14 struct io_completion_data {
17 int error; /* output */
18 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
19 struct timespec time; /* output */
23 * The ->io_axmap contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 static bool random_map_free(struct fio_file *f, const uint64_t block)
28 return !axmap_isset(f->io_axmap, block);
32 * Mark a given offset as used in the map.
34 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36 unsigned int min_bs = td->o.min_bs[io_u->ddir];
37 struct fio_file *f = io_u->file;
38 unsigned int nr_blocks;
41 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
42 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44 if (!(io_u->flags & IO_U_F_BUSY_OK))
45 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47 if ((nr_blocks * min_bs) < io_u->buflen)
48 io_u->buflen = nr_blocks * min_bs;
51 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
57 assert(ddir_rw(ddir));
60 * Hmm, should we make sure that ->io_size <= ->real_file_size?
61 * -> not for now since there is code assuming it could go either.
63 max_size = f->io_size;
64 if (max_size > f->real_file_size)
65 max_size = f->real_file_size;
68 max_size = td->o.zone_range;
70 if (td->o.min_bs[ddir] > td->o.ba[ddir])
71 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
73 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
81 struct flist_head list;
85 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
86 enum fio_ddir ddir, uint64_t *b,
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
92 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
94 r = __rand(&td->random_state);
96 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
98 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
102 assert(fio_file_lfsr(f));
104 if (lfsr_next(&f->lfsr, &off))
111 * if we are not maintaining a random map, we are done.
113 if (!file_randommap(td, f))
117 * calculate map offset and check if it's free
119 if (random_map_free(f, *b))
122 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
123 (unsigned long long) *b);
125 *b = axmap_next_free(f->io_axmap, *b);
126 if (*b == (uint64_t) -1ULL)
132 static int __get_next_rand_offset_zipf(struct thread_data *td,
133 struct fio_file *f, enum fio_ddir ddir,
136 *b = zipf_next(&f->zipf);
140 static int __get_next_rand_offset_pareto(struct thread_data *td,
141 struct fio_file *f, enum fio_ddir ddir,
144 *b = pareto_next(&f->zipf);
148 static int __get_next_rand_offset_gauss(struct thread_data *td,
149 struct fio_file *f, enum fio_ddir ddir,
152 *b = gauss_next(&f->gauss);
156 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
158 enum fio_ddir ddir, uint64_t *b)
160 struct zone_split_index *zsi;
161 uint64_t lastb, send, stotal;
164 lastb = last_block(td, f, ddir);
168 if (!td->o.zone_split_nr[ddir]) {
170 return __get_next_rand_offset(td, f, ddir, b, lastb);
174 * Generate a value, v, between 1 and 100, both inclusive
176 v = rand32_between(&td->zone_state, 1, 100);
179 * Find our generated table. 'send' is the end block of this zone,
180 * 'stotal' is our start offset.
182 zsi = &td->zone_state_index[ddir][v - 1];
183 stotal = zsi->size_prev / td->o.ba[ddir];
184 send = zsi->size / td->o.ba[ddir];
187 * Should never happen
190 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
191 log_err("fio: bug in zoned generation\n");
193 } else if (send > lastb) {
195 * This happens if the user specifies ranges that exceed
196 * the file/device size. We can't handle that gracefully,
199 log_err("fio: zoned_abs sizes exceed file size\n");
204 * Generate index from 0..send-stotal
206 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
213 static int __get_next_rand_offset_zoned(struct thread_data *td,
214 struct fio_file *f, enum fio_ddir ddir,
217 unsigned int v, send, stotal;
218 uint64_t offset, lastb;
219 struct zone_split_index *zsi;
221 lastb = last_block(td, f, ddir);
225 if (!td->o.zone_split_nr[ddir]) {
227 return __get_next_rand_offset(td, f, ddir, b, lastb);
231 * Generate a value, v, between 1 and 100, both inclusive
233 v = rand32_between(&td->zone_state, 1, 100);
235 zsi = &td->zone_state_index[ddir][v - 1];
236 stotal = zsi->size_perc_prev;
237 send = zsi->size_perc;
240 * Should never happen
243 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
244 log_err("fio: bug in zoned generation\n");
249 * 'send' is some percentage below or equal to 100 that
250 * marks the end of the current IO range. 'stotal' marks
251 * the start, in percent.
254 offset = stotal * lastb / 100ULL;
258 lastb = lastb * (send - stotal) / 100ULL;
261 * Generate index from 0..send-of-lastb
263 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
267 * Add our start offset, if any
275 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
277 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
278 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
280 return r1->off - r2->off;
283 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
284 enum fio_ddir ddir, uint64_t *b)
286 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
289 lastb = last_block(td, f, ddir);
293 return __get_next_rand_offset(td, f, ddir, b, lastb);
294 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
295 return __get_next_rand_offset_zipf(td, f, ddir, b);
296 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
297 return __get_next_rand_offset_pareto(td, f, ddir, b);
298 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
299 return __get_next_rand_offset_gauss(td, f, ddir, b);
300 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
301 return __get_next_rand_offset_zoned(td, f, ddir, b);
302 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
303 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
305 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
310 * Sort the reads for a verify phase in batches of verifysort_nr, if
313 static inline bool should_sort_io(struct thread_data *td)
315 if (!td->o.verifysort_nr || !td->o.do_verify)
319 if (td->runstate != TD_VERIFYING)
321 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
322 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
328 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
332 if (td->o.perc_rand[ddir] == 100)
335 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
337 return v <= td->o.perc_rand[ddir];
340 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
341 enum fio_ddir ddir, uint64_t *b)
346 if (!should_sort_io(td))
347 return get_off_from_method(td, f, ddir, b);
349 if (!flist_empty(&td->next_rand_list)) {
351 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
358 for (i = 0; i < td->o.verifysort_nr; i++) {
359 r = malloc(sizeof(*r));
361 ret = get_off_from_method(td, f, ddir, &r->off);
367 flist_add(&r->list, &td->next_rand_list);
373 assert(!flist_empty(&td->next_rand_list));
374 flist_sort(NULL, &td->next_rand_list, flist_cmp);
378 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
380 struct thread_options *o = &td->o;
382 if (o->invalidate_cache && !o->odirect) {
385 ret = file_invalidate_cache(td, f);
389 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
390 enum fio_ddir ddir, uint64_t *b)
392 if (!get_next_rand_offset(td, f, ddir, b))
395 if (td->o.time_based ||
396 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
397 fio_file_reset(td, f);
398 if (!get_next_rand_offset(td, f, ddir, b))
400 loop_cache_invalidate(td, f);
403 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
404 f->file_name, (unsigned long long) f->last_pos[ddir],
405 (unsigned long long) f->real_file_size);
409 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
410 enum fio_ddir ddir, uint64_t *offset)
412 struct thread_options *o = &td->o;
414 assert(ddir_rw(ddir));
417 * If we reach the end for a time based run, reset us back to 0
418 * and invalidate the cache, if we need to.
420 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
422 f->last_pos[ddir] = f->file_offset;
423 loop_cache_invalidate(td, f);
426 if (f->last_pos[ddir] < f->real_file_size) {
430 * Only rewind if we already hit the end
432 if (f->last_pos[ddir] == f->file_offset &&
433 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,
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,
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 inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
578 struct fio_file *f = io_u->file;
580 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
583 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
586 int ddir = io_u->ddir;
587 unsigned int buflen = 0;
588 unsigned int minbs, maxbs;
589 uint64_t frand_max, r;
592 assert(ddir_rw(ddir));
594 if (td->o.bs_is_seq_rand)
595 ddir = is_random ? DDIR_WRITE : DDIR_READ;
597 minbs = td->o.min_bs[ddir];
598 maxbs = td->o.max_bs[ddir];
604 * If we can't satisfy the min block size from here, then fail
606 if (!io_u_fits(td, io_u, minbs))
609 frand_max = rand_max(&td->bsrange_state[ddir]);
611 r = __rand(&td->bsrange_state[ddir]);
613 if (!td->o.bssplit_nr[ddir]) {
614 buflen = 1 + (unsigned int) ((double) maxbs *
615 (r / (frand_max + 1.0)));
622 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
623 struct bssplit *bsp = &td->o.bssplit[ddir][i];
629 if ((r / perc <= frand_max / 100ULL) &&
630 io_u_fits(td, io_u, buflen))
635 power_2 = is_power_of_2(minbs);
636 if (!td->o.bs_unaligned && power_2)
637 buflen &= ~(minbs - 1);
638 else if (!td->o.bs_unaligned && !power_2)
639 buflen -= buflen % minbs;
640 } while (!io_u_fits(td, io_u, buflen));
645 static void set_rwmix_bytes(struct thread_data *td)
650 * we do time or byte based switch. this is needed because
651 * buffered writes may issue a lot quicker than they complete,
652 * whereas reads do not.
654 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
655 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
658 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
662 v = rand32_between(&td->rwmix_state, 1, 100);
664 if (v <= td->o.rwmix[DDIR_READ])
670 int io_u_quiesce(struct thread_data *td)
675 * We are going to sleep, ensure that we flush anything pending as
676 * not to skew our latency numbers.
678 * Changed to only monitor 'in flight' requests here instead of the
679 * td->cur_depth, b/c td->cur_depth does not accurately represent
680 * io's that have been actually submitted to an async engine,
681 * and cur_depth is meaningless for sync engines.
683 if (td->io_u_queued || td->cur_depth) {
686 ret = td_io_commit(td);
689 while (td->io_u_in_flight) {
692 ret = io_u_queued_complete(td, 1);
697 if (td->flags & TD_F_REGROW_LOGS)
703 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
705 enum fio_ddir odir = ddir ^ 1;
709 assert(ddir_rw(ddir));
710 now = utime_since_now(&td->start);
713 * if rate_next_io_time is in the past, need to catch up to rate
715 if (td->rate_next_io_time[ddir] <= now)
719 * We are ahead of rate in this direction. See if we
722 if (td_rw(td) && td->o.rwmix[odir]) {
724 * Other direction is behind rate, switch
726 if (td->rate_next_io_time[odir] <= now)
730 * Both directions are ahead of rate. sleep the min,
731 * switch if necessary
733 if (td->rate_next_io_time[ddir] <=
734 td->rate_next_io_time[odir]) {
735 usec = td->rate_next_io_time[ddir] - now;
737 usec = td->rate_next_io_time[odir] - now;
741 usec = td->rate_next_io_time[ddir] - now;
743 if (td->o.io_submit_mode == IO_MODE_INLINE)
746 usec_sleep(td, usec);
751 * Return the data direction for the next io_u. If the job is a
752 * mixed read/write workload, check the rwmix cycle and switch if
755 static enum fio_ddir get_rw_ddir(struct thread_data *td)
760 * See if it's time to fsync/fdatasync/sync_file_range first,
761 * and if not then move on to check regular I/Os.
763 if (should_fsync(td)) {
764 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
765 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
768 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
769 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
770 return DDIR_DATASYNC;
772 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
773 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
774 return DDIR_SYNC_FILE_RANGE;
779 * Check if it's time to seed a new data direction.
781 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
783 * Put a top limit on how many bytes we do for
784 * one data direction, to avoid overflowing the
787 ddir = get_rand_ddir(td);
789 if (ddir != td->rwmix_ddir)
792 td->rwmix_ddir = ddir;
794 ddir = td->rwmix_ddir;
795 } else if (td_read(td))
797 else if (td_write(td))
799 else if (td_trim(td))
804 td->rwmix_ddir = rate_ddir(td, ddir);
805 return td->rwmix_ddir;
808 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
810 enum fio_ddir ddir = get_rw_ddir(td);
812 if (td_trimwrite(td)) {
813 struct fio_file *f = io_u->file;
814 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
820 io_u->ddir = io_u->acct_ddir = ddir;
822 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
823 td->o.barrier_blocks &&
824 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
825 td->io_issues[DDIR_WRITE])
826 io_u_set(td, io_u, IO_U_F_BARRIER);
829 void put_file_log(struct thread_data *td, struct fio_file *f)
831 unsigned int ret = put_file(td, f);
834 td_verror(td, ret, "file close");
837 void put_io_u(struct thread_data *td, struct io_u *io_u)
844 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
845 put_file_log(td, io_u->file);
848 io_u_set(td, io_u, IO_U_F_FREE);
850 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
852 assert(!(td->flags & TD_F_CHILD));
854 io_u_qpush(&td->io_u_freelist, io_u);
855 td_io_u_free_notify(td);
859 void clear_io_u(struct thread_data *td, struct io_u *io_u)
861 io_u_clear(td, io_u, IO_U_F_FLIGHT);
865 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
867 struct io_u *__io_u = *io_u;
868 enum fio_ddir ddir = acct_ddir(__io_u);
870 dprint(FD_IO, "requeue %p\n", __io_u);
877 io_u_set(td, __io_u, IO_U_F_FREE);
878 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
879 td->io_issues[ddir]--;
881 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
882 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
884 assert(!(td->flags & TD_F_CHILD));
887 io_u_rpush(&td->io_u_requeues, __io_u);
888 td_io_u_free_notify(td);
893 static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
895 struct fio_file *f = io_u->file;
898 * See if it's time to switch to a new zone
900 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
902 f->file_offset += td->o.zone_range + td->o.zone_skip;
905 * Wrap from the beginning, if we exceed the file size
907 if (f->file_offset >= f->real_file_size)
908 f->file_offset = f->real_file_size - f->file_offset;
909 f->last_pos[io_u->ddir] = f->file_offset;
910 td->io_skip_bytes += td->o.zone_skip;
914 * If zone_size > zone_range, then maintain the same zone until
915 * zone_bytes >= zone_size.
917 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
918 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
919 f->file_offset, f->last_pos[io_u->ddir]);
920 f->last_pos[io_u->ddir] = f->file_offset;
924 * For random: if 'norandommap' is not set and zone_size > zone_range,
925 * map needs to be reset as it's done with zone_range everytime.
927 if ((td->zone_bytes % td->o.zone_range) == 0) {
928 fio_file_reset(td, f);
932 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
936 if (td_ioengine_flagged(td, FIO_NOIO))
939 set_rw_ddir(td, io_u);
942 * fsync() or fdatasync() or trim etc, we are done
944 if (!ddir_rw(io_u->ddir))
948 * When file is zoned zone_range is always positive
950 if (td->o.zone_range) {
951 __fill_io_u_zone(td, io_u);
955 * No log, let the seq/rand engine retrieve the next buflen and
958 if (get_next_offset(td, io_u, &is_random)) {
959 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
963 io_u->buflen = get_next_buflen(td, io_u, is_random);
965 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
969 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
970 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
972 (unsigned long long) io_u->offset, io_u->buflen,
973 (unsigned long long) io_u->file->real_file_size);
978 * mark entry before potentially trimming io_u
980 if (td_random(td) && file_randommap(td, io_u->file))
981 mark_random_map(td, io_u);
984 dprint_io_u(io_u, "fill");
985 td->zone_bytes += io_u->buflen;
989 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
1018 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1020 __io_u_mark_map(td->ts.io_u_submit, nr);
1021 td->ts.total_submit++;
1024 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1026 __io_u_mark_map(td->ts.io_u_complete, nr);
1027 td->ts.total_complete++;
1030 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1034 switch (td->cur_depth) {
1056 td->ts.io_u_map[idx] += nr;
1059 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1063 assert(nsec < 1000);
1096 assert(idx < FIO_IO_U_LAT_N_NR);
1097 td->ts.io_u_lat_n[idx]++;
1100 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1104 assert(usec < 1000 && usec >= 1);
1137 assert(idx < FIO_IO_U_LAT_U_NR);
1138 td->ts.io_u_lat_u[idx]++;
1141 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1184 assert(idx < FIO_IO_U_LAT_M_NR);
1185 td->ts.io_u_lat_m[idx]++;
1188 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1191 io_u_mark_lat_nsec(td, nsec);
1192 else if (nsec < 1000000)
1193 io_u_mark_lat_usec(td, nsec / 1000);
1195 io_u_mark_lat_msec(td, nsec / 1000000);
1198 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1200 unsigned long fileno;
1202 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1203 uint64_t frand_max = rand_max(&td->next_file_state);
1206 r = __rand(&td->next_file_state);
1207 return (unsigned int) ((double) td->o.nr_files
1208 * (r / (frand_max + 1.0)));
1211 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1212 fileno = zipf_next(&td->next_file_zipf);
1213 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1214 fileno = pareto_next(&td->next_file_zipf);
1215 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1216 fileno = gauss_next(&td->next_file_gauss);
1218 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1223 return fileno >> FIO_FSERVICE_SHIFT;
1227 * Get next file to service by choosing one at random
1229 static struct fio_file *get_next_file_rand(struct thread_data *td,
1230 enum fio_file_flags goodf,
1231 enum fio_file_flags badf)
1239 fno = __get_next_fileno_rand(td);
1242 if (fio_file_done(f))
1245 if (!fio_file_open(f)) {
1248 if (td->nr_open_files >= td->o.open_files)
1249 return ERR_PTR(-EBUSY);
1251 err = td_io_open_file(td, f);
1257 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1258 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1262 td_io_close_file(td, f);
1267 * Get next file to service by doing round robin between all available ones
1269 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1272 unsigned int old_next_file = td->next_file;
1278 f = td->files[td->next_file];
1281 if (td->next_file >= td->o.nr_files)
1284 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1285 if (fio_file_done(f)) {
1290 if (!fio_file_open(f)) {
1293 if (td->nr_open_files >= td->o.open_files)
1294 return ERR_PTR(-EBUSY);
1296 err = td_io_open_file(td, f);
1298 dprint(FD_FILE, "error %d on open of %s\n",
1306 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1308 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1312 td_io_close_file(td, f);
1315 } while (td->next_file != old_next_file);
1317 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1321 static struct fio_file *__get_next_file(struct thread_data *td)
1325 assert(td->o.nr_files <= td->files_index);
1327 if (td->nr_done_files >= td->o.nr_files) {
1328 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1329 " nr_files=%d\n", td->nr_open_files,
1335 f = td->file_service_file;
1336 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1337 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1339 if (td->file_service_left--)
1343 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1344 td->o.file_service_type == FIO_FSERVICE_SEQ)
1345 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1347 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1352 td->file_service_file = f;
1353 td->file_service_left = td->file_service_nr - 1;
1356 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1358 dprint(FD_FILE, "get_next_file: NULL\n");
1362 static struct fio_file *get_next_file(struct thread_data *td)
1364 return __get_next_file(td);
1367 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1372 f = get_next_file(td);
1373 if (IS_ERR_OR_NULL(f))
1379 if (!fill_io_u(td, io_u))
1382 put_file_log(td, f);
1383 td_io_close_file(td, f);
1385 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1386 fio_file_reset(td, f);
1388 fio_file_set_done(f);
1389 td->nr_done_files++;
1390 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1391 td->nr_done_files, td->o.nr_files);
1398 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1399 unsigned long long tnsec, unsigned long long max_nsec)
1402 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1403 td_verror(td, ETIMEDOUT, "max latency exceeded");
1404 icd->error = ETIMEDOUT;
1407 static void lat_new_cycle(struct thread_data *td)
1409 fio_gettime(&td->latency_ts, NULL);
1410 td->latency_ios = ddir_rw_sum(td->io_blocks);
1411 td->latency_failed = 0;
1415 * We had an IO outside the latency target. Reduce the queue depth. If we
1416 * are at QD=1, then it's time to give up.
1418 static bool __lat_target_failed(struct thread_data *td)
1420 if (td->latency_qd == 1)
1423 td->latency_qd_high = td->latency_qd;
1425 if (td->latency_qd == td->latency_qd_low)
1426 td->latency_qd_low--;
1428 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1430 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1433 * When we ramp QD down, quiesce existing IO to prevent
1434 * a storm of ramp downs due to pending higher depth.
1441 static bool lat_target_failed(struct thread_data *td)
1443 if (td->o.latency_percentile.u.f == 100.0)
1444 return __lat_target_failed(td);
1446 td->latency_failed++;
1450 void lat_target_init(struct thread_data *td)
1452 td->latency_end_run = 0;
1454 if (td->o.latency_target) {
1455 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1456 fio_gettime(&td->latency_ts, NULL);
1458 td->latency_qd_high = td->o.iodepth;
1459 td->latency_qd_low = 1;
1460 td->latency_ios = ddir_rw_sum(td->io_blocks);
1462 td->latency_qd = td->o.iodepth;
1465 void lat_target_reset(struct thread_data *td)
1467 if (!td->latency_end_run)
1468 lat_target_init(td);
1471 static void lat_target_success(struct thread_data *td)
1473 const unsigned int qd = td->latency_qd;
1474 struct thread_options *o = &td->o;
1476 td->latency_qd_low = td->latency_qd;
1479 * If we haven't failed yet, we double up to a failing value instead
1480 * of bisecting from highest possible queue depth. If we have set
1481 * a limit other than td->o.iodepth, bisect between that.
1483 if (td->latency_qd_high != o->iodepth)
1484 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1486 td->latency_qd *= 2;
1488 if (td->latency_qd > o->iodepth)
1489 td->latency_qd = o->iodepth;
1491 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1494 * Same as last one, we are done. Let it run a latency cycle, so
1495 * we get only the results from the targeted depth.
1497 if (td->latency_qd == qd) {
1498 if (td->latency_end_run) {
1499 dprint(FD_RATE, "We are done\n");
1502 dprint(FD_RATE, "Quiesce and final run\n");
1504 td->latency_end_run = 1;
1505 reset_all_stats(td);
1514 * Check if we can bump the queue depth
1516 void lat_target_check(struct thread_data *td)
1518 uint64_t usec_window;
1522 usec_window = utime_since_now(&td->latency_ts);
1523 if (usec_window < td->o.latency_window)
1526 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1527 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1528 success_ios *= 100.0;
1530 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1532 if (success_ios >= td->o.latency_percentile.u.f)
1533 lat_target_success(td);
1535 __lat_target_failed(td);
1539 * If latency target is enabled, we might be ramping up or down and not
1540 * using the full queue depth available.
1542 bool queue_full(const struct thread_data *td)
1544 const int qempty = io_u_qempty(&td->io_u_freelist);
1548 if (!td->o.latency_target)
1551 return td->cur_depth >= td->latency_qd;
1554 struct io_u *__get_io_u(struct thread_data *td)
1556 struct io_u *io_u = NULL;
1565 if (!io_u_rempty(&td->io_u_requeues))
1566 io_u = io_u_rpop(&td->io_u_requeues);
1567 else if (!queue_full(td)) {
1568 io_u = io_u_qpop(&td->io_u_freelist);
1573 io_u->end_io = NULL;
1577 assert(io_u->flags & IO_U_F_FREE);
1578 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1579 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1583 io_u->acct_ddir = -1;
1585 assert(!(td->flags & TD_F_CHILD));
1586 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1588 } else if (td_async_processing(td)) {
1590 * We ran out, wait for async verify threads to finish and
1593 assert(!(td->flags & TD_F_CHILD));
1594 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1603 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1605 if (!(td->flags & TD_F_TRIM_BACKLOG))
1607 if (!td->trim_entries)
1610 if (td->trim_batch) {
1612 if (get_next_trim(td, io_u))
1614 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1615 td->last_ddir != DDIR_READ) {
1616 td->trim_batch = td->o.trim_batch;
1617 if (!td->trim_batch)
1618 td->trim_batch = td->o.trim_backlog;
1619 if (get_next_trim(td, io_u))
1626 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1628 if (!(td->flags & TD_F_VER_BACKLOG))
1631 if (td->io_hist_len) {
1634 if (td->verify_batch)
1636 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1637 td->last_ddir != DDIR_READ) {
1638 td->verify_batch = td->o.verify_batch;
1639 if (!td->verify_batch)
1640 td->verify_batch = td->o.verify_backlog;
1644 if (get_verify && !get_next_verify(td, io_u)) {
1654 * Fill offset and start time into the buffer content, to prevent too
1655 * easy compressible data for simple de-dupe attempts. Do this for every
1656 * 512b block in the range, since that should be the smallest block size
1657 * we can expect from a device.
1659 static void small_content_scramble(struct io_u *io_u)
1661 unsigned int i, nr_blocks = io_u->buflen >> 9;
1662 unsigned int offset;
1663 uint64_t boffset, *iptr;
1670 boffset = io_u->offset;
1672 if (io_u->buf_filled_len)
1673 io_u->buf_filled_len = 0;
1676 * Generate random index between 0..7. We do chunks of 512b, if
1677 * we assume a cacheline is 64 bytes, then we have 8 of those.
1678 * Scramble content within the blocks in the same cacheline to
1681 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1683 for (i = 0; i < nr_blocks; i++) {
1685 * Fill offset into start of cacheline, time into end
1688 iptr = (void *) p + (offset << 6);
1691 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1692 iptr[0] = io_u->start_time.tv_sec;
1693 iptr[1] = io_u->start_time.tv_nsec;
1701 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1702 * etc. The returned io_u is fully ready to be prepped and submitted.
1704 struct io_u *get_io_u(struct thread_data *td)
1708 int do_scramble = 0;
1711 io_u = __get_io_u(td);
1713 dprint(FD_IO, "__get_io_u failed\n");
1717 if (check_get_verify(td, io_u))
1719 if (check_get_trim(td, io_u))
1723 * from a requeue, io_u already setup
1729 * If using an iolog, grab next piece if any available.
1731 if (td->flags & TD_F_READ_IOLOG) {
1732 if (read_iolog_get(td, io_u))
1734 } else if (set_io_u_file(td, io_u)) {
1736 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1742 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1746 assert(fio_file_open(f));
1748 if (ddir_rw(io_u->ddir)) {
1749 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1750 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1754 f->last_start[io_u->ddir] = io_u->offset;
1755 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1757 if (io_u->ddir == DDIR_WRITE) {
1758 if (td->flags & TD_F_REFILL_BUFFERS) {
1759 io_u_fill_buffer(td, io_u,
1760 td->o.min_bs[DDIR_WRITE],
1762 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1763 !(td->flags & TD_F_COMPRESS))
1765 if (td->flags & TD_F_VER_NONE) {
1766 populate_verify_io_u(td, io_u);
1769 } else if (io_u->ddir == DDIR_READ) {
1771 * Reset the buf_filled parameters so next time if the
1772 * buffer is used for writes it is refilled.
1774 io_u->buf_filled_len = 0;
1779 * Set io data pointers.
1781 io_u->xfer_buf = io_u->buf;
1782 io_u->xfer_buflen = io_u->buflen;
1786 if (!td_io_prep(td, io_u)) {
1787 if (!td->o.disable_lat)
1788 fio_gettime(&io_u->start_time, NULL);
1791 small_content_scramble(io_u);
1796 dprint(FD_IO, "get_io_u failed\n");
1798 return ERR_PTR(ret);
1801 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1803 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1805 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1808 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1809 io_u->file ? " on file " : "",
1810 io_u->file ? io_u->file->file_name : "",
1811 strerror(io_u->error),
1812 io_ddir_name(io_u->ddir),
1813 io_u->offset, io_u->xfer_buflen);
1815 if (td->io_ops->errdetails) {
1816 char *err = td->io_ops->errdetails(io_u);
1818 log_err("fio: %s\n", err);
1823 td_verror(td, io_u->error, "io_u error");
1826 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1828 __io_u_log_error(td, io_u);
1830 __io_u_log_error(td->parent, io_u);
1833 static inline bool gtod_reduce(struct thread_data *td)
1835 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1836 || td->o.gtod_reduce;
1839 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1840 struct io_completion_data *icd,
1841 const enum fio_ddir idx, unsigned int bytes)
1843 const int no_reduce = !gtod_reduce(td);
1844 unsigned long long llnsec = 0;
1849 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1853 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1855 if (!td->o.disable_lat) {
1856 unsigned long long tnsec;
1858 tnsec = ntime_since(&io_u->start_time, &icd->time);
1859 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1861 if (td->flags & TD_F_PROFILE_OPS) {
1862 struct prof_io_ops *ops = &td->prof_io_ops;
1865 icd->error = ops->io_u_lat(td, tnsec);
1868 if (td->o.max_latency && tnsec > td->o.max_latency)
1869 lat_fatal(td, icd, tnsec, td->o.max_latency);
1870 if (td->o.latency_target && tnsec > td->o.latency_target) {
1871 if (lat_target_failed(td))
1872 lat_fatal(td, icd, tnsec, td->o.latency_target);
1877 if (!td->o.disable_clat) {
1878 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1879 io_u_mark_latency(td, llnsec);
1882 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1883 add_bw_sample(td, io_u, bytes, llnsec);
1885 if (no_reduce && per_unit_log(td->iops_log))
1886 add_iops_sample(td, io_u, bytes);
1887 } else if (ddir_sync(idx) && !td->o.disable_clat)
1888 add_sync_clat_sample(&td->ts, llnsec);
1890 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1891 uint32_t *info = io_u_block_info(td, io_u);
1892 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1893 if (io_u->ddir == DDIR_TRIM) {
1894 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1895 BLOCK_INFO_TRIMS(*info) + 1);
1896 } else if (io_u->ddir == DDIR_WRITE) {
1897 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1904 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1905 uint64_t offset, unsigned int bytes)
1912 if (f->first_write == -1ULL || offset < f->first_write)
1913 f->first_write = offset;
1914 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1915 f->last_write = offset + bytes;
1917 if (!f->last_write_comp)
1920 idx = f->last_write_idx++;
1921 f->last_write_comp[idx] = offset;
1922 if (f->last_write_idx == td->o.iodepth)
1923 f->last_write_idx = 0;
1926 static bool should_account(struct thread_data *td)
1928 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1929 td->runstate == TD_VERIFYING);
1932 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1933 struct io_completion_data *icd)
1935 struct io_u *io_u = *io_u_ptr;
1936 enum fio_ddir ddir = io_u->ddir;
1937 struct fio_file *f = io_u->file;
1939 dprint_io_u(io_u, "complete");
1941 assert(io_u->flags & IO_U_F_FLIGHT);
1942 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1945 * Mark IO ok to verify
1949 * Remove errored entry from the verification list
1952 unlog_io_piece(td, io_u);
1954 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1959 if (ddir_sync(ddir)) {
1960 td->last_was_sync = true;
1962 f->first_write = -1ULL;
1963 f->last_write = -1ULL;
1965 if (should_account(td))
1966 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1970 td->last_was_sync = false;
1971 td->last_ddir = ddir;
1973 if (!io_u->error && ddir_rw(ddir)) {
1974 unsigned int bytes = io_u->buflen - io_u->resid;
1977 td->io_blocks[ddir]++;
1978 td->io_bytes[ddir] += bytes;
1980 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1981 td->this_io_blocks[ddir]++;
1982 td->this_io_bytes[ddir] += bytes;
1985 if (ddir == DDIR_WRITE)
1986 file_log_write_comp(td, f, io_u->offset, bytes);
1988 if (should_account(td))
1989 account_io_completion(td, io_u, icd, ddir, bytes);
1991 icd->bytes_done[ddir] += bytes;
1994 ret = io_u->end_io(td, io_u_ptr);
1996 if (ret && !icd->error)
1999 } else if (io_u->error) {
2000 icd->error = io_u->error;
2001 io_u_log_error(td, io_u);
2004 enum error_type_bit eb = td_error_type(ddir, icd->error);
2006 if (!td_non_fatal_error(td, eb, icd->error))
2010 * If there is a non_fatal error, then add to the error count
2011 * and clear all the errors.
2013 update_error_count(td, icd->error);
2021 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2026 if (!gtod_reduce(td))
2027 fio_gettime(&icd->time, NULL);
2032 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2033 icd->bytes_done[ddir] = 0;
2036 static void ios_completed(struct thread_data *td,
2037 struct io_completion_data *icd)
2042 for (i = 0; i < icd->nr; i++) {
2043 io_u = td->io_ops->event(td, i);
2045 io_completed(td, &io_u, icd);
2053 * Complete a single io_u for the sync engines.
2055 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2057 struct io_completion_data icd;
2060 init_icd(td, &icd, 1);
2061 io_completed(td, &io_u, &icd);
2067 td_verror(td, icd.error, "io_u_sync_complete");
2071 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2072 td->bytes_done[ddir] += icd.bytes_done[ddir];
2078 * Called to complete min_events number of io for the async engines.
2080 int io_u_queued_complete(struct thread_data *td, int min_evts)
2082 struct io_completion_data icd;
2083 struct timespec *tvp = NULL;
2085 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2087 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2091 else if (min_evts > td->cur_depth)
2092 min_evts = td->cur_depth;
2094 /* No worries, td_io_getevents fixes min and max if they are
2095 * set incorrectly */
2096 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2098 td_verror(td, -ret, "td_io_getevents");
2103 init_icd(td, &icd, ret);
2104 ios_completed(td, &icd);
2106 td_verror(td, icd.error, "io_u_queued_complete");
2110 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2111 td->bytes_done[ddir] += icd.bytes_done[ddir];
2117 * Call when io_u is really queued, to update the submission latency.
2119 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2121 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2122 unsigned long slat_time;
2124 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2129 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2135 * See if we should reuse the last seed, if dedupe is enabled
2137 static struct frand_state *get_buf_state(struct thread_data *td)
2141 if (!td->o.dedupe_percentage)
2142 return &td->buf_state;
2143 else if (td->o.dedupe_percentage == 100) {
2144 frand_copy(&td->buf_state_prev, &td->buf_state);
2145 return &td->buf_state;
2148 v = rand32_between(&td->dedupe_state, 1, 100);
2150 if (v <= td->o.dedupe_percentage)
2151 return &td->buf_state_prev;
2153 return &td->buf_state;
2156 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2158 if (td->o.dedupe_percentage == 100)
2159 frand_copy(rs, &td->buf_state_prev);
2160 else if (rs == &td->buf_state)
2161 frand_copy(&td->buf_state_prev, rs);
2164 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2165 unsigned int max_bs)
2167 struct thread_options *o = &td->o;
2169 if (o->mem_type == MEM_CUDA_MALLOC)
2172 if (o->compress_percentage || o->dedupe_percentage) {
2173 unsigned int perc = td->o.compress_percentage;
2174 struct frand_state *rs;
2175 unsigned int left = max_bs;
2176 unsigned int this_write;
2179 rs = get_buf_state(td);
2181 min_write = min(min_write, left);
2184 this_write = min_not_zero(min_write,
2185 td->o.compress_chunk);
2187 fill_random_buf_percentage(rs, buf, perc,
2188 this_write, this_write,
2190 o->buffer_pattern_bytes);
2192 fill_random_buf(rs, buf, min_write);
2193 this_write = min_write;
2198 save_buf_state(td, rs);
2200 } else if (o->buffer_pattern_bytes)
2201 fill_buffer_pattern(td, buf, max_bs);
2202 else if (o->zero_buffers)
2203 memset(buf, 0, max_bs);
2205 fill_random_buf(get_buf_state(td), buf, max_bs);
2209 * "randomly" fill the buffer contents
2211 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2212 unsigned int min_write, unsigned int max_bs)
2214 io_u->buf_filled_len = 0;
2215 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2218 static int do_sync_file_range(const struct thread_data *td,
2221 off64_t offset, nbytes;
2223 offset = f->first_write;
2224 nbytes = f->last_write - f->first_write;
2229 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2232 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2236 if (io_u->ddir == DDIR_SYNC) {
2237 ret = fsync(io_u->file->fd);
2238 } else if (io_u->ddir == DDIR_DATASYNC) {
2239 #ifdef CONFIG_FDATASYNC
2240 ret = fdatasync(io_u->file->fd);
2242 ret = io_u->xfer_buflen;
2243 io_u->error = EINVAL;
2245 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2246 ret = do_sync_file_range(td, io_u->file);
2248 ret = io_u->xfer_buflen;
2249 io_u->error = EINVAL;
2253 io_u->error = errno;
2258 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2260 #ifndef FIO_HAVE_TRIM
2261 io_u->error = EINVAL;
2264 struct fio_file *f = io_u->file;
2267 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2269 return io_u->xfer_buflen;