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_first_entry(&td->next_rand_list, 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 struct thread_options *o = &td->o;
276 assert(ddir_rw(ddir));
278 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
280 f->last_pos = f->last_pos - f->io_size;
282 if (f->last_pos < f->real_file_size) {
285 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
286 f->last_pos = f->real_file_size;
288 pos = f->last_pos - f->file_offset;
289 if (pos && o->ddir_seq_add) {
290 pos += o->ddir_seq_add;
293 * If we reach beyond the end of the file
294 * with holed IO, wrap around to the
297 if (pos >= f->real_file_size)
298 pos = f->file_offset;
308 static int get_next_block(struct thread_data *td, struct io_u *io_u,
309 enum fio_ddir ddir, int rw_seq,
310 unsigned int *is_random)
312 struct fio_file *f = io_u->file;
316 assert(ddir_rw(ddir));
322 if (should_do_random(td, ddir)) {
323 ret = get_next_rand_block(td, f, ddir, &b);
327 io_u->flags |= IO_U_F_BUSY_OK;
328 ret = get_next_seq_offset(td, f, ddir, &offset);
330 ret = get_next_rand_block(td, f, ddir, &b);
334 ret = get_next_seq_offset(td, f, ddir, &offset);
337 io_u->flags |= IO_U_F_BUSY_OK;
340 if (td->o.rw_seq == RW_SEQ_SEQ) {
341 ret = get_next_seq_offset(td, f, ddir, &offset);
343 ret = get_next_rand_block(td, f, ddir, &b);
346 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
347 if (f->last_start != -1ULL)
348 offset = f->last_start - f->file_offset;
353 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
360 io_u->offset = offset;
362 io_u->offset = b * td->o.ba[ddir];
364 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
373 * For random io, generate a random new block and see if it's used. Repeat
374 * until we find a free one. For sequential io, just return the end of
375 * the last io issued.
377 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
378 unsigned int *is_random)
380 struct fio_file *f = io_u->file;
381 enum fio_ddir ddir = io_u->ddir;
384 assert(ddir_rw(ddir));
386 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
388 td->ddir_seq_nr = td->o.ddir_seq_nr;
391 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
394 if (io_u->offset >= f->io_size) {
395 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
396 (unsigned long long) io_u->offset,
397 (unsigned long long) f->io_size);
401 io_u->offset += f->file_offset;
402 if (io_u->offset >= f->real_file_size) {
403 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
404 (unsigned long long) io_u->offset,
405 (unsigned long long) f->real_file_size);
412 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
413 unsigned int *is_random)
415 if (td->flags & TD_F_PROFILE_OPS) {
416 struct prof_io_ops *ops = &td->prof_io_ops;
418 if (ops->fill_io_u_off)
419 return ops->fill_io_u_off(td, io_u, is_random);
422 return __get_next_offset(td, io_u, is_random);
425 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
428 struct fio_file *f = io_u->file;
430 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
433 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
434 unsigned int is_random)
436 int ddir = io_u->ddir;
437 unsigned int buflen = 0;
438 unsigned int minbs, maxbs;
439 unsigned long r, rand_max;
441 assert(ddir_rw(ddir));
443 if (td->o.bs_is_seq_rand)
444 ddir = is_random ? DDIR_WRITE: DDIR_READ;
446 minbs = td->o.min_bs[ddir];
447 maxbs = td->o.max_bs[ddir];
453 * If we can't satisfy the min block size from here, then fail
455 if (!io_u_fits(td, io_u, minbs))
458 if (td->o.use_os_rand)
459 rand_max = OS_RAND_MAX;
461 rand_max = FRAND_MAX;
464 if (td->o.use_os_rand)
465 r = os_random_long(&td->bsrange_state);
467 r = __rand(&td->__bsrange_state);
469 if (!td->o.bssplit_nr[ddir]) {
470 buflen = 1 + (unsigned int) ((double) maxbs *
471 (r / (rand_max + 1.0)));
478 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
479 struct bssplit *bsp = &td->o.bssplit[ddir][i];
483 if ((r <= ((rand_max / 100L) * perc)) &&
484 io_u_fits(td, io_u, buflen))
489 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
490 buflen = (buflen + td->o.verify_interval - 1) &
491 ~(td->o.verify_interval - 1);
493 if (!td->o.bs_unaligned && is_power_of_2(minbs))
494 buflen = (buflen + minbs - 1) & ~(minbs - 1);
496 } while (!io_u_fits(td, io_u, buflen));
501 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
502 unsigned int is_random)
504 if (td->flags & TD_F_PROFILE_OPS) {
505 struct prof_io_ops *ops = &td->prof_io_ops;
507 if (ops->fill_io_u_size)
508 return ops->fill_io_u_size(td, io_u, is_random);
511 return __get_next_buflen(td, io_u, is_random);
514 static void set_rwmix_bytes(struct thread_data *td)
519 * we do time or byte based switch. this is needed because
520 * buffered writes may issue a lot quicker than they complete,
521 * whereas reads do not.
523 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
524 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
527 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
532 if (td->o.use_os_rand) {
533 r = os_random_long(&td->rwmix_state);
534 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
536 r = __rand(&td->__rwmix_state);
537 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
540 if (v <= td->o.rwmix[DDIR_READ])
546 void io_u_quiesce(struct thread_data *td)
549 * We are going to sleep, ensure that we flush anything pending as
550 * not to skew our latency numbers.
552 * Changed to only monitor 'in flight' requests here instead of the
553 * td->cur_depth, b/c td->cur_depth does not accurately represent
554 * io's that have been actually submitted to an async engine,
555 * and cur_depth is meaningless for sync engines.
557 while (td->io_u_in_flight) {
560 ret = io_u_queued_complete(td, 1, NULL);
564 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
566 enum fio_ddir odir = ddir ^ 1;
570 assert(ddir_rw(ddir));
572 if (td->rate_pending_usleep[ddir] <= 0)
576 * We have too much pending sleep in this direction. See if we
579 if (td_rw(td) && td->o.rwmix[odir]) {
581 * Other direction does not have too much pending, switch
583 if (td->rate_pending_usleep[odir] < 100000)
587 * Both directions have pending sleep. Sleep the minimum time
588 * and deduct from both.
590 if (td->rate_pending_usleep[ddir] <=
591 td->rate_pending_usleep[odir]) {
592 usec = td->rate_pending_usleep[ddir];
594 usec = td->rate_pending_usleep[odir];
598 usec = td->rate_pending_usleep[ddir];
602 fio_gettime(&t, NULL);
603 usec_sleep(td, usec);
604 usec = utime_since_now(&t);
606 td->rate_pending_usleep[ddir] -= usec;
609 if (td_rw(td) && __should_check_rate(td, odir))
610 td->rate_pending_usleep[odir] -= usec;
619 * Return the data direction for the next io_u. If the job is a
620 * mixed read/write workload, check the rwmix cycle and switch if
623 static enum fio_ddir get_rw_ddir(struct thread_data *td)
628 * see if it's time to fsync
630 if (td->o.fsync_blocks &&
631 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
632 td->io_issues[DDIR_WRITE] && should_fsync(td))
636 * see if it's time to fdatasync
638 if (td->o.fdatasync_blocks &&
639 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
640 td->io_issues[DDIR_WRITE] && should_fsync(td))
641 return DDIR_DATASYNC;
644 * see if it's time to sync_file_range
646 if (td->sync_file_range_nr &&
647 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
648 td->io_issues[DDIR_WRITE] && should_fsync(td))
649 return DDIR_SYNC_FILE_RANGE;
653 * Check if it's time to seed a new data direction.
655 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
657 * Put a top limit on how many bytes we do for
658 * one data direction, to avoid overflowing the
661 ddir = get_rand_ddir(td);
663 if (ddir != td->rwmix_ddir)
666 td->rwmix_ddir = ddir;
668 ddir = td->rwmix_ddir;
669 } else if (td_read(td))
671 else if (td_write(td))
676 td->rwmix_ddir = rate_ddir(td, ddir);
677 return td->rwmix_ddir;
680 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
682 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
684 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
685 td->o.barrier_blocks &&
686 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
687 td->io_issues[DDIR_WRITE])
688 io_u->flags |= IO_U_F_BARRIER;
691 void put_file_log(struct thread_data *td, struct fio_file *f)
693 unsigned int ret = put_file(td, f);
696 td_verror(td, ret, "file close");
699 void put_io_u(struct thread_data *td, struct io_u *io_u)
703 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
704 put_file_log(td, io_u->file);
707 io_u->flags |= IO_U_F_FREE;
709 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
711 io_u_qpush(&td->io_u_freelist, io_u);
713 td_io_u_free_notify(td);
716 void clear_io_u(struct thread_data *td, struct io_u *io_u)
718 io_u->flags &= ~IO_U_F_FLIGHT;
722 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
724 struct io_u *__io_u = *io_u;
725 enum fio_ddir ddir = acct_ddir(__io_u);
727 dprint(FD_IO, "requeue %p\n", __io_u);
731 __io_u->flags |= IO_U_F_FREE;
732 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
733 td->io_issues[ddir]--;
735 __io_u->flags &= ~IO_U_F_FLIGHT;
736 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
739 io_u_rpush(&td->io_u_requeues, __io_u);
744 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
746 unsigned int is_random;
748 if (td->io_ops->flags & FIO_NOIO)
751 set_rw_ddir(td, io_u);
754 * fsync() or fdatasync() or trim etc, we are done
756 if (!ddir_rw(io_u->ddir))
760 * See if it's time to switch to a new zone
762 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
763 struct fio_file *f = io_u->file;
766 f->file_offset += td->o.zone_range + td->o.zone_skip;
769 * Wrap from the beginning, if we exceed the file size
771 if (f->file_offset >= f->real_file_size)
772 f->file_offset = f->real_file_size - f->file_offset;
773 f->last_pos = f->file_offset;
774 td->io_skip_bytes += td->o.zone_skip;
778 * No log, let the seq/rand engine retrieve the next buflen and
781 if (get_next_offset(td, io_u, &is_random)) {
782 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
786 io_u->buflen = get_next_buflen(td, io_u, is_random);
788 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
792 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
793 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
794 dprint(FD_IO, " off=%llu/%lu > %llu\n",
795 (unsigned long long) io_u->offset, io_u->buflen,
796 (unsigned long long) io_u->file->real_file_size);
801 * mark entry before potentially trimming io_u
803 if (td_random(td) && file_randommap(td, io_u->file))
804 mark_random_map(td, io_u);
807 dprint_io_u(io_u, "fill_io_u");
808 td->zone_bytes += io_u->buflen;
812 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
841 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
843 __io_u_mark_map(td->ts.io_u_submit, nr);
844 td->ts.total_submit++;
847 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
849 __io_u_mark_map(td->ts.io_u_complete, nr);
850 td->ts.total_complete++;
853 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
857 switch (td->cur_depth) {
879 td->ts.io_u_map[idx] += nr;
882 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
919 assert(idx < FIO_IO_U_LAT_U_NR);
920 td->ts.io_u_lat_u[idx]++;
923 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
964 assert(idx < FIO_IO_U_LAT_M_NR);
965 td->ts.io_u_lat_m[idx]++;
968 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
971 io_u_mark_lat_usec(td, usec);
973 io_u_mark_lat_msec(td, usec / 1000);
977 * Get next file to service by choosing one at random
979 static struct fio_file *get_next_file_rand(struct thread_data *td,
980 enum fio_file_flags goodf,
981 enum fio_file_flags badf)
990 if (td->o.use_os_rand) {
991 r = os_random_long(&td->next_file_state);
992 fno = (unsigned int) ((double) td->o.nr_files
993 * (r / (OS_RAND_MAX + 1.0)));
995 r = __rand(&td->__next_file_state);
996 fno = (unsigned int) ((double) td->o.nr_files
997 * (r / (FRAND_MAX + 1.0)));
1001 if (fio_file_done(f))
1004 if (!fio_file_open(f)) {
1007 if (td->nr_open_files >= td->o.open_files)
1008 return ERR_PTR(-EBUSY);
1010 err = td_io_open_file(td, f);
1016 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1017 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1021 td_io_close_file(td, f);
1026 * Get next file to service by doing round robin between all available ones
1028 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1031 unsigned int old_next_file = td->next_file;
1037 f = td->files[td->next_file];
1040 if (td->next_file >= td->o.nr_files)
1043 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1044 if (fio_file_done(f)) {
1049 if (!fio_file_open(f)) {
1052 if (td->nr_open_files >= td->o.open_files)
1053 return ERR_PTR(-EBUSY);
1055 err = td_io_open_file(td, f);
1057 dprint(FD_FILE, "error %d on open of %s\n",
1065 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1067 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1071 td_io_close_file(td, f);
1074 } while (td->next_file != old_next_file);
1076 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1080 static struct fio_file *__get_next_file(struct thread_data *td)
1084 assert(td->o.nr_files <= td->files_index);
1086 if (td->nr_done_files >= td->o.nr_files) {
1087 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1088 " nr_files=%d\n", td->nr_open_files,
1094 f = td->file_service_file;
1095 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1096 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1098 if (td->file_service_left--)
1102 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1103 td->o.file_service_type == FIO_FSERVICE_SEQ)
1104 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1106 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1111 td->file_service_file = f;
1112 td->file_service_left = td->file_service_nr - 1;
1115 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1117 dprint(FD_FILE, "get_next_file: NULL\n");
1121 static struct fio_file *get_next_file(struct thread_data *td)
1123 if (td->flags & TD_F_PROFILE_OPS) {
1124 struct prof_io_ops *ops = &td->prof_io_ops;
1126 if (ops->get_next_file)
1127 return ops->get_next_file(td);
1130 return __get_next_file(td);
1133 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1138 f = get_next_file(td);
1139 if (IS_ERR_OR_NULL(f))
1145 if (!fill_io_u(td, io_u))
1148 put_file_log(td, f);
1149 td_io_close_file(td, f);
1151 fio_file_set_done(f);
1152 td->nr_done_files++;
1153 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1154 td->nr_done_files, td->o.nr_files);
1160 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1161 unsigned long tusec, unsigned long max_usec)
1164 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1165 td_verror(td, ETIMEDOUT, "max latency exceeded");
1166 icd->error = ETIMEDOUT;
1169 static void lat_new_cycle(struct thread_data *td)
1171 fio_gettime(&td->latency_ts, NULL);
1172 td->latency_ios = ddir_rw_sum(td->io_blocks);
1173 td->latency_failed = 0;
1177 * We had an IO outside the latency target. Reduce the queue depth. If we
1178 * are at QD=1, then it's time to give up.
1180 static int __lat_target_failed(struct thread_data *td)
1182 if (td->latency_qd == 1)
1185 td->latency_qd_high = td->latency_qd;
1187 if (td->latency_qd == td->latency_qd_low)
1188 td->latency_qd_low--;
1190 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1192 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1195 * When we ramp QD down, quiesce existing IO to prevent
1196 * a storm of ramp downs due to pending higher depth.
1203 static int lat_target_failed(struct thread_data *td)
1205 if (td->o.latency_percentile.u.f == 100.0)
1206 return __lat_target_failed(td);
1208 td->latency_failed++;
1212 void lat_target_init(struct thread_data *td)
1214 td->latency_end_run = 0;
1216 if (td->o.latency_target) {
1217 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1218 fio_gettime(&td->latency_ts, NULL);
1220 td->latency_qd_high = td->o.iodepth;
1221 td->latency_qd_low = 1;
1222 td->latency_ios = ddir_rw_sum(td->io_blocks);
1224 td->latency_qd = td->o.iodepth;
1227 void lat_target_reset(struct thread_data *td)
1229 if (!td->latency_end_run)
1230 lat_target_init(td);
1233 static void lat_target_success(struct thread_data *td)
1235 const unsigned int qd = td->latency_qd;
1236 struct thread_options *o = &td->o;
1238 td->latency_qd_low = td->latency_qd;
1241 * If we haven't failed yet, we double up to a failing value instead
1242 * of bisecting from highest possible queue depth. If we have set
1243 * a limit other than td->o.iodepth, bisect between that.
1245 if (td->latency_qd_high != o->iodepth)
1246 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1248 td->latency_qd *= 2;
1250 if (td->latency_qd > o->iodepth)
1251 td->latency_qd = o->iodepth;
1253 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1256 * Same as last one, we are done. Let it run a latency cycle, so
1257 * we get only the results from the targeted depth.
1259 if (td->latency_qd == qd) {
1260 if (td->latency_end_run) {
1261 dprint(FD_RATE, "We are done\n");
1264 dprint(FD_RATE, "Quiesce and final run\n");
1266 td->latency_end_run = 1;
1267 reset_all_stats(td);
1276 * Check if we can bump the queue depth
1278 void lat_target_check(struct thread_data *td)
1280 uint64_t usec_window;
1284 usec_window = utime_since_now(&td->latency_ts);
1285 if (usec_window < td->o.latency_window)
1288 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1289 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1290 success_ios *= 100.0;
1292 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1294 if (success_ios >= td->o.latency_percentile.u.f)
1295 lat_target_success(td);
1297 __lat_target_failed(td);
1301 * If latency target is enabled, we might be ramping up or down and not
1302 * using the full queue depth available.
1304 int queue_full(const struct thread_data *td)
1306 const int qempty = io_u_qempty(&td->io_u_freelist);
1310 if (!td->o.latency_target)
1313 return td->cur_depth >= td->latency_qd;
1316 struct io_u *__get_io_u(struct thread_data *td)
1318 struct io_u *io_u = NULL;
1323 if (!io_u_rempty(&td->io_u_requeues))
1324 io_u = io_u_rpop(&td->io_u_requeues);
1325 else if (!queue_full(td)) {
1326 io_u = io_u_qpop(&td->io_u_freelist);
1331 io_u->end_io = NULL;
1335 assert(io_u->flags & IO_U_F_FREE);
1336 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1337 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1341 io_u->acct_ddir = -1;
1343 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1345 } else if (td->o.verify_async) {
1347 * We ran out, wait for async verify threads to finish and
1350 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1358 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1360 if (!(td->flags & TD_F_TRIM_BACKLOG))
1363 if (td->trim_entries) {
1366 if (td->trim_batch) {
1369 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1370 td->last_ddir != DDIR_READ) {
1371 td->trim_batch = td->o.trim_batch;
1372 if (!td->trim_batch)
1373 td->trim_batch = td->o.trim_backlog;
1377 if (get_trim && !get_next_trim(td, io_u))
1384 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1386 if (!(td->flags & TD_F_VER_BACKLOG))
1389 if (td->io_hist_len) {
1392 if (td->verify_batch)
1394 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1395 td->last_ddir != DDIR_READ) {
1396 td->verify_batch = td->o.verify_batch;
1397 if (!td->verify_batch)
1398 td->verify_batch = td->o.verify_backlog;
1402 if (get_verify && !get_next_verify(td, io_u)) {
1412 * Fill offset and start time into the buffer content, to prevent too
1413 * easy compressible data for simple de-dupe attempts. Do this for every
1414 * 512b block in the range, since that should be the smallest block size
1415 * we can expect from a device.
1417 static void small_content_scramble(struct io_u *io_u)
1419 unsigned int i, nr_blocks = io_u->buflen / 512;
1421 unsigned int offset;
1428 boffset = io_u->offset;
1429 io_u->buf_filled_len = 0;
1431 for (i = 0; i < nr_blocks; i++) {
1433 * Fill the byte offset into a "random" start offset of
1434 * the buffer, given by the product of the usec time
1435 * and the actual offset.
1437 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1438 offset &= ~(sizeof(uint64_t) - 1);
1439 if (offset >= 512 - sizeof(uint64_t))
1440 offset -= sizeof(uint64_t);
1441 memcpy(p + offset, &boffset, sizeof(boffset));
1443 end = p + 512 - sizeof(io_u->start_time);
1444 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1451 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1452 * etc. The returned io_u is fully ready to be prepped and submitted.
1454 struct io_u *get_io_u(struct thread_data *td)
1458 int do_scramble = 0;
1461 io_u = __get_io_u(td);
1463 dprint(FD_IO, "__get_io_u failed\n");
1467 if (check_get_verify(td, io_u))
1469 if (check_get_trim(td, io_u))
1473 * from a requeue, io_u already setup
1479 * If using an iolog, grab next piece if any available.
1481 if (td->flags & TD_F_READ_IOLOG) {
1482 if (read_iolog_get(td, io_u))
1484 } else if (set_io_u_file(td, io_u)) {
1486 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1492 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1496 assert(fio_file_open(f));
1498 if (ddir_rw(io_u->ddir)) {
1499 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1500 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1504 f->last_start = io_u->offset;
1505 f->last_pos = io_u->offset + io_u->buflen;
1507 if (io_u->ddir == DDIR_WRITE) {
1508 if (td->flags & TD_F_REFILL_BUFFERS) {
1509 io_u_fill_buffer(td, io_u,
1510 td->o.min_bs[DDIR_WRITE],
1512 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1513 !(td->flags & TD_F_COMPRESS))
1515 if (td->flags & TD_F_VER_NONE) {
1516 populate_verify_io_u(td, io_u);
1519 } else if (io_u->ddir == DDIR_READ) {
1521 * Reset the buf_filled parameters so next time if the
1522 * buffer is used for writes it is refilled.
1524 io_u->buf_filled_len = 0;
1529 * Set io data pointers.
1531 io_u->xfer_buf = io_u->buf;
1532 io_u->xfer_buflen = io_u->buflen;
1536 if (!td_io_prep(td, io_u)) {
1537 if (!td->o.disable_slat)
1538 fio_gettime(&io_u->start_time, NULL);
1540 small_content_scramble(io_u);
1544 dprint(FD_IO, "get_io_u failed\n");
1546 return ERR_PTR(ret);
1549 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1551 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1553 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1556 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1557 io_u->file ? " on file " : "",
1558 io_u->file ? io_u->file->file_name : "",
1559 strerror(io_u->error),
1560 io_ddir_name(io_u->ddir),
1561 io_u->offset, io_u->xfer_buflen);
1564 td_verror(td, io_u->error, "io_u error");
1567 static inline int gtod_reduce(struct thread_data *td)
1569 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1570 && td->o.disable_bw;
1573 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1574 struct io_completion_data *icd,
1575 const enum fio_ddir idx, unsigned int bytes)
1577 unsigned long lusec = 0;
1579 if (!gtod_reduce(td))
1580 lusec = utime_since(&io_u->issue_time, &icd->time);
1582 if (!td->o.disable_lat) {
1583 unsigned long tusec;
1585 tusec = utime_since(&io_u->start_time, &icd->time);
1586 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1588 if (td->flags & TD_F_PROFILE_OPS) {
1589 struct prof_io_ops *ops = &td->prof_io_ops;
1592 icd->error = ops->io_u_lat(td, tusec);
1595 if (td->o.max_latency && tusec > td->o.max_latency)
1596 lat_fatal(td, icd, tusec, td->o.max_latency);
1597 if (td->o.latency_target && tusec > td->o.latency_target) {
1598 if (lat_target_failed(td))
1599 lat_fatal(td, icd, tusec, td->o.latency_target);
1603 if (!td->o.disable_clat) {
1604 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1605 io_u_mark_latency(td, lusec);
1608 if (!td->o.disable_bw)
1609 add_bw_sample(td, idx, bytes, &icd->time);
1611 if (!gtod_reduce(td))
1612 add_iops_sample(td, idx, bytes, &icd->time);
1615 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1617 uint64_t secs, remainder, bps, bytes;
1619 bytes = td->this_io_bytes[ddir];
1620 bps = td->rate_bps[ddir];
1622 remainder = bytes % bps;
1623 return remainder * 1000000 / bps + secs * 1000000;
1626 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1627 struct io_completion_data *icd)
1629 struct io_u *io_u = *io_u_ptr;
1630 enum fio_ddir ddir = io_u->ddir;
1631 struct fio_file *f = io_u->file;
1633 dprint_io_u(io_u, "io complete");
1636 assert(io_u->flags & IO_U_F_FLIGHT);
1637 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1640 * Mark IO ok to verify
1644 * Remove errored entry from the verification list
1647 unlog_io_piece(td, io_u);
1649 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1656 if (ddir_sync(ddir)) {
1657 td->last_was_sync = 1;
1659 f->first_write = -1ULL;
1660 f->last_write = -1ULL;
1665 td->last_was_sync = 0;
1666 td->last_ddir = ddir;
1668 if (!io_u->error && ddir_rw(ddir)) {
1669 unsigned int bytes = io_u->buflen - io_u->resid;
1670 const enum fio_ddir oddir = ddir ^ 1;
1673 td->io_blocks[ddir]++;
1674 td->this_io_blocks[ddir]++;
1675 td->io_bytes[ddir] += bytes;
1677 if (!(io_u->flags & IO_U_F_VER_LIST))
1678 td->this_io_bytes[ddir] += bytes;
1680 if (ddir == DDIR_WRITE && f) {
1681 if (f->first_write == -1ULL ||
1682 io_u->offset < f->first_write)
1683 f->first_write = io_u->offset;
1684 if (f->last_write == -1ULL ||
1685 ((io_u->offset + bytes) > f->last_write))
1686 f->last_write = io_u->offset + bytes;
1689 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1690 td->runstate == TD_VERIFYING)) {
1691 account_io_completion(td, io_u, icd, ddir, bytes);
1693 if (__should_check_rate(td, ddir)) {
1694 td->rate_pending_usleep[ddir] =
1695 (usec_for_io(td, ddir) -
1696 utime_since_now(&td->start));
1698 if (ddir != DDIR_TRIM &&
1699 __should_check_rate(td, oddir)) {
1700 td->rate_pending_usleep[oddir] =
1701 (usec_for_io(td, oddir) -
1702 utime_since_now(&td->start));
1706 icd->bytes_done[ddir] += bytes;
1709 ret = io_u->end_io(td, io_u_ptr);
1711 if (ret && !icd->error)
1714 } else if (io_u->error) {
1715 icd->error = io_u->error;
1716 io_u_log_error(td, io_u);
1719 enum error_type_bit eb = td_error_type(ddir, icd->error);
1721 if (!td_non_fatal_error(td, eb, icd->error))
1725 * If there is a non_fatal error, then add to the error count
1726 * and clear all the errors.
1728 update_error_count(td, icd->error);
1736 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1741 if (!gtod_reduce(td))
1742 fio_gettime(&icd->time, NULL);
1747 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1748 icd->bytes_done[ddir] = 0;
1751 static void ios_completed(struct thread_data *td,
1752 struct io_completion_data *icd)
1757 for (i = 0; i < icd->nr; i++) {
1758 io_u = td->io_ops->event(td, i);
1760 io_completed(td, &io_u, icd);
1768 * Complete a single io_u for the sync engines.
1770 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1773 struct io_completion_data icd;
1775 init_icd(td, &icd, 1);
1776 io_completed(td, &io_u, &icd);
1782 td_verror(td, icd.error, "io_u_sync_complete");
1789 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1790 bytes[ddir] += icd.bytes_done[ddir];
1797 * Called to complete min_events number of io for the async engines.
1799 int io_u_queued_complete(struct thread_data *td, int min_evts,
1802 struct io_completion_data icd;
1803 struct timespec *tvp = NULL;
1805 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1807 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1811 else if (min_evts > td->cur_depth)
1812 min_evts = td->cur_depth;
1814 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1816 td_verror(td, -ret, "td_io_getevents");
1821 init_icd(td, &icd, ret);
1822 ios_completed(td, &icd);
1824 td_verror(td, icd.error, "io_u_queued_complete");
1831 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1832 bytes[ddir] += icd.bytes_done[ddir];
1839 * Call when io_u is really queued, to update the submission latency.
1841 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1843 if (!td->o.disable_slat) {
1844 unsigned long slat_time;
1846 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1847 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1853 * See if we should reuse the last seed, if dedupe is enabled
1855 static struct frand_state *get_buf_state(struct thread_data *td)
1860 if (!td->o.dedupe_percentage)
1861 return &td->buf_state;
1862 else if (td->o.dedupe_percentage == 100)
1863 return &td->buf_state_prev;
1865 r = __rand(&td->dedupe_state);
1866 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1868 if (v <= td->o.dedupe_percentage)
1869 return &td->buf_state_prev;
1871 return &td->buf_state;
1874 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1876 if (rs == &td->buf_state)
1877 frand_copy(&td->buf_state_prev, rs);
1880 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1881 unsigned int max_bs)
1883 if (td->o.buffer_pattern_bytes)
1884 fill_buffer_pattern(td, buf, max_bs);
1885 else if (!td->o.zero_buffers) {
1886 unsigned int perc = td->o.compress_percentage;
1887 struct frand_state *rs;
1888 unsigned int left = max_bs;
1891 rs = get_buf_state(td);
1893 min_write = min(min_write, left);
1896 unsigned int seg = min_write;
1898 seg = min(min_write, td->o.compress_chunk);
1902 fill_random_buf_percentage(rs, buf, perc, seg,
1905 fill_random_buf(rs, buf, min_write);
1909 save_buf_state(td, rs);
1912 memset(buf, 0, max_bs);
1916 * "randomly" fill the buffer contents
1918 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1919 unsigned int min_write, unsigned int max_bs)
1921 io_u->buf_filled_len = 0;
1922 fill_io_buffer(td, io_u->buf, min_write, max_bs);