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) {
95 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
97 if (td->o.use_os_rand) {
99 r = os_random_long(&td->random_state);
102 r = __rand(&td->__random_state);
105 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
107 *b = lastb * (r / ((uint64_t) rmax + 1.0));
111 if (lfsr_next(&f->lfsr, &off, lastb))
118 * if we are not maintaining a random map, we are done.
120 if (!file_randommap(td, f))
124 * calculate map offset and check if it's free
126 if (random_map_free(f, *b))
129 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
130 (unsigned long long) *b);
132 *b = axmap_next_free(f->io_axmap, *b);
133 if (*b == (uint64_t) -1ULL)
139 static int __get_next_rand_offset_zipf(struct thread_data *td,
140 struct fio_file *f, enum fio_ddir ddir,
143 *b = zipf_next(&f->zipf);
147 static int __get_next_rand_offset_pareto(struct thread_data *td,
148 struct fio_file *f, enum fio_ddir ddir,
151 *b = pareto_next(&f->zipf);
155 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
157 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
158 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
160 return r1->off - r2->off;
163 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
164 enum fio_ddir ddir, uint64_t *b)
166 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
167 return __get_next_rand_offset(td, f, ddir, b);
168 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
169 return __get_next_rand_offset_zipf(td, f, ddir, b);
170 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
171 return __get_next_rand_offset_pareto(td, f, ddir, b);
173 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
178 * Sort the reads for a verify phase in batches of verifysort_nr, if
181 static inline int should_sort_io(struct thread_data *td)
183 if (!td->o.verifysort_nr || !td->o.do_verify)
187 if (td->runstate != TD_VERIFYING)
189 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
195 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
200 if (td->o.perc_rand[ddir] == 100)
203 if (td->o.use_os_rand) {
204 r = os_random_long(&td->seq_rand_state[ddir]);
205 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
207 r = __rand(&td->__seq_rand_state[ddir]);
208 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
211 return v <= td->o.perc_rand[ddir];
214 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
215 enum fio_ddir ddir, uint64_t *b)
220 if (!should_sort_io(td))
221 return get_off_from_method(td, f, ddir, b);
223 if (!flist_empty(&td->next_rand_list)) {
226 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
233 for (i = 0; i < td->o.verifysort_nr; i++) {
234 r = malloc(sizeof(*r));
236 ret = get_off_from_method(td, f, ddir, &r->off);
242 flist_add(&r->list, &td->next_rand_list);
248 assert(!flist_empty(&td->next_rand_list));
249 flist_sort(NULL, &td->next_rand_list, flist_cmp);
253 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
254 enum fio_ddir ddir, uint64_t *b)
256 if (!get_next_rand_offset(td, f, ddir, b))
259 if (td->o.time_based) {
260 fio_file_reset(td, f);
261 if (!get_next_rand_offset(td, f, ddir, b))
265 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
266 f->file_name, (unsigned long long) f->last_pos,
267 (unsigned long long) f->real_file_size);
271 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
272 enum fio_ddir ddir, uint64_t *offset)
274 assert(ddir_rw(ddir));
276 if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->o.time_based)
277 f->last_pos = f->last_pos - f->io_size;
279 if (f->last_pos < f->real_file_size) {
282 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
283 f->last_pos = f->real_file_size;
285 pos = f->last_pos - f->file_offset;
287 pos += td->o.ddir_seq_add;
296 static int get_next_block(struct thread_data *td, struct io_u *io_u,
297 enum fio_ddir ddir, int rw_seq,
298 unsigned int *is_random)
300 struct fio_file *f = io_u->file;
304 assert(ddir_rw(ddir));
310 if (should_do_random(td, ddir)) {
311 ret = get_next_rand_block(td, f, ddir, &b);
315 io_u->flags |= IO_U_F_BUSY_OK;
316 ret = get_next_seq_offset(td, f, ddir, &offset);
318 ret = get_next_rand_block(td, f, ddir, &b);
322 ret = get_next_seq_offset(td, f, ddir, &offset);
325 io_u->flags |= IO_U_F_BUSY_OK;
328 if (td->o.rw_seq == RW_SEQ_SEQ) {
329 ret = get_next_seq_offset(td, f, ddir, &offset);
331 ret = get_next_rand_block(td, f, ddir, &b);
334 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
335 if (f->last_start != -1ULL)
336 offset = f->last_start - f->file_offset;
341 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
348 io_u->offset = offset;
350 io_u->offset = b * td->o.ba[ddir];
352 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
361 * For random io, generate a random new block and see if it's used. Repeat
362 * until we find a free one. For sequential io, just return the end of
363 * the last io issued.
365 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
366 unsigned int *is_random)
368 struct fio_file *f = io_u->file;
369 enum fio_ddir ddir = io_u->ddir;
372 assert(ddir_rw(ddir));
374 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
376 td->ddir_seq_nr = td->o.ddir_seq_nr;
379 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
382 if (io_u->offset >= f->io_size) {
383 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
384 (unsigned long long) io_u->offset,
385 (unsigned long long) f->io_size);
389 io_u->offset += f->file_offset;
390 if (io_u->offset >= f->real_file_size) {
391 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
392 (unsigned long long) io_u->offset,
393 (unsigned long long) f->real_file_size);
400 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
401 unsigned int *is_random)
403 if (td->flags & TD_F_PROFILE_OPS) {
404 struct prof_io_ops *ops = &td->prof_io_ops;
406 if (ops->fill_io_u_off)
407 return ops->fill_io_u_off(td, io_u, is_random);
410 return __get_next_offset(td, io_u, is_random);
413 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
416 struct fio_file *f = io_u->file;
418 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
421 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
422 unsigned int is_random)
424 int ddir = io_u->ddir;
425 unsigned int buflen = 0;
426 unsigned int minbs, maxbs;
427 unsigned long r, rand_max;
429 assert(ddir_rw(io_u->ddir));
431 if (td->o.bs_is_seq_rand)
432 ddir = is_random ? DDIR_WRITE: DDIR_READ;
436 minbs = td->o.min_bs[ddir];
437 maxbs = td->o.max_bs[ddir];
443 * If we can't satisfy the min block size from here, then fail
445 if (!io_u_fits(td, io_u, minbs))
448 if (td->o.use_os_rand)
449 rand_max = OS_RAND_MAX;
451 rand_max = FRAND_MAX;
454 if (td->o.use_os_rand)
455 r = os_random_long(&td->bsrange_state);
457 r = __rand(&td->__bsrange_state);
459 if (!td->o.bssplit_nr[ddir]) {
460 buflen = 1 + (unsigned int) ((double) maxbs *
461 (r / (rand_max + 1.0)));
468 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
469 struct bssplit *bsp = &td->o.bssplit[ddir][i];
473 if ((r <= ((rand_max / 100L) * perc)) &&
474 io_u_fits(td, io_u, buflen))
479 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
480 buflen = (buflen + td->o.verify_interval - 1) &
481 ~(td->o.verify_interval - 1);
483 if (!td->o.bs_unaligned && is_power_of_2(minbs))
484 buflen = (buflen + minbs - 1) & ~(minbs - 1);
486 } while (!io_u_fits(td, io_u, buflen));
491 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
492 unsigned int is_random)
494 if (td->flags & TD_F_PROFILE_OPS) {
495 struct prof_io_ops *ops = &td->prof_io_ops;
497 if (ops->fill_io_u_size)
498 return ops->fill_io_u_size(td, io_u, is_random);
501 return __get_next_buflen(td, io_u, is_random);
504 static void set_rwmix_bytes(struct thread_data *td)
509 * we do time or byte based switch. this is needed because
510 * buffered writes may issue a lot quicker than they complete,
511 * whereas reads do not.
513 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
514 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
517 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
522 if (td->o.use_os_rand) {
523 r = os_random_long(&td->rwmix_state);
524 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
526 r = __rand(&td->__rwmix_state);
527 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
530 if (v <= td->o.rwmix[DDIR_READ])
536 void io_u_quiesce(struct thread_data *td)
539 * We are going to sleep, ensure that we flush anything pending as
540 * not to skew our latency numbers.
542 * Changed to only monitor 'in flight' requests here instead of the
543 * td->cur_depth, b/c td->cur_depth does not accurately represent
544 * io's that have been actually submitted to an async engine,
545 * and cur_depth is meaningless for sync engines.
547 while (td->io_u_in_flight) {
550 ret = io_u_queued_complete(td, 1, NULL);
554 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
556 enum fio_ddir odir = ddir ^ 1;
560 assert(ddir_rw(ddir));
562 if (td->rate_pending_usleep[ddir] <= 0)
566 * We have too much pending sleep in this direction. See if we
569 if (td_rw(td) && td->o.rwmix[odir]) {
571 * Other direction does not have too much pending, switch
573 if (td->rate_pending_usleep[odir] < 100000)
577 * Both directions have pending sleep. Sleep the minimum time
578 * and deduct from both.
580 if (td->rate_pending_usleep[ddir] <=
581 td->rate_pending_usleep[odir]) {
582 usec = td->rate_pending_usleep[ddir];
584 usec = td->rate_pending_usleep[odir];
588 usec = td->rate_pending_usleep[ddir];
592 fio_gettime(&t, NULL);
593 usec_sleep(td, usec);
594 usec = utime_since_now(&t);
596 td->rate_pending_usleep[ddir] -= usec;
599 if (td_rw(td) && __should_check_rate(td, odir))
600 td->rate_pending_usleep[odir] -= usec;
609 * Return the data direction for the next io_u. If the job is a
610 * mixed read/write workload, check the rwmix cycle and switch if
613 static enum fio_ddir get_rw_ddir(struct thread_data *td)
618 * see if it's time to fsync
620 if (td->o.fsync_blocks &&
621 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
622 td->io_issues[DDIR_WRITE] && should_fsync(td))
626 * see if it's time to fdatasync
628 if (td->o.fdatasync_blocks &&
629 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
630 td->io_issues[DDIR_WRITE] && should_fsync(td))
631 return DDIR_DATASYNC;
634 * see if it's time to sync_file_range
636 if (td->sync_file_range_nr &&
637 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
638 td->io_issues[DDIR_WRITE] && should_fsync(td))
639 return DDIR_SYNC_FILE_RANGE;
643 * Check if it's time to seed a new data direction.
645 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
647 * Put a top limit on how many bytes we do for
648 * one data direction, to avoid overflowing the
651 ddir = get_rand_ddir(td);
653 if (ddir != td->rwmix_ddir)
656 td->rwmix_ddir = ddir;
658 ddir = td->rwmix_ddir;
659 } else if (td_read(td))
661 else if (td_write(td))
666 td->rwmix_ddir = rate_ddir(td, ddir);
667 return td->rwmix_ddir;
670 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
672 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
674 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
675 td->o.barrier_blocks &&
676 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
677 td->io_issues[DDIR_WRITE])
678 io_u->flags |= IO_U_F_BARRIER;
681 void put_file_log(struct thread_data *td, struct fio_file *f)
683 int ret = put_file(td, f);
686 td_verror(td, ret, "file close");
689 void put_io_u(struct thread_data *td, struct io_u *io_u)
693 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
694 put_file_log(td, io_u->file);
696 io_u->flags &= ~IO_U_F_FREE_DEF;
697 io_u->flags |= IO_U_F_FREE;
699 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
701 io_u_qpush(&td->io_u_freelist, io_u);
703 td_io_u_free_notify(td);
706 void clear_io_u(struct thread_data *td, struct io_u *io_u)
708 io_u->flags &= ~IO_U_F_FLIGHT;
712 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
714 struct io_u *__io_u = *io_u;
715 enum fio_ddir ddir = acct_ddir(__io_u);
717 dprint(FD_IO, "requeue %p\n", __io_u);
721 __io_u->flags |= IO_U_F_FREE;
722 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
723 td->io_issues[ddir]--;
725 __io_u->flags &= ~IO_U_F_FLIGHT;
726 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
729 io_u_rpush(&td->io_u_requeues, __io_u);
734 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
736 unsigned int is_random;
738 if (td->io_ops->flags & FIO_NOIO)
741 set_rw_ddir(td, io_u);
744 * fsync() or fdatasync() or trim etc, we are done
746 if (!ddir_rw(io_u->ddir))
750 * See if it's time to switch to a new zone
752 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
754 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
755 io_u->file->last_pos = io_u->file->file_offset;
756 td->io_skip_bytes += td->o.zone_skip;
760 * No log, let the seq/rand engine retrieve the next buflen and
763 if (get_next_offset(td, io_u, &is_random)) {
764 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
768 io_u->buflen = get_next_buflen(td, io_u, is_random);
770 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
774 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
775 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
776 dprint(FD_IO, " off=%llu/%lu > %llu\n",
777 (unsigned long long) io_u->offset, io_u->buflen,
778 (unsigned long long) io_u->file->real_file_size);
783 * mark entry before potentially trimming io_u
785 if (td_random(td) && file_randommap(td, io_u->file))
786 mark_random_map(td, io_u);
789 dprint_io_u(io_u, "fill_io_u");
790 td->zone_bytes += io_u->buflen;
794 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
823 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
825 __io_u_mark_map(td->ts.io_u_submit, nr);
826 td->ts.total_submit++;
829 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
831 __io_u_mark_map(td->ts.io_u_complete, nr);
832 td->ts.total_complete++;
835 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
839 switch (td->cur_depth) {
861 td->ts.io_u_map[idx] += nr;
864 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
901 assert(idx < FIO_IO_U_LAT_U_NR);
902 td->ts.io_u_lat_u[idx]++;
905 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
946 assert(idx < FIO_IO_U_LAT_M_NR);
947 td->ts.io_u_lat_m[idx]++;
950 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
953 io_u_mark_lat_usec(td, usec);
955 io_u_mark_lat_msec(td, usec / 1000);
959 * Get next file to service by choosing one at random
961 static struct fio_file *get_next_file_rand(struct thread_data *td,
962 enum fio_file_flags goodf,
963 enum fio_file_flags badf)
972 if (td->o.use_os_rand) {
973 r = os_random_long(&td->next_file_state);
974 fno = (unsigned int) ((double) td->o.nr_files
975 * (r / (OS_RAND_MAX + 1.0)));
977 r = __rand(&td->__next_file_state);
978 fno = (unsigned int) ((double) td->o.nr_files
979 * (r / (FRAND_MAX + 1.0)));
983 if (fio_file_done(f))
986 if (!fio_file_open(f)) {
989 if (td->nr_open_files >= td->o.open_files)
990 return ERR_PTR(-EBUSY);
992 err = td_io_open_file(td, f);
998 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
999 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1003 td_io_close_file(td, f);
1008 * Get next file to service by doing round robin between all available ones
1010 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1013 unsigned int old_next_file = td->next_file;
1019 f = td->files[td->next_file];
1022 if (td->next_file >= td->o.nr_files)
1025 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1026 if (fio_file_done(f)) {
1031 if (!fio_file_open(f)) {
1034 if (td->nr_open_files >= td->o.open_files)
1035 return ERR_PTR(-EBUSY);
1037 err = td_io_open_file(td, f);
1039 dprint(FD_FILE, "error %d on open of %s\n",
1047 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1049 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1053 td_io_close_file(td, f);
1056 } while (td->next_file != old_next_file);
1058 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1062 static struct fio_file *__get_next_file(struct thread_data *td)
1066 assert(td->o.nr_files <= td->files_index);
1068 if (td->nr_done_files >= td->o.nr_files) {
1069 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1070 " nr_files=%d\n", td->nr_open_files,
1076 f = td->file_service_file;
1077 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1078 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1080 if (td->file_service_left--)
1084 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1085 td->o.file_service_type == FIO_FSERVICE_SEQ)
1086 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1088 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1093 td->file_service_file = f;
1094 td->file_service_left = td->file_service_nr - 1;
1097 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1099 dprint(FD_FILE, "get_next_file: NULL\n");
1103 static struct fio_file *get_next_file(struct thread_data *td)
1105 if (!(td->flags & TD_F_PROFILE_OPS)) {
1106 struct prof_io_ops *ops = &td->prof_io_ops;
1108 if (ops->get_next_file)
1109 return ops->get_next_file(td);
1112 return __get_next_file(td);
1115 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1120 f = get_next_file(td);
1121 if (IS_ERR_OR_NULL(f))
1127 if (!fill_io_u(td, io_u))
1130 put_file_log(td, f);
1131 td_io_close_file(td, f);
1133 fio_file_set_done(f);
1134 td->nr_done_files++;
1135 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1136 td->nr_done_files, td->o.nr_files);
1142 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1143 unsigned long tusec, unsigned long max_usec)
1146 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1147 td_verror(td, ETIMEDOUT, "max latency exceeded");
1148 icd->error = ETIMEDOUT;
1151 static void lat_new_cycle(struct thread_data *td)
1153 fio_gettime(&td->latency_ts, NULL);
1154 td->latency_ios = ddir_rw_sum(td->io_blocks);
1155 td->latency_failed = 0;
1159 * We had an IO outside the latency target. Reduce the queue depth. If we
1160 * are at QD=1, then it's time to give up.
1162 static int __lat_target_failed(struct thread_data *td)
1164 if (td->latency_qd == 1)
1167 td->latency_qd_high = td->latency_qd;
1169 if (td->latency_qd == td->latency_qd_low)
1170 td->latency_qd_low--;
1172 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1174 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1177 * When we ramp QD down, quiesce existing IO to prevent
1178 * a storm of ramp downs due to pending higher depth.
1185 static int lat_target_failed(struct thread_data *td)
1187 if (td->o.latency_percentile.u.f == 100.0)
1188 return __lat_target_failed(td);
1190 td->latency_failed++;
1194 void lat_target_init(struct thread_data *td)
1196 td->latency_end_run = 0;
1198 if (td->o.latency_target) {
1199 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1200 fio_gettime(&td->latency_ts, NULL);
1202 td->latency_qd_high = td->o.iodepth;
1203 td->latency_qd_low = 1;
1204 td->latency_ios = ddir_rw_sum(td->io_blocks);
1206 td->latency_qd = td->o.iodepth;
1209 void lat_target_reset(struct thread_data *td)
1211 if (!td->latency_end_run)
1212 lat_target_init(td);
1215 static void lat_target_success(struct thread_data *td)
1217 const unsigned int qd = td->latency_qd;
1218 struct thread_options *o = &td->o;
1220 td->latency_qd_low = td->latency_qd;
1223 * If we haven't failed yet, we double up to a failing value instead
1224 * of bisecting from highest possible queue depth. If we have set
1225 * a limit other than td->o.iodepth, bisect between that.
1227 if (td->latency_qd_high != o->iodepth)
1228 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1230 td->latency_qd *= 2;
1232 if (td->latency_qd > o->iodepth)
1233 td->latency_qd = o->iodepth;
1235 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1238 * Same as last one, we are done. Let it run a latency cycle, so
1239 * we get only the results from the targeted depth.
1241 if (td->latency_qd == qd) {
1242 if (td->latency_end_run) {
1243 dprint(FD_RATE, "We are done\n");
1246 dprint(FD_RATE, "Quiesce and final run\n");
1248 td->latency_end_run = 1;
1249 reset_all_stats(td);
1258 * Check if we can bump the queue depth
1260 void lat_target_check(struct thread_data *td)
1262 uint64_t usec_window;
1266 usec_window = utime_since_now(&td->latency_ts);
1267 if (usec_window < td->o.latency_window)
1270 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1271 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1272 success_ios *= 100.0;
1274 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1276 if (success_ios >= td->o.latency_percentile.u.f)
1277 lat_target_success(td);
1279 __lat_target_failed(td);
1283 * If latency target is enabled, we might be ramping up or down and not
1284 * using the full queue depth available.
1286 int queue_full(struct thread_data *td)
1288 const int qempty = io_u_qempty(&td->io_u_freelist);
1292 if (!td->o.latency_target)
1295 return td->cur_depth >= td->latency_qd;
1298 struct io_u *__get_io_u(struct thread_data *td)
1300 struct io_u *io_u = NULL;
1305 if (!io_u_rempty(&td->io_u_requeues))
1306 io_u = io_u_rpop(&td->io_u_requeues);
1307 else if (!queue_full(td)) {
1308 io_u = io_u_qpop(&td->io_u_freelist);
1313 io_u->end_io = NULL;
1317 assert(io_u->flags & IO_U_F_FREE);
1318 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1319 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1320 io_u->flags &= ~IO_U_F_VER_LIST;
1323 io_u->acct_ddir = -1;
1325 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1327 } else if (td->o.verify_async) {
1329 * We ran out, wait for async verify threads to finish and
1332 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1340 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1342 if (!(td->flags & TD_F_TRIM_BACKLOG))
1345 if (td->trim_entries) {
1348 if (td->trim_batch) {
1351 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1352 td->last_ddir != DDIR_READ) {
1353 td->trim_batch = td->o.trim_batch;
1354 if (!td->trim_batch)
1355 td->trim_batch = td->o.trim_backlog;
1359 if (get_trim && !get_next_trim(td, io_u))
1366 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1368 if (!(td->flags & TD_F_VER_BACKLOG))
1371 if (td->io_hist_len) {
1374 if (td->verify_batch)
1376 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1377 td->last_ddir != DDIR_READ) {
1378 td->verify_batch = td->o.verify_batch;
1379 if (!td->verify_batch)
1380 td->verify_batch = td->o.verify_backlog;
1384 if (get_verify && !get_next_verify(td, io_u)) {
1394 * Fill offset and start time into the buffer content, to prevent too
1395 * easy compressible data for simple de-dupe attempts. Do this for every
1396 * 512b block in the range, since that should be the smallest block size
1397 * we can expect from a device.
1399 static void small_content_scramble(struct io_u *io_u)
1401 unsigned int i, nr_blocks = io_u->buflen / 512;
1403 unsigned int offset;
1410 boffset = io_u->offset;
1411 io_u->buf_filled_len = 0;
1413 for (i = 0; i < nr_blocks; i++) {
1415 * Fill the byte offset into a "random" start offset of
1416 * the buffer, given by the product of the usec time
1417 * and the actual offset.
1419 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1420 offset &= ~(sizeof(uint64_t) - 1);
1421 if (offset >= 512 - sizeof(uint64_t))
1422 offset -= sizeof(uint64_t);
1423 memcpy(p + offset, &boffset, sizeof(boffset));
1425 end = p + 512 - sizeof(io_u->start_time);
1426 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1433 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1434 * etc. The returned io_u is fully ready to be prepped and submitted.
1436 struct io_u *get_io_u(struct thread_data *td)
1440 int do_scramble = 0;
1443 io_u = __get_io_u(td);
1445 dprint(FD_IO, "__get_io_u failed\n");
1449 if (check_get_verify(td, io_u))
1451 if (check_get_trim(td, io_u))
1455 * from a requeue, io_u already setup
1461 * If using an iolog, grab next piece if any available.
1463 if (td->flags & TD_F_READ_IOLOG) {
1464 if (read_iolog_get(td, io_u))
1466 } else if (set_io_u_file(td, io_u)) {
1468 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1474 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1478 assert(fio_file_open(f));
1480 if (ddir_rw(io_u->ddir)) {
1481 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1482 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1486 f->last_start = io_u->offset;
1487 f->last_pos = io_u->offset + io_u->buflen;
1489 if (io_u->ddir == DDIR_WRITE) {
1490 if (td->flags & TD_F_REFILL_BUFFERS) {
1491 io_u_fill_buffer(td, io_u,
1492 io_u->xfer_buflen, io_u->xfer_buflen);
1493 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1494 !(td->flags & TD_F_COMPRESS))
1496 if (td->flags & TD_F_VER_NONE) {
1497 populate_verify_io_u(td, io_u);
1500 } else if (io_u->ddir == DDIR_READ) {
1502 * Reset the buf_filled parameters so next time if the
1503 * buffer is used for writes it is refilled.
1505 io_u->buf_filled_len = 0;
1510 * Set io data pointers.
1512 io_u->xfer_buf = io_u->buf;
1513 io_u->xfer_buflen = io_u->buflen;
1517 if (!td_io_prep(td, io_u)) {
1518 if (!td->o.disable_slat)
1519 fio_gettime(&io_u->start_time, NULL);
1521 small_content_scramble(io_u);
1525 dprint(FD_IO, "get_io_u failed\n");
1527 return ERR_PTR(ret);
1530 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1532 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1533 const char *msg[] = { "read", "write", "sync", "datasync",
1534 "sync_file_range", "wait", "trim" };
1536 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1539 log_err("fio: io_u error");
1542 log_err(" on file %s", io_u->file->file_name);
1544 log_err(": %s\n", strerror(io_u->error));
1546 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1547 io_u->offset, io_u->xfer_buflen);
1550 td_verror(td, io_u->error, "io_u error");
1553 static inline int gtod_reduce(struct thread_data *td)
1555 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1556 && td->o.disable_bw;
1559 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1560 struct io_completion_data *icd,
1561 const enum fio_ddir idx, unsigned int bytes)
1563 unsigned long lusec = 0;
1565 if (!gtod_reduce(td))
1566 lusec = utime_since(&io_u->issue_time, &icd->time);
1568 if (!td->o.disable_lat) {
1569 unsigned long tusec;
1571 tusec = utime_since(&io_u->start_time, &icd->time);
1572 add_lat_sample(td, idx, tusec, bytes);
1574 if (td->flags & TD_F_PROFILE_OPS) {
1575 struct prof_io_ops *ops = &td->prof_io_ops;
1578 icd->error = ops->io_u_lat(td, tusec);
1581 if (td->o.max_latency && tusec > td->o.max_latency)
1582 lat_fatal(td, icd, tusec, td->o.max_latency);
1583 if (td->o.latency_target && tusec > td->o.latency_target) {
1584 if (lat_target_failed(td))
1585 lat_fatal(td, icd, tusec, td->o.latency_target);
1589 if (!td->o.disable_clat) {
1590 add_clat_sample(td, idx, lusec, bytes);
1591 io_u_mark_latency(td, lusec);
1594 if (!td->o.disable_bw)
1595 add_bw_sample(td, idx, bytes, &icd->time);
1597 if (!gtod_reduce(td))
1598 add_iops_sample(td, idx, bytes, &icd->time);
1601 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1603 uint64_t secs, remainder, bps, bytes;
1605 bytes = td->this_io_bytes[ddir];
1606 bps = td->rate_bps[ddir];
1608 remainder = bytes % bps;
1609 return remainder * 1000000 / bps + secs * 1000000;
1612 static void io_completed(struct thread_data *td, struct io_u *io_u,
1613 struct io_completion_data *icd)
1617 dprint_io_u(io_u, "io complete");
1620 assert(io_u->flags & IO_U_F_FLIGHT);
1621 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1624 * Mark IO ok to verify
1627 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1633 if (ddir_sync(io_u->ddir)) {
1634 td->last_was_sync = 1;
1637 f->first_write = -1ULL;
1638 f->last_write = -1ULL;
1643 td->last_was_sync = 0;
1644 td->last_ddir = io_u->ddir;
1646 if (!io_u->error && ddir_rw(io_u->ddir)) {
1647 unsigned int bytes = io_u->buflen - io_u->resid;
1648 const enum fio_ddir idx = io_u->ddir;
1649 const enum fio_ddir odx = io_u->ddir ^ 1;
1652 td->io_blocks[idx]++;
1653 td->this_io_blocks[idx]++;
1654 td->io_bytes[idx] += bytes;
1656 if (!(io_u->flags & IO_U_F_VER_LIST))
1657 td->this_io_bytes[idx] += bytes;
1659 if (idx == DDIR_WRITE) {
1662 if (f->first_write == -1ULL ||
1663 io_u->offset < f->first_write)
1664 f->first_write = io_u->offset;
1665 if (f->last_write == -1ULL ||
1666 ((io_u->offset + bytes) > f->last_write))
1667 f->last_write = io_u->offset + bytes;
1671 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1672 td->runstate == TD_VERIFYING)) {
1673 account_io_completion(td, io_u, icd, idx, bytes);
1675 if (__should_check_rate(td, idx)) {
1676 td->rate_pending_usleep[idx] =
1677 (usec_for_io(td, idx) -
1678 utime_since_now(&td->start));
1680 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1681 td->rate_pending_usleep[odx] =
1682 (usec_for_io(td, odx) -
1683 utime_since_now(&td->start));
1686 icd->bytes_done[idx] += bytes;
1689 ret = io_u->end_io(td, io_u);
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(io_u->ddir, icd->error);
1699 if (!td_non_fatal_error(td, eb, icd->error))
1702 * If there is a non_fatal error, then add to the error count
1703 * and clear all the errors.
1705 update_error_count(td, icd->error);
1712 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1717 if (!gtod_reduce(td))
1718 fio_gettime(&icd->time, NULL);
1723 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1724 icd->bytes_done[ddir] = 0;
1727 static void ios_completed(struct thread_data *td,
1728 struct io_completion_data *icd)
1733 for (i = 0; i < icd->nr; i++) {
1734 io_u = td->io_ops->event(td, i);
1736 io_completed(td, io_u, icd);
1738 if (!(io_u->flags & IO_U_F_FREE_DEF))
1744 * Complete a single io_u for the sync engines.
1746 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1749 struct io_completion_data icd;
1751 init_icd(td, &icd, 1);
1752 io_completed(td, io_u, &icd);
1754 if (!(io_u->flags & IO_U_F_FREE_DEF))
1758 td_verror(td, icd.error, "io_u_sync_complete");
1765 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1766 bytes[ddir] += icd.bytes_done[ddir];
1773 * Called to complete min_events number of io for the async engines.
1775 int io_u_queued_complete(struct thread_data *td, int min_evts,
1778 struct io_completion_data icd;
1779 struct timespec *tvp = NULL;
1781 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1783 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1788 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1790 td_verror(td, -ret, "td_io_getevents");
1795 init_icd(td, &icd, ret);
1796 ios_completed(td, &icd);
1798 td_verror(td, icd.error, "io_u_queued_complete");
1805 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1806 bytes[ddir] += icd.bytes_done[ddir];
1813 * Call when io_u is really queued, to update the submission latency.
1815 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1817 if (!td->o.disable_slat) {
1818 unsigned long slat_time;
1820 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1821 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1825 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1826 unsigned int max_bs)
1828 if (td->o.buffer_pattern_bytes)
1829 fill_buffer_pattern(td, buf, max_bs);
1830 else if (!td->o.zero_buffers) {
1831 unsigned int perc = td->o.compress_percentage;
1834 unsigned int seg = min_write;
1836 seg = min(min_write, td->o.compress_chunk);
1840 fill_random_buf_percentage(&td->buf_state, buf,
1843 fill_random_buf(&td->buf_state, buf, max_bs);
1845 memset(buf, 0, max_bs);
1849 * "randomly" fill the buffer contents
1851 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1852 unsigned int min_write, unsigned int max_bs)
1854 io_u->buf_filled_len = 0;
1855 fill_io_buffer(td, io_u->buf, min_write, max_bs);