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 if (td->o.min_bs[ddir] > td->o.ba[ddir])
72 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
74 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
82 struct flist_head list;
86 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir, uint64_t *b)
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
94 lastb = last_block(td, f, ddir);
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
106 assert(fio_file_lfsr(f));
108 if (lfsr_next(&f->lfsr, &off))
115 * if we are not maintaining a random map, we are done.
117 if (!file_randommap(td, f))
121 * calculate map offset and check if it's free
123 if (random_map_free(f, *b))
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
136 static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
140 *b = zipf_next(&f->zipf);
144 static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
148 *b = pareto_next(&f->zipf);
152 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
154 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
155 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
157 return r1->off - r2->off;
160 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
161 enum fio_ddir ddir, uint64_t *b)
163 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
164 return __get_next_rand_offset(td, f, ddir, b);
165 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
166 return __get_next_rand_offset_zipf(td, f, ddir, b);
167 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
168 return __get_next_rand_offset_pareto(td, f, ddir, b);
170 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
175 * Sort the reads for a verify phase in batches of verifysort_nr, if
178 static inline int should_sort_io(struct thread_data *td)
180 if (!td->o.verifysort_nr || !td->o.do_verify)
184 if (td->runstate != TD_VERIFYING)
186 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
192 static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
197 if (td->o.perc_rand[ddir] == 100)
200 r = __rand(&td->seq_rand_state[ddir]);
201 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
203 return v <= td->o.perc_rand[ddir];
206 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
207 enum fio_ddir ddir, uint64_t *b)
212 if (!should_sort_io(td))
213 return get_off_from_method(td, f, ddir, b);
215 if (!flist_empty(&td->next_rand_list)) {
217 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
224 for (i = 0; i < td->o.verifysort_nr; i++) {
225 r = malloc(sizeof(*r));
227 ret = get_off_from_method(td, f, ddir, &r->off);
233 flist_add(&r->list, &td->next_rand_list);
239 assert(!flist_empty(&td->next_rand_list));
240 flist_sort(NULL, &td->next_rand_list, flist_cmp);
244 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
245 enum fio_ddir ddir, uint64_t *b)
247 if (!get_next_rand_offset(td, f, ddir, b))
250 if (td->o.time_based) {
251 fio_file_reset(td, f);
252 if (!get_next_rand_offset(td, f, ddir, b))
256 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
257 f->file_name, (unsigned long long) f->last_pos[ddir],
258 (unsigned long long) f->real_file_size);
262 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
263 enum fio_ddir ddir, uint64_t *offset)
265 struct thread_options *o = &td->o;
267 assert(ddir_rw(ddir));
269 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
271 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
273 if (f->last_pos[ddir] < f->real_file_size) {
276 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
277 f->last_pos[ddir] = f->real_file_size;
279 pos = f->last_pos[ddir] - f->file_offset;
280 if (pos && o->ddir_seq_add) {
281 pos += o->ddir_seq_add;
284 * If we reach beyond the end of the file
285 * with holed IO, wrap around to the
288 if (pos >= f->real_file_size)
289 pos = f->file_offset;
299 static int get_next_block(struct thread_data *td, struct io_u *io_u,
300 enum fio_ddir ddir, int rw_seq,
301 unsigned int *is_random)
303 struct fio_file *f = io_u->file;
307 assert(ddir_rw(ddir));
313 if (should_do_random(td, ddir)) {
314 ret = get_next_rand_block(td, f, ddir, &b);
318 io_u->flags |= IO_U_F_BUSY_OK;
319 ret = get_next_seq_offset(td, f, ddir, &offset);
321 ret = get_next_rand_block(td, f, ddir, &b);
325 ret = get_next_seq_offset(td, f, ddir, &offset);
328 io_u->flags |= IO_U_F_BUSY_OK;
331 if (td->o.rw_seq == RW_SEQ_SEQ) {
332 ret = get_next_seq_offset(td, f, ddir, &offset);
334 ret = get_next_rand_block(td, f, ddir, &b);
337 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
338 if (f->last_start[ddir] != -1ULL)
339 offset = f->last_start[ddir] - f->file_offset;
344 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
351 io_u->offset = offset;
353 io_u->offset = b * td->o.ba[ddir];
355 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
364 * For random io, generate a random new block and see if it's used. Repeat
365 * until we find a free one. For sequential io, just return the end of
366 * the last io issued.
368 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
369 unsigned int *is_random)
371 struct fio_file *f = io_u->file;
372 enum fio_ddir ddir = io_u->ddir;
375 assert(ddir_rw(ddir));
377 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
379 td->ddir_seq_nr = td->o.ddir_seq_nr;
382 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
385 if (io_u->offset >= f->io_size) {
386 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
387 (unsigned long long) io_u->offset,
388 (unsigned long long) f->io_size);
392 io_u->offset += f->file_offset;
393 if (io_u->offset >= f->real_file_size) {
394 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
395 (unsigned long long) io_u->offset,
396 (unsigned long long) f->real_file_size);
403 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
404 unsigned int *is_random)
406 if (td->flags & TD_F_PROFILE_OPS) {
407 struct prof_io_ops *ops = &td->prof_io_ops;
409 if (ops->fill_io_u_off)
410 return ops->fill_io_u_off(td, io_u, is_random);
413 return __get_next_offset(td, io_u, is_random);
416 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
419 struct fio_file *f = io_u->file;
421 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
424 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
425 unsigned int is_random)
427 int ddir = io_u->ddir;
428 unsigned int buflen = 0;
429 unsigned int minbs, maxbs;
432 assert(ddir_rw(ddir));
434 if (td->o.bs_is_seq_rand)
435 ddir = is_random ? DDIR_WRITE: DDIR_READ;
437 minbs = td->o.min_bs[ddir];
438 maxbs = td->o.max_bs[ddir];
444 * If we can't satisfy the min block size from here, then fail
446 if (!io_u_fits(td, io_u, minbs))
450 r = __rand(&td->bsrange_state);
452 if (!td->o.bssplit_nr[ddir]) {
453 buflen = 1 + (unsigned int) ((double) maxbs *
454 (r / (FRAND_MAX + 1.0)));
461 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
462 struct bssplit *bsp = &td->o.bssplit[ddir][i];
466 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
467 io_u_fits(td, io_u, buflen))
472 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
473 buflen = (buflen + td->o.verify_interval - 1) &
474 ~(td->o.verify_interval - 1);
476 if (!td->o.bs_unaligned && is_power_of_2(minbs))
477 buflen = (buflen + minbs - 1) & ~(minbs - 1);
479 } while (!io_u_fits(td, io_u, buflen));
484 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
485 unsigned int is_random)
487 if (td->flags & TD_F_PROFILE_OPS) {
488 struct prof_io_ops *ops = &td->prof_io_ops;
490 if (ops->fill_io_u_size)
491 return ops->fill_io_u_size(td, io_u, is_random);
494 return __get_next_buflen(td, io_u, is_random);
497 static void set_rwmix_bytes(struct thread_data *td)
502 * we do time or byte based switch. this is needed because
503 * buffered writes may issue a lot quicker than they complete,
504 * whereas reads do not.
506 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
507 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
510 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
515 r = __rand(&td->rwmix_state);
516 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
518 if (v <= td->o.rwmix[DDIR_READ])
524 void io_u_quiesce(struct thread_data *td)
527 * We are going to sleep, ensure that we flush anything pending as
528 * not to skew our latency numbers.
530 * Changed to only monitor 'in flight' requests here instead of the
531 * td->cur_depth, b/c td->cur_depth does not accurately represent
532 * io's that have been actually submitted to an async engine,
533 * and cur_depth is meaningless for sync engines.
535 if (td->io_u_queued || td->cur_depth) {
538 ret = td_io_commit(td);
541 while (td->io_u_in_flight) {
544 ret = io_u_queued_complete(td, 1, NULL);
548 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
550 enum fio_ddir odir = ddir ^ 1;
553 assert(ddir_rw(ddir));
555 if (td->rate_pending_usleep[ddir] <= 0)
559 * We have too much pending sleep in this direction. See if we
562 if (td_rw(td) && td->o.rwmix[odir]) {
564 * Other direction does not have too much pending, switch
566 if (td->rate_pending_usleep[odir] < 100000)
570 * Both directions have pending sleep. Sleep the minimum time
571 * and deduct from both.
573 if (td->rate_pending_usleep[ddir] <=
574 td->rate_pending_usleep[odir]) {
575 usec = td->rate_pending_usleep[ddir];
577 usec = td->rate_pending_usleep[odir];
581 usec = td->rate_pending_usleep[ddir];
585 usec = usec_sleep(td, usec);
587 td->rate_pending_usleep[ddir] -= usec;
590 if (td_rw(td) && __should_check_rate(td, odir))
591 td->rate_pending_usleep[odir] -= usec;
593 if (ddir == DDIR_TRIM)
600 * Return the data direction for the next io_u. If the job is a
601 * mixed read/write workload, check the rwmix cycle and switch if
604 static enum fio_ddir get_rw_ddir(struct thread_data *td)
609 * see if it's time to fsync
611 if (td->o.fsync_blocks &&
612 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
613 td->io_issues[DDIR_WRITE] && should_fsync(td))
617 * see if it's time to fdatasync
619 if (td->o.fdatasync_blocks &&
620 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
621 td->io_issues[DDIR_WRITE] && should_fsync(td))
622 return DDIR_DATASYNC;
625 * see if it's time to sync_file_range
627 if (td->sync_file_range_nr &&
628 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
629 td->io_issues[DDIR_WRITE] && should_fsync(td))
630 return DDIR_SYNC_FILE_RANGE;
634 * Check if it's time to seed a new data direction.
636 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
638 * Put a top limit on how many bytes we do for
639 * one data direction, to avoid overflowing the
642 ddir = get_rand_ddir(td);
644 if (ddir != td->rwmix_ddir)
647 td->rwmix_ddir = ddir;
649 ddir = td->rwmix_ddir;
650 } else if (td_read(td))
652 else if (td_write(td))
657 td->rwmix_ddir = rate_ddir(td, ddir);
658 return td->rwmix_ddir;
661 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
663 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
665 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
666 td->o.barrier_blocks &&
667 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
668 td->io_issues[DDIR_WRITE])
669 io_u->flags |= IO_U_F_BARRIER;
672 void put_file_log(struct thread_data *td, struct fio_file *f)
674 unsigned int ret = put_file(td, f);
677 td_verror(td, ret, "file close");
680 void put_io_u(struct thread_data *td, struct io_u *io_u)
684 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
685 put_file_log(td, io_u->file);
688 io_u->flags |= IO_U_F_FREE;
690 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
692 io_u_qpush(&td->io_u_freelist, io_u);
694 td_io_u_free_notify(td);
697 void clear_io_u(struct thread_data *td, struct io_u *io_u)
699 io_u->flags &= ~IO_U_F_FLIGHT;
703 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
705 struct io_u *__io_u = *io_u;
706 enum fio_ddir ddir = acct_ddir(__io_u);
708 dprint(FD_IO, "requeue %p\n", __io_u);
712 __io_u->flags |= IO_U_F_FREE;
713 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
714 td->io_issues[ddir]--;
716 __io_u->flags &= ~IO_U_F_FLIGHT;
717 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
720 io_u_rpush(&td->io_u_requeues, __io_u);
725 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
727 unsigned int is_random;
729 if (td->io_ops->flags & FIO_NOIO)
732 set_rw_ddir(td, io_u);
735 * fsync() or fdatasync() or trim etc, we are done
737 if (!ddir_rw(io_u->ddir))
741 * See if it's time to switch to a new zone
743 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
744 struct fio_file *f = io_u->file;
747 f->file_offset += td->o.zone_range + td->o.zone_skip;
750 * Wrap from the beginning, if we exceed the file size
752 if (f->file_offset >= f->real_file_size)
753 f->file_offset = f->real_file_size - f->file_offset;
754 f->last_pos[io_u->ddir] = f->file_offset;
755 td->io_skip_bytes += td->o.zone_skip;
759 * No log, let the seq/rand engine retrieve the next buflen and
762 if (get_next_offset(td, io_u, &is_random)) {
763 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
767 io_u->buflen = get_next_buflen(td, io_u, is_random);
769 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
773 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
774 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
775 dprint(FD_IO, " off=%llu/%lu > %llu\n",
776 (unsigned long long) io_u->offset, io_u->buflen,
777 (unsigned long long) io_u->file->real_file_size);
782 * mark entry before potentially trimming io_u
784 if (td_random(td) && file_randommap(td, io_u->file))
785 mark_random_map(td, io_u);
788 dprint_io_u(io_u, "fill_io_u");
789 td->zone_bytes += io_u->buflen;
793 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
822 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
824 __io_u_mark_map(td->ts.io_u_submit, nr);
825 td->ts.total_submit++;
828 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
830 __io_u_mark_map(td->ts.io_u_complete, nr);
831 td->ts.total_complete++;
834 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
838 switch (td->cur_depth) {
860 td->ts.io_u_map[idx] += nr;
863 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
900 assert(idx < FIO_IO_U_LAT_U_NR);
901 td->ts.io_u_lat_u[idx]++;
904 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
945 assert(idx < FIO_IO_U_LAT_M_NR);
946 td->ts.io_u_lat_m[idx]++;
949 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
952 io_u_mark_lat_usec(td, usec);
954 io_u_mark_lat_msec(td, usec / 1000);
958 * Get next file to service by choosing one at random
960 static struct fio_file *get_next_file_rand(struct thread_data *td,
961 enum fio_file_flags goodf,
962 enum fio_file_flags badf)
971 r = __rand(&td->next_file_state);
972 fno = (unsigned int) ((double) td->o.nr_files
973 * (r / (FRAND_MAX + 1.0)));
976 if (fio_file_done(f))
979 if (!fio_file_open(f)) {
982 if (td->nr_open_files >= td->o.open_files)
983 return ERR_PTR(-EBUSY);
985 err = td_io_open_file(td, f);
991 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
992 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
996 td_io_close_file(td, f);
1001 * Get next file to service by doing round robin between all available ones
1003 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1006 unsigned int old_next_file = td->next_file;
1012 f = td->files[td->next_file];
1015 if (td->next_file >= td->o.nr_files)
1018 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1019 if (fio_file_done(f)) {
1024 if (!fio_file_open(f)) {
1027 if (td->nr_open_files >= td->o.open_files)
1028 return ERR_PTR(-EBUSY);
1030 err = td_io_open_file(td, f);
1032 dprint(FD_FILE, "error %d on open of %s\n",
1040 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1042 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1046 td_io_close_file(td, f);
1049 } while (td->next_file != old_next_file);
1051 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1055 static struct fio_file *__get_next_file(struct thread_data *td)
1059 assert(td->o.nr_files <= td->files_index);
1061 if (td->nr_done_files >= td->o.nr_files) {
1062 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1063 " nr_files=%d\n", td->nr_open_files,
1069 f = td->file_service_file;
1070 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1071 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1073 if (td->file_service_left--)
1077 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1078 td->o.file_service_type == FIO_FSERVICE_SEQ)
1079 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1081 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1086 td->file_service_file = f;
1087 td->file_service_left = td->file_service_nr - 1;
1090 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1092 dprint(FD_FILE, "get_next_file: NULL\n");
1096 static struct fio_file *get_next_file(struct thread_data *td)
1098 if (td->flags & TD_F_PROFILE_OPS) {
1099 struct prof_io_ops *ops = &td->prof_io_ops;
1101 if (ops->get_next_file)
1102 return ops->get_next_file(td);
1105 return __get_next_file(td);
1108 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1113 f = get_next_file(td);
1114 if (IS_ERR_OR_NULL(f))
1120 if (!fill_io_u(td, io_u))
1123 put_file_log(td, f);
1124 td_io_close_file(td, f);
1126 fio_file_set_done(f);
1127 td->nr_done_files++;
1128 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1129 td->nr_done_files, td->o.nr_files);
1135 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1136 unsigned long tusec, unsigned long max_usec)
1139 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1140 td_verror(td, ETIMEDOUT, "max latency exceeded");
1141 icd->error = ETIMEDOUT;
1144 static void lat_new_cycle(struct thread_data *td)
1146 fio_gettime(&td->latency_ts, NULL);
1147 td->latency_ios = ddir_rw_sum(td->io_blocks);
1148 td->latency_failed = 0;
1152 * We had an IO outside the latency target. Reduce the queue depth. If we
1153 * are at QD=1, then it's time to give up.
1155 static int __lat_target_failed(struct thread_data *td)
1157 if (td->latency_qd == 1)
1160 td->latency_qd_high = td->latency_qd;
1162 if (td->latency_qd == td->latency_qd_low)
1163 td->latency_qd_low--;
1165 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1167 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1170 * When we ramp QD down, quiesce existing IO to prevent
1171 * a storm of ramp downs due to pending higher depth.
1178 static int lat_target_failed(struct thread_data *td)
1180 if (td->o.latency_percentile.u.f == 100.0)
1181 return __lat_target_failed(td);
1183 td->latency_failed++;
1187 void lat_target_init(struct thread_data *td)
1189 td->latency_end_run = 0;
1191 if (td->o.latency_target) {
1192 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1193 fio_gettime(&td->latency_ts, NULL);
1195 td->latency_qd_high = td->o.iodepth;
1196 td->latency_qd_low = 1;
1197 td->latency_ios = ddir_rw_sum(td->io_blocks);
1199 td->latency_qd = td->o.iodepth;
1202 void lat_target_reset(struct thread_data *td)
1204 if (!td->latency_end_run)
1205 lat_target_init(td);
1208 static void lat_target_success(struct thread_data *td)
1210 const unsigned int qd = td->latency_qd;
1211 struct thread_options *o = &td->o;
1213 td->latency_qd_low = td->latency_qd;
1216 * If we haven't failed yet, we double up to a failing value instead
1217 * of bisecting from highest possible queue depth. If we have set
1218 * a limit other than td->o.iodepth, bisect between that.
1220 if (td->latency_qd_high != o->iodepth)
1221 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1223 td->latency_qd *= 2;
1225 if (td->latency_qd > o->iodepth)
1226 td->latency_qd = o->iodepth;
1228 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1231 * Same as last one, we are done. Let it run a latency cycle, so
1232 * we get only the results from the targeted depth.
1234 if (td->latency_qd == qd) {
1235 if (td->latency_end_run) {
1236 dprint(FD_RATE, "We are done\n");
1239 dprint(FD_RATE, "Quiesce and final run\n");
1241 td->latency_end_run = 1;
1242 reset_all_stats(td);
1251 * Check if we can bump the queue depth
1253 void lat_target_check(struct thread_data *td)
1255 uint64_t usec_window;
1259 usec_window = utime_since_now(&td->latency_ts);
1260 if (usec_window < td->o.latency_window)
1263 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1264 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1265 success_ios *= 100.0;
1267 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1269 if (success_ios >= td->o.latency_percentile.u.f)
1270 lat_target_success(td);
1272 __lat_target_failed(td);
1276 * If latency target is enabled, we might be ramping up or down and not
1277 * using the full queue depth available.
1279 int queue_full(const struct thread_data *td)
1281 const int qempty = io_u_qempty(&td->io_u_freelist);
1285 if (!td->o.latency_target)
1288 return td->cur_depth >= td->latency_qd;
1291 struct io_u *__get_io_u(struct thread_data *td)
1293 struct io_u *io_u = NULL;
1301 if (!io_u_rempty(&td->io_u_requeues))
1302 io_u = io_u_rpop(&td->io_u_requeues);
1303 else if (!queue_full(td)) {
1304 io_u = io_u_qpop(&td->io_u_freelist);
1309 io_u->end_io = NULL;
1313 assert(io_u->flags & IO_U_F_FREE);
1314 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1315 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1319 io_u->acct_ddir = -1;
1321 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1323 } else if (td->o.verify_async) {
1325 * We ran out, wait for async verify threads to finish and
1328 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1336 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1338 if (!(td->flags & TD_F_TRIM_BACKLOG))
1341 if (td->trim_entries) {
1344 if (td->trim_batch) {
1347 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1348 td->last_ddir != DDIR_READ) {
1349 td->trim_batch = td->o.trim_batch;
1350 if (!td->trim_batch)
1351 td->trim_batch = td->o.trim_backlog;
1355 if (get_trim && !get_next_trim(td, io_u))
1362 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1364 if (!(td->flags & TD_F_VER_BACKLOG))
1367 if (td->io_hist_len) {
1370 if (td->verify_batch)
1372 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1373 td->last_ddir != DDIR_READ) {
1374 td->verify_batch = td->o.verify_batch;
1375 if (!td->verify_batch)
1376 td->verify_batch = td->o.verify_backlog;
1380 if (get_verify && !get_next_verify(td, io_u)) {
1390 * Fill offset and start time into the buffer content, to prevent too
1391 * easy compressible data for simple de-dupe attempts. Do this for every
1392 * 512b block in the range, since that should be the smallest block size
1393 * we can expect from a device.
1395 static void small_content_scramble(struct io_u *io_u)
1397 unsigned int i, nr_blocks = io_u->buflen / 512;
1399 unsigned int offset;
1406 boffset = io_u->offset;
1407 io_u->buf_filled_len = 0;
1409 for (i = 0; i < nr_blocks; i++) {
1411 * Fill the byte offset into a "random" start offset of
1412 * the buffer, given by the product of the usec time
1413 * and the actual offset.
1415 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1416 offset &= ~(sizeof(uint64_t) - 1);
1417 if (offset >= 512 - sizeof(uint64_t))
1418 offset -= sizeof(uint64_t);
1419 memcpy(p + offset, &boffset, sizeof(boffset));
1421 end = p + 512 - sizeof(io_u->start_time);
1422 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1429 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1430 * etc. The returned io_u is fully ready to be prepped and submitted.
1432 struct io_u *get_io_u(struct thread_data *td)
1436 int do_scramble = 0;
1439 io_u = __get_io_u(td);
1441 dprint(FD_IO, "__get_io_u failed\n");
1445 if (check_get_verify(td, io_u))
1447 if (check_get_trim(td, io_u))
1451 * from a requeue, io_u already setup
1457 * If using an iolog, grab next piece if any available.
1459 if (td->flags & TD_F_READ_IOLOG) {
1460 if (read_iolog_get(td, io_u))
1462 } else if (set_io_u_file(td, io_u)) {
1464 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1470 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1474 assert(fio_file_open(f));
1476 if (ddir_rw(io_u->ddir)) {
1477 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1478 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1482 f->last_start[io_u->ddir] = io_u->offset;
1483 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1485 if (io_u->ddir == DDIR_WRITE) {
1486 if (td->flags & TD_F_REFILL_BUFFERS) {
1487 io_u_fill_buffer(td, io_u,
1488 td->o.min_bs[DDIR_WRITE],
1490 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1491 !(td->flags & TD_F_COMPRESS))
1493 if (td->flags & TD_F_VER_NONE) {
1494 populate_verify_io_u(td, io_u);
1497 } else if (io_u->ddir == DDIR_READ) {
1499 * Reset the buf_filled parameters so next time if the
1500 * buffer is used for writes it is refilled.
1502 io_u->buf_filled_len = 0;
1507 * Set io data pointers.
1509 io_u->xfer_buf = io_u->buf;
1510 io_u->xfer_buflen = io_u->buflen;
1514 if (!td_io_prep(td, io_u)) {
1515 if (!td->o.disable_slat)
1516 fio_gettime(&io_u->start_time, NULL);
1518 small_content_scramble(io_u);
1522 dprint(FD_IO, "get_io_u failed\n");
1524 return ERR_PTR(ret);
1527 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1529 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1531 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1534 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1535 io_u->file ? " on file " : "",
1536 io_u->file ? io_u->file->file_name : "",
1537 strerror(io_u->error),
1538 io_ddir_name(io_u->ddir),
1539 io_u->offset, io_u->xfer_buflen);
1542 td_verror(td, io_u->error, "io_u error");
1545 static inline int gtod_reduce(struct thread_data *td)
1547 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1548 && td->o.disable_bw;
1551 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1552 struct io_completion_data *icd,
1553 const enum fio_ddir idx, unsigned int bytes)
1555 unsigned long lusec = 0;
1557 if (!gtod_reduce(td))
1558 lusec = utime_since(&io_u->issue_time, &icd->time);
1560 if (!td->o.disable_lat) {
1561 unsigned long tusec;
1563 tusec = utime_since(&io_u->start_time, &icd->time);
1564 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1566 if (td->flags & TD_F_PROFILE_OPS) {
1567 struct prof_io_ops *ops = &td->prof_io_ops;
1570 icd->error = ops->io_u_lat(td, tusec);
1573 if (td->o.max_latency && tusec > td->o.max_latency)
1574 lat_fatal(td, icd, tusec, td->o.max_latency);
1575 if (td->o.latency_target && tusec > td->o.latency_target) {
1576 if (lat_target_failed(td))
1577 lat_fatal(td, icd, tusec, td->o.latency_target);
1581 if (!td->o.disable_clat) {
1582 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1583 io_u_mark_latency(td, lusec);
1586 if (!td->o.disable_bw)
1587 add_bw_sample(td, idx, bytes, &icd->time);
1589 if (!gtod_reduce(td))
1590 add_iops_sample(td, idx, bytes, &icd->time);
1593 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1595 uint64_t secs, remainder, bps, bytes;
1597 bytes = td->this_io_bytes[ddir];
1598 bps = td->rate_bps[ddir];
1600 remainder = bytes % bps;
1601 return remainder * 1000000 / bps + secs * 1000000;
1604 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1605 struct io_completion_data *icd)
1607 struct io_u *io_u = *io_u_ptr;
1608 enum fio_ddir ddir = io_u->ddir;
1609 struct fio_file *f = io_u->file;
1611 dprint_io_u(io_u, "io complete");
1614 assert(io_u->flags & IO_U_F_FLIGHT);
1615 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1618 * Mark IO ok to verify
1622 * Remove errored entry from the verification list
1625 unlog_io_piece(td, io_u);
1627 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1634 if (ddir_sync(ddir)) {
1635 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 = ddir;
1646 if (!io_u->error && ddir_rw(ddir)) {
1647 unsigned int bytes = io_u->buflen - io_u->resid;
1648 const enum fio_ddir oddir = ddir ^ 1;
1651 td->io_blocks[ddir]++;
1652 td->this_io_blocks[ddir]++;
1653 td->io_bytes[ddir] += bytes;
1655 if (!(io_u->flags & IO_U_F_VER_LIST))
1656 td->this_io_bytes[ddir] += bytes;
1658 if (ddir == DDIR_WRITE) {
1660 if (f->first_write == -1ULL ||
1661 io_u->offset < f->first_write)
1662 f->first_write = io_u->offset;
1663 if (f->last_write == -1ULL ||
1664 ((io_u->offset + bytes) > f->last_write))
1665 f->last_write = io_u->offset + bytes;
1667 if (td->last_write_comp) {
1668 int idx = td->last_write_idx++;
1670 td->last_write_comp[idx] = io_u->offset;
1671 if (td->last_write_idx == td->o.iodepth)
1672 td->last_write_idx = 0;
1676 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1677 td->runstate == TD_VERIFYING)) {
1678 account_io_completion(td, io_u, icd, ddir, bytes);
1680 if (__should_check_rate(td, ddir)) {
1681 td->rate_pending_usleep[ddir] =
1682 (usec_for_io(td, ddir) -
1683 utime_since_now(&td->start));
1685 if (ddir != DDIR_TRIM &&
1686 __should_check_rate(td, oddir)) {
1687 td->rate_pending_usleep[oddir] =
1688 (usec_for_io(td, oddir) -
1689 utime_since_now(&td->start));
1693 icd->bytes_done[ddir] += bytes;
1696 ret = io_u->end_io(td, io_u_ptr);
1698 if (ret && !icd->error)
1701 } else if (io_u->error) {
1702 icd->error = io_u->error;
1703 io_u_log_error(td, io_u);
1706 enum error_type_bit eb = td_error_type(ddir, icd->error);
1708 if (!td_non_fatal_error(td, eb, icd->error))
1712 * If there is a non_fatal error, then add to the error count
1713 * and clear all the errors.
1715 update_error_count(td, icd->error);
1723 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1728 if (!gtod_reduce(td))
1729 fio_gettime(&icd->time, NULL);
1734 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1735 icd->bytes_done[ddir] = 0;
1738 static void ios_completed(struct thread_data *td,
1739 struct io_completion_data *icd)
1744 for (i = 0; i < icd->nr; i++) {
1745 io_u = td->io_ops->event(td, i);
1747 io_completed(td, &io_u, icd);
1755 * Complete a single io_u for the sync engines.
1757 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1760 struct io_completion_data icd;
1762 init_icd(td, &icd, 1);
1763 io_completed(td, &io_u, &icd);
1769 td_verror(td, icd.error, "io_u_sync_complete");
1776 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1777 bytes[ddir] += icd.bytes_done[ddir];
1784 * Called to complete min_events number of io for the async engines.
1786 int io_u_queued_complete(struct thread_data *td, int min_evts,
1789 struct io_completion_data icd;
1790 struct timespec *tvp = NULL;
1792 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1794 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1798 else if (min_evts > td->cur_depth)
1799 min_evts = td->cur_depth;
1801 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1803 td_verror(td, -ret, "td_io_getevents");
1808 init_icd(td, &icd, ret);
1809 ios_completed(td, &icd);
1811 td_verror(td, icd.error, "io_u_queued_complete");
1818 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1819 bytes[ddir] += icd.bytes_done[ddir];
1826 * Call when io_u is really queued, to update the submission latency.
1828 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1830 if (!td->o.disable_slat) {
1831 unsigned long slat_time;
1833 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1834 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1840 * See if we should reuse the last seed, if dedupe is enabled
1842 static struct frand_state *get_buf_state(struct thread_data *td)
1847 if (!td->o.dedupe_percentage)
1848 return &td->buf_state;
1849 else if (td->o.dedupe_percentage == 100)
1850 return &td->buf_state_prev;
1852 r = __rand(&td->dedupe_state);
1853 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1855 if (v <= td->o.dedupe_percentage)
1856 return &td->buf_state_prev;
1858 return &td->buf_state;
1861 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1863 if (rs == &td->buf_state)
1864 frand_copy(&td->buf_state_prev, rs);
1867 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1868 unsigned int max_bs)
1870 struct thread_options *o = &td->o;
1872 if (o->compress_percentage || o->dedupe_percentage) {
1873 unsigned int perc = td->o.compress_percentage;
1874 struct frand_state *rs;
1875 unsigned int left = max_bs;
1878 rs = get_buf_state(td);
1880 min_write = min(min_write, left);
1883 unsigned int seg = min_write;
1885 seg = min(min_write, td->o.compress_chunk);
1889 fill_random_buf_percentage(rs, buf, perc, seg,
1890 min_write, o->buffer_pattern,
1891 o->buffer_pattern_bytes);
1893 fill_random_buf(rs, buf, min_write);
1897 save_buf_state(td, rs);
1899 } else if (o->buffer_pattern_bytes)
1900 fill_buffer_pattern(td, buf, max_bs);
1902 memset(buf, 0, max_bs);
1906 * "randomly" fill the buffer contents
1908 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1909 unsigned int min_write, unsigned int max_bs)
1911 io_u->buf_filled_len = 0;
1912 fill_io_buffer(td, io_u->buf, min_write, max_bs);