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) &&
358 o->time_based && o->nr_files == 1) {
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));
420 if (td_randtrimwrite(td) && ddir == DDIR_WRITE) {
421 /* don't mark randommap for these writes */
422 io_u_set(td, io_u, IO_U_F_BUSY_OK);
423 offset = f->last_start[DDIR_TRIM];
428 if (should_do_random(td, ddir)) {
429 ret = get_next_rand_block(td, f, ddir, &b);
433 io_u_set(td, io_u, IO_U_F_BUSY_OK);
434 ret = get_next_seq_offset(td, f, ddir, &offset);
436 ret = get_next_rand_block(td, f, ddir, &b);
440 ret = get_next_seq_offset(td, f, ddir, &offset);
443 io_u_set(td, io_u, IO_U_F_BUSY_OK);
446 if (td->o.rw_seq == RW_SEQ_SEQ) {
447 ret = get_next_seq_offset(td, f, ddir, &offset);
449 ret = get_next_rand_block(td, f, ddir, &b);
452 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
453 if (f->last_start[ddir] != -1ULL)
454 offset = f->last_start[ddir] - f->file_offset;
459 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
466 io_u->offset = offset;
468 io_u->offset = b * td->o.ba[ddir];
470 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
473 io_u->verify_offset = io_u->offset;
480 * For random io, generate a random new block and see if it's used. Repeat
481 * until we find a free one. For sequential io, just return the end of
482 * the last io issued.
484 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
487 struct fio_file *f = io_u->file;
488 enum fio_ddir ddir = io_u->ddir;
491 assert(ddir_rw(ddir));
493 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
495 td->ddir_seq_nr = td->o.ddir_seq_nr;
498 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
501 if (io_u->offset >= f->io_size) {
502 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
503 (unsigned long long) io_u->offset,
504 (unsigned long long) f->io_size);
508 io_u->offset += f->file_offset;
509 if (io_u->offset >= f->real_file_size) {
510 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
511 (unsigned long long) io_u->offset,
512 (unsigned long long) f->real_file_size);
517 * For randtrimwrite, we decide whether to issue a trim or a write
518 * based on whether the offsets for the most recent trim and write
519 * operations match. If they don't match that means we just issued a
520 * new trim and the next operation should be a write. If they *do*
521 * match that means we just completed a trim+write pair and the next
522 * command should be a trim.
524 * This works fine for sequential workloads but for random workloads
525 * it's possible to complete a trim+write pair and then have the next
526 * randomly generated offset match the previous offset. If that happens
527 * we need to alter the offset for the last write operation in order
528 * to ensure that we issue a write operation the next time through.
530 if (td_randtrimwrite(td) && ddir == DDIR_TRIM &&
531 f->last_start[DDIR_TRIM] == io_u->offset)
532 f->last_start[DDIR_WRITE]--;
534 io_u->verify_offset = io_u->offset;
538 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
539 unsigned long long buflen)
541 struct fio_file *f = io_u->file;
543 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
546 static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
549 int ddir = io_u->ddir;
550 unsigned long long buflen = 0;
551 unsigned long long minbs, maxbs;
552 uint64_t frand_max, r;
555 assert(ddir_rw(ddir));
557 if (td_randtrimwrite(td) && ddir == DDIR_WRITE) {
558 struct fio_file *f = io_u->file;
560 return f->last_pos[DDIR_TRIM] - f->last_start[DDIR_TRIM];
563 if (td->o.bs_is_seq_rand)
564 ddir = is_random ? DDIR_WRITE : DDIR_READ;
566 minbs = td->o.min_bs[ddir];
567 maxbs = td->o.max_bs[ddir];
573 * If we can't satisfy the min block size from here, then fail
575 if (!io_u_fits(td, io_u, minbs))
578 frand_max = rand_max(&td->bsrange_state[ddir]);
580 r = __rand(&td->bsrange_state[ddir]);
582 if (!td->o.bssplit_nr[ddir]) {
583 buflen = minbs + (unsigned long long) ((double) maxbs *
584 (r / (frand_max + 1.0)));
589 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
590 struct bssplit *bsp = &td->o.bssplit[ddir][i];
596 if ((r / perc <= frand_max / 100ULL) &&
597 io_u_fits(td, io_u, buflen))
602 power_2 = is_power_of_2(minbs);
603 if (!td->o.bs_unaligned && power_2)
604 buflen &= ~(minbs - 1);
605 else if (!td->o.bs_unaligned && !power_2)
606 buflen -= buflen % minbs;
609 } while (!io_u_fits(td, io_u, buflen));
614 static void set_rwmix_bytes(struct thread_data *td)
619 * we do time or byte based switch. this is needed because
620 * buffered writes may issue a lot quicker than they complete,
621 * whereas reads do not.
623 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
624 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
627 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
631 v = rand_between(&td->rwmix_state, 1, 100);
633 if (v <= td->o.rwmix[DDIR_READ])
639 int io_u_quiesce(struct thread_data *td)
641 int ret = 0, completed = 0, err = 0;
644 * We are going to sleep, ensure that we flush anything pending as
645 * not to skew our latency numbers.
647 * Changed to only monitor 'in flight' requests here instead of the
648 * td->cur_depth, b/c td->cur_depth does not accurately represent
649 * io's that have been actually submitted to an async engine,
650 * and cur_depth is meaningless for sync engines.
652 if (td->io_u_queued || td->cur_depth)
655 while (td->io_u_in_flight) {
656 ret = io_u_queued_complete(td, 1);
663 if (td->flags & TD_F_REGROW_LOGS)
672 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
674 enum fio_ddir odir = ddir ^ 1;
678 assert(ddir_rw(ddir));
679 now = utime_since_now(&td->epoch);
682 * if rate_next_io_time is in the past, need to catch up to rate
684 if (td->rate_next_io_time[ddir] <= now)
688 * We are ahead of rate in this direction. See if we
691 if (td_rw(td) && td->o.rwmix[odir]) {
693 * Other direction is behind rate, switch
695 if (td->rate_next_io_time[odir] <= now)
699 * Both directions are ahead of rate. sleep the min,
700 * switch if necessary
702 if (td->rate_next_io_time[ddir] <=
703 td->rate_next_io_time[odir]) {
704 usec = td->rate_next_io_time[ddir] - now;
706 usec = td->rate_next_io_time[odir] - now;
710 usec = td->rate_next_io_time[ddir] - now;
712 if (td->o.io_submit_mode == IO_MODE_INLINE)
715 if (td->o.timeout && ((usec + now) > td->o.timeout)) {
717 * check if the usec is capable of taking negative values
719 if (now > td->o.timeout) {
723 usec = td->o.timeout - now;
725 usec_sleep(td, usec);
727 now = utime_since_now(&td->epoch);
728 if ((td->o.timeout && (now > td->o.timeout)) || td->terminate)
735 * Return the data direction for the next io_u. If the job is a
736 * mixed read/write workload, check the rwmix cycle and switch if
739 static enum fio_ddir get_rw_ddir(struct thread_data *td)
744 * See if it's time to fsync/fdatasync/sync_file_range first,
745 * and if not then move on to check regular I/Os.
747 if (should_fsync(td)) {
748 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
749 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
752 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
753 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
754 return DDIR_DATASYNC;
756 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
757 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
758 return DDIR_SYNC_FILE_RANGE;
763 * Check if it's time to seed a new data direction.
765 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
767 * Put a top limit on how many bytes we do for
768 * one data direction, to avoid overflowing the
771 ddir = get_rand_ddir(td);
773 if (ddir != td->rwmix_ddir)
776 td->rwmix_ddir = ddir;
778 ddir = td->rwmix_ddir;
779 } else if (td_read(td))
781 else if (td_write(td))
783 else if (td_trim(td))
788 if (!should_check_rate(td)) {
790 * avoid time-consuming call to utime_since_now() if rate checking
791 * isn't being used. this imrpoves IOPs 50%. See:
792 * https://github.com/axboe/fio/issues/1501#issuecomment-1418327049
794 td->rwmix_ddir = ddir;
796 td->rwmix_ddir = rate_ddir(td, ddir);
797 return td->rwmix_ddir;
800 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
802 enum fio_ddir ddir = get_rw_ddir(td);
804 if (td->o.zone_mode == ZONE_MODE_ZBD)
805 ddir = zbd_adjust_ddir(td, io_u, ddir);
807 if (td_trimwrite(td)) {
808 struct fio_file *f = io_u->file;
809 if (f->last_start[DDIR_WRITE] == f->last_start[DDIR_TRIM])
815 io_u->ddir = io_u->acct_ddir = ddir;
817 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
818 td->o.barrier_blocks &&
819 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
820 td->io_issues[DDIR_WRITE])
821 io_u_set(td, io_u, IO_U_F_BARRIER);
824 void put_file_log(struct thread_data *td, struct fio_file *f)
826 unsigned int ret = put_file(td, f);
829 td_verror(td, ret, "file close");
832 void put_io_u(struct thread_data *td, struct io_u *io_u)
834 const bool needs_lock = td_async_processing(td);
836 zbd_put_io_u(td, 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);
858 __td_io_u_unlock(td);
861 void clear_io_u(struct thread_data *td, struct io_u *io_u)
863 io_u_clear(td, io_u, IO_U_F_FLIGHT);
867 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
869 const bool needs_lock = td_async_processing(td);
870 struct io_u *__io_u = *io_u;
871 enum fio_ddir ddir = acct_ddir(__io_u);
873 dprint(FD_IO, "requeue %p\n", __io_u);
881 io_u_set(td, __io_u, IO_U_F_FREE);
882 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
883 td->io_issues[ddir]--;
885 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
886 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
888 assert(!(td->flags & TD_F_CHILD));
891 io_u_rpush(&td->io_u_requeues, __io_u);
892 td_io_u_free_notify(td);
895 __td_io_u_unlock(td);
900 static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
902 struct fio_file *f = io_u->file;
904 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
905 assert(td->o.zone_size);
906 assert(td->o.zone_range);
909 * See if it's time to switch to a new zone
911 if (td->zone_bytes >= td->o.zone_size) {
913 f->file_offset += td->o.zone_range + td->o.zone_skip;
916 * Wrap from the beginning, if we exceed the file size
918 if (f->file_offset >= f->real_file_size)
919 f->file_offset = get_start_offset(td, f);
921 f->last_pos[io_u->ddir] = f->file_offset;
922 td->io_skip_bytes += td->o.zone_skip;
926 * If zone_size > zone_range, then maintain the same zone until
927 * zone_bytes >= zone_size.
929 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
930 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
931 f->file_offset, f->last_pos[io_u->ddir]);
932 f->last_pos[io_u->ddir] = f->file_offset;
936 * For random: if 'norandommap' is not set and zone_size > zone_range,
937 * map needs to be reset as it's done with zone_range everytime.
939 if ((td->zone_bytes % td->o.zone_range) == 0)
940 fio_file_reset(td, f);
943 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
947 enum io_u_action ret;
949 if (td_ioengine_flagged(td, FIO_NOIO))
952 set_rw_ddir(td, io_u);
954 if (io_u->ddir == DDIR_INVAL) {
955 dprint(FD_IO, "invalid direction received ddir = %d", io_u->ddir);
959 * fsync() or fdatasync() or trim etc, we are done
961 if (!ddir_rw(io_u->ddir))
964 if (td->o.zone_mode == ZONE_MODE_STRIDED)
965 setup_strided_zone_mode(td, io_u);
966 else if (td->o.zone_mode == ZONE_MODE_ZBD)
967 setup_zbd_zone_mode(td, io_u);
970 * No log, let the seq/rand engine retrieve the next buflen and
973 if (get_next_offset(td, io_u, &is_random)) {
974 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
978 io_u->buflen = get_next_buflen(td, io_u, is_random);
980 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
984 offset = io_u->offset;
985 if (td->o.zone_mode == ZONE_MODE_ZBD) {
986 ret = zbd_adjust_block(td, io_u);
991 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
992 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
994 (unsigned long long) io_u->offset, io_u->buflen,
995 (unsigned long long) io_u->file->real_file_size);
1000 * mark entry before potentially trimming io_u
1002 if (td_random(td) && file_randommap(td, io_u->file))
1003 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
1006 dprint_io_u(io_u, "fill");
1007 io_u->verify_offset = io_u->offset;
1008 td->zone_bytes += io_u->buflen;
1012 static void __io_u_mark_map(uint64_t *map, unsigned int nr)
1042 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1044 __io_u_mark_map(td->ts.io_u_submit, nr);
1045 td->ts.total_submit++;
1048 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1050 __io_u_mark_map(td->ts.io_u_complete, nr);
1051 td->ts.total_complete++;
1054 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1058 switch (td->cur_depth) {
1081 td->ts.io_u_map[idx] += nr;
1084 static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1088 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);
1164 assert(idx < FIO_IO_U_LAT_U_NR);
1165 td->ts.io_u_lat_u[idx]++;
1168 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1212 assert(idx < FIO_IO_U_LAT_M_NR);
1213 td->ts.io_u_lat_m[idx]++;
1216 static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1219 io_u_mark_lat_nsec(td, nsec);
1220 else if (nsec < 1000000)
1221 io_u_mark_lat_usec(td, nsec / 1000);
1223 io_u_mark_lat_msec(td, nsec / 1000000);
1226 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1228 unsigned long fileno;
1230 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1231 uint64_t frand_max = rand_max(&td->next_file_state);
1234 r = __rand(&td->next_file_state);
1235 return (unsigned int) ((double) td->o.nr_files
1236 * (r / (frand_max + 1.0)));
1239 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1240 fileno = zipf_next(&td->next_file_zipf);
1241 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1242 fileno = pareto_next(&td->next_file_zipf);
1243 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1244 fileno = gauss_next(&td->next_file_gauss);
1246 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1251 return fileno >> FIO_FSERVICE_SHIFT;
1255 * Get next file to service by choosing one at random
1257 static struct fio_file *get_next_file_rand(struct thread_data *td,
1258 enum fio_file_flags goodf,
1259 enum fio_file_flags badf)
1267 fno = __get_next_fileno_rand(td);
1270 if (fio_file_done(f))
1273 if (!fio_file_open(f)) {
1276 if (td->nr_open_files >= td->o.open_files)
1277 return ERR_PTR(-EBUSY);
1279 err = td_io_open_file(td, f);
1285 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1286 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1290 td_io_close_file(td, f);
1295 * Get next file to service by doing round robin between all available ones
1297 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1300 unsigned int old_next_file = td->next_file;
1306 f = td->files[td->next_file];
1309 if (td->next_file >= td->o.nr_files)
1312 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1313 if (fio_file_done(f)) {
1318 if (!fio_file_open(f)) {
1321 if (td->nr_open_files >= td->o.open_files)
1322 return ERR_PTR(-EBUSY);
1324 err = td_io_open_file(td, f);
1326 dprint(FD_FILE, "error %d on open of %s\n",
1334 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1336 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1340 td_io_close_file(td, f);
1343 } while (td->next_file != old_next_file);
1345 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1349 static struct fio_file *__get_next_file(struct thread_data *td)
1353 assert(td->o.nr_files <= td->files_index);
1355 if (td->nr_done_files >= td->o.nr_files) {
1356 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1357 " nr_files=%d\n", td->nr_open_files,
1363 f = td->file_service_file;
1364 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1365 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1367 if (td->file_service_left) {
1368 td->file_service_left--;
1373 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1374 td->o.file_service_type == FIO_FSERVICE_SEQ)
1375 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1377 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1382 td->file_service_file = f;
1383 td->file_service_left = td->file_service_nr - 1;
1386 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1388 dprint(FD_FILE, "get_next_file: NULL\n");
1392 static struct fio_file *get_next_file(struct thread_data *td)
1394 return __get_next_file(td);
1397 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1402 f = get_next_file(td);
1403 if (IS_ERR_OR_NULL(f))
1409 if (!fill_io_u(td, io_u))
1412 zbd_put_io_u(td, io_u);
1414 put_file_log(td, f);
1415 td_io_close_file(td, f);
1417 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1418 fio_file_reset(td, f);
1420 fio_file_set_done(f);
1421 td->nr_done_files++;
1422 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1423 td->nr_done_files, td->o.nr_files);
1430 static void lat_fatal(struct thread_data *td, struct io_u *io_u, struct io_completion_data *icd,
1431 unsigned long long tnsec, unsigned long long max_nsec)
1434 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec): %s %s %llu %llu\n",
1436 io_u->file->file_name,
1437 io_ddir_name(io_u->ddir),
1438 io_u->offset, io_u->buflen);
1440 td_verror(td, ETIMEDOUT, "max latency exceeded");
1441 icd->error = ETIMEDOUT;
1444 static void lat_new_cycle(struct thread_data *td)
1446 fio_gettime(&td->latency_ts, NULL);
1447 td->latency_ios = ddir_rw_sum(td->io_blocks);
1448 td->latency_failed = 0;
1452 * We had an IO outside the latency target. Reduce the queue depth. If we
1453 * are at QD=1, then it's time to give up.
1455 static bool __lat_target_failed(struct thread_data *td)
1457 if (td->latency_qd == 1)
1460 td->latency_qd_high = td->latency_qd;
1462 if (td->latency_qd == td->latency_qd_low)
1463 td->latency_qd_low--;
1465 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1466 td->latency_stable_count = 0;
1468 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1471 * When we ramp QD down, quiesce existing IO to prevent
1472 * a storm of ramp downs due to pending higher depth.
1479 static bool lat_target_failed(struct thread_data *td)
1481 if (td->o.latency_percentile.u.f == 100.0)
1482 return __lat_target_failed(td);
1484 td->latency_failed++;
1488 void lat_target_init(struct thread_data *td)
1490 td->latency_end_run = 0;
1492 if (td->o.latency_target) {
1493 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1494 fio_gettime(&td->latency_ts, NULL);
1496 td->latency_qd_high = td->o.iodepth;
1497 td->latency_qd_low = 1;
1498 td->latency_ios = ddir_rw_sum(td->io_blocks);
1500 td->latency_qd = td->o.iodepth;
1503 void lat_target_reset(struct thread_data *td)
1505 if (!td->latency_end_run)
1506 lat_target_init(td);
1509 static void lat_target_success(struct thread_data *td)
1511 const unsigned int qd = td->latency_qd;
1512 struct thread_options *o = &td->o;
1514 td->latency_qd_low = td->latency_qd;
1516 if (td->latency_qd + 1 == td->latency_qd_high) {
1518 * latency_qd will not incease on lat_target_success(), so
1519 * called stable. If we stick with this queue depth, the
1520 * final latency is likely lower than latency_target. Fix
1521 * this by increasing latency_qd_high slowly. Use a naive
1522 * heuristic here. If we get lat_target_success() 3 times
1523 * in a row, increase latency_qd_high by 1.
1525 if (++td->latency_stable_count >= 3) {
1526 td->latency_qd_high++;
1527 td->latency_stable_count = 0;
1532 * If we haven't failed yet, we double up to a failing value instead
1533 * of bisecting from highest possible queue depth. If we have set
1534 * a limit other than td->o.iodepth, bisect between that.
1536 if (td->latency_qd_high != o->iodepth)
1537 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1539 td->latency_qd *= 2;
1541 if (td->latency_qd > o->iodepth)
1542 td->latency_qd = o->iodepth;
1544 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1547 * Same as last one, we are done. Let it run a latency cycle, so
1548 * we get only the results from the targeted depth.
1550 if (!o->latency_run && td->latency_qd == qd) {
1551 if (td->latency_end_run) {
1552 dprint(FD_RATE, "We are done\n");
1555 dprint(FD_RATE, "Quiesce and final run\n");
1557 td->latency_end_run = 1;
1558 reset_all_stats(td);
1567 * Check if we can bump the queue depth
1569 void lat_target_check(struct thread_data *td)
1571 uint64_t usec_window;
1575 usec_window = utime_since_now(&td->latency_ts);
1576 if (usec_window < td->o.latency_window)
1579 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1580 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1581 success_ios *= 100.0;
1583 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1585 if (success_ios >= td->o.latency_percentile.u.f)
1586 lat_target_success(td);
1588 __lat_target_failed(td);
1592 * If latency target is enabled, we might be ramping up or down and not
1593 * using the full queue depth available.
1595 bool queue_full(const struct thread_data *td)
1597 const int qempty = io_u_qempty(&td->io_u_freelist);
1601 if (!td->o.latency_target)
1604 return td->cur_depth >= td->latency_qd;
1607 struct io_u *__get_io_u(struct thread_data *td)
1609 const bool needs_lock = td_async_processing(td);
1610 struct io_u *io_u = NULL;
1620 if (!io_u_rempty(&td->io_u_requeues)) {
1621 io_u = io_u_rpop(&td->io_u_requeues);
1623 } else if (!queue_full(td)) {
1624 io_u = io_u_qpop(&td->io_u_freelist);
1629 io_u->end_io = NULL;
1633 assert(io_u->flags & IO_U_F_FREE);
1634 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1635 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1639 io_u->acct_ddir = -1;
1641 assert(!(td->flags & TD_F_CHILD));
1642 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1644 } else if (td_async_processing(td)) {
1646 * We ran out, wait for async verify threads to finish and
1649 assert(!(td->flags & TD_F_CHILD));
1650 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1657 __td_io_u_unlock(td);
1662 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1664 if (!(td->flags & TD_F_TRIM_BACKLOG))
1666 if (!td->trim_entries)
1669 if (td->trim_batch) {
1671 if (get_next_trim(td, io_u))
1673 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1674 td->last_ddir != DDIR_READ) {
1675 td->trim_batch = td->o.trim_batch;
1676 if (!td->trim_batch)
1677 td->trim_batch = td->o.trim_backlog;
1678 if (get_next_trim(td, io_u))
1685 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1687 if (!(td->flags & TD_F_VER_BACKLOG))
1690 if (td->io_hist_len) {
1693 if (td->verify_batch)
1695 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1696 td->last_ddir != DDIR_READ) {
1697 td->verify_batch = td->o.verify_batch;
1698 if (!td->verify_batch)
1699 td->verify_batch = td->o.verify_backlog;
1703 if (get_verify && !get_next_verify(td, io_u)) {
1713 * Fill offset and start time into the buffer content, to prevent too
1714 * easy compressible data for simple de-dupe attempts. Do this for every
1715 * 512b block in the range, since that should be the smallest block size
1716 * we can expect from a device.
1718 static void small_content_scramble(struct io_u *io_u)
1720 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1721 unsigned int offset;
1722 uint64_t boffset, *iptr;
1729 boffset = io_u->offset;
1731 if (io_u->buf_filled_len)
1732 io_u->buf_filled_len = 0;
1735 * Generate random index between 0..7. We do chunks of 512b, if
1736 * we assume a cacheline is 64 bytes, then we have 8 of those.
1737 * Scramble content within the blocks in the same cacheline to
1740 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1742 for (i = 0; i < nr_blocks; i++) {
1744 * Fill offset into start of cacheline, time into end
1747 iptr = (void *) p + (offset << 6);
1750 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1751 iptr[0] = io_u->start_time.tv_sec;
1752 iptr[1] = io_u->start_time.tv_nsec;
1760 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1761 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1763 struct io_u *get_io_u(struct thread_data *td)
1767 int do_scramble = 0;
1770 io_u = __get_io_u(td);
1772 dprint(FD_IO, "__get_io_u failed\n");
1776 if (check_get_verify(td, io_u))
1778 if (check_get_trim(td, io_u))
1782 * from a requeue, io_u already setup
1788 * If using an iolog, grab next piece if any available.
1790 if (td->flags & TD_F_READ_IOLOG) {
1791 if (read_iolog_get(td, io_u))
1793 } else if (set_io_u_file(td, io_u)) {
1795 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1801 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1805 assert(fio_file_open(f));
1807 if (ddir_rw(io_u->ddir)) {
1808 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1809 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1813 f->last_start[io_u->ddir] = io_u->offset;
1814 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1816 if (io_u->ddir == DDIR_WRITE) {
1817 if (td->flags & TD_F_REFILL_BUFFERS) {
1818 io_u_fill_buffer(td, io_u,
1819 td->o.min_bs[DDIR_WRITE],
1821 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1822 !(td->flags & TD_F_COMPRESS) &&
1823 !(td->flags & TD_F_DO_VERIFY))
1825 } else if (io_u->ddir == DDIR_READ) {
1827 * Reset the buf_filled parameters so next time if the
1828 * buffer is used for writes it is refilled.
1830 io_u->buf_filled_len = 0;
1835 * Set io data pointers.
1837 io_u->xfer_buf = io_u->buf;
1838 io_u->xfer_buflen = io_u->buflen;
1841 * Remember the issuing context priority. The IO engine may change this.
1843 io_u->ioprio = td->ioprio;
1844 io_u->clat_prio_index = 0;
1847 if (!td_io_prep(td, io_u)) {
1848 if (!td->o.disable_lat)
1849 fio_gettime(&io_u->start_time, NULL);
1852 small_content_scramble(io_u);
1857 dprint(FD_IO, "get_io_u failed\n");
1859 return ERR_PTR(ret);
1862 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1864 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1866 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1869 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1870 io_u->file ? " on file " : "",
1871 io_u->file ? io_u->file->file_name : "",
1872 strerror(io_u->error),
1873 io_ddir_name(io_u->ddir),
1874 io_u->offset, io_u->xfer_buflen);
1876 if (td->io_ops->errdetails) {
1877 char *err = td->io_ops->errdetails(io_u);
1879 log_err("fio: %s\n", err);
1884 td_verror(td, io_u->error, "io_u error");
1887 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1889 __io_u_log_error(td, io_u);
1891 __io_u_log_error(td->parent, io_u);
1894 static inline bool gtod_reduce(struct thread_data *td)
1896 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1897 || td->o.gtod_reduce;
1900 static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1902 uint32_t *info = io_u_block_info(td, io_u);
1904 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1907 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1910 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1911 struct io_completion_data *icd,
1912 const enum fio_ddir idx, unsigned int bytes)
1914 const int no_reduce = !gtod_reduce(td);
1915 unsigned long long llnsec = 0;
1920 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1924 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1926 if (!td->o.disable_lat) {
1927 unsigned long long tnsec;
1929 tnsec = ntime_since(&io_u->start_time, &icd->time);
1930 add_lat_sample(td, idx, tnsec, bytes, io_u->offset,
1931 io_u->ioprio, io_u->clat_prio_index);
1933 if (td->flags & TD_F_PROFILE_OPS) {
1934 struct prof_io_ops *ops = &td->prof_io_ops;
1937 icd->error = ops->io_u_lat(td, tnsec);
1941 if (td->o.max_latency[idx] && tnsec > td->o.max_latency[idx])
1942 lat_fatal(td, io_u, icd, tnsec, td->o.max_latency[idx]);
1943 if (td->o.latency_target && tnsec > td->o.latency_target) {
1944 if (lat_target_failed(td))
1945 lat_fatal(td, io_u, icd, tnsec, td->o.latency_target);
1951 if (!td->o.disable_clat) {
1952 add_clat_sample(td, idx, llnsec, bytes, io_u->offset,
1953 io_u->ioprio, io_u->clat_prio_index);
1954 io_u_mark_latency(td, llnsec);
1957 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1958 add_bw_sample(td, io_u, bytes, llnsec);
1960 if (no_reduce && per_unit_log(td->iops_log))
1961 add_iops_sample(td, io_u, bytes);
1962 } else if (ddir_sync(idx) && !td->o.disable_clat)
1963 add_sync_clat_sample(&td->ts, llnsec);
1965 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1966 trim_block_info(td, io_u);
1969 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1970 uint64_t offset, unsigned int bytes)
1977 if (f->first_write == -1ULL || offset < f->first_write)
1978 f->first_write = offset;
1979 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1980 f->last_write = offset + bytes;
1982 if (!f->last_write_comp)
1985 idx = f->last_write_idx++;
1986 f->last_write_comp[idx] = offset;
1987 if (f->last_write_idx == td->o.iodepth)
1988 f->last_write_idx = 0;
1991 static bool should_account(struct thread_data *td)
1993 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1994 td->runstate == TD_VERIFYING);
1997 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1998 struct io_completion_data *icd)
2000 struct io_u *io_u = *io_u_ptr;
2001 enum fio_ddir ddir = io_u->ddir;
2002 struct fio_file *f = io_u->file;
2004 dprint_io_u(io_u, "complete");
2006 assert(io_u->flags & IO_U_F_FLIGHT);
2007 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
2010 * Mark IO ok to verify
2014 * Remove errored entry from the verification list
2017 unlog_io_piece(td, io_u);
2019 atomic_store_release(&io_u->ipo->flags,
2020 io_u->ipo->flags & ~IP_F_IN_FLIGHT);
2024 if (ddir_sync(ddir)) {
2025 td->last_was_sync = true;
2027 f->first_write = -1ULL;
2028 f->last_write = -1ULL;
2030 if (should_account(td))
2031 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
2035 td->last_was_sync = false;
2036 td->last_ddir = ddir;
2038 if (!io_u->error && ddir_rw(ddir)) {
2039 unsigned long long bytes = io_u->xfer_buflen - io_u->resid;
2043 * Make sure we notice short IO from here, and requeue them
2046 if (bytes && io_u->resid) {
2047 io_u->xfer_buflen = io_u->resid;
2048 io_u->xfer_buf += bytes;
2049 io_u->offset += bytes;
2050 td->ts.short_io_u[io_u->ddir]++;
2051 if (io_u->offset < io_u->file->real_file_size) {
2052 requeue_io_u(td, io_u_ptr);
2057 td->io_blocks[ddir]++;
2058 td->io_bytes[ddir] += bytes;
2060 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2061 td->this_io_blocks[ddir]++;
2062 td->this_io_bytes[ddir] += bytes;
2065 if (ddir == DDIR_WRITE)
2066 file_log_write_comp(td, f, io_u->offset, bytes);
2068 if (should_account(td))
2069 account_io_completion(td, io_u, icd, ddir, bytes);
2071 icd->bytes_done[ddir] += bytes;
2074 ret = io_u->end_io(td, io_u_ptr);
2076 if (ret && !icd->error)
2079 } else if (io_u->error) {
2080 icd->error = io_u->error;
2081 io_u_log_error(td, io_u);
2084 enum error_type_bit eb = td_error_type(ddir, icd->error);
2086 if (!td_non_fatal_error(td, eb, icd->error))
2090 * If there is a non_fatal error, then add to the error count
2091 * and clear all the errors.
2093 update_error_count(td, icd->error);
2101 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2106 if (!gtod_reduce(td))
2107 fio_gettime(&icd->time, NULL);
2112 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2113 icd->bytes_done[ddir] = 0;
2116 static void ios_completed(struct thread_data *td,
2117 struct io_completion_data *icd)
2122 for (i = 0; i < icd->nr; i++) {
2123 io_u = td->io_ops->event(td, i);
2125 io_completed(td, &io_u, icd);
2132 static void io_u_update_bytes_done(struct thread_data *td,
2133 struct io_completion_data *icd)
2137 if (td->runstate == TD_VERIFYING) {
2138 td->bytes_verified += icd->bytes_done[DDIR_READ];
2142 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2143 td->bytes_done[ddir] += icd->bytes_done[ddir];
2147 * Complete a single io_u for the sync engines.
2149 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2151 struct io_completion_data icd;
2153 init_icd(td, &icd, 1);
2154 io_completed(td, &io_u, &icd);
2160 td_verror(td, icd.error, "io_u_sync_complete");
2164 io_u_update_bytes_done(td, &icd);
2170 * Called to complete min_events number of io for the async engines.
2172 int io_u_queued_complete(struct thread_data *td, int min_evts)
2174 struct io_completion_data icd;
2175 struct timespec *tvp = NULL;
2177 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2179 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2183 else if (min_evts > td->cur_depth)
2184 min_evts = td->cur_depth;
2186 /* No worries, td_io_getevents fixes min and max if they are
2187 * set incorrectly */
2188 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2190 td_verror(td, -ret, "td_io_getevents");
2195 init_icd(td, &icd, ret);
2196 ios_completed(td, &icd);
2198 td_verror(td, icd.error, "io_u_queued_complete");
2202 io_u_update_bytes_done(td, &icd);
2208 * Call when io_u is really queued, to update the submission latency.
2210 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2212 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2213 unsigned long slat_time;
2215 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2220 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2221 io_u->offset, io_u->ioprio);
2226 * See if we should reuse the last seed, if dedupe is enabled
2228 static struct frand_state *get_buf_state(struct thread_data *td)
2231 unsigned long long i;
2233 if (!td->o.dedupe_percentage)
2234 return &td->buf_state;
2235 else if (td->o.dedupe_percentage == 100) {
2236 frand_copy(&td->buf_state_prev, &td->buf_state);
2237 return &td->buf_state;
2240 v = rand_between(&td->dedupe_state, 1, 100);
2242 if (v <= td->o.dedupe_percentage)
2243 switch (td->o.dedupe_mode) {
2244 case DEDUPE_MODE_REPEAT:
2246 * The caller advances the returned frand_state.
2247 * A copy of prev should be returned instead since
2248 * a subsequent intention to generate a deduped buffer
2249 * might result in generating a unique one
2251 frand_copy(&td->buf_state_ret, &td->buf_state_prev);
2252 return &td->buf_state_ret;
2253 case DEDUPE_MODE_WORKING_SET:
2254 i = rand_between(&td->dedupe_working_set_index_state, 0, td->num_unique_pages - 1);
2255 frand_copy(&td->buf_state_ret, &td->dedupe_working_set_states[i]);
2256 return &td->buf_state_ret;
2258 log_err("unexpected dedupe mode %u\n", td->o.dedupe_mode);
2262 return &td->buf_state;
2265 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2267 if (td->o.dedupe_percentage == 100)
2268 frand_copy(rs, &td->buf_state_prev);
2269 else if (rs == &td->buf_state)
2270 frand_copy(&td->buf_state_prev, rs);
2273 void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2274 unsigned long long max_bs)
2276 struct thread_options *o = &td->o;
2278 if (o->mem_type == MEM_CUDA_MALLOC)
2281 if (o->compress_percentage || o->dedupe_percentage) {
2282 unsigned int perc = td->o.compress_percentage;
2283 struct frand_state *rs = NULL;
2284 unsigned long long left = max_bs;
2285 unsigned long long this_write;
2289 * Buffers are either entirely dedupe-able or not.
2290 * If we choose to dedup, the buffer should undergo
2291 * the same manipulation as the original write. Which
2292 * means we should retrack the steps we took for compression
2296 rs = get_buf_state(td);
2298 min_write = min(min_write, left);
2300 this_write = min_not_zero(min_write,
2301 (unsigned long long) td->o.compress_chunk);
2303 fill_random_buf_percentage(rs, buf, perc,
2304 this_write, this_write,
2306 o->buffer_pattern_bytes);
2310 save_buf_state(td, rs);
2312 } else if (o->buffer_pattern_bytes)
2313 fill_buffer_pattern(td, buf, max_bs);
2314 else if (o->zero_buffers)
2315 memset(buf, 0, max_bs);
2317 fill_random_buf(get_buf_state(td), buf, max_bs);
2321 * "randomly" fill the buffer contents
2323 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2324 unsigned long long min_write, unsigned long long max_bs)
2326 io_u->buf_filled_len = 0;
2327 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2330 static int do_sync_file_range(const struct thread_data *td,
2333 uint64_t offset, nbytes;
2335 offset = f->first_write;
2336 nbytes = f->last_write - f->first_write;
2341 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2344 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2348 if (io_u->ddir == DDIR_SYNC) {
2349 #ifdef CONFIG_FCNTL_SYNC
2350 ret = fcntl(io_u->file->fd, F_FULLFSYNC);
2352 ret = fsync(io_u->file->fd);
2354 } else if (io_u->ddir == DDIR_DATASYNC) {
2355 #ifdef CONFIG_FDATASYNC
2356 ret = fdatasync(io_u->file->fd);
2358 ret = io_u->xfer_buflen;
2359 io_u->error = EINVAL;
2361 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2362 ret = do_sync_file_range(td, io_u->file);
2364 ret = io_u->xfer_buflen;
2365 io_u->error = EINVAL;
2369 io_u->error = errno;
2374 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2376 #ifndef FIO_HAVE_TRIM
2377 io_u->error = EINVAL;
2380 struct fio_file *f = io_u->file;
2383 if (td->o.zone_mode == ZONE_MODE_ZBD) {
2384 ret = zbd_do_io_u_trim(td, io_u);
2385 if (ret == io_u_completed)
2386 return io_u->xfer_buflen;
2391 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2393 return io_u->xfer_buflen;