13 #include "lib/axmap.h"
16 struct io_completion_data {
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 static int random_map_free(struct fio_file *f, const uint64_t block)
30 return !axmap_isset(f->io_axmap, block);
34 * Mark a given offset as used in the map.
36 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
53 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 assert(ddir_rw(ddir));
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
69 max_size = td->o.zone_range;
71 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
79 struct flist_head list;
83 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b)
88 lastb = last_block(td, f, ddir);
92 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
93 r = __rand(&td->__random_state);
95 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
97 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
101 if (lfsr_next(&f->lfsr, &off, lastb))
108 * if we are not maintaining a random map, we are done.
110 if (!file_randommap(td, f))
114 * calculate map offset and check if it's free
116 if (random_map_free(f, *b))
119 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
120 (unsigned long long) *b);
122 *b = axmap_next_free(f->io_axmap, *b);
123 if (*b == (uint64_t) -1ULL)
129 static int __get_next_rand_offset_zipf(struct thread_data *td,
130 struct fio_file *f, enum fio_ddir ddir,
133 *b = zipf_next(&f->zipf);
137 static int __get_next_rand_offset_pareto(struct thread_data *td,
138 struct fio_file *f, enum fio_ddir ddir,
141 *b = pareto_next(&f->zipf);
145 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
147 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
148 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
150 return r1->off - r2->off;
153 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
154 enum fio_ddir ddir, uint64_t *b)
156 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
157 return __get_next_rand_offset(td, f, ddir, b);
158 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
159 return __get_next_rand_offset_zipf(td, f, ddir, b);
160 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
161 return __get_next_rand_offset_pareto(td, f, ddir, b);
163 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
168 * Sort the reads for a verify phase in batches of verifysort_nr, if
171 static inline int should_sort_io(struct thread_data *td)
173 if (!td->o.verifysort_nr || !td->o.do_verify)
177 if (td->runstate != TD_VERIFYING)
179 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
185 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
190 if (td->o.perc_rand[ddir] == 100)
193 r = __rand(&td->__seq_rand_state[ddir]);
194 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
196 return v <= td->o.perc_rand[ddir];
199 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
200 enum fio_ddir ddir, uint64_t *b)
205 if (!should_sort_io(td))
206 return get_off_from_method(td, f, ddir, b);
208 if (!flist_empty(&td->next_rand_list)) {
211 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
218 for (i = 0; i < td->o.verifysort_nr; i++) {
219 r = malloc(sizeof(*r));
221 ret = get_off_from_method(td, f, ddir, &r->off);
227 flist_add(&r->list, &td->next_rand_list);
233 assert(!flist_empty(&td->next_rand_list));
234 flist_sort(NULL, &td->next_rand_list, flist_cmp);
238 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239 enum fio_ddir ddir, uint64_t *b)
241 if (!get_next_rand_offset(td, f, ddir, b))
244 if (td->o.time_based) {
245 fio_file_reset(td, f);
246 if (!get_next_rand_offset(td, f, ddir, b))
250 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
251 f->file_name, (unsigned long long) f->last_pos,
252 (unsigned long long) f->real_file_size);
256 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
257 enum fio_ddir ddir, uint64_t *offset)
259 struct thread_options *o = &td->o;
261 assert(ddir_rw(ddir));
263 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
265 f->last_pos = f->last_pos - f->io_size;
267 if (f->last_pos < f->real_file_size) {
270 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
271 f->last_pos = f->real_file_size;
273 pos = f->last_pos - f->file_offset;
274 if (pos && o->ddir_seq_add) {
275 pos += o->ddir_seq_add;
278 * If we reach beyond the end of the file
279 * with holed IO, wrap around to the
282 if (pos >= f->real_file_size)
283 pos = f->file_offset;
293 static int get_next_block(struct thread_data *td, struct io_u *io_u,
294 enum fio_ddir ddir, int rw_seq,
295 unsigned int *is_random)
297 struct fio_file *f = io_u->file;
301 assert(ddir_rw(ddir));
307 if (should_do_random(td, ddir)) {
308 ret = get_next_rand_block(td, f, ddir, &b);
312 io_u->flags |= IO_U_F_BUSY_OK;
313 ret = get_next_seq_offset(td, f, ddir, &offset);
315 ret = get_next_rand_block(td, f, ddir, &b);
319 ret = get_next_seq_offset(td, f, ddir, &offset);
322 io_u->flags |= IO_U_F_BUSY_OK;
325 if (td->o.rw_seq == RW_SEQ_SEQ) {
326 ret = get_next_seq_offset(td, f, ddir, &offset);
328 ret = get_next_rand_block(td, f, ddir, &b);
331 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
332 if (f->last_start != -1ULL)
333 offset = f->last_start - f->file_offset;
338 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
345 io_u->offset = offset;
347 io_u->offset = b * td->o.ba[ddir];
349 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
358 * For random io, generate a random new block and see if it's used. Repeat
359 * until we find a free one. For sequential io, just return the end of
360 * the last io issued.
362 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
363 unsigned int *is_random)
365 struct fio_file *f = io_u->file;
366 enum fio_ddir ddir = io_u->ddir;
369 assert(ddir_rw(ddir));
371 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
373 td->ddir_seq_nr = td->o.ddir_seq_nr;
376 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
379 if (io_u->offset >= f->io_size) {
380 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
381 (unsigned long long) io_u->offset,
382 (unsigned long long) f->io_size);
386 io_u->offset += f->file_offset;
387 if (io_u->offset >= f->real_file_size) {
388 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
389 (unsigned long long) io_u->offset,
390 (unsigned long long) f->real_file_size);
397 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
398 unsigned int *is_random)
400 if (td->flags & TD_F_PROFILE_OPS) {
401 struct prof_io_ops *ops = &td->prof_io_ops;
403 if (ops->fill_io_u_off)
404 return ops->fill_io_u_off(td, io_u, is_random);
407 return __get_next_offset(td, io_u, is_random);
410 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
413 struct fio_file *f = io_u->file;
415 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
418 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
419 unsigned int is_random)
421 int ddir = io_u->ddir;
422 unsigned int buflen = 0;
423 unsigned int minbs, maxbs;
426 assert(ddir_rw(ddir));
428 if (td->o.bs_is_seq_rand)
429 ddir = is_random ? DDIR_WRITE: DDIR_READ;
431 minbs = td->o.min_bs[ddir];
432 maxbs = td->o.max_bs[ddir];
438 * If we can't satisfy the min block size from here, then fail
440 if (!io_u_fits(td, io_u, minbs))
444 r = __rand(&td->__bsrange_state);
446 if (!td->o.bssplit_nr[ddir]) {
447 buflen = 1 + (unsigned int) ((double) maxbs *
448 (r / (FRAND_MAX + 1.0)));
455 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
456 struct bssplit *bsp = &td->o.bssplit[ddir][i];
460 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
461 io_u_fits(td, io_u, buflen))
466 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
467 buflen = (buflen + td->o.verify_interval - 1) &
468 ~(td->o.verify_interval - 1);
470 if (!td->o.bs_unaligned && is_power_of_2(minbs))
471 buflen = (buflen + minbs - 1) & ~(minbs - 1);
473 } while (!io_u_fits(td, io_u, buflen));
478 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
479 unsigned int is_random)
481 if (td->flags & TD_F_PROFILE_OPS) {
482 struct prof_io_ops *ops = &td->prof_io_ops;
484 if (ops->fill_io_u_size)
485 return ops->fill_io_u_size(td, io_u, is_random);
488 return __get_next_buflen(td, io_u, is_random);
491 static void set_rwmix_bytes(struct thread_data *td)
496 * we do time or byte based switch. this is needed because
497 * buffered writes may issue a lot quicker than they complete,
498 * whereas reads do not.
500 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
501 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
504 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
509 r = __rand(&td->__rwmix_state);
510 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
512 if (v <= td->o.rwmix[DDIR_READ])
518 void io_u_quiesce(struct thread_data *td)
521 * We are going to sleep, ensure that we flush anything pending as
522 * not to skew our latency numbers.
524 * Changed to only monitor 'in flight' requests here instead of the
525 * td->cur_depth, b/c td->cur_depth does not accurately represent
526 * io's that have been actually submitted to an async engine,
527 * and cur_depth is meaningless for sync engines.
529 while (td->io_u_in_flight) {
532 ret = io_u_queued_complete(td, 1, NULL);
536 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
538 enum fio_ddir odir = ddir ^ 1;
542 assert(ddir_rw(ddir));
544 if (td->rate_pending_usleep[ddir] <= 0)
548 * We have too much pending sleep in this direction. See if we
551 if (td_rw(td) && td->o.rwmix[odir]) {
553 * Other direction does not have too much pending, switch
555 if (td->rate_pending_usleep[odir] < 100000)
559 * Both directions have pending sleep. Sleep the minimum time
560 * and deduct from both.
562 if (td->rate_pending_usleep[ddir] <=
563 td->rate_pending_usleep[odir]) {
564 usec = td->rate_pending_usleep[ddir];
566 usec = td->rate_pending_usleep[odir];
570 usec = td->rate_pending_usleep[ddir];
574 fio_gettime(&t, NULL);
575 usec_sleep(td, usec);
576 usec = utime_since_now(&t);
578 td->rate_pending_usleep[ddir] -= usec;
581 if (td_rw(td) && __should_check_rate(td, odir))
582 td->rate_pending_usleep[odir] -= usec;
591 * Return the data direction for the next io_u. If the job is a
592 * mixed read/write workload, check the rwmix cycle and switch if
595 static enum fio_ddir get_rw_ddir(struct thread_data *td)
600 * see if it's time to fsync
602 if (td->o.fsync_blocks &&
603 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
604 td->io_issues[DDIR_WRITE] && should_fsync(td))
608 * see if it's time to fdatasync
610 if (td->o.fdatasync_blocks &&
611 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
612 td->io_issues[DDIR_WRITE] && should_fsync(td))
613 return DDIR_DATASYNC;
616 * see if it's time to sync_file_range
618 if (td->sync_file_range_nr &&
619 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
620 td->io_issues[DDIR_WRITE] && should_fsync(td))
621 return DDIR_SYNC_FILE_RANGE;
625 * Check if it's time to seed a new data direction.
627 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
629 * Put a top limit on how many bytes we do for
630 * one data direction, to avoid overflowing the
633 ddir = get_rand_ddir(td);
635 if (ddir != td->rwmix_ddir)
638 td->rwmix_ddir = ddir;
640 ddir = td->rwmix_ddir;
641 } else if (td_read(td))
643 else if (td_write(td))
648 td->rwmix_ddir = rate_ddir(td, ddir);
649 return td->rwmix_ddir;
652 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
654 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
656 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
657 td->o.barrier_blocks &&
658 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
659 td->io_issues[DDIR_WRITE])
660 io_u->flags |= IO_U_F_BARRIER;
663 void put_file_log(struct thread_data *td, struct fio_file *f)
665 unsigned int ret = put_file(td, f);
668 td_verror(td, ret, "file close");
671 void put_io_u(struct thread_data *td, struct io_u *io_u)
675 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
676 put_file_log(td, io_u->file);
679 io_u->flags |= IO_U_F_FREE;
681 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
683 io_u_qpush(&td->io_u_freelist, io_u);
685 td_io_u_free_notify(td);
688 void clear_io_u(struct thread_data *td, struct io_u *io_u)
690 io_u->flags &= ~IO_U_F_FLIGHT;
694 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
696 struct io_u *__io_u = *io_u;
697 enum fio_ddir ddir = acct_ddir(__io_u);
699 dprint(FD_IO, "requeue %p\n", __io_u);
703 __io_u->flags |= IO_U_F_FREE;
704 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
705 td->io_issues[ddir]--;
707 __io_u->flags &= ~IO_U_F_FLIGHT;
708 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
711 io_u_rpush(&td->io_u_requeues, __io_u);
716 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
718 unsigned int is_random;
720 if (td->io_ops->flags & FIO_NOIO)
723 set_rw_ddir(td, io_u);
726 * fsync() or fdatasync() or trim etc, we are done
728 if (!ddir_rw(io_u->ddir))
732 * See if it's time to switch to a new zone
734 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
735 struct fio_file *f = io_u->file;
738 f->file_offset += td->o.zone_range + td->o.zone_skip;
741 * Wrap from the beginning, if we exceed the file size
743 if (f->file_offset >= f->real_file_size)
744 f->file_offset = f->real_file_size - f->file_offset;
745 f->last_pos = f->file_offset;
746 td->io_skip_bytes += td->o.zone_skip;
750 * No log, let the seq/rand engine retrieve the next buflen and
753 if (get_next_offset(td, io_u, &is_random)) {
754 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
758 io_u->buflen = get_next_buflen(td, io_u, is_random);
760 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
764 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
765 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
766 dprint(FD_IO, " off=%llu/%lu > %llu\n",
767 (unsigned long long) io_u->offset, io_u->buflen,
768 (unsigned long long) io_u->file->real_file_size);
773 * mark entry before potentially trimming io_u
775 if (td_random(td) && file_randommap(td, io_u->file))
776 mark_random_map(td, io_u);
779 dprint_io_u(io_u, "fill_io_u");
780 td->zone_bytes += io_u->buflen;
784 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
813 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
815 __io_u_mark_map(td->ts.io_u_submit, nr);
816 td->ts.total_submit++;
819 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
821 __io_u_mark_map(td->ts.io_u_complete, nr);
822 td->ts.total_complete++;
825 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
829 switch (td->cur_depth) {
851 td->ts.io_u_map[idx] += nr;
854 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
891 assert(idx < FIO_IO_U_LAT_U_NR);
892 td->ts.io_u_lat_u[idx]++;
895 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
936 assert(idx < FIO_IO_U_LAT_M_NR);
937 td->ts.io_u_lat_m[idx]++;
940 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
943 io_u_mark_lat_usec(td, usec);
945 io_u_mark_lat_msec(td, usec / 1000);
949 * Get next file to service by choosing one at random
951 static struct fio_file *get_next_file_rand(struct thread_data *td,
952 enum fio_file_flags goodf,
953 enum fio_file_flags badf)
962 r = __rand(&td->__next_file_state);
963 fno = (unsigned int) ((double) td->o.nr_files
964 * (r / (FRAND_MAX + 1.0)));
967 if (fio_file_done(f))
970 if (!fio_file_open(f)) {
973 if (td->nr_open_files >= td->o.open_files)
974 return ERR_PTR(-EBUSY);
976 err = td_io_open_file(td, f);
982 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
983 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
987 td_io_close_file(td, f);
992 * Get next file to service by doing round robin between all available ones
994 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
997 unsigned int old_next_file = td->next_file;
1003 f = td->files[td->next_file];
1006 if (td->next_file >= td->o.nr_files)
1009 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1010 if (fio_file_done(f)) {
1015 if (!fio_file_open(f)) {
1018 if (td->nr_open_files >= td->o.open_files)
1019 return ERR_PTR(-EBUSY);
1021 err = td_io_open_file(td, f);
1023 dprint(FD_FILE, "error %d on open of %s\n",
1031 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1033 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1037 td_io_close_file(td, f);
1040 } while (td->next_file != old_next_file);
1042 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1046 static struct fio_file *__get_next_file(struct thread_data *td)
1050 assert(td->o.nr_files <= td->files_index);
1052 if (td->nr_done_files >= td->o.nr_files) {
1053 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1054 " nr_files=%d\n", td->nr_open_files,
1060 f = td->file_service_file;
1061 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1062 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1064 if (td->file_service_left--)
1068 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1069 td->o.file_service_type == FIO_FSERVICE_SEQ)
1070 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1072 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1077 td->file_service_file = f;
1078 td->file_service_left = td->file_service_nr - 1;
1081 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1083 dprint(FD_FILE, "get_next_file: NULL\n");
1087 static struct fio_file *get_next_file(struct thread_data *td)
1089 if (td->flags & TD_F_PROFILE_OPS) {
1090 struct prof_io_ops *ops = &td->prof_io_ops;
1092 if (ops->get_next_file)
1093 return ops->get_next_file(td);
1096 return __get_next_file(td);
1099 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1104 f = get_next_file(td);
1105 if (IS_ERR_OR_NULL(f))
1111 if (!fill_io_u(td, io_u))
1114 put_file_log(td, f);
1115 td_io_close_file(td, f);
1117 fio_file_set_done(f);
1118 td->nr_done_files++;
1119 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1120 td->nr_done_files, td->o.nr_files);
1126 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1127 unsigned long tusec, unsigned long max_usec)
1130 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1131 td_verror(td, ETIMEDOUT, "max latency exceeded");
1132 icd->error = ETIMEDOUT;
1135 static void lat_new_cycle(struct thread_data *td)
1137 fio_gettime(&td->latency_ts, NULL);
1138 td->latency_ios = ddir_rw_sum(td->io_blocks);
1139 td->latency_failed = 0;
1143 * We had an IO outside the latency target. Reduce the queue depth. If we
1144 * are at QD=1, then it's time to give up.
1146 static int __lat_target_failed(struct thread_data *td)
1148 if (td->latency_qd == 1)
1151 td->latency_qd_high = td->latency_qd;
1153 if (td->latency_qd == td->latency_qd_low)
1154 td->latency_qd_low--;
1156 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1158 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1161 * When we ramp QD down, quiesce existing IO to prevent
1162 * a storm of ramp downs due to pending higher depth.
1169 static int lat_target_failed(struct thread_data *td)
1171 if (td->o.latency_percentile.u.f == 100.0)
1172 return __lat_target_failed(td);
1174 td->latency_failed++;
1178 void lat_target_init(struct thread_data *td)
1180 td->latency_end_run = 0;
1182 if (td->o.latency_target) {
1183 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1184 fio_gettime(&td->latency_ts, NULL);
1186 td->latency_qd_high = td->o.iodepth;
1187 td->latency_qd_low = 1;
1188 td->latency_ios = ddir_rw_sum(td->io_blocks);
1190 td->latency_qd = td->o.iodepth;
1193 void lat_target_reset(struct thread_data *td)
1195 if (!td->latency_end_run)
1196 lat_target_init(td);
1199 static void lat_target_success(struct thread_data *td)
1201 const unsigned int qd = td->latency_qd;
1202 struct thread_options *o = &td->o;
1204 td->latency_qd_low = td->latency_qd;
1207 * If we haven't failed yet, we double up to a failing value instead
1208 * of bisecting from highest possible queue depth. If we have set
1209 * a limit other than td->o.iodepth, bisect between that.
1211 if (td->latency_qd_high != o->iodepth)
1212 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1214 td->latency_qd *= 2;
1216 if (td->latency_qd > o->iodepth)
1217 td->latency_qd = o->iodepth;
1219 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1222 * Same as last one, we are done. Let it run a latency cycle, so
1223 * we get only the results from the targeted depth.
1225 if (td->latency_qd == qd) {
1226 if (td->latency_end_run) {
1227 dprint(FD_RATE, "We are done\n");
1230 dprint(FD_RATE, "Quiesce and final run\n");
1232 td->latency_end_run = 1;
1233 reset_all_stats(td);
1242 * Check if we can bump the queue depth
1244 void lat_target_check(struct thread_data *td)
1246 uint64_t usec_window;
1250 usec_window = utime_since_now(&td->latency_ts);
1251 if (usec_window < td->o.latency_window)
1254 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1255 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1256 success_ios *= 100.0;
1258 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1260 if (success_ios >= td->o.latency_percentile.u.f)
1261 lat_target_success(td);
1263 __lat_target_failed(td);
1267 * If latency target is enabled, we might be ramping up or down and not
1268 * using the full queue depth available.
1270 int queue_full(const struct thread_data *td)
1272 const int qempty = io_u_qempty(&td->io_u_freelist);
1276 if (!td->o.latency_target)
1279 return td->cur_depth >= td->latency_qd;
1282 struct io_u *__get_io_u(struct thread_data *td)
1284 struct io_u *io_u = NULL;
1289 if (!io_u_rempty(&td->io_u_requeues))
1290 io_u = io_u_rpop(&td->io_u_requeues);
1291 else if (!queue_full(td)) {
1292 io_u = io_u_qpop(&td->io_u_freelist);
1297 io_u->end_io = NULL;
1301 assert(io_u->flags & IO_U_F_FREE);
1302 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1303 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1307 io_u->acct_ddir = -1;
1309 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1311 } else if (td->o.verify_async) {
1313 * We ran out, wait for async verify threads to finish and
1316 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1324 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1326 if (!(td->flags & TD_F_TRIM_BACKLOG))
1329 if (td->trim_entries) {
1332 if (td->trim_batch) {
1335 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1336 td->last_ddir != DDIR_READ) {
1337 td->trim_batch = td->o.trim_batch;
1338 if (!td->trim_batch)
1339 td->trim_batch = td->o.trim_backlog;
1343 if (get_trim && !get_next_trim(td, io_u))
1350 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1352 if (!(td->flags & TD_F_VER_BACKLOG))
1355 if (td->io_hist_len) {
1358 if (td->verify_batch)
1360 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1361 td->last_ddir != DDIR_READ) {
1362 td->verify_batch = td->o.verify_batch;
1363 if (!td->verify_batch)
1364 td->verify_batch = td->o.verify_backlog;
1368 if (get_verify && !get_next_verify(td, io_u)) {
1378 * Fill offset and start time into the buffer content, to prevent too
1379 * easy compressible data for simple de-dupe attempts. Do this for every
1380 * 512b block in the range, since that should be the smallest block size
1381 * we can expect from a device.
1383 static void small_content_scramble(struct io_u *io_u)
1385 unsigned int i, nr_blocks = io_u->buflen / 512;
1387 unsigned int offset;
1394 boffset = io_u->offset;
1395 io_u->buf_filled_len = 0;
1397 for (i = 0; i < nr_blocks; i++) {
1399 * Fill the byte offset into a "random" start offset of
1400 * the buffer, given by the product of the usec time
1401 * and the actual offset.
1403 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1404 offset &= ~(sizeof(uint64_t) - 1);
1405 if (offset >= 512 - sizeof(uint64_t))
1406 offset -= sizeof(uint64_t);
1407 memcpy(p + offset, &boffset, sizeof(boffset));
1409 end = p + 512 - sizeof(io_u->start_time);
1410 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1417 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1418 * etc. The returned io_u is fully ready to be prepped and submitted.
1420 struct io_u *get_io_u(struct thread_data *td)
1424 int do_scramble = 0;
1427 io_u = __get_io_u(td);
1429 dprint(FD_IO, "__get_io_u failed\n");
1433 if (check_get_verify(td, io_u))
1435 if (check_get_trim(td, io_u))
1439 * from a requeue, io_u already setup
1445 * If using an iolog, grab next piece if any available.
1447 if (td->flags & TD_F_READ_IOLOG) {
1448 if (read_iolog_get(td, io_u))
1450 } else if (set_io_u_file(td, io_u)) {
1452 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1458 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1462 assert(fio_file_open(f));
1464 if (ddir_rw(io_u->ddir)) {
1465 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1466 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1470 f->last_start = io_u->offset;
1471 f->last_pos = io_u->offset + io_u->buflen;
1473 if (io_u->ddir == DDIR_WRITE) {
1474 if (td->flags & TD_F_REFILL_BUFFERS) {
1475 io_u_fill_buffer(td, io_u,
1476 td->o.min_bs[DDIR_WRITE],
1478 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1479 !(td->flags & TD_F_COMPRESS))
1481 if (td->flags & TD_F_VER_NONE) {
1482 populate_verify_io_u(td, io_u);
1485 } else if (io_u->ddir == DDIR_READ) {
1487 * Reset the buf_filled parameters so next time if the
1488 * buffer is used for writes it is refilled.
1490 io_u->buf_filled_len = 0;
1495 * Set io data pointers.
1497 io_u->xfer_buf = io_u->buf;
1498 io_u->xfer_buflen = io_u->buflen;
1502 if (!td_io_prep(td, io_u)) {
1503 if (!td->o.disable_slat)
1504 fio_gettime(&io_u->start_time, NULL);
1506 small_content_scramble(io_u);
1510 dprint(FD_IO, "get_io_u failed\n");
1512 return ERR_PTR(ret);
1515 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1517 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1519 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1522 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1523 io_u->file ? " on file " : "",
1524 io_u->file ? io_u->file->file_name : "",
1525 strerror(io_u->error),
1526 io_ddir_name(io_u->ddir),
1527 io_u->offset, io_u->xfer_buflen);
1530 td_verror(td, io_u->error, "io_u error");
1533 static inline int gtod_reduce(struct thread_data *td)
1535 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1536 && td->o.disable_bw;
1539 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1540 struct io_completion_data *icd,
1541 const enum fio_ddir idx, unsigned int bytes)
1543 unsigned long lusec = 0;
1545 if (!gtod_reduce(td))
1546 lusec = utime_since(&io_u->issue_time, &icd->time);
1548 if (!td->o.disable_lat) {
1549 unsigned long tusec;
1551 tusec = utime_since(&io_u->start_time, &icd->time);
1552 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1554 if (td->flags & TD_F_PROFILE_OPS) {
1555 struct prof_io_ops *ops = &td->prof_io_ops;
1558 icd->error = ops->io_u_lat(td, tusec);
1561 if (td->o.max_latency && tusec > td->o.max_latency)
1562 lat_fatal(td, icd, tusec, td->o.max_latency);
1563 if (td->o.latency_target && tusec > td->o.latency_target) {
1564 if (lat_target_failed(td))
1565 lat_fatal(td, icd, tusec, td->o.latency_target);
1569 if (!td->o.disable_clat) {
1570 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1571 io_u_mark_latency(td, lusec);
1574 if (!td->o.disable_bw)
1575 add_bw_sample(td, idx, bytes, &icd->time);
1577 if (!gtod_reduce(td))
1578 add_iops_sample(td, idx, bytes, &icd->time);
1581 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1583 uint64_t secs, remainder, bps, bytes;
1585 bytes = td->this_io_bytes[ddir];
1586 bps = td->rate_bps[ddir];
1588 remainder = bytes % bps;
1589 return remainder * 1000000 / bps + secs * 1000000;
1592 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1593 struct io_completion_data *icd)
1595 struct io_u *io_u = *io_u_ptr;
1596 enum fio_ddir ddir = io_u->ddir;
1597 struct fio_file *f = io_u->file;
1599 dprint_io_u(io_u, "io complete");
1602 assert(io_u->flags & IO_U_F_FLIGHT);
1603 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1606 * Mark IO ok to verify
1610 * Remove errored entry from the verification list
1613 unlog_io_piece(td, io_u);
1615 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1622 if (ddir_sync(ddir)) {
1623 td->last_was_sync = 1;
1625 f->first_write = -1ULL;
1626 f->last_write = -1ULL;
1631 td->last_was_sync = 0;
1632 td->last_ddir = ddir;
1634 if (!io_u->error && ddir_rw(ddir)) {
1635 unsigned int bytes = io_u->buflen - io_u->resid;
1636 const enum fio_ddir oddir = ddir ^ 1;
1639 td->io_blocks[ddir]++;
1640 td->this_io_blocks[ddir]++;
1641 td->io_bytes[ddir] += bytes;
1643 if (!(io_u->flags & IO_U_F_VER_LIST))
1644 td->this_io_bytes[ddir] += bytes;
1646 if (ddir == DDIR_WRITE && f) {
1647 if (f->first_write == -1ULL ||
1648 io_u->offset < f->first_write)
1649 f->first_write = io_u->offset;
1650 if (f->last_write == -1ULL ||
1651 ((io_u->offset + bytes) > f->last_write))
1652 f->last_write = io_u->offset + bytes;
1655 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1656 td->runstate == TD_VERIFYING)) {
1657 account_io_completion(td, io_u, icd, ddir, bytes);
1659 if (__should_check_rate(td, ddir)) {
1660 td->rate_pending_usleep[ddir] =
1661 (usec_for_io(td, ddir) -
1662 utime_since_now(&td->start));
1664 if (ddir != DDIR_TRIM &&
1665 __should_check_rate(td, oddir)) {
1666 td->rate_pending_usleep[oddir] =
1667 (usec_for_io(td, oddir) -
1668 utime_since_now(&td->start));
1672 icd->bytes_done[ddir] += bytes;
1675 ret = io_u->end_io(td, io_u_ptr);
1677 if (ret && !icd->error)
1680 } else if (io_u->error) {
1681 icd->error = io_u->error;
1682 io_u_log_error(td, io_u);
1685 enum error_type_bit eb = td_error_type(ddir, icd->error);
1687 if (!td_non_fatal_error(td, eb, icd->error))
1691 * If there is a non_fatal error, then add to the error count
1692 * and clear all the errors.
1694 update_error_count(td, icd->error);
1702 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1707 if (!gtod_reduce(td))
1708 fio_gettime(&icd->time, NULL);
1713 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1714 icd->bytes_done[ddir] = 0;
1717 static void ios_completed(struct thread_data *td,
1718 struct io_completion_data *icd)
1723 for (i = 0; i < icd->nr; i++) {
1724 io_u = td->io_ops->event(td, i);
1726 io_completed(td, &io_u, icd);
1734 * Complete a single io_u for the sync engines.
1736 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1739 struct io_completion_data icd;
1741 init_icd(td, &icd, 1);
1742 io_completed(td, &io_u, &icd);
1748 td_verror(td, icd.error, "io_u_sync_complete");
1755 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1756 bytes[ddir] += icd.bytes_done[ddir];
1763 * Called to complete min_events number of io for the async engines.
1765 int io_u_queued_complete(struct thread_data *td, int min_evts,
1768 struct io_completion_data icd;
1769 struct timespec *tvp = NULL;
1771 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1773 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1777 else if (min_evts > td->cur_depth)
1778 min_evts = td->cur_depth;
1780 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1782 td_verror(td, -ret, "td_io_getevents");
1787 init_icd(td, &icd, ret);
1788 ios_completed(td, &icd);
1790 td_verror(td, icd.error, "io_u_queued_complete");
1797 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1798 bytes[ddir] += icd.bytes_done[ddir];
1805 * Call when io_u is really queued, to update the submission latency.
1807 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1809 if (!td->o.disable_slat) {
1810 unsigned long slat_time;
1812 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1813 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1819 * See if we should reuse the last seed, if dedupe is enabled
1821 static struct frand_state *get_buf_state(struct thread_data *td)
1826 if (!td->o.dedupe_percentage)
1827 return &td->buf_state;
1828 else if (td->o.dedupe_percentage == 100)
1829 return &td->buf_state_prev;
1831 r = __rand(&td->dedupe_state);
1832 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1834 if (v <= td->o.dedupe_percentage)
1835 return &td->buf_state_prev;
1837 return &td->buf_state;
1840 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1842 if (rs == &td->buf_state)
1843 frand_copy(&td->buf_state_prev, rs);
1846 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1847 unsigned int max_bs)
1849 if (td->o.buffer_pattern_bytes)
1850 fill_buffer_pattern(td, buf, max_bs);
1851 else if (!td->o.zero_buffers) {
1852 unsigned int perc = td->o.compress_percentage;
1853 struct frand_state *rs;
1854 unsigned int left = max_bs;
1857 rs = get_buf_state(td);
1859 min_write = min(min_write, left);
1862 unsigned int seg = min_write;
1864 seg = min(min_write, td->o.compress_chunk);
1868 fill_random_buf_percentage(rs, buf, perc, seg,
1871 fill_random_buf(rs, buf, min_write);
1875 save_buf_state(td, rs);
1878 memset(buf, 0, max_bs);
1882 * "randomly" fill the buffer contents
1884 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1885 unsigned int min_write, unsigned int max_bs)
1887 io_u->buf_filled_len = 0;
1888 fill_io_buffer(td, io_u->buf, min_write, max_bs);