13 #include "lib/axmap.h"
15 struct io_completion_data {
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timeval time; /* output */
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 static int random_map_free(struct fio_file *f, const uint64_t block)
29 return !axmap_isset(f->io_axmap, block);
33 * Mark a given offset as used in the map.
35 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37 unsigned int min_bs = td->o.rw_min_bs;
38 struct fio_file *f = io_u->file;
39 unsigned int nr_blocks;
42 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
43 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45 if (!(io_u->flags & IO_U_F_BUSY_OK))
46 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48 if ((nr_blocks * min_bs) < io_u->buflen)
49 io_u->buflen = nr_blocks * min_bs;
52 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
58 assert(ddir_rw(ddir));
61 * Hmm, should we make sure that ->io_size <= ->real_file_size?
63 max_size = f->io_size;
64 if (max_size > f->real_file_size)
65 max_size = f->real_file_size;
68 max_size = td->o.zone_range;
70 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
78 struct flist_head list;
82 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
83 enum fio_ddir ddir, uint64_t *b)
87 lastb = last_block(td, f, ddir);
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
94 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
96 if (td->o.use_os_rand) {
98 r = os_random_long(&td->random_state);
101 r = __rand(&td->__random_state);
104 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
106 *b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
110 if (lfsr_next(&f->lfsr, &off, lastb))
117 * if we are not maintaining a random map, we are done.
119 if (!file_randommap(td, f))
123 * calculate map offset and check if it's free
125 if (random_map_free(f, *b))
128 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
129 (unsigned long long) *b);
131 *b = axmap_next_free(f->io_axmap, *b);
132 if (*b == (uint64_t) -1ULL)
138 static int __get_next_rand_offset_zipf(struct thread_data *td,
139 struct fio_file *f, enum fio_ddir ddir,
142 *b = zipf_next(&f->zipf);
146 static int __get_next_rand_offset_pareto(struct thread_data *td,
147 struct fio_file *f, enum fio_ddir ddir,
150 *b = pareto_next(&f->zipf);
154 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
156 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
157 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
159 return r1->off - r2->off;
162 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
163 enum fio_ddir ddir, uint64_t *b)
165 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
166 return __get_next_rand_offset(td, f, ddir, b);
167 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
168 return __get_next_rand_offset_zipf(td, f, ddir, b);
169 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
170 return __get_next_rand_offset_pareto(td, f, ddir, b);
172 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
177 * Sort the reads for a verify phase in batches of verifysort_nr, if
180 static inline int should_sort_io(struct thread_data *td)
182 if (!td->o.verifysort_nr || !td->o.do_verify)
186 if (td->runstate != TD_VERIFYING)
188 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
194 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
199 if (td->o.perc_rand[ddir] == 100)
202 if (td->o.use_os_rand) {
203 r = os_random_long(&td->seq_rand_state[ddir]);
204 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
206 r = __rand(&td->__seq_rand_state[ddir]);
207 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
210 return v <= td->o.perc_rand[ddir];
213 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
214 enum fio_ddir ddir, uint64_t *b)
219 if (!should_sort_io(td))
220 return get_off_from_method(td, f, ddir, b);
222 if (!flist_empty(&td->next_rand_list)) {
225 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
232 for (i = 0; i < td->o.verifysort_nr; i++) {
233 r = malloc(sizeof(*r));
235 ret = get_off_from_method(td, f, ddir, &r->off);
241 flist_add(&r->list, &td->next_rand_list);
247 assert(!flist_empty(&td->next_rand_list));
248 flist_sort(NULL, &td->next_rand_list, flist_cmp);
252 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
253 enum fio_ddir ddir, uint64_t *b)
255 if (!get_next_rand_offset(td, f, ddir, b))
258 if (td->o.time_based) {
259 fio_file_reset(td, f);
260 if (!get_next_rand_offset(td, f, ddir, b))
264 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
265 f->file_name, (unsigned long long) f->last_pos,
266 (unsigned long long) f->real_file_size);
270 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
271 enum fio_ddir ddir, uint64_t *offset)
273 assert(ddir_rw(ddir));
275 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
276 f->last_pos = f->last_pos - f->io_size;
278 if (f->last_pos < f->real_file_size) {
281 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
282 f->last_pos = f->real_file_size;
284 pos = f->last_pos - f->file_offset;
286 pos += td->o.ddir_seq_add;
295 static int get_next_block(struct thread_data *td, struct io_u *io_u,
296 enum fio_ddir ddir, int rw_seq,
297 unsigned int *is_random)
299 struct fio_file *f = io_u->file;
303 assert(ddir_rw(ddir));
309 if (should_do_random(td, ddir)) {
310 ret = get_next_rand_block(td, f, ddir, &b);
314 io_u->flags |= IO_U_F_BUSY_OK;
315 ret = get_next_seq_offset(td, f, ddir, &offset);
317 ret = get_next_rand_block(td, f, ddir, &b);
321 ret = get_next_seq_offset(td, f, ddir, &offset);
324 io_u->flags |= IO_U_F_BUSY_OK;
327 if (td->o.rw_seq == RW_SEQ_SEQ) {
328 ret = get_next_seq_offset(td, f, ddir, &offset);
330 ret = get_next_rand_block(td, f, ddir, &b);
333 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
334 if (f->last_start != -1ULL)
335 offset = f->last_start - f->file_offset;
340 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
347 io_u->offset = offset;
349 io_u->offset = b * td->o.ba[ddir];
351 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
360 * For random io, generate a random new block and see if it's used. Repeat
361 * until we find a free one. For sequential io, just return the end of
362 * the last io issued.
364 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
365 unsigned int *is_random)
367 struct fio_file *f = io_u->file;
368 enum fio_ddir ddir = io_u->ddir;
371 assert(ddir_rw(ddir));
373 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
375 td->ddir_seq_nr = td->o.ddir_seq_nr;
378 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
381 if (io_u->offset >= f->io_size) {
382 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
383 (unsigned long long) io_u->offset,
384 (unsigned long long) f->io_size);
388 io_u->offset += f->file_offset;
389 if (io_u->offset >= f->real_file_size) {
390 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
391 (unsigned long long) io_u->offset,
392 (unsigned long long) f->real_file_size);
399 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
400 unsigned int *is_random)
402 if (td->flags & TD_F_PROFILE_OPS) {
403 struct prof_io_ops *ops = &td->prof_io_ops;
405 if (ops->fill_io_u_off)
406 return ops->fill_io_u_off(td, io_u, is_random);
409 return __get_next_offset(td, io_u, is_random);
412 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
415 struct fio_file *f = io_u->file;
417 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
420 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
421 unsigned int is_random)
423 int ddir = io_u->ddir;
424 unsigned int buflen = 0;
425 unsigned int minbs, maxbs;
426 unsigned long r, rand_max;
428 assert(ddir_rw(io_u->ddir));
430 if (td->o.bs_is_seq_rand)
431 ddir = is_random ? DDIR_WRITE: DDIR_READ;
435 minbs = td->o.min_bs[ddir];
436 maxbs = td->o.max_bs[ddir];
442 * If we can't satisfy the min block size from here, then fail
444 if (!io_u_fits(td, io_u, minbs))
447 if (td->o.use_os_rand)
448 rand_max = OS_RAND_MAX;
450 rand_max = FRAND_MAX;
453 if (td->o.use_os_rand)
454 r = os_random_long(&td->bsrange_state);
456 r = __rand(&td->__bsrange_state);
458 if (!td->o.bssplit_nr[ddir]) {
459 buflen = 1 + (unsigned int) ((double) maxbs *
460 (r / (rand_max + 1.0)));
467 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
468 struct bssplit *bsp = &td->o.bssplit[ddir][i];
472 if ((r <= ((rand_max / 100L) * perc)) &&
473 io_u_fits(td, io_u, buflen))
478 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
479 buflen = (buflen + td->o.verify_interval - 1) &
480 ~(td->o.verify_interval - 1);
482 if (!td->o.bs_unaligned && is_power_of_2(minbs))
483 buflen = (buflen + minbs - 1) & ~(minbs - 1);
485 } while (!io_u_fits(td, io_u, buflen));
490 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
491 unsigned int is_random)
493 if (td->flags & TD_F_PROFILE_OPS) {
494 struct prof_io_ops *ops = &td->prof_io_ops;
496 if (ops->fill_io_u_size)
497 return ops->fill_io_u_size(td, io_u, is_random);
500 return __get_next_buflen(td, io_u, is_random);
503 static void set_rwmix_bytes(struct thread_data *td)
508 * we do time or byte based switch. this is needed because
509 * buffered writes may issue a lot quicker than they complete,
510 * whereas reads do not.
512 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
513 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
516 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
521 if (td->o.use_os_rand) {
522 r = os_random_long(&td->rwmix_state);
523 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
525 r = __rand(&td->__rwmix_state);
526 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
529 if (v <= td->o.rwmix[DDIR_READ])
535 void io_u_quiesce(struct thread_data *td)
538 * We are going to sleep, ensure that we flush anything pending as
539 * not to skew our latency numbers.
541 * Changed to only monitor 'in flight' requests here instead of the
542 * td->cur_depth, b/c td->cur_depth does not accurately represent
543 * io's that have been actually submitted to an async engine,
544 * and cur_depth is meaningless for sync engines.
546 while (td->io_u_in_flight) {
549 ret = io_u_queued_complete(td, 1, NULL);
553 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
555 enum fio_ddir odir = ddir ^ 1;
559 assert(ddir_rw(ddir));
561 if (td->rate_pending_usleep[ddir] <= 0)
565 * We have too much pending sleep in this direction. See if we
568 if (td_rw(td) && td->o.rwmix[odir]) {
570 * Other direction does not have too much pending, switch
572 if (td->rate_pending_usleep[odir] < 100000)
576 * Both directions have pending sleep. Sleep the minimum time
577 * and deduct from both.
579 if (td->rate_pending_usleep[ddir] <=
580 td->rate_pending_usleep[odir]) {
581 usec = td->rate_pending_usleep[ddir];
583 usec = td->rate_pending_usleep[odir];
587 usec = td->rate_pending_usleep[ddir];
591 fio_gettime(&t, NULL);
592 usec_sleep(td, usec);
593 usec = utime_since_now(&t);
595 td->rate_pending_usleep[ddir] -= usec;
598 if (td_rw(td) && __should_check_rate(td, odir))
599 td->rate_pending_usleep[odir] -= usec;
608 * Return the data direction for the next io_u. If the job is a
609 * mixed read/write workload, check the rwmix cycle and switch if
612 static enum fio_ddir get_rw_ddir(struct thread_data *td)
617 * see if it's time to fsync
619 if (td->o.fsync_blocks &&
620 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
621 td->io_issues[DDIR_WRITE] && should_fsync(td))
625 * see if it's time to fdatasync
627 if (td->o.fdatasync_blocks &&
628 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
629 td->io_issues[DDIR_WRITE] && should_fsync(td))
630 return DDIR_DATASYNC;
633 * see if it's time to sync_file_range
635 if (td->sync_file_range_nr &&
636 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
637 td->io_issues[DDIR_WRITE] && should_fsync(td))
638 return DDIR_SYNC_FILE_RANGE;
642 * Check if it's time to seed a new data direction.
644 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
646 * Put a top limit on how many bytes we do for
647 * one data direction, to avoid overflowing the
650 ddir = get_rand_ddir(td);
652 if (ddir != td->rwmix_ddir)
655 td->rwmix_ddir = ddir;
657 ddir = td->rwmix_ddir;
658 } else if (td_read(td))
660 else if (td_write(td))
665 td->rwmix_ddir = rate_ddir(td, ddir);
666 return td->rwmix_ddir;
669 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
671 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
673 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
674 td->o.barrier_blocks &&
675 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
676 td->io_issues[DDIR_WRITE])
677 io_u->flags |= IO_U_F_BARRIER;
680 void put_file_log(struct thread_data *td, struct fio_file *f)
682 int ret = put_file(td, f);
685 td_verror(td, ret, "file close");
688 void put_io_u(struct thread_data *td, struct io_u *io_u)
692 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
693 put_file_log(td, io_u->file);
695 io_u->flags &= ~IO_U_F_FREE_DEF;
696 io_u->flags |= IO_U_F_FREE;
698 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
700 io_u_qpush(&td->io_u_freelist, io_u);
702 td_io_u_free_notify(td);
705 void clear_io_u(struct thread_data *td, struct io_u *io_u)
707 io_u->flags &= ~IO_U_F_FLIGHT;
711 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
713 struct io_u *__io_u = *io_u;
714 enum fio_ddir ddir = acct_ddir(__io_u);
716 dprint(FD_IO, "requeue %p\n", __io_u);
720 __io_u->flags |= IO_U_F_FREE;
721 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
722 td->io_issues[ddir]--;
724 __io_u->flags &= ~IO_U_F_FLIGHT;
725 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
728 io_u_rpush(&td->io_u_requeues, __io_u);
733 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
735 unsigned int is_random;
737 if (td->io_ops->flags & FIO_NOIO)
740 set_rw_ddir(td, io_u);
743 * fsync() or fdatasync() or trim etc, we are done
745 if (!ddir_rw(io_u->ddir))
749 * See if it's time to switch to a new zone
751 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
753 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
754 io_u->file->last_pos = io_u->file->file_offset;
755 td->io_skip_bytes += td->o.zone_skip;
759 * No log, let the seq/rand engine retrieve the next buflen and
762 if (get_next_offset(td, io_u, &is_random)) {
763 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
767 io_u->buflen = get_next_buflen(td, io_u, is_random);
769 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
773 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
774 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
775 dprint(FD_IO, " off=%llu/%lu > %llu\n",
776 (unsigned long long) io_u->offset, io_u->buflen,
777 (unsigned long long) io_u->file->real_file_size);
782 * mark entry before potentially trimming io_u
784 if (td_random(td) && file_randommap(td, io_u->file))
785 mark_random_map(td, io_u);
788 dprint_io_u(io_u, "fill_io_u");
789 td->zone_bytes += io_u->buflen;
793 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
822 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
824 __io_u_mark_map(td->ts.io_u_submit, nr);
825 td->ts.total_submit++;
828 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
830 __io_u_mark_map(td->ts.io_u_complete, nr);
831 td->ts.total_complete++;
834 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
838 switch (td->cur_depth) {
860 td->ts.io_u_map[idx] += nr;
863 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
900 assert(idx < FIO_IO_U_LAT_U_NR);
901 td->ts.io_u_lat_u[idx]++;
904 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
945 assert(idx < FIO_IO_U_LAT_M_NR);
946 td->ts.io_u_lat_m[idx]++;
949 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
952 io_u_mark_lat_usec(td, usec);
954 io_u_mark_lat_msec(td, usec / 1000);
958 * Get next file to service by choosing one at random
960 static struct fio_file *get_next_file_rand(struct thread_data *td,
961 enum fio_file_flags goodf,
962 enum fio_file_flags badf)
971 if (td->o.use_os_rand) {
972 r = os_random_long(&td->next_file_state);
973 fno = (unsigned int) ((double) td->o.nr_files
974 * (r / (OS_RAND_MAX + 1.0)));
976 r = __rand(&td->__next_file_state);
977 fno = (unsigned int) ((double) td->o.nr_files
978 * (r / (FRAND_MAX + 1.0)));
982 if (fio_file_done(f))
985 if (!fio_file_open(f)) {
988 err = td_io_open_file(td, f);
994 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
995 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
999 td_io_close_file(td, f);
1004 * Get next file to service by doing round robin between all available ones
1006 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1009 unsigned int old_next_file = td->next_file;
1015 f = td->files[td->next_file];
1018 if (td->next_file >= td->o.nr_files)
1021 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1022 if (fio_file_done(f)) {
1027 if (!fio_file_open(f)) {
1030 err = td_io_open_file(td, f);
1032 dprint(FD_FILE, "error %d on open of %s\n",
1040 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1042 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1046 td_io_close_file(td, f);
1049 } while (td->next_file != old_next_file);
1051 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1055 static struct fio_file *__get_next_file(struct thread_data *td)
1059 assert(td->o.nr_files <= td->files_index);
1061 if (td->nr_done_files >= td->o.nr_files) {
1062 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1063 " nr_files=%d\n", td->nr_open_files,
1069 f = td->file_service_file;
1070 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1071 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1073 if (td->file_service_left--)
1077 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1078 td->o.file_service_type == FIO_FSERVICE_SEQ)
1079 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1081 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1083 td->file_service_file = f;
1084 td->file_service_left = td->file_service_nr - 1;
1086 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1090 static struct fio_file *get_next_file(struct thread_data *td)
1092 if (!(td->flags & TD_F_PROFILE_OPS)) {
1093 struct prof_io_ops *ops = &td->prof_io_ops;
1095 if (ops->get_next_file)
1096 return ops->get_next_file(td);
1099 return __get_next_file(td);
1102 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1107 f = get_next_file(td);
1114 if (!fill_io_u(td, io_u))
1117 put_file_log(td, f);
1118 td_io_close_file(td, f);
1120 fio_file_set_done(f);
1121 td->nr_done_files++;
1122 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1123 td->nr_done_files, td->o.nr_files);
1129 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1130 unsigned long tusec, unsigned long max_usec)
1133 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1134 td_verror(td, ETIMEDOUT, "max latency exceeded");
1135 icd->error = ETIMEDOUT;
1138 static void lat_new_cycle(struct thread_data *td)
1140 fio_gettime(&td->latency_ts, NULL);
1141 td->latency_ios = ddir_rw_sum(td->io_blocks);
1142 td->latency_failed = 0;
1146 * We had an IO outside the latency target. Reduce the queue depth. If we
1147 * are at QD=1, then it's time to give up.
1149 static int __lat_target_failed(struct thread_data *td)
1151 if (td->latency_qd == 1)
1154 td->latency_qd_high = td->latency_qd;
1155 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1157 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1160 * When we ramp QD down, quiesce existing IO to prevent
1161 * a storm of ramp downs due to pending higher depth.
1168 static int lat_target_failed(struct thread_data *td)
1170 if (td->o.latency_percentile.u.f == 100.0)
1171 return __lat_target_failed(td);
1173 td->latency_failed++;
1177 void lat_target_init(struct thread_data *td)
1179 if (td->o.latency_target) {
1180 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1181 fio_gettime(&td->latency_ts, NULL);
1183 td->latency_qd_high = td->o.iodepth;
1184 td->latency_qd_low = 1;
1185 td->latency_ios = ddir_rw_sum(td->io_blocks);
1187 td->latency_qd = td->o.iodepth;
1190 static void lat_target_success(struct thread_data *td)
1192 const unsigned int qd = td->latency_qd;
1194 td->latency_qd_low = td->latency_qd;
1197 * If we haven't failed yet, we double up to a failing value instead
1198 * of bisecting from highest possible queue depth. If we have set
1199 * a limit other than td->o.iodepth, bisect between that.
1201 if (td->latency_qd_high != td->o.iodepth)
1202 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1204 td->latency_qd *= 2;
1206 if (td->latency_qd > td->o.iodepth)
1207 td->latency_qd = td->o.iodepth;
1209 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1211 * Same as last one, we are done
1213 if (td->latency_qd == qd)
1220 * Check if we can bump the queue depth
1222 void lat_target_check(struct thread_data *td)
1224 uint64_t usec_window;
1228 usec_window = utime_since_now(&td->latency_ts);
1229 if (usec_window < td->o.latency_window)
1232 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1233 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1234 success_ios *= 100.0;
1236 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1238 if (success_ios >= td->o.latency_percentile.u.f)
1239 lat_target_success(td);
1241 __lat_target_failed(td);
1245 * If latency target is enabled, we might be ramping up or down and not
1246 * using the full queue depth available.
1248 int queue_full(struct thread_data *td)
1250 const int qempty = io_u_qempty(&td->io_u_freelist);
1254 if (!td->o.latency_target)
1257 return td->cur_depth >= td->latency_qd;
1260 struct io_u *__get_io_u(struct thread_data *td)
1267 if (!io_u_rempty(&td->io_u_requeues))
1268 io_u = io_u_rpop(&td->io_u_requeues);
1269 else if (!queue_full(td)) {
1270 io_u = io_u_qpop(&td->io_u_freelist);
1275 io_u->end_io = NULL;
1279 assert(io_u->flags & IO_U_F_FREE);
1280 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1281 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1282 io_u->flags &= ~IO_U_F_VER_LIST;
1285 io_u->acct_ddir = -1;
1287 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1288 } else if (td->o.verify_async) {
1290 * We ran out, wait for async verify threads to finish and
1293 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1301 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1303 if (!(td->flags & TD_F_TRIM_BACKLOG))
1306 if (td->trim_entries) {
1309 if (td->trim_batch) {
1312 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1313 td->last_ddir != DDIR_READ) {
1314 td->trim_batch = td->o.trim_batch;
1315 if (!td->trim_batch)
1316 td->trim_batch = td->o.trim_backlog;
1320 if (get_trim && !get_next_trim(td, io_u))
1327 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1329 if (!(td->flags & TD_F_VER_BACKLOG))
1332 if (td->io_hist_len) {
1335 if (td->verify_batch)
1337 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1338 td->last_ddir != DDIR_READ) {
1339 td->verify_batch = td->o.verify_batch;
1340 if (!td->verify_batch)
1341 td->verify_batch = td->o.verify_backlog;
1345 if (get_verify && !get_next_verify(td, io_u)) {
1355 * Fill offset and start time into the buffer content, to prevent too
1356 * easy compressible data for simple de-dupe attempts. Do this for every
1357 * 512b block in the range, since that should be the smallest block size
1358 * we can expect from a device.
1360 static void small_content_scramble(struct io_u *io_u)
1362 unsigned int i, nr_blocks = io_u->buflen / 512;
1364 unsigned int offset;
1371 boffset = io_u->offset;
1372 io_u->buf_filled_len = 0;
1374 for (i = 0; i < nr_blocks; i++) {
1376 * Fill the byte offset into a "random" start offset of
1377 * the buffer, given by the product of the usec time
1378 * and the actual offset.
1380 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1381 offset &= ~(sizeof(uint64_t) - 1);
1382 if (offset >= 512 - sizeof(uint64_t))
1383 offset -= sizeof(uint64_t);
1384 memcpy(p + offset, &boffset, sizeof(boffset));
1386 end = p + 512 - sizeof(io_u->start_time);
1387 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1394 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1395 * etc. The returned io_u is fully ready to be prepped and submitted.
1397 struct io_u *get_io_u(struct thread_data *td)
1401 int do_scramble = 0;
1403 io_u = __get_io_u(td);
1405 dprint(FD_IO, "__get_io_u failed\n");
1409 if (check_get_verify(td, io_u))
1411 if (check_get_trim(td, io_u))
1415 * from a requeue, io_u already setup
1421 * If using an iolog, grab next piece if any available.
1423 if (td->flags & TD_F_READ_IOLOG) {
1424 if (read_iolog_get(td, io_u))
1426 } else if (set_io_u_file(td, io_u)) {
1427 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1432 assert(fio_file_open(f));
1434 if (ddir_rw(io_u->ddir)) {
1435 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1436 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1440 f->last_start = io_u->offset;
1441 f->last_pos = io_u->offset + io_u->buflen;
1443 if (io_u->ddir == DDIR_WRITE) {
1444 if (td->flags & TD_F_REFILL_BUFFERS) {
1445 io_u_fill_buffer(td, io_u,
1446 io_u->xfer_buflen, io_u->xfer_buflen);
1447 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1449 if (td->flags & TD_F_VER_NONE) {
1450 populate_verify_io_u(td, io_u);
1453 } else if (io_u->ddir == DDIR_READ) {
1455 * Reset the buf_filled parameters so next time if the
1456 * buffer is used for writes it is refilled.
1458 io_u->buf_filled_len = 0;
1463 * Set io data pointers.
1465 io_u->xfer_buf = io_u->buf;
1466 io_u->xfer_buflen = io_u->buflen;
1470 if (!td_io_prep(td, io_u)) {
1471 if (!td->o.disable_slat)
1472 fio_gettime(&io_u->start_time, NULL);
1474 small_content_scramble(io_u);
1478 dprint(FD_IO, "get_io_u failed\n");
1483 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1485 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1486 const char *msg[] = { "read", "write", "sync", "datasync",
1487 "sync_file_range", "wait", "trim" };
1489 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1492 log_err("fio: io_u error");
1495 log_err(" on file %s", io_u->file->file_name);
1497 log_err(": %s\n", strerror(io_u->error));
1499 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1500 io_u->offset, io_u->xfer_buflen);
1503 td_verror(td, io_u->error, "io_u error");
1506 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1507 struct io_completion_data *icd,
1508 const enum fio_ddir idx, unsigned int bytes)
1510 unsigned long lusec = 0;
1512 if (!td->o.disable_clat || !td->o.disable_bw)
1513 lusec = utime_since(&io_u->issue_time, &icd->time);
1515 if (!td->o.disable_lat) {
1516 unsigned long tusec;
1518 tusec = utime_since(&io_u->start_time, &icd->time);
1519 add_lat_sample(td, idx, tusec, bytes);
1521 if (td->flags & TD_F_PROFILE_OPS) {
1522 struct prof_io_ops *ops = &td->prof_io_ops;
1525 icd->error = ops->io_u_lat(td, tusec);
1528 if (td->o.max_latency && tusec > td->o.max_latency)
1529 lat_fatal(td, icd, tusec, td->o.max_latency);
1530 if (td->o.latency_target && tusec > td->o.latency_target) {
1531 if (lat_target_failed(td))
1532 lat_fatal(td, icd, tusec, td->o.latency_target);
1536 if (!td->o.disable_clat) {
1537 add_clat_sample(td, idx, lusec, bytes);
1538 io_u_mark_latency(td, lusec);
1541 if (!td->o.disable_bw)
1542 add_bw_sample(td, idx, bytes, &icd->time);
1544 add_iops_sample(td, idx, bytes, &icd->time);
1546 if (td->o.number_ios && !--td->o.number_ios)
1550 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1552 uint64_t secs, remainder, bps, bytes;
1554 bytes = td->this_io_bytes[ddir];
1555 bps = td->rate_bps[ddir];
1557 remainder = bytes % bps;
1558 return remainder * 1000000 / bps + secs * 1000000;
1561 static void io_completed(struct thread_data *td, struct io_u *io_u,
1562 struct io_completion_data *icd)
1566 dprint_io_u(io_u, "io complete");
1569 assert(io_u->flags & IO_U_F_FLIGHT);
1570 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1573 if (ddir_sync(io_u->ddir)) {
1574 td->last_was_sync = 1;
1577 f->first_write = -1ULL;
1578 f->last_write = -1ULL;
1583 td->last_was_sync = 0;
1584 td->last_ddir = io_u->ddir;
1586 if (!io_u->error && ddir_rw(io_u->ddir)) {
1587 unsigned int bytes = io_u->buflen - io_u->resid;
1588 const enum fio_ddir idx = io_u->ddir;
1589 const enum fio_ddir odx = io_u->ddir ^ 1;
1592 td->io_blocks[idx]++;
1593 td->this_io_blocks[idx]++;
1594 td->io_bytes[idx] += bytes;
1596 if (!(io_u->flags & IO_U_F_VER_LIST))
1597 td->this_io_bytes[idx] += bytes;
1599 if (idx == DDIR_WRITE) {
1602 if (f->first_write == -1ULL ||
1603 io_u->offset < f->first_write)
1604 f->first_write = io_u->offset;
1605 if (f->last_write == -1ULL ||
1606 ((io_u->offset + bytes) > f->last_write))
1607 f->last_write = io_u->offset + bytes;
1611 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1612 td->runstate == TD_VERIFYING)) {
1613 account_io_completion(td, io_u, icd, idx, bytes);
1615 if (__should_check_rate(td, idx)) {
1616 td->rate_pending_usleep[idx] =
1617 (usec_for_io(td, idx) -
1618 utime_since_now(&td->start));
1620 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1621 td->rate_pending_usleep[odx] =
1622 (usec_for_io(td, odx) -
1623 utime_since_now(&td->start));
1626 icd->bytes_done[idx] += bytes;
1629 ret = io_u->end_io(td, io_u);
1630 if (ret && !icd->error)
1633 } else if (io_u->error) {
1634 icd->error = io_u->error;
1635 io_u_log_error(td, io_u);
1638 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1639 if (!td_non_fatal_error(td, eb, icd->error))
1642 * If there is a non_fatal error, then add to the error count
1643 * and clear all the errors.
1645 update_error_count(td, icd->error);
1652 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1656 if (!td->o.disable_clat || !td->o.disable_bw)
1657 fio_gettime(&icd->time, NULL);
1662 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1663 icd->bytes_done[ddir] = 0;
1666 static void ios_completed(struct thread_data *td,
1667 struct io_completion_data *icd)
1672 for (i = 0; i < icd->nr; i++) {
1673 io_u = td->io_ops->event(td, i);
1675 io_completed(td, io_u, icd);
1677 if (!(io_u->flags & IO_U_F_FREE_DEF))
1683 * Complete a single io_u for the sync engines.
1685 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1688 struct io_completion_data icd;
1690 init_icd(td, &icd, 1);
1691 io_completed(td, io_u, &icd);
1693 if (!(io_u->flags & IO_U_F_FREE_DEF))
1697 td_verror(td, icd.error, "io_u_sync_complete");
1704 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1705 bytes[ddir] += icd.bytes_done[ddir];
1712 * Called to complete min_events number of io for the async engines.
1714 int io_u_queued_complete(struct thread_data *td, int min_evts,
1717 struct io_completion_data icd;
1718 struct timespec *tvp = NULL;
1720 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1722 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1727 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1729 td_verror(td, -ret, "td_io_getevents");
1734 init_icd(td, &icd, ret);
1735 ios_completed(td, &icd);
1737 td_verror(td, icd.error, "io_u_queued_complete");
1744 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1745 bytes[ddir] += icd.bytes_done[ddir];
1752 * Call when io_u is really queued, to update the submission latency.
1754 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1756 if (!td->o.disable_slat) {
1757 unsigned long slat_time;
1759 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1760 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1764 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1765 unsigned int max_bs)
1767 if (td->o.buffer_pattern_bytes)
1768 fill_buffer_pattern(td, buf, max_bs);
1769 else if (!td->o.zero_buffers) {
1770 unsigned int perc = td->o.compress_percentage;
1773 unsigned int seg = min_write;
1775 seg = min(min_write, td->o.compress_chunk);
1779 fill_random_buf_percentage(&td->buf_state, buf,
1782 fill_random_buf(&td->buf_state, buf, max_bs);
1784 memset(buf, 0, max_bs);
1788 * "randomly" fill the buffer contents
1790 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1791 unsigned int min_write, unsigned int max_bs)
1793 io_u->buf_filled_len = 0;
1794 fill_io_buffer(td, io_u->buf, min_write, max_bs);