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 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
333 fio_file_reset(td, f);
334 if (!get_next_rand_offset(td, f, ddir, b))
338 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
339 f->file_name, (unsigned long long) f->last_pos[ddir],
340 (unsigned long long) f->real_file_size);
344 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
345 enum fio_ddir ddir, uint64_t *offset)
347 struct thread_options *o = &td->o;
349 assert(ddir_rw(ddir));
351 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
353 struct thread_options *o = &td->o;
354 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
356 if (io_size > f->last_pos[ddir])
357 f->last_pos[ddir] = 0;
359 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
362 if (f->last_pos[ddir] < f->real_file_size) {
365 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
366 if (f->real_file_size > f->io_size)
367 f->last_pos[ddir] = f->io_size;
369 f->last_pos[ddir] = f->real_file_size;
372 pos = f->last_pos[ddir] - f->file_offset;
373 if (pos && o->ddir_seq_add) {
374 pos += o->ddir_seq_add;
377 * If we reach beyond the end of the file
378 * with holed IO, wrap around to the
379 * beginning again. If we're doing backwards IO,
382 if (pos >= f->real_file_size) {
383 if (o->ddir_seq_add > 0)
384 pos = f->file_offset;
386 if (f->real_file_size > f->io_size)
389 pos = f->real_file_size;
391 pos += o->ddir_seq_add;
403 static int get_next_block(struct thread_data *td, struct io_u *io_u,
404 enum fio_ddir ddir, int rw_seq,
405 unsigned int *is_random)
407 struct fio_file *f = io_u->file;
411 assert(ddir_rw(ddir));
417 if (should_do_random(td, ddir)) {
418 ret = get_next_rand_block(td, f, ddir, &b);
422 io_u_set(td, io_u, IO_U_F_BUSY_OK);
423 ret = get_next_seq_offset(td, f, ddir, &offset);
425 ret = get_next_rand_block(td, f, ddir, &b);
429 ret = get_next_seq_offset(td, f, ddir, &offset);
432 io_u_set(td, io_u, IO_U_F_BUSY_OK);
435 if (td->o.rw_seq == RW_SEQ_SEQ) {
436 ret = get_next_seq_offset(td, f, ddir, &offset);
438 ret = get_next_rand_block(td, f, ddir, &b);
441 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
442 if (f->last_start[ddir] != -1ULL)
443 offset = f->last_start[ddir] - f->file_offset;
448 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
455 io_u->offset = offset;
457 io_u->offset = b * td->o.ba[ddir];
459 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
468 * For random io, generate a random new block and see if it's used. Repeat
469 * until we find a free one. For sequential io, just return the end of
470 * the last io issued.
472 static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
473 unsigned int *is_random)
475 struct fio_file *f = io_u->file;
476 enum fio_ddir ddir = io_u->ddir;
479 assert(ddir_rw(ddir));
481 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
483 td->ddir_seq_nr = td->o.ddir_seq_nr;
486 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
489 if (io_u->offset >= f->io_size) {
490 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
491 (unsigned long long) io_u->offset,
492 (unsigned long long) f->io_size);
496 io_u->offset += f->file_offset;
497 if (io_u->offset >= f->real_file_size) {
498 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
499 (unsigned long long) io_u->offset,
500 (unsigned long long) f->real_file_size);
507 static int get_next_offset(struct thread_data *td, struct io_u *io_u,
508 unsigned int *is_random)
510 if (td->flags & TD_F_PROFILE_OPS) {
511 struct prof_io_ops *ops = &td->prof_io_ops;
513 if (ops->fill_io_u_off)
514 return ops->fill_io_u_off(td, io_u, is_random);
517 return __get_next_offset(td, io_u, is_random);
520 static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
523 struct fio_file *f = io_u->file;
525 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
528 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
529 unsigned int is_random)
531 int ddir = io_u->ddir;
532 unsigned int buflen = 0;
533 unsigned int minbs, maxbs;
534 uint64_t frand_max, r;
536 assert(ddir_rw(ddir));
538 if (td->o.bs_is_seq_rand)
539 ddir = is_random ? DDIR_WRITE: DDIR_READ;
541 minbs = td->o.min_bs[ddir];
542 maxbs = td->o.max_bs[ddir];
548 * If we can't satisfy the min block size from here, then fail
550 if (!io_u_fits(td, io_u, minbs))
553 frand_max = rand_max(&td->bsrange_state);
555 r = __rand(&td->bsrange_state);
557 if (!td->o.bssplit_nr[ddir]) {
558 buflen = 1 + (unsigned int) ((double) maxbs *
559 (r / (frand_max + 1.0)));
566 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
567 struct bssplit *bsp = &td->o.bssplit[ddir][i];
573 if ((r / perc <= frand_max / 100ULL) &&
574 io_u_fits(td, io_u, buflen))
579 if (td->o.verify != VERIFY_NONE)
580 buflen = (buflen + td->o.verify_interval - 1) &
581 ~(td->o.verify_interval - 1);
583 if (!td->o.bs_unaligned && is_power_of_2(minbs))
584 buflen &= ~(minbs - 1);
586 } while (!io_u_fits(td, io_u, buflen));
591 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
592 unsigned int is_random)
594 if (td->flags & TD_F_PROFILE_OPS) {
595 struct prof_io_ops *ops = &td->prof_io_ops;
597 if (ops->fill_io_u_size)
598 return ops->fill_io_u_size(td, io_u, is_random);
601 return __get_next_buflen(td, io_u, is_random);
604 static void set_rwmix_bytes(struct thread_data *td)
609 * we do time or byte based switch. this is needed because
610 * buffered writes may issue a lot quicker than they complete,
611 * whereas reads do not.
613 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
614 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
617 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
621 v = rand32_between(&td->rwmix_state, 1, 100);
623 if (v <= td->o.rwmix[DDIR_READ])
629 int io_u_quiesce(struct thread_data *td)
634 * We are going to sleep, ensure that we flush anything pending as
635 * not to skew our latency numbers.
637 * Changed to only monitor 'in flight' requests here instead of the
638 * td->cur_depth, b/c td->cur_depth does not accurately represent
639 * io's that have been actually submitted to an async engine,
640 * and cur_depth is meaningless for sync engines.
642 if (td->io_u_queued || td->cur_depth) {
645 ret = td_io_commit(td);
648 while (td->io_u_in_flight) {
651 ret = io_u_queued_complete(td, 1);
656 if (td->flags & TD_F_REGROW_LOGS)
662 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
664 enum fio_ddir odir = ddir ^ 1;
668 assert(ddir_rw(ddir));
669 now = utime_since_now(&td->start);
672 * if rate_next_io_time is in the past, need to catch up to rate
674 if (td->rate_next_io_time[ddir] <= now)
678 * We are ahead of rate in this direction. See if we
681 if (td_rw(td) && td->o.rwmix[odir]) {
683 * Other direction is behind rate, switch
685 if (td->rate_next_io_time[odir] <= now)
689 * Both directions are ahead of rate. sleep the min
690 * switch if necissary
692 if (td->rate_next_io_time[ddir] <=
693 td->rate_next_io_time[odir]) {
694 usec = td->rate_next_io_time[ddir] - now;
696 usec = td->rate_next_io_time[odir] - now;
700 usec = td->rate_next_io_time[ddir] - now;
702 if (td->o.io_submit_mode == IO_MODE_INLINE)
705 usec = usec_sleep(td, usec);
711 * Return the data direction for the next io_u. If the job is a
712 * mixed read/write workload, check the rwmix cycle and switch if
715 static enum fio_ddir get_rw_ddir(struct thread_data *td)
720 * see if it's time to fsync
722 if (td->o.fsync_blocks &&
723 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
724 td->io_issues[DDIR_WRITE] && should_fsync(td))
728 * see if it's time to fdatasync
730 if (td->o.fdatasync_blocks &&
731 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
732 td->io_issues[DDIR_WRITE] && should_fsync(td))
733 return DDIR_DATASYNC;
736 * see if it's time to sync_file_range
738 if (td->sync_file_range_nr &&
739 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
740 td->io_issues[DDIR_WRITE] && should_fsync(td))
741 return DDIR_SYNC_FILE_RANGE;
745 * Check if it's time to seed a new data direction.
747 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
749 * Put a top limit on how many bytes we do for
750 * one data direction, to avoid overflowing the
753 ddir = get_rand_ddir(td);
755 if (ddir != td->rwmix_ddir)
758 td->rwmix_ddir = ddir;
760 ddir = td->rwmix_ddir;
761 } else if (td_read(td))
763 else if (td_write(td))
768 td->rwmix_ddir = rate_ddir(td, ddir);
769 return td->rwmix_ddir;
772 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
774 enum fio_ddir ddir = get_rw_ddir(td);
776 if (td_trimwrite(td)) {
777 struct fio_file *f = io_u->file;
778 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
784 io_u->ddir = io_u->acct_ddir = ddir;
786 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
787 td->o.barrier_blocks &&
788 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
789 td->io_issues[DDIR_WRITE])
790 io_u_set(td, io_u, IO_U_F_BARRIER);
793 void put_file_log(struct thread_data *td, struct fio_file *f)
795 unsigned int ret = put_file(td, f);
798 td_verror(td, ret, "file close");
801 void put_io_u(struct thread_data *td, struct io_u *io_u)
808 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
809 put_file_log(td, io_u->file);
812 io_u_set(td, io_u, IO_U_F_FREE);
814 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
816 assert(!(td->flags & TD_F_CHILD));
818 io_u_qpush(&td->io_u_freelist, io_u);
820 td_io_u_free_notify(td);
823 void clear_io_u(struct thread_data *td, struct io_u *io_u)
825 io_u_clear(td, io_u, IO_U_F_FLIGHT);
829 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
831 struct io_u *__io_u = *io_u;
832 enum fio_ddir ddir = acct_ddir(__io_u);
834 dprint(FD_IO, "requeue %p\n", __io_u);
841 io_u_set(td, __io_u, IO_U_F_FREE);
842 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
843 td->io_issues[ddir]--;
845 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
846 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
848 assert(!(td->flags & TD_F_CHILD));
851 io_u_rpush(&td->io_u_requeues, __io_u);
853 td_io_u_free_notify(td);
857 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
859 unsigned int is_random;
861 if (td_ioengine_flagged(td, FIO_NOIO))
864 set_rw_ddir(td, io_u);
867 * fsync() or fdatasync() or trim etc, we are done
869 if (!ddir_rw(io_u->ddir))
873 * See if it's time to switch to a new zone
875 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
876 struct fio_file *f = io_u->file;
879 f->file_offset += td->o.zone_range + td->o.zone_skip;
882 * Wrap from the beginning, if we exceed the file size
884 if (f->file_offset >= f->real_file_size)
885 f->file_offset = f->real_file_size - f->file_offset;
886 f->last_pos[io_u->ddir] = f->file_offset;
887 td->io_skip_bytes += td->o.zone_skip;
891 * No log, let the seq/rand engine retrieve the next buflen and
894 if (get_next_offset(td, io_u, &is_random)) {
895 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
899 io_u->buflen = get_next_buflen(td, io_u, is_random);
901 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
905 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
906 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
907 dprint(FD_IO, " off=%llu/%lu > %llu\n",
908 (unsigned long long) io_u->offset, io_u->buflen,
909 (unsigned long long) io_u->file->real_file_size);
914 * mark entry before potentially trimming io_u
916 if (td_random(td) && file_randommap(td, io_u->file))
917 mark_random_map(td, io_u);
920 dprint_io_u(io_u, "fill_io_u");
921 td->zone_bytes += io_u->buflen;
925 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
954 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
956 __io_u_mark_map(td->ts.io_u_submit, nr);
957 td->ts.total_submit++;
960 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
962 __io_u_mark_map(td->ts.io_u_complete, nr);
963 td->ts.total_complete++;
966 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
970 switch (td->cur_depth) {
992 td->ts.io_u_map[idx] += nr;
995 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
1032 assert(idx < FIO_IO_U_LAT_U_NR);
1033 td->ts.io_u_lat_u[idx]++;
1036 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
1077 assert(idx < FIO_IO_U_LAT_M_NR);
1078 td->ts.io_u_lat_m[idx]++;
1081 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1084 io_u_mark_lat_usec(td, usec);
1086 io_u_mark_lat_msec(td, usec / 1000);
1089 static unsigned int __get_next_fileno_rand(struct thread_data *td)
1091 unsigned long fileno;
1093 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1094 uint64_t frand_max = rand_max(&td->next_file_state);
1097 r = __rand(&td->next_file_state);
1098 return (unsigned int) ((double) td->o.nr_files
1099 * (r / (frand_max + 1.0)));
1102 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1103 fileno = zipf_next(&td->next_file_zipf);
1104 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1105 fileno = pareto_next(&td->next_file_zipf);
1106 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1107 fileno = gauss_next(&td->next_file_gauss);
1109 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1114 return fileno >> FIO_FSERVICE_SHIFT;
1118 * Get next file to service by choosing one at random
1120 static struct fio_file *get_next_file_rand(struct thread_data *td,
1121 enum fio_file_flags goodf,
1122 enum fio_file_flags badf)
1130 fno = __get_next_fileno_rand(td);
1133 if (fio_file_done(f))
1136 if (!fio_file_open(f)) {
1139 if (td->nr_open_files >= td->o.open_files)
1140 return ERR_PTR(-EBUSY);
1142 err = td_io_open_file(td, f);
1148 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1149 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1153 td_io_close_file(td, f);
1158 * Get next file to service by doing round robin between all available ones
1160 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1163 unsigned int old_next_file = td->next_file;
1169 f = td->files[td->next_file];
1172 if (td->next_file >= td->o.nr_files)
1175 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1176 if (fio_file_done(f)) {
1181 if (!fio_file_open(f)) {
1184 if (td->nr_open_files >= td->o.open_files)
1185 return ERR_PTR(-EBUSY);
1187 err = td_io_open_file(td, f);
1189 dprint(FD_FILE, "error %d on open of %s\n",
1197 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1199 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1203 td_io_close_file(td, f);
1206 } while (td->next_file != old_next_file);
1208 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1212 static struct fio_file *__get_next_file(struct thread_data *td)
1216 assert(td->o.nr_files <= td->files_index);
1218 if (td->nr_done_files >= td->o.nr_files) {
1219 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1220 " nr_files=%d\n", td->nr_open_files,
1226 f = td->file_service_file;
1227 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1228 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1230 if (td->file_service_left--)
1234 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1235 td->o.file_service_type == FIO_FSERVICE_SEQ)
1236 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1238 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1243 td->file_service_file = f;
1244 td->file_service_left = td->file_service_nr - 1;
1247 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1249 dprint(FD_FILE, "get_next_file: NULL\n");
1253 static struct fio_file *get_next_file(struct thread_data *td)
1255 if (td->flags & TD_F_PROFILE_OPS) {
1256 struct prof_io_ops *ops = &td->prof_io_ops;
1258 if (ops->get_next_file)
1259 return ops->get_next_file(td);
1262 return __get_next_file(td);
1265 static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1270 f = get_next_file(td);
1271 if (IS_ERR_OR_NULL(f))
1277 if (!fill_io_u(td, io_u))
1280 put_file_log(td, f);
1281 td_io_close_file(td, f);
1283 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1284 fio_file_reset(td, f);
1286 fio_file_set_done(f);
1287 td->nr_done_files++;
1288 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1289 td->nr_done_files, td->o.nr_files);
1296 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1297 unsigned long tusec, unsigned long max_usec)
1300 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1301 td_verror(td, ETIMEDOUT, "max latency exceeded");
1302 icd->error = ETIMEDOUT;
1305 static void lat_new_cycle(struct thread_data *td)
1307 fio_gettime(&td->latency_ts, NULL);
1308 td->latency_ios = ddir_rw_sum(td->io_blocks);
1309 td->latency_failed = 0;
1313 * We had an IO outside the latency target. Reduce the queue depth. If we
1314 * are at QD=1, then it's time to give up.
1316 static bool __lat_target_failed(struct thread_data *td)
1318 if (td->latency_qd == 1)
1321 td->latency_qd_high = td->latency_qd;
1323 if (td->latency_qd == td->latency_qd_low)
1324 td->latency_qd_low--;
1326 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1328 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1331 * When we ramp QD down, quiesce existing IO to prevent
1332 * a storm of ramp downs due to pending higher depth.
1339 static bool lat_target_failed(struct thread_data *td)
1341 if (td->o.latency_percentile.u.f == 100.0)
1342 return __lat_target_failed(td);
1344 td->latency_failed++;
1348 void lat_target_init(struct thread_data *td)
1350 td->latency_end_run = 0;
1352 if (td->o.latency_target) {
1353 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1354 fio_gettime(&td->latency_ts, NULL);
1356 td->latency_qd_high = td->o.iodepth;
1357 td->latency_qd_low = 1;
1358 td->latency_ios = ddir_rw_sum(td->io_blocks);
1360 td->latency_qd = td->o.iodepth;
1363 void lat_target_reset(struct thread_data *td)
1365 if (!td->latency_end_run)
1366 lat_target_init(td);
1369 static void lat_target_success(struct thread_data *td)
1371 const unsigned int qd = td->latency_qd;
1372 struct thread_options *o = &td->o;
1374 td->latency_qd_low = td->latency_qd;
1377 * If we haven't failed yet, we double up to a failing value instead
1378 * of bisecting from highest possible queue depth. If we have set
1379 * a limit other than td->o.iodepth, bisect between that.
1381 if (td->latency_qd_high != o->iodepth)
1382 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1384 td->latency_qd *= 2;
1386 if (td->latency_qd > o->iodepth)
1387 td->latency_qd = o->iodepth;
1389 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1392 * Same as last one, we are done. Let it run a latency cycle, so
1393 * we get only the results from the targeted depth.
1395 if (td->latency_qd == qd) {
1396 if (td->latency_end_run) {
1397 dprint(FD_RATE, "We are done\n");
1400 dprint(FD_RATE, "Quiesce and final run\n");
1402 td->latency_end_run = 1;
1403 reset_all_stats(td);
1412 * Check if we can bump the queue depth
1414 void lat_target_check(struct thread_data *td)
1416 uint64_t usec_window;
1420 usec_window = utime_since_now(&td->latency_ts);
1421 if (usec_window < td->o.latency_window)
1424 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1425 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1426 success_ios *= 100.0;
1428 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1430 if (success_ios >= td->o.latency_percentile.u.f)
1431 lat_target_success(td);
1433 __lat_target_failed(td);
1437 * If latency target is enabled, we might be ramping up or down and not
1438 * using the full queue depth available.
1440 bool queue_full(const struct thread_data *td)
1442 const int qempty = io_u_qempty(&td->io_u_freelist);
1446 if (!td->o.latency_target)
1449 return td->cur_depth >= td->latency_qd;
1452 struct io_u *__get_io_u(struct thread_data *td)
1454 struct io_u *io_u = NULL;
1462 if (!io_u_rempty(&td->io_u_requeues))
1463 io_u = io_u_rpop(&td->io_u_requeues);
1464 else if (!queue_full(td)) {
1465 io_u = io_u_qpop(&td->io_u_freelist);
1470 io_u->end_io = NULL;
1474 assert(io_u->flags & IO_U_F_FREE);
1475 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1476 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1480 io_u->acct_ddir = -1;
1482 assert(!(td->flags & TD_F_CHILD));
1483 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1485 } else if (td_async_processing(td)) {
1487 * We ran out, wait for async verify threads to finish and
1490 assert(!(td->flags & TD_F_CHILD));
1491 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1499 static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1501 if (!(td->flags & TD_F_TRIM_BACKLOG))
1504 if (td->trim_entries) {
1507 if (td->trim_batch) {
1510 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1511 td->last_ddir != DDIR_READ) {
1512 td->trim_batch = td->o.trim_batch;
1513 if (!td->trim_batch)
1514 td->trim_batch = td->o.trim_backlog;
1518 if (get_trim && get_next_trim(td, io_u))
1525 static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1527 if (!(td->flags & TD_F_VER_BACKLOG))
1530 if (td->io_hist_len) {
1533 if (td->verify_batch)
1535 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1536 td->last_ddir != DDIR_READ) {
1537 td->verify_batch = td->o.verify_batch;
1538 if (!td->verify_batch)
1539 td->verify_batch = td->o.verify_backlog;
1543 if (get_verify && !get_next_verify(td, io_u)) {
1553 * Fill offset and start time into the buffer content, to prevent too
1554 * easy compressible data for simple de-dupe attempts. Do this for every
1555 * 512b block in the range, since that should be the smallest block size
1556 * we can expect from a device.
1558 static void small_content_scramble(struct io_u *io_u)
1560 unsigned int i, nr_blocks = io_u->buflen / 512;
1562 unsigned int offset;
1569 boffset = io_u->offset;
1570 io_u->buf_filled_len = 0;
1572 for (i = 0; i < nr_blocks; i++) {
1574 * Fill the byte offset into a "random" start offset of
1575 * the buffer, given by the product of the usec time
1576 * and the actual offset.
1578 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1579 offset &= ~(sizeof(uint64_t) - 1);
1580 if (offset >= 512 - sizeof(uint64_t))
1581 offset -= sizeof(uint64_t);
1582 memcpy(p + offset, &boffset, sizeof(boffset));
1584 end = p + 512 - sizeof(io_u->start_time);
1585 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1592 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1593 * etc. The returned io_u is fully ready to be prepped and submitted.
1595 struct io_u *get_io_u(struct thread_data *td)
1599 int do_scramble = 0;
1602 io_u = __get_io_u(td);
1604 dprint(FD_IO, "__get_io_u failed\n");
1608 if (check_get_verify(td, io_u))
1610 if (check_get_trim(td, io_u))
1614 * from a requeue, io_u already setup
1620 * If using an iolog, grab next piece if any available.
1622 if (td->flags & TD_F_READ_IOLOG) {
1623 if (read_iolog_get(td, io_u))
1625 } else if (set_io_u_file(td, io_u)) {
1627 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1633 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1637 assert(fio_file_open(f));
1639 if (ddir_rw(io_u->ddir)) {
1640 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1641 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1645 f->last_start[io_u->ddir] = io_u->offset;
1646 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1648 if (io_u->ddir == DDIR_WRITE) {
1649 if (td->flags & TD_F_REFILL_BUFFERS) {
1650 io_u_fill_buffer(td, io_u,
1651 td->o.min_bs[DDIR_WRITE],
1653 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1654 !(td->flags & TD_F_COMPRESS))
1656 if (td->flags & TD_F_VER_NONE) {
1657 populate_verify_io_u(td, io_u);
1660 } else if (io_u->ddir == DDIR_READ) {
1662 * Reset the buf_filled parameters so next time if the
1663 * buffer is used for writes it is refilled.
1665 io_u->buf_filled_len = 0;
1670 * Set io data pointers.
1672 io_u->xfer_buf = io_u->buf;
1673 io_u->xfer_buflen = io_u->buflen;
1677 if (!td_io_prep(td, io_u)) {
1678 if (!td->o.disable_lat)
1679 fio_gettime(&io_u->start_time, NULL);
1681 small_content_scramble(io_u);
1685 dprint(FD_IO, "get_io_u failed\n");
1687 return ERR_PTR(ret);
1690 static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1692 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1694 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1697 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1698 io_u->file ? " on file " : "",
1699 io_u->file ? io_u->file->file_name : "",
1700 strerror(io_u->error),
1701 io_ddir_name(io_u->ddir),
1702 io_u->offset, io_u->xfer_buflen);
1704 if (td->io_ops->errdetails) {
1705 char *err = td->io_ops->errdetails(io_u);
1707 log_err("fio: %s\n", err);
1712 td_verror(td, io_u->error, "io_u error");
1715 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1717 __io_u_log_error(td, io_u);
1719 __io_u_log_error(td->parent, io_u);
1722 static inline bool gtod_reduce(struct thread_data *td)
1724 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1725 || td->o.gtod_reduce;
1728 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1729 struct io_completion_data *icd,
1730 const enum fio_ddir idx, unsigned int bytes)
1732 const int no_reduce = !gtod_reduce(td);
1733 unsigned long lusec = 0;
1739 lusec = utime_since(&io_u->issue_time, &icd->time);
1741 if (!td->o.disable_lat) {
1742 unsigned long tusec;
1744 tusec = utime_since(&io_u->start_time, &icd->time);
1745 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1747 if (td->flags & TD_F_PROFILE_OPS) {
1748 struct prof_io_ops *ops = &td->prof_io_ops;
1751 icd->error = ops->io_u_lat(td, tusec);
1754 if (td->o.max_latency && tusec > td->o.max_latency)
1755 lat_fatal(td, icd, tusec, td->o.max_latency);
1756 if (td->o.latency_target && tusec > td->o.latency_target) {
1757 if (lat_target_failed(td))
1758 lat_fatal(td, icd, tusec, td->o.latency_target);
1763 if (!td->o.disable_clat) {
1764 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1765 io_u_mark_latency(td, lusec);
1768 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1769 add_bw_sample(td, io_u, bytes, lusec);
1771 if (no_reduce && per_unit_log(td->iops_log))
1772 add_iops_sample(td, io_u, bytes);
1775 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1776 uint32_t *info = io_u_block_info(td, io_u);
1777 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1778 if (io_u->ddir == DDIR_TRIM) {
1779 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1780 BLOCK_INFO_TRIMS(*info) + 1);
1781 } else if (io_u->ddir == DDIR_WRITE) {
1782 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1789 static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1790 uint64_t offset, unsigned int bytes)
1797 if (f->first_write == -1ULL || offset < f->first_write)
1798 f->first_write = offset;
1799 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1800 f->last_write = offset + bytes;
1802 if (!f->last_write_comp)
1805 idx = f->last_write_idx++;
1806 f->last_write_comp[idx] = offset;
1807 if (f->last_write_idx == td->o.iodepth)
1808 f->last_write_idx = 0;
1811 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1812 struct io_completion_data *icd)
1814 struct io_u *io_u = *io_u_ptr;
1815 enum fio_ddir ddir = io_u->ddir;
1816 struct fio_file *f = io_u->file;
1818 dprint_io_u(io_u, "io complete");
1820 assert(io_u->flags & IO_U_F_FLIGHT);
1821 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1824 * Mark IO ok to verify
1828 * Remove errored entry from the verification list
1831 unlog_io_piece(td, io_u);
1833 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1838 if (ddir_sync(ddir)) {
1839 td->last_was_sync = 1;
1841 f->first_write = -1ULL;
1842 f->last_write = -1ULL;
1847 td->last_was_sync = 0;
1848 td->last_ddir = ddir;
1850 if (!io_u->error && ddir_rw(ddir)) {
1851 unsigned int bytes = io_u->buflen - io_u->resid;
1854 td->io_blocks[ddir]++;
1855 td->this_io_blocks[ddir]++;
1856 td->io_bytes[ddir] += bytes;
1858 if (!(io_u->flags & IO_U_F_VER_LIST))
1859 td->this_io_bytes[ddir] += bytes;
1861 if (ddir == DDIR_WRITE)
1862 file_log_write_comp(td, f, io_u->offset, bytes);
1864 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1865 td->runstate == TD_VERIFYING))
1866 account_io_completion(td, io_u, icd, ddir, bytes);
1868 icd->bytes_done[ddir] += bytes;
1871 ret = io_u->end_io(td, io_u_ptr);
1873 if (ret && !icd->error)
1876 } else if (io_u->error) {
1877 icd->error = io_u->error;
1878 io_u_log_error(td, io_u);
1881 enum error_type_bit eb = td_error_type(ddir, icd->error);
1883 if (!td_non_fatal_error(td, eb, icd->error))
1887 * If there is a non_fatal error, then add to the error count
1888 * and clear all the errors.
1890 update_error_count(td, icd->error);
1898 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1903 if (!gtod_reduce(td))
1904 fio_gettime(&icd->time, NULL);
1909 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1910 icd->bytes_done[ddir] = 0;
1913 static void ios_completed(struct thread_data *td,
1914 struct io_completion_data *icd)
1919 for (i = 0; i < icd->nr; i++) {
1920 io_u = td->io_ops->event(td, i);
1922 io_completed(td, &io_u, icd);
1930 * Complete a single io_u for the sync engines.
1932 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1934 struct io_completion_data icd;
1937 init_icd(td, &icd, 1);
1938 io_completed(td, &io_u, &icd);
1944 td_verror(td, icd.error, "io_u_sync_complete");
1948 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1949 td->bytes_done[ddir] += icd.bytes_done[ddir];
1955 * Called to complete min_events number of io for the async engines.
1957 int io_u_queued_complete(struct thread_data *td, int min_evts)
1959 struct io_completion_data icd;
1960 struct timespec *tvp = NULL;
1962 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1964 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1968 else if (min_evts > td->cur_depth)
1969 min_evts = td->cur_depth;
1971 /* No worries, td_io_getevents fixes min and max if they are
1972 * set incorrectly */
1973 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1975 td_verror(td, -ret, "td_io_getevents");
1980 init_icd(td, &icd, ret);
1981 ios_completed(td, &icd);
1983 td_verror(td, icd.error, "io_u_queued_complete");
1987 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1988 td->bytes_done[ddir] += icd.bytes_done[ddir];
1994 * Call when io_u is really queued, to update the submission latency.
1996 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1998 if (!td->o.disable_slat) {
1999 unsigned long slat_time;
2001 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
2006 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2012 * See if we should reuse the last seed, if dedupe is enabled
2014 static struct frand_state *get_buf_state(struct thread_data *td)
2018 if (!td->o.dedupe_percentage)
2019 return &td->buf_state;
2020 else if (td->o.dedupe_percentage == 100) {
2021 frand_copy(&td->buf_state_prev, &td->buf_state);
2022 return &td->buf_state;
2025 v = rand32_between(&td->dedupe_state, 1, 100);
2027 if (v <= td->o.dedupe_percentage)
2028 return &td->buf_state_prev;
2030 return &td->buf_state;
2033 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2035 if (td->o.dedupe_percentage == 100)
2036 frand_copy(rs, &td->buf_state_prev);
2037 else if (rs == &td->buf_state)
2038 frand_copy(&td->buf_state_prev, rs);
2041 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2042 unsigned int max_bs)
2044 struct thread_options *o = &td->o;
2046 if (o->compress_percentage || o->dedupe_percentage) {
2047 unsigned int perc = td->o.compress_percentage;
2048 struct frand_state *rs;
2049 unsigned int left = max_bs;
2050 unsigned int this_write;
2053 rs = get_buf_state(td);
2055 min_write = min(min_write, left);
2058 this_write = min_not_zero(min_write,
2059 td->o.compress_chunk);
2061 fill_random_buf_percentage(rs, buf, perc,
2062 this_write, this_write,
2064 o->buffer_pattern_bytes);
2066 fill_random_buf(rs, buf, min_write);
2067 this_write = min_write;
2072 save_buf_state(td, rs);
2074 } else if (o->buffer_pattern_bytes)
2075 fill_buffer_pattern(td, buf, max_bs);
2076 else if (o->zero_buffers)
2077 memset(buf, 0, max_bs);
2079 fill_random_buf(get_buf_state(td), buf, max_bs);
2083 * "randomly" fill the buffer contents
2085 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2086 unsigned int min_write, unsigned int max_bs)
2088 io_u->buf_filled_len = 0;
2089 fill_io_buffer(td, io_u->buf, min_write, max_bs);