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) {
433 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
434 if (f->real_file_size > f->io_size)
435 f->last_pos[ddir] = f->io_size;
437 f->last_pos[ddir] = f->real_file_size;
440 pos = f->last_pos[ddir] - f->file_offset;
441 if (pos && o->ddir_seq_add) {
442 pos += o->ddir_seq_add;
445 * If we reach beyond the end of the file
446 * with holed IO, wrap around to the
447 * beginning again. If we're doing backwards IO,
450 if (pos >= f->real_file_size) {
451 if (o->ddir_seq_add > 0)
452 pos = f->file_offset;
454 if (f->real_file_size > f->io_size)
457 pos = f->real_file_size;
459 pos += o->ddir_seq_add;
471 static int get_next_block(struct thread_data *td, struct io_u *io_u,
472 enum fio_ddir ddir, int rw_seq,
473 unsigned int *is_random)
475 struct fio_file *f = io_u->file;
479 assert(ddir_rw(ddir));
485 if (should_do_random(td, ddir)) {
486 ret = get_next_rand_block(td, f, ddir, &b);
490 io_u_set(td, io_u, IO_U_F_BUSY_OK);
491 ret = get_next_seq_offset(td, f, ddir, &offset);
493 ret = get_next_rand_block(td, f, ddir, &b);
497 ret = get_next_seq_offset(td, f, ddir, &offset);
500 io_u_set(td, io_u, IO_U_F_BUSY_OK);
503 if (td->o.rw_seq == RW_SEQ_SEQ) {
504 ret = get_next_seq_offset(td, f, ddir, &offset);
506 ret = get_next_rand_block(td, f, ddir, &b);
509 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
510 if (f->last_start[ddir] != -1ULL)
511 offset = f->last_start[ddir] - f->file_offset;
516 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
523 io_u->offset = offset;
525 io_u->offset = b * td->o.ba[ddir];
527 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
536 * For random io, generate a random new block and see if it's used. Repeat
537 * until we find a free one. For sequential io, just return the end of
538 * the last io issued.
540 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
541 unsigned int *is_random)
543 struct fio_file *f = io_u->file;
544 enum fio_ddir ddir = io_u->ddir;
547 assert(ddir_rw(ddir));
549 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
551 td->ddir_seq_nr = td->o.ddir_seq_nr;
554 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
557 if (io_u->offset >= f->io_size) {
558 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
559 (unsigned long long) io_u->offset,
560 (unsigned long long) f->io_size);
564 io_u->offset += f->file_offset;
565 if (io_u->offset >= f->real_file_size) {
566 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
567 (unsigned long long) io_u->offset,
568 (unsigned long long) f->real_file_size);
575 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
576 unsigned int *is_random)
578 if (td->flags & TD_F_PROFILE_OPS) {
579 struct prof_io_ops *ops = &td->prof_io_ops;
581 if (ops->fill_io_u_off)
582 return ops->fill_io_u_off(td, io_u, is_random);
585 return __get_next_offset(td, io_u, is_random);
588 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
591 struct fio_file *f = io_u->file;
593 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
596 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
597 unsigned int is_random)
599 int ddir = io_u->ddir;
600 unsigned int buflen = 0;
601 unsigned int minbs, maxbs;
602 uint64_t frand_max, r;
605 assert(ddir_rw(ddir));
607 if (td->o.bs_is_seq_rand)
608 ddir = is_random ? DDIR_WRITE: DDIR_READ;
610 minbs = td->o.min_bs[ddir];
611 maxbs = td->o.max_bs[ddir];
617 * If we can't satisfy the min block size from here, then fail
619 if (!io_u_fits(td, io_u, minbs))
622 frand_max = rand_max(&td->bsrange_state[ddir]);
624 r = __rand(&td->bsrange_state[ddir]);
626 if (!td->o.bssplit_nr[ddir]) {
627 buflen = 1 + (unsigned int) ((double) maxbs *
628 (r / (frand_max + 1.0)));
635 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
636 struct bssplit *bsp = &td->o.bssplit[ddir][i];
642 if ((r / perc <= frand_max / 100ULL) &&
643 io_u_fits(td, io_u, buflen))
648 power_2 = is_power_of_2(minbs);
649 if (!td->o.bs_unaligned && power_2)
650 buflen &= ~(minbs - 1);
651 else if (!td->o.bs_unaligned && !power_2)
652 buflen -= buflen % minbs;
653 } while (!io_u_fits(td, io_u, buflen));
658 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
659 unsigned int is_random)
661 if (td->flags & TD_F_PROFILE_OPS) {
662 struct prof_io_ops *ops = &td->prof_io_ops;
664 if (ops->fill_io_u_size)
665 return ops->fill_io_u_size(td, io_u, is_random);
668 return __get_next_buflen(td, io_u, is_random);
671 static void set_rwmix_bytes(struct thread_data *td)
676 * we do time or byte based switch. this is needed because
677 * buffered writes may issue a lot quicker than they complete,
678 * whereas reads do not.
680 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
681 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
684 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
688 v = rand32_between(&td->rwmix_state, 1, 100);
690 if (v <= td->o.rwmix[DDIR_READ])
696 int io_u_quiesce(struct thread_data *td)
701 * We are going to sleep, ensure that we flush anything pending as
702 * not to skew our latency numbers.
704 * Changed to only monitor 'in flight' requests here instead of the
705 * td->cur_depth, b/c td->cur_depth does not accurately represent
706 * io's that have been actually submitted to an async engine,
707 * and cur_depth is meaningless for sync engines.
709 if (td->io_u_queued || td->cur_depth) {
712 ret = td_io_commit(td);
715 while (td->io_u_in_flight) {
718 ret = io_u_queued_complete(td, 1);
723 if (td->flags & TD_F_REGROW_LOGS)
729 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
731 enum fio_ddir odir = ddir ^ 1;
735 assert(ddir_rw(ddir));
736 now = utime_since_now(&td->start);
739 * if rate_next_io_time is in the past, need to catch up to rate
741 if (td->rate_next_io_time[ddir] <= now)
745 * We are ahead of rate in this direction. See if we
748 if (td_rw(td) && td->o.rwmix[odir]) {
750 * Other direction is behind rate, switch
752 if (td->rate_next_io_time[odir] <= now)
756 * Both directions are ahead of rate. sleep the min,
757 * switch if necessary
759 if (td->rate_next_io_time[ddir] <=
760 td->rate_next_io_time[odir]) {
761 usec = td->rate_next_io_time[ddir] - now;
763 usec = td->rate_next_io_time[odir] - now;
767 usec = td->rate_next_io_time[ddir] - now;
769 if (td->o.io_submit_mode == IO_MODE_INLINE)
772 usec_sleep(td, usec);
777 * Return the data direction for the next io_u. If the job is a
778 * mixed read/write workload, check the rwmix cycle and switch if
781 static enum fio_ddir get_rw_ddir(struct thread_data *td)
786 * See if it's time to fsync/fdatasync/sync_file_range first,
787 * and if not then move on to check regular I/Os.
789 if (should_fsync(td)) {
790 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
791 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
794 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
795 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
796 return DDIR_DATASYNC;
798 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
799 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
800 return DDIR_SYNC_FILE_RANGE;
805 * Check if it's time to seed a new data direction.
807 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
809 * Put a top limit on how many bytes we do for
810 * one data direction, to avoid overflowing the
813 ddir = get_rand_ddir(td);
815 if (ddir != td->rwmix_ddir)
818 td->rwmix_ddir = ddir;
820 ddir = td->rwmix_ddir;
821 } else if (td_read(td))
823 else if (td_write(td))
825 else if (td_trim(td))
830 td->rwmix_ddir = rate_ddir(td, ddir);
831 return td->rwmix_ddir;
834 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
836 enum fio_ddir ddir = get_rw_ddir(td);
838 if (td_trimwrite(td)) {
839 struct fio_file *f = io_u->file;
840 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
846 io_u->ddir = io_u->acct_ddir = ddir;
848 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
849 td->o.barrier_blocks &&
850 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
851 td->io_issues[DDIR_WRITE])
852 io_u_set(td, io_u, IO_U_F_BARRIER);
855 void put_file_log(struct thread_data *td, struct fio_file *f)
857 unsigned int ret = put_file(td, f);
860 td_verror(td, ret, "file close");
863 void put_io_u(struct thread_data *td, struct io_u *io_u)
870 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
871 put_file_log(td, io_u->file);
874 io_u_set(td, io_u, IO_U_F_FREE);
876 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
878 assert(!(td->flags & TD_F_CHILD));
880 io_u_qpush(&td->io_u_freelist, io_u);
882 td_io_u_free_notify(td);
885 void clear_io_u(struct thread_data *td, struct io_u *io_u)
887 io_u_clear(td, io_u, IO_U_F_FLIGHT);
891 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
893 struct io_u *__io_u = *io_u;
894 enum fio_ddir ddir = acct_ddir(__io_u);
896 dprint(FD_IO, "requeue %p\n", __io_u);
903 io_u_set(td, __io_u, IO_U_F_FREE);
904 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
905 td->io_issues[ddir]--;
907 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
908 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
910 assert(!(td->flags & TD_F_CHILD));
913 io_u_rpush(&td->io_u_requeues, __io_u);
915 td_io_u_free_notify(td);
919 static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
921 struct fio_file *f = io_u->file;
924 * See if it's time to switch to a new zone
926 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
928 f->file_offset += td->o.zone_range + td->o.zone_skip;
931 * Wrap from the beginning, if we exceed the file size
933 if (f->file_offset >= f->real_file_size)
934 f->file_offset = f->real_file_size - f->file_offset;
935 f->last_pos[io_u->ddir] = f->file_offset;
936 td->io_skip_bytes += td->o.zone_skip;
940 * If zone_size > zone_range, then maintain the same zone until
941 * zone_bytes >= zone_size.
943 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
944 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
945 f->file_offset, f->last_pos[io_u->ddir]);
946 f->last_pos[io_u->ddir] = f->file_offset;
950 * For random: if 'norandommap' is not set and zone_size > zone_range,
951 * map needs to be reset as it's done with zone_range everytime.
953 if ((td->zone_bytes % td->o.zone_range) == 0) {
954 fio_file_reset(td, f);
958 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
960 unsigned int is_random;
962 if (td_ioengine_flagged(td, FIO_NOIO))
965 set_rw_ddir(td, io_u);
968 * fsync() or fdatasync() or trim etc, we are done
970 if (!ddir_rw(io_u->ddir))
974 * When file is zoned zone_range is always positive
976 if (td->o.zone_range) {
977 __fill_io_u_zone(td, io_u);
981 * No log, let the seq/rand engine retrieve the next buflen and
984 if (get_next_offset(td, io_u, &is_random)) {
985 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
989 io_u->buflen = get_next_buflen(td, io_u, is_random);
991 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
995 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
996 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
998 (unsigned long long) io_u->offset, io_u->buflen,
999 (unsigned long long) io_u->file->real_file_size);
1004 * mark entry before potentially trimming io_u
1006 if (td_random(td) && file_randommap(td, io_u->file))
1007 mark_random_map(td, io_u);
1010 dprint_io_u(io_u, "fill");
1011 td->zone_bytes += io_u->buflen;
1015 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
1044 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1046 __io_u_mark_map(td->ts.io_u_submit, nr);
1047 td->ts.total_submit++;
1050 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1052 __io_u_mark_map(td->ts.io_u_complete, nr);
1053 td->ts.total_complete++;
1056 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1060 switch (td->cur_depth) {
1082 td->ts.io_u_map[idx] += nr;
1085 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1089 assert(nsec < 1000);
1122 assert(idx < FIO_IO_U_LAT_N_NR);
1123 td->ts.io_u_lat_n[idx]++;
1126 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1130 assert(usec < 1000 && usec >= 1);
1163 assert(idx < FIO_IO_U_LAT_U_NR);
1164 td->ts.io_u_lat_u[idx]++;
1167 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1210 assert(idx < FIO_IO_U_LAT_M_NR);
1211 td->ts.io_u_lat_m[idx]++;
1214 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1217 io_u_mark_lat_nsec(td, nsec);
1218 else if (nsec < 1000000)
1219 io_u_mark_lat_usec(td, nsec / 1000);
1221 io_u_mark_lat_msec(td, nsec / 1000000);
1224 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1226 unsigned long fileno;
1228 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1229 uint64_t frand_max = rand_max(&td->next_file_state);
1232 r = __rand(&td->next_file_state);
1233 return (unsigned int) ((double) td->o.nr_files
1234 * (r / (frand_max + 1.0)));
1237 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1238 fileno = zipf_next(&td->next_file_zipf);
1239 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1240 fileno = pareto_next(&td->next_file_zipf);
1241 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1242 fileno = gauss_next(&td->next_file_gauss);
1244 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1249 return fileno >> FIO_FSERVICE_SHIFT;
1253 * Get next file to service by choosing one at random
1255 static struct fio_file *get_next_file_rand(struct thread_data *td,
1256 enum fio_file_flags goodf,
1257 enum fio_file_flags badf)
1265 fno = __get_next_fileno_rand(td);
1268 if (fio_file_done(f))
1271 if (!fio_file_open(f)) {
1274 if (td->nr_open_files >= td->o.open_files)
1275 return ERR_PTR(-EBUSY);
1277 err = td_io_open_file(td, f);
1283 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1284 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1288 td_io_close_file(td, f);
1293 * Get next file to service by doing round robin between all available ones
1295 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1298 unsigned int old_next_file = td->next_file;
1304 f = td->files[td->next_file];
1307 if (td->next_file >= td->o.nr_files)
1310 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1311 if (fio_file_done(f)) {
1316 if (!fio_file_open(f)) {
1319 if (td->nr_open_files >= td->o.open_files)
1320 return ERR_PTR(-EBUSY);
1322 err = td_io_open_file(td, f);
1324 dprint(FD_FILE, "error %d on open of %s\n",
1332 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1334 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1338 td_io_close_file(td, f);
1341 } while (td->next_file != old_next_file);
1343 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1347 static struct fio_file *__get_next_file(struct thread_data *td)
1351 assert(td->o.nr_files <= td->files_index);
1353 if (td->nr_done_files >= td->o.nr_files) {
1354 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1355 " nr_files=%d\n", td->nr_open_files,
1361 f = td->file_service_file;
1362 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1363 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1365 if (td->file_service_left--)
1369 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1370 td->o.file_service_type == FIO_FSERVICE_SEQ)
1371 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1373 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1378 td->file_service_file = f;
1379 td->file_service_left = td->file_service_nr - 1;
1382 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1384 dprint(FD_FILE, "get_next_file: NULL\n");
1388 static struct fio_file *get_next_file(struct thread_data *td)
1390 if (td->flags & TD_F_PROFILE_OPS) {
1391 struct prof_io_ops *ops = &td->prof_io_ops;
1393 if (ops->get_next_file)
1394 return ops->get_next_file(td);
1397 return __get_next_file(td);
1400 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1405 f = get_next_file(td);
1406 if (IS_ERR_OR_NULL(f))
1412 if (!fill_io_u(td, io_u))
1415 put_file_log(td, f);
1416 td_io_close_file(td, f);
1418 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1419 fio_file_reset(td, f);
1421 fio_file_set_done(f);
1422 td->nr_done_files++;
1423 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1424 td->nr_done_files, td->o.nr_files);
1431 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1432 unsigned long long tnsec, unsigned long long max_nsec)
1435 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1436 td_verror(td, ETIMEDOUT, "max latency exceeded");
1437 icd->error = ETIMEDOUT;
1440 static void lat_new_cycle(struct thread_data *td)
1442 fio_gettime(&td->latency_ts, NULL);
1443 td->latency_ios = ddir_rw_sum(td->io_blocks);
1444 td->latency_failed = 0;
1448 * We had an IO outside the latency target. Reduce the queue depth. If we
1449 * are at QD=1, then it's time to give up.
1451 static bool __lat_target_failed(struct thread_data *td)
1453 if (td->latency_qd == 1)
1456 td->latency_qd_high = td->latency_qd;
1458 if (td->latency_qd == td->latency_qd_low)
1459 td->latency_qd_low--;
1461 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1463 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1466 * When we ramp QD down, quiesce existing IO to prevent
1467 * a storm of ramp downs due to pending higher depth.
1474 static bool lat_target_failed(struct thread_data *td)
1476 if (td->o.latency_percentile.u.f == 100.0)
1477 return __lat_target_failed(td);
1479 td->latency_failed++;
1483 void lat_target_init(struct thread_data *td)
1485 td->latency_end_run = 0;
1487 if (td->o.latency_target) {
1488 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1489 fio_gettime(&td->latency_ts, NULL);
1491 td->latency_qd_high = td->o.iodepth;
1492 td->latency_qd_low = 1;
1493 td->latency_ios = ddir_rw_sum(td->io_blocks);
1495 td->latency_qd = td->o.iodepth;
1498 void lat_target_reset(struct thread_data *td)
1500 if (!td->latency_end_run)
1501 lat_target_init(td);
1504 static void lat_target_success(struct thread_data *td)
1506 const unsigned int qd = td->latency_qd;
1507 struct thread_options *o = &td->o;
1509 td->latency_qd_low = td->latency_qd;
1512 * If we haven't failed yet, we double up to a failing value instead
1513 * of bisecting from highest possible queue depth. If we have set
1514 * a limit other than td->o.iodepth, bisect between that.
1516 if (td->latency_qd_high != o->iodepth)
1517 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1519 td->latency_qd *= 2;
1521 if (td->latency_qd > o->iodepth)
1522 td->latency_qd = o->iodepth;
1524 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1527 * Same as last one, we are done. Let it run a latency cycle, so
1528 * we get only the results from the targeted depth.
1530 if (td->latency_qd == qd) {
1531 if (td->latency_end_run) {
1532 dprint(FD_RATE, "We are done\n");
1535 dprint(FD_RATE, "Quiesce and final run\n");
1537 td->latency_end_run = 1;
1538 reset_all_stats(td);
1547 * Check if we can bump the queue depth
1549 void lat_target_check(struct thread_data *td)
1551 uint64_t usec_window;
1555 usec_window = utime_since_now(&td->latency_ts);
1556 if (usec_window < td->o.latency_window)
1559 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1560 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1561 success_ios *= 100.0;
1563 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1565 if (success_ios >= td->o.latency_percentile.u.f)
1566 lat_target_success(td);
1568 __lat_target_failed(td);
1572 * If latency target is enabled, we might be ramping up or down and not
1573 * using the full queue depth available.
1575 bool queue_full(const struct thread_data *td)
1577 const int qempty = io_u_qempty(&td->io_u_freelist);
1581 if (!td->o.latency_target)
1584 return td->cur_depth >= td->latency_qd;
1587 struct io_u *__get_io_u(struct thread_data *td)
1589 struct io_u *io_u = NULL;
1597 if (!io_u_rempty(&td->io_u_requeues))
1598 io_u = io_u_rpop(&td->io_u_requeues);
1599 else if (!queue_full(td)) {
1600 io_u = io_u_qpop(&td->io_u_freelist);
1605 io_u->end_io = NULL;
1609 assert(io_u->flags & IO_U_F_FREE);
1610 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1611 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1615 io_u->acct_ddir = -1;
1617 assert(!(td->flags & TD_F_CHILD));
1618 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1620 } else if (td_async_processing(td)) {
1622 * We ran out, wait for async verify threads to finish and
1625 assert(!(td->flags & TD_F_CHILD));
1626 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1634 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1636 if (!(td->flags & TD_F_TRIM_BACKLOG))
1638 if (!td->trim_entries)
1641 if (td->trim_batch) {
1643 if (get_next_trim(td, io_u))
1645 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1646 td->last_ddir != DDIR_READ) {
1647 td->trim_batch = td->o.trim_batch;
1648 if (!td->trim_batch)
1649 td->trim_batch = td->o.trim_backlog;
1650 if (get_next_trim(td, io_u))
1657 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1659 if (!(td->flags & TD_F_VER_BACKLOG))
1662 if (td->io_hist_len) {
1665 if (td->verify_batch)
1667 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1668 td->last_ddir != DDIR_READ) {
1669 td->verify_batch = td->o.verify_batch;
1670 if (!td->verify_batch)
1671 td->verify_batch = td->o.verify_backlog;
1675 if (get_verify && !get_next_verify(td, io_u)) {
1685 * Fill offset and start time into the buffer content, to prevent too
1686 * easy compressible data for simple de-dupe attempts. Do this for every
1687 * 512b block in the range, since that should be the smallest block size
1688 * we can expect from a device.
1690 static void small_content_scramble(struct io_u *io_u)
1692 unsigned int i, nr_blocks = io_u->buflen >> 9;
1693 unsigned int offset;
1694 uint64_t boffset, *iptr;
1701 boffset = io_u->offset;
1703 if (io_u->buf_filled_len)
1704 io_u->buf_filled_len = 0;
1707 * Generate random index between 0..7. We do chunks of 512b, if
1708 * we assume a cacheline is 64 bytes, then we have 8 of those.
1709 * Scramble content within the blocks in the same cacheline to
1712 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1714 for (i = 0; i < nr_blocks; i++) {
1716 * Fill offset into start of cacheline, time into end
1719 iptr = (void *) p + (offset << 6);
1722 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1723 iptr[0] = io_u->start_time.tv_sec;
1724 iptr[1] = io_u->start_time.tv_nsec;
1732 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1733 * etc. The returned io_u is fully ready to be prepped and submitted.
1735 struct io_u *get_io_u(struct thread_data *td)
1739 int do_scramble = 0;
1742 io_u = __get_io_u(td);
1744 dprint(FD_IO, "__get_io_u failed\n");
1748 if (check_get_verify(td, io_u))
1750 if (check_get_trim(td, io_u))
1754 * from a requeue, io_u already setup
1760 * If using an iolog, grab next piece if any available.
1762 if (td->flags & TD_F_READ_IOLOG) {
1763 if (read_iolog_get(td, io_u))
1765 } else if (set_io_u_file(td, io_u)) {
1767 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1773 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1777 assert(fio_file_open(f));
1779 if (ddir_rw(io_u->ddir)) {
1780 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1781 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1785 f->last_start[io_u->ddir] = io_u->offset;
1786 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1788 if (io_u->ddir == DDIR_WRITE) {
1789 if (td->flags & TD_F_REFILL_BUFFERS) {
1790 io_u_fill_buffer(td, io_u,
1791 td->o.min_bs[DDIR_WRITE],
1793 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1794 !(td->flags & TD_F_COMPRESS))
1796 if (td->flags & TD_F_VER_NONE) {
1797 populate_verify_io_u(td, io_u);
1800 } else if (io_u->ddir == DDIR_READ) {
1802 * Reset the buf_filled parameters so next time if the
1803 * buffer is used for writes it is refilled.
1805 io_u->buf_filled_len = 0;
1810 * Set io data pointers.
1812 io_u->xfer_buf = io_u->buf;
1813 io_u->xfer_buflen = io_u->buflen;
1817 if (!td_io_prep(td, io_u)) {
1818 if (!td->o.disable_lat)
1819 fio_gettime(&io_u->start_time, NULL);
1822 small_content_scramble(io_u);
1827 dprint(FD_IO, "get_io_u failed\n");
1829 return ERR_PTR(ret);
1832 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1834 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1836 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1839 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1840 io_u->file ? " on file " : "",
1841 io_u->file ? io_u->file->file_name : "",
1842 strerror(io_u->error),
1843 io_ddir_name(io_u->ddir),
1844 io_u->offset, io_u->xfer_buflen);
1846 if (td->io_ops->errdetails) {
1847 char *err = td->io_ops->errdetails(io_u);
1849 log_err("fio: %s\n", err);
1854 td_verror(td, io_u->error, "io_u error");
1857 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1859 __io_u_log_error(td, io_u);
1861 __io_u_log_error(td->parent, io_u);
1864 static inline bool gtod_reduce(struct thread_data *td)
1866 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1867 || td->o.gtod_reduce;
1870 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1871 struct io_completion_data *icd,
1872 const enum fio_ddir idx, unsigned int bytes)
1874 const int no_reduce = !gtod_reduce(td);
1875 unsigned long long llnsec = 0;
1880 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1884 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1886 if (!td->o.disable_lat) {
1887 unsigned long long tnsec;
1889 tnsec = ntime_since(&io_u->start_time, &icd->time);
1890 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1892 if (td->flags & TD_F_PROFILE_OPS) {
1893 struct prof_io_ops *ops = &td->prof_io_ops;
1896 icd->error = ops->io_u_lat(td, tnsec);
1899 if (td->o.max_latency && tnsec > td->o.max_latency)
1900 lat_fatal(td, icd, tnsec, td->o.max_latency);
1901 if (td->o.latency_target && tnsec > td->o.latency_target) {
1902 if (lat_target_failed(td))
1903 lat_fatal(td, icd, tnsec, td->o.latency_target);
1908 if (!td->o.disable_clat) {
1909 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1910 io_u_mark_latency(td, llnsec);
1913 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1914 add_bw_sample(td, io_u, bytes, llnsec);
1916 if (no_reduce && per_unit_log(td->iops_log))
1917 add_iops_sample(td, io_u, bytes);
1918 } else if (ddir_sync(idx) && !td->o.disable_clat)
1919 add_sync_clat_sample(&td->ts, llnsec);
1921 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1922 uint32_t *info = io_u_block_info(td, io_u);
1923 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1924 if (io_u->ddir == DDIR_TRIM) {
1925 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1926 BLOCK_INFO_TRIMS(*info) + 1);
1927 } else if (io_u->ddir == DDIR_WRITE) {
1928 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1935 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1936 uint64_t offset, unsigned int bytes)
1943 if (f->first_write == -1ULL || offset < f->first_write)
1944 f->first_write = offset;
1945 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1946 f->last_write = offset + bytes;
1948 if (!f->last_write_comp)
1951 idx = f->last_write_idx++;
1952 f->last_write_comp[idx] = offset;
1953 if (f->last_write_idx == td->o.iodepth)
1954 f->last_write_idx = 0;
1957 static bool should_account(struct thread_data *td)
1959 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1960 td->runstate == TD_VERIFYING);
1963 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1964 struct io_completion_data *icd)
1966 struct io_u *io_u = *io_u_ptr;
1967 enum fio_ddir ddir = io_u->ddir;
1968 struct fio_file *f = io_u->file;
1970 dprint_io_u(io_u, "complete");
1972 assert(io_u->flags & IO_U_F_FLIGHT);
1973 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1976 * Mark IO ok to verify
1980 * Remove errored entry from the verification list
1983 unlog_io_piece(td, io_u);
1985 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1990 if (ddir_sync(ddir)) {
1991 td->last_was_sync = true;
1993 f->first_write = -1ULL;
1994 f->last_write = -1ULL;
1996 if (should_account(td))
1997 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
2001 td->last_was_sync = false;
2002 td->last_ddir = ddir;
2004 if (!io_u->error && ddir_rw(ddir)) {
2005 unsigned int bytes = io_u->buflen - io_u->resid;
2008 td->io_blocks[ddir]++;
2009 td->io_bytes[ddir] += bytes;
2011 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2012 td->this_io_blocks[ddir]++;
2013 td->this_io_bytes[ddir] += bytes;
2016 if (ddir == DDIR_WRITE)
2017 file_log_write_comp(td, f, io_u->offset, bytes);
2019 if (should_account(td))
2020 account_io_completion(td, io_u, icd, ddir, bytes);
2022 icd->bytes_done[ddir] += bytes;
2025 ret = io_u->end_io(td, io_u_ptr);
2027 if (ret && !icd->error)
2030 } else if (io_u->error) {
2031 icd->error = io_u->error;
2032 io_u_log_error(td, io_u);
2035 enum error_type_bit eb = td_error_type(ddir, icd->error);
2037 if (!td_non_fatal_error(td, eb, icd->error))
2041 * If there is a non_fatal error, then add to the error count
2042 * and clear all the errors.
2044 update_error_count(td, icd->error);
2052 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2057 if (!gtod_reduce(td))
2058 fio_gettime(&icd->time, NULL);
2063 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2064 icd->bytes_done[ddir] = 0;
2067 static void ios_completed(struct thread_data *td,
2068 struct io_completion_data *icd)
2073 for (i = 0; i < icd->nr; i++) {
2074 io_u = td->io_ops->event(td, i);
2076 io_completed(td, &io_u, icd);
2084 * Complete a single io_u for the sync engines.
2086 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2088 struct io_completion_data icd;
2091 init_icd(td, &icd, 1);
2092 io_completed(td, &io_u, &icd);
2098 td_verror(td, icd.error, "io_u_sync_complete");
2102 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2103 td->bytes_done[ddir] += icd.bytes_done[ddir];
2109 * Called to complete min_events number of io for the async engines.
2111 int io_u_queued_complete(struct thread_data *td, int min_evts)
2113 struct io_completion_data icd;
2114 struct timespec *tvp = NULL;
2116 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2118 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2122 else if (min_evts > td->cur_depth)
2123 min_evts = td->cur_depth;
2125 /* No worries, td_io_getevents fixes min and max if they are
2126 * set incorrectly */
2127 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2129 td_verror(td, -ret, "td_io_getevents");
2134 init_icd(td, &icd, ret);
2135 ios_completed(td, &icd);
2137 td_verror(td, icd.error, "io_u_queued_complete");
2141 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2142 td->bytes_done[ddir] += icd.bytes_done[ddir];
2148 * Call when io_u is really queued, to update the submission latency.
2150 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2152 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2153 unsigned long slat_time;
2155 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2160 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2166 * See if we should reuse the last seed, if dedupe is enabled
2168 static struct frand_state *get_buf_state(struct thread_data *td)
2172 if (!td->o.dedupe_percentage)
2173 return &td->buf_state;
2174 else if (td->o.dedupe_percentage == 100) {
2175 frand_copy(&td->buf_state_prev, &td->buf_state);
2176 return &td->buf_state;
2179 v = rand32_between(&td->dedupe_state, 1, 100);
2181 if (v <= td->o.dedupe_percentage)
2182 return &td->buf_state_prev;
2184 return &td->buf_state;
2187 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2189 if (td->o.dedupe_percentage == 100)
2190 frand_copy(rs, &td->buf_state_prev);
2191 else if (rs == &td->buf_state)
2192 frand_copy(&td->buf_state_prev, rs);
2195 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2196 unsigned int max_bs)
2198 struct thread_options *o = &td->o;
2200 if (o->mem_type == MEM_CUDA_MALLOC)
2203 if (o->compress_percentage || o->dedupe_percentage) {
2204 unsigned int perc = td->o.compress_percentage;
2205 struct frand_state *rs;
2206 unsigned int left = max_bs;
2207 unsigned int this_write;
2210 rs = get_buf_state(td);
2212 min_write = min(min_write, left);
2215 this_write = min_not_zero(min_write,
2216 td->o.compress_chunk);
2218 fill_random_buf_percentage(rs, buf, perc,
2219 this_write, this_write,
2221 o->buffer_pattern_bytes);
2223 fill_random_buf(rs, buf, min_write);
2224 this_write = min_write;
2229 save_buf_state(td, rs);
2231 } else if (o->buffer_pattern_bytes)
2232 fill_buffer_pattern(td, buf, max_bs);
2233 else if (o->zero_buffers)
2234 memset(buf, 0, max_bs);
2236 fill_random_buf(get_buf_state(td), buf, max_bs);
2240 * "randomly" fill the buffer contents
2242 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2243 unsigned int min_write, unsigned int max_bs)
2245 io_u->buf_filled_len = 0;
2246 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2249 static int do_sync_file_range(const struct thread_data *td,
2252 off64_t offset, nbytes;
2254 offset = f->first_write;
2255 nbytes = f->last_write - f->first_write;
2260 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2263 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2267 if (io_u->ddir == DDIR_SYNC) {
2268 ret = fsync(io_u->file->fd);
2269 } else if (io_u->ddir == DDIR_DATASYNC) {
2270 #ifdef CONFIG_FDATASYNC
2271 ret = fdatasync(io_u->file->fd);
2273 ret = io_u->xfer_buflen;
2274 io_u->error = EINVAL;
2276 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2277 ret = do_sync_file_range(td, io_u->file);
2279 ret = io_u->xfer_buflen;
2280 io_u->error = EINVAL;
2284 io_u->error = errno;
2289 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2291 #ifndef FIO_HAVE_TRIM
2292 io_u->error = EINVAL;
2295 struct fio_file *f = io_u->file;
2298 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2300 return io_u->xfer_buflen;