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 f->last_pos[ddir] = f->real_file_size;
368 pos = f->last_pos[ddir] - f->file_offset;
369 if (pos && o->ddir_seq_add) {
370 pos += o->ddir_seq_add;
373 * If we reach beyond the end of the file
374 * with holed IO, wrap around to the
375 * beginning again. If we're doing backwards IO,
378 if (pos >= f->real_file_size) {
379 if (o->ddir_seq_add > 0)
380 pos = f->file_offset;
382 pos = f->real_file_size + o->ddir_seq_add;
393 static int get_next_block(struct thread_data *td, struct io_u *io_u,
394 enum fio_ddir ddir, int rw_seq,
395 unsigned int *is_random)
397 struct fio_file *f = io_u->file;
401 assert(ddir_rw(ddir));
407 if (should_do_random(td, ddir)) {
408 ret = get_next_rand_block(td, f, ddir, &b);
412 io_u_set(td, io_u, IO_U_F_BUSY_OK);
413 ret = get_next_seq_offset(td, f, ddir, &offset);
415 ret = get_next_rand_block(td, f, ddir, &b);
419 ret = get_next_seq_offset(td, f, ddir, &offset);
422 io_u_set(td, io_u, IO_U_F_BUSY_OK);
425 if (td->o.rw_seq == RW_SEQ_SEQ) {
426 ret = get_next_seq_offset(td, f, ddir, &offset);
428 ret = get_next_rand_block(td, f, ddir, &b);
431 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
432 if (f->last_start[ddir] != -1ULL)
433 offset = f->last_start[ddir] - f->file_offset;
438 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
445 io_u->offset = offset;
447 io_u->offset = b * td->o.ba[ddir];
449 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
458 * For random io, generate a random new block and see if it's used. Repeat
459 * until we find a free one. For sequential io, just return the end of
460 * the last io issued.
462 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
463 unsigned int *is_random)
465 struct fio_file *f = io_u->file;
466 enum fio_ddir ddir = io_u->ddir;
469 assert(ddir_rw(ddir));
471 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
473 td->ddir_seq_nr = td->o.ddir_seq_nr;
476 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
479 if (io_u->offset >= f->io_size) {
480 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
481 (unsigned long long) io_u->offset,
482 (unsigned long long) f->io_size);
486 io_u->offset += f->file_offset;
487 if (io_u->offset >= f->real_file_size) {
488 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
489 (unsigned long long) io_u->offset,
490 (unsigned long long) f->real_file_size);
497 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
498 unsigned int *is_random)
500 if (td->flags & TD_F_PROFILE_OPS) {
501 struct prof_io_ops *ops = &td->prof_io_ops;
503 if (ops->fill_io_u_off)
504 return ops->fill_io_u_off(td, io_u, is_random);
507 return __get_next_offset(td, io_u, is_random);
510 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
513 struct fio_file *f = io_u->file;
515 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
518 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
519 unsigned int is_random)
521 int ddir = io_u->ddir;
522 unsigned int buflen = 0;
523 unsigned int minbs, maxbs;
527 assert(ddir_rw(ddir));
529 if (td->o.bs_is_seq_rand)
530 ddir = is_random ? DDIR_WRITE: DDIR_READ;
532 minbs = td->o.min_bs[ddir];
533 maxbs = td->o.max_bs[ddir];
539 * If we can't satisfy the min block size from here, then fail
541 if (!io_u_fits(td, io_u, minbs))
544 frand_max = rand_max(&td->bsrange_state);
546 r = __rand(&td->bsrange_state);
548 if (!td->o.bssplit_nr[ddir]) {
549 buflen = 1 + (unsigned int) ((double) maxbs *
550 (r / (frand_max + 1.0)));
557 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
558 struct bssplit *bsp = &td->o.bssplit[ddir][i];
562 if ((r * 100UL <= frand_max * perc) &&
563 io_u_fits(td, io_u, buflen))
568 if (td->o.verify != VERIFY_NONE)
569 buflen = (buflen + td->o.verify_interval - 1) &
570 ~(td->o.verify_interval - 1);
572 if (!td->o.bs_unaligned && is_power_of_2(minbs))
573 buflen &= ~(minbs - 1);
575 } while (!io_u_fits(td, io_u, buflen));
580 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
581 unsigned int is_random)
583 if (td->flags & TD_F_PROFILE_OPS) {
584 struct prof_io_ops *ops = &td->prof_io_ops;
586 if (ops->fill_io_u_size)
587 return ops->fill_io_u_size(td, io_u, is_random);
590 return __get_next_buflen(td, io_u, is_random);
593 static void set_rwmix_bytes(struct thread_data *td)
598 * we do time or byte based switch. this is needed because
599 * buffered writes may issue a lot quicker than they complete,
600 * whereas reads do not.
602 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
603 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
606 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
610 v = rand32_between(&td->rwmix_state, 1, 100);
612 if (v <= td->o.rwmix[DDIR_READ])
618 int io_u_quiesce(struct thread_data *td)
623 * We are going to sleep, ensure that we flush anything pending as
624 * not to skew our latency numbers.
626 * Changed to only monitor 'in flight' requests here instead of the
627 * td->cur_depth, b/c td->cur_depth does not accurately represent
628 * io's that have been actually submitted to an async engine,
629 * and cur_depth is meaningless for sync engines.
631 if (td->io_u_queued || td->cur_depth) {
634 ret = td_io_commit(td);
637 while (td->io_u_in_flight) {
640 ret = io_u_queued_complete(td, 1);
648 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
650 enum fio_ddir odir = ddir ^ 1;
653 assert(ddir_rw(ddir));
654 now = utime_since_now(&td->start);
657 * if rate_next_io_time is in the past, need to catch up to rate
659 if (td->rate_next_io_time[ddir] <= now)
663 * We are ahead of rate in this direction. See if we
666 if (td_rw(td) && td->o.rwmix[odir]) {
668 * Other direction is behind rate, switch
670 if (td->rate_next_io_time[odir] <= now)
674 * Both directions are ahead of rate. sleep the min
675 * switch if necissary
677 if (td->rate_next_io_time[ddir] <=
678 td->rate_next_io_time[odir]) {
679 usec = td->rate_next_io_time[ddir] - now;
681 usec = td->rate_next_io_time[odir] - now;
685 usec = td->rate_next_io_time[ddir] - now;
687 if (td->o.io_submit_mode == IO_MODE_INLINE)
690 usec = usec_sleep(td, usec);
696 * Return the data direction for the next io_u. If the job is a
697 * mixed read/write workload, check the rwmix cycle and switch if
700 static enum fio_ddir get_rw_ddir(struct thread_data *td)
705 * see if it's time to fsync
707 if (td->o.fsync_blocks &&
708 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
709 td->io_issues[DDIR_WRITE] && should_fsync(td))
713 * see if it's time to fdatasync
715 if (td->o.fdatasync_blocks &&
716 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
717 td->io_issues[DDIR_WRITE] && should_fsync(td))
718 return DDIR_DATASYNC;
721 * see if it's time to sync_file_range
723 if (td->sync_file_range_nr &&
724 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
725 td->io_issues[DDIR_WRITE] && should_fsync(td))
726 return DDIR_SYNC_FILE_RANGE;
730 * Check if it's time to seed a new data direction.
732 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
734 * Put a top limit on how many bytes we do for
735 * one data direction, to avoid overflowing the
738 ddir = get_rand_ddir(td);
740 if (ddir != td->rwmix_ddir)
743 td->rwmix_ddir = ddir;
745 ddir = td->rwmix_ddir;
746 } else if (td_read(td))
748 else if (td_write(td))
753 td->rwmix_ddir = rate_ddir(td, ddir);
754 return td->rwmix_ddir;
757 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
759 enum fio_ddir ddir = get_rw_ddir(td);
761 if (td_trimwrite(td)) {
762 struct fio_file *f = io_u->file;
763 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
769 io_u->ddir = io_u->acct_ddir = ddir;
771 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
772 td->o.barrier_blocks &&
773 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
774 td->io_issues[DDIR_WRITE])
775 io_u_set(td, io_u, IO_U_F_BARRIER);
778 void put_file_log(struct thread_data *td, struct fio_file *f)
780 unsigned int ret = put_file(td, f);
783 td_verror(td, ret, "file close");
786 void put_io_u(struct thread_data *td, struct io_u *io_u)
793 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
794 put_file_log(td, io_u->file);
797 io_u_set(td, io_u, IO_U_F_FREE);
799 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
801 assert(!(td->flags & TD_F_CHILD));
803 io_u_qpush(&td->io_u_freelist, io_u);
805 td_io_u_free_notify(td);
808 void clear_io_u(struct thread_data *td, struct io_u *io_u)
810 io_u_clear(td, io_u, IO_U_F_FLIGHT);
814 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
816 struct io_u *__io_u = *io_u;
817 enum fio_ddir ddir = acct_ddir(__io_u);
819 dprint(FD_IO, "requeue %p\n", __io_u);
826 io_u_set(td, __io_u, IO_U_F_FREE);
827 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
828 td->io_issues[ddir]--;
830 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
831 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
833 assert(!(td->flags & TD_F_CHILD));
836 io_u_rpush(&td->io_u_requeues, __io_u);
838 td_io_u_free_notify(td);
842 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
844 unsigned int is_random;
846 if (td->io_ops->flags & FIO_NOIO)
849 set_rw_ddir(td, io_u);
852 * fsync() or fdatasync() or trim etc, we are done
854 if (!ddir_rw(io_u->ddir))
858 * See if it's time to switch to a new zone
860 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
861 struct fio_file *f = io_u->file;
864 f->file_offset += td->o.zone_range + td->o.zone_skip;
867 * Wrap from the beginning, if we exceed the file size
869 if (f->file_offset >= f->real_file_size)
870 f->file_offset = f->real_file_size - f->file_offset;
871 f->last_pos[io_u->ddir] = f->file_offset;
872 td->io_skip_bytes += td->o.zone_skip;
876 * No log, let the seq/rand engine retrieve the next buflen and
879 if (get_next_offset(td, io_u, &is_random)) {
880 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
884 io_u->buflen = get_next_buflen(td, io_u, is_random);
886 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
890 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
891 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
892 dprint(FD_IO, " off=%llu/%lu > %llu\n",
893 (unsigned long long) io_u->offset, io_u->buflen,
894 (unsigned long long) io_u->file->real_file_size);
899 * mark entry before potentially trimming io_u
901 if (td_random(td) && file_randommap(td, io_u->file))
902 mark_random_map(td, io_u);
905 dprint_io_u(io_u, "fill_io_u");
906 td->zone_bytes += io_u->buflen;
910 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
939 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
941 __io_u_mark_map(td->ts.io_u_submit, nr);
942 td->ts.total_submit++;
945 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
947 __io_u_mark_map(td->ts.io_u_complete, nr);
948 td->ts.total_complete++;
951 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
955 switch (td->cur_depth) {
977 td->ts.io_u_map[idx] += nr;
980 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1017 assert(idx < FIO_IO_U_LAT_U_NR);
1018 td->ts.io_u_lat_u[idx]++;
1021 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1062 assert(idx < FIO_IO_U_LAT_M_NR);
1063 td->ts.io_u_lat_m[idx]++;
1066 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1069 io_u_mark_lat_usec(td, usec);
1071 io_u_mark_lat_msec(td, usec / 1000);
1074 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1076 unsigned long fileno;
1078 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1079 uint64_t frand_max = rand_max(&td->next_file_state);
1082 r = __rand(&td->next_file_state);
1083 return (unsigned int) ((double) td->o.nr_files
1084 * (r / (frand_max + 1.0)));
1087 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1088 fileno = zipf_next(&td->next_file_zipf);
1089 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1090 fileno = pareto_next(&td->next_file_zipf);
1091 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1092 fileno = gauss_next(&td->next_file_gauss);
1094 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1099 return fileno >> FIO_FSERVICE_SHIFT;
1103 * Get next file to service by choosing one at random
1105 static struct fio_file *get_next_file_rand(struct thread_data *td,
1106 enum fio_file_flags goodf,
1107 enum fio_file_flags badf)
1115 fno = __get_next_fileno_rand(td);
1118 if (fio_file_done(f))
1121 if (!fio_file_open(f)) {
1124 if (td->nr_open_files >= td->o.open_files)
1125 return ERR_PTR(-EBUSY);
1127 err = td_io_open_file(td, f);
1133 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1134 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1138 td_io_close_file(td, f);
1143 * Get next file to service by doing round robin between all available ones
1145 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1148 unsigned int old_next_file = td->next_file;
1154 f = td->files[td->next_file];
1157 if (td->next_file >= td->o.nr_files)
1160 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1161 if (fio_file_done(f)) {
1166 if (!fio_file_open(f)) {
1169 if (td->nr_open_files >= td->o.open_files)
1170 return ERR_PTR(-EBUSY);
1172 err = td_io_open_file(td, f);
1174 dprint(FD_FILE, "error %d on open of %s\n",
1182 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1184 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1188 td_io_close_file(td, f);
1191 } while (td->next_file != old_next_file);
1193 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1197 static struct fio_file *__get_next_file(struct thread_data *td)
1201 assert(td->o.nr_files <= td->files_index);
1203 if (td->nr_done_files >= td->o.nr_files) {
1204 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1205 " nr_files=%d\n", td->nr_open_files,
1211 f = td->file_service_file;
1212 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1213 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1215 if (td->file_service_left--)
1219 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1220 td->o.file_service_type == FIO_FSERVICE_SEQ)
1221 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1223 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1228 td->file_service_file = f;
1229 td->file_service_left = td->file_service_nr - 1;
1232 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1234 dprint(FD_FILE, "get_next_file: NULL\n");
1238 static struct fio_file *get_next_file(struct thread_data *td)
1240 if (td->flags & TD_F_PROFILE_OPS) {
1241 struct prof_io_ops *ops = &td->prof_io_ops;
1243 if (ops->get_next_file)
1244 return ops->get_next_file(td);
1247 return __get_next_file(td);
1250 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1255 f = get_next_file(td);
1256 if (IS_ERR_OR_NULL(f))
1262 if (!fill_io_u(td, io_u))
1265 put_file_log(td, f);
1266 td_io_close_file(td, f);
1268 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1269 fio_file_reset(td, f);
1271 fio_file_set_done(f);
1272 td->nr_done_files++;
1273 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1274 td->nr_done_files, td->o.nr_files);
1281 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1282 unsigned long tusec, unsigned long max_usec)
1285 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1286 td_verror(td, ETIMEDOUT, "max latency exceeded");
1287 icd->error = ETIMEDOUT;
1290 static void lat_new_cycle(struct thread_data *td)
1292 fio_gettime(&td->latency_ts, NULL);
1293 td->latency_ios = ddir_rw_sum(td->io_blocks);
1294 td->latency_failed = 0;
1298 * We had an IO outside the latency target. Reduce the queue depth. If we
1299 * are at QD=1, then it's time to give up.
1301 static bool __lat_target_failed(struct thread_data *td)
1303 if (td->latency_qd == 1)
1306 td->latency_qd_high = td->latency_qd;
1308 if (td->latency_qd == td->latency_qd_low)
1309 td->latency_qd_low--;
1311 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1313 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1316 * When we ramp QD down, quiesce existing IO to prevent
1317 * a storm of ramp downs due to pending higher depth.
1324 static bool lat_target_failed(struct thread_data *td)
1326 if (td->o.latency_percentile.u.f == 100.0)
1327 return __lat_target_failed(td);
1329 td->latency_failed++;
1333 void lat_target_init(struct thread_data *td)
1335 td->latency_end_run = 0;
1337 if (td->o.latency_target) {
1338 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1339 fio_gettime(&td->latency_ts, NULL);
1341 td->latency_qd_high = td->o.iodepth;
1342 td->latency_qd_low = 1;
1343 td->latency_ios = ddir_rw_sum(td->io_blocks);
1345 td->latency_qd = td->o.iodepth;
1348 void lat_target_reset(struct thread_data *td)
1350 if (!td->latency_end_run)
1351 lat_target_init(td);
1354 static void lat_target_success(struct thread_data *td)
1356 const unsigned int qd = td->latency_qd;
1357 struct thread_options *o = &td->o;
1359 td->latency_qd_low = td->latency_qd;
1362 * If we haven't failed yet, we double up to a failing value instead
1363 * of bisecting from highest possible queue depth. If we have set
1364 * a limit other than td->o.iodepth, bisect between that.
1366 if (td->latency_qd_high != o->iodepth)
1367 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1369 td->latency_qd *= 2;
1371 if (td->latency_qd > o->iodepth)
1372 td->latency_qd = o->iodepth;
1374 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1377 * Same as last one, we are done. Let it run a latency cycle, so
1378 * we get only the results from the targeted depth.
1380 if (td->latency_qd == qd) {
1381 if (td->latency_end_run) {
1382 dprint(FD_RATE, "We are done\n");
1385 dprint(FD_RATE, "Quiesce and final run\n");
1387 td->latency_end_run = 1;
1388 reset_all_stats(td);
1397 * Check if we can bump the queue depth
1399 void lat_target_check(struct thread_data *td)
1401 uint64_t usec_window;
1405 usec_window = utime_since_now(&td->latency_ts);
1406 if (usec_window < td->o.latency_window)
1409 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1410 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1411 success_ios *= 100.0;
1413 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1415 if (success_ios >= td->o.latency_percentile.u.f)
1416 lat_target_success(td);
1418 __lat_target_failed(td);
1422 * If latency target is enabled, we might be ramping up or down and not
1423 * using the full queue depth available.
1425 bool queue_full(const struct thread_data *td)
1427 const int qempty = io_u_qempty(&td->io_u_freelist);
1431 if (!td->o.latency_target)
1434 return td->cur_depth >= td->latency_qd;
1437 struct io_u *__get_io_u(struct thread_data *td)
1439 struct io_u *io_u = NULL;
1447 if (!io_u_rempty(&td->io_u_requeues))
1448 io_u = io_u_rpop(&td->io_u_requeues);
1449 else if (!queue_full(td)) {
1450 io_u = io_u_qpop(&td->io_u_freelist);
1455 io_u->end_io = NULL;
1459 assert(io_u->flags & IO_U_F_FREE);
1460 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1461 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1465 io_u->acct_ddir = -1;
1467 assert(!(td->flags & TD_F_CHILD));
1468 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1470 } else if (td_async_processing(td)) {
1472 * We ran out, wait for async verify threads to finish and
1475 assert(!(td->flags & TD_F_CHILD));
1476 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1484 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1486 if (!(td->flags & TD_F_TRIM_BACKLOG))
1489 if (td->trim_entries) {
1492 if (td->trim_batch) {
1495 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1496 td->last_ddir != DDIR_READ) {
1497 td->trim_batch = td->o.trim_batch;
1498 if (!td->trim_batch)
1499 td->trim_batch = td->o.trim_backlog;
1503 if (get_trim && !get_next_trim(td, io_u))
1510 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1512 if (!(td->flags & TD_F_VER_BACKLOG))
1515 if (td->io_hist_len) {
1518 if (td->verify_batch)
1520 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1521 td->last_ddir != DDIR_READ) {
1522 td->verify_batch = td->o.verify_batch;
1523 if (!td->verify_batch)
1524 td->verify_batch = td->o.verify_backlog;
1528 if (get_verify && !get_next_verify(td, io_u)) {
1538 * Fill offset and start time into the buffer content, to prevent too
1539 * easy compressible data for simple de-dupe attempts. Do this for every
1540 * 512b block in the range, since that should be the smallest block size
1541 * we can expect from a device.
1543 static void small_content_scramble(struct io_u *io_u)
1545 unsigned int i, nr_blocks = io_u->buflen / 512;
1547 unsigned int offset;
1554 boffset = io_u->offset;
1555 io_u->buf_filled_len = 0;
1557 for (i = 0; i < nr_blocks; i++) {
1559 * Fill the byte offset into a "random" start offset of
1560 * the buffer, given by the product of the usec time
1561 * and the actual offset.
1563 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1564 offset &= ~(sizeof(uint64_t) - 1);
1565 if (offset >= 512 - sizeof(uint64_t))
1566 offset -= sizeof(uint64_t);
1567 memcpy(p + offset, &boffset, sizeof(boffset));
1569 end = p + 512 - sizeof(io_u->start_time);
1570 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1577 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1578 * etc. The returned io_u is fully ready to be prepped and submitted.
1580 struct io_u *get_io_u(struct thread_data *td)
1584 int do_scramble = 0;
1587 io_u = __get_io_u(td);
1589 dprint(FD_IO, "__get_io_u failed\n");
1593 if (check_get_verify(td, io_u))
1595 if (check_get_trim(td, io_u))
1599 * from a requeue, io_u already setup
1605 * If using an iolog, grab next piece if any available.
1607 if (td->flags & TD_F_READ_IOLOG) {
1608 if (read_iolog_get(td, io_u))
1610 } else if (set_io_u_file(td, io_u)) {
1612 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1618 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1622 assert(fio_file_open(f));
1624 if (ddir_rw(io_u->ddir)) {
1625 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1626 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1630 f->last_start[io_u->ddir] = io_u->offset;
1631 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1633 if (io_u->ddir == DDIR_WRITE) {
1634 if (td->flags & TD_F_REFILL_BUFFERS) {
1635 io_u_fill_buffer(td, io_u,
1636 td->o.min_bs[DDIR_WRITE],
1638 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1639 !(td->flags & TD_F_COMPRESS))
1641 if (td->flags & TD_F_VER_NONE) {
1642 populate_verify_io_u(td, io_u);
1645 } else if (io_u->ddir == DDIR_READ) {
1647 * Reset the buf_filled parameters so next time if the
1648 * buffer is used for writes it is refilled.
1650 io_u->buf_filled_len = 0;
1655 * Set io data pointers.
1657 io_u->xfer_buf = io_u->buf;
1658 io_u->xfer_buflen = io_u->buflen;
1662 if (!td_io_prep(td, io_u)) {
1663 if (!td->o.disable_lat)
1664 fio_gettime(&io_u->start_time, NULL);
1666 small_content_scramble(io_u);
1670 dprint(FD_IO, "get_io_u failed\n");
1672 return ERR_PTR(ret);
1675 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1677 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1679 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1682 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1683 io_u->file ? " on file " : "",
1684 io_u->file ? io_u->file->file_name : "",
1685 strerror(io_u->error),
1686 io_ddir_name(io_u->ddir),
1687 io_u->offset, io_u->xfer_buflen);
1689 if (td->io_ops->errdetails) {
1690 char *err = td->io_ops->errdetails(io_u);
1692 log_err("fio: %s\n", err);
1697 td_verror(td, io_u->error, "io_u error");
1700 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1702 __io_u_log_error(td, io_u);
1704 __io_u_log_error(td->parent, io_u);
1707 static inline bool gtod_reduce(struct thread_data *td)
1709 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1710 || td->o.gtod_reduce;
1713 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1714 struct io_completion_data *icd,
1715 const enum fio_ddir idx, unsigned int bytes)
1717 const int no_reduce = !gtod_reduce(td);
1718 unsigned long lusec = 0;
1724 lusec = utime_since(&io_u->issue_time, &icd->time);
1726 if (!td->o.disable_lat) {
1727 unsigned long tusec;
1729 tusec = utime_since(&io_u->start_time, &icd->time);
1730 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1732 if (td->flags & TD_F_PROFILE_OPS) {
1733 struct prof_io_ops *ops = &td->prof_io_ops;
1736 icd->error = ops->io_u_lat(td, tusec);
1739 if (td->o.max_latency && tusec > td->o.max_latency)
1740 lat_fatal(td, icd, tusec, td->o.max_latency);
1741 if (td->o.latency_target && tusec > td->o.latency_target) {
1742 if (lat_target_failed(td))
1743 lat_fatal(td, icd, tusec, td->o.latency_target);
1748 if (!td->o.disable_clat) {
1749 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1750 io_u_mark_latency(td, lusec);
1753 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1754 add_bw_sample(td, io_u, bytes, lusec);
1756 if (no_reduce && per_unit_log(td->iops_log))
1757 add_iops_sample(td, io_u, bytes);
1760 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1761 uint32_t *info = io_u_block_info(td, io_u);
1762 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1763 if (io_u->ddir == DDIR_TRIM) {
1764 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1765 BLOCK_INFO_TRIMS(*info) + 1);
1766 } else if (io_u->ddir == DDIR_WRITE) {
1767 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1774 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1775 uint64_t offset, unsigned int bytes)
1782 if (f->first_write == -1ULL || offset < f->first_write)
1783 f->first_write = offset;
1784 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1785 f->last_write = offset + bytes;
1787 if (!f->last_write_comp)
1790 idx = f->last_write_idx++;
1791 f->last_write_comp[idx] = offset;
1792 if (f->last_write_idx == td->o.iodepth)
1793 f->last_write_idx = 0;
1796 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1797 struct io_completion_data *icd)
1799 struct io_u *io_u = *io_u_ptr;
1800 enum fio_ddir ddir = io_u->ddir;
1801 struct fio_file *f = io_u->file;
1803 dprint_io_u(io_u, "io complete");
1805 assert(io_u->flags & IO_U_F_FLIGHT);
1806 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1809 * Mark IO ok to verify
1813 * Remove errored entry from the verification list
1816 unlog_io_piece(td, io_u);
1818 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1823 if (ddir_sync(ddir)) {
1824 td->last_was_sync = 1;
1826 f->first_write = -1ULL;
1827 f->last_write = -1ULL;
1832 td->last_was_sync = 0;
1833 td->last_ddir = ddir;
1835 if (!io_u->error && ddir_rw(ddir)) {
1836 unsigned int bytes = io_u->buflen - io_u->resid;
1839 td->io_blocks[ddir]++;
1840 td->this_io_blocks[ddir]++;
1841 td->io_bytes[ddir] += bytes;
1843 if (!(io_u->flags & IO_U_F_VER_LIST))
1844 td->this_io_bytes[ddir] += bytes;
1846 if (ddir == DDIR_WRITE)
1847 file_log_write_comp(td, f, io_u->offset, bytes);
1849 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1850 td->runstate == TD_VERIFYING))
1851 account_io_completion(td, io_u, icd, ddir, bytes);
1853 icd->bytes_done[ddir] += bytes;
1856 ret = io_u->end_io(td, io_u_ptr);
1858 if (ret && !icd->error)
1861 } else if (io_u->error) {
1862 icd->error = io_u->error;
1863 io_u_log_error(td, io_u);
1866 enum error_type_bit eb = td_error_type(ddir, icd->error);
1868 if (!td_non_fatal_error(td, eb, icd->error))
1872 * If there is a non_fatal error, then add to the error count
1873 * and clear all the errors.
1875 update_error_count(td, icd->error);
1883 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1888 if (!gtod_reduce(td))
1889 fio_gettime(&icd->time, NULL);
1894 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1895 icd->bytes_done[ddir] = 0;
1898 static void ios_completed(struct thread_data *td,
1899 struct io_completion_data *icd)
1904 for (i = 0; i < icd->nr; i++) {
1905 io_u = td->io_ops->event(td, i);
1907 io_completed(td, &io_u, icd);
1915 * Complete a single io_u for the sync engines.
1917 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1919 struct io_completion_data icd;
1922 init_icd(td, &icd, 1);
1923 io_completed(td, &io_u, &icd);
1929 td_verror(td, icd.error, "io_u_sync_complete");
1933 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1934 td->bytes_done[ddir] += icd.bytes_done[ddir];
1940 * Called to complete min_events number of io for the async engines.
1942 int io_u_queued_complete(struct thread_data *td, int min_evts)
1944 struct io_completion_data icd;
1945 struct timespec *tvp = NULL;
1947 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1949 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1953 else if (min_evts > td->cur_depth)
1954 min_evts = td->cur_depth;
1956 /* No worries, td_io_getevents fixes min and max if they are
1957 * set incorrectly */
1958 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1960 td_verror(td, -ret, "td_io_getevents");
1965 init_icd(td, &icd, ret);
1966 ios_completed(td, &icd);
1968 td_verror(td, icd.error, "io_u_queued_complete");
1972 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1973 td->bytes_done[ddir] += icd.bytes_done[ddir];
1979 * Call when io_u is really queued, to update the submission latency.
1981 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1983 if (!td->o.disable_slat) {
1984 unsigned long slat_time;
1986 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1991 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1997 * See if we should reuse the last seed, if dedupe is enabled
1999 static struct frand_state *get_buf_state(struct thread_data *td)
2003 if (!td->o.dedupe_percentage)
2004 return &td->buf_state;
2005 else if (td->o.dedupe_percentage == 100) {
2006 frand_copy(&td->buf_state_prev, &td->buf_state);
2007 return &td->buf_state;
2010 v = rand32_between(&td->dedupe_state, 1, 100);
2012 if (v <= td->o.dedupe_percentage)
2013 return &td->buf_state_prev;
2015 return &td->buf_state;
2018 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2020 if (td->o.dedupe_percentage == 100)
2021 frand_copy(rs, &td->buf_state_prev);
2022 else if (rs == &td->buf_state)
2023 frand_copy(&td->buf_state_prev, rs);
2026 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2027 unsigned int max_bs)
2029 struct thread_options *o = &td->o;
2031 if (o->compress_percentage || o->dedupe_percentage) {
2032 unsigned int perc = td->o.compress_percentage;
2033 struct frand_state *rs;
2034 unsigned int left = max_bs;
2035 unsigned int this_write;
2038 rs = get_buf_state(td);
2040 min_write = min(min_write, left);
2043 this_write = min_not_zero(min_write,
2044 td->o.compress_chunk);
2046 fill_random_buf_percentage(rs, buf, perc,
2047 this_write, this_write,
2049 o->buffer_pattern_bytes);
2051 fill_random_buf(rs, buf, min_write);
2052 this_write = min_write;
2057 save_buf_state(td, rs);
2059 } else if (o->buffer_pattern_bytes)
2060 fill_buffer_pattern(td, buf, max_bs);
2061 else if (o->zero_buffers)
2062 memset(buf, 0, max_bs);
2064 fill_random_buf(get_buf_state(td), buf, max_bs);
2068 * "randomly" fill the buffer contents
2070 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2071 unsigned int min_write, unsigned int max_bs)
2073 io_u->buf_filled_len = 0;
2074 fill_io_buffer(td, io_u->buf, min_write, max_bs);