13 #include "lib/axmap.h"
16 struct io_completion_data {
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 static int random_map_free(struct fio_file *f, const uint64_t block)
30 return !axmap_isset(f->io_axmap, block);
34 * Mark a given offset as used in the map.
36 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
53 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 assert(ddir_rw(ddir));
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
69 max_size = td->o.zone_range;
71 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
79 struct flist_head list;
83 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b)
88 lastb = last_block(td, f, ddir);
92 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
93 r = __rand(&td->random_state);
95 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
97 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
101 if (lfsr_next(&f->lfsr, &off, lastb))
108 * if we are not maintaining a random map, we are done.
110 if (!file_randommap(td, f))
114 * calculate map offset and check if it's free
116 if (random_map_free(f, *b))
119 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
120 (unsigned long long) *b);
122 *b = axmap_next_free(f->io_axmap, *b);
123 if (*b == (uint64_t) -1ULL)
129 static int __get_next_rand_offset_zipf(struct thread_data *td,
130 struct fio_file *f, enum fio_ddir ddir,
133 *b = zipf_next(&f->zipf);
137 static int __get_next_rand_offset_pareto(struct thread_data *td,
138 struct fio_file *f, enum fio_ddir ddir,
141 *b = pareto_next(&f->zipf);
145 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
147 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
148 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
150 return r1->off - r2->off;
153 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
154 enum fio_ddir ddir, uint64_t *b)
156 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
157 return __get_next_rand_offset(td, f, ddir, b);
158 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
159 return __get_next_rand_offset_zipf(td, f, ddir, b);
160 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
161 return __get_next_rand_offset_pareto(td, f, ddir, b);
163 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
168 * Sort the reads for a verify phase in batches of verifysort_nr, if
171 static inline int should_sort_io(struct thread_data *td)
173 if (!td->o.verifysort_nr || !td->o.do_verify)
177 if (td->runstate != TD_VERIFYING)
179 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
185 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
190 if (td->o.perc_rand[ddir] == 100)
193 r = __rand(&td->seq_rand_state[ddir]);
194 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
196 return v <= td->o.perc_rand[ddir];
199 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
200 enum fio_ddir ddir, uint64_t *b)
205 if (!should_sort_io(td))
206 return get_off_from_method(td, f, ddir, b);
208 if (!flist_empty(&td->next_rand_list)) {
210 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
217 for (i = 0; i < td->o.verifysort_nr; i++) {
218 r = malloc(sizeof(*r));
220 ret = get_off_from_method(td, f, ddir, &r->off);
226 flist_add(&r->list, &td->next_rand_list);
232 assert(!flist_empty(&td->next_rand_list));
233 flist_sort(NULL, &td->next_rand_list, flist_cmp);
237 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, uint64_t *b)
240 if (!get_next_rand_offset(td, f, ddir, b))
243 if (td->o.time_based) {
244 fio_file_reset(td, f);
245 if (!get_next_rand_offset(td, f, ddir, b))
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, (unsigned long long) f->last_pos,
251 (unsigned long long) f->real_file_size);
255 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
256 enum fio_ddir ddir, uint64_t *offset)
258 struct thread_options *o = &td->o;
260 assert(ddir_rw(ddir));
262 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
264 f->last_pos = f->last_pos - f->io_size;
266 if (f->last_pos < f->real_file_size) {
269 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
270 f->last_pos = f->real_file_size;
272 pos = f->last_pos - f->file_offset;
273 if (pos && o->ddir_seq_add) {
274 pos += o->ddir_seq_add;
277 * If we reach beyond the end of the file
278 * with holed IO, wrap around to the
281 if (pos >= f->real_file_size)
282 pos = f->file_offset;
292 static int get_next_block(struct thread_data *td, struct io_u *io_u,
293 enum fio_ddir ddir, int rw_seq,
294 unsigned int *is_random)
296 struct fio_file *f = io_u->file;
300 assert(ddir_rw(ddir));
306 if (should_do_random(td, ddir)) {
307 ret = get_next_rand_block(td, f, ddir, &b);
311 io_u->flags |= IO_U_F_BUSY_OK;
312 ret = get_next_seq_offset(td, f, ddir, &offset);
314 ret = get_next_rand_block(td, f, ddir, &b);
318 ret = get_next_seq_offset(td, f, ddir, &offset);
321 io_u->flags |= IO_U_F_BUSY_OK;
324 if (td->o.rw_seq == RW_SEQ_SEQ) {
325 ret = get_next_seq_offset(td, f, ddir, &offset);
327 ret = get_next_rand_block(td, f, ddir, &b);
330 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
331 if (f->last_start != -1ULL)
332 offset = f->last_start - f->file_offset;
337 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
344 io_u->offset = offset;
346 io_u->offset = b * td->o.ba[ddir];
348 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
357 * For random io, generate a random new block and see if it's used. Repeat
358 * until we find a free one. For sequential io, just return the end of
359 * the last io issued.
361 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
362 unsigned int *is_random)
364 struct fio_file *f = io_u->file;
365 enum fio_ddir ddir = io_u->ddir;
368 assert(ddir_rw(ddir));
370 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
372 td->ddir_seq_nr = td->o.ddir_seq_nr;
375 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
378 if (io_u->offset >= f->io_size) {
379 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
380 (unsigned long long) io_u->offset,
381 (unsigned long long) f->io_size);
385 io_u->offset += f->file_offset;
386 if (io_u->offset >= f->real_file_size) {
387 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
388 (unsigned long long) io_u->offset,
389 (unsigned long long) f->real_file_size);
396 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
397 unsigned int *is_random)
399 if (td->flags & TD_F_PROFILE_OPS) {
400 struct prof_io_ops *ops = &td->prof_io_ops;
402 if (ops->fill_io_u_off)
403 return ops->fill_io_u_off(td, io_u, is_random);
406 return __get_next_offset(td, io_u, is_random);
409 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
412 struct fio_file *f = io_u->file;
414 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
417 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
418 unsigned int is_random)
420 int ddir = io_u->ddir;
421 unsigned int buflen = 0;
422 unsigned int minbs, maxbs;
425 assert(ddir_rw(ddir));
427 if (td->o.bs_is_seq_rand)
428 ddir = is_random ? DDIR_WRITE: DDIR_READ;
430 minbs = td->o.min_bs[ddir];
431 maxbs = td->o.max_bs[ddir];
437 * If we can't satisfy the min block size from here, then fail
439 if (!io_u_fits(td, io_u, minbs))
443 r = __rand(&td->bsrange_state);
445 if (!td->o.bssplit_nr[ddir]) {
446 buflen = 1 + (unsigned int) ((double) maxbs *
447 (r / (FRAND_MAX + 1.0)));
454 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
455 struct bssplit *bsp = &td->o.bssplit[ddir][i];
459 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
460 io_u_fits(td, io_u, buflen))
465 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
466 buflen = (buflen + td->o.verify_interval - 1) &
467 ~(td->o.verify_interval - 1);
469 if (!td->o.bs_unaligned && is_power_of_2(minbs))
470 buflen = (buflen + minbs - 1) & ~(minbs - 1);
472 } while (!io_u_fits(td, io_u, buflen));
477 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
478 unsigned int is_random)
480 if (td->flags & TD_F_PROFILE_OPS) {
481 struct prof_io_ops *ops = &td->prof_io_ops;
483 if (ops->fill_io_u_size)
484 return ops->fill_io_u_size(td, io_u, is_random);
487 return __get_next_buflen(td, io_u, is_random);
490 static void set_rwmix_bytes(struct thread_data *td)
495 * we do time or byte based switch. this is needed because
496 * buffered writes may issue a lot quicker than they complete,
497 * whereas reads do not.
499 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
500 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
503 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
508 r = __rand(&td->rwmix_state);
509 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
511 if (v <= td->o.rwmix[DDIR_READ])
517 void io_u_quiesce(struct thread_data *td)
520 * We are going to sleep, ensure that we flush anything pending as
521 * not to skew our latency numbers.
523 * Changed to only monitor 'in flight' requests here instead of the
524 * td->cur_depth, b/c td->cur_depth does not accurately represent
525 * io's that have been actually submitted to an async engine,
526 * and cur_depth is meaningless for sync engines.
528 while (td->io_u_in_flight) {
531 ret = io_u_queued_complete(td, 1, NULL);
535 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
537 enum fio_ddir odir = ddir ^ 1;
541 assert(ddir_rw(ddir));
543 if (td->rate_pending_usleep[ddir] <= 0)
547 * We have too much pending sleep in this direction. See if we
550 if (td_rw(td) && td->o.rwmix[odir]) {
552 * Other direction does not have too much pending, switch
554 if (td->rate_pending_usleep[odir] < 100000)
558 * Both directions have pending sleep. Sleep the minimum time
559 * and deduct from both.
561 if (td->rate_pending_usleep[ddir] <=
562 td->rate_pending_usleep[odir]) {
563 usec = td->rate_pending_usleep[ddir];
565 usec = td->rate_pending_usleep[odir];
569 usec = td->rate_pending_usleep[ddir];
573 fio_gettime(&t, NULL);
574 usec_sleep(td, usec);
575 usec = utime_since_now(&t);
577 td->rate_pending_usleep[ddir] -= usec;
580 if (td_rw(td) && __should_check_rate(td, odir))
581 td->rate_pending_usleep[odir] -= usec;
583 if (ddir == DDIR_TRIM)
590 * Return the data direction for the next io_u. If the job is a
591 * mixed read/write workload, check the rwmix cycle and switch if
594 static enum fio_ddir get_rw_ddir(struct thread_data *td)
599 * see if it's time to fsync
601 if (td->o.fsync_blocks &&
602 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
603 td->io_issues[DDIR_WRITE] && should_fsync(td))
607 * see if it's time to fdatasync
609 if (td->o.fdatasync_blocks &&
610 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
611 td->io_issues[DDIR_WRITE] && should_fsync(td))
612 return DDIR_DATASYNC;
615 * see if it's time to sync_file_range
617 if (td->sync_file_range_nr &&
618 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
619 td->io_issues[DDIR_WRITE] && should_fsync(td))
620 return DDIR_SYNC_FILE_RANGE;
624 * Check if it's time to seed a new data direction.
626 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
628 * Put a top limit on how many bytes we do for
629 * one data direction, to avoid overflowing the
632 ddir = get_rand_ddir(td);
634 if (ddir != td->rwmix_ddir)
637 td->rwmix_ddir = ddir;
639 ddir = td->rwmix_ddir;
640 } else if (td_read(td))
642 else if (td_write(td))
647 td->rwmix_ddir = rate_ddir(td, ddir);
648 return td->rwmix_ddir;
651 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
653 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
655 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
656 td->o.barrier_blocks &&
657 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
658 td->io_issues[DDIR_WRITE])
659 io_u->flags |= IO_U_F_BARRIER;
662 void put_file_log(struct thread_data *td, struct fio_file *f)
664 unsigned int ret = put_file(td, f);
667 td_verror(td, ret, "file close");
670 void put_io_u(struct thread_data *td, struct io_u *io_u)
674 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
675 put_file_log(td, io_u->file);
678 io_u->flags |= IO_U_F_FREE;
680 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
682 io_u_qpush(&td->io_u_freelist, io_u);
684 td_io_u_free_notify(td);
687 void clear_io_u(struct thread_data *td, struct io_u *io_u)
689 io_u->flags &= ~IO_U_F_FLIGHT;
693 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
695 struct io_u *__io_u = *io_u;
696 enum fio_ddir ddir = acct_ddir(__io_u);
698 dprint(FD_IO, "requeue %p\n", __io_u);
702 __io_u->flags |= IO_U_F_FREE;
703 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
704 td->io_issues[ddir]--;
706 __io_u->flags &= ~IO_U_F_FLIGHT;
707 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
710 io_u_rpush(&td->io_u_requeues, __io_u);
715 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
717 unsigned int is_random;
719 if (td->io_ops->flags & FIO_NOIO)
722 set_rw_ddir(td, io_u);
725 * fsync() or fdatasync() or trim etc, we are done
727 if (!ddir_rw(io_u->ddir))
731 * See if it's time to switch to a new zone
733 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
734 struct fio_file *f = io_u->file;
737 f->file_offset += td->o.zone_range + td->o.zone_skip;
740 * Wrap from the beginning, if we exceed the file size
742 if (f->file_offset >= f->real_file_size)
743 f->file_offset = f->real_file_size - f->file_offset;
744 f->last_pos = f->file_offset;
745 td->io_skip_bytes += td->o.zone_skip;
749 * No log, let the seq/rand engine retrieve the next buflen and
752 if (get_next_offset(td, io_u, &is_random)) {
753 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
757 io_u->buflen = get_next_buflen(td, io_u, is_random);
759 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
763 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
764 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
765 dprint(FD_IO, " off=%llu/%lu > %llu\n",
766 (unsigned long long) io_u->offset, io_u->buflen,
767 (unsigned long long) io_u->file->real_file_size);
772 * mark entry before potentially trimming io_u
774 if (td_random(td) && file_randommap(td, io_u->file))
775 mark_random_map(td, io_u);
778 dprint_io_u(io_u, "fill_io_u");
779 td->zone_bytes += io_u->buflen;
783 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
812 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
814 __io_u_mark_map(td->ts.io_u_submit, nr);
815 td->ts.total_submit++;
818 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
820 __io_u_mark_map(td->ts.io_u_complete, nr);
821 td->ts.total_complete++;
824 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
828 switch (td->cur_depth) {
850 td->ts.io_u_map[idx] += nr;
853 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
890 assert(idx < FIO_IO_U_LAT_U_NR);
891 td->ts.io_u_lat_u[idx]++;
894 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
935 assert(idx < FIO_IO_U_LAT_M_NR);
936 td->ts.io_u_lat_m[idx]++;
939 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
942 io_u_mark_lat_usec(td, usec);
944 io_u_mark_lat_msec(td, usec / 1000);
948 * Get next file to service by choosing one at random
950 static struct fio_file *get_next_file_rand(struct thread_data *td,
951 enum fio_file_flags goodf,
952 enum fio_file_flags badf)
961 r = __rand(&td->next_file_state);
962 fno = (unsigned int) ((double) td->o.nr_files
963 * (r / (FRAND_MAX + 1.0)));
966 if (fio_file_done(f))
969 if (!fio_file_open(f)) {
972 if (td->nr_open_files >= td->o.open_files)
973 return ERR_PTR(-EBUSY);
975 err = td_io_open_file(td, f);
981 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
982 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
986 td_io_close_file(td, f);
991 * Get next file to service by doing round robin between all available ones
993 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
996 unsigned int old_next_file = td->next_file;
1002 f = td->files[td->next_file];
1005 if (td->next_file >= td->o.nr_files)
1008 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1009 if (fio_file_done(f)) {
1014 if (!fio_file_open(f)) {
1017 if (td->nr_open_files >= td->o.open_files)
1018 return ERR_PTR(-EBUSY);
1020 err = td_io_open_file(td, f);
1022 dprint(FD_FILE, "error %d on open of %s\n",
1030 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1032 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1036 td_io_close_file(td, f);
1039 } while (td->next_file != old_next_file);
1041 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1045 static struct fio_file *__get_next_file(struct thread_data *td)
1049 assert(td->o.nr_files <= td->files_index);
1051 if (td->nr_done_files >= td->o.nr_files) {
1052 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1053 " nr_files=%d\n", td->nr_open_files,
1059 f = td->file_service_file;
1060 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1061 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1063 if (td->file_service_left--)
1067 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1068 td->o.file_service_type == FIO_FSERVICE_SEQ)
1069 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1071 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1076 td->file_service_file = f;
1077 td->file_service_left = td->file_service_nr - 1;
1080 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1082 dprint(FD_FILE, "get_next_file: NULL\n");
1086 static struct fio_file *get_next_file(struct thread_data *td)
1088 if (td->flags & TD_F_PROFILE_OPS) {
1089 struct prof_io_ops *ops = &td->prof_io_ops;
1091 if (ops->get_next_file)
1092 return ops->get_next_file(td);
1095 return __get_next_file(td);
1098 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1103 f = get_next_file(td);
1104 if (IS_ERR_OR_NULL(f))
1110 if (!fill_io_u(td, io_u))
1113 put_file_log(td, f);
1114 td_io_close_file(td, f);
1116 fio_file_set_done(f);
1117 td->nr_done_files++;
1118 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1119 td->nr_done_files, td->o.nr_files);
1125 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1126 unsigned long tusec, unsigned long max_usec)
1129 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1130 td_verror(td, ETIMEDOUT, "max latency exceeded");
1131 icd->error = ETIMEDOUT;
1134 static void lat_new_cycle(struct thread_data *td)
1136 fio_gettime(&td->latency_ts, NULL);
1137 td->latency_ios = ddir_rw_sum(td->io_blocks);
1138 td->latency_failed = 0;
1142 * We had an IO outside the latency target. Reduce the queue depth. If we
1143 * are at QD=1, then it's time to give up.
1145 static int __lat_target_failed(struct thread_data *td)
1147 if (td->latency_qd == 1)
1150 td->latency_qd_high = td->latency_qd;
1152 if (td->latency_qd == td->latency_qd_low)
1153 td->latency_qd_low--;
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 td->latency_end_run = 0;
1181 if (td->o.latency_target) {
1182 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1183 fio_gettime(&td->latency_ts, NULL);
1185 td->latency_qd_high = td->o.iodepth;
1186 td->latency_qd_low = 1;
1187 td->latency_ios = ddir_rw_sum(td->io_blocks);
1189 td->latency_qd = td->o.iodepth;
1192 void lat_target_reset(struct thread_data *td)
1194 if (!td->latency_end_run)
1195 lat_target_init(td);
1198 static void lat_target_success(struct thread_data *td)
1200 const unsigned int qd = td->latency_qd;
1201 struct thread_options *o = &td->o;
1203 td->latency_qd_low = td->latency_qd;
1206 * If we haven't failed yet, we double up to a failing value instead
1207 * of bisecting from highest possible queue depth. If we have set
1208 * a limit other than td->o.iodepth, bisect between that.
1210 if (td->latency_qd_high != o->iodepth)
1211 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1213 td->latency_qd *= 2;
1215 if (td->latency_qd > o->iodepth)
1216 td->latency_qd = o->iodepth;
1218 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1221 * Same as last one, we are done. Let it run a latency cycle, so
1222 * we get only the results from the targeted depth.
1224 if (td->latency_qd == qd) {
1225 if (td->latency_end_run) {
1226 dprint(FD_RATE, "We are done\n");
1229 dprint(FD_RATE, "Quiesce and final run\n");
1231 td->latency_end_run = 1;
1232 reset_all_stats(td);
1241 * Check if we can bump the queue depth
1243 void lat_target_check(struct thread_data *td)
1245 uint64_t usec_window;
1249 usec_window = utime_since_now(&td->latency_ts);
1250 if (usec_window < td->o.latency_window)
1253 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1254 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1255 success_ios *= 100.0;
1257 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1259 if (success_ios >= td->o.latency_percentile.u.f)
1260 lat_target_success(td);
1262 __lat_target_failed(td);
1266 * If latency target is enabled, we might be ramping up or down and not
1267 * using the full queue depth available.
1269 int queue_full(const struct thread_data *td)
1271 const int qempty = io_u_qempty(&td->io_u_freelist);
1275 if (!td->o.latency_target)
1278 return td->cur_depth >= td->latency_qd;
1281 struct io_u *__get_io_u(struct thread_data *td)
1283 struct io_u *io_u = NULL;
1288 if (!io_u_rempty(&td->io_u_requeues))
1289 io_u = io_u_rpop(&td->io_u_requeues);
1290 else if (!queue_full(td)) {
1291 io_u = io_u_qpop(&td->io_u_freelist);
1296 io_u->end_io = NULL;
1300 assert(io_u->flags & IO_U_F_FREE);
1301 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1302 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1306 io_u->acct_ddir = -1;
1308 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1310 } else if (td->o.verify_async) {
1312 * We ran out, wait for async verify threads to finish and
1315 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1323 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1325 if (!(td->flags & TD_F_TRIM_BACKLOG))
1328 if (td->trim_entries) {
1331 if (td->trim_batch) {
1334 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1335 td->last_ddir != DDIR_READ) {
1336 td->trim_batch = td->o.trim_batch;
1337 if (!td->trim_batch)
1338 td->trim_batch = td->o.trim_backlog;
1342 if (get_trim && !get_next_trim(td, io_u))
1349 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1351 if (!(td->flags & TD_F_VER_BACKLOG))
1354 if (td->io_hist_len) {
1357 if (td->verify_batch)
1359 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1360 td->last_ddir != DDIR_READ) {
1361 td->verify_batch = td->o.verify_batch;
1362 if (!td->verify_batch)
1363 td->verify_batch = td->o.verify_backlog;
1367 if (get_verify && !get_next_verify(td, io_u)) {
1377 * Fill offset and start time into the buffer content, to prevent too
1378 * easy compressible data for simple de-dupe attempts. Do this for every
1379 * 512b block in the range, since that should be the smallest block size
1380 * we can expect from a device.
1382 static void small_content_scramble(struct io_u *io_u)
1384 unsigned int i, nr_blocks = io_u->buflen / 512;
1386 unsigned int offset;
1393 boffset = io_u->offset;
1394 io_u->buf_filled_len = 0;
1396 for (i = 0; i < nr_blocks; i++) {
1398 * Fill the byte offset into a "random" start offset of
1399 * the buffer, given by the product of the usec time
1400 * and the actual offset.
1402 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1403 offset &= ~(sizeof(uint64_t) - 1);
1404 if (offset >= 512 - sizeof(uint64_t))
1405 offset -= sizeof(uint64_t);
1406 memcpy(p + offset, &boffset, sizeof(boffset));
1408 end = p + 512 - sizeof(io_u->start_time);
1409 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1416 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1417 * etc. The returned io_u is fully ready to be prepped and submitted.
1419 struct io_u *get_io_u(struct thread_data *td)
1423 int do_scramble = 0;
1426 io_u = __get_io_u(td);
1428 dprint(FD_IO, "__get_io_u failed\n");
1432 if (check_get_verify(td, io_u))
1434 if (check_get_trim(td, io_u))
1438 * from a requeue, io_u already setup
1444 * If using an iolog, grab next piece if any available.
1446 if (td->flags & TD_F_READ_IOLOG) {
1447 if (read_iolog_get(td, io_u))
1449 } else if (set_io_u_file(td, io_u)) {
1451 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1457 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1461 assert(fio_file_open(f));
1463 if (ddir_rw(io_u->ddir)) {
1464 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1465 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1469 f->last_start = io_u->offset;
1470 f->last_pos = io_u->offset + io_u->buflen;
1472 if (io_u->ddir == DDIR_WRITE) {
1473 if (td->flags & TD_F_REFILL_BUFFERS) {
1474 io_u_fill_buffer(td, io_u,
1475 td->o.min_bs[DDIR_WRITE],
1477 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1478 !(td->flags & TD_F_COMPRESS))
1480 if (td->flags & TD_F_VER_NONE) {
1481 populate_verify_io_u(td, io_u);
1484 } else if (io_u->ddir == DDIR_READ) {
1486 * Reset the buf_filled parameters so next time if the
1487 * buffer is used for writes it is refilled.
1489 io_u->buf_filled_len = 0;
1494 * Set io data pointers.
1496 io_u->xfer_buf = io_u->buf;
1497 io_u->xfer_buflen = io_u->buflen;
1501 if (!td_io_prep(td, io_u)) {
1502 if (!td->o.disable_slat)
1503 fio_gettime(&io_u->start_time, NULL);
1505 small_content_scramble(io_u);
1509 dprint(FD_IO, "get_io_u failed\n");
1511 return ERR_PTR(ret);
1514 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1516 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1518 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1521 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1522 io_u->file ? " on file " : "",
1523 io_u->file ? io_u->file->file_name : "",
1524 strerror(io_u->error),
1525 io_ddir_name(io_u->ddir),
1526 io_u->offset, io_u->xfer_buflen);
1529 td_verror(td, io_u->error, "io_u error");
1532 static inline int gtod_reduce(struct thread_data *td)
1534 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1535 && td->o.disable_bw;
1538 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1539 struct io_completion_data *icd,
1540 const enum fio_ddir idx, unsigned int bytes)
1542 unsigned long lusec = 0;
1544 if (!gtod_reduce(td))
1545 lusec = utime_since(&io_u->issue_time, &icd->time);
1547 if (!td->o.disable_lat) {
1548 unsigned long tusec;
1550 tusec = utime_since(&io_u->start_time, &icd->time);
1551 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1553 if (td->flags & TD_F_PROFILE_OPS) {
1554 struct prof_io_ops *ops = &td->prof_io_ops;
1557 icd->error = ops->io_u_lat(td, tusec);
1560 if (td->o.max_latency && tusec > td->o.max_latency)
1561 lat_fatal(td, icd, tusec, td->o.max_latency);
1562 if (td->o.latency_target && tusec > td->o.latency_target) {
1563 if (lat_target_failed(td))
1564 lat_fatal(td, icd, tusec, td->o.latency_target);
1568 if (!td->o.disable_clat) {
1569 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1570 io_u_mark_latency(td, lusec);
1573 if (!td->o.disable_bw)
1574 add_bw_sample(td, idx, bytes, &icd->time);
1576 if (!gtod_reduce(td))
1577 add_iops_sample(td, idx, bytes, &icd->time);
1580 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1582 uint64_t secs, remainder, bps, bytes;
1584 bytes = td->this_io_bytes[ddir];
1585 bps = td->rate_bps[ddir];
1587 remainder = bytes % bps;
1588 return remainder * 1000000 / bps + secs * 1000000;
1591 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1592 struct io_completion_data *icd)
1594 struct io_u *io_u = *io_u_ptr;
1595 enum fio_ddir ddir = io_u->ddir;
1596 struct fio_file *f = io_u->file;
1598 dprint_io_u(io_u, "io complete");
1601 assert(io_u->flags & IO_U_F_FLIGHT);
1602 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1605 * Mark IO ok to verify
1609 * Remove errored entry from the verification list
1612 unlog_io_piece(td, io_u);
1614 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1621 if (ddir_sync(ddir)) {
1622 td->last_was_sync = 1;
1624 f->first_write = -1ULL;
1625 f->last_write = -1ULL;
1630 td->last_was_sync = 0;
1631 td->last_ddir = ddir;
1633 if (!io_u->error && ddir_rw(ddir)) {
1634 unsigned int bytes = io_u->buflen - io_u->resid;
1635 const enum fio_ddir oddir = ddir ^ 1;
1638 td->io_blocks[ddir]++;
1639 td->this_io_blocks[ddir]++;
1640 td->io_bytes[ddir] += bytes;
1642 if (!(io_u->flags & IO_U_F_VER_LIST))
1643 td->this_io_bytes[ddir] += bytes;
1645 if (ddir == DDIR_WRITE && f) {
1646 if (f->first_write == -1ULL ||
1647 io_u->offset < f->first_write)
1648 f->first_write = io_u->offset;
1649 if (f->last_write == -1ULL ||
1650 ((io_u->offset + bytes) > f->last_write))
1651 f->last_write = io_u->offset + bytes;
1654 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1655 td->runstate == TD_VERIFYING)) {
1656 account_io_completion(td, io_u, icd, ddir, bytes);
1658 if (__should_check_rate(td, ddir)) {
1659 td->rate_pending_usleep[ddir] =
1660 (usec_for_io(td, ddir) -
1661 utime_since_now(&td->start));
1663 if (ddir != DDIR_TRIM &&
1664 __should_check_rate(td, oddir)) {
1665 td->rate_pending_usleep[oddir] =
1666 (usec_for_io(td, oddir) -
1667 utime_since_now(&td->start));
1671 icd->bytes_done[ddir] += bytes;
1674 ret = io_u->end_io(td, io_u_ptr);
1676 if (ret && !icd->error)
1679 } else if (io_u->error) {
1680 icd->error = io_u->error;
1681 io_u_log_error(td, io_u);
1684 enum error_type_bit eb = td_error_type(ddir, icd->error);
1686 if (!td_non_fatal_error(td, eb, icd->error))
1690 * If there is a non_fatal error, then add to the error count
1691 * and clear all the errors.
1693 update_error_count(td, icd->error);
1701 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1706 if (!gtod_reduce(td))
1707 fio_gettime(&icd->time, NULL);
1712 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1713 icd->bytes_done[ddir] = 0;
1716 static void ios_completed(struct thread_data *td,
1717 struct io_completion_data *icd)
1722 for (i = 0; i < icd->nr; i++) {
1723 io_u = td->io_ops->event(td, i);
1725 io_completed(td, &io_u, icd);
1733 * Complete a single io_u for the sync engines.
1735 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1738 struct io_completion_data icd;
1740 init_icd(td, &icd, 1);
1741 io_completed(td, &io_u, &icd);
1747 td_verror(td, icd.error, "io_u_sync_complete");
1754 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1755 bytes[ddir] += icd.bytes_done[ddir];
1762 * Called to complete min_events number of io for the async engines.
1764 int io_u_queued_complete(struct thread_data *td, int min_evts,
1767 struct io_completion_data icd;
1768 struct timespec *tvp = NULL;
1770 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1772 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1776 else if (min_evts > td->cur_depth)
1777 min_evts = td->cur_depth;
1779 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1781 td_verror(td, -ret, "td_io_getevents");
1786 init_icd(td, &icd, ret);
1787 ios_completed(td, &icd);
1789 td_verror(td, icd.error, "io_u_queued_complete");
1796 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1797 bytes[ddir] += icd.bytes_done[ddir];
1804 * Call when io_u is really queued, to update the submission latency.
1806 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1808 if (!td->o.disable_slat) {
1809 unsigned long slat_time;
1811 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1812 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1818 * See if we should reuse the last seed, if dedupe is enabled
1820 static struct frand_state *get_buf_state(struct thread_data *td)
1825 if (!td->o.dedupe_percentage)
1826 return &td->buf_state;
1827 else if (td->o.dedupe_percentage == 100)
1828 return &td->buf_state_prev;
1830 r = __rand(&td->dedupe_state);
1831 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1833 if (v <= td->o.dedupe_percentage)
1834 return &td->buf_state_prev;
1836 return &td->buf_state;
1839 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1841 if (rs == &td->buf_state)
1842 frand_copy(&td->buf_state_prev, rs);
1845 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1846 unsigned int max_bs)
1848 if (td->o.buffer_pattern_bytes)
1849 fill_buffer_pattern(td, buf, max_bs);
1850 else if (!td->o.zero_buffers) {
1851 unsigned int perc = td->o.compress_percentage;
1852 struct frand_state *rs;
1853 unsigned int left = max_bs;
1856 rs = get_buf_state(td);
1858 min_write = min(min_write, left);
1861 unsigned int seg = min_write;
1863 seg = min(min_write, td->o.compress_chunk);
1867 fill_random_buf_percentage(rs, buf, perc, seg,
1870 fill_random_buf(rs, buf, min_write);
1874 save_buf_state(td, rs);
1877 memset(buf, 0, max_bs);
1881 * "randomly" fill the buffer contents
1883 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1884 unsigned int min_write, unsigned int max_bs)
1886 io_u->buf_filled_len = 0;
1887 fill_io_buffer(td, io_u->buf, min_write, max_bs);