15 struct io_completion_data {
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timespec time; /* output */
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 static bool random_map_free(struct fio_file *f, const uint64_t block)
29 return !axmap_isset(f->io_axmap, block);
33 * Mark a given offset as used in the map.
35 static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
36 uint64_t offset, uint64_t buflen)
38 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
39 struct fio_file *f = io_u->file;
40 unsigned long long nr_blocks;
43 block = (offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (buflen + min_bs - 1) / min_bs;
45 assert(nr_blocks > 0);
47 if (!(io_u->flags & IO_U_F_BUSY_OK)) {
48 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 assert(nr_blocks > 0);
52 if ((nr_blocks * min_bs) < buflen)
53 buflen = nr_blocks * min_bs;
58 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
64 assert(ddir_rw(ddir));
67 * Hmm, should we make sure that ->io_size <= ->real_file_size?
68 * -> not for now since there is code assuming it could go either.
70 max_size = f->io_size;
71 if (max_size > f->real_file_size)
72 max_size = f->real_file_size;
74 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
75 max_size = td->o.zone_range;
77 if (td->o.min_bs[ddir] > td->o.ba[ddir])
78 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
80 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
87 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
88 enum fio_ddir ddir, uint64_t *b,
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
96 r = __rand(&td->random_state);
98 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
100 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
104 assert(fio_file_lfsr(f));
106 if (lfsr_next(&f->lfsr, &off))
113 * if we are not maintaining a random map, we are done.
115 if (!file_randommap(td, f))
119 * calculate map offset and check if it's free
121 if (random_map_free(f, *b))
124 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
125 (unsigned long long) *b);
127 *b = axmap_next_free(f->io_axmap, *b);
128 if (*b == (uint64_t) -1ULL)
134 static int __get_next_rand_offset_zipf(struct thread_data *td,
135 struct fio_file *f, enum fio_ddir ddir,
138 *b = zipf_next(&f->zipf);
142 static int __get_next_rand_offset_pareto(struct thread_data *td,
143 struct fio_file *f, enum fio_ddir ddir,
146 *b = pareto_next(&f->zipf);
150 static int __get_next_rand_offset_gauss(struct thread_data *td,
151 struct fio_file *f, enum fio_ddir ddir,
154 *b = gauss_next(&f->gauss);
158 static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
160 enum fio_ddir ddir, uint64_t *b)
162 struct zone_split_index *zsi;
163 uint64_t lastb, send, stotal;
166 lastb = last_block(td, f, ddir);
170 if (!td->o.zone_split_nr[ddir]) {
172 return __get_next_rand_offset(td, f, ddir, b, lastb);
176 * Generate a value, v, between 1 and 100, both inclusive
178 v = rand_between(&td->zone_state, 1, 100);
181 * Find our generated table. 'send' is the end block of this zone,
182 * 'stotal' is our start offset.
184 zsi = &td->zone_state_index[ddir][v - 1];
185 stotal = zsi->size_prev / td->o.ba[ddir];
186 send = zsi->size / td->o.ba[ddir];
189 * Should never happen
192 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
193 log_err("fio: bug in zoned generation\n");
195 } else if (send > lastb) {
197 * This happens if the user specifies ranges that exceed
198 * the file/device size. We can't handle that gracefully,
201 log_err("fio: zoned_abs sizes exceed file size\n");
206 * Generate index from 0..send-stotal
208 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
215 static int __get_next_rand_offset_zoned(struct thread_data *td,
216 struct fio_file *f, enum fio_ddir ddir,
219 unsigned int v, send, stotal;
220 uint64_t offset, lastb;
221 struct zone_split_index *zsi;
223 lastb = last_block(td, f, ddir);
227 if (!td->o.zone_split_nr[ddir]) {
229 return __get_next_rand_offset(td, f, ddir, b, lastb);
233 * Generate a value, v, between 1 and 100, both inclusive
235 v = rand_between(&td->zone_state, 1, 100);
237 zsi = &td->zone_state_index[ddir][v - 1];
238 stotal = zsi->size_perc_prev;
239 send = zsi->size_perc;
242 * Should never happen
245 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
246 log_err("fio: bug in zoned generation\n");
251 * 'send' is some percentage below or equal to 100 that
252 * marks the end of the current IO range. 'stotal' marks
253 * the start, in percent.
256 offset = stotal * lastb / 100ULL;
260 lastb = lastb * (send - stotal) / 100ULL;
263 * Generate index from 0..send-of-lastb
265 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
269 * Add our start offset, if any
277 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
278 enum fio_ddir ddir, uint64_t *b)
280 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
283 lastb = last_block(td, f, ddir);
287 return __get_next_rand_offset(td, f, ddir, b, lastb);
288 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
289 return __get_next_rand_offset_zipf(td, f, ddir, b);
290 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
291 return __get_next_rand_offset_pareto(td, f, ddir, b);
292 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
293 return __get_next_rand_offset_gauss(td, f, ddir, b);
294 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
295 return __get_next_rand_offset_zoned(td, f, ddir, b);
296 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
297 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
299 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
303 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
307 if (td->o.perc_rand[ddir] == 100)
310 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
312 return v <= td->o.perc_rand[ddir];
315 static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
317 struct thread_options *o = &td->o;
319 if (o->invalidate_cache && !o->odirect) {
322 ret = file_invalidate_cache(td, f);
326 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
327 enum fio_ddir ddir, uint64_t *b)
329 if (!get_next_rand_offset(td, f, ddir, b))
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
334 fio_file_reset(td, f);
335 loop_cache_invalidate(td, f);
336 if (!get_next_rand_offset(td, f, ddir, b))
340 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
341 f->file_name, (unsigned long long) f->last_pos[ddir],
342 (unsigned long long) f->real_file_size);
346 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
347 enum fio_ddir ddir, uint64_t *offset)
349 struct thread_options *o = &td->o;
351 assert(ddir_rw(ddir));
354 * If we reach the end for a time based run, reset us back to 0
355 * and invalidate the cache, if we need to.
357 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
359 f->last_pos[ddir] = f->file_offset;
360 loop_cache_invalidate(td, f);
363 if (f->last_pos[ddir] < f->real_file_size) {
367 * Only rewind if we already hit the end
369 if (f->last_pos[ddir] == f->file_offset &&
370 f->file_offset && o->ddir_seq_add < 0) {
371 if (f->real_file_size > f->io_size)
372 f->last_pos[ddir] = f->io_size;
374 f->last_pos[ddir] = f->real_file_size;
377 pos = f->last_pos[ddir] - f->file_offset;
378 if (pos && o->ddir_seq_add) {
379 pos += o->ddir_seq_add;
382 * If we reach beyond the end of the file
383 * with holed IO, wrap around to the
384 * beginning again. If we're doing backwards IO,
387 if (pos >= f->real_file_size) {
388 if (o->ddir_seq_add > 0)
389 pos = f->file_offset;
391 if (f->real_file_size > f->io_size)
394 pos = f->real_file_size;
396 pos += o->ddir_seq_add;
408 static int get_next_block(struct thread_data *td, struct io_u *io_u,
409 enum fio_ddir ddir, int rw_seq,
412 struct fio_file *f = io_u->file;
416 assert(ddir_rw(ddir));
422 if (should_do_random(td, ddir)) {
423 ret = get_next_rand_block(td, f, ddir, &b);
427 io_u_set(td, io_u, IO_U_F_BUSY_OK);
428 ret = get_next_seq_offset(td, f, ddir, &offset);
430 ret = get_next_rand_block(td, f, ddir, &b);
434 ret = get_next_seq_offset(td, f, ddir, &offset);
437 io_u_set(td, io_u, IO_U_F_BUSY_OK);
440 if (td->o.rw_seq == RW_SEQ_SEQ) {
441 ret = get_next_seq_offset(td, f, ddir, &offset);
443 ret = get_next_rand_block(td, f, ddir, &b);
446 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
447 if (f->last_start[ddir] != -1ULL)
448 offset = f->last_start[ddir] - f->file_offset;
453 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
460 io_u->offset = offset;
462 io_u->offset = b * td->o.ba[ddir];
464 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
473 * For random io, generate a random new block and see if it's used. Repeat
474 * until we find a free one. For sequential io, just return the end of
475 * the last io issued.
477 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
480 struct fio_file *f = io_u->file;
481 enum fio_ddir ddir = io_u->ddir;
484 assert(ddir_rw(ddir));
486 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
488 td->ddir_seq_nr = td->o.ddir_seq_nr;
491 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
494 if (io_u->offset >= f->io_size) {
495 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
496 (unsigned long long) io_u->offset,
497 (unsigned long long) f->io_size);
501 io_u->offset += f->file_offset;
502 if (io_u->offset >= f->real_file_size) {
503 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
504 (unsigned long long) io_u->offset,
505 (unsigned long long) f->real_file_size);
512 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
513 unsigned long long buflen)
515 struct fio_file *f = io_u->file;
517 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
520 static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
523 int ddir = io_u->ddir;
524 unsigned long long buflen = 0;
525 unsigned long long minbs, maxbs;
526 uint64_t frand_max, r;
529 assert(ddir_rw(ddir));
531 if (td->o.bs_is_seq_rand)
532 ddir = is_random ? DDIR_WRITE : DDIR_READ;
534 minbs = td->o.min_bs[ddir];
535 maxbs = td->o.max_bs[ddir];
541 * If we can't satisfy the min block size from here, then fail
543 if (!io_u_fits(td, io_u, minbs))
546 frand_max = rand_max(&td->bsrange_state[ddir]);
548 r = __rand(&td->bsrange_state[ddir]);
550 if (!td->o.bssplit_nr[ddir]) {
551 buflen = minbs + (unsigned long long) ((double) maxbs *
552 (r / (frand_max + 1.0)));
557 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
558 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564 if ((r / perc <= frand_max / 100ULL) &&
565 io_u_fits(td, io_u, buflen))
570 power_2 = is_power_of_2(minbs);
571 if (!td->o.bs_unaligned && power_2)
572 buflen &= ~(minbs - 1);
573 else if (!td->o.bs_unaligned && !power_2)
574 buflen -= buflen % minbs;
577 } while (!io_u_fits(td, io_u, buflen));
582 static void set_rwmix_bytes(struct thread_data *td)
587 * we do time or byte based switch. this is needed because
588 * buffered writes may issue a lot quicker than they complete,
589 * whereas reads do not.
591 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
592 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
595 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
599 v = rand_between(&td->rwmix_state, 1, 100);
601 if (v <= td->o.rwmix[DDIR_READ])
607 int io_u_quiesce(struct thread_data *td)
609 int ret = 0, completed = 0, err = 0;
612 * We are going to sleep, ensure that we flush anything pending as
613 * not to skew our latency numbers.
615 * Changed to only monitor 'in flight' requests here instead of the
616 * td->cur_depth, b/c td->cur_depth does not accurately represent
617 * io's that have been actually submitted to an async engine,
618 * and cur_depth is meaningless for sync engines.
620 if (td->io_u_queued || td->cur_depth)
623 while (td->io_u_in_flight) {
624 ret = io_u_queued_complete(td, 1);
631 if (td->flags & TD_F_REGROW_LOGS)
640 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
642 enum fio_ddir odir = ddir ^ 1;
646 assert(ddir_rw(ddir));
647 now = utime_since_now(&td->epoch);
650 * if rate_next_io_time is in the past, need to catch up to rate
652 if (td->rate_next_io_time[ddir] <= now)
656 * We are ahead of rate in this direction. See if we
659 if (td_rw(td) && td->o.rwmix[odir]) {
661 * Other direction is behind rate, switch
663 if (td->rate_next_io_time[odir] <= now)
667 * Both directions are ahead of rate. sleep the min,
668 * switch if necessary
670 if (td->rate_next_io_time[ddir] <=
671 td->rate_next_io_time[odir]) {
672 usec = td->rate_next_io_time[ddir] - now;
674 usec = td->rate_next_io_time[odir] - now;
678 usec = td->rate_next_io_time[ddir] - now;
680 if (td->o.io_submit_mode == IO_MODE_INLINE)
683 if (td->o.timeout && ((usec + now) > td->o.timeout)) {
685 * check if the usec is capable of taking negative values
687 if (now > td->o.timeout) {
691 usec = td->o.timeout - now;
693 usec_sleep(td, usec);
695 now = utime_since_now(&td->epoch);
696 if ((td->o.timeout && (now > td->o.timeout)) || td->terminate)
703 * Return the data direction for the next io_u. If the job is a
704 * mixed read/write workload, check the rwmix cycle and switch if
707 static enum fio_ddir get_rw_ddir(struct thread_data *td)
712 * See if it's time to fsync/fdatasync/sync_file_range first,
713 * and if not then move on to check regular I/Os.
715 if (should_fsync(td)) {
716 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
717 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
720 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
721 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
722 return DDIR_DATASYNC;
724 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
725 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
726 return DDIR_SYNC_FILE_RANGE;
731 * Check if it's time to seed a new data direction.
733 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
735 * Put a top limit on how many bytes we do for
736 * one data direction, to avoid overflowing the
739 ddir = get_rand_ddir(td);
741 if (ddir != td->rwmix_ddir)
744 td->rwmix_ddir = ddir;
746 ddir = td->rwmix_ddir;
747 } else if (td_read(td))
749 else if (td_write(td))
751 else if (td_trim(td))
756 td->rwmix_ddir = rate_ddir(td, ddir);
757 return td->rwmix_ddir;
760 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
762 enum fio_ddir ddir = get_rw_ddir(td);
764 if (td->o.zone_mode == ZONE_MODE_ZBD)
765 ddir = zbd_adjust_ddir(td, io_u, ddir);
767 if (td_trimwrite(td)) {
768 struct fio_file *f = io_u->file;
769 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
775 io_u->ddir = io_u->acct_ddir = ddir;
777 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
778 td->o.barrier_blocks &&
779 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
780 td->io_issues[DDIR_WRITE])
781 io_u_set(td, io_u, IO_U_F_BARRIER);
784 void put_file_log(struct thread_data *td, struct fio_file *f)
786 unsigned int ret = put_file(td, f);
789 td_verror(td, ret, "file close");
792 void put_io_u(struct thread_data *td, struct io_u *io_u)
794 const bool needs_lock = td_async_processing(td);
804 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
805 put_file_log(td, io_u->file);
808 io_u_set(td, io_u, IO_U_F_FREE);
810 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
812 assert(!(td->flags & TD_F_CHILD));
814 io_u_qpush(&td->io_u_freelist, io_u);
815 td_io_u_free_notify(td);
818 __td_io_u_unlock(td);
821 void clear_io_u(struct thread_data *td, struct io_u *io_u)
823 io_u_clear(td, io_u, IO_U_F_FLIGHT);
827 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
829 const bool needs_lock = td_async_processing(td);
830 struct io_u *__io_u = *io_u;
831 enum fio_ddir ddir = acct_ddir(__io_u);
833 dprint(FD_IO, "requeue %p\n", __io_u);
841 io_u_set(td, __io_u, IO_U_F_FREE);
842 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
843 td->io_issues[ddir]--;
845 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
846 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
848 assert(!(td->flags & TD_F_CHILD));
851 io_u_rpush(&td->io_u_requeues, __io_u);
852 td_io_u_free_notify(td);
855 __td_io_u_unlock(td);
860 static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
862 struct fio_file *f = io_u->file;
864 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
865 assert(td->o.zone_size);
866 assert(td->o.zone_range);
869 * See if it's time to switch to a new zone
871 if (td->zone_bytes >= td->o.zone_size) {
873 f->file_offset += td->o.zone_range + td->o.zone_skip;
876 * Wrap from the beginning, if we exceed the file size
878 if (f->file_offset >= f->real_file_size)
879 f->file_offset = get_start_offset(td, f);
881 f->last_pos[io_u->ddir] = f->file_offset;
882 td->io_skip_bytes += td->o.zone_skip;
886 * If zone_size > zone_range, then maintain the same zone until
887 * zone_bytes >= zone_size.
889 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
890 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
891 f->file_offset, f->last_pos[io_u->ddir]);
892 f->last_pos[io_u->ddir] = f->file_offset;
896 * For random: if 'norandommap' is not set and zone_size > zone_range,
897 * map needs to be reset as it's done with zone_range everytime.
899 if ((td->zone_bytes % td->o.zone_range) == 0)
900 fio_file_reset(td, f);
903 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
907 enum io_u_action ret;
909 if (td_ioengine_flagged(td, FIO_NOIO))
912 set_rw_ddir(td, io_u);
914 if (io_u->ddir == DDIR_INVAL) {
915 dprint(FD_IO, "invalid direction received ddir = %d", io_u->ddir);
919 * fsync() or fdatasync() or trim etc, we are done
921 if (!ddir_rw(io_u->ddir))
924 if (td->o.zone_mode == ZONE_MODE_STRIDED)
925 setup_strided_zone_mode(td, io_u);
926 else if (td->o.zone_mode == ZONE_MODE_ZBD)
927 setup_zbd_zone_mode(td, io_u);
930 * No log, let the seq/rand engine retrieve the next buflen and
933 if (get_next_offset(td, io_u, &is_random)) {
934 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
938 io_u->buflen = get_next_buflen(td, io_u, is_random);
940 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
944 offset = io_u->offset;
945 if (td->o.zone_mode == ZONE_MODE_ZBD) {
946 ret = zbd_adjust_block(td, io_u);
951 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
952 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
954 (unsigned long long) io_u->offset, io_u->buflen,
955 (unsigned long long) io_u->file->real_file_size);
960 * mark entry before potentially trimming io_u
962 if (td_random(td) && file_randommap(td, io_u->file))
963 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
966 dprint_io_u(io_u, "fill");
967 td->zone_bytes += io_u->buflen;
971 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
1000 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1002 __io_u_mark_map(td->ts.io_u_submit, nr);
1003 td->ts.total_submit++;
1006 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1008 __io_u_mark_map(td->ts.io_u_complete, nr);
1009 td->ts.total_complete++;
1012 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1016 switch (td->cur_depth) {
1038 td->ts.io_u_map[idx] += nr;
1041 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1045 assert(nsec < 1000);
1078 assert(idx < FIO_IO_U_LAT_N_NR);
1079 td->ts.io_u_lat_n[idx]++;
1082 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1086 assert(usec < 1000 && usec >= 1);
1119 assert(idx < FIO_IO_U_LAT_U_NR);
1120 td->ts.io_u_lat_u[idx]++;
1123 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1166 assert(idx < FIO_IO_U_LAT_M_NR);
1167 td->ts.io_u_lat_m[idx]++;
1170 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1173 io_u_mark_lat_nsec(td, nsec);
1174 else if (nsec < 1000000)
1175 io_u_mark_lat_usec(td, nsec / 1000);
1177 io_u_mark_lat_msec(td, nsec / 1000000);
1180 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1182 unsigned long fileno;
1184 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1185 uint64_t frand_max = rand_max(&td->next_file_state);
1188 r = __rand(&td->next_file_state);
1189 return (unsigned int) ((double) td->o.nr_files
1190 * (r / (frand_max + 1.0)));
1193 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1194 fileno = zipf_next(&td->next_file_zipf);
1195 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1196 fileno = pareto_next(&td->next_file_zipf);
1197 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1198 fileno = gauss_next(&td->next_file_gauss);
1200 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1205 return fileno >> FIO_FSERVICE_SHIFT;
1209 * Get next file to service by choosing one at random
1211 static struct fio_file *get_next_file_rand(struct thread_data *td,
1212 enum fio_file_flags goodf,
1213 enum fio_file_flags badf)
1221 fno = __get_next_fileno_rand(td);
1224 if (fio_file_done(f))
1227 if (!fio_file_open(f)) {
1230 if (td->nr_open_files >= td->o.open_files)
1231 return ERR_PTR(-EBUSY);
1233 err = td_io_open_file(td, f);
1239 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1240 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1244 td_io_close_file(td, f);
1249 * Get next file to service by doing round robin between all available ones
1251 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1254 unsigned int old_next_file = td->next_file;
1260 f = td->files[td->next_file];
1263 if (td->next_file >= td->o.nr_files)
1266 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1267 if (fio_file_done(f)) {
1272 if (!fio_file_open(f)) {
1275 if (td->nr_open_files >= td->o.open_files)
1276 return ERR_PTR(-EBUSY);
1278 err = td_io_open_file(td, f);
1280 dprint(FD_FILE, "error %d on open of %s\n",
1288 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1290 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1294 td_io_close_file(td, f);
1297 } while (td->next_file != old_next_file);
1299 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1303 static struct fio_file *__get_next_file(struct thread_data *td)
1307 assert(td->o.nr_files <= td->files_index);
1309 if (td->nr_done_files >= td->o.nr_files) {
1310 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1311 " nr_files=%d\n", td->nr_open_files,
1317 f = td->file_service_file;
1318 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1319 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1321 if (td->file_service_left--)
1325 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1326 td->o.file_service_type == FIO_FSERVICE_SEQ)
1327 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1329 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1334 td->file_service_file = f;
1335 td->file_service_left = td->file_service_nr - 1;
1338 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1340 dprint(FD_FILE, "get_next_file: NULL\n");
1344 static struct fio_file *get_next_file(struct thread_data *td)
1346 return __get_next_file(td);
1349 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1354 f = get_next_file(td);
1355 if (IS_ERR_OR_NULL(f))
1361 if (!fill_io_u(td, io_u))
1366 put_file_log(td, f);
1367 td_io_close_file(td, f);
1369 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1370 fio_file_reset(td, f);
1372 fio_file_set_done(f);
1373 td->nr_done_files++;
1374 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1375 td->nr_done_files, td->o.nr_files);
1382 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1383 unsigned long long tnsec, unsigned long long max_nsec)
1386 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1387 td_verror(td, ETIMEDOUT, "max latency exceeded");
1388 icd->error = ETIMEDOUT;
1391 static void lat_new_cycle(struct thread_data *td)
1393 fio_gettime(&td->latency_ts, NULL);
1394 td->latency_ios = ddir_rw_sum(td->io_blocks);
1395 td->latency_failed = 0;
1399 * We had an IO outside the latency target. Reduce the queue depth. If we
1400 * are at QD=1, then it's time to give up.
1402 static bool __lat_target_failed(struct thread_data *td)
1404 if (td->latency_qd == 1)
1407 td->latency_qd_high = td->latency_qd;
1409 if (td->latency_qd == td->latency_qd_low)
1410 td->latency_qd_low--;
1412 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1413 td->latency_stable_count = 0;
1415 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1418 * When we ramp QD down, quiesce existing IO to prevent
1419 * a storm of ramp downs due to pending higher depth.
1426 static bool lat_target_failed(struct thread_data *td)
1428 if (td->o.latency_percentile.u.f == 100.0)
1429 return __lat_target_failed(td);
1431 td->latency_failed++;
1435 void lat_target_init(struct thread_data *td)
1437 td->latency_end_run = 0;
1439 if (td->o.latency_target) {
1440 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1441 fio_gettime(&td->latency_ts, NULL);
1443 td->latency_qd_high = td->o.iodepth;
1444 td->latency_qd_low = 1;
1445 td->latency_ios = ddir_rw_sum(td->io_blocks);
1447 td->latency_qd = td->o.iodepth;
1450 void lat_target_reset(struct thread_data *td)
1452 if (!td->latency_end_run)
1453 lat_target_init(td);
1456 static void lat_target_success(struct thread_data *td)
1458 const unsigned int qd = td->latency_qd;
1459 struct thread_options *o = &td->o;
1461 td->latency_qd_low = td->latency_qd;
1463 if (td->latency_qd + 1 == td->latency_qd_high) {
1465 * latency_qd will not incease on lat_target_success(), so
1466 * called stable. If we stick with this queue depth, the
1467 * final latency is likely lower than latency_target. Fix
1468 * this by increasing latency_qd_high slowly. Use a naive
1469 * heuristic here. If we get lat_target_success() 3 times
1470 * in a row, increase latency_qd_high by 1.
1472 if (++td->latency_stable_count >= 3) {
1473 td->latency_qd_high++;
1474 td->latency_stable_count = 0;
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 (!o->latency_run && 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 const bool needs_lock = td_async_processing(td);
1557 struct io_u *io_u = NULL;
1567 if (!io_u_rempty(&td->io_u_requeues))
1568 io_u = io_u_rpop(&td->io_u_requeues);
1569 else if (!queue_full(td)) {
1570 io_u = io_u_qpop(&td->io_u_freelist);
1575 io_u->end_io = NULL;
1579 assert(io_u->flags & IO_U_F_FREE);
1580 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1581 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1582 IO_U_F_VER_LIST | IO_U_F_PRIORITY);
1585 io_u->acct_ddir = -1;
1587 assert(!(td->flags & TD_F_CHILD));
1588 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1590 } else if (td_async_processing(td)) {
1592 * We ran out, wait for async verify threads to finish and
1595 assert(!(td->flags & TD_F_CHILD));
1596 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1603 __td_io_u_unlock(td);
1608 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1610 if (!(td->flags & TD_F_TRIM_BACKLOG))
1612 if (!td->trim_entries)
1615 if (td->trim_batch) {
1617 if (get_next_trim(td, io_u))
1619 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1620 td->last_ddir != DDIR_READ) {
1621 td->trim_batch = td->o.trim_batch;
1622 if (!td->trim_batch)
1623 td->trim_batch = td->o.trim_backlog;
1624 if (get_next_trim(td, io_u))
1631 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1633 if (!(td->flags & TD_F_VER_BACKLOG))
1636 if (td->io_hist_len) {
1639 if (td->verify_batch)
1641 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1642 td->last_ddir != DDIR_READ) {
1643 td->verify_batch = td->o.verify_batch;
1644 if (!td->verify_batch)
1645 td->verify_batch = td->o.verify_backlog;
1649 if (get_verify && !get_next_verify(td, io_u)) {
1659 * Fill offset and start time into the buffer content, to prevent too
1660 * easy compressible data for simple de-dupe attempts. Do this for every
1661 * 512b block in the range, since that should be the smallest block size
1662 * we can expect from a device.
1664 static void small_content_scramble(struct io_u *io_u)
1666 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1667 unsigned int offset;
1668 uint64_t boffset, *iptr;
1675 boffset = io_u->offset;
1677 if (io_u->buf_filled_len)
1678 io_u->buf_filled_len = 0;
1681 * Generate random index between 0..7. We do chunks of 512b, if
1682 * we assume a cacheline is 64 bytes, then we have 8 of those.
1683 * Scramble content within the blocks in the same cacheline to
1686 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1688 for (i = 0; i < nr_blocks; i++) {
1690 * Fill offset into start of cacheline, time into end
1693 iptr = (void *) p + (offset << 6);
1696 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1697 iptr[0] = io_u->start_time.tv_sec;
1698 iptr[1] = io_u->start_time.tv_nsec;
1706 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1707 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1709 struct io_u *get_io_u(struct thread_data *td)
1713 int do_scramble = 0;
1716 io_u = __get_io_u(td);
1718 dprint(FD_IO, "__get_io_u failed\n");
1722 if (check_get_verify(td, io_u))
1724 if (check_get_trim(td, io_u))
1728 * from a requeue, io_u already setup
1734 * If using an iolog, grab next piece if any available.
1736 if (td->flags & TD_F_READ_IOLOG) {
1737 if (read_iolog_get(td, io_u))
1739 } else if (set_io_u_file(td, io_u)) {
1741 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1747 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1751 assert(fio_file_open(f));
1753 if (ddir_rw(io_u->ddir)) {
1754 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1755 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1759 f->last_start[io_u->ddir] = io_u->offset;
1760 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1762 if (io_u->ddir == DDIR_WRITE) {
1763 if (td->flags & TD_F_REFILL_BUFFERS) {
1764 io_u_fill_buffer(td, io_u,
1765 td->o.min_bs[DDIR_WRITE],
1767 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1768 !(td->flags & TD_F_COMPRESS) &&
1769 !(td->flags & TD_F_DO_VERIFY))
1771 } else if (io_u->ddir == DDIR_READ) {
1773 * Reset the buf_filled parameters so next time if the
1774 * buffer is used for writes it is refilled.
1776 io_u->buf_filled_len = 0;
1781 * Set io data pointers.
1783 io_u->xfer_buf = io_u->buf;
1784 io_u->xfer_buflen = io_u->buflen;
1788 if (!td_io_prep(td, io_u)) {
1789 if (!td->o.disable_lat)
1790 fio_gettime(&io_u->start_time, NULL);
1793 small_content_scramble(io_u);
1798 dprint(FD_IO, "get_io_u failed\n");
1800 return ERR_PTR(ret);
1803 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1805 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1807 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1810 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1811 io_u->file ? " on file " : "",
1812 io_u->file ? io_u->file->file_name : "",
1813 strerror(io_u->error),
1814 io_ddir_name(io_u->ddir),
1815 io_u->offset, io_u->xfer_buflen);
1817 if (td->io_ops->errdetails) {
1818 char *err = td->io_ops->errdetails(io_u);
1820 log_err("fio: %s\n", err);
1825 td_verror(td, io_u->error, "io_u error");
1828 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1830 __io_u_log_error(td, io_u);
1832 __io_u_log_error(td->parent, io_u);
1835 static inline bool gtod_reduce(struct thread_data *td)
1837 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1838 || td->o.gtod_reduce;
1841 static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1843 uint32_t *info = io_u_block_info(td, io_u);
1845 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1848 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1851 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1852 struct io_completion_data *icd,
1853 const enum fio_ddir idx, unsigned int bytes)
1855 const int no_reduce = !gtod_reduce(td);
1856 unsigned long long llnsec = 0;
1861 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1865 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1867 if (!td->o.disable_lat) {
1868 unsigned long long tnsec;
1870 tnsec = ntime_since(&io_u->start_time, &icd->time);
1871 add_lat_sample(td, idx, tnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1873 if (td->flags & TD_F_PROFILE_OPS) {
1874 struct prof_io_ops *ops = &td->prof_io_ops;
1877 icd->error = ops->io_u_lat(td, tnsec);
1880 if (td->o.max_latency && tnsec > td->o.max_latency)
1881 lat_fatal(td, icd, tnsec, td->o.max_latency);
1882 if (td->o.latency_target && tnsec > td->o.latency_target) {
1883 if (lat_target_failed(td))
1884 lat_fatal(td, icd, tnsec, td->o.latency_target);
1889 if (!td->o.disable_clat) {
1890 add_clat_sample(td, idx, llnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1891 io_u_mark_latency(td, llnsec);
1894 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1895 add_bw_sample(td, io_u, bytes, llnsec);
1897 if (no_reduce && per_unit_log(td->iops_log))
1898 add_iops_sample(td, io_u, bytes);
1899 } else if (ddir_sync(idx) && !td->o.disable_clat)
1900 add_sync_clat_sample(&td->ts, llnsec);
1902 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1903 trim_block_info(td, io_u);
1906 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1907 uint64_t offset, unsigned int bytes)
1914 if (f->first_write == -1ULL || offset < f->first_write)
1915 f->first_write = offset;
1916 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1917 f->last_write = offset + bytes;
1919 if (!f->last_write_comp)
1922 idx = f->last_write_idx++;
1923 f->last_write_comp[idx] = offset;
1924 if (f->last_write_idx == td->o.iodepth)
1925 f->last_write_idx = 0;
1928 static bool should_account(struct thread_data *td)
1930 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1931 td->runstate == TD_VERIFYING);
1934 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1935 struct io_completion_data *icd)
1937 struct io_u *io_u = *io_u_ptr;
1938 enum fio_ddir ddir = io_u->ddir;
1939 struct fio_file *f = io_u->file;
1941 dprint_io_u(io_u, "complete");
1943 assert(io_u->flags & IO_U_F_FLIGHT);
1944 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1947 * Mark IO ok to verify
1951 * Remove errored entry from the verification list
1954 unlog_io_piece(td, io_u);
1956 atomic_store_release(&io_u->ipo->flags,
1957 io_u->ipo->flags & ~IP_F_IN_FLIGHT);
1961 if (ddir_sync(ddir)) {
1962 td->last_was_sync = true;
1964 f->first_write = -1ULL;
1965 f->last_write = -1ULL;
1967 if (should_account(td))
1968 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1972 td->last_was_sync = false;
1973 td->last_ddir = ddir;
1975 if (!io_u->error && ddir_rw(ddir)) {
1976 unsigned long long bytes = io_u->buflen - io_u->resid;
1979 td->io_blocks[ddir]++;
1980 td->io_bytes[ddir] += bytes;
1982 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1983 td->this_io_blocks[ddir]++;
1984 td->this_io_bytes[ddir] += bytes;
1987 if (ddir == DDIR_WRITE)
1988 file_log_write_comp(td, f, io_u->offset, bytes);
1990 if (should_account(td))
1991 account_io_completion(td, io_u, icd, ddir, bytes);
1993 icd->bytes_done[ddir] += bytes;
1996 ret = io_u->end_io(td, io_u_ptr);
1998 if (ret && !icd->error)
2001 } else if (io_u->error) {
2002 icd->error = io_u->error;
2003 io_u_log_error(td, io_u);
2006 enum error_type_bit eb = td_error_type(ddir, icd->error);
2008 if (!td_non_fatal_error(td, eb, icd->error))
2012 * If there is a non_fatal error, then add to the error count
2013 * and clear all the errors.
2015 update_error_count(td, icd->error);
2023 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2028 if (!gtod_reduce(td))
2029 fio_gettime(&icd->time, NULL);
2034 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2035 icd->bytes_done[ddir] = 0;
2038 static void ios_completed(struct thread_data *td,
2039 struct io_completion_data *icd)
2044 for (i = 0; i < icd->nr; i++) {
2045 io_u = td->io_ops->event(td, i);
2047 io_completed(td, &io_u, icd);
2055 * Complete a single io_u for the sync engines.
2057 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2059 struct io_completion_data icd;
2062 init_icd(td, &icd, 1);
2063 io_completed(td, &io_u, &icd);
2069 td_verror(td, icd.error, "io_u_sync_complete");
2073 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2074 td->bytes_done[ddir] += icd.bytes_done[ddir];
2080 * Called to complete min_events number of io for the async engines.
2082 int io_u_queued_complete(struct thread_data *td, int min_evts)
2084 struct io_completion_data icd;
2085 struct timespec *tvp = NULL;
2087 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2089 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2093 else if (min_evts > td->cur_depth)
2094 min_evts = td->cur_depth;
2096 /* No worries, td_io_getevents fixes min and max if they are
2097 * set incorrectly */
2098 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2100 td_verror(td, -ret, "td_io_getevents");
2105 init_icd(td, &icd, ret);
2106 ios_completed(td, &icd);
2108 td_verror(td, icd.error, "io_u_queued_complete");
2112 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2113 td->bytes_done[ddir] += icd.bytes_done[ddir];
2119 * Call when io_u is really queued, to update the submission latency.
2121 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2123 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2124 unsigned long slat_time;
2126 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2131 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2132 io_u->offset, io_u_is_prio(io_u));
2137 * See if we should reuse the last seed, if dedupe is enabled
2139 static struct frand_state *get_buf_state(struct thread_data *td)
2143 if (!td->o.dedupe_percentage)
2144 return &td->buf_state;
2145 else if (td->o.dedupe_percentage == 100) {
2146 frand_copy(&td->buf_state_prev, &td->buf_state);
2147 return &td->buf_state;
2150 v = rand_between(&td->dedupe_state, 1, 100);
2152 if (v <= td->o.dedupe_percentage)
2153 return &td->buf_state_prev;
2155 return &td->buf_state;
2158 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2160 if (td->o.dedupe_percentage == 100)
2161 frand_copy(rs, &td->buf_state_prev);
2162 else if (rs == &td->buf_state)
2163 frand_copy(&td->buf_state_prev, rs);
2166 void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2167 unsigned long long max_bs)
2169 struct thread_options *o = &td->o;
2171 if (o->mem_type == MEM_CUDA_MALLOC)
2174 if (o->compress_percentage || o->dedupe_percentage) {
2175 unsigned int perc = td->o.compress_percentage;
2176 struct frand_state *rs;
2177 unsigned long long left = max_bs;
2178 unsigned long long this_write;
2181 rs = get_buf_state(td);
2183 min_write = min(min_write, left);
2186 this_write = min_not_zero(min_write,
2187 (unsigned long long) td->o.compress_chunk);
2189 fill_random_buf_percentage(rs, buf, perc,
2190 this_write, this_write,
2192 o->buffer_pattern_bytes);
2194 fill_random_buf(rs, buf, min_write);
2195 this_write = min_write;
2200 save_buf_state(td, rs);
2202 } else if (o->buffer_pattern_bytes)
2203 fill_buffer_pattern(td, buf, max_bs);
2204 else if (o->zero_buffers)
2205 memset(buf, 0, max_bs);
2207 fill_random_buf(get_buf_state(td), buf, max_bs);
2211 * "randomly" fill the buffer contents
2213 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2214 unsigned long long min_write, unsigned long long max_bs)
2216 io_u->buf_filled_len = 0;
2217 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2220 static int do_sync_file_range(const struct thread_data *td,
2223 uint64_t offset, nbytes;
2225 offset = f->first_write;
2226 nbytes = f->last_write - f->first_write;
2231 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2234 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2238 if (io_u->ddir == DDIR_SYNC) {
2239 ret = fsync(io_u->file->fd);
2240 } else if (io_u->ddir == DDIR_DATASYNC) {
2241 #ifdef CONFIG_FDATASYNC
2242 ret = fdatasync(io_u->file->fd);
2244 ret = io_u->xfer_buflen;
2245 io_u->error = EINVAL;
2247 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2248 ret = do_sync_file_range(td, io_u->file);
2250 ret = io_u->xfer_buflen;
2251 io_u->error = EINVAL;
2255 io_u->error = errno;
2260 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2262 #ifndef FIO_HAVE_TRIM
2263 io_u->error = EINVAL;
2266 struct fio_file *f = io_u->file;
2269 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2271 return io_u->xfer_buflen;