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?
66 max_size = f->io_size;
67 if (max_size > f->real_file_size)
68 max_size = f->real_file_size;
71 max_size = td->o.zone_range;
73 if (td->o.min_bs[ddir] > td->o.ba[ddir])
74 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
84 struct flist_head list;
88 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
89 enum fio_ddir ddir, uint64_t *b,
94 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
95 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
97 r = __rand(&td->random_state);
99 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
101 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
105 assert(fio_file_lfsr(f));
107 if (lfsr_next(&f->lfsr, &off))
114 * if we are not maintaining a random map, we are done.
116 if (!file_randommap(td, f))
120 * calculate map offset and check if it's free
122 if (random_map_free(f, *b))
125 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
126 (unsigned long long) *b);
128 *b = axmap_next_free(f->io_axmap, *b);
129 if (*b == (uint64_t) -1ULL)
135 static int __get_next_rand_offset_zipf(struct thread_data *td,
136 struct fio_file *f, enum fio_ddir ddir,
139 *b = zipf_next(&f->zipf);
143 static int __get_next_rand_offset_pareto(struct thread_data *td,
144 struct fio_file *f, enum fio_ddir ddir,
147 *b = pareto_next(&f->zipf);
151 static int __get_next_rand_offset_gauss(struct thread_data *td,
152 struct fio_file *f, enum fio_ddir ddir,
155 *b = gauss_next(&f->gauss);
159 static int __get_next_rand_offset_zoned(struct thread_data *td,
160 struct fio_file *f, enum fio_ddir ddir,
163 unsigned int v, send, stotal;
164 uint64_t offset, lastb;
166 struct zone_split_index *zsi;
168 lastb = last_block(td, f, ddir);
172 if (!td->o.zone_split_nr[ddir]) {
174 return __get_next_rand_offset(td, f, ddir, b, lastb);
178 * Generate a value, v, between 1 and 100, both inclusive
180 v = rand32_between(&td->zone_state, 1, 100);
182 zsi = &td->zone_state_index[ddir][v - 1];
183 stotal = zsi->size_perc_prev;
184 send = zsi->size_perc;
187 * Should never happen
191 log_err("fio: bug in zoned generation\n");
198 * 'send' is some percentage below or equal to 100 that
199 * marks the end of the current IO range. 'stotal' marks
200 * the start, in percent.
203 offset = stotal * lastb / 100ULL;
207 lastb = lastb * (send - stotal) / 100ULL;
210 * Generate index from 0..send-of-lastb
212 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
216 * Add our start offset, if any
224 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
226 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
227 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
229 return r1->off - r2->off;
232 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
233 enum fio_ddir ddir, uint64_t *b)
235 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
238 lastb = last_block(td, f, ddir);
242 return __get_next_rand_offset(td, f, ddir, b, lastb);
243 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
244 return __get_next_rand_offset_zipf(td, f, ddir, b);
245 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
246 return __get_next_rand_offset_pareto(td, f, ddir, b);
247 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
248 return __get_next_rand_offset_gauss(td, f, ddir, b);
249 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
250 return __get_next_rand_offset_zoned(td, f, ddir, b);
252 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
257 * Sort the reads for a verify phase in batches of verifysort_nr, if
260 static inline bool should_sort_io(struct thread_data *td)
262 if (!td->o.verifysort_nr || !td->o.do_verify)
266 if (td->runstate != TD_VERIFYING)
268 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
269 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
275 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
279 if (td->o.perc_rand[ddir] == 100)
282 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
284 return v <= td->o.perc_rand[ddir];
287 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
288 enum fio_ddir ddir, uint64_t *b)
293 if (!should_sort_io(td))
294 return get_off_from_method(td, f, ddir, b);
296 if (!flist_empty(&td->next_rand_list)) {
298 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
305 for (i = 0; i < td->o.verifysort_nr; i++) {
306 r = malloc(sizeof(*r));
308 ret = get_off_from_method(td, f, ddir, &r->off);
314 flist_add(&r->list, &td->next_rand_list);
320 assert(!flist_empty(&td->next_rand_list));
321 flist_sort(NULL, &td->next_rand_list, flist_cmp);
325 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
326 enum fio_ddir ddir, uint64_t *b)
328 if (!get_next_rand_offset(td, f, ddir, b))
331 if (td->o.time_based ||
332 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
333 fio_file_reset(td, f);
334 if (!get_next_rand_offset(td, f, ddir, b))
338 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
339 f->file_name, (unsigned long long) f->last_pos[ddir],
340 (unsigned long long) f->real_file_size);
344 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
345 enum fio_ddir ddir, uint64_t *offset)
347 struct thread_options *o = &td->o;
349 assert(ddir_rw(ddir));
351 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
353 struct thread_options *o = &td->o;
354 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
356 if (io_size > f->last_pos[ddir])
357 f->last_pos[ddir] = 0;
359 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
362 if (f->last_pos[ddir] < f->real_file_size) {
365 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
366 if (f->real_file_size > f->io_size)
367 f->last_pos[ddir] = f->io_size;
369 f->last_pos[ddir] = f->real_file_size;
372 pos = f->last_pos[ddir] - f->file_offset;
373 if (pos && o->ddir_seq_add) {
374 pos += o->ddir_seq_add;
377 * If we reach beyond the end of the file
378 * with holed IO, wrap around to the
379 * beginning again. If we're doing backwards IO,
382 if (pos >= f->real_file_size) {
383 if (o->ddir_seq_add > 0)
384 pos = f->file_offset;
386 if (f->real_file_size > f->io_size)
389 pos = f->real_file_size;
391 pos += o->ddir_seq_add;
403 static int get_next_block(struct thread_data *td, struct io_u *io_u,
404 enum fio_ddir ddir, int rw_seq,
405 unsigned int *is_random)
407 struct fio_file *f = io_u->file;
411 assert(ddir_rw(ddir));
417 if (should_do_random(td, ddir)) {
418 ret = get_next_rand_block(td, f, ddir, &b);
422 io_u_set(td, io_u, IO_U_F_BUSY_OK);
423 ret = get_next_seq_offset(td, f, ddir, &offset);
425 ret = get_next_rand_block(td, f, ddir, &b);
429 ret = get_next_seq_offset(td, f, ddir, &offset);
432 io_u_set(td, io_u, IO_U_F_BUSY_OK);
435 if (td->o.rw_seq == RW_SEQ_SEQ) {
436 ret = get_next_seq_offset(td, f, ddir, &offset);
438 ret = get_next_rand_block(td, f, ddir, &b);
441 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
442 if (f->last_start[ddir] != -1ULL)
443 offset = f->last_start[ddir] - f->file_offset;
448 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
455 io_u->offset = offset;
457 io_u->offset = b * td->o.ba[ddir];
459 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
468 * For random io, generate a random new block and see if it's used. Repeat
469 * until we find a free one. For sequential io, just return the end of
470 * the last io issued.
472 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
473 unsigned int *is_random)
475 struct fio_file *f = io_u->file;
476 enum fio_ddir ddir = io_u->ddir;
479 assert(ddir_rw(ddir));
481 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
483 td->ddir_seq_nr = td->o.ddir_seq_nr;
486 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
489 if (io_u->offset >= f->io_size) {
490 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
491 (unsigned long long) io_u->offset,
492 (unsigned long long) f->io_size);
496 io_u->offset += f->file_offset;
497 if (io_u->offset >= f->real_file_size) {
498 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
499 (unsigned long long) io_u->offset,
500 (unsigned long long) f->real_file_size);
507 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
508 unsigned int *is_random)
510 if (td->flags & TD_F_PROFILE_OPS) {
511 struct prof_io_ops *ops = &td->prof_io_ops;
513 if (ops->fill_io_u_off)
514 return ops->fill_io_u_off(td, io_u, is_random);
517 return __get_next_offset(td, io_u, is_random);
520 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
523 struct fio_file *f = io_u->file;
525 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
528 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
529 unsigned int is_random)
531 int ddir = io_u->ddir;
532 unsigned int buflen = 0;
533 unsigned int minbs, maxbs;
537 assert(ddir_rw(ddir));
539 if (td->o.bs_is_seq_rand)
540 ddir = is_random ? DDIR_WRITE: DDIR_READ;
542 minbs = td->o.min_bs[ddir];
543 maxbs = td->o.max_bs[ddir];
549 * If we can't satisfy the min block size from here, then fail
551 if (!io_u_fits(td, io_u, minbs))
554 frand_max = rand_max(&td->bsrange_state);
556 r = __rand(&td->bsrange_state);
558 if (!td->o.bssplit_nr[ddir]) {
559 buflen = 1 + (unsigned int) ((double) maxbs *
560 (r / (frand_max + 1.0)));
567 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
568 struct bssplit *bsp = &td->o.bssplit[ddir][i];
572 if ((r * 100UL <= frand_max * perc) &&
573 io_u_fits(td, io_u, buflen))
578 if (td->o.verify != VERIFY_NONE)
579 buflen = (buflen + td->o.verify_interval - 1) &
580 ~(td->o.verify_interval - 1);
582 if (!td->o.bs_unaligned && is_power_of_2(minbs))
583 buflen &= ~(minbs - 1);
585 } while (!io_u_fits(td, io_u, buflen));
590 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
591 unsigned int is_random)
593 if (td->flags & TD_F_PROFILE_OPS) {
594 struct prof_io_ops *ops = &td->prof_io_ops;
596 if (ops->fill_io_u_size)
597 return ops->fill_io_u_size(td, io_u, is_random);
600 return __get_next_buflen(td, io_u, is_random);
603 static void set_rwmix_bytes(struct thread_data *td)
608 * we do time or byte based switch. this is needed because
609 * buffered writes may issue a lot quicker than they complete,
610 * whereas reads do not.
612 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
613 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
616 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
620 v = rand32_between(&td->rwmix_state, 1, 100);
622 if (v <= td->o.rwmix[DDIR_READ])
628 int io_u_quiesce(struct thread_data *td)
633 * We are going to sleep, ensure that we flush anything pending as
634 * not to skew our latency numbers.
636 * Changed to only monitor 'in flight' requests here instead of the
637 * td->cur_depth, b/c td->cur_depth does not accurately represent
638 * io's that have been actually submitted to an async engine,
639 * and cur_depth is meaningless for sync engines.
641 if (td->io_u_queued || td->cur_depth) {
644 ret = td_io_commit(td);
647 while (td->io_u_in_flight) {
650 ret = io_u_queued_complete(td, 1);
658 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
660 enum fio_ddir odir = ddir ^ 1;
663 assert(ddir_rw(ddir));
664 now = utime_since_now(&td->start);
667 * if rate_next_io_time is in the past, need to catch up to rate
669 if (td->rate_next_io_time[ddir] <= now)
673 * We are ahead of rate in this direction. See if we
676 if (td_rw(td) && td->o.rwmix[odir]) {
678 * Other direction is behind rate, switch
680 if (td->rate_next_io_time[odir] <= now)
684 * Both directions are ahead of rate. sleep the min
685 * switch if necissary
687 if (td->rate_next_io_time[ddir] <=
688 td->rate_next_io_time[odir]) {
689 usec = td->rate_next_io_time[ddir] - now;
691 usec = td->rate_next_io_time[odir] - now;
695 usec = td->rate_next_io_time[ddir] - now;
697 if (td->o.io_submit_mode == IO_MODE_INLINE)
700 usec = usec_sleep(td, usec);
706 * Return the data direction for the next io_u. If the job is a
707 * mixed read/write workload, check the rwmix cycle and switch if
710 static enum fio_ddir get_rw_ddir(struct thread_data *td)
715 * see if it's time to fsync
717 if (td->o.fsync_blocks &&
718 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
719 td->io_issues[DDIR_WRITE] && should_fsync(td))
723 * see if it's time to fdatasync
725 if (td->o.fdatasync_blocks &&
726 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
727 td->io_issues[DDIR_WRITE] && should_fsync(td))
728 return DDIR_DATASYNC;
731 * see if it's time to sync_file_range
733 if (td->sync_file_range_nr &&
734 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
735 td->io_issues[DDIR_WRITE] && should_fsync(td))
736 return DDIR_SYNC_FILE_RANGE;
740 * Check if it's time to seed a new data direction.
742 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
744 * Put a top limit on how many bytes we do for
745 * one data direction, to avoid overflowing the
748 ddir = get_rand_ddir(td);
750 if (ddir != td->rwmix_ddir)
753 td->rwmix_ddir = ddir;
755 ddir = td->rwmix_ddir;
756 } else if (td_read(td))
758 else if (td_write(td))
763 td->rwmix_ddir = rate_ddir(td, ddir);
764 return td->rwmix_ddir;
767 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
769 enum fio_ddir ddir = get_rw_ddir(td);
771 if (td_trimwrite(td)) {
772 struct fio_file *f = io_u->file;
773 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
779 io_u->ddir = io_u->acct_ddir = ddir;
781 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
782 td->o.barrier_blocks &&
783 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
784 td->io_issues[DDIR_WRITE])
785 io_u_set(td, io_u, IO_U_F_BARRIER);
788 void put_file_log(struct thread_data *td, struct fio_file *f)
790 unsigned int ret = put_file(td, f);
793 td_verror(td, ret, "file close");
796 void put_io_u(struct thread_data *td, struct io_u *io_u)
803 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
804 put_file_log(td, io_u->file);
807 io_u_set(td, io_u, IO_U_F_FREE);
809 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
811 assert(!(td->flags & TD_F_CHILD));
813 io_u_qpush(&td->io_u_freelist, io_u);
815 td_io_u_free_notify(td);
818 void clear_io_u(struct thread_data *td, struct io_u *io_u)
820 io_u_clear(td, io_u, IO_U_F_FLIGHT);
824 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
826 struct io_u *__io_u = *io_u;
827 enum fio_ddir ddir = acct_ddir(__io_u);
829 dprint(FD_IO, "requeue %p\n", __io_u);
836 io_u_set(td, __io_u, IO_U_F_FREE);
837 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
838 td->io_issues[ddir]--;
840 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
841 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
843 assert(!(td->flags & TD_F_CHILD));
846 io_u_rpush(&td->io_u_requeues, __io_u);
848 td_io_u_free_notify(td);
852 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
854 unsigned int is_random;
856 if (td_ioengine_flagged(td, FIO_NOIO))
859 set_rw_ddir(td, io_u);
862 * fsync() or fdatasync() or trim etc, we are done
864 if (!ddir_rw(io_u->ddir))
868 * See if it's time to switch to a new zone
870 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
871 struct fio_file *f = io_u->file;
874 f->file_offset += td->o.zone_range + td->o.zone_skip;
877 * Wrap from the beginning, if we exceed the file size
879 if (f->file_offset >= f->real_file_size)
880 f->file_offset = f->real_file_size - f->file_offset;
881 f->last_pos[io_u->ddir] = f->file_offset;
882 td->io_skip_bytes += td->o.zone_skip;
886 * No log, let the seq/rand engine retrieve the next buflen and
889 if (get_next_offset(td, io_u, &is_random)) {
890 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
894 io_u->buflen = get_next_buflen(td, io_u, is_random);
896 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
900 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
901 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
902 dprint(FD_IO, " off=%llu/%lu > %llu\n",
903 (unsigned long long) io_u->offset, io_u->buflen,
904 (unsigned long long) io_u->file->real_file_size);
909 * mark entry before potentially trimming io_u
911 if (td_random(td) && file_randommap(td, io_u->file))
912 mark_random_map(td, io_u);
915 dprint_io_u(io_u, "fill_io_u");
916 td->zone_bytes += io_u->buflen;
920 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
949 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
951 __io_u_mark_map(td->ts.io_u_submit, nr);
952 td->ts.total_submit++;
955 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
957 __io_u_mark_map(td->ts.io_u_complete, nr);
958 td->ts.total_complete++;
961 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
965 switch (td->cur_depth) {
987 td->ts.io_u_map[idx] += nr;
990 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1027 assert(idx < FIO_IO_U_LAT_U_NR);
1028 td->ts.io_u_lat_u[idx]++;
1031 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1072 assert(idx < FIO_IO_U_LAT_M_NR);
1073 td->ts.io_u_lat_m[idx]++;
1076 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1079 io_u_mark_lat_usec(td, usec);
1081 io_u_mark_lat_msec(td, usec / 1000);
1084 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1086 unsigned long fileno;
1088 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1089 uint64_t frand_max = rand_max(&td->next_file_state);
1092 r = __rand(&td->next_file_state);
1093 return (unsigned int) ((double) td->o.nr_files
1094 * (r / (frand_max + 1.0)));
1097 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1098 fileno = zipf_next(&td->next_file_zipf);
1099 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1100 fileno = pareto_next(&td->next_file_zipf);
1101 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1102 fileno = gauss_next(&td->next_file_gauss);
1104 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1109 return fileno >> FIO_FSERVICE_SHIFT;
1113 * Get next file to service by choosing one at random
1115 static struct fio_file *get_next_file_rand(struct thread_data *td,
1116 enum fio_file_flags goodf,
1117 enum fio_file_flags badf)
1125 fno = __get_next_fileno_rand(td);
1128 if (fio_file_done(f))
1131 if (!fio_file_open(f)) {
1134 if (td->nr_open_files >= td->o.open_files)
1135 return ERR_PTR(-EBUSY);
1137 err = td_io_open_file(td, f);
1143 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1144 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1148 td_io_close_file(td, f);
1153 * Get next file to service by doing round robin between all available ones
1155 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1158 unsigned int old_next_file = td->next_file;
1164 f = td->files[td->next_file];
1167 if (td->next_file >= td->o.nr_files)
1170 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1171 if (fio_file_done(f)) {
1176 if (!fio_file_open(f)) {
1179 if (td->nr_open_files >= td->o.open_files)
1180 return ERR_PTR(-EBUSY);
1182 err = td_io_open_file(td, f);
1184 dprint(FD_FILE, "error %d on open of %s\n",
1192 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1194 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1198 td_io_close_file(td, f);
1201 } while (td->next_file != old_next_file);
1203 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1207 static struct fio_file *__get_next_file(struct thread_data *td)
1211 assert(td->o.nr_files <= td->files_index);
1213 if (td->nr_done_files >= td->o.nr_files) {
1214 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1215 " nr_files=%d\n", td->nr_open_files,
1221 f = td->file_service_file;
1222 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1223 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1225 if (td->file_service_left--)
1229 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1230 td->o.file_service_type == FIO_FSERVICE_SEQ)
1231 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1233 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1238 td->file_service_file = f;
1239 td->file_service_left = td->file_service_nr - 1;
1242 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1244 dprint(FD_FILE, "get_next_file: NULL\n");
1248 static struct fio_file *get_next_file(struct thread_data *td)
1250 if (td->flags & TD_F_PROFILE_OPS) {
1251 struct prof_io_ops *ops = &td->prof_io_ops;
1253 if (ops->get_next_file)
1254 return ops->get_next_file(td);
1257 return __get_next_file(td);
1260 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1265 f = get_next_file(td);
1266 if (IS_ERR_OR_NULL(f))
1272 if (!fill_io_u(td, io_u))
1275 put_file_log(td, f);
1276 td_io_close_file(td, f);
1278 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1279 fio_file_reset(td, f);
1281 fio_file_set_done(f);
1282 td->nr_done_files++;
1283 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1284 td->nr_done_files, td->o.nr_files);
1291 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1292 unsigned long tusec, unsigned long max_usec)
1295 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1296 td_verror(td, ETIMEDOUT, "max latency exceeded");
1297 icd->error = ETIMEDOUT;
1300 static void lat_new_cycle(struct thread_data *td)
1302 fio_gettime(&td->latency_ts, NULL);
1303 td->latency_ios = ddir_rw_sum(td->io_blocks);
1304 td->latency_failed = 0;
1308 * We had an IO outside the latency target. Reduce the queue depth. If we
1309 * are at QD=1, then it's time to give up.
1311 static bool __lat_target_failed(struct thread_data *td)
1313 if (td->latency_qd == 1)
1316 td->latency_qd_high = td->latency_qd;
1318 if (td->latency_qd == td->latency_qd_low)
1319 td->latency_qd_low--;
1321 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1323 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1326 * When we ramp QD down, quiesce existing IO to prevent
1327 * a storm of ramp downs due to pending higher depth.
1334 static bool lat_target_failed(struct thread_data *td)
1336 if (td->o.latency_percentile.u.f == 100.0)
1337 return __lat_target_failed(td);
1339 td->latency_failed++;
1343 void lat_target_init(struct thread_data *td)
1345 td->latency_end_run = 0;
1347 if (td->o.latency_target) {
1348 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1349 fio_gettime(&td->latency_ts, NULL);
1351 td->latency_qd_high = td->o.iodepth;
1352 td->latency_qd_low = 1;
1353 td->latency_ios = ddir_rw_sum(td->io_blocks);
1355 td->latency_qd = td->o.iodepth;
1358 void lat_target_reset(struct thread_data *td)
1360 if (!td->latency_end_run)
1361 lat_target_init(td);
1364 static void lat_target_success(struct thread_data *td)
1366 const unsigned int qd = td->latency_qd;
1367 struct thread_options *o = &td->o;
1369 td->latency_qd_low = td->latency_qd;
1372 * If we haven't failed yet, we double up to a failing value instead
1373 * of bisecting from highest possible queue depth. If we have set
1374 * a limit other than td->o.iodepth, bisect between that.
1376 if (td->latency_qd_high != o->iodepth)
1377 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1379 td->latency_qd *= 2;
1381 if (td->latency_qd > o->iodepth)
1382 td->latency_qd = o->iodepth;
1384 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1387 * Same as last one, we are done. Let it run a latency cycle, so
1388 * we get only the results from the targeted depth.
1390 if (td->latency_qd == qd) {
1391 if (td->latency_end_run) {
1392 dprint(FD_RATE, "We are done\n");
1395 dprint(FD_RATE, "Quiesce and final run\n");
1397 td->latency_end_run = 1;
1398 reset_all_stats(td);
1407 * Check if we can bump the queue depth
1409 void lat_target_check(struct thread_data *td)
1411 uint64_t usec_window;
1415 usec_window = utime_since_now(&td->latency_ts);
1416 if (usec_window < td->o.latency_window)
1419 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1420 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1421 success_ios *= 100.0;
1423 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1425 if (success_ios >= td->o.latency_percentile.u.f)
1426 lat_target_success(td);
1428 __lat_target_failed(td);
1432 * If latency target is enabled, we might be ramping up or down and not
1433 * using the full queue depth available.
1435 bool queue_full(const struct thread_data *td)
1437 const int qempty = io_u_qempty(&td->io_u_freelist);
1441 if (!td->o.latency_target)
1444 return td->cur_depth >= td->latency_qd;
1447 struct io_u *__get_io_u(struct thread_data *td)
1449 struct io_u *io_u = NULL;
1457 if (!io_u_rempty(&td->io_u_requeues))
1458 io_u = io_u_rpop(&td->io_u_requeues);
1459 else if (!queue_full(td)) {
1460 io_u = io_u_qpop(&td->io_u_freelist);
1465 io_u->end_io = NULL;
1469 assert(io_u->flags & IO_U_F_FREE);
1470 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1471 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1475 io_u->acct_ddir = -1;
1477 assert(!(td->flags & TD_F_CHILD));
1478 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1480 } else if (td_async_processing(td)) {
1482 * We ran out, wait for async verify threads to finish and
1485 assert(!(td->flags & TD_F_CHILD));
1486 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1494 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1496 if (!(td->flags & TD_F_TRIM_BACKLOG))
1499 if (td->trim_entries) {
1502 if (td->trim_batch) {
1505 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1506 td->last_ddir != DDIR_READ) {
1507 td->trim_batch = td->o.trim_batch;
1508 if (!td->trim_batch)
1509 td->trim_batch = td->o.trim_backlog;
1513 if (get_trim && get_next_trim(td, io_u))
1520 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1522 if (!(td->flags & TD_F_VER_BACKLOG))
1525 if (td->io_hist_len) {
1528 if (td->verify_batch)
1530 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1531 td->last_ddir != DDIR_READ) {
1532 td->verify_batch = td->o.verify_batch;
1533 if (!td->verify_batch)
1534 td->verify_batch = td->o.verify_backlog;
1538 if (get_verify && !get_next_verify(td, io_u)) {
1548 * Fill offset and start time into the buffer content, to prevent too
1549 * easy compressible data for simple de-dupe attempts. Do this for every
1550 * 512b block in the range, since that should be the smallest block size
1551 * we can expect from a device.
1553 static void small_content_scramble(struct io_u *io_u)
1555 unsigned int i, nr_blocks = io_u->buflen / 512;
1557 unsigned int offset;
1564 boffset = io_u->offset;
1565 io_u->buf_filled_len = 0;
1567 for (i = 0; i < nr_blocks; i++) {
1569 * Fill the byte offset into a "random" start offset of
1570 * the buffer, given by the product of the usec time
1571 * and the actual offset.
1573 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1574 offset &= ~(sizeof(uint64_t) - 1);
1575 if (offset >= 512 - sizeof(uint64_t))
1576 offset -= sizeof(uint64_t);
1577 memcpy(p + offset, &boffset, sizeof(boffset));
1579 end = p + 512 - sizeof(io_u->start_time);
1580 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1587 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1588 * etc. The returned io_u is fully ready to be prepped and submitted.
1590 struct io_u *get_io_u(struct thread_data *td)
1594 int do_scramble = 0;
1597 io_u = __get_io_u(td);
1599 dprint(FD_IO, "__get_io_u failed\n");
1603 if (check_get_verify(td, io_u))
1605 if (check_get_trim(td, io_u))
1609 * from a requeue, io_u already setup
1615 * If using an iolog, grab next piece if any available.
1617 if (td->flags & TD_F_READ_IOLOG) {
1618 if (read_iolog_get(td, io_u))
1620 } else if (set_io_u_file(td, io_u)) {
1622 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1628 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1632 assert(fio_file_open(f));
1634 if (ddir_rw(io_u->ddir)) {
1635 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1636 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1640 f->last_start[io_u->ddir] = io_u->offset;
1641 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1643 if (io_u->ddir == DDIR_WRITE) {
1644 if (td->flags & TD_F_REFILL_BUFFERS) {
1645 io_u_fill_buffer(td, io_u,
1646 td->o.min_bs[DDIR_WRITE],
1648 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1649 !(td->flags & TD_F_COMPRESS))
1651 if (td->flags & TD_F_VER_NONE) {
1652 populate_verify_io_u(td, io_u);
1655 } else if (io_u->ddir == DDIR_READ) {
1657 * Reset the buf_filled parameters so next time if the
1658 * buffer is used for writes it is refilled.
1660 io_u->buf_filled_len = 0;
1665 * Set io data pointers.
1667 io_u->xfer_buf = io_u->buf;
1668 io_u->xfer_buflen = io_u->buflen;
1672 if (!td_io_prep(td, io_u)) {
1673 if (!td->o.disable_lat)
1674 fio_gettime(&io_u->start_time, NULL);
1676 small_content_scramble(io_u);
1680 dprint(FD_IO, "get_io_u failed\n");
1682 return ERR_PTR(ret);
1685 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1687 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1689 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1692 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1693 io_u->file ? " on file " : "",
1694 io_u->file ? io_u->file->file_name : "",
1695 strerror(io_u->error),
1696 io_ddir_name(io_u->ddir),
1697 io_u->offset, io_u->xfer_buflen);
1699 if (td->io_ops->errdetails) {
1700 char *err = td->io_ops->errdetails(io_u);
1702 log_err("fio: %s\n", err);
1707 td_verror(td, io_u->error, "io_u error");
1710 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1712 __io_u_log_error(td, io_u);
1714 __io_u_log_error(td->parent, io_u);
1717 static inline bool gtod_reduce(struct thread_data *td)
1719 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1720 || td->o.gtod_reduce;
1723 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1724 struct io_completion_data *icd,
1725 const enum fio_ddir idx, unsigned int bytes)
1727 const int no_reduce = !gtod_reduce(td);
1728 unsigned long lusec = 0;
1734 lusec = utime_since(&io_u->issue_time, &icd->time);
1736 if (!td->o.disable_lat) {
1737 unsigned long tusec;
1739 tusec = utime_since(&io_u->start_time, &icd->time);
1740 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1742 if (td->flags & TD_F_PROFILE_OPS) {
1743 struct prof_io_ops *ops = &td->prof_io_ops;
1746 icd->error = ops->io_u_lat(td, tusec);
1749 if (td->o.max_latency && tusec > td->o.max_latency)
1750 lat_fatal(td, icd, tusec, td->o.max_latency);
1751 if (td->o.latency_target && tusec > td->o.latency_target) {
1752 if (lat_target_failed(td))
1753 lat_fatal(td, icd, tusec, td->o.latency_target);
1758 if (!td->o.disable_clat) {
1759 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1760 io_u_mark_latency(td, lusec);
1763 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1764 add_bw_sample(td, io_u, bytes, lusec);
1766 if (no_reduce && per_unit_log(td->iops_log))
1767 add_iops_sample(td, io_u, bytes);
1770 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1771 uint32_t *info = io_u_block_info(td, io_u);
1772 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1773 if (io_u->ddir == DDIR_TRIM) {
1774 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1775 BLOCK_INFO_TRIMS(*info) + 1);
1776 } else if (io_u->ddir == DDIR_WRITE) {
1777 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1784 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1785 uint64_t offset, unsigned int bytes)
1792 if (f->first_write == -1ULL || offset < f->first_write)
1793 f->first_write = offset;
1794 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1795 f->last_write = offset + bytes;
1797 if (!f->last_write_comp)
1800 idx = f->last_write_idx++;
1801 f->last_write_comp[idx] = offset;
1802 if (f->last_write_idx == td->o.iodepth)
1803 f->last_write_idx = 0;
1806 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1807 struct io_completion_data *icd)
1809 struct io_u *io_u = *io_u_ptr;
1810 enum fio_ddir ddir = io_u->ddir;
1811 struct fio_file *f = io_u->file;
1813 dprint_io_u(io_u, "io complete");
1815 assert(io_u->flags & IO_U_F_FLIGHT);
1816 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1819 * Mark IO ok to verify
1823 * Remove errored entry from the verification list
1826 unlog_io_piece(td, io_u);
1828 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1833 if (ddir_sync(ddir)) {
1834 td->last_was_sync = 1;
1836 f->first_write = -1ULL;
1837 f->last_write = -1ULL;
1842 td->last_was_sync = 0;
1843 td->last_ddir = ddir;
1845 if (!io_u->error && ddir_rw(ddir)) {
1846 unsigned int bytes = io_u->buflen - io_u->resid;
1849 td->io_blocks[ddir]++;
1850 td->this_io_blocks[ddir]++;
1851 td->io_bytes[ddir] += bytes;
1853 if (!(io_u->flags & IO_U_F_VER_LIST))
1854 td->this_io_bytes[ddir] += bytes;
1856 if (ddir == DDIR_WRITE)
1857 file_log_write_comp(td, f, io_u->offset, bytes);
1859 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1860 td->runstate == TD_VERIFYING))
1861 account_io_completion(td, io_u, icd, ddir, bytes);
1863 icd->bytes_done[ddir] += bytes;
1866 ret = io_u->end_io(td, io_u_ptr);
1868 if (ret && !icd->error)
1871 } else if (io_u->error) {
1872 icd->error = io_u->error;
1873 io_u_log_error(td, io_u);
1876 enum error_type_bit eb = td_error_type(ddir, icd->error);
1878 if (!td_non_fatal_error(td, eb, icd->error))
1882 * If there is a non_fatal error, then add to the error count
1883 * and clear all the errors.
1885 update_error_count(td, icd->error);
1893 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1898 if (!gtod_reduce(td))
1899 fio_gettime(&icd->time, NULL);
1904 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1905 icd->bytes_done[ddir] = 0;
1908 static void ios_completed(struct thread_data *td,
1909 struct io_completion_data *icd)
1914 for (i = 0; i < icd->nr; i++) {
1915 io_u = td->io_ops->event(td, i);
1917 io_completed(td, &io_u, icd);
1925 * Complete a single io_u for the sync engines.
1927 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1929 struct io_completion_data icd;
1932 init_icd(td, &icd, 1);
1933 io_completed(td, &io_u, &icd);
1939 td_verror(td, icd.error, "io_u_sync_complete");
1943 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1944 td->bytes_done[ddir] += icd.bytes_done[ddir];
1950 * Called to complete min_events number of io for the async engines.
1952 int io_u_queued_complete(struct thread_data *td, int min_evts)
1954 struct io_completion_data icd;
1955 struct timespec *tvp = NULL;
1957 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1959 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1963 else if (min_evts > td->cur_depth)
1964 min_evts = td->cur_depth;
1966 /* No worries, td_io_getevents fixes min and max if they are
1967 * set incorrectly */
1968 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1970 td_verror(td, -ret, "td_io_getevents");
1975 init_icd(td, &icd, ret);
1976 ios_completed(td, &icd);
1978 td_verror(td, icd.error, "io_u_queued_complete");
1982 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1983 td->bytes_done[ddir] += icd.bytes_done[ddir];
1989 * Call when io_u is really queued, to update the submission latency.
1991 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1993 if (!td->o.disable_slat) {
1994 unsigned long slat_time;
1996 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
2001 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2007 * See if we should reuse the last seed, if dedupe is enabled
2009 static struct frand_state *get_buf_state(struct thread_data *td)
2013 if (!td->o.dedupe_percentage)
2014 return &td->buf_state;
2015 else if (td->o.dedupe_percentage == 100) {
2016 frand_copy(&td->buf_state_prev, &td->buf_state);
2017 return &td->buf_state;
2020 v = rand32_between(&td->dedupe_state, 1, 100);
2022 if (v <= td->o.dedupe_percentage)
2023 return &td->buf_state_prev;
2025 return &td->buf_state;
2028 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2030 if (td->o.dedupe_percentage == 100)
2031 frand_copy(rs, &td->buf_state_prev);
2032 else if (rs == &td->buf_state)
2033 frand_copy(&td->buf_state_prev, rs);
2036 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2037 unsigned int max_bs)
2039 struct thread_options *o = &td->o;
2041 if (o->compress_percentage || o->dedupe_percentage) {
2042 unsigned int perc = td->o.compress_percentage;
2043 struct frand_state *rs;
2044 unsigned int left = max_bs;
2045 unsigned int this_write;
2048 rs = get_buf_state(td);
2050 min_write = min(min_write, left);
2053 this_write = min_not_zero(min_write,
2054 td->o.compress_chunk);
2056 fill_random_buf_percentage(rs, buf, perc,
2057 this_write, this_write,
2059 o->buffer_pattern_bytes);
2061 fill_random_buf(rs, buf, min_write);
2062 this_write = min_write;
2067 save_buf_state(td, rs);
2069 } else if (o->buffer_pattern_bytes)
2070 fill_buffer_pattern(td, buf, max_bs);
2071 else if (o->zero_buffers)
2072 memset(buf, 0, max_bs);
2074 fill_random_buf(get_buf_state(td), buf, max_bs);
2078 * "randomly" fill the buffer contents
2080 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2081 unsigned int min_write, unsigned int max_bs)
2083 io_u->buf_filled_len = 0;
2084 fill_io_buffer(td, io_u->buf, min_write, max_bs);