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?
66 max_size = f->io_size;
67 if (max_size > f->real_file_size)
68 max_size = f->real_file_size;
71 max_size = td->o.zone_range;
73 if (td->o.min_bs[ddir] > td->o.ba[ddir])
74 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
84 struct flist_head list;
88 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
89 enum fio_ddir ddir, uint64_t *b,
94 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
95 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
97 r = __rand(&td->random_state);
99 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
101 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
105 assert(fio_file_lfsr(f));
107 if (lfsr_next(&f->lfsr, &off))
114 * if we are not maintaining a random map, we are done.
116 if (!file_randommap(td, f))
120 * calculate map offset and check if it's free
122 if (random_map_free(f, *b))
125 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
126 (unsigned long long) *b);
128 *b = axmap_next_free(f->io_axmap, *b);
129 if (*b == (uint64_t) -1ULL)
135 static int __get_next_rand_offset_zipf(struct thread_data *td,
136 struct fio_file *f, enum fio_ddir ddir,
139 *b = zipf_next(&f->zipf);
143 static int __get_next_rand_offset_pareto(struct thread_data *td,
144 struct fio_file *f, enum fio_ddir ddir,
147 *b = pareto_next(&f->zipf);
151 static int __get_next_rand_offset_gauss(struct thread_data *td,
152 struct fio_file *f, enum fio_ddir ddir,
155 *b = gauss_next(&f->gauss);
159 static int __get_next_rand_offset_zoned(struct thread_data *td,
160 struct fio_file *f, enum fio_ddir ddir,
163 unsigned int v, send, stotal;
164 uint64_t offset, lastb;
166 struct zone_split_index *zsi;
168 lastb = last_block(td, f, ddir);
172 if (!td->o.zone_split_nr[ddir]) {
174 return __get_next_rand_offset(td, f, ddir, b, lastb);
178 * Generate a value, v, between 1 and 100, both inclusive
180 v = rand32_between(&td->zone_state, 1, 100);
182 zsi = &td->zone_state_index[ddir][v - 1];
183 stotal = zsi->size_perc_prev;
184 send = zsi->size_perc;
187 * Should never happen
191 log_err("fio: bug in zoned generation\n");
198 * 'send' is some percentage below or equal to 100 that
199 * marks the end of the current IO range. 'stotal' marks
200 * the start, in percent.
203 offset = stotal * lastb / 100ULL;
207 lastb = lastb * (send - stotal) / 100ULL;
210 * Generate index from 0..send-of-lastb
212 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
216 * Add our start offset, if any
224 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
226 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
227 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
229 return r1->off - r2->off;
232 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
233 enum fio_ddir ddir, uint64_t *b)
235 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
238 lastb = last_block(td, f, ddir);
242 return __get_next_rand_offset(td, f, ddir, b, lastb);
243 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
244 return __get_next_rand_offset_zipf(td, f, ddir, b);
245 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
246 return __get_next_rand_offset_pareto(td, f, ddir, b);
247 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
248 return __get_next_rand_offset_gauss(td, f, ddir, b);
249 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
250 return __get_next_rand_offset_zoned(td, f, ddir, b);
252 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
257 * Sort the reads for a verify phase in batches of verifysort_nr, if
260 static inline bool should_sort_io(struct thread_data *td)
262 if (!td->o.verifysort_nr || !td->o.do_verify)
266 if (td->runstate != TD_VERIFYING)
268 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
269 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
275 static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
279 if (td->o.perc_rand[ddir] == 100)
282 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
284 return v <= td->o.perc_rand[ddir];
287 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
288 enum fio_ddir ddir, uint64_t *b)
293 if (!should_sort_io(td))
294 return get_off_from_method(td, f, ddir, b);
296 if (!flist_empty(&td->next_rand_list)) {
298 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
305 for (i = 0; i < td->o.verifysort_nr; i++) {
306 r = malloc(sizeof(*r));
308 ret = get_off_from_method(td, f, ddir, &r->off);
314 flist_add(&r->list, &td->next_rand_list);
320 assert(!flist_empty(&td->next_rand_list));
321 flist_sort(NULL, &td->next_rand_list, flist_cmp);
325 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
326 enum fio_ddir ddir, uint64_t *b)
328 if (!get_next_rand_offset(td, f, ddir, b))
331 if (td->o.time_based) {
332 fio_file_reset(td, f);
333 if (!get_next_rand_offset(td, f, ddir, b))
337 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
338 f->file_name, (unsigned long long) f->last_pos[ddir],
339 (unsigned long long) f->real_file_size);
343 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
344 enum fio_ddir ddir, uint64_t *offset)
346 struct thread_options *o = &td->o;
348 assert(ddir_rw(ddir));
350 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
352 struct thread_options *o = &td->o;
353 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
355 if (io_size > f->last_pos[ddir])
356 f->last_pos[ddir] = 0;
358 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
361 if (f->last_pos[ddir] < f->real_file_size) {
364 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
365 f->last_pos[ddir] = f->real_file_size;
367 pos = f->last_pos[ddir] - f->file_offset;
368 if (pos && o->ddir_seq_add) {
369 pos += o->ddir_seq_add;
372 * If we reach beyond the end of the file
373 * with holed IO, wrap around to the
374 * beginning again. If we're doing backwards IO,
377 if (pos >= f->real_file_size) {
378 if (o->ddir_seq_add > 0)
379 pos = f->file_offset;
381 pos = f->real_file_size + o->ddir_seq_add;
392 static int get_next_block(struct thread_data *td, struct io_u *io_u,
393 enum fio_ddir ddir, int rw_seq,
394 unsigned int *is_random)
396 struct fio_file *f = io_u->file;
400 assert(ddir_rw(ddir));
406 if (should_do_random(td, ddir)) {
407 ret = get_next_rand_block(td, f, ddir, &b);
411 io_u_set(io_u, IO_U_F_BUSY_OK);
412 ret = get_next_seq_offset(td, f, ddir, &offset);
414 ret = get_next_rand_block(td, f, ddir, &b);
418 ret = get_next_seq_offset(td, f, ddir, &offset);
421 io_u_set(io_u, IO_U_F_BUSY_OK);
424 if (td->o.rw_seq == RW_SEQ_SEQ) {
425 ret = get_next_seq_offset(td, f, ddir, &offset);
427 ret = get_next_rand_block(td, f, ddir, &b);
430 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
431 if (f->last_start[ddir] != -1ULL)
432 offset = f->last_start[ddir] - f->file_offset;
437 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
444 io_u->offset = offset;
446 io_u->offset = b * td->o.ba[ddir];
448 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
457 * For random io, generate a random new block and see if it's used. Repeat
458 * until we find a free one. For sequential io, just return the end of
459 * the last io issued.
461 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
462 unsigned int *is_random)
464 struct fio_file *f = io_u->file;
465 enum fio_ddir ddir = io_u->ddir;
468 assert(ddir_rw(ddir));
470 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
472 td->ddir_seq_nr = td->o.ddir_seq_nr;
475 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
478 if (io_u->offset >= f->io_size) {
479 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
480 (unsigned long long) io_u->offset,
481 (unsigned long long) f->io_size);
485 io_u->offset += f->file_offset;
486 if (io_u->offset >= f->real_file_size) {
487 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
488 (unsigned long long) io_u->offset,
489 (unsigned long long) f->real_file_size);
496 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
497 unsigned int *is_random)
499 if (td->flags & TD_F_PROFILE_OPS) {
500 struct prof_io_ops *ops = &td->prof_io_ops;
502 if (ops->fill_io_u_off)
503 return ops->fill_io_u_off(td, io_u, is_random);
506 return __get_next_offset(td, io_u, is_random);
509 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
512 struct fio_file *f = io_u->file;
514 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
517 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
518 unsigned int is_random)
520 int ddir = io_u->ddir;
521 unsigned int buflen = 0;
522 unsigned int minbs, maxbs;
526 assert(ddir_rw(ddir));
528 if (td->o.bs_is_seq_rand)
529 ddir = is_random ? DDIR_WRITE: DDIR_READ;
531 minbs = td->o.min_bs[ddir];
532 maxbs = td->o.max_bs[ddir];
538 * If we can't satisfy the min block size from here, then fail
540 if (!io_u_fits(td, io_u, minbs))
543 frand_max = rand_max(&td->bsrange_state);
545 r = __rand(&td->bsrange_state);
547 if (!td->o.bssplit_nr[ddir]) {
548 buflen = 1 + (unsigned int) ((double) maxbs *
549 (r / (frand_max + 1.0)));
556 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
557 struct bssplit *bsp = &td->o.bssplit[ddir][i];
561 if ((r * 100UL <= frand_max * perc) &&
562 io_u_fits(td, io_u, buflen))
567 if (td->o.verify != VERIFY_NONE)
568 buflen = (buflen + td->o.verify_interval - 1) &
569 ~(td->o.verify_interval - 1);
571 if (!td->o.bs_unaligned && is_power_of_2(minbs))
572 buflen &= ~(minbs - 1);
574 } while (!io_u_fits(td, io_u, buflen));
579 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
580 unsigned int is_random)
582 if (td->flags & TD_F_PROFILE_OPS) {
583 struct prof_io_ops *ops = &td->prof_io_ops;
585 if (ops->fill_io_u_size)
586 return ops->fill_io_u_size(td, io_u, is_random);
589 return __get_next_buflen(td, io_u, is_random);
592 static void set_rwmix_bytes(struct thread_data *td)
597 * we do time or byte based switch. this is needed because
598 * buffered writes may issue a lot quicker than they complete,
599 * whereas reads do not.
601 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
602 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
605 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
609 v = rand32_between(&td->rwmix_state, 1, 100);
611 if (v <= td->o.rwmix[DDIR_READ])
617 int io_u_quiesce(struct thread_data *td)
622 * We are going to sleep, ensure that we flush anything pending as
623 * not to skew our latency numbers.
625 * Changed to only monitor 'in flight' requests here instead of the
626 * td->cur_depth, b/c td->cur_depth does not accurately represent
627 * io's that have been actually submitted to an async engine,
628 * and cur_depth is meaningless for sync engines.
630 if (td->io_u_queued || td->cur_depth) {
633 ret = td_io_commit(td);
636 while (td->io_u_in_flight) {
639 ret = io_u_queued_complete(td, 1);
647 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
649 enum fio_ddir odir = ddir ^ 1;
652 assert(ddir_rw(ddir));
653 now = utime_since_now(&td->start);
656 * if rate_next_io_time is in the past, need to catch up to rate
658 if (td->rate_next_io_time[ddir] <= now)
662 * We are ahead of rate in this direction. See if we
665 if (td_rw(td) && td->o.rwmix[odir]) {
667 * Other direction is behind rate, switch
669 if (td->rate_next_io_time[odir] <= now)
673 * Both directions are ahead of rate. sleep the min
674 * switch if necissary
676 if (td->rate_next_io_time[ddir] <=
677 td->rate_next_io_time[odir]) {
678 usec = td->rate_next_io_time[ddir] - now;
680 usec = td->rate_next_io_time[odir] - now;
684 usec = td->rate_next_io_time[ddir] - now;
686 if (td->o.io_submit_mode == IO_MODE_INLINE)
689 usec = usec_sleep(td, usec);
695 * Return the data direction for the next io_u. If the job is a
696 * mixed read/write workload, check the rwmix cycle and switch if
699 static enum fio_ddir get_rw_ddir(struct thread_data *td)
704 * see if it's time to fsync
706 if (td->o.fsync_blocks &&
707 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
708 td->io_issues[DDIR_WRITE] && should_fsync(td))
712 * see if it's time to fdatasync
714 if (td->o.fdatasync_blocks &&
715 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
716 td->io_issues[DDIR_WRITE] && should_fsync(td))
717 return DDIR_DATASYNC;
720 * see if it's time to sync_file_range
722 if (td->sync_file_range_nr &&
723 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
724 td->io_issues[DDIR_WRITE] && should_fsync(td))
725 return DDIR_SYNC_FILE_RANGE;
729 * Check if it's time to seed a new data direction.
731 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
733 * Put a top limit on how many bytes we do for
734 * one data direction, to avoid overflowing the
737 ddir = get_rand_ddir(td);
739 if (ddir != td->rwmix_ddir)
742 td->rwmix_ddir = ddir;
744 ddir = td->rwmix_ddir;
745 } else if (td_read(td))
747 else if (td_write(td))
752 td->rwmix_ddir = rate_ddir(td, ddir);
753 return td->rwmix_ddir;
756 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
758 enum fio_ddir ddir = get_rw_ddir(td);
760 if (td_trimwrite(td)) {
761 struct fio_file *f = io_u->file;
762 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
768 io_u->ddir = io_u->acct_ddir = ddir;
770 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
771 td->o.barrier_blocks &&
772 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
773 td->io_issues[DDIR_WRITE])
774 io_u_set(io_u, IO_U_F_BARRIER);
777 void put_file_log(struct thread_data *td, struct fio_file *f)
779 unsigned int ret = put_file(td, f);
782 td_verror(td, ret, "file close");
785 void put_io_u(struct thread_data *td, struct io_u *io_u)
792 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
793 put_file_log(td, io_u->file);
796 io_u_set(io_u, IO_U_F_FREE);
798 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
800 assert(!(td->flags & TD_F_CHILD));
802 io_u_qpush(&td->io_u_freelist, io_u);
804 td_io_u_free_notify(td);
807 void clear_io_u(struct thread_data *td, struct io_u *io_u)
809 io_u_clear(io_u, IO_U_F_FLIGHT);
813 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
815 struct io_u *__io_u = *io_u;
816 enum fio_ddir ddir = acct_ddir(__io_u);
818 dprint(FD_IO, "requeue %p\n", __io_u);
825 io_u_set(__io_u, IO_U_F_FREE);
826 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
827 td->io_issues[ddir]--;
829 io_u_clear(__io_u, IO_U_F_FLIGHT);
830 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
832 assert(!(td->flags & TD_F_CHILD));
835 io_u_rpush(&td->io_u_requeues, __io_u);
837 td_io_u_free_notify(td);
841 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
843 unsigned int is_random;
845 if (td->io_ops->flags & FIO_NOIO)
848 set_rw_ddir(td, io_u);
851 * fsync() or fdatasync() or trim etc, we are done
853 if (!ddir_rw(io_u->ddir))
857 * See if it's time to switch to a new zone
859 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
860 struct fio_file *f = io_u->file;
863 f->file_offset += td->o.zone_range + td->o.zone_skip;
866 * Wrap from the beginning, if we exceed the file size
868 if (f->file_offset >= f->real_file_size)
869 f->file_offset = f->real_file_size - f->file_offset;
870 f->last_pos[io_u->ddir] = f->file_offset;
871 td->io_skip_bytes += td->o.zone_skip;
875 * No log, let the seq/rand engine retrieve the next buflen and
878 if (get_next_offset(td, io_u, &is_random)) {
879 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
883 io_u->buflen = get_next_buflen(td, io_u, is_random);
885 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
889 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
890 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
891 dprint(FD_IO, " off=%llu/%lu > %llu\n",
892 (unsigned long long) io_u->offset, io_u->buflen,
893 (unsigned long long) io_u->file->real_file_size);
898 * mark entry before potentially trimming io_u
900 if (td_random(td) && file_randommap(td, io_u->file))
901 mark_random_map(td, io_u);
904 dprint_io_u(io_u, "fill_io_u");
905 td->zone_bytes += io_u->buflen;
909 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
938 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
940 __io_u_mark_map(td->ts.io_u_submit, nr);
941 td->ts.total_submit++;
944 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
946 __io_u_mark_map(td->ts.io_u_complete, nr);
947 td->ts.total_complete++;
950 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
954 switch (td->cur_depth) {
976 td->ts.io_u_map[idx] += nr;
979 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1016 assert(idx < FIO_IO_U_LAT_U_NR);
1017 td->ts.io_u_lat_u[idx]++;
1020 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1061 assert(idx < FIO_IO_U_LAT_M_NR);
1062 td->ts.io_u_lat_m[idx]++;
1065 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1068 io_u_mark_lat_usec(td, usec);
1070 io_u_mark_lat_msec(td, usec / 1000);
1074 * Get next file to service by choosing one at random
1076 static struct fio_file *get_next_file_rand(struct thread_data *td,
1077 enum fio_file_flags goodf,
1078 enum fio_file_flags badf)
1080 uint64_t frand_max = rand_max(&td->next_file_state);
1088 r = __rand(&td->next_file_state);
1089 fno = (unsigned int) ((double) td->o.nr_files
1090 * (r / (frand_max + 1.0)));
1093 if (fio_file_done(f))
1096 if (!fio_file_open(f)) {
1099 if (td->nr_open_files >= td->o.open_files)
1100 return ERR_PTR(-EBUSY);
1102 err = td_io_open_file(td, f);
1108 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1109 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1113 td_io_close_file(td, f);
1118 * Get next file to service by doing round robin between all available ones
1120 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1123 unsigned int old_next_file = td->next_file;
1129 f = td->files[td->next_file];
1132 if (td->next_file >= td->o.nr_files)
1135 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1136 if (fio_file_done(f)) {
1141 if (!fio_file_open(f)) {
1144 if (td->nr_open_files >= td->o.open_files)
1145 return ERR_PTR(-EBUSY);
1147 err = td_io_open_file(td, f);
1149 dprint(FD_FILE, "error %d on open of %s\n",
1157 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1159 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1163 td_io_close_file(td, f);
1166 } while (td->next_file != old_next_file);
1168 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1172 static struct fio_file *__get_next_file(struct thread_data *td)
1176 assert(td->o.nr_files <= td->files_index);
1178 if (td->nr_done_files >= td->o.nr_files) {
1179 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1180 " nr_files=%d\n", td->nr_open_files,
1186 f = td->file_service_file;
1187 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1188 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1190 if (td->file_service_left--)
1194 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1195 td->o.file_service_type == FIO_FSERVICE_SEQ)
1196 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1198 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1203 td->file_service_file = f;
1204 td->file_service_left = td->file_service_nr - 1;
1207 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1209 dprint(FD_FILE, "get_next_file: NULL\n");
1213 static struct fio_file *get_next_file(struct thread_data *td)
1215 if (td->flags & TD_F_PROFILE_OPS) {
1216 struct prof_io_ops *ops = &td->prof_io_ops;
1218 if (ops->get_next_file)
1219 return ops->get_next_file(td);
1222 return __get_next_file(td);
1225 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1230 f = get_next_file(td);
1231 if (IS_ERR_OR_NULL(f))
1237 if (!fill_io_u(td, io_u))
1240 put_file_log(td, f);
1241 td_io_close_file(td, f);
1243 fio_file_set_done(f);
1244 td->nr_done_files++;
1245 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1246 td->nr_done_files, td->o.nr_files);
1252 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1253 unsigned long tusec, unsigned long max_usec)
1256 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1257 td_verror(td, ETIMEDOUT, "max latency exceeded");
1258 icd->error = ETIMEDOUT;
1261 static void lat_new_cycle(struct thread_data *td)
1263 fio_gettime(&td->latency_ts, NULL);
1264 td->latency_ios = ddir_rw_sum(td->io_blocks);
1265 td->latency_failed = 0;
1269 * We had an IO outside the latency target. Reduce the queue depth. If we
1270 * are at QD=1, then it's time to give up.
1272 static bool __lat_target_failed(struct thread_data *td)
1274 if (td->latency_qd == 1)
1277 td->latency_qd_high = td->latency_qd;
1279 if (td->latency_qd == td->latency_qd_low)
1280 td->latency_qd_low--;
1282 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1284 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1287 * When we ramp QD down, quiesce existing IO to prevent
1288 * a storm of ramp downs due to pending higher depth.
1295 static bool lat_target_failed(struct thread_data *td)
1297 if (td->o.latency_percentile.u.f == 100.0)
1298 return __lat_target_failed(td);
1300 td->latency_failed++;
1304 void lat_target_init(struct thread_data *td)
1306 td->latency_end_run = 0;
1308 if (td->o.latency_target) {
1309 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1310 fio_gettime(&td->latency_ts, NULL);
1312 td->latency_qd_high = td->o.iodepth;
1313 td->latency_qd_low = 1;
1314 td->latency_ios = ddir_rw_sum(td->io_blocks);
1316 td->latency_qd = td->o.iodepth;
1319 void lat_target_reset(struct thread_data *td)
1321 if (!td->latency_end_run)
1322 lat_target_init(td);
1325 static void lat_target_success(struct thread_data *td)
1327 const unsigned int qd = td->latency_qd;
1328 struct thread_options *o = &td->o;
1330 td->latency_qd_low = td->latency_qd;
1333 * If we haven't failed yet, we double up to a failing value instead
1334 * of bisecting from highest possible queue depth. If we have set
1335 * a limit other than td->o.iodepth, bisect between that.
1337 if (td->latency_qd_high != o->iodepth)
1338 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1340 td->latency_qd *= 2;
1342 if (td->latency_qd > o->iodepth)
1343 td->latency_qd = o->iodepth;
1345 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1348 * Same as last one, we are done. Let it run a latency cycle, so
1349 * we get only the results from the targeted depth.
1351 if (td->latency_qd == qd) {
1352 if (td->latency_end_run) {
1353 dprint(FD_RATE, "We are done\n");
1356 dprint(FD_RATE, "Quiesce and final run\n");
1358 td->latency_end_run = 1;
1359 reset_all_stats(td);
1368 * Check if we can bump the queue depth
1370 void lat_target_check(struct thread_data *td)
1372 uint64_t usec_window;
1376 usec_window = utime_since_now(&td->latency_ts);
1377 if (usec_window < td->o.latency_window)
1380 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1381 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1382 success_ios *= 100.0;
1384 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1386 if (success_ios >= td->o.latency_percentile.u.f)
1387 lat_target_success(td);
1389 __lat_target_failed(td);
1393 * If latency target is enabled, we might be ramping up or down and not
1394 * using the full queue depth available.
1396 bool queue_full(const struct thread_data *td)
1398 const int qempty = io_u_qempty(&td->io_u_freelist);
1402 if (!td->o.latency_target)
1405 return td->cur_depth >= td->latency_qd;
1408 struct io_u *__get_io_u(struct thread_data *td)
1410 struct io_u *io_u = NULL;
1418 if (!io_u_rempty(&td->io_u_requeues))
1419 io_u = io_u_rpop(&td->io_u_requeues);
1420 else if (!queue_full(td)) {
1421 io_u = io_u_qpop(&td->io_u_freelist);
1426 io_u->end_io = NULL;
1430 assert(io_u->flags & IO_U_F_FREE);
1431 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1432 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1436 io_u->acct_ddir = -1;
1438 assert(!(td->flags & TD_F_CHILD));
1439 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1441 } else if (td_async_processing(td)) {
1443 * We ran out, wait for async verify threads to finish and
1446 assert(!(td->flags & TD_F_CHILD));
1447 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1455 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1457 if (!(td->flags & TD_F_TRIM_BACKLOG))
1460 if (td->trim_entries) {
1463 if (td->trim_batch) {
1466 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1467 td->last_ddir != DDIR_READ) {
1468 td->trim_batch = td->o.trim_batch;
1469 if (!td->trim_batch)
1470 td->trim_batch = td->o.trim_backlog;
1474 if (get_trim && !get_next_trim(td, io_u))
1481 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1483 if (!(td->flags & TD_F_VER_BACKLOG))
1486 if (td->io_hist_len) {
1489 if (td->verify_batch)
1491 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1492 td->last_ddir != DDIR_READ) {
1493 td->verify_batch = td->o.verify_batch;
1494 if (!td->verify_batch)
1495 td->verify_batch = td->o.verify_backlog;
1499 if (get_verify && !get_next_verify(td, io_u)) {
1509 * Fill offset and start time into the buffer content, to prevent too
1510 * easy compressible data for simple de-dupe attempts. Do this for every
1511 * 512b block in the range, since that should be the smallest block size
1512 * we can expect from a device.
1514 static void small_content_scramble(struct io_u *io_u)
1516 unsigned int i, nr_blocks = io_u->buflen / 512;
1518 unsigned int offset;
1525 boffset = io_u->offset;
1526 io_u->buf_filled_len = 0;
1528 for (i = 0; i < nr_blocks; i++) {
1530 * Fill the byte offset into a "random" start offset of
1531 * the buffer, given by the product of the usec time
1532 * and the actual offset.
1534 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1535 offset &= ~(sizeof(uint64_t) - 1);
1536 if (offset >= 512 - sizeof(uint64_t))
1537 offset -= sizeof(uint64_t);
1538 memcpy(p + offset, &boffset, sizeof(boffset));
1540 end = p + 512 - sizeof(io_u->start_time);
1541 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1548 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1549 * etc. The returned io_u is fully ready to be prepped and submitted.
1551 struct io_u *get_io_u(struct thread_data *td)
1555 int do_scramble = 0;
1558 io_u = __get_io_u(td);
1560 dprint(FD_IO, "__get_io_u failed\n");
1564 if (check_get_verify(td, io_u))
1566 if (check_get_trim(td, io_u))
1570 * from a requeue, io_u already setup
1576 * If using an iolog, grab next piece if any available.
1578 if (td->flags & TD_F_READ_IOLOG) {
1579 if (read_iolog_get(td, io_u))
1581 } else if (set_io_u_file(td, io_u)) {
1583 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1589 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1593 assert(fio_file_open(f));
1595 if (ddir_rw(io_u->ddir)) {
1596 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1597 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1601 f->last_start[io_u->ddir] = io_u->offset;
1602 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1604 if (io_u->ddir == DDIR_WRITE) {
1605 if (td->flags & TD_F_REFILL_BUFFERS) {
1606 io_u_fill_buffer(td, io_u,
1607 td->o.min_bs[DDIR_WRITE],
1609 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1610 !(td->flags & TD_F_COMPRESS))
1612 if (td->flags & TD_F_VER_NONE) {
1613 populate_verify_io_u(td, io_u);
1616 } else if (io_u->ddir == DDIR_READ) {
1618 * Reset the buf_filled parameters so next time if the
1619 * buffer is used for writes it is refilled.
1621 io_u->buf_filled_len = 0;
1626 * Set io data pointers.
1628 io_u->xfer_buf = io_u->buf;
1629 io_u->xfer_buflen = io_u->buflen;
1633 if (!td_io_prep(td, io_u)) {
1634 if (!td->o.disable_lat)
1635 fio_gettime(&io_u->start_time, NULL);
1637 small_content_scramble(io_u);
1641 dprint(FD_IO, "get_io_u failed\n");
1643 return ERR_PTR(ret);
1646 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1648 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1650 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1653 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1654 io_u->file ? " on file " : "",
1655 io_u->file ? io_u->file->file_name : "",
1656 strerror(io_u->error),
1657 io_ddir_name(io_u->ddir),
1658 io_u->offset, io_u->xfer_buflen);
1660 if (td->io_ops->errdetails) {
1661 char *err = td->io_ops->errdetails(io_u);
1663 log_err("fio: %s\n", err);
1668 td_verror(td, io_u->error, "io_u error");
1671 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1673 __io_u_log_error(td, io_u);
1675 __io_u_log_error(td->parent, io_u);
1678 static inline bool gtod_reduce(struct thread_data *td)
1680 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1681 || td->o.gtod_reduce;
1684 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1685 struct io_completion_data *icd,
1686 const enum fio_ddir idx, unsigned int bytes)
1688 const int no_reduce = !gtod_reduce(td);
1689 unsigned long lusec = 0;
1695 lusec = utime_since(&io_u->issue_time, &icd->time);
1697 if (!td->o.disable_lat) {
1698 unsigned long tusec;
1700 tusec = utime_since(&io_u->start_time, &icd->time);
1701 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1703 if (td->flags & TD_F_PROFILE_OPS) {
1704 struct prof_io_ops *ops = &td->prof_io_ops;
1707 icd->error = ops->io_u_lat(td, tusec);
1710 if (td->o.max_latency && tusec > td->o.max_latency)
1711 lat_fatal(td, icd, tusec, td->o.max_latency);
1712 if (td->o.latency_target && tusec > td->o.latency_target) {
1713 if (lat_target_failed(td))
1714 lat_fatal(td, icd, tusec, td->o.latency_target);
1719 if (!td->o.disable_clat) {
1720 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1721 io_u_mark_latency(td, lusec);
1724 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1725 add_bw_sample(td, io_u, bytes, lusec);
1727 if (no_reduce && per_unit_log(td->iops_log))
1728 add_iops_sample(td, io_u, bytes);
1731 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1732 uint32_t *info = io_u_block_info(td, io_u);
1733 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1734 if (io_u->ddir == DDIR_TRIM) {
1735 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1736 BLOCK_INFO_TRIMS(*info) + 1);
1737 } else if (io_u->ddir == DDIR_WRITE) {
1738 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1745 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1746 uint64_t offset, unsigned int bytes)
1753 if (f->first_write == -1ULL || offset < f->first_write)
1754 f->first_write = offset;
1755 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1756 f->last_write = offset + bytes;
1758 if (!f->last_write_comp)
1761 idx = f->last_write_idx++;
1762 f->last_write_comp[idx] = offset;
1763 if (f->last_write_idx == td->o.iodepth)
1764 f->last_write_idx = 0;
1767 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1768 struct io_completion_data *icd)
1770 struct io_u *io_u = *io_u_ptr;
1771 enum fio_ddir ddir = io_u->ddir;
1772 struct fio_file *f = io_u->file;
1774 dprint_io_u(io_u, "io complete");
1776 assert(io_u->flags & IO_U_F_FLIGHT);
1777 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1780 * Mark IO ok to verify
1784 * Remove errored entry from the verification list
1787 unlog_io_piece(td, io_u);
1789 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1794 if (ddir_sync(ddir)) {
1795 td->last_was_sync = 1;
1797 f->first_write = -1ULL;
1798 f->last_write = -1ULL;
1803 td->last_was_sync = 0;
1804 td->last_ddir = ddir;
1806 if (!io_u->error && ddir_rw(ddir)) {
1807 unsigned int bytes = io_u->buflen - io_u->resid;
1810 td->io_blocks[ddir]++;
1811 td->this_io_blocks[ddir]++;
1812 td->io_bytes[ddir] += bytes;
1814 if (!(io_u->flags & IO_U_F_VER_LIST))
1815 td->this_io_bytes[ddir] += bytes;
1817 if (ddir == DDIR_WRITE)
1818 file_log_write_comp(td, f, io_u->offset, bytes);
1820 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1821 td->runstate == TD_VERIFYING))
1822 account_io_completion(td, io_u, icd, ddir, bytes);
1824 icd->bytes_done[ddir] += bytes;
1827 ret = io_u->end_io(td, io_u_ptr);
1829 if (ret && !icd->error)
1832 } else if (io_u->error) {
1833 icd->error = io_u->error;
1834 io_u_log_error(td, io_u);
1837 enum error_type_bit eb = td_error_type(ddir, icd->error);
1839 if (!td_non_fatal_error(td, eb, icd->error))
1843 * If there is a non_fatal error, then add to the error count
1844 * and clear all the errors.
1846 update_error_count(td, icd->error);
1854 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1859 if (!gtod_reduce(td))
1860 fio_gettime(&icd->time, NULL);
1865 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1866 icd->bytes_done[ddir] = 0;
1869 static void ios_completed(struct thread_data *td,
1870 struct io_completion_data *icd)
1875 for (i = 0; i < icd->nr; i++) {
1876 io_u = td->io_ops->event(td, i);
1878 io_completed(td, &io_u, icd);
1886 * Complete a single io_u for the sync engines.
1888 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1890 struct io_completion_data icd;
1893 init_icd(td, &icd, 1);
1894 io_completed(td, &io_u, &icd);
1900 td_verror(td, icd.error, "io_u_sync_complete");
1904 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1905 td->bytes_done[ddir] += icd.bytes_done[ddir];
1911 * Called to complete min_events number of io for the async engines.
1913 int io_u_queued_complete(struct thread_data *td, int min_evts)
1915 struct io_completion_data icd;
1916 struct timespec *tvp = NULL;
1918 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1920 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1924 else if (min_evts > td->cur_depth)
1925 min_evts = td->cur_depth;
1927 /* No worries, td_io_getevents fixes min and max if they are
1928 * set incorrectly */
1929 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1931 td_verror(td, -ret, "td_io_getevents");
1936 init_icd(td, &icd, ret);
1937 ios_completed(td, &icd);
1939 td_verror(td, icd.error, "io_u_queued_complete");
1943 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1944 td->bytes_done[ddir] += icd.bytes_done[ddir];
1950 * Call when io_u is really queued, to update the submission latency.
1952 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1954 if (!td->o.disable_slat) {
1955 unsigned long slat_time;
1957 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1962 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1968 * See if we should reuse the last seed, if dedupe is enabled
1970 static struct frand_state *get_buf_state(struct thread_data *td)
1974 if (!td->o.dedupe_percentage)
1975 return &td->buf_state;
1976 else if (td->o.dedupe_percentage == 100) {
1977 frand_copy(&td->buf_state_prev, &td->buf_state);
1978 return &td->buf_state;
1981 v = rand32_between(&td->dedupe_state, 1, 100);
1983 if (v <= td->o.dedupe_percentage)
1984 return &td->buf_state_prev;
1986 return &td->buf_state;
1989 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1991 if (td->o.dedupe_percentage == 100)
1992 frand_copy(rs, &td->buf_state_prev);
1993 else if (rs == &td->buf_state)
1994 frand_copy(&td->buf_state_prev, rs);
1997 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1998 unsigned int max_bs)
2000 struct thread_options *o = &td->o;
2002 if (o->compress_percentage || o->dedupe_percentage) {
2003 unsigned int perc = td->o.compress_percentage;
2004 struct frand_state *rs;
2005 unsigned int left = max_bs;
2006 unsigned int this_write;
2009 rs = get_buf_state(td);
2011 min_write = min(min_write, left);
2014 this_write = min_not_zero(min_write,
2015 td->o.compress_chunk);
2017 fill_random_buf_percentage(rs, buf, perc,
2018 this_write, this_write,
2020 o->buffer_pattern_bytes);
2022 fill_random_buf(rs, buf, min_write);
2023 this_write = min_write;
2028 save_buf_state(td, rs);
2030 } else if (o->buffer_pattern_bytes)
2031 fill_buffer_pattern(td, buf, max_bs);
2032 else if (o->zero_buffers)
2033 memset(buf, 0, max_bs);
2035 fill_random_buf(get_buf_state(td), buf, max_bs);
2039 * "randomly" fill the buffer contents
2041 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2042 unsigned int min_write, unsigned int max_bs)
2044 io_u->buf_filled_len = 0;
2045 fill_io_buffer(td, io_u->buf, min_write, max_bs);