_service_nr - 1;
1376 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1378 dprint(FD_FILE, "get_next_file: NULL\n");
1382 static struct fio_file *get_next_file(struct thread_data *td)
1384 if (td->flags & TD_F_PROFILE_OPS) {
1385 struct prof_io_ops *ops = &td->prof_io_ops;
1387 if (ops->get_next_file)
1388 return ops->get_next_file(td);
1391 return __get_next_file(td);
1394 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1399 f = get_next_file(td);
1400 if (IS_ERR_OR_NULL(f))
1406 if (!fill_io_u(td, io_u))
1409 put_file_log(td, f);
1410 td_io_close_file(td, f);
1412 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1413 fio_file_reset(td, f);
1415 fio_file_set_done(f);
1416 td->nr_done_files++;
1417 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1418 td->nr_done_files, td->o.nr_files);
1425 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1426 unsigned long tusec, unsigned long max_usec)
1429 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1430 td_verror(td, ETIMEDOUT, "max latency exceeded");
1431 icd->error = ETIMEDOUT;
1434 static void lat_new_cycle(struct thread_data *td)
1436 fio_gettime(&td->latency_ts, NULL);
1437 td->latency_ios = ddir_rw_sum(td->io_blocks);
1438 td->latency_failed = 0;
1442 * We had an IO outside the latency target. Reduce the queue depth. If we
1443 * are at QD=1, then it's time to give up.
1445 static bool __lat_target_failed(struct thread_data *td)
1447 if (td->latency_qd == 1)
1450 td->latency_qd_high = td->latency_qd;
1452 if (td->latency_qd == td->latency_qd_low)
1453 td->latency_qd_low--;
1455 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1457 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1460 * When we ramp QD down, quiesce existing IO to prevent
1461 * a storm of ramp downs due to pending higher depth.
1468 static bool lat_target_failed(struct thread_data *td)
1470 if (td->o.latency_percentile.u.f == 100.0)
1471 return __lat_target_failed(td);
1473 td->latency_failed++;
1477 void lat_target_init(struct thread_data *td)
1479 td->latency_end_run = 0;
1481 if (td->o.latency_target) {
1482 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1483 fio_gettime(&td->latency_ts, NULL);
1485 td->latency_qd_high = td->o.iodepth;
1486 td->latency_qd_low = 1;
1487 td->latency_ios = ddir_rw_sum(td->io_blocks);
1489 td->latency_qd = td->o.iodepth;
1492 void lat_target_reset(struct thread_data *td)
1494 if (!td->latency_end_run)
1495 lat_target_init(td);
1498 static void lat_target_success(struct thread_data *td)
1500 const unsigned int qd = td->latency_qd;
1501 struct thread_options *o = &td->o;
1503 td->latency_qd_low = td->latency_qd;
1506 * If we haven't failed yet, we double up to a failing value instead
1507 * of bisecting from highest possible queue depth. If we have set
1508 * a limit other than td->o.iodepth, bisect between that.
1510 if (td->latency_qd_high != o->iodepth)
1511 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1513 td->latency_qd *= 2;
1515 if (td->latency_qd > o->iodepth)
1516 td->latency_qd = o->iodepth;
1518 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1521 * Same as last one, we are done. Let it run a latency cycle, so
1522 * we get only the results from the targeted depth.
1524 if (td->latency_qd == qd) {
1525 if (td->latency_end_run) {
1526 dprint(FD_RATE, "We are done\n");
1529 dprint(FD_RATE, "Quiesce and final run\n");
1531 td->latency_end_run = 1;
1532 reset_all_stats(td);
1541 * Check if we can bump the queue depth
1543 void lat_target_check(struct thread_data *td)
1545 uint64_t usec_window;
1549 usec_window = utime_since_now(&td->latency_ts);
1550 if (usec_window < td->o.latency_window)
1553 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1554 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1555 success_ios *= 100.0;
1557 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1559 if (success_ios >= td->o.latency_percentile.u.f)
1560 lat_target_success(td);
1562 __lat_target_failed(td);
1566 * If latency target is enabled, we might be ramping up or down and not
1567 * using the full queue depth available.
1569 bool queue_full(const struct thread_data *td)
1571 const int qempty = io_u_qempty(&td->io_u_freelist);
1575 if (!td->o.latency_target)
1578 return td->cur_depth >= td->latency_qd;
1581 struct io_u *__get_io_u(struct thread_data *td)
1583 struct io_u *io_u = NULL;
1591 if (!io_u_rempty(&td->io_u_requeues))
1592 io_u = io_u_rpop(&td->io_u_requeues);
1593 else if (!queue_full(td)) {
1594 io_u = io_u_qpop(&td->io_u_freelist);
1599 io_u->end_io = NULL;
1603 assert(io_u->flags & IO_U_F_FREE);
1604 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1605 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1609 io_u->acct_ddir = -1;
1611 assert(!(td->flags & TD_F_CHILD));
1612 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1614 } else if (td_async_processing(td)) {
1616 * We ran out, wait for async verify threads to finish and
1619 assert(!(td->flags & TD_F_CHILD));
1620 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1628 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1630 if (!(td->flags & TD_F_TRIM_BACKLOG))
1633 if (td->trim_entries) {
1636 if (td->trim_batch) {
1639 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1640 td->last_ddir != DDIR_READ) {
1641 td->trim_batch = td->o.trim_batch;
1642 if (!td->trim_batch)
1643 td->trim_batch = td->o.trim_backlog;
1647 if (get_trim && get_next_trim(td, io_u))
1654 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1656 if (!(td->flags & TD_F_VER_BACKLOG))
1659 if (td->io_hist_len) {
1662 if (td->verify_batch)
1664 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1665 td->last_ddir != DDIR_READ) {
1666 td->verify_batch = td->o.verify_batch;
1667 if (!td->verify_batch)
1668 td->verify_batch = td->o.verify_backlog;
1672 if (get_verify && !get_next_verify(td, io_u)) {
1682 * Fill offset and start time into the buffer content, to prevent too
1683 * easy compressible data for simple de-dupe attempts. Do this for every
1684 * 512b block in the range, since that should be the smallest block size
1685 * we can expect from a device.
1687 static void small_content_scramble(struct io_u *io_u)
1689 unsigned int i, nr_blocks = io_u->buflen / 512;
1691 unsigned int offset;
1698 boffset = io_u->offset;
1699 io_u->buf_filled_len = 0;
1701 for (i = 0; i < nr_blocks; i++) {
1703 * Fill the byte offset into a "random" start offset of
1704 * the buffer, given by the product of the usec time
1705 * and the actual offset.
1707 offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511;
1708 offset &= ~(sizeof(uint64_t) - 1);
1709 if (offset >= 512 - sizeof(uint64_t))
1710 offset -= sizeof(uint64_t);
1711 memcpy(p + offset, &boffset, sizeof(boffset));
1713 end = p + 512 - sizeof(io_u->start_time);
1714 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1721 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1722 * etc. The returned io_u is fully ready to be prepped and submitted.
1724 struct io_u *get_io_u(struct thread_data *td)
1728 int do_scramble = 0;
1731 io_u = __get_io_u(td);
1733 dprint(FD_IO, "__get_io_u failed\n");
1737 if (check_get_verify(td, io_u))
1739 if (check_get_trim(td, io_u))
1743 * from a requeue, io_u already setup
1749 * If using an iolog, grab next piece if any available.
1751 if (td->flags & TD_F_READ_IOLOG) {
1752 if (read_iolog_get(td, io_u))
1754 } else if (set_io_u_file(td, io_u)) {
1756 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1762 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1766 assert(fio_file_open(f));
1768 if (ddir_rw(io_u->ddir)) {
1769 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1770 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1774 f->last_start[io_u->ddir] = io_u->offset;
1775 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1777 if (io_u->ddir == DDIR_WRITE) {
1778 if (td->flags & TD_F_REFILL_BUFFERS) {
1779 io_u_fill_buffer(td, io_u,
1780 td->o.min_bs[DDIR_WRITE],
1782 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1783 !(td->flags & TD_F_COMPRESS))
1785 if (td->flags & TD_F_VER_NONE) {
1786 populate_verify_io_u(td, io_u);
1789 } else if (io_u->ddir == DDIR_READ) {
1791 * Reset the buf_filled parameters so next time if the
1792 * buffer is used for writes it is refilled.
1794 io_u->buf_filled_len = 0;
1799 * Set io data pointers.
1801 io_u->xfer_buf = io_u->buf;
1802 io_u->xfer_buflen = io_u->buflen;
1806 if (!td_io_prep(td, io_u)) {
1807 if (!td->o.disable_lat)
1808 fio_gettime(&io_u->start_time, NULL);
1811 small_content_scramble(io_u);
1816 dprint(FD_IO, "get_io_u failed\n");
1818 return ERR_PTR(ret);
1821 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1823 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1825 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1828 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1829 io_u->file ? " on file " : "",
1830 io_u->file ? io_u->file->file_name : "",
1831 strerror(io_u->error),
1832 io_ddir_name(io_u->ddir),
1833 io_u->offset, io_u->xfer_buflen);
1835 if (td->io_ops->errdetails) {
1836 char *err = td->io_ops->errdetails(io_u);
1838 log_err("fio: %s\n", err);
1843 td_verror(td, io_u->error, "io_u error");
1846 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1848 __io_u_log_error(td, io_u);
1850 __io_u_log_error(td->parent, io_u);
1853 static inline bool gtod_reduce(struct thread_data *td)
1855 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1856 || td->o.gtod_reduce;
1859 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1860 struct io_completion_data *icd,
1861 const enum fio_ddir idx, unsigned int bytes)
1863 const int no_reduce = !gtod_reduce(td);
1864 unsigned long long llnsec = 0;
1869 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1873 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1875 if (!td->o.disable_lat) {
1876 unsigned long long tnsec;
1878 tnsec = ntime_since(&io_u->start_time, &icd->time);
1879 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1881 if (td->flags & TD_F_PROFILE_OPS) {
1882 struct prof_io_ops *ops = &td->prof_io_ops;
1885 icd->error = ops->io_u_lat(td, tnsec/1000);
1888 if (td->o.max_latency && tnsec/1000 > td->o.max_latency)
1889 lat_fatal(td, icd, tnsec/1000, td->o.max_latency);
1890 if (td->o.latency_target && tnsec/1000 > td->o.latency_target) {
1891 if (lat_target_failed(td))
1892 lat_fatal(td, icd, tnsec/1000, td->o.latency_target);
1897 if (!td->o.disable_clat) {
1898 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1899 io_u_mark_latency(td, llnsec);
1902 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1903 add_bw_sample(td, io_u, bytes, llnsec);
1905 if (no_reduce && per_unit_log(td->iops_log))
1906 add_iops_sample(td, io_u, bytes);
1909 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1910 uint32_t *info = io_u_block_info(td, io_u);
1911 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1912 if (io_u->ddir == DDIR_TRIM) {
1913 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1914 BLOCK_INFO_TRIMS(*info) + 1);
1915 } else if (io_u->ddir == DDIR_WRITE) {
1916 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1923 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1924 uint64_t offset, unsigned int bytes)
1931 if (f->first_write == -1ULL || offset < f->first_write)
1932 f->first_write = offset;
1933 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1934 f->last_write = offset + bytes;
1936 if (!f->last_write_comp)
1939 idx = f->last_write_idx++;
1940 f->last_write_comp[idx] = offset;
1941 if (f->last_write_idx == td->o.iodepth)
1942 f->last_write_idx = 0;
1945 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1946 struct io_completion_data *icd)
1948 struct io_u *io_u = *io_u_ptr;
1949 enum fio_ddir ddir = io_u->ddir;
1950 struct fio_file *f = io_u->file;
1952 dprint_io_u(io_u, "io complete");
1954 assert(io_u->flags & IO_U_F_FLIGHT);
1955 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1958 * Mark IO ok to verify
1962 * Remove errored entry from the verification list
1965 unlog_io_piece(td, io_u);
1967 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1972 if (ddir_sync(ddir)) {
1973 td->last_was_sync = 1;
1975 f->first_write = -1ULL;
1976 f->last_write = -1ULL;
1981 td->last_was_sync = 0;
1982 td->last_ddir = ddir;
1984 if (!io_u->error && ddir_rw(ddir)) {
1985 unsigned int bytes = io_u->buflen - io_u->resid;
1988 td->io_blocks[ddir]++;
1989 td->this_io_blocks[ddir]++;
1990 td->io_bytes[ddir] += bytes;
1992 if (!(io_u->flags & IO_U_F_VER_LIST))
1993 td->this_io_bytes[ddir] += bytes;
1995 if (ddir == DDIR_WRITE)
1996 file_log_write_comp(td, f, io_u->offset, bytes);
1998 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1999 td->runstate == TD_VERIFYING))
2000 account_io_completion(td, io_u, icd, ddir, bytes);
2002 icd->bytes_done[ddir] += bytes;
2005 ret = io_u->end_io(td, io_u_ptr);
2007 if (ret && !icd->error)
2010 } else if (io_u->error) {
2011 icd->error = io_u->error;
2012 io_u_log_error(td, io_u);
2015 enum error_type_bit eb = td_error_type(ddir, icd->error);
2017 if (!td_non_fatal_error(td, eb, icd->error))
2021 * If there is a non_fatal error, then add to the error count
2022 * and clear all the errors.
2024 update_error_count(td, icd->error);
2032 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2037 if (!gtod_reduce(td))
2038 fio_gettime(&icd->time, NULL);
2043 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2044 icd->bytes_done[ddir] = 0;
2047 static void ios_completed(struct thread_data *td,
2048 struct io_completion_data *icd)
2053 for (i = 0; i < icd->nr; i++) {
2054 io_u = td->io_ops->event(td, i);
2056 io_completed(td, &io_u, icd);
2064 * Complete a single io_u for the sync engines.
2066 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2068 struct io_completion_data icd;
2071 init_icd(td, &icd, 1);
2072 io_completed(td, &io_u, &icd);
2078 td_verror(td, icd.error, "io_u_sync_complete");
2082 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2083 td->bytes_done[ddir] += icd.bytes_done[ddir];
2089 * Called to complete min_events number of io for the async engines.
2091 int io_u_queued_complete(struct thread_data *td, int min_evts)
2093 struct io_completion_data icd;
2094 struct timespec *tvp = NULL;
2096 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2098 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2102 else if (min_evts > td->cur_depth)
2103 min_evts = td->cur_depth;
2105 /* No worries, td_io_getevents fixes min and max if they are
2106 * set incorrectly */
2107 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2109 td_verror(td, -ret, "td_io_getevents");
2114 init_icd(td, &icd, ret);
2115 ios_completed(td, &icd);
2117 td_verror(td, icd.error, "io_u_queued_complete");
2121 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2122 td->bytes_done[ddir] += icd.bytes_done[ddir];
2128 * Call when io_u is really queued, to update the submission latency.
2130 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2132 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2133 unsigned long slat_time;
2135 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2140 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2146 * See if we should reuse the last seed, if dedupe is enabled
2148 static struct frand_state *get_buf_state(struct thread_data *td)
2152 if (!td->o.dedupe_percentage)
2153 return &td->buf_state;
2154 else if (td->o.dedupe_percentage == 100) {
2155 frand_copy(&td->buf_state_prev, &td->buf_state);
2156 return &td->buf_state;
2159 v = rand32_between(&td->dedupe_state, 1, 100);
2161 if (v <= td->o.dedupe_percentage)
2162 return &td->buf_state_prev;
2164 return &td->buf_state;
2167 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2169 if (td->o.dedupe_percentage == 100)
2170 frand_copy(rs, &td->buf_state_prev);
2171 else if (rs == &td->buf_state)
2172 frand_copy(&td->buf_state_prev, rs);
2175 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2176 unsigned int max_bs)
2178 struct thread_options *o = &td->o;
2180 if (o->mem_type == MEM_CUDA_MALLOC)
2183 if (o->compress_percentage || o->dedupe_percentage) {
2184 unsigned int perc = td->o.compress_percentage;
2185 struct frand_state *rs;
2186 unsigned int left = max_bs;
2187 unsigned int this_write;
2190 rs = get_buf_state(td);
2192 min_write = min(min_write, left);
2195 this_write = min_not_zero(min_write,
2196 td->o.compress_chunk);
2198 fill_random_buf_percentage(rs, buf, perc,
2199 this_write, this_write,
2201 o->buffer_pattern_bytes);
2203 fill_random_buf(rs, buf, min_write);
2204 this_write = min_write;
2209 save_buf_state(td, rs);
2211 } else if (o->buffer_pattern_bytes)
2212 fill_buffer_pattern(td, buf, max_bs);
2213 else if (o->zero_buffers)
2214 memset(buf, 0, max_bs);
2216 fill_random_buf(get_buf_state(td), buf, max_bs);
2220 * "randomly" fill the buffer contents
2222 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2223 unsigned int min_write, unsigned int max_bs)
2225 io_u->buf_filled_len = 0;
2226 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2229 static int do_sync_file_range(const struct thread_data *td,
2232 off64_t offset, nbytes;
2234 offset = f->first_write;
2235 nbytes = f->last_write - f->first_write;
2240 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2243 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2247 if (io_u->ddir == DDIR_SYNC) {
2248 ret = fsync(io_u->file->fd);
2249 } else if (io_u->ddir == DDIR_DATASYNC) {
2250 #ifdef CONFIG_FDATASYNC
2251 ret = fdatasync(io_u->file->fd);
2253 ret = io_u->xfer_buflen;
2254 io_u->error = EINVAL;
2256 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2257 ret = do_sync_file_range(td, io_u->file);
2259 ret = io_u->xfer_buflen;
2260 io_u->error = EINVAL;
2264 io_u->error = errno;
2269 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2271 #ifndef FIO_HAVE_TRIM
2272 io_u->error = EINVAL;
2275 struct fio_file *f = io_u->file;
2278 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2280 return io_u->xfer_buflen;