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 lastb, send, stotal;
168 lastb = last_block(td, f, ddir);
172 if (!td->o.zone_split_nr[ddir]) {
174 return __get_next_rand_offset(td, f, ddir, b, lastb);
178 * Generate a value, v, between 1 and 100, both inclusive
180 v = rand32_between(&td->zone_state, 1, 100);
183 * Find our generated table. 'send' is the end block of this zone,
184 * 'stotal' is our start offset.
186 zsi = &td->zone_state_index[ddir][v - 1];
187 stotal = zsi->size_prev / td->o.ba[ddir];
188 send = zsi->size / td->o.ba[ddir];
191 * Should never happen
194 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
195 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 * Generate index from 0..send-stotal
210 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
217 static int __get_next_rand_offset_zoned(struct thread_data *td,
218 struct fio_file *f, enum fio_ddir ddir,
221 unsigned int v, send, stotal;
222 uint64_t offset, lastb;
223 struct zone_split_index *zsi;
225 lastb = last_block(td, f, ddir);
229 if (!td->o.zone_split_nr[ddir]) {
231 return __get_next_rand_offset(td, f, ddir, b, lastb);
235 * Generate a value, v, between 1 and 100, both inclusive
237 v = rand32_between(&td->zone_state, 1, 100);
239 zsi = &td->zone_state_index[ddir][v - 1];
240 stotal = zsi->size_perc_prev;
241 send = zsi->size_perc;
244 * Should never happen
247 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
248 log_err("fio: bug in zoned generation\n");
253 * 'send' is some percentage below or equal to 100 that
254 * marks the end of the current IO range. 'stotal' marks
255 * the start, in percent.
258 offset = stotal * lastb / 100ULL;
262 lastb = lastb * (send - stotal) / 100ULL;
265 * Generate index from 0..send-of-lastb
267 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
271 * Add our start offset, if any
279 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
281 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
282 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
284 return r1->off - r2->off;
287 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
288 enum fio_ddir ddir, uint64_t *b)
290 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
293 lastb = last_block(td, f, ddir);
297 return __get_next_rand_offset(td, f, ddir, b, lastb);
298 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
299 return __get_next_rand_offset_zipf(td, f, ddir, b);
300 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
301 return __get_next_rand_offset_pareto(td, f, ddir, b);
302 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
303 return __get_next_rand_offset_gauss(td, f, ddir, b);
304 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
305 return __get_next_rand_offset_zoned(td, f, ddir, b);
306 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
307 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
309 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
314 * Sort the reads for a verify phase in batches of verifysort_nr, if
317 static inline bool should_sort_io(struct thread_data *td)
319 if (!td->o.verifysort_nr || !td->o.do_verify)
323 if (td->runstate != TD_VERIFYING)
325 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
326 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
332 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
336 if (td->o.perc_rand[ddir] == 100)
339 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
341 return v <= td->o.perc_rand[ddir];
344 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
345 enum fio_ddir ddir, uint64_t *b)
350 if (!should_sort_io(td))
351 return get_off_from_method(td, f, ddir, b);
353 if (!flist_empty(&td->next_rand_list)) {
355 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
362 for (i = 0; i < td->o.verifysort_nr; i++) {
363 r = malloc(sizeof(*r));
365 ret = get_off_from_method(td, f, ddir, &r->off);
371 flist_add(&r->list, &td->next_rand_list);
377 assert(!flist_empty(&td->next_rand_list));
378 flist_sort(NULL, &td->next_rand_list, flist_cmp);
382 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
384 struct thread_options *o = &td->o;
386 if (o->invalidate_cache && !o->odirect) {
389 ret = file_invalidate_cache(td, f);
393 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
394 enum fio_ddir ddir, uint64_t *b)
396 if (!get_next_rand_offset(td, f, ddir, b))
399 if (td->o.time_based ||
400 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
401 fio_file_reset(td, f);
402 if (!get_next_rand_offset(td, f, ddir, b))
404 loop_cache_invalidate(td, f);
407 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
408 f->file_name, (unsigned long long) f->last_pos[ddir],
409 (unsigned long long) f->real_file_size);
413 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
414 enum fio_ddir ddir, uint64_t *offset)
416 struct thread_options *o = &td->o;
418 assert(ddir_rw(ddir));
421 * If we reach the end for a time based run, reset us back to 0
422 * and invalidate the cache, if we need to.
424 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
426 f->last_pos[ddir] = f->file_offset;
427 loop_cache_invalidate(td, f);
430 if (f->last_pos[ddir] < f->real_file_size) {
434 * Only rewind if we already hit the end
436 if (f->last_pos[ddir] == f->file_offset &&
437 f->file_offset && o->ddir_seq_add < 0) {
438 if (f->real_file_size > f->io_size)
439 f->last_pos[ddir] = f->io_size;
441 f->last_pos[ddir] = f->real_file_size;
444 pos = f->last_pos[ddir] - f->file_offset;
445 if (pos && o->ddir_seq_add) {
446 pos += o->ddir_seq_add;
449 * If we reach beyond the end of the file
450 * with holed IO, wrap around to the
451 * beginning again. If we're doing backwards IO,
454 if (pos >= f->real_file_size) {
455 if (o->ddir_seq_add > 0)
456 pos = f->file_offset;
458 if (f->real_file_size > f->io_size)
461 pos = f->real_file_size;
463 pos += o->ddir_seq_add;
475 static int get_next_block(struct thread_data *td, struct io_u *io_u,
476 enum fio_ddir ddir, int rw_seq,
479 struct fio_file *f = io_u->file;
483 assert(ddir_rw(ddir));
489 if (should_do_random(td, ddir)) {
490 ret = get_next_rand_block(td, f, ddir, &b);
494 io_u_set(td, io_u, IO_U_F_BUSY_OK);
495 ret = get_next_seq_offset(td, f, ddir, &offset);
497 ret = get_next_rand_block(td, f, ddir, &b);
501 ret = get_next_seq_offset(td, f, ddir, &offset);
504 io_u_set(td, io_u, IO_U_F_BUSY_OK);
507 if (td->o.rw_seq == RW_SEQ_SEQ) {
508 ret = get_next_seq_offset(td, f, ddir, &offset);
510 ret = get_next_rand_block(td, f, ddir, &b);
513 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
514 if (f->last_start[ddir] != -1ULL)
515 offset = f->last_start[ddir] - f->file_offset;
520 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
527 io_u->offset = offset;
529 io_u->offset = b * td->o.ba[ddir];
531 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
540 * For random io, generate a random new block and see if it's used. Repeat
541 * until we find a free one. For sequential io, just return the end of
542 * the last io issued.
544 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
547 struct fio_file *f = io_u->file;
548 enum fio_ddir ddir = io_u->ddir;
551 assert(ddir_rw(ddir));
553 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
555 td->ddir_seq_nr = td->o.ddir_seq_nr;
558 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
561 if (io_u->offset >= f->io_size) {
562 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
563 (unsigned long long) io_u->offset,
564 (unsigned long long) f->io_size);
568 io_u->offset += f->file_offset;
569 if (io_u->offset >= f->real_file_size) {
570 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
571 (unsigned long long) io_u->offset,
572 (unsigned long long) f->real_file_size);
579 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
582 struct fio_file *f = io_u->file;
584 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
587 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
590 int ddir = io_u->ddir;
591 unsigned int buflen = 0;
592 unsigned int minbs, maxbs;
593 uint64_t frand_max, r;
596 assert(ddir_rw(ddir));
598 if (td->o.bs_is_seq_rand)
599 ddir = is_random ? DDIR_WRITE : DDIR_READ;
601 minbs = td->o.min_bs[ddir];
602 maxbs = td->o.max_bs[ddir];
608 * If we can't satisfy the min block size from here, then fail
610 if (!io_u_fits(td, io_u, minbs))
613 frand_max = rand_max(&td->bsrange_state[ddir]);
615 r = __rand(&td->bsrange_state[ddir]);
617 if (!td->o.bssplit_nr[ddir]) {
618 buflen = 1 + (unsigned int) ((double) maxbs *
619 (r / (frand_max + 1.0)));
626 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
627 struct bssplit *bsp = &td->o.bssplit[ddir][i];
633 if ((r / perc <= frand_max / 100ULL) &&
634 io_u_fits(td, io_u, buflen))
639 power_2 = is_power_of_2(minbs);
640 if (!td->o.bs_unaligned && power_2)
641 buflen &= ~(minbs - 1);
642 else if (!td->o.bs_unaligned && !power_2)
643 buflen -= buflen % minbs;
644 } while (!io_u_fits(td, io_u, buflen));
649 static void set_rwmix_bytes(struct thread_data *td)
654 * we do time or byte based switch. this is needed because
655 * buffered writes may issue a lot quicker than they complete,
656 * whereas reads do not.
658 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
659 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
662 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
666 v = rand32_between(&td->rwmix_state, 1, 100);
668 if (v <= td->o.rwmix[DDIR_READ])
674 int io_u_quiesce(struct thread_data *td)
679 * We are going to sleep, ensure that we flush anything pending as
680 * not to skew our latency numbers.
682 * Changed to only monitor 'in flight' requests here instead of the
683 * td->cur_depth, b/c td->cur_depth does not accurately represent
684 * io's that have been actually submitted to an async engine,
685 * and cur_depth is meaningless for sync engines.
687 if (td->io_u_queued || td->cur_depth) {
690 ret = td_io_commit(td);
693 while (td->io_u_in_flight) {
696 ret = io_u_queued_complete(td, 1);
701 if (td->flags & TD_F_REGROW_LOGS)
707 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
709 enum fio_ddir odir = ddir ^ 1;
713 assert(ddir_rw(ddir));
714 now = utime_since_now(&td->start);
717 * if rate_next_io_time is in the past, need to catch up to rate
719 if (td->rate_next_io_time[ddir] <= now)
723 * We are ahead of rate in this direction. See if we
726 if (td_rw(td) && td->o.rwmix[odir]) {
728 * Other direction is behind rate, switch
730 if (td->rate_next_io_time[odir] <= now)
734 * Both directions are ahead of rate. sleep the min,
735 * switch if necessary
737 if (td->rate_next_io_time[ddir] <=
738 td->rate_next_io_time[odir]) {
739 usec = td->rate_next_io_time[ddir] - now;
741 usec = td->rate_next_io_time[odir] - now;
745 usec = td->rate_next_io_time[ddir] - now;
747 if (td->o.io_submit_mode == IO_MODE_INLINE)
750 usec_sleep(td, usec);
755 * Return the data direction for the next io_u. If the job is a
756 * mixed read/write workload, check the rwmix cycle and switch if
759 static enum fio_ddir get_rw_ddir(struct thread_data *td)
764 * See if it's time to fsync/fdatasync/sync_file_range first,
765 * and if not then move on to check regular I/Os.
767 if (should_fsync(td)) {
768 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
769 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
772 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
773 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
774 return DDIR_DATASYNC;
776 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
777 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
778 return DDIR_SYNC_FILE_RANGE;
783 * Check if it's time to seed a new data direction.
785 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
787 * Put a top limit on how many bytes we do for
788 * one data direction, to avoid overflowing the
791 ddir = get_rand_ddir(td);
793 if (ddir != td->rwmix_ddir)
796 td->rwmix_ddir = ddir;
798 ddir = td->rwmix_ddir;
799 } else if (td_read(td))
801 else if (td_write(td))
803 else if (td_trim(td))
808 td->rwmix_ddir = rate_ddir(td, ddir);
809 return td->rwmix_ddir;
812 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
814 enum fio_ddir ddir = get_rw_ddir(td);
816 if (td_trimwrite(td)) {
817 struct fio_file *f = io_u->file;
818 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
824 io_u->ddir = io_u->acct_ddir = ddir;
826 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
827 td->o.barrier_blocks &&
828 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
829 td->io_issues[DDIR_WRITE])
830 io_u_set(td, io_u, IO_U_F_BARRIER);
833 void put_file_log(struct thread_data *td, struct fio_file *f)
835 unsigned int ret = put_file(td, f);
838 td_verror(td, ret, "file close");
841 void put_io_u(struct thread_data *td, struct io_u *io_u)
848 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
849 put_file_log(td, io_u->file);
852 io_u_set(td, io_u, IO_U_F_FREE);
854 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
856 assert(!(td->flags & TD_F_CHILD));
858 io_u_qpush(&td->io_u_freelist, io_u);
859 td_io_u_free_notify(td);
863 void clear_io_u(struct thread_data *td, struct io_u *io_u)
865 io_u_clear(td, io_u, IO_U_F_FLIGHT);
869 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
871 struct io_u *__io_u = *io_u;
872 enum fio_ddir ddir = acct_ddir(__io_u);
874 dprint(FD_IO, "requeue %p\n", __io_u);
881 io_u_set(td, __io_u, IO_U_F_FREE);
882 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
883 td->io_issues[ddir]--;
885 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
886 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
888 assert(!(td->flags & TD_F_CHILD));
891 io_u_rpush(&td->io_u_requeues, __io_u);
892 td_io_u_free_notify(td);
897 static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
899 struct fio_file *f = io_u->file;
902 * See if it's time to switch to a new zone
904 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
906 f->file_offset += td->o.zone_range + td->o.zone_skip;
909 * Wrap from the beginning, if we exceed the file size
911 if (f->file_offset >= f->real_file_size)
912 f->file_offset = f->real_file_size - f->file_offset;
913 f->last_pos[io_u->ddir] = f->file_offset;
914 td->io_skip_bytes += td->o.zone_skip;
918 * If zone_size > zone_range, then maintain the same zone until
919 * zone_bytes >= zone_size.
921 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
922 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
923 f->file_offset, f->last_pos[io_u->ddir]);
924 f->last_pos[io_u->ddir] = f->file_offset;
928 * For random: if 'norandommap' is not set and zone_size > zone_range,
929 * map needs to be reset as it's done with zone_range everytime.
931 if ((td->zone_bytes % td->o.zone_range) == 0) {
932 fio_file_reset(td, f);
936 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
940 if (td_ioengine_flagged(td, FIO_NOIO))
943 set_rw_ddir(td, io_u);
946 * fsync() or fdatasync() or trim etc, we are done
948 if (!ddir_rw(io_u->ddir))
952 * When file is zoned zone_range is always positive
954 if (td->o.zone_range) {
955 __fill_io_u_zone(td, io_u);
959 * No log, let the seq/rand engine retrieve the next buflen and
962 if (get_next_offset(td, io_u, &is_random)) {
963 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
967 io_u->buflen = get_next_buflen(td, io_u, is_random);
969 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
973 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
974 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
976 (unsigned long long) io_u->offset, io_u->buflen,
977 (unsigned long long) io_u->file->real_file_size);
982 * mark entry before potentially trimming io_u
984 if (td_random(td) && file_randommap(td, io_u->file))
985 mark_random_map(td, io_u);
988 dprint_io_u(io_u, "fill");
989 td->zone_bytes += io_u->buflen;
993 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
1022 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1024 __io_u_mark_map(td->ts.io_u_submit, nr);
1025 td->ts.total_submit++;
1028 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1030 __io_u_mark_map(td->ts.io_u_complete, nr);
1031 td->ts.total_complete++;
1034 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1038 switch (td->cur_depth) {
1060 td->ts.io_u_map[idx] += nr;
1063 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1067 assert(nsec < 1000);
1100 assert(idx < FIO_IO_U_LAT_N_NR);
1101 td->ts.io_u_lat_n[idx]++;
1104 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1108 assert(usec < 1000 && usec >= 1);
1141 assert(idx < FIO_IO_U_LAT_U_NR);
1142 td->ts.io_u_lat_u[idx]++;
1145 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1188 assert(idx < FIO_IO_U_LAT_M_NR);
1189 td->ts.io_u_lat_m[idx]++;
1192 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1195 io_u_mark_lat_nsec(td, nsec);
1196 else if (nsec < 1000000)
1197 io_u_mark_lat_usec(td, nsec / 1000);
1199 io_u_mark_lat_msec(td, nsec / 1000000);
1202 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1204 unsigned long fileno;
1206 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1207 uint64_t frand_max = rand_max(&td->next_file_state);
1210 r = __rand(&td->next_file_state);
1211 return (unsigned int) ((double) td->o.nr_files
1212 * (r / (frand_max + 1.0)));
1215 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1216 fileno = zipf_next(&td->next_file_zipf);
1217 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1218 fileno = pareto_next(&td->next_file_zipf);
1219 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1220 fileno = gauss_next(&td->next_file_gauss);
1222 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1227 return fileno >> FIO_FSERVICE_SHIFT;
1231 * Get next file to service by choosing one at random
1233 static struct fio_file *get_next_file_rand(struct thread_data *td,
1234 enum fio_file_flags goodf,
1235 enum fio_file_flags badf)
1243 fno = __get_next_fileno_rand(td);
1246 if (fio_file_done(f))
1249 if (!fio_file_open(f)) {
1252 if (td->nr_open_files >= td->o.open_files)
1253 return ERR_PTR(-EBUSY);
1255 err = td_io_open_file(td, f);
1261 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1262 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1266 td_io_close_file(td, f);
1271 * Get next file to service by doing round robin between all available ones
1273 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1276 unsigned int old_next_file = td->next_file;
1282 f = td->files[td->next_file];
1285 if (td->next_file >= td->o.nr_files)
1288 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1289 if (fio_file_done(f)) {
1294 if (!fio_file_open(f)) {
1297 if (td->nr_open_files >= td->o.open_files)
1298 return ERR_PTR(-EBUSY);
1300 err = td_io_open_file(td, f);
1302 dprint(FD_FILE, "error %d on open of %s\n",
1310 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1312 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1316 td_io_close_file(td, f);
1319 } while (td->next_file != old_next_file);
1321 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1325 static struct fio_file *__get_next_file(struct thread_data *td)
1329 assert(td->o.nr_files <= td->files_index);
1331 if (td->nr_done_files >= td->o.nr_files) {
1332 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1333 " nr_files=%d\n", td->nr_open_files,
1339 f = td->file_service_file;
1340 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1341 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1343 if (td->file_service_left--)
1347 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1348 td->o.file_service_type == FIO_FSERVICE_SEQ)
1349 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1351 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1356 td->file_service_file = f;
1357 td->file_service_left = td->file_service_nr - 1;
1360 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1362 dprint(FD_FILE, "get_next_file: NULL\n");
1366 static struct fio_file *get_next_file(struct thread_data *td)
1368 return __get_next_file(td);
1371 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1376 f = get_next_file(td);
1377 if (IS_ERR_OR_NULL(f))
1383 if (!fill_io_u(td, io_u))
1386 put_file_log(td, f);
1387 td_io_close_file(td, f);
1389 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1390 fio_file_reset(td, f);
1392 fio_file_set_done(f);
1393 td->nr_done_files++;
1394 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1395 td->nr_done_files, td->o.nr_files);
1402 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1403 unsigned long long tnsec, unsigned long long max_nsec)
1406 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1407 td_verror(td, ETIMEDOUT, "max latency exceeded");
1408 icd->error = ETIMEDOUT;
1411 static void lat_new_cycle(struct thread_data *td)
1413 fio_gettime(&td->latency_ts, NULL);
1414 td->latency_ios = ddir_rw_sum(td->io_blocks);
1415 td->latency_failed = 0;
1419 * We had an IO outside the latency target. Reduce the queue depth. If we
1420 * are at QD=1, then it's time to give up.
1422 static bool __lat_target_failed(struct thread_data *td)
1424 if (td->latency_qd == 1)
1427 td->latency_qd_high = td->latency_qd;
1429 if (td->latency_qd == td->latency_qd_low)
1430 td->latency_qd_low--;
1432 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1434 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1437 * When we ramp QD down, quiesce existing IO to prevent
1438 * a storm of ramp downs due to pending higher depth.
1445 static bool lat_target_failed(struct thread_data *td)
1447 if (td->o.latency_percentile.u.f == 100.0)
1448 return __lat_target_failed(td);
1450 td->latency_failed++;
1454 void lat_target_init(struct thread_data *td)
1456 td->latency_end_run = 0;
1458 if (td->o.latency_target) {
1459 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1460 fio_gettime(&td->latency_ts, NULL);
1462 td->latency_qd_high = td->o.iodepth;
1463 td->latency_qd_low = 1;
1464 td->latency_ios = ddir_rw_sum(td->io_blocks);
1466 td->latency_qd = td->o.iodepth;
1469 void lat_target_reset(struct thread_data *td)
1471 if (!td->latency_end_run)
1472 lat_target_init(td);
1475 static void lat_target_success(struct thread_data *td)
1477 const unsigned int qd = td->latency_qd;
1478 struct thread_options *o = &td->o;
1480 td->latency_qd_low = td->latency_qd;
1483 * If we haven't failed yet, we double up to a failing value instead
1484 * of bisecting from highest possible queue depth. If we have set
1485 * a limit other than td->o.iodepth, bisect between that.
1487 if (td->latency_qd_high != o->iodepth)
1488 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1490 td->latency_qd *= 2;
1492 if (td->latency_qd > o->iodepth)
1493 td->latency_qd = o->iodepth;
1495 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1498 * Same as last one, we are done. Let it run a latency cycle, so
1499 * we get only the results from the targeted depth.
1501 if (td->latency_qd == qd) {
1502 if (td->latency_end_run) {
1503 dprint(FD_RATE, "We are done\n");
1506 dprint(FD_RATE, "Quiesce and final run\n");
1508 td->latency_end_run = 1;
1509 reset_all_stats(td);
1518 * Check if we can bump the queue depth
1520 void lat_target_check(struct thread_data *td)
1522 uint64_t usec_window;
1526 usec_window = utime_since_now(&td->latency_ts);
1527 if (usec_window < td->o.latency_window)
1530 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1531 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1532 success_ios *= 100.0;
1534 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1536 if (success_ios >= td->o.latency_percentile.u.f)
1537 lat_target_success(td);
1539 __lat_target_failed(td);
1543 * If latency target is enabled, we might be ramping up or down and not
1544 * using the full queue depth available.
1546 bool queue_full(const struct thread_data *td)
1548 const int qempty = io_u_qempty(&td->io_u_freelist);
1552 if (!td->o.latency_target)
1555 return td->cur_depth >= td->latency_qd;
1558 struct io_u *__get_io_u(struct thread_data *td)
1560 struct io_u *io_u = NULL;
1569 if (!io_u_rempty(&td->io_u_requeues))
1570 io_u = io_u_rpop(&td->io_u_requeues);
1571 else if (!queue_full(td)) {
1572 io_u = io_u_qpop(&td->io_u_freelist);
1577 io_u->end_io = NULL;
1581 assert(io_u->flags & IO_U_F_FREE);
1582 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1583 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1587 io_u->acct_ddir = -1;
1589 assert(!(td->flags & TD_F_CHILD));
1590 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1592 } else if (td_async_processing(td)) {
1594 * We ran out, wait for async verify threads to finish and
1597 assert(!(td->flags & TD_F_CHILD));
1598 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1607 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1609 if (!(td->flags & TD_F_TRIM_BACKLOG))
1611 if (!td->trim_entries)
1614 if (td->trim_batch) {
1616 if (get_next_trim(td, io_u))
1618 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1619 td->last_ddir != DDIR_READ) {
1620 td->trim_batch = td->o.trim_batch;
1621 if (!td->trim_batch)
1622 td->trim_batch = td->o.trim_backlog;
1623 if (get_next_trim(td, io_u))
1630 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1632 if (!(td->flags & TD_F_VER_BACKLOG))
1635 if (td->io_hist_len) {
1638 if (td->verify_batch)
1640 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1641 td->last_ddir != DDIR_READ) {
1642 td->verify_batch = td->o.verify_batch;
1643 if (!td->verify_batch)
1644 td->verify_batch = td->o.verify_backlog;
1648 if (get_verify && !get_next_verify(td, io_u)) {
1658 * Fill offset and start time into the buffer content, to prevent too
1659 * easy compressible data for simple de-dupe attempts. Do this for every
1660 * 512b block in the range, since that should be the smallest block size
1661 * we can expect from a device.
1663 static void small_content_scramble(struct io_u *io_u)
1665 unsigned int i, nr_blocks = io_u->buflen >> 9;
1666 unsigned int offset;
1667 uint64_t boffset, *iptr;
1674 boffset = io_u->offset;
1676 if (io_u->buf_filled_len)
1677 io_u->buf_filled_len = 0;
1680 * Generate random index between 0..7. We do chunks of 512b, if
1681 * we assume a cacheline is 64 bytes, then we have 8 of those.
1682 * Scramble content within the blocks in the same cacheline to
1685 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1687 for (i = 0; i < nr_blocks; i++) {
1689 * Fill offset into start of cacheline, time into end
1692 iptr = (void *) p + (offset << 6);
1695 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1696 iptr[0] = io_u->start_time.tv_sec;
1697 iptr[1] = io_u->start_time.tv_nsec;
1705 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1706 * etc. The returned io_u is fully ready to be prepped and submitted.
1708 struct io_u *get_io_u(struct thread_data *td)
1712 int do_scramble = 0;
1715 io_u = __get_io_u(td);
1717 dprint(FD_IO, "__get_io_u failed\n");
1721 if (check_get_verify(td, io_u))
1723 if (check_get_trim(td, io_u))
1727 * from a requeue, io_u already setup
1733 * If using an iolog, grab next piece if any available.
1735 if (td->flags & TD_F_READ_IOLOG) {
1736 if (read_iolog_get(td, io_u))
1738 } else if (set_io_u_file(td, io_u)) {
1740 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1746 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1750 assert(fio_file_open(f));
1752 if (ddir_rw(io_u->ddir)) {
1753 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1754 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1758 f->last_start[io_u->ddir] = io_u->offset;
1759 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1761 if (io_u->ddir == DDIR_WRITE) {
1762 if (td->flags & TD_F_REFILL_BUFFERS) {
1763 io_u_fill_buffer(td, io_u,
1764 td->o.min_bs[DDIR_WRITE],
1766 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1767 !(td->flags & TD_F_COMPRESS))
1769 if (td->flags & TD_F_VER_NONE) {
1770 populate_verify_io_u(td, io_u);
1773 } else if (io_u->ddir == DDIR_READ) {
1775 * Reset the buf_filled parameters so next time if the
1776 * buffer is used for writes it is refilled.
1778 io_u->buf_filled_len = 0;
1783 * Set io data pointers.
1785 io_u->xfer_buf = io_u->buf;
1786 io_u->xfer_buflen = io_u->buflen;
1790 if (!td_io_prep(td, io_u)) {
1791 if (!td->o.disable_lat)
1792 fio_gettime(&io_u->start_time, NULL);
1795 small_content_scramble(io_u);
1800 dprint(FD_IO, "get_io_u failed\n");
1802 return ERR_PTR(ret);
1805 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1807 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1809 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1812 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1813 io_u->file ? " on file " : "",
1814 io_u->file ? io_u->file->file_name : "",
1815 strerror(io_u->error),
1816 io_ddir_name(io_u->ddir),
1817 io_u->offset, io_u->xfer_buflen);
1819 if (td->io_ops->errdetails) {
1820 char *err = td->io_ops->errdetails(io_u);
1822 log_err("fio: %s\n", err);
1827 td_verror(td, io_u->error, "io_u error");
1830 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1832 __io_u_log_error(td, io_u);
1834 __io_u_log_error(td->parent, io_u);
1837 static inline bool gtod_reduce(struct thread_data *td)
1839 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1840 || td->o.gtod_reduce;
1843 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1844 struct io_completion_data *icd,
1845 const enum fio_ddir idx, unsigned int bytes)
1847 const int no_reduce = !gtod_reduce(td);
1848 unsigned long long llnsec = 0;
1853 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1857 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1859 if (!td->o.disable_lat) {
1860 unsigned long long tnsec;
1862 tnsec = ntime_since(&io_u->start_time, &icd->time);
1863 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1865 if (td->flags & TD_F_PROFILE_OPS) {
1866 struct prof_io_ops *ops = &td->prof_io_ops;
1869 icd->error = ops->io_u_lat(td, tnsec);
1872 if (td->o.max_latency && tnsec > td->o.max_latency)
1873 lat_fatal(td, icd, tnsec, td->o.max_latency);
1874 if (td->o.latency_target && tnsec > td->o.latency_target) {
1875 if (lat_target_failed(td))
1876 lat_fatal(td, icd, tnsec, td->o.latency_target);
1881 if (!td->o.disable_clat) {
1882 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1883 io_u_mark_latency(td, llnsec);
1886 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1887 add_bw_sample(td, io_u, bytes, llnsec);
1889 if (no_reduce && per_unit_log(td->iops_log))
1890 add_iops_sample(td, io_u, bytes);
1891 } else if (ddir_sync(idx) && !td->o.disable_clat)
1892 add_sync_clat_sample(&td->ts, llnsec);
1894 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1895 uint32_t *info = io_u_block_info(td, io_u);
1896 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1897 if (io_u->ddir == DDIR_TRIM) {
1898 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1899 BLOCK_INFO_TRIMS(*info) + 1);
1900 } else if (io_u->ddir == DDIR_WRITE) {
1901 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1908 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1909 uint64_t offset, unsigned int bytes)
1916 if (f->first_write == -1ULL || offset < f->first_write)
1917 f->first_write = offset;
1918 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1919 f->last_write = offset + bytes;
1921 if (!f->last_write_comp)
1924 idx = f->last_write_idx++;
1925 f->last_write_comp[idx] = offset;
1926 if (f->last_write_idx == td->o.iodepth)
1927 f->last_write_idx = 0;
1930 static bool should_account(struct thread_data *td)
1932 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1933 td->runstate == TD_VERIFYING);
1936 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1937 struct io_completion_data *icd)
1939 struct io_u *io_u = *io_u_ptr;
1940 enum fio_ddir ddir = io_u->ddir;
1941 struct fio_file *f = io_u->file;
1943 dprint_io_u(io_u, "complete");
1945 assert(io_u->flags & IO_U_F_FLIGHT);
1946 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1949 * Mark IO ok to verify
1953 * Remove errored entry from the verification list
1956 unlog_io_piece(td, io_u);
1958 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1963 if (ddir_sync(ddir)) {
1964 td->last_was_sync = true;
1966 f->first_write = -1ULL;
1967 f->last_write = -1ULL;
1969 if (should_account(td))
1970 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1974 td->last_was_sync = false;
1975 td->last_ddir = ddir;
1977 if (!io_u->error && ddir_rw(ddir)) {
1978 unsigned int bytes = io_u->buflen - io_u->resid;
1981 td->io_blocks[ddir]++;
1982 td->io_bytes[ddir] += bytes;
1984 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1985 td->this_io_blocks[ddir]++;
1986 td->this_io_bytes[ddir] += bytes;
1989 if (ddir == DDIR_WRITE)
1990 file_log_write_comp(td, f, io_u->offset, bytes);
1992 if (should_account(td))
1993 account_io_completion(td, io_u, icd, ddir, bytes);
1995 icd->bytes_done[ddir] += bytes;
1998 ret = io_u->end_io(td, io_u_ptr);
2000 if (ret && !icd->error)
2003 } else if (io_u->error) {
2004 icd->error = io_u->error;
2005 io_u_log_error(td, io_u);
2008 enum error_type_bit eb = td_error_type(ddir, icd->error);
2010 if (!td_non_fatal_error(td, eb, icd->error))
2014 * If there is a non_fatal error, then add to the error count
2015 * and clear all the errors.
2017 update_error_count(td, icd->error);
2025 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2030 if (!gtod_reduce(td))
2031 fio_gettime(&icd->time, NULL);
2036 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2037 icd->bytes_done[ddir] = 0;
2040 static void ios_completed(struct thread_data *td,
2041 struct io_completion_data *icd)
2046 for (i = 0; i < icd->nr; i++) {
2047 io_u = td->io_ops->event(td, i);
2049 io_completed(td, &io_u, icd);
2057 * Complete a single io_u for the sync engines.
2059 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2061 struct io_completion_data icd;
2064 init_icd(td, &icd, 1);
2065 io_completed(td, &io_u, &icd);
2071 td_verror(td, icd.error, "io_u_sync_complete");
2075 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2076 td->bytes_done[ddir] += icd.bytes_done[ddir];
2082 * Called to complete min_events number of io for the async engines.
2084 int io_u_queued_complete(struct thread_data *td, int min_evts)
2086 struct io_completion_data icd;
2087 struct timespec *tvp = NULL;
2089 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2091 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2095 else if (min_evts > td->cur_depth)
2096 min_evts = td->cur_depth;
2098 /* No worries, td_io_getevents fixes min and max if they are
2099 * set incorrectly */
2100 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2102 td_verror(td, -ret, "td_io_getevents");
2107 init_icd(td, &icd, ret);
2108 ios_completed(td, &icd);
2110 td_verror(td, icd.error, "io_u_queued_complete");
2114 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2115 td->bytes_done[ddir] += icd.bytes_done[ddir];
2121 * Call when io_u is really queued, to update the submission latency.
2123 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2125 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2126 unsigned long slat_time;
2128 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2133 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2139 * See if we should reuse the last seed, if dedupe is enabled
2141 static struct frand_state *get_buf_state(struct thread_data *td)
2145 if (!td->o.dedupe_percentage)
2146 return &td->buf_state;
2147 else if (td->o.dedupe_percentage == 100) {
2148 frand_copy(&td->buf_state_prev, &td->buf_state);
2149 return &td->buf_state;
2152 v = rand32_between(&td->dedupe_state, 1, 100);
2154 if (v <= td->o.dedupe_percentage)
2155 return &td->buf_state_prev;
2157 return &td->buf_state;
2160 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2162 if (td->o.dedupe_percentage == 100)
2163 frand_copy(rs, &td->buf_state_prev);
2164 else if (rs == &td->buf_state)
2165 frand_copy(&td->buf_state_prev, rs);
2168 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2169 unsigned int max_bs)
2171 struct thread_options *o = &td->o;
2173 if (o->mem_type == MEM_CUDA_MALLOC)
2176 if (o->compress_percentage || o->dedupe_percentage) {
2177 unsigned int perc = td->o.compress_percentage;
2178 struct frand_state *rs;
2179 unsigned int left = max_bs;
2180 unsigned int this_write;
2183 rs = get_buf_state(td);
2185 min_write = min(min_write, left);
2188 this_write = min_not_zero(min_write,
2189 td->o.compress_chunk);
2191 fill_random_buf_percentage(rs, buf, perc,
2192 this_write, this_write,
2194 o->buffer_pattern_bytes);
2196 fill_random_buf(rs, buf, min_write);
2197 this_write = min_write;
2202 save_buf_state(td, rs);
2204 } else if (o->buffer_pattern_bytes)
2205 fill_buffer_pattern(td, buf, max_bs);
2206 else if (o->zero_buffers)
2207 memset(buf, 0, max_bs);
2209 fill_random_buf(get_buf_state(td), buf, max_bs);
2213 * "randomly" fill the buffer contents
2215 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2216 unsigned int min_write, unsigned int max_bs)
2218 io_u->buf_filled_len = 0;
2219 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2222 static int do_sync_file_range(const struct thread_data *td,
2225 off64_t offset, nbytes;
2227 offset = f->first_write;
2228 nbytes = f->last_write - f->first_write;
2233 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2236 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2240 if (io_u->ddir == DDIR_SYNC) {
2241 ret = fsync(io_u->file->fd);
2242 } else if (io_u->ddir == DDIR_DATASYNC) {
2243 #ifdef CONFIG_FDATASYNC
2244 ret = fdatasync(io_u->file->fd);
2246 ret = io_u->xfer_buflen;
2247 io_u->error = EINVAL;
2249 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2250 ret = do_sync_file_range(td, io_u->file);
2252 ret = io_u->xfer_buflen;
2253 io_u->error = EINVAL;
2257 io_u->error = errno;
2262 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2264 #ifndef FIO_HAVE_TRIM
2265 io_u->error = EINVAL;
2268 struct fio_file *f = io_u->file;
2271 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2273 return io_u->xfer_buflen;