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 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
91 lastb = last_block(td, f, ddir);
95 r = __rand(&td->random_state);
97 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
99 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
103 if (lfsr_next(&f->lfsr, &off))
110 * if we are not maintaining a random map, we are done.
112 if (!file_randommap(td, f))
116 * calculate map offset and check if it's free
118 if (random_map_free(f, *b))
121 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
122 (unsigned long long) *b);
124 *b = axmap_next_free(f->io_axmap, *b);
125 if (*b == (uint64_t) -1ULL)
131 static int __get_next_rand_offset_zipf(struct thread_data *td,
132 struct fio_file *f, enum fio_ddir ddir,
135 *b = zipf_next(&f->zipf);
139 static int __get_next_rand_offset_pareto(struct thread_data *td,
140 struct fio_file *f, enum fio_ddir ddir,
143 *b = pareto_next(&f->zipf);
147 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
149 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
150 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
152 return r1->off - r2->off;
155 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
156 enum fio_ddir ddir, uint64_t *b)
158 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
159 return __get_next_rand_offset(td, f, ddir, b);
160 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
161 return __get_next_rand_offset_zipf(td, f, ddir, b);
162 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
163 return __get_next_rand_offset_pareto(td, f, ddir, b);
165 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
170 * Sort the reads for a verify phase in batches of verifysort_nr, if
173 static inline int should_sort_io(struct thread_data *td)
175 if (!td->o.verifysort_nr || !td->o.do_verify)
179 if (td->runstate != TD_VERIFYING)
181 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
187 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
192 if (td->o.perc_rand[ddir] == 100)
195 r = __rand(&td->seq_rand_state[ddir]);
196 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
198 return v <= td->o.perc_rand[ddir];
201 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
202 enum fio_ddir ddir, uint64_t *b)
207 if (!should_sort_io(td))
208 return get_off_from_method(td, f, ddir, b);
210 if (!flist_empty(&td->next_rand_list)) {
212 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
219 for (i = 0; i < td->o.verifysort_nr; i++) {
220 r = malloc(sizeof(*r));
222 ret = get_off_from_method(td, f, ddir, &r->off);
228 flist_add(&r->list, &td->next_rand_list);
234 assert(!flist_empty(&td->next_rand_list));
235 flist_sort(NULL, &td->next_rand_list, flist_cmp);
239 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
240 enum fio_ddir ddir, uint64_t *b)
242 if (!get_next_rand_offset(td, f, ddir, b))
245 if (td->o.time_based) {
246 fio_file_reset(td, f);
247 if (!get_next_rand_offset(td, f, ddir, b))
251 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
252 f->file_name, (unsigned long long) f->last_pos,
253 (unsigned long long) f->real_file_size);
257 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
258 enum fio_ddir ddir, uint64_t *offset)
260 struct thread_options *o = &td->o;
262 assert(ddir_rw(ddir));
264 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
266 f->last_pos = f->last_pos - f->io_size;
268 if (f->last_pos < f->real_file_size) {
271 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
272 f->last_pos = f->real_file_size;
274 pos = f->last_pos - f->file_offset;
275 if (pos && o->ddir_seq_add) {
276 pos += o->ddir_seq_add;
279 * If we reach beyond the end of the file
280 * with holed IO, wrap around to the
283 if (pos >= f->real_file_size)
284 pos = f->file_offset;
294 static int get_next_block(struct thread_data *td, struct io_u *io_u,
295 enum fio_ddir ddir, int rw_seq,
296 unsigned int *is_random)
298 struct fio_file *f = io_u->file;
302 assert(ddir_rw(ddir));
308 if (should_do_random(td, ddir)) {
309 ret = get_next_rand_block(td, f, ddir, &b);
313 io_u->flags |= IO_U_F_BUSY_OK;
314 ret = get_next_seq_offset(td, f, ddir, &offset);
316 ret = get_next_rand_block(td, f, ddir, &b);
320 ret = get_next_seq_offset(td, f, ddir, &offset);
323 io_u->flags |= IO_U_F_BUSY_OK;
326 if (td->o.rw_seq == RW_SEQ_SEQ) {
327 ret = get_next_seq_offset(td, f, ddir, &offset);
329 ret = get_next_rand_block(td, f, ddir, &b);
332 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
333 if (f->last_start != -1ULL)
334 offset = f->last_start - f->file_offset;
339 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
346 io_u->offset = offset;
348 io_u->offset = b * td->o.ba[ddir];
350 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
359 * For random io, generate a random new block and see if it's used. Repeat
360 * until we find a free one. For sequential io, just return the end of
361 * the last io issued.
363 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
364 unsigned int *is_random)
366 struct fio_file *f = io_u->file;
367 enum fio_ddir ddir = io_u->ddir;
370 assert(ddir_rw(ddir));
372 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
374 td->ddir_seq_nr = td->o.ddir_seq_nr;
377 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
380 if (io_u->offset >= f->io_size) {
381 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
382 (unsigned long long) io_u->offset,
383 (unsigned long long) f->io_size);
387 io_u->offset += f->file_offset;
388 if (io_u->offset >= f->real_file_size) {
389 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
390 (unsigned long long) io_u->offset,
391 (unsigned long long) f->real_file_size);
398 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
399 unsigned int *is_random)
401 if (td->flags & TD_F_PROFILE_OPS) {
402 struct prof_io_ops *ops = &td->prof_io_ops;
404 if (ops->fill_io_u_off)
405 return ops->fill_io_u_off(td, io_u, is_random);
408 return __get_next_offset(td, io_u, is_random);
411 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
414 struct fio_file *f = io_u->file;
416 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
419 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
420 unsigned int is_random)
422 int ddir = io_u->ddir;
423 unsigned int buflen = 0;
424 unsigned int minbs, maxbs;
427 assert(ddir_rw(ddir));
429 if (td->o.bs_is_seq_rand)
430 ddir = is_random ? DDIR_WRITE: DDIR_READ;
432 minbs = td->o.min_bs[ddir];
433 maxbs = td->o.max_bs[ddir];
439 * If we can't satisfy the min block size from here, then fail
441 if (!io_u_fits(td, io_u, minbs))
445 r = __rand(&td->bsrange_state);
447 if (!td->o.bssplit_nr[ddir]) {
448 buflen = 1 + (unsigned int) ((double) maxbs *
449 (r / (FRAND_MAX + 1.0)));
456 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
457 struct bssplit *bsp = &td->o.bssplit[ddir][i];
461 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
462 io_u_fits(td, io_u, buflen))
467 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
468 buflen = (buflen + td->o.verify_interval - 1) &
469 ~(td->o.verify_interval - 1);
471 if (!td->o.bs_unaligned && is_power_of_2(minbs))
472 buflen = (buflen + minbs - 1) & ~(minbs - 1);
474 } while (!io_u_fits(td, io_u, buflen));
479 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
480 unsigned int is_random)
482 if (td->flags & TD_F_PROFILE_OPS) {
483 struct prof_io_ops *ops = &td->prof_io_ops;
485 if (ops->fill_io_u_size)
486 return ops->fill_io_u_size(td, io_u, is_random);
489 return __get_next_buflen(td, io_u, is_random);
492 static void set_rwmix_bytes(struct thread_data *td)
497 * we do time or byte based switch. this is needed because
498 * buffered writes may issue a lot quicker than they complete,
499 * whereas reads do not.
501 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
502 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
505 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
510 r = __rand(&td->rwmix_state);
511 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
513 if (v <= td->o.rwmix[DDIR_READ])
519 void io_u_quiesce(struct thread_data *td)
522 * We are going to sleep, ensure that we flush anything pending as
523 * not to skew our latency numbers.
525 * Changed to only monitor 'in flight' requests here instead of the
526 * td->cur_depth, b/c td->cur_depth does not accurately represent
527 * io's that have been actually submitted to an async engine,
528 * and cur_depth is meaningless for sync engines.
530 while (td->io_u_in_flight) {
533 ret = io_u_queued_complete(td, 1, NULL);
537 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
539 enum fio_ddir odir = ddir ^ 1;
543 assert(ddir_rw(ddir));
545 if (td->rate_pending_usleep[ddir] <= 0)
549 * We have too much pending sleep in this direction. See if we
552 if (td_rw(td) && td->o.rwmix[odir]) {
554 * Other direction does not have too much pending, switch
556 if (td->rate_pending_usleep[odir] < 100000)
560 * Both directions have pending sleep. Sleep the minimum time
561 * and deduct from both.
563 if (td->rate_pending_usleep[ddir] <=
564 td->rate_pending_usleep[odir]) {
565 usec = td->rate_pending_usleep[ddir];
567 usec = td->rate_pending_usleep[odir];
571 usec = td->rate_pending_usleep[ddir];
575 fio_gettime(&t, NULL);
576 usec_sleep(td, usec);
577 usec = utime_since_now(&t);
579 td->rate_pending_usleep[ddir] -= usec;
582 if (td_rw(td) && __should_check_rate(td, odir))
583 td->rate_pending_usleep[odir] -= usec;
585 if (ddir == DDIR_TRIM)
592 * Return the data direction for the next io_u. If the job is a
593 * mixed read/write workload, check the rwmix cycle and switch if
596 static enum fio_ddir get_rw_ddir(struct thread_data *td)
601 * see if it's time to fsync
603 if (td->o.fsync_blocks &&
604 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
605 td->io_issues[DDIR_WRITE] && should_fsync(td))
609 * see if it's time to fdatasync
611 if (td->o.fdatasync_blocks &&
612 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
613 td->io_issues[DDIR_WRITE] && should_fsync(td))
614 return DDIR_DATASYNC;
617 * see if it's time to sync_file_range
619 if (td->sync_file_range_nr &&
620 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
621 td->io_issues[DDIR_WRITE] && should_fsync(td))
622 return DDIR_SYNC_FILE_RANGE;
626 * Check if it's time to seed a new data direction.
628 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
630 * Put a top limit on how many bytes we do for
631 * one data direction, to avoid overflowing the
634 ddir = get_rand_ddir(td);
636 if (ddir != td->rwmix_ddir)
639 td->rwmix_ddir = ddir;
641 ddir = td->rwmix_ddir;
642 } else if (td_read(td))
644 else if (td_write(td))
649 td->rwmix_ddir = rate_ddir(td, ddir);
650 return td->rwmix_ddir;
653 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
655 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
657 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
658 td->o.barrier_blocks &&
659 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
660 td->io_issues[DDIR_WRITE])
661 io_u->flags |= IO_U_F_BARRIER;
664 void put_file_log(struct thread_data *td, struct fio_file *f)
666 unsigned int ret = put_file(td, f);
669 td_verror(td, ret, "file close");
672 void put_io_u(struct thread_data *td, struct io_u *io_u)
676 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
677 put_file_log(td, io_u->file);
680 io_u->flags |= IO_U_F_FREE;
682 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
684 io_u_qpush(&td->io_u_freelist, io_u);
686 td_io_u_free_notify(td);
689 void clear_io_u(struct thread_data *td, struct io_u *io_u)
691 io_u->flags &= ~IO_U_F_FLIGHT;
695 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
697 struct io_u *__io_u = *io_u;
698 enum fio_ddir ddir = acct_ddir(__io_u);
700 dprint(FD_IO, "requeue %p\n", __io_u);
704 __io_u->flags |= IO_U_F_FREE;
705 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
706 td->io_issues[ddir]--;
708 __io_u->flags &= ~IO_U_F_FLIGHT;
709 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
712 io_u_rpush(&td->io_u_requeues, __io_u);
717 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
719 unsigned int is_random;
721 if (td->io_ops->flags & FIO_NOIO)
724 set_rw_ddir(td, io_u);
727 * fsync() or fdatasync() or trim etc, we are done
729 if (!ddir_rw(io_u->ddir))
733 * See if it's time to switch to a new zone
735 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
736 struct fio_file *f = io_u->file;
739 f->file_offset += td->o.zone_range + td->o.zone_skip;
742 * Wrap from the beginning, if we exceed the file size
744 if (f->file_offset >= f->real_file_size)
745 f->file_offset = f->real_file_size - f->file_offset;
746 f->last_pos = f->file_offset;
747 td->io_skip_bytes += td->o.zone_skip;
751 * No log, let the seq/rand engine retrieve the next buflen and
754 if (get_next_offset(td, io_u, &is_random)) {
755 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
759 io_u->buflen = get_next_buflen(td, io_u, is_random);
761 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
765 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
766 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
767 dprint(FD_IO, " off=%llu/%lu > %llu\n",
768 (unsigned long long) io_u->offset, io_u->buflen,
769 (unsigned long long) io_u->file->real_file_size);
774 * mark entry before potentially trimming io_u
776 if (td_random(td) && file_randommap(td, io_u->file))
777 mark_random_map(td, io_u);
780 dprint_io_u(io_u, "fill_io_u");
781 td->zone_bytes += io_u->buflen;
785 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
814 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
816 __io_u_mark_map(td->ts.io_u_submit, nr);
817 td->ts.total_submit++;
820 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
822 __io_u_mark_map(td->ts.io_u_complete, nr);
823 td->ts.total_complete++;
826 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
830 switch (td->cur_depth) {
852 td->ts.io_u_map[idx] += nr;
855 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
892 assert(idx < FIO_IO_U_LAT_U_NR);
893 td->ts.io_u_lat_u[idx]++;
896 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
937 assert(idx < FIO_IO_U_LAT_M_NR);
938 td->ts.io_u_lat_m[idx]++;
941 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
944 io_u_mark_lat_usec(td, usec);
946 io_u_mark_lat_msec(td, usec / 1000);
950 * Get next file to service by choosing one at random
952 static struct fio_file *get_next_file_rand(struct thread_data *td,
953 enum fio_file_flags goodf,
954 enum fio_file_flags badf)
963 r = __rand(&td->next_file_state);
964 fno = (unsigned int) ((double) td->o.nr_files
965 * (r / (FRAND_MAX + 1.0)));
968 if (fio_file_done(f))
971 if (!fio_file_open(f)) {
974 if (td->nr_open_files >= td->o.open_files)
975 return ERR_PTR(-EBUSY);
977 err = td_io_open_file(td, f);
983 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
984 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
988 td_io_close_file(td, f);
993 * Get next file to service by doing round robin between all available ones
995 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
998 unsigned int old_next_file = td->next_file;
1004 f = td->files[td->next_file];
1007 if (td->next_file >= td->o.nr_files)
1010 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1011 if (fio_file_done(f)) {
1016 if (!fio_file_open(f)) {
1019 if (td->nr_open_files >= td->o.open_files)
1020 return ERR_PTR(-EBUSY);
1022 err = td_io_open_file(td, f);
1024 dprint(FD_FILE, "error %d on open of %s\n",
1032 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1034 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1038 td_io_close_file(td, f);
1041 } while (td->next_file != old_next_file);
1043 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1047 static struct fio_file *__get_next_file(struct thread_data *td)
1051 assert(td->o.nr_files <= td->files_index);
1053 if (td->nr_done_files >= td->o.nr_files) {
1054 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1055 " nr_files=%d\n", td->nr_open_files,
1061 f = td->file_service_file;
1062 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1063 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1065 if (td->file_service_left--)
1069 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1070 td->o.file_service_type == FIO_FSERVICE_SEQ)
1071 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1073 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1078 td->file_service_file = f;
1079 td->file_service_left = td->file_service_nr - 1;
1082 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1084 dprint(FD_FILE, "get_next_file: NULL\n");
1088 static struct fio_file *get_next_file(struct thread_data *td)
1090 if (td->flags & TD_F_PROFILE_OPS) {
1091 struct prof_io_ops *ops = &td->prof_io_ops;
1093 if (ops->get_next_file)
1094 return ops->get_next_file(td);
1097 return __get_next_file(td);
1100 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1105 f = get_next_file(td);
1106 if (IS_ERR_OR_NULL(f))
1112 if (!fill_io_u(td, io_u))
1115 put_file_log(td, f);
1116 td_io_close_file(td, f);
1118 fio_file_set_done(f);
1119 td->nr_done_files++;
1120 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1121 td->nr_done_files, td->o.nr_files);
1127 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1128 unsigned long tusec, unsigned long max_usec)
1131 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1132 td_verror(td, ETIMEDOUT, "max latency exceeded");
1133 icd->error = ETIMEDOUT;
1136 static void lat_new_cycle(struct thread_data *td)
1138 fio_gettime(&td->latency_ts, NULL);
1139 td->latency_ios = ddir_rw_sum(td->io_blocks);
1140 td->latency_failed = 0;
1144 * We had an IO outside the latency target. Reduce the queue depth. If we
1145 * are at QD=1, then it's time to give up.
1147 static int __lat_target_failed(struct thread_data *td)
1149 if (td->latency_qd == 1)
1152 td->latency_qd_high = td->latency_qd;
1154 if (td->latency_qd == td->latency_qd_low)
1155 td->latency_qd_low--;
1157 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1159 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1162 * When we ramp QD down, quiesce existing IO to prevent
1163 * a storm of ramp downs due to pending higher depth.
1170 static int lat_target_failed(struct thread_data *td)
1172 if (td->o.latency_percentile.u.f == 100.0)
1173 return __lat_target_failed(td);
1175 td->latency_failed++;
1179 void lat_target_init(struct thread_data *td)
1181 td->latency_end_run = 0;
1183 if (td->o.latency_target) {
1184 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1185 fio_gettime(&td->latency_ts, NULL);
1187 td->latency_qd_high = td->o.iodepth;
1188 td->latency_qd_low = 1;
1189 td->latency_ios = ddir_rw_sum(td->io_blocks);
1191 td->latency_qd = td->o.iodepth;
1194 void lat_target_reset(struct thread_data *td)
1196 if (!td->latency_end_run)
1197 lat_target_init(td);
1200 static void lat_target_success(struct thread_data *td)
1202 const unsigned int qd = td->latency_qd;
1203 struct thread_options *o = &td->o;
1205 td->latency_qd_low = td->latency_qd;
1208 * If we haven't failed yet, we double up to a failing value instead
1209 * of bisecting from highest possible queue depth. If we have set
1210 * a limit other than td->o.iodepth, bisect between that.
1212 if (td->latency_qd_high != o->iodepth)
1213 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1215 td->latency_qd *= 2;
1217 if (td->latency_qd > o->iodepth)
1218 td->latency_qd = o->iodepth;
1220 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1223 * Same as last one, we are done. Let it run a latency cycle, so
1224 * we get only the results from the targeted depth.
1226 if (td->latency_qd == qd) {
1227 if (td->latency_end_run) {
1228 dprint(FD_RATE, "We are done\n");
1231 dprint(FD_RATE, "Quiesce and final run\n");
1233 td->latency_end_run = 1;
1234 reset_all_stats(td);
1243 * Check if we can bump the queue depth
1245 void lat_target_check(struct thread_data *td)
1247 uint64_t usec_window;
1251 usec_window = utime_since_now(&td->latency_ts);
1252 if (usec_window < td->o.latency_window)
1255 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1256 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1257 success_ios *= 100.0;
1259 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1261 if (success_ios >= td->o.latency_percentile.u.f)
1262 lat_target_success(td);
1264 __lat_target_failed(td);
1268 * If latency target is enabled, we might be ramping up or down and not
1269 * using the full queue depth available.
1271 int queue_full(const struct thread_data *td)
1273 const int qempty = io_u_qempty(&td->io_u_freelist);
1277 if (!td->o.latency_target)
1280 return td->cur_depth >= td->latency_qd;
1283 struct io_u *__get_io_u(struct thread_data *td)
1285 struct io_u *io_u = NULL;
1293 if (!io_u_rempty(&td->io_u_requeues))
1294 io_u = io_u_rpop(&td->io_u_requeues);
1295 else if (!queue_full(td)) {
1296 io_u = io_u_qpop(&td->io_u_freelist);
1301 io_u->end_io = NULL;
1305 assert(io_u->flags & IO_U_F_FREE);
1306 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1307 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1311 io_u->acct_ddir = -1;
1313 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1315 } else if (td->o.verify_async) {
1317 * We ran out, wait for async verify threads to finish and
1320 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1328 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1330 if (!(td->flags & TD_F_TRIM_BACKLOG))
1333 if (td->trim_entries) {
1336 if (td->trim_batch) {
1339 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1340 td->last_ddir != DDIR_READ) {
1341 td->trim_batch = td->o.trim_batch;
1342 if (!td->trim_batch)
1343 td->trim_batch = td->o.trim_backlog;
1347 if (get_trim && !get_next_trim(td, io_u))
1354 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1356 if (!(td->flags & TD_F_VER_BACKLOG))
1359 if (td->io_hist_len) {
1362 if (td->verify_batch)
1364 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1365 td->last_ddir != DDIR_READ) {
1366 td->verify_batch = td->o.verify_batch;
1367 if (!td->verify_batch)
1368 td->verify_batch = td->o.verify_backlog;
1372 if (get_verify && !get_next_verify(td, io_u)) {
1382 * Fill offset and start time into the buffer content, to prevent too
1383 * easy compressible data for simple de-dupe attempts. Do this for every
1384 * 512b block in the range, since that should be the smallest block size
1385 * we can expect from a device.
1387 static void small_content_scramble(struct io_u *io_u)
1389 unsigned int i, nr_blocks = io_u->buflen / 512;
1391 unsigned int offset;
1398 boffset = io_u->offset;
1399 io_u->buf_filled_len = 0;
1401 for (i = 0; i < nr_blocks; i++) {
1403 * Fill the byte offset into a "random" start offset of
1404 * the buffer, given by the product of the usec time
1405 * and the actual offset.
1407 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1408 offset &= ~(sizeof(uint64_t) - 1);
1409 if (offset >= 512 - sizeof(uint64_t))
1410 offset -= sizeof(uint64_t);
1411 memcpy(p + offset, &boffset, sizeof(boffset));
1413 end = p + 512 - sizeof(io_u->start_time);
1414 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1421 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1422 * etc. The returned io_u is fully ready to be prepped and submitted.
1424 struct io_u *get_io_u(struct thread_data *td)
1428 int do_scramble = 0;
1431 io_u = __get_io_u(td);
1433 dprint(FD_IO, "__get_io_u failed\n");
1437 if (check_get_verify(td, io_u))
1439 if (check_get_trim(td, io_u))
1443 * from a requeue, io_u already setup
1449 * If using an iolog, grab next piece if any available.
1451 if (td->flags & TD_F_READ_IOLOG) {
1452 if (read_iolog_get(td, io_u))
1454 } else if (set_io_u_file(td, io_u)) {
1456 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1462 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1466 assert(fio_file_open(f));
1468 if (ddir_rw(io_u->ddir)) {
1469 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1470 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1474 f->last_start = io_u->offset;
1475 f->last_pos = io_u->offset + io_u->buflen;
1477 if (io_u->ddir == DDIR_WRITE) {
1478 if (td->flags & TD_F_REFILL_BUFFERS) {
1479 io_u_fill_buffer(td, io_u,
1480 td->o.min_bs[DDIR_WRITE],
1482 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1483 !(td->flags & TD_F_COMPRESS))
1485 if (td->flags & TD_F_VER_NONE) {
1486 populate_verify_io_u(td, io_u);
1489 } else if (io_u->ddir == DDIR_READ) {
1491 * Reset the buf_filled parameters so next time if the
1492 * buffer is used for writes it is refilled.
1494 io_u->buf_filled_len = 0;
1499 * Set io data pointers.
1501 io_u->xfer_buf = io_u->buf;
1502 io_u->xfer_buflen = io_u->buflen;
1506 if (!td_io_prep(td, io_u)) {
1507 if (!td->o.disable_slat)
1508 fio_gettime(&io_u->start_time, NULL);
1510 small_content_scramble(io_u);
1514 dprint(FD_IO, "get_io_u failed\n");
1516 return ERR_PTR(ret);
1519 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1521 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1523 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1526 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1527 io_u->file ? " on file " : "",
1528 io_u->file ? io_u->file->file_name : "",
1529 strerror(io_u->error),
1530 io_ddir_name(io_u->ddir),
1531 io_u->offset, io_u->xfer_buflen);
1534 td_verror(td, io_u->error, "io_u error");
1537 static inline int gtod_reduce(struct thread_data *td)
1539 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1540 && td->o.disable_bw;
1543 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1544 struct io_completion_data *icd,
1545 const enum fio_ddir idx, unsigned int bytes)
1547 unsigned long lusec = 0;
1549 if (!gtod_reduce(td))
1550 lusec = utime_since(&io_u->issue_time, &icd->time);
1552 if (!td->o.disable_lat) {
1553 unsigned long tusec;
1555 tusec = utime_since(&io_u->start_time, &icd->time);
1556 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1558 if (td->flags & TD_F_PROFILE_OPS) {
1559 struct prof_io_ops *ops = &td->prof_io_ops;
1562 icd->error = ops->io_u_lat(td, tusec);
1565 if (td->o.max_latency && tusec > td->o.max_latency)
1566 lat_fatal(td, icd, tusec, td->o.max_latency);
1567 if (td->o.latency_target && tusec > td->o.latency_target) {
1568 if (lat_target_failed(td))
1569 lat_fatal(td, icd, tusec, td->o.latency_target);
1573 if (!td->o.disable_clat) {
1574 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1575 io_u_mark_latency(td, lusec);
1578 if (!td->o.disable_bw)
1579 add_bw_sample(td, idx, bytes, &icd->time);
1581 if (!gtod_reduce(td))
1582 add_iops_sample(td, idx, bytes, &icd->time);
1585 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1587 uint64_t secs, remainder, bps, bytes;
1589 bytes = td->this_io_bytes[ddir];
1590 bps = td->rate_bps[ddir];
1592 remainder = bytes % bps;
1593 return remainder * 1000000 / bps + secs * 1000000;
1596 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1597 struct io_completion_data *icd)
1599 struct io_u *io_u = *io_u_ptr;
1600 enum fio_ddir ddir = io_u->ddir;
1601 struct fio_file *f = io_u->file;
1603 dprint_io_u(io_u, "io complete");
1606 assert(io_u->flags & IO_U_F_FLIGHT);
1607 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1610 * Mark IO ok to verify
1614 * Remove errored entry from the verification list
1617 unlog_io_piece(td, io_u);
1619 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1626 if (ddir_sync(ddir)) {
1627 td->last_was_sync = 1;
1629 f->first_write = -1ULL;
1630 f->last_write = -1ULL;
1635 td->last_was_sync = 0;
1636 td->last_ddir = ddir;
1638 if (!io_u->error && ddir_rw(ddir)) {
1639 unsigned int bytes = io_u->buflen - io_u->resid;
1640 const enum fio_ddir oddir = ddir ^ 1;
1643 td->io_blocks[ddir]++;
1644 td->this_io_blocks[ddir]++;
1645 td->io_bytes[ddir] += bytes;
1647 if (!(io_u->flags & IO_U_F_VER_LIST))
1648 td->this_io_bytes[ddir] += bytes;
1650 if (ddir == DDIR_WRITE) {
1652 if (f->first_write == -1ULL ||
1653 io_u->offset < f->first_write)
1654 f->first_write = io_u->offset;
1655 if (f->last_write == -1ULL ||
1656 ((io_u->offset + bytes) > f->last_write))
1657 f->last_write = io_u->offset + bytes;
1659 if (td->last_write_comp) {
1660 int idx = td->last_write_idx++;
1662 td->last_write_comp[idx] = io_u->offset;
1663 if (td->last_write_idx == td->o.iodepth)
1664 td->last_write_idx = 0;
1668 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1669 td->runstate == TD_VERIFYING)) {
1670 account_io_completion(td, io_u, icd, ddir, bytes);
1672 if (__should_check_rate(td, ddir)) {
1673 td->rate_pending_usleep[ddir] =
1674 (usec_for_io(td, ddir) -
1675 utime_since_now(&td->start));
1677 if (ddir != DDIR_TRIM &&
1678 __should_check_rate(td, oddir)) {
1679 td->rate_pending_usleep[oddir] =
1680 (usec_for_io(td, oddir) -
1681 utime_since_now(&td->start));
1685 icd->bytes_done[ddir] += bytes;
1688 ret = io_u->end_io(td, io_u_ptr);
1690 if (ret && !icd->error)
1693 } else if (io_u->error) {
1694 icd->error = io_u->error;
1695 io_u_log_error(td, io_u);
1698 enum error_type_bit eb = td_error_type(ddir, icd->error);
1700 if (!td_non_fatal_error(td, eb, icd->error))
1704 * If there is a non_fatal error, then add to the error count
1705 * and clear all the errors.
1707 update_error_count(td, icd->error);
1715 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1720 if (!gtod_reduce(td))
1721 fio_gettime(&icd->time, NULL);
1726 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1727 icd->bytes_done[ddir] = 0;
1730 static void ios_completed(struct thread_data *td,
1731 struct io_completion_data *icd)
1736 for (i = 0; i < icd->nr; i++) {
1737 io_u = td->io_ops->event(td, i);
1739 io_completed(td, &io_u, icd);
1747 * Complete a single io_u for the sync engines.
1749 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1752 struct io_completion_data icd;
1754 init_icd(td, &icd, 1);
1755 io_completed(td, &io_u, &icd);
1761 td_verror(td, icd.error, "io_u_sync_complete");
1768 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1769 bytes[ddir] += icd.bytes_done[ddir];
1776 * Called to complete min_events number of io for the async engines.
1778 int io_u_queued_complete(struct thread_data *td, int min_evts,
1781 struct io_completion_data icd;
1782 struct timespec *tvp = NULL;
1784 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1786 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1790 else if (min_evts > td->cur_depth)
1791 min_evts = td->cur_depth;
1793 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1795 td_verror(td, -ret, "td_io_getevents");
1800 init_icd(td, &icd, ret);
1801 ios_completed(td, &icd);
1803 td_verror(td, icd.error, "io_u_queued_complete");
1810 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1811 bytes[ddir] += icd.bytes_done[ddir];
1818 * Call when io_u is really queued, to update the submission latency.
1820 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1822 if (!td->o.disable_slat) {
1823 unsigned long slat_time;
1825 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1826 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1832 * See if we should reuse the last seed, if dedupe is enabled
1834 static struct frand_state *get_buf_state(struct thread_data *td)
1839 if (!td->o.dedupe_percentage)
1840 return &td->buf_state;
1841 else if (td->o.dedupe_percentage == 100)
1842 return &td->buf_state_prev;
1844 r = __rand(&td->dedupe_state);
1845 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1847 if (v <= td->o.dedupe_percentage)
1848 return &td->buf_state_prev;
1850 return &td->buf_state;
1853 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1855 if (rs == &td->buf_state)
1856 frand_copy(&td->buf_state_prev, rs);
1859 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1860 unsigned int max_bs)
1862 if (td->o.buffer_pattern_bytes)
1863 fill_buffer_pattern(td, buf, max_bs);
1864 else if (!td->o.zero_buffers) {
1865 unsigned int perc = td->o.compress_percentage;
1866 struct frand_state *rs;
1867 unsigned int left = max_bs;
1870 rs = get_buf_state(td);
1872 min_write = min(min_write, left);
1875 unsigned int seg = min_write;
1877 seg = min(min_write, td->o.compress_chunk);
1881 fill_random_buf_percentage(rs, buf, perc, seg,
1884 fill_random_buf(rs, buf, min_write);
1888 save_buf_state(td, rs);
1891 memset(buf, 0, max_bs);
1895 * "randomly" fill the buffer contents
1897 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1898 unsigned int min_write, unsigned int max_bs)
1900 io_u->buf_filled_len = 0;
1901 fill_io_buffer(td, io_u->buf, min_write, max_bs);