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) {
764 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
765 io_u->file->last_pos = io_u->file->file_offset;
766 td->io_skip_bytes += td->o.zone_skip;
770 * No log, let the seq/rand engine retrieve the next buflen and
773 if (get_next_offset(td, io_u, &is_random)) {
774 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
778 io_u->buflen = get_next_buflen(td, io_u, is_random);
780 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
784 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
785 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
786 dprint(FD_IO, " off=%llu/%lu > %llu\n",
787 (unsigned long long) io_u->offset, io_u->buflen,
788 (unsigned long long) io_u->file->real_file_size);
793 * mark entry before potentially trimming io_u
795 if (td_random(td) && file_randommap(td, io_u->file))
796 mark_random_map(td, io_u);
799 dprint_io_u(io_u, "fill_io_u");
800 td->zone_bytes += io_u->buflen;
804 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
833 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
835 __io_u_mark_map(td->ts.io_u_submit, nr);
836 td->ts.total_submit++;
839 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
841 __io_u_mark_map(td->ts.io_u_complete, nr);
842 td->ts.total_complete++;
845 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
849 switch (td->cur_depth) {
871 td->ts.io_u_map[idx] += nr;
874 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
911 assert(idx < FIO_IO_U_LAT_U_NR);
912 td->ts.io_u_lat_u[idx]++;
915 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
956 assert(idx < FIO_IO_U_LAT_M_NR);
957 td->ts.io_u_lat_m[idx]++;
960 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
963 io_u_mark_lat_usec(td, usec);
965 io_u_mark_lat_msec(td, usec / 1000);
969 * Get next file to service by choosing one at random
971 static struct fio_file *get_next_file_rand(struct thread_data *td,
972 enum fio_file_flags goodf,
973 enum fio_file_flags badf)
982 if (td->o.use_os_rand) {
983 r = os_random_long(&td->next_file_state);
984 fno = (unsigned int) ((double) td->o.nr_files
985 * (r / (OS_RAND_MAX + 1.0)));
987 r = __rand(&td->__next_file_state);
988 fno = (unsigned int) ((double) td->o.nr_files
989 * (r / (FRAND_MAX + 1.0)));
993 if (fio_file_done(f))
996 if (!fio_file_open(f)) {
999 if (td->nr_open_files >= td->o.open_files)
1000 return ERR_PTR(-EBUSY);
1002 err = td_io_open_file(td, f);
1008 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1009 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1013 td_io_close_file(td, f);
1018 * Get next file to service by doing round robin between all available ones
1020 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1023 unsigned int old_next_file = td->next_file;
1029 f = td->files[td->next_file];
1032 if (td->next_file >= td->o.nr_files)
1035 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1036 if (fio_file_done(f)) {
1041 if (!fio_file_open(f)) {
1044 if (td->nr_open_files >= td->o.open_files)
1045 return ERR_PTR(-EBUSY);
1047 err = td_io_open_file(td, f);
1049 dprint(FD_FILE, "error %d on open of %s\n",
1057 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1059 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1063 td_io_close_file(td, f);
1066 } while (td->next_file != old_next_file);
1068 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1072 static struct fio_file *__get_next_file(struct thread_data *td)
1076 assert(td->o.nr_files <= td->files_index);
1078 if (td->nr_done_files >= td->o.nr_files) {
1079 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1080 " nr_files=%d\n", td->nr_open_files,
1086 f = td->file_service_file;
1087 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1088 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1090 if (td->file_service_left--)
1094 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1095 td->o.file_service_type == FIO_FSERVICE_SEQ)
1096 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1098 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1103 td->file_service_file = f;
1104 td->file_service_left = td->file_service_nr - 1;
1107 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1109 dprint(FD_FILE, "get_next_file: NULL\n");
1113 static struct fio_file *get_next_file(struct thread_data *td)
1115 if (td->flags & TD_F_PROFILE_OPS) {
1116 struct prof_io_ops *ops = &td->prof_io_ops;
1118 if (ops->get_next_file)
1119 return ops->get_next_file(td);
1122 return __get_next_file(td);
1125 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1130 f = get_next_file(td);
1131 if (IS_ERR_OR_NULL(f))
1137 if (!fill_io_u(td, io_u))
1140 put_file_log(td, f);
1141 td_io_close_file(td, f);
1143 fio_file_set_done(f);
1144 td->nr_done_files++;
1145 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1146 td->nr_done_files, td->o.nr_files);
1152 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1153 unsigned long tusec, unsigned long max_usec)
1156 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1157 td_verror(td, ETIMEDOUT, "max latency exceeded");
1158 icd->error = ETIMEDOUT;
1161 static void lat_new_cycle(struct thread_data *td)
1163 fio_gettime(&td->latency_ts, NULL);
1164 td->latency_ios = ddir_rw_sum(td->io_blocks);
1165 td->latency_failed = 0;
1169 * We had an IO outside the latency target. Reduce the queue depth. If we
1170 * are at QD=1, then it's time to give up.
1172 static int __lat_target_failed(struct thread_data *td)
1174 if (td->latency_qd == 1)
1177 td->latency_qd_high = td->latency_qd;
1179 if (td->latency_qd == td->latency_qd_low)
1180 td->latency_qd_low--;
1182 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1184 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1187 * When we ramp QD down, quiesce existing IO to prevent
1188 * a storm of ramp downs due to pending higher depth.
1195 static int lat_target_failed(struct thread_data *td)
1197 if (td->o.latency_percentile.u.f == 100.0)
1198 return __lat_target_failed(td);
1200 td->latency_failed++;
1204 void lat_target_init(struct thread_data *td)
1206 td->latency_end_run = 0;
1208 if (td->o.latency_target) {
1209 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1210 fio_gettime(&td->latency_ts, NULL);
1212 td->latency_qd_high = td->o.iodepth;
1213 td->latency_qd_low = 1;
1214 td->latency_ios = ddir_rw_sum(td->io_blocks);
1216 td->latency_qd = td->o.iodepth;
1219 void lat_target_reset(struct thread_data *td)
1221 if (!td->latency_end_run)
1222 lat_target_init(td);
1225 static void lat_target_success(struct thread_data *td)
1227 const unsigned int qd = td->latency_qd;
1228 struct thread_options *o = &td->o;
1230 td->latency_qd_low = td->latency_qd;
1233 * If we haven't failed yet, we double up to a failing value instead
1234 * of bisecting from highest possible queue depth. If we have set
1235 * a limit other than td->o.iodepth, bisect between that.
1237 if (td->latency_qd_high != o->iodepth)
1238 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1240 td->latency_qd *= 2;
1242 if (td->latency_qd > o->iodepth)
1243 td->latency_qd = o->iodepth;
1245 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1248 * Same as last one, we are done. Let it run a latency cycle, so
1249 * we get only the results from the targeted depth.
1251 if (td->latency_qd == qd) {
1252 if (td->latency_end_run) {
1253 dprint(FD_RATE, "We are done\n");
1256 dprint(FD_RATE, "Quiesce and final run\n");
1258 td->latency_end_run = 1;
1259 reset_all_stats(td);
1268 * Check if we can bump the queue depth
1270 void lat_target_check(struct thread_data *td)
1272 uint64_t usec_window;
1276 usec_window = utime_since_now(&td->latency_ts);
1277 if (usec_window < td->o.latency_window)
1280 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1281 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1282 success_ios *= 100.0;
1284 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1286 if (success_ios >= td->o.latency_percentile.u.f)
1287 lat_target_success(td);
1289 __lat_target_failed(td);
1293 * If latency target is enabled, we might be ramping up or down and not
1294 * using the full queue depth available.
1296 int queue_full(struct thread_data *td)
1298 const int qempty = io_u_qempty(&td->io_u_freelist);
1302 if (!td->o.latency_target)
1305 return td->cur_depth >= td->latency_qd;
1308 struct io_u *__get_io_u(struct thread_data *td)
1310 struct io_u *io_u = NULL;
1315 if (!io_u_rempty(&td->io_u_requeues))
1316 io_u = io_u_rpop(&td->io_u_requeues);
1317 else if (!queue_full(td)) {
1318 io_u = io_u_qpop(&td->io_u_freelist);
1323 io_u->end_io = NULL;
1327 assert(io_u->flags & IO_U_F_FREE);
1328 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1329 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1333 io_u->acct_ddir = -1;
1335 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1337 } else if (td->o.verify_async) {
1339 * We ran out, wait for async verify threads to finish and
1342 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1350 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1352 if (!(td->flags & TD_F_TRIM_BACKLOG))
1355 if (td->trim_entries) {
1358 if (td->trim_batch) {
1361 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1362 td->last_ddir != DDIR_READ) {
1363 td->trim_batch = td->o.trim_batch;
1364 if (!td->trim_batch)
1365 td->trim_batch = td->o.trim_backlog;
1369 if (get_trim && !get_next_trim(td, io_u))
1376 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1378 if (!(td->flags & TD_F_VER_BACKLOG))
1381 if (td->io_hist_len) {
1384 if (td->verify_batch)
1386 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1387 td->last_ddir != DDIR_READ) {
1388 td->verify_batch = td->o.verify_batch;
1389 if (!td->verify_batch)
1390 td->verify_batch = td->o.verify_backlog;
1394 if (get_verify && !get_next_verify(td, io_u)) {
1404 * Fill offset and start time into the buffer content, to prevent too
1405 * easy compressible data for simple de-dupe attempts. Do this for every
1406 * 512b block in the range, since that should be the smallest block size
1407 * we can expect from a device.
1409 static void small_content_scramble(struct io_u *io_u)
1411 unsigned int i, nr_blocks = io_u->buflen / 512;
1413 unsigned int offset;
1420 boffset = io_u->offset;
1421 io_u->buf_filled_len = 0;
1423 for (i = 0; i < nr_blocks; i++) {
1425 * Fill the byte offset into a "random" start offset of
1426 * the buffer, given by the product of the usec time
1427 * and the actual offset.
1429 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1430 offset &= ~(sizeof(uint64_t) - 1);
1431 if (offset >= 512 - sizeof(uint64_t))
1432 offset -= sizeof(uint64_t);
1433 memcpy(p + offset, &boffset, sizeof(boffset));
1435 end = p + 512 - sizeof(io_u->start_time);
1436 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1443 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1444 * etc. The returned io_u is fully ready to be prepped and submitted.
1446 struct io_u *get_io_u(struct thread_data *td)
1450 int do_scramble = 0;
1453 io_u = __get_io_u(td);
1455 dprint(FD_IO, "__get_io_u failed\n");
1459 if (check_get_verify(td, io_u))
1461 if (check_get_trim(td, io_u))
1465 * from a requeue, io_u already setup
1471 * If using an iolog, grab next piece if any available.
1473 if (td->flags & TD_F_READ_IOLOG) {
1474 if (read_iolog_get(td, io_u))
1476 } else if (set_io_u_file(td, io_u)) {
1478 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1484 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1488 assert(fio_file_open(f));
1490 if (ddir_rw(io_u->ddir)) {
1491 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1492 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1496 f->last_start = io_u->offset;
1497 f->last_pos = io_u->offset + io_u->buflen;
1499 if (io_u->ddir == DDIR_WRITE) {
1500 if (td->flags & TD_F_REFILL_BUFFERS) {
1501 io_u_fill_buffer(td, io_u,
1502 td->o.min_bs[DDIR_WRITE],
1504 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1505 !(td->flags & TD_F_COMPRESS))
1507 if (td->flags & TD_F_VER_NONE) {
1508 populate_verify_io_u(td, io_u);
1511 } else if (io_u->ddir == DDIR_READ) {
1513 * Reset the buf_filled parameters so next time if the
1514 * buffer is used for writes it is refilled.
1516 io_u->buf_filled_len = 0;
1521 * Set io data pointers.
1523 io_u->xfer_buf = io_u->buf;
1524 io_u->xfer_buflen = io_u->buflen;
1528 if (!td_io_prep(td, io_u)) {
1529 if (!td->o.disable_slat)
1530 fio_gettime(&io_u->start_time, NULL);
1532 small_content_scramble(io_u);
1536 dprint(FD_IO, "get_io_u failed\n");
1538 return ERR_PTR(ret);
1541 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1543 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1545 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1548 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1549 io_u->file ? " on file " : "",
1550 io_u->file ? io_u->file->file_name : "",
1551 strerror(io_u->error),
1552 io_ddir_name(io_u->ddir),
1553 io_u->offset, io_u->xfer_buflen);
1556 td_verror(td, io_u->error, "io_u error");
1559 static inline int gtod_reduce(struct thread_data *td)
1561 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1562 && td->o.disable_bw;
1565 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1566 struct io_completion_data *icd,
1567 const enum fio_ddir idx, unsigned int bytes)
1569 unsigned long lusec = 0;
1571 if (!gtod_reduce(td))
1572 lusec = utime_since(&io_u->issue_time, &icd->time);
1574 if (!td->o.disable_lat) {
1575 unsigned long tusec;
1577 tusec = utime_since(&io_u->start_time, &icd->time);
1578 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1580 if (td->flags & TD_F_PROFILE_OPS) {
1581 struct prof_io_ops *ops = &td->prof_io_ops;
1584 icd->error = ops->io_u_lat(td, tusec);
1587 if (td->o.max_latency && tusec > td->o.max_latency)
1588 lat_fatal(td, icd, tusec, td->o.max_latency);
1589 if (td->o.latency_target && tusec > td->o.latency_target) {
1590 if (lat_target_failed(td))
1591 lat_fatal(td, icd, tusec, td->o.latency_target);
1595 if (!td->o.disable_clat) {
1596 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1597 io_u_mark_latency(td, lusec);
1600 if (!td->o.disable_bw)
1601 add_bw_sample(td, idx, bytes, &icd->time);
1603 if (!gtod_reduce(td))
1604 add_iops_sample(td, idx, bytes, &icd->time);
1607 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1609 uint64_t secs, remainder, bps, bytes;
1611 bytes = td->this_io_bytes[ddir];
1612 bps = td->rate_bps[ddir];
1614 remainder = bytes % bps;
1615 return remainder * 1000000 / bps + secs * 1000000;
1618 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1619 struct io_completion_data *icd)
1621 struct io_u *io_u = *io_u_ptr;
1622 enum fio_ddir ddir = io_u->ddir;
1623 struct fio_file *f = io_u->file;
1625 dprint_io_u(io_u, "io complete");
1628 assert(io_u->flags & IO_U_F_FLIGHT);
1629 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1632 * Mark IO ok to verify
1636 * Remove errored entry from the verification list
1639 unlog_io_piece(td, io_u);
1641 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1648 if (ddir_sync(ddir)) {
1649 td->last_was_sync = 1;
1651 f->first_write = -1ULL;
1652 f->last_write = -1ULL;
1657 td->last_was_sync = 0;
1658 td->last_ddir = ddir;
1660 if (!io_u->error && ddir_rw(ddir)) {
1661 unsigned int bytes = io_u->buflen - io_u->resid;
1662 const enum fio_ddir oddir = ddir ^ 1;
1665 td->io_blocks[ddir]++;
1666 td->this_io_blocks[ddir]++;
1667 td->io_bytes[ddir] += bytes;
1669 if (!(io_u->flags & IO_U_F_VER_LIST))
1670 td->this_io_bytes[ddir] += bytes;
1672 if (ddir == DDIR_WRITE && f) {
1673 if (f->first_write == -1ULL ||
1674 io_u->offset < f->first_write)
1675 f->first_write = io_u->offset;
1676 if (f->last_write == -1ULL ||
1677 ((io_u->offset + bytes) > f->last_write))
1678 f->last_write = io_u->offset + bytes;
1681 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1682 td->runstate == TD_VERIFYING)) {
1683 account_io_completion(td, io_u, icd, ddir, bytes);
1685 if (__should_check_rate(td, ddir)) {
1686 td->rate_pending_usleep[ddir] =
1687 (usec_for_io(td, ddir) -
1688 utime_since_now(&td->start));
1690 if (ddir != DDIR_TRIM &&
1691 __should_check_rate(td, oddir)) {
1692 td->rate_pending_usleep[oddir] =
1693 (usec_for_io(td, oddir) -
1694 utime_since_now(&td->start));
1698 icd->bytes_done[ddir] += bytes;
1701 ret = io_u->end_io(td, io_u_ptr);
1703 if (ret && !icd->error)
1706 } else if (io_u->error) {
1707 icd->error = io_u->error;
1708 io_u_log_error(td, io_u);
1711 enum error_type_bit eb = td_error_type(ddir, icd->error);
1713 if (!td_non_fatal_error(td, eb, icd->error))
1717 * If there is a non_fatal error, then add to the error count
1718 * and clear all the errors.
1720 update_error_count(td, icd->error);
1728 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1733 if (!gtod_reduce(td))
1734 fio_gettime(&icd->time, NULL);
1739 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1740 icd->bytes_done[ddir] = 0;
1743 static void ios_completed(struct thread_data *td,
1744 struct io_completion_data *icd)
1749 for (i = 0; i < icd->nr; i++) {
1750 io_u = td->io_ops->event(td, i);
1752 io_completed(td, &io_u, icd);
1760 * Complete a single io_u for the sync engines.
1762 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1765 struct io_completion_data icd;
1767 init_icd(td, &icd, 1);
1768 io_completed(td, &io_u, &icd);
1774 td_verror(td, icd.error, "io_u_sync_complete");
1781 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1782 bytes[ddir] += icd.bytes_done[ddir];
1789 * Called to complete min_events number of io for the async engines.
1791 int io_u_queued_complete(struct thread_data *td, int min_evts,
1794 struct io_completion_data icd;
1795 struct timespec *tvp = NULL;
1797 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1799 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1803 else if (min_evts > td->cur_depth)
1804 min_evts = td->cur_depth;
1806 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1808 td_verror(td, -ret, "td_io_getevents");
1813 init_icd(td, &icd, ret);
1814 ios_completed(td, &icd);
1816 td_verror(td, icd.error, "io_u_queued_complete");
1823 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1824 bytes[ddir] += icd.bytes_done[ddir];
1831 * Call when io_u is really queued, to update the submission latency.
1833 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1835 if (!td->o.disable_slat) {
1836 unsigned long slat_time;
1838 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1839 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1845 * See if we should reuse the last seed, if dedupe is enabled
1847 static struct frand_state *get_buf_state(struct thread_data *td)
1852 if (!td->o.dedupe_percentage)
1853 return &td->buf_state;
1854 else if (td->o.dedupe_percentage == 100)
1855 return &td->buf_state_prev;
1857 r = __rand(&td->dedupe_state);
1858 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1860 if (v <= td->o.dedupe_percentage)
1861 return &td->buf_state_prev;
1863 return &td->buf_state;
1866 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1868 if (rs == &td->buf_state)
1869 frand_copy(&td->buf_state_prev, rs);
1872 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1873 unsigned int max_bs)
1875 if (td->o.buffer_pattern_bytes)
1876 fill_buffer_pattern(td, buf, max_bs);
1877 else if (!td->o.zero_buffers) {
1878 unsigned int perc = td->o.compress_percentage;
1879 struct frand_state *rs;
1880 unsigned int left = max_bs;
1883 rs = get_buf_state(td);
1885 min_write = min(min_write, left);
1888 unsigned int seg = min_write;
1890 seg = min(min_write, td->o.compress_chunk);
1894 fill_random_buf_percentage(rs, buf, perc, seg,
1897 fill_random_buf(rs, buf, min_write);
1901 save_buf_state(td, rs);
1904 memset(buf, 0, max_bs);
1908 * "randomly" fill the buffer contents
1910 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1911 unsigned int min_write, unsigned int max_bs)
1913 io_u->buf_filled_len = 0;
1914 fill_io_buffer(td, io_u->buf, min_write, max_bs);