13 #include "lib/axmap.h"
18 struct io_completion_data {
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timeval time; /* output */
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
30 static bool random_map_free(struct fio_file *f, const uint64_t block)
32 return !axmap_isset(f->io_axmap, block);
36 * Mark a given offset as used in the map.
38 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40 unsigned int min_bs = td->o.rw_min_bs;
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
55 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
61 assert(ddir_rw(ddir));
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 * -> not for now since there is code assuming it could go either.
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
72 max_size = td->o.zone_range;
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
85 struct flist_head list;
89 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
90 enum fio_ddir ddir, uint64_t *b,
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
160 static int __get_next_rand_offset_zoned(struct thread_data *td,
161 struct fio_file *f, enum fio_ddir ddir,
164 unsigned int v, send, stotal;
165 uint64_t offset, lastb;
167 struct zone_split_index *zsi;
169 lastb = last_block(td, f, ddir);
173 if (!td->o.zone_split_nr[ddir]) {
175 return __get_next_rand_offset(td, f, ddir, b, lastb);
179 * Generate a value, v, between 1 and 100, both inclusive
181 v = rand32_between(&td->zone_state, 1, 100);
183 zsi = &td->zone_state_index[ddir][v - 1];
184 stotal = zsi->size_perc_prev;
185 send = zsi->size_perc;
188 * Should never happen
192 log_err("fio: bug in zoned generation\n");
199 * 'send' is some percentage below or equal to 100 that
200 * marks the end of the current IO range. 'stotal' marks
201 * the start, in percent.
204 offset = stotal * lastb / 100ULL;
208 lastb = lastb * (send - stotal) / 100ULL;
211 * Generate index from 0..send-of-lastb
213 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
217 * Add our start offset, if any
225 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
227 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
228 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
230 return r1->off - r2->off;
233 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
234 enum fio_ddir ddir, uint64_t *b)
236 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
239 lastb = last_block(td, f, ddir);
243 return __get_next_rand_offset(td, f, ddir, b, lastb);
244 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
245 return __get_next_rand_offset_zipf(td, f, ddir, b);
246 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
247 return __get_next_rand_offset_pareto(td, f, ddir, b);
248 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
249 return __get_next_rand_offset_gauss(td, f, ddir, b);
250 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
251 return __get_next_rand_offset_zoned(td, f, ddir, b);
253 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
258 * Sort the reads for a verify phase in batches of verifysort_nr, if
261 static inline bool should_sort_io(struct thread_data *td)
263 if (!td->o.verifysort_nr || !td->o.do_verify)
267 if (td->runstate != TD_VERIFYING)
269 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
270 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
276 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
280 if (td->o.perc_rand[ddir] == 100)
283 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
285 return v <= td->o.perc_rand[ddir];
288 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
289 enum fio_ddir ddir, uint64_t *b)
294 if (!should_sort_io(td))
295 return get_off_from_method(td, f, ddir, b);
297 if (!flist_empty(&td->next_rand_list)) {
299 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
306 for (i = 0; i < td->o.verifysort_nr; i++) {
307 r = malloc(sizeof(*r));
309 ret = get_off_from_method(td, f, ddir, &r->off);
315 flist_add(&r->list, &td->next_rand_list);
321 assert(!flist_empty(&td->next_rand_list));
322 flist_sort(NULL, &td->next_rand_list, flist_cmp);
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 if (!get_next_rand_offset(td, f, ddir, b))
339 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
340 f->file_name, (unsigned long long) f->last_pos[ddir],
341 (unsigned long long) f->real_file_size);
345 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
346 enum fio_ddir ddir, uint64_t *offset)
348 struct thread_options *o = &td->o;
350 assert(ddir_rw(ddir));
352 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
354 struct thread_options *o = &td->o;
355 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
357 if (io_size > f->last_pos[ddir])
358 f->last_pos[ddir] = 0;
360 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
363 if (f->last_pos[ddir] < f->real_file_size) {
366 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
367 if (f->real_file_size > f->io_size)
368 f->last_pos[ddir] = f->io_size;
370 f->last_pos[ddir] = f->real_file_size;
373 pos = f->last_pos[ddir] - f->file_offset;
374 if (pos && o->ddir_seq_add) {
375 pos += o->ddir_seq_add;
378 * If we reach beyond the end of the file
379 * with holed IO, wrap around to the
380 * beginning again. If we're doing backwards IO,
383 if (pos >= f->real_file_size) {
384 if (o->ddir_seq_add > 0)
385 pos = f->file_offset;
387 if (f->real_file_size > f->io_size)
390 pos = f->real_file_size;
392 pos += o->ddir_seq_add;
404 static int get_next_block(struct thread_data *td, struct io_u *io_u,
405 enum fio_ddir ddir, int rw_seq,
406 unsigned int *is_random)
408 struct fio_file *f = io_u->file;
412 assert(ddir_rw(ddir));
418 if (should_do_random(td, ddir)) {
419 ret = get_next_rand_block(td, f, ddir, &b);
423 io_u_set(td, io_u, IO_U_F_BUSY_OK);
424 ret = get_next_seq_offset(td, f, ddir, &offset);
426 ret = get_next_rand_block(td, f, ddir, &b);
430 ret = get_next_seq_offset(td, f, ddir, &offset);
433 io_u_set(td, io_u, IO_U_F_BUSY_OK);
436 if (td->o.rw_seq == RW_SEQ_SEQ) {
437 ret = get_next_seq_offset(td, f, ddir, &offset);
439 ret = get_next_rand_block(td, f, ddir, &b);
442 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
443 if (f->last_start[ddir] != -1ULL)
444 offset = f->last_start[ddir] - f->file_offset;
449 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
456 io_u->offset = offset;
458 io_u->offset = b * td->o.ba[ddir];
460 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
469 * For random io, generate a random new block and see if it's used. Repeat
470 * until we find a free one. For sequential io, just return the end of
471 * the last io issued.
473 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
474 unsigned int *is_random)
476 struct fio_file *f = io_u->file;
477 enum fio_ddir ddir = io_u->ddir;
480 assert(ddir_rw(ddir));
482 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
484 td->ddir_seq_nr = td->o.ddir_seq_nr;
487 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
490 if (io_u->offset >= f->io_size) {
491 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
492 (unsigned long long) io_u->offset,
493 (unsigned long long) f->io_size);
497 io_u->offset += f->file_offset;
498 if (io_u->offset >= f->real_file_size) {
499 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
500 (unsigned long long) io_u->offset,
501 (unsigned long long) f->real_file_size);
508 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
509 unsigned int *is_random)
511 if (td->flags & TD_F_PROFILE_OPS) {
512 struct prof_io_ops *ops = &td->prof_io_ops;
514 if (ops->fill_io_u_off)
515 return ops->fill_io_u_off(td, io_u, is_random);
518 return __get_next_offset(td, io_u, is_random);
521 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
524 struct fio_file *f = io_u->file;
526 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
529 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
530 unsigned int is_random)
532 int ddir = io_u->ddir;
533 unsigned int buflen = 0;
534 unsigned int minbs, maxbs;
535 uint64_t frand_max, r;
536 bool power_2 = false;
538 assert(ddir_rw(ddir));
540 if (td->o.bs_is_seq_rand)
541 ddir = is_random ? DDIR_WRITE: DDIR_READ;
543 minbs = td->o.min_bs[ddir];
544 maxbs = td->o.max_bs[ddir];
550 * If we can't satisfy the min block size from here, then fail
552 if (!io_u_fits(td, io_u, minbs))
555 frand_max = rand_max(&td->bsrange_state);
557 r = __rand(&td->bsrange_state);
559 if (!td->o.bssplit_nr[ddir]) {
560 buflen = 1 + (unsigned int) ((double) maxbs *
561 (r / (frand_max + 1.0)));
568 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
569 struct bssplit *bsp = &td->o.bssplit[ddir][i];
575 if ((r / perc <= frand_max / 100ULL) &&
576 io_u_fits(td, io_u, buflen))
581 power_2 = is_power_of_2(minbs);
582 if (!td->o.bs_unaligned && power_2)
583 buflen &= ~(minbs - 1);
584 else if (!td->o.bs_unaligned && !power_2)
585 buflen -= buflen % minbs;
586 } while (!io_u_fits(td, io_u, buflen));
591 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
592 unsigned int is_random)
594 if (td->flags & TD_F_PROFILE_OPS) {
595 struct prof_io_ops *ops = &td->prof_io_ops;
597 if (ops->fill_io_u_size)
598 return ops->fill_io_u_size(td, io_u, is_random);
601 return __get_next_buflen(td, io_u, is_random);
604 static void set_rwmix_bytes(struct thread_data *td)
609 * we do time or byte based switch. this is needed because
610 * buffered writes may issue a lot quicker than they complete,
611 * whereas reads do not.
613 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
614 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
617 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
621 v = rand32_between(&td->rwmix_state, 1, 100);
623 if (v <= td->o.rwmix[DDIR_READ])
629 int io_u_quiesce(struct thread_data *td)
634 * We are going to sleep, ensure that we flush anything pending as
635 * not to skew our latency numbers.
637 * Changed to only monitor 'in flight' requests here instead of the
638 * td->cur_depth, b/c td->cur_depth does not accurately represent
639 * io's that have been actually submitted to an async engine,
640 * and cur_depth is meaningless for sync engines.
642 if (td->io_u_queued || td->cur_depth) {
645 ret = td_io_commit(td);
648 while (td->io_u_in_flight) {
651 ret = io_u_queued_complete(td, 1);
656 if (td->flags & TD_F_REGROW_LOGS)
662 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
664 enum fio_ddir odir = ddir ^ 1;
668 assert(ddir_rw(ddir));
669 now = utime_since_now(&td->start);
672 * if rate_next_io_time is in the past, need to catch up to rate
674 if (td->rate_next_io_time[ddir] <= now)
678 * We are ahead of rate in this direction. See if we
681 if (td_rw(td) && td->o.rwmix[odir]) {
683 * Other direction is behind rate, switch
685 if (td->rate_next_io_time[odir] <= now)
689 * Both directions are ahead of rate. sleep the min
690 * switch if necissary
692 if (td->rate_next_io_time[ddir] <=
693 td->rate_next_io_time[odir]) {
694 usec = td->rate_next_io_time[ddir] - now;
696 usec = td->rate_next_io_time[odir] - now;
700 usec = td->rate_next_io_time[ddir] - now;
702 if (td->o.io_submit_mode == IO_MODE_INLINE)
705 usec = usec_sleep(td, usec);
711 * Return the data direction for the next io_u. If the job is a
712 * mixed read/write workload, check the rwmix cycle and switch if
715 static enum fio_ddir get_rw_ddir(struct thread_data *td)
720 * see if it's time to fsync
722 if (td->o.fsync_blocks &&
723 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
724 td->io_issues[DDIR_WRITE] && should_fsync(td))
728 * see if it's time to fdatasync
730 if (td->o.fdatasync_blocks &&
731 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
732 td->io_issues[DDIR_WRITE] && should_fsync(td))
733 return DDIR_DATASYNC;
736 * see if it's time to sync_file_range
738 if (td->sync_file_range_nr &&
739 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
740 td->io_issues[DDIR_WRITE] && should_fsync(td))
741 return DDIR_SYNC_FILE_RANGE;
745 * Check if it's time to seed a new data direction.
747 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
749 * Put a top limit on how many bytes we do for
750 * one data direction, to avoid overflowing the
753 ddir = get_rand_ddir(td);
755 if (ddir != td->rwmix_ddir)
758 td->rwmix_ddir = ddir;
760 ddir = td->rwmix_ddir;
761 } else if (td_read(td))
763 else if (td_write(td))
765 else if (td_trim(td))
770 td->rwmix_ddir = rate_ddir(td, ddir);
771 return td->rwmix_ddir;
774 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
776 enum fio_ddir ddir = get_rw_ddir(td);
778 if (td_trimwrite(td)) {
779 struct fio_file *f = io_u->file;
780 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
786 io_u->ddir = io_u->acct_ddir = ddir;
788 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
789 td->o.barrier_blocks &&
790 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
791 td->io_issues[DDIR_WRITE])
792 io_u_set(td, io_u, IO_U_F_BARRIER);
795 void put_file_log(struct thread_data *td, struct fio_file *f)
797 unsigned int ret = put_file(td, f);
800 td_verror(td, ret, "file close");
803 void put_io_u(struct thread_data *td, struct io_u *io_u)
810 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
811 put_file_log(td, io_u->file);
814 io_u_set(td, io_u, IO_U_F_FREE);
816 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
818 assert(!(td->flags & TD_F_CHILD));
820 io_u_qpush(&td->io_u_freelist, io_u);
822 td_io_u_free_notify(td);
825 void clear_io_u(struct thread_data *td, struct io_u *io_u)
827 io_u_clear(td, io_u, IO_U_F_FLIGHT);
831 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
833 struct io_u *__io_u = *io_u;
834 enum fio_ddir ddir = acct_ddir(__io_u);
836 dprint(FD_IO, "requeue %p\n", __io_u);
843 io_u_set(td, __io_u, IO_U_F_FREE);
844 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
845 td->io_issues[ddir]--;
847 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
848 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
850 assert(!(td->flags & TD_F_CHILD));
853 io_u_rpush(&td->io_u_requeues, __io_u);
855 td_io_u_free_notify(td);
859 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
861 unsigned int is_random;
863 if (td_ioengine_flagged(td, FIO_NOIO))
866 set_rw_ddir(td, io_u);
869 * fsync() or fdatasync() or trim etc, we are done
871 if (!ddir_rw(io_u->ddir))
875 * See if it's time to switch to a new zone
877 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
878 struct fio_file *f = io_u->file;
881 f->file_offset += td->o.zone_range + td->o.zone_skip;
884 * Wrap from the beginning, if we exceed the file size
886 if (f->file_offset >= f->real_file_size)
887 f->file_offset = f->real_file_size - f->file_offset;
888 f->last_pos[io_u->ddir] = f->file_offset;
889 td->io_skip_bytes += td->o.zone_skip;
893 * No log, let the seq/rand engine retrieve the next buflen and
896 if (get_next_offset(td, io_u, &is_random)) {
897 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
901 io_u->buflen = get_next_buflen(td, io_u, is_random);
903 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
907 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
908 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
909 dprint(FD_IO, " off=%llu/%lu > %llu\n",
910 (unsigned long long) io_u->offset, io_u->buflen,
911 (unsigned long long) io_u->file->real_file_size);
916 * mark entry before potentially trimming io_u
918 if (td_random(td) && file_randommap(td, io_u->file))
919 mark_random_map(td, io_u);
922 dprint_io_u(io_u, "fill_io_u");
923 td->zone_bytes += io_u->buflen;
927 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
956 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
958 __io_u_mark_map(td->ts.io_u_submit, nr);
959 td->ts.total_submit++;
962 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
964 __io_u_mark_map(td->ts.io_u_complete, nr);
965 td->ts.total_complete++;
968 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
972 switch (td->cur_depth) {
994 td->ts.io_u_map[idx] += nr;
997 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1001 assert(usec < 1000);
1034 assert(idx < FIO_IO_U_LAT_U_NR);
1035 td->ts.io_u_lat_u[idx]++;
1038 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1079 assert(idx < FIO_IO_U_LAT_M_NR);
1080 td->ts.io_u_lat_m[idx]++;
1083 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1086 io_u_mark_lat_usec(td, usec);
1088 io_u_mark_lat_msec(td, usec / 1000);
1091 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1093 unsigned long fileno;
1095 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1096 uint64_t frand_max = rand_max(&td->next_file_state);
1099 r = __rand(&td->next_file_state);
1100 return (unsigned int) ((double) td->o.nr_files
1101 * (r / (frand_max + 1.0)));
1104 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1105 fileno = zipf_next(&td->next_file_zipf);
1106 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1107 fileno = pareto_next(&td->next_file_zipf);
1108 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1109 fileno = gauss_next(&td->next_file_gauss);
1111 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1116 return fileno >> FIO_FSERVICE_SHIFT;
1120 * Get next file to service by choosing one at random
1122 static struct fio_file *get_next_file_rand(struct thread_data *td,
1123 enum fio_file_flags goodf,
1124 enum fio_file_flags badf)
1132 fno = __get_next_fileno_rand(td);
1135 if (fio_file_done(f))
1138 if (!fio_file_open(f)) {
1141 if (td->nr_open_files >= td->o.open_files)
1142 return ERR_PTR(-EBUSY);
1144 err = td_io_open_file(td, f);
1150 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1151 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1155 td_io_close_file(td, f);
1160 * Get next file to service by doing round robin between all available ones
1162 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1165 unsigned int old_next_file = td->next_file;
1171 f = td->files[td->next_file];
1174 if (td->next_file >= td->o.nr_files)
1177 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1178 if (fio_file_done(f)) {
1183 if (!fio_file_open(f)) {
1186 if (td->nr_open_files >= td->o.open_files)
1187 return ERR_PTR(-EBUSY);
1189 err = td_io_open_file(td, f);
1191 dprint(FD_FILE, "error %d on open of %s\n",
1199 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1201 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1205 td_io_close_file(td, f);
1208 } while (td->next_file != old_next_file);
1210 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1214 static struct fio_file *__get_next_file(struct thread_data *td)
1218 assert(td->o.nr_files <= td->files_index);
1220 if (td->nr_done_files >= td->o.nr_files) {
1221 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1222 " nr_files=%d\n", td->nr_open_files,
1228 f = td->file_service_file;
1229 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1230 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1232 if (td->file_service_left--)
1236 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1237 td->o.file_service_type == FIO_FSERVICE_SEQ)
1238 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1240 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1245 td->file_service_file = f;
1246 td->file_service_left = td->file_service_nr - 1;
1249 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1251 dprint(FD_FILE, "get_next_file: NULL\n");
1255 static struct fio_file *get_next_file(struct thread_data *td)
1257 if (td->flags & TD_F_PROFILE_OPS) {
1258 struct prof_io_ops *ops = &td->prof_io_ops;
1260 if (ops->get_next_file)
1261 return ops->get_next_file(td);
1264 return __get_next_file(td);
1267 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1272 f = get_next_file(td);
1273 if (IS_ERR_OR_NULL(f))
1279 if (!fill_io_u(td, io_u))
1282 put_file_log(td, f);
1283 td_io_close_file(td, f);
1285 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1286 fio_file_reset(td, f);
1288 fio_file_set_done(f);
1289 td->nr_done_files++;
1290 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1291 td->nr_done_files, td->o.nr_files);
1298 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1299 unsigned long tusec, unsigned long max_usec)
1302 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1303 td_verror(td, ETIMEDOUT, "max latency exceeded");
1304 icd->error = ETIMEDOUT;
1307 static void lat_new_cycle(struct thread_data *td)
1309 fio_gettime(&td->latency_ts, NULL);
1310 td->latency_ios = ddir_rw_sum(td->io_blocks);
1311 td->latency_failed = 0;
1315 * We had an IO outside the latency target. Reduce the queue depth. If we
1316 * are at QD=1, then it's time to give up.
1318 static bool __lat_target_failed(struct thread_data *td)
1320 if (td->latency_qd == 1)
1323 td->latency_qd_high = td->latency_qd;
1325 if (td->latency_qd == td->latency_qd_low)
1326 td->latency_qd_low--;
1328 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1330 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1333 * When we ramp QD down, quiesce existing IO to prevent
1334 * a storm of ramp downs due to pending higher depth.
1341 static bool lat_target_failed(struct thread_data *td)
1343 if (td->o.latency_percentile.u.f == 100.0)
1344 return __lat_target_failed(td);
1346 td->latency_failed++;
1350 void lat_target_init(struct thread_data *td)
1352 td->latency_end_run = 0;
1354 if (td->o.latency_target) {
1355 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1356 fio_gettime(&td->latency_ts, NULL);
1358 td->latency_qd_high = td->o.iodepth;
1359 td->latency_qd_low = 1;
1360 td->latency_ios = ddir_rw_sum(td->io_blocks);
1362 td->latency_qd = td->o.iodepth;
1365 void lat_target_reset(struct thread_data *td)
1367 if (!td->latency_end_run)
1368 lat_target_init(td);
1371 static void lat_target_success(struct thread_data *td)
1373 const unsigned int qd = td->latency_qd;
1374 struct thread_options *o = &td->o;
1376 td->latency_qd_low = td->latency_qd;
1379 * If we haven't failed yet, we double up to a failing value instead
1380 * of bisecting from highest possible queue depth. If we have set
1381 * a limit other than td->o.iodepth, bisect between that.
1383 if (td->latency_qd_high != o->iodepth)
1384 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1386 td->latency_qd *= 2;
1388 if (td->latency_qd > o->iodepth)
1389 td->latency_qd = o->iodepth;
1391 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1394 * Same as last one, we are done. Let it run a latency cycle, so
1395 * we get only the results from the targeted depth.
1397 if (td->latency_qd == qd) {
1398 if (td->latency_end_run) {
1399 dprint(FD_RATE, "We are done\n");
1402 dprint(FD_RATE, "Quiesce and final run\n");
1404 td->latency_end_run = 1;
1405 reset_all_stats(td);
1414 * Check if we can bump the queue depth
1416 void lat_target_check(struct thread_data *td)
1418 uint64_t usec_window;
1422 usec_window = utime_since_now(&td->latency_ts);
1423 if (usec_window < td->o.latency_window)
1426 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1427 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1428 success_ios *= 100.0;
1430 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1432 if (success_ios >= td->o.latency_percentile.u.f)
1433 lat_target_success(td);
1435 __lat_target_failed(td);
1439 * If latency target is enabled, we might be ramping up or down and not
1440 * using the full queue depth available.
1442 bool queue_full(const struct thread_data *td)
1444 const int qempty = io_u_qempty(&td->io_u_freelist);
1448 if (!td->o.latency_target)
1451 return td->cur_depth >= td->latency_qd;
1454 struct io_u *__get_io_u(struct thread_data *td)
1456 struct io_u *io_u = NULL;
1464 if (!io_u_rempty(&td->io_u_requeues))
1465 io_u = io_u_rpop(&td->io_u_requeues);
1466 else if (!queue_full(td)) {
1467 io_u = io_u_qpop(&td->io_u_freelist);
1472 io_u->end_io = NULL;
1476 assert(io_u->flags & IO_U_F_FREE);
1477 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1478 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1482 io_u->acct_ddir = -1;
1484 assert(!(td->flags & TD_F_CHILD));
1485 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1487 } else if (td_async_processing(td)) {
1489 * We ran out, wait for async verify threads to finish and
1492 assert(!(td->flags & TD_F_CHILD));
1493 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1501 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1503 if (!(td->flags & TD_F_TRIM_BACKLOG))
1506 if (td->trim_entries) {
1509 if (td->trim_batch) {
1512 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1513 td->last_ddir != DDIR_READ) {
1514 td->trim_batch = td->o.trim_batch;
1515 if (!td->trim_batch)
1516 td->trim_batch = td->o.trim_backlog;
1520 if (get_trim && get_next_trim(td, io_u))
1527 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1529 if (!(td->flags & TD_F_VER_BACKLOG))
1532 if (td->io_hist_len) {
1535 if (td->verify_batch)
1537 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1538 td->last_ddir != DDIR_READ) {
1539 td->verify_batch = td->o.verify_batch;
1540 if (!td->verify_batch)
1541 td->verify_batch = td->o.verify_backlog;
1545 if (get_verify && !get_next_verify(td, io_u)) {
1555 * Fill offset and start time into the buffer content, to prevent too
1556 * easy compressible data for simple de-dupe attempts. Do this for every
1557 * 512b block in the range, since that should be the smallest block size
1558 * we can expect from a device.
1560 static void small_content_scramble(struct io_u *io_u)
1562 unsigned int i, nr_blocks = io_u->buflen / 512;
1564 unsigned int offset;
1571 boffset = io_u->offset;
1572 io_u->buf_filled_len = 0;
1574 for (i = 0; i < nr_blocks; i++) {
1576 * Fill the byte offset into a "random" start offset of
1577 * the buffer, given by the product of the usec time
1578 * and the actual offset.
1580 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1581 offset &= ~(sizeof(uint64_t) - 1);
1582 if (offset >= 512 - sizeof(uint64_t))
1583 offset -= sizeof(uint64_t);
1584 memcpy(p + offset, &boffset, sizeof(boffset));
1586 end = p + 512 - sizeof(io_u->start_time);
1587 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1594 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1595 * etc. The returned io_u is fully ready to be prepped and submitted.
1597 struct io_u *get_io_u(struct thread_data *td)
1601 int do_scramble = 0;
1604 io_u = __get_io_u(td);
1606 dprint(FD_IO, "__get_io_u failed\n");
1610 if (check_get_verify(td, io_u))
1612 if (check_get_trim(td, io_u))
1616 * from a requeue, io_u already setup
1622 * If using an iolog, grab next piece if any available.
1624 if (td->flags & TD_F_READ_IOLOG) {
1625 if (read_iolog_get(td, io_u))
1627 } else if (set_io_u_file(td, io_u)) {
1629 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1635 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1639 assert(fio_file_open(f));
1641 if (ddir_rw(io_u->ddir)) {
1642 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1643 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1647 f->last_start[io_u->ddir] = io_u->offset;
1648 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1650 if (io_u->ddir == DDIR_WRITE) {
1651 if (td->flags & TD_F_REFILL_BUFFERS) {
1652 io_u_fill_buffer(td, io_u,
1653 td->o.min_bs[DDIR_WRITE],
1655 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1656 !(td->flags & TD_F_COMPRESS))
1658 if (td->flags & TD_F_VER_NONE) {
1659 populate_verify_io_u(td, io_u);
1662 } else if (io_u->ddir == DDIR_READ) {
1664 * Reset the buf_filled parameters so next time if the
1665 * buffer is used for writes it is refilled.
1667 io_u->buf_filled_len = 0;
1672 * Set io data pointers.
1674 io_u->xfer_buf = io_u->buf;
1675 io_u->xfer_buflen = io_u->buflen;
1679 if (!td_io_prep(td, io_u)) {
1680 if (!td->o.disable_lat)
1681 fio_gettime(&io_u->start_time, NULL);
1683 small_content_scramble(io_u);
1687 dprint(FD_IO, "get_io_u failed\n");
1689 return ERR_PTR(ret);
1692 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1694 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1696 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1699 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1700 io_u->file ? " on file " : "",
1701 io_u->file ? io_u->file->file_name : "",
1702 strerror(io_u->error),
1703 io_ddir_name(io_u->ddir),
1704 io_u->offset, io_u->xfer_buflen);
1706 if (td->io_ops->errdetails) {
1707 char *err = td->io_ops->errdetails(io_u);
1709 log_err("fio: %s\n", err);
1714 td_verror(td, io_u->error, "io_u error");
1717 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1719 __io_u_log_error(td, io_u);
1721 __io_u_log_error(td->parent, io_u);
1724 static inline bool gtod_reduce(struct thread_data *td)
1726 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1727 || td->o.gtod_reduce;
1730 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1731 struct io_completion_data *icd,
1732 const enum fio_ddir idx, unsigned int bytes)
1734 const int no_reduce = !gtod_reduce(td);
1735 unsigned long lusec = 0;
1744 lusec = utime_since(&io_u->issue_time, &icd->time);
1746 if (!td->o.disable_lat) {
1747 unsigned long tusec;
1749 tusec = utime_since(&io_u->start_time, &icd->time);
1750 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1752 if (td->flags & TD_F_PROFILE_OPS) {
1753 struct prof_io_ops *ops = &td->prof_io_ops;
1756 icd->error = ops->io_u_lat(td, tusec);
1759 if (td->o.max_latency && tusec > td->o.max_latency)
1760 lat_fatal(td, icd, tusec, td->o.max_latency);
1761 if (td->o.latency_target && tusec > td->o.latency_target) {
1762 if (lat_target_failed(td))
1763 lat_fatal(td, icd, tusec, td->o.latency_target);
1768 if (!td->o.disable_clat) {
1769 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1770 io_u_mark_latency(td, lusec);
1773 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1774 add_bw_sample(td, io_u, bytes, lusec);
1776 if (no_reduce && per_unit_log(td->iops_log))
1777 add_iops_sample(td, io_u, bytes);
1780 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1781 uint32_t *info = io_u_block_info(td, io_u);
1782 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1783 if (io_u->ddir == DDIR_TRIM) {
1784 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1785 BLOCK_INFO_TRIMS(*info) + 1);
1786 } else if (io_u->ddir == DDIR_WRITE) {
1787 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1794 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1795 uint64_t offset, unsigned int bytes)
1802 if (f->first_write == -1ULL || offset < f->first_write)
1803 f->first_write = offset;
1804 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1805 f->last_write = offset + bytes;
1807 if (!f->last_write_comp)
1810 idx = f->last_write_idx++;
1811 f->last_write_comp[idx] = offset;
1812 if (f->last_write_idx == td->o.iodepth)
1813 f->last_write_idx = 0;
1816 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1817 struct io_completion_data *icd)
1819 struct io_u *io_u = *io_u_ptr;
1820 enum fio_ddir ddir = io_u->ddir;
1821 struct fio_file *f = io_u->file;
1823 dprint_io_u(io_u, "io complete");
1825 assert(io_u->flags & IO_U_F_FLIGHT);
1826 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1829 * Mark IO ok to verify
1833 * Remove errored entry from the verification list
1836 unlog_io_piece(td, io_u);
1838 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1843 if (ddir_sync(ddir)) {
1844 td->last_was_sync = 1;
1846 f->first_write = -1ULL;
1847 f->last_write = -1ULL;
1852 td->last_was_sync = 0;
1853 td->last_ddir = ddir;
1855 if (!io_u->error && ddir_rw(ddir)) {
1856 unsigned int bytes = io_u->buflen - io_u->resid;
1859 td->io_blocks[ddir]++;
1860 td->this_io_blocks[ddir]++;
1861 td->io_bytes[ddir] += bytes;
1863 if (!(io_u->flags & IO_U_F_VER_LIST))
1864 td->this_io_bytes[ddir] += bytes;
1866 if (ddir == DDIR_WRITE)
1867 file_log_write_comp(td, f, io_u->offset, bytes);
1869 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1870 td->runstate == TD_VERIFYING))
1871 account_io_completion(td, io_u, icd, ddir, bytes);
1873 icd->bytes_done[ddir] += bytes;
1876 ret = io_u->end_io(td, io_u_ptr);
1878 if (ret && !icd->error)
1881 } else if (io_u->error) {
1882 icd->error = io_u->error;
1883 io_u_log_error(td, io_u);
1886 enum error_type_bit eb = td_error_type(ddir, icd->error);
1888 if (!td_non_fatal_error(td, eb, icd->error))
1892 * If there is a non_fatal error, then add to the error count
1893 * and clear all the errors.
1895 update_error_count(td, icd->error);
1903 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1908 if (!gtod_reduce(td))
1909 fio_gettime(&icd->time, NULL);
1914 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1915 icd->bytes_done[ddir] = 0;
1918 static void ios_completed(struct thread_data *td,
1919 struct io_completion_data *icd)
1924 for (i = 0; i < icd->nr; i++) {
1925 io_u = td->io_ops->event(td, i);
1927 io_completed(td, &io_u, icd);
1935 * Complete a single io_u for the sync engines.
1937 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1939 struct io_completion_data icd;
1942 init_icd(td, &icd, 1);
1943 io_completed(td, &io_u, &icd);
1949 td_verror(td, icd.error, "io_u_sync_complete");
1953 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1954 td->bytes_done[ddir] += icd.bytes_done[ddir];
1960 * Called to complete min_events number of io for the async engines.
1962 int io_u_queued_complete(struct thread_data *td, int min_evts)
1964 struct io_completion_data icd;
1965 struct timespec *tvp = NULL;
1967 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1969 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
1973 else if (min_evts > td->cur_depth)
1974 min_evts = td->cur_depth;
1976 /* No worries, td_io_getevents fixes min and max if they are
1977 * set incorrectly */
1978 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1980 td_verror(td, -ret, "td_io_getevents");
1985 init_icd(td, &icd, ret);
1986 ios_completed(td, &icd);
1988 td_verror(td, icd.error, "io_u_queued_complete");
1992 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1993 td->bytes_done[ddir] += icd.bytes_done[ddir];
1999 * Call when io_u is really queued, to update the submission latency.
2001 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2003 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2004 unsigned long slat_time;
2006 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
2011 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2017 * See if we should reuse the last seed, if dedupe is enabled
2019 static struct frand_state *get_buf_state(struct thread_data *td)
2023 if (!td->o.dedupe_percentage)
2024 return &td->buf_state;
2025 else if (td->o.dedupe_percentage == 100) {
2026 frand_copy(&td->buf_state_prev, &td->buf_state);
2027 return &td->buf_state;
2030 v = rand32_between(&td->dedupe_state, 1, 100);
2032 if (v <= td->o.dedupe_percentage)
2033 return &td->buf_state_prev;
2035 return &td->buf_state;
2038 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2040 if (td->o.dedupe_percentage == 100)
2041 frand_copy(rs, &td->buf_state_prev);
2042 else if (rs == &td->buf_state)
2043 frand_copy(&td->buf_state_prev, rs);
2046 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2047 unsigned int max_bs)
2049 struct thread_options *o = &td->o;
2051 if (o->compress_percentage || o->dedupe_percentage) {
2052 unsigned int perc = td->o.compress_percentage;
2053 struct frand_state *rs;
2054 unsigned int left = max_bs;
2055 unsigned int this_write;
2058 rs = get_buf_state(td);
2060 min_write = min(min_write, left);
2063 this_write = min_not_zero(min_write,
2064 td->o.compress_chunk);
2066 fill_random_buf_percentage(rs, buf, perc,
2067 this_write, this_write,
2069 o->buffer_pattern_bytes);
2071 fill_random_buf(rs, buf, min_write);
2072 this_write = min_write;
2077 save_buf_state(td, rs);
2079 } else if (o->buffer_pattern_bytes)
2080 fill_buffer_pattern(td, buf, max_bs);
2081 else if (o->zero_buffers)
2082 memset(buf, 0, max_bs);
2084 fill_random_buf(get_buf_state(td), buf, max_bs);
2088 * "randomly" fill the buffer contents
2090 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2091 unsigned int min_write, unsigned int max_bs)
2093 io_u->buf_filled_len = 0;
2094 fill_io_buffer(td, io_u->buf, min_write, max_bs);