13 #include "lib/axmap.h"
18 struct io_completion_data {
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timeval time; /* output */
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
30 static bool random_map_free(struct fio_file *f, const uint64_t block)
32 return !axmap_isset(f->io_axmap, block);
36 * Mark a given offset as used in the map.
38 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40 unsigned int min_bs = td->o.rw_min_bs;
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
55 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
61 assert(ddir_rw(ddir));
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 * -> not for now since there is code assuming it could go either.
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
72 max_size = td->o.zone_range;
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
85 struct flist_head list;
89 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
90 enum fio_ddir ddir, uint64_t *b,
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
98 r = __rand(&td->random_state);
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
102 *b = lastb * (r / (rand_max(&td->random_state) + 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 __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
156 *b = gauss_next(&f->gauss);
160 static int __get_next_rand_offset_zoned(struct thread_data *td,
161 struct fio_file *f, enum fio_ddir ddir,
164 unsigned int v, send, stotal;
165 uint64_t offset, lastb;
167 struct zone_split_index *zsi;
169 lastb = last_block(td, f, ddir);
173 if (!td->o.zone_split_nr[ddir]) {
175 return __get_next_rand_offset(td, f, ddir, b, lastb);
179 * Generate a value, v, between 1 and 100, both inclusive
181 v = rand32_between(&td->zone_state, 1, 100);
183 zsi = &td->zone_state_index[ddir][v - 1];
184 stotal = zsi->size_perc_prev;
185 send = zsi->size_perc;
188 * Should never happen
192 log_err("fio: bug in zoned generation\n");
199 * 'send' is some percentage below or equal to 100 that
200 * marks the end of the current IO range. 'stotal' marks
201 * the start, in percent.
204 offset = stotal * lastb / 100ULL;
208 lastb = lastb * (send - stotal) / 100ULL;
211 * Generate index from 0..send-of-lastb
213 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
217 * Add our start offset, if any
225 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
227 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
228 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
230 return r1->off - r2->off;
233 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
234 enum fio_ddir ddir, uint64_t *b)
236 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
239 lastb = last_block(td, f, ddir);
243 return __get_next_rand_offset(td, f, ddir, b, lastb);
244 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
245 return __get_next_rand_offset_zipf(td, f, ddir, b);
246 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
247 return __get_next_rand_offset_pareto(td, f, ddir, b);
248 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
249 return __get_next_rand_offset_gauss(td, f, ddir, b);
250 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
251 return __get_next_rand_offset_zoned(td, f, ddir, b);
253 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
258 * Sort the reads for a verify phase in batches of verifysort_nr, if
261 static inline bool should_sort_io(struct thread_data *td)
263 if (!td->o.verifysort_nr || !td->o.do_verify)
267 if (td->runstate != TD_VERIFYING)
269 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
270 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
276 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
280 if (td->o.perc_rand[ddir] == 100)
283 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
285 return v <= td->o.perc_rand[ddir];
288 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
289 enum fio_ddir ddir, uint64_t *b)
294 if (!should_sort_io(td))
295 return get_off_from_method(td, f, ddir, b);
297 if (!flist_empty(&td->next_rand_list)) {
299 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
306 for (i = 0; i < td->o.verifysort_nr; i++) {
307 r = malloc(sizeof(*r));
309 ret = get_off_from_method(td, f, ddir, &r->off);
315 flist_add(&r->list, &td->next_rand_list);
321 assert(!flist_empty(&td->next_rand_list));
322 flist_sort(NULL, &td->next_rand_list, flist_cmp);
326 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
327 enum fio_ddir ddir, uint64_t *b)
329 if (!get_next_rand_offset(td, f, ddir, b))
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
334 fio_file_reset(td, f);
335 if (!get_next_rand_offset(td, f, ddir, b))
339 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
340 f->file_name, (unsigned long long) f->last_pos[ddir],
341 (unsigned long long) f->real_file_size);
345 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
346 enum fio_ddir ddir, uint64_t *offset)
348 struct thread_options *o = &td->o;
350 assert(ddir_rw(ddir));
352 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
354 struct thread_options *o = &td->o;
355 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
357 if (io_size > f->last_pos[ddir])
358 f->last_pos[ddir] = 0;
360 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
363 if (f->last_pos[ddir] < f->real_file_size) {
366 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
367 if (f->real_file_size > f->io_size)
368 f->last_pos[ddir] = f->io_size;
370 f->last_pos[ddir] = f->real_file_size;
373 pos = f->last_pos[ddir] - f->file_offset;
374 if (pos && o->ddir_seq_add) {
375 pos += o->ddir_seq_add;
378 * If we reach beyond the end of the file
379 * with holed IO, wrap around to the
380 * beginning again. If we're doing backwards IO,
383 if (pos >= f->real_file_size) {
384 if (o->ddir_seq_add > 0)
385 pos = f->file_offset;
387 if (f->real_file_size > f->io_size)
390 pos = f->real_file_size;
392 pos += o->ddir_seq_add;
404 static int get_next_block(struct thread_data *td, struct io_u *io_u,
405 enum fio_ddir ddir, int rw_seq,
406 unsigned int *is_random)
408 struct fio_file *f = io_u->file;
412 assert(ddir_rw(ddir));
418 if (should_do_random(td, ddir)) {
419 ret = get_next_rand_block(td, f, ddir, &b);
423 io_u_set(td, io_u, IO_U_F_BUSY_OK);
424 ret = get_next_seq_offset(td, f, ddir, &offset);
426 ret = get_next_rand_block(td, f, ddir, &b);
430 ret = get_next_seq_offset(td, f, ddir, &offset);
433 io_u_set(td, io_u, IO_U_F_BUSY_OK);
436 if (td->o.rw_seq == RW_SEQ_SEQ) {
437 ret = get_next_seq_offset(td, f, ddir, &offset);
439 ret = get_next_rand_block(td, f, ddir, &b);
442 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
443 if (f->last_start[ddir] != -1ULL)
444 offset = f->last_start[ddir] - f->file_offset;
449 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
456 io_u->offset = offset;
458 io_u->offset = b * td->o.ba[ddir];
460 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
469 * For random io, generate a random new block and see if it's used. Repeat
470 * until we find a free one. For sequential io, just return the end of
471 * the last io issued.
473 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
474 unsigned int *is_random)
476 struct fio_file *f = io_u->file;
477 enum fio_ddir ddir = io_u->ddir;
480 assert(ddir_rw(ddir));
482 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
484 td->ddir_seq_nr = td->o.ddir_seq_nr;
487 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
490 if (io_u->offset >= f->io_size) {
491 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
492 (unsigned long long) io_u->offset,
493 (unsigned long long) f->io_size);
497 io_u->offset += f->file_offset;
498 if (io_u->offset >= f->real_file_size) {
499 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
500 (unsigned long long) io_u->offset,
501 (unsigned long long) f->real_file_size);
508 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
509 unsigned int *is_random)
511 if (td->flags & TD_F_PROFILE_OPS) {
512 struct prof_io_ops *ops = &td->prof_io_ops;
514 if (ops->fill_io_u_off)
515 return ops->fill_io_u_off(td, io_u, is_random);
518 return __get_next_offset(td, io_u, is_random);
521 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
524 struct fio_file *f = io_u->file;
526 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
529 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
530 unsigned int is_random)
532 int ddir = io_u->ddir;
533 unsigned int buflen = 0;
534 unsigned int minbs, maxbs;
535 uint64_t frand_max, r;
537 assert(ddir_rw(ddir));
539 if (td->o.bs_is_seq_rand)
540 ddir = is_random ? DDIR_WRITE: DDIR_READ;
542 minbs = td->o.min_bs[ddir];
543 maxbs = td->o.max_bs[ddir];
549 * If we can't satisfy the min block size from here, then fail
551 if (!io_u_fits(td, io_u, minbs))
554 frand_max = rand_max(&td->bsrange_state);
556 r = __rand(&td->bsrange_state);
558 if (!td->o.bssplit_nr[ddir]) {
559 buflen = 1 + (unsigned int) ((double) maxbs *
560 (r / (frand_max + 1.0)));
567 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
568 struct bssplit *bsp = &td->o.bssplit[ddir][i];
574 if ((r / perc <= frand_max / 100ULL) &&
575 io_u_fits(td, io_u, buflen))
580 if (!td->o.bs_unaligned && is_power_of_2(minbs))
581 buflen &= ~(minbs - 1);
583 } while (!io_u_fits(td, io_u, buflen));
588 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
589 unsigned int is_random)
591 if (td->flags & TD_F_PROFILE_OPS) {
592 struct prof_io_ops *ops = &td->prof_io_ops;
594 if (ops->fill_io_u_size)
595 return ops->fill_io_u_size(td, io_u, is_random);
598 return __get_next_buflen(td, io_u, is_random);
601 static void set_rwmix_bytes(struct thread_data *td)
606 * we do time or byte based switch. this is needed because
607 * buffered writes may issue a lot quicker than they complete,
608 * whereas reads do not.
610 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
611 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
614 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
618 v = rand32_between(&td->rwmix_state, 1, 100);
620 if (v <= td->o.rwmix[DDIR_READ])
626 int io_u_quiesce(struct thread_data *td)
631 * We are going to sleep, ensure that we flush anything pending as
632 * not to skew our latency numbers.
634 * Changed to only monitor 'in flight' requests here instead of the
635 * td->cur_depth, b/c td->cur_depth does not accurately represent
636 * io's that have been actually submitted to an async engine,
637 * and cur_depth is meaningless for sync engines.
639 if (td->io_u_queued || td->cur_depth) {
642 ret = td_io_commit(td);
645 while (td->io_u_in_flight) {
648 ret = io_u_queued_complete(td, 1);
653 if (td->flags & TD_F_REGROW_LOGS)
659 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
661 enum fio_ddir odir = ddir ^ 1;
665 assert(ddir_rw(ddir));
666 now = utime_since_now(&td->start);
669 * if rate_next_io_time is in the past, need to catch up to rate
671 if (td->rate_next_io_time[ddir] <= now)
675 * We are ahead of rate in this direction. See if we
678 if (td_rw(td) && td->o.rwmix[odir]) {
680 * Other direction is behind rate, switch
682 if (td->rate_next_io_time[odir] <= now)
686 * Both directions are ahead of rate. sleep the min
687 * switch if necissary
689 if (td->rate_next_io_time[ddir] <=
690 td->rate_next_io_time[odir]) {
691 usec = td->rate_next_io_time[ddir] - now;
693 usec = td->rate_next_io_time[odir] - now;
697 usec = td->rate_next_io_time[ddir] - now;
699 if (td->o.io_submit_mode == IO_MODE_INLINE)
702 usec = usec_sleep(td, usec);
708 * Return the data direction for the next io_u. If the job is a
709 * mixed read/write workload, check the rwmix cycle and switch if
712 static enum fio_ddir get_rw_ddir(struct thread_data *td)
717 * see if it's time to fsync
719 if (td->o.fsync_blocks &&
720 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
721 td->io_issues[DDIR_WRITE] && should_fsync(td))
725 * see if it's time to fdatasync
727 if (td->o.fdatasync_blocks &&
728 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
729 td->io_issues[DDIR_WRITE] && should_fsync(td))
730 return DDIR_DATASYNC;
733 * see if it's time to sync_file_range
735 if (td->sync_file_range_nr &&
736 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
737 td->io_issues[DDIR_WRITE] && should_fsync(td))
738 return DDIR_SYNC_FILE_RANGE;
742 * Check if it's time to seed a new data direction.
744 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
746 * Put a top limit on how many bytes we do for
747 * one data direction, to avoid overflowing the
750 ddir = get_rand_ddir(td);
752 if (ddir != td->rwmix_ddir)
755 td->rwmix_ddir = ddir;
757 ddir = td->rwmix_ddir;
758 } else if (td_read(td))
760 else if (td_write(td))
762 else if (td_trim(td))
767 td->rwmix_ddir = rate_ddir(td, ddir);
768 return td->rwmix_ddir;
771 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
773 enum fio_ddir ddir = get_rw_ddir(td);
775 if (td_trimwrite(td)) {
776 struct fio_file *f = io_u->file;
777 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
783 io_u->ddir = io_u->acct_ddir = ddir;
785 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
786 td->o.barrier_blocks &&
787 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
788 td->io_issues[DDIR_WRITE])
789 io_u_set(td, io_u, IO_U_F_BARRIER);
792 void put_file_log(struct thread_data *td, struct fio_file *f)
794 unsigned int ret = put_file(td, f);
797 td_verror(td, ret, "file close");
800 void put_io_u(struct thread_data *td, struct io_u *io_u)
807 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
808 put_file_log(td, io_u->file);
811 io_u_set(td, io_u, IO_U_F_FREE);
813 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
815 assert(!(td->flags & TD_F_CHILD));
817 io_u_qpush(&td->io_u_freelist, io_u);
819 td_io_u_free_notify(td);
822 void clear_io_u(struct thread_data *td, struct io_u *io_u)
824 io_u_clear(td, io_u, IO_U_F_FLIGHT);
828 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
830 struct io_u *__io_u = *io_u;
831 enum fio_ddir ddir = acct_ddir(__io_u);
833 dprint(FD_IO, "requeue %p\n", __io_u);
840 io_u_set(td, __io_u, IO_U_F_FREE);
841 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
842 td->io_issues[ddir]--;
844 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
845 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
847 assert(!(td->flags & TD_F_CHILD));
850 io_u_rpush(&td->io_u_requeues, __io_u);
852 td_io_u_free_notify(td);
856 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
858 unsigned int is_random;
860 if (td_ioengine_flagged(td, FIO_NOIO))
863 set_rw_ddir(td, io_u);
866 * fsync() or fdatasync() or trim etc, we are done
868 if (!ddir_rw(io_u->ddir))
872 * See if it's time to switch to a new zone
874 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
875 struct fio_file *f = io_u->file;
878 f->file_offset += td->o.zone_range + td->o.zone_skip;
881 * Wrap from the beginning, if we exceed the file size
883 if (f->file_offset >= f->real_file_size)
884 f->file_offset = f->real_file_size - f->file_offset;
885 f->last_pos[io_u->ddir] = f->file_offset;
886 td->io_skip_bytes += td->o.zone_skip;
890 * No log, let the seq/rand engine retrieve the next buflen and
893 if (get_next_offset(td, io_u, &is_random)) {
894 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
898 io_u->buflen = get_next_buflen(td, io_u, is_random);
900 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
904 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
905 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
906 dprint(FD_IO, " off=%llu/%lu > %llu\n",
907 (unsigned long long) io_u->offset, io_u->buflen,
908 (unsigned long long) io_u->file->real_file_size);
913 * mark entry before potentially trimming io_u
915 if (td_random(td) && file_randommap(td, io_u->file))
916 mark_random_map(td, io_u);
919 dprint_io_u(io_u, "fill_io_u");
920 td->zone_bytes += io_u->buflen;
924 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
953 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
955 __io_u_mark_map(td->ts.io_u_submit, nr);
956 td->ts.total_submit++;
959 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
961 __io_u_mark_map(td->ts.io_u_complete, nr);
962 td->ts.total_complete++;
965 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
969 switch (td->cur_depth) {
991 td->ts.io_u_map[idx] += nr;
994 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1031 assert(idx < FIO_IO_U_LAT_U_NR);
1032 td->ts.io_u_lat_u[idx]++;
1035 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1076 assert(idx < FIO_IO_U_LAT_M_NR);
1077 td->ts.io_u_lat_m[idx]++;
1080 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1083 io_u_mark_lat_usec(td, usec);
1085 io_u_mark_lat_msec(td, usec / 1000);
1088 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1090 unsigned long fileno;
1092 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1093 uint64_t frand_max = rand_max(&td->next_file_state);
1096 r = __rand(&td->next_file_state);
1097 return (unsigned int) ((double) td->o.nr_files
1098 * (r / (frand_max + 1.0)));
1101 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1102 fileno = zipf_next(&td->next_file_zipf);
1103 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1104 fileno = pareto_next(&td->next_file_zipf);
1105 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1106 fileno = gauss_next(&td->next_file_gauss);
1108 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1113 return fileno >> FIO_FSERVICE_SHIFT;
1117 * Get next file to service by choosing one at random
1119 static struct fio_file *get_next_file_rand(struct thread_data *td,
1120 enum fio_file_flags goodf,
1121 enum fio_file_flags badf)
1129 fno = __get_next_fileno_rand(td);
1132 if (fio_file_done(f))
1135 if (!fio_file_open(f)) {
1138 if (td->nr_open_files >= td->o.open_files)
1139 return ERR_PTR(-EBUSY);
1141 err = td_io_open_file(td, f);
1147 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1148 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1152 td_io_close_file(td, f);
1157 * Get next file to service by doing round robin between all available ones
1159 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1162 unsigned int old_next_file = td->next_file;
1168 f = td->files[td->next_file];
1171 if (td->next_file >= td->o.nr_files)
1174 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1175 if (fio_file_done(f)) {
1180 if (!fio_file_open(f)) {
1183 if (td->nr_open_files >= td->o.open_files)
1184 return ERR_PTR(-EBUSY);
1186 err = td_io_open_file(td, f);
1188 dprint(FD_FILE, "error %d on open of %s\n",
1196 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1198 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1202 td_io_close_file(td, f);
1205 } while (td->next_file != old_next_file);
1207 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1211 static struct fio_file *__get_next_file(struct thread_data *td)
1215 assert(td->o.nr_files <= td->files_index);
1217 if (td->nr_done_files >= td->o.nr_files) {
1218 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1219 " nr_files=%d\n", td->nr_open_files,
1225 f = td->file_service_file;
1226 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1227 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1229 if (td->file_service_left--)
1233 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1234 td->o.file_service_type == FIO_FSERVICE_SEQ)
1235 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1237 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1242 td->file_service_file = f;
1243 td->file_service_left = td->file_service_nr - 1;
1246 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1248 dprint(FD_FILE, "get_next_file: NULL\n");
1252 static struct fio_file *get_next_file(struct thread_data *td)
1254 if (td->flags & TD_F_PROFILE_OPS) {
1255 struct prof_io_ops *ops = &td->prof_io_ops;
1257 if (ops->get_next_file)
1258 return ops->get_next_file(td);
1261 return __get_next_file(td);
1264 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1269 f = get_next_file(td);
1270 if (IS_ERR_OR_NULL(f))
1276 if (!fill_io_u(td, io_u))
1279 put_file_log(td, f);
1280 td_io_close_file(td, f);
1282 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1283 fio_file_reset(td, f);
1285 fio_file_set_done(f);
1286 td->nr_done_files++;
1287 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1288 td->nr_done_files, td->o.nr_files);
1295 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1296 unsigned long tusec, unsigned long max_usec)
1299 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1300 td_verror(td, ETIMEDOUT, "max latency exceeded");
1301 icd->error = ETIMEDOUT;
1304 static void lat_new_cycle(struct thread_data *td)
1306 fio_gettime(&td->latency_ts, NULL);
1307 td->latency_ios = ddir_rw_sum(td->io_blocks);
1308 td->latency_failed = 0;
1312 * We had an IO outside the latency target. Reduce the queue depth. If we
1313 * are at QD=1, then it's time to give up.
1315 static bool __lat_target_failed(struct thread_data *td)
1317 if (td->latency_qd == 1)
1320 td->latency_qd_high = td->latency_qd;
1322 if (td->latency_qd == td->latency_qd_low)
1323 td->latency_qd_low--;
1325 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1327 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1330 * When we ramp QD down, quiesce existing IO to prevent
1331 * a storm of ramp downs due to pending higher depth.
1338 static bool lat_target_failed(struct thread_data *td)
1340 if (td->o.latency_percentile.u.f == 100.0)
1341 return __lat_target_failed(td);
1343 td->latency_failed++;
1347 void lat_target_init(struct thread_data *td)
1349 td->latency_end_run = 0;
1351 if (td->o.latency_target) {
1352 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1353 fio_gettime(&td->latency_ts, NULL);
1355 td->latency_qd_high = td->o.iodepth;
1356 td->latency_qd_low = 1;
1357 td->latency_ios = ddir_rw_sum(td->io_blocks);
1359 td->latency_qd = td->o.iodepth;
1362 void lat_target_reset(struct thread_data *td)
1364 if (!td->latency_end_run)
1365 lat_target_init(td);
1368 static void lat_target_success(struct thread_data *td)
1370 const unsigned int qd = td->latency_qd;
1371 struct thread_options *o = &td->o;
1373 td->latency_qd_low = td->latency_qd;
1376 * If we haven't failed yet, we double up to a failing value instead
1377 * of bisecting from highest possible queue depth. If we have set
1378 * a limit other than td->o.iodepth, bisect between that.
1380 if (td->latency_qd_high != o->iodepth)
1381 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1383 td->latency_qd *= 2;
1385 if (td->latency_qd > o->iodepth)
1386 td->latency_qd = o->iodepth;
1388 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1391 * Same as last one, we are done. Let it run a latency cycle, so
1392 * we get only the results from the targeted depth.
1394 if (td->latency_qd == qd) {
1395 if (td->latency_end_run) {
1396 dprint(FD_RATE, "We are done\n");
1399 dprint(FD_RATE, "Quiesce and final run\n");
1401 td->latency_end_run = 1;
1402 reset_all_stats(td);
1411 * Check if we can bump the queue depth
1413 void lat_target_check(struct thread_data *td)
1415 uint64_t usec_window;
1419 usec_window = utime_since_now(&td->latency_ts);
1420 if (usec_window < td->o.latency_window)
1423 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1424 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1425 success_ios *= 100.0;
1427 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1429 if (success_ios >= td->o.latency_percentile.u.f)
1430 lat_target_success(td);
1432 __lat_target_failed(td);
1436 * If latency target is enabled, we might be ramping up or down and not
1437 * using the full queue depth available.
1439 bool queue_full(const struct thread_data *td)
1441 const int qempty = io_u_qempty(&td->io_u_freelist);
1445 if (!td->o.latency_target)
1448 return td->cur_depth >= td->latency_qd;
1451 struct io_u *__get_io_u(struct thread_data *td)
1453 struct io_u *io_u = NULL;
1461 if (!io_u_rempty(&td->io_u_requeues))
1462 io_u = io_u_rpop(&td->io_u_requeues);
1463 else if (!queue_full(td)) {
1464 io_u = io_u_qpop(&td->io_u_freelist);
1469 io_u->end_io = NULL;
1473 assert(io_u->flags & IO_U_F_FREE);
1474 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1475 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1479 io_u->acct_ddir = -1;
1481 assert(!(td->flags & TD_F_CHILD));
1482 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1484 } else if (td_async_processing(td)) {
1486 * We ran out, wait for async verify threads to finish and
1489 assert(!(td->flags & TD_F_CHILD));
1490 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1498 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1500 if (!(td->flags & TD_F_TRIM_BACKLOG))
1503 if (td->trim_entries) {
1506 if (td->trim_batch) {
1509 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1510 td->last_ddir != DDIR_READ) {
1511 td->trim_batch = td->o.trim_batch;
1512 if (!td->trim_batch)
1513 td->trim_batch = td->o.trim_backlog;
1517 if (get_trim && get_next_trim(td, io_u))
1524 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1526 if (!(td->flags & TD_F_VER_BACKLOG))
1529 if (td->io_hist_len) {
1532 if (td->verify_batch)
1534 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1535 td->last_ddir != DDIR_READ) {
1536 td->verify_batch = td->o.verify_batch;
1537 if (!td->verify_batch)
1538 td->verify_batch = td->o.verify_backlog;
1542 if (get_verify && !get_next_verify(td, io_u)) {
1552 * Fill offset and start time into the buffer content, to prevent too
1553 * easy compressible data for simple de-dupe attempts. Do this for every
1554 * 512b block in the range, since that should be the smallest block size
1555 * we can expect from a device.
1557 static void small_content_scramble(struct io_u *io_u)
1559 unsigned int i, nr_blocks = io_u->buflen / 512;
1561 unsigned int offset;
1568 boffset = io_u->offset;
1569 io_u->buf_filled_len = 0;
1571 for (i = 0; i < nr_blocks; i++) {
1573 * Fill the byte offset into a "random" start offset of
1574 * the buffer, given by the product of the usec time
1575 * and the actual offset.
1577 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1578 offset &= ~(sizeof(uint64_t) - 1);
1579 if (offset >= 512 - sizeof(uint64_t))
1580 offset -= sizeof(uint64_t);
1581 memcpy(p + offset, &boffset, sizeof(boffset));
1583 end = p + 512 - sizeof(io_u->start_time);
1584 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1591 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1592 * etc. The returned io_u is fully ready to be prepped and submitted.
1594 struct io_u *get_io_u(struct thread_data *td)
1598 int do_scramble = 0;
1601 io_u = __get_io_u(td);
1603 dprint(FD_IO, "__get_io_u failed\n");
1607 if (check_get_verify(td, io_u))
1609 if (check_get_trim(td, io_u))
1613 * from a requeue, io_u already setup
1619 * If using an iolog, grab next piece if any available.
1621 if (td->flags & TD_F_READ_IOLOG) {
1622 if (read_iolog_get(td, io_u))
1624 } else if (set_io_u_file(td, io_u)) {
1626 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1632 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1636 assert(fio_file_open(f));
1638 if (ddir_rw(io_u->ddir)) {
1639 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1640 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1644 f->last_start[io_u->ddir] = io_u->offset;
1645 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1647 if (io_u->ddir == DDIR_WRITE) {
1648 if (td->flags & TD_F_REFILL_BUFFERS) {
1649 io_u_fill_buffer(td, io_u,
1650 td->o.min_bs[DDIR_WRITE],
1652 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1653 !(td->flags & TD_F_COMPRESS))
1655 if (td->flags & TD_F_VER_NONE) {
1656 populate_verify_io_u(td, io_u);
1659 } else if (io_u->ddir == DDIR_READ) {
1661 * Reset the buf_filled parameters so next time if the
1662 * buffer is used for writes it is refilled.
1664 io_u->buf_filled_len = 0;
1669 * Set io data pointers.
1671 io_u->xfer_buf = io_u->buf;
1672 io_u->xfer_buflen = io_u->buflen;
1676 if (!td_io_prep(td, io_u)) {
1677 if (!td->o.disable_lat)
1678 fio_gettime(&io_u->start_time, NULL);
1680 small_content_scramble(io_u);
1684 dprint(FD_IO, "get_io_u failed\n");
1686 return ERR_PTR(ret);
1689 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1691 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1693 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1696 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1697 io_u->file ? " on file " : "",
1698 io_u->file ? io_u->file->file_name : "",
1699 strerror(io_u->error),
1700 io_ddir_name(io_u->ddir),
1701 io_u->offset, io_u->xfer_buflen);
1703 if (td->io_ops->errdetails) {
1704 char *err = td->io_ops->errdetails(io_u);
1706 log_err("fio: %s\n", err);
1711 td_verror(td, io_u->error, "io_u error");
1714 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1716 __io_u_log_error(td, io_u);
1718 __io_u_log_error(td->parent, io_u);
1721 static inline bool gtod_reduce(struct thread_data *td)
1723 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1724 || td->o.gtod_reduce;
1727 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1728 struct io_completion_data *icd,
1729 const enum fio_ddir idx, unsigned int bytes)
1731 const int no_reduce = !gtod_reduce(td);
1732 unsigned long lusec = 0;
1741 lusec = utime_since(&io_u->issue_time, &icd->time);
1743 if (!td->o.disable_lat) {
1744 unsigned long tusec;
1746 tusec = utime_since(&io_u->start_time, &icd->time);
1747 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1749 if (td->flags & TD_F_PROFILE_OPS) {
1750 struct prof_io_ops *ops = &td->prof_io_ops;
1753 icd->error = ops->io_u_lat(td, tusec);
1756 if (td->o.max_latency && tusec > td->o.max_latency)
1757 lat_fatal(td, icd, tusec, td->o.max_latency);
1758 if (td->o.latency_target && tusec > td->o.latency_target) {
1759 if (lat_target_failed(td))
1760 lat_fatal(td, icd, tusec, td->o.latency_target);
1765 if (!td->o.disable_clat) {
1766 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1767 io_u_mark_latency(td, lusec);
1770 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1771 add_bw_sample(td, io_u, bytes, lusec);
1773 if (no_reduce && per_unit_log(td->iops_log))
1774 add_iops_sample(td, io_u, bytes);
1777 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1778 uint32_t *info = io_u_block_info(td, io_u);
1779 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1780 if (io_u->ddir == DDIR_TRIM) {
1781 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1782 BLOCK_INFO_TRIMS(*info) + 1);
1783 } else if (io_u->ddir == DDIR_WRITE) {
1784 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1791 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1792 uint64_t offset, unsigned int bytes)
1799 if (f->first_write == -1ULL || offset < f->first_write)
1800 f->first_write = offset;
1801 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1802 f->last_write = offset + bytes;
1804 if (!f->last_write_comp)
1807 idx = f->last_write_idx++;
1808 f->last_write_comp[idx] = offset;
1809 if (f->last_write_idx == td->o.iodepth)
1810 f->last_write_idx = 0;
1813 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1814 struct io_completion_data *icd)
1816 struct io_u *io_u = *io_u_ptr;
1817 enum fio_ddir ddir = io_u->ddir;
1818 struct fio_file *f = io_u->file;
1820 dprint_io_u(io_u, "io complete");
1822 assert(io_u->flags & IO_U_F_FLIGHT);
1823 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1826 * Mark IO ok to verify
1830 * Remove errored entry from the verification list
1833 unlog_io_piece(td, io_u);
1835 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1840 if (ddir_sync(ddir)) {
1841 td->last_was_sync = 1;
1843 f->first_write = -1ULL;
1844 f->last_write = -1ULL;
1849 td->last_was_sync = 0;
1850 td->last_ddir = ddir;
1852 if (!io_u->error && ddir_rw(ddir)) {
1853 unsigned int bytes = io_u->buflen - io_u->resid;
1856 td->io_blocks[ddir]++;
1857 td->this_io_blocks[ddir]++;
1858 td->io_bytes[ddir] += bytes;
1860 if (!(io_u->flags & IO_U_F_VER_LIST))
1861 td->this_io_bytes[ddir] += bytes;
1863 if (ddir == DDIR_WRITE)
1864 file_log_write_comp(td, f, io_u->offset, bytes);
1866 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1867 td->runstate == TD_VERIFYING))
1868 account_io_completion(td, io_u, icd, ddir, bytes);
1870 icd->bytes_done[ddir] += bytes;
1873 ret = io_u->end_io(td, io_u_ptr);
1875 if (ret && !icd->error)
1878 } else if (io_u->error) {
1879 icd->error = io_u->error;
1880 io_u_log_error(td, io_u);
1883 enum error_type_bit eb = td_error_type(ddir, icd->error);
1885 if (!td_non_fatal_error(td, eb, icd->error))
1889 * If there is a non_fatal error, then add to the error count
1890 * and clear all the errors.
1892 update_error_count(td, icd->error);
1900 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1905 if (!gtod_reduce(td))
1906 fio_gettime(&icd->time, NULL);
1911 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1912 icd->bytes_done[ddir] = 0;
1915 static void ios_completed(struct thread_data *td,
1916 struct io_completion_data *icd)
1921 for (i = 0; i < icd->nr; i++) {
1922 io_u = td->io_ops->event(td, i);
1924 io_completed(td, &io_u, icd);
1932 * Complete a single io_u for the sync engines.
1934 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1936 struct io_completion_data icd;
1939 init_icd(td, &icd, 1);
1940 io_completed(td, &io_u, &icd);
1946 td_verror(td, icd.error, "io_u_sync_complete");
1950 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1951 td->bytes_done[ddir] += icd.bytes_done[ddir];
1957 * Called to complete min_events number of io for the async engines.
1959 int io_u_queued_complete(struct thread_data *td, int min_evts)
1961 struct io_completion_data icd;
1962 struct timespec *tvp = NULL;
1964 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1966 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
1970 else if (min_evts > td->cur_depth)
1971 min_evts = td->cur_depth;
1973 /* No worries, td_io_getevents fixes min and max if they are
1974 * set incorrectly */
1975 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1977 td_verror(td, -ret, "td_io_getevents");
1982 init_icd(td, &icd, ret);
1983 ios_completed(td, &icd);
1985 td_verror(td, icd.error, "io_u_queued_complete");
1989 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1990 td->bytes_done[ddir] += icd.bytes_done[ddir];
1996 * Call when io_u is really queued, to update the submission latency.
1998 void io_u_queued(struct thread_data *td, struct io_u *io_u)
2000 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2001 unsigned long slat_time;
2003 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
2008 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2014 * See if we should reuse the last seed, if dedupe is enabled
2016 static struct frand_state *get_buf_state(struct thread_data *td)
2020 if (!td->o.dedupe_percentage)
2021 return &td->buf_state;
2022 else if (td->o.dedupe_percentage == 100) {
2023 frand_copy(&td->buf_state_prev, &td->buf_state);
2024 return &td->buf_state;
2027 v = rand32_between(&td->dedupe_state, 1, 100);
2029 if (v <= td->o.dedupe_percentage)
2030 return &td->buf_state_prev;
2032 return &td->buf_state;
2035 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2037 if (td->o.dedupe_percentage == 100)
2038 frand_copy(rs, &td->buf_state_prev);
2039 else if (rs == &td->buf_state)
2040 frand_copy(&td->buf_state_prev, rs);
2043 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2044 unsigned int max_bs)
2046 struct thread_options *o = &td->o;
2048 if (o->compress_percentage || o->dedupe_percentage) {
2049 unsigned int perc = td->o.compress_percentage;
2050 struct frand_state *rs;
2051 unsigned int left = max_bs;
2052 unsigned int this_write;
2055 rs = get_buf_state(td);
2057 min_write = min(min_write, left);
2060 this_write = min_not_zero(min_write,
2061 td->o.compress_chunk);
2063 fill_random_buf_percentage(rs, buf, perc,
2064 this_write, this_write,
2066 o->buffer_pattern_bytes);
2068 fill_random_buf(rs, buf, min_write);
2069 this_write = min_write;
2074 save_buf_state(td, rs);
2076 } else if (o->buffer_pattern_bytes)
2077 fill_buffer_pattern(td, buf, max_bs);
2078 else if (o->zero_buffers)
2079 memset(buf, 0, max_bs);
2081 fill_random_buf(get_buf_state(td), buf, max_bs);
2085 * "randomly" fill the buffer contents
2087 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2088 unsigned int min_write, unsigned int max_bs)
2090 io_u->buf_filled_len = 0;
2091 fill_io_buffer(td, io_u->buf, min_write, max_bs);